[
{
"path": ".gitignore",
"content": ".vs/\nComLightLib/x64/\nWhisper/x64/\nx64/\nTools/CompressShaders/bin/\nTools/CompressShaders/obj/\nWhisper/D3D/shaderData-Debug.inl\nWhisper/D3D/shaderData-Release.inl\nWhisperNet/bin/\nWhisperNet/obj/\nExamples/TranscribeCS/bin/\nExamples/TranscribeCS/obj/\n*.aps\n*.json\n*.user\nExamples/MicrophoneCS/obj/\nExamples/MicrophoneCS/bin/\nTools/PerfSummary/bin/\nTools/PerfSummary/obj/\npackages/\nWhisperPS/obj/\nWhisperPS/bin/\nTools/CompressTables/bin/\nTools/CompressTables/obj/"
},
{
"path": "ComLightLib/ComLightLib.vcxproj",
"content": "\n\n \n \n Debug\n x64\n \n \n Release\n x64\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n 15.0\n {52F486E7-830C-45D8-BE47-E76B5AAB2772}\n Win32Proj\n ComLightLib\n 10.0\n \n \n \n StaticLibrary\n true\n v143\n Unicode\n \n \n StaticLibrary\n false\n v143\n true\n Unicode\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n true\n $(Platform)\\$(Configuration)\\\n \n \n false\n $(Platform)\\$(Configuration)\\\n \n \n \n NotUsing\n Level3\n Disabled\n true\n _DEBUG;_LIB;%(PreprocessorDefinitions)\n true\n stdcpp20\n AdvancedVectorExtensions\n \n \n Windows\n true\n \n \n \n \n NotUsing\n Level3\n MaxSpeed\n true\n true\n true\n NDEBUG;_LIB;%(PreprocessorDefinitions)\n true\n stdcpp20\n AdvancedVectorExtensions\n \n \n Windows\n true\n true\n true\n \n \n \n \n \n"
},
{
"path": "ComLightLib/ComLightLib.vcxproj.filters",
"content": "\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n"
},
{
"path": "ComLightLib/Exception.hpp",
"content": "#pragma once\n\nnamespace ComLight\n{\n\tclass Exception : public std::runtime_error\n\t{\n\t\t// I don't like C++ exceptions too much, but for some cases they are useful.\n\t\t// You can throw ComLight::Exception from constructor, or from FinalConstruct() method, the library will catch & return the code from the class factory function.\n\t\t// Unfortunately, for interface methods this doesn't work, the C++ parts of the library can't catch them without very complex trickery like code generation.\n\t\t// You can still use this class in methods, but you'll need to catch them manually near the API boundary or the app will crash.\n\t\t// C++ doesn't have an ABI, the framework can't catch C++ exception across the modules.\n\t\tconst HRESULT m_code;\n\n\tpublic:\n\n\t\tException( HRESULT hr ) : runtime_error( \"ComLight HRESULT exception\" ), m_code( hr ) { }\n\n\t\tHRESULT code() const { return m_code; }\n\t};\n}"
},
{
"path": "ComLightLib/Readme.txt",
"content": "Copy-pasted from there:\nhttps://github.com/Const-me/ComLightInterop/tree/master/ComLightLib\nWith only a few minor changes."
},
{
"path": "ComLightLib/client/CComPtr.hpp",
"content": "#pragma once\n\nnamespace ComLight\n{\n\t// COM smart pointer, very comparable to CComPtr from ATL\n\ttemplate \n\tclass CComPtr\n\t{\n\t\tI* p;\n\n\t\tvoid callAddRef() const\n\t\t{\n\t\t\tif( nullptr == p )\n\t\t\t\treturn;\n\t\t\tp->AddRef();\n\t\t}\n\n\tpublic:\n\n\t\t// Construct with nullptr\n\t\tCComPtr() : p( nullptr ) { }\n\n\t\t// Release the pointer\n\t\tvoid release()\n\t\t{\n\t\t\tif( nullptr == p )\n\t\t\t\treturn;\n\t\t\tp->Release();\n\t\t\tp = nullptr;\n\t\t}\n\n\t\t~CComPtr()\n\t\t{\n\t\t\trelease();\n\t\t}\n\n\t\t// Attach without AddRef()\n\t\tvoid attach( I* raw )\n\t\t{\n\t\t\trelease();\n\t\t\tp = raw;\n\t\t}\n\n\t\t// Detach without Release(), set this pointer to nullptr\n\t\tI* detach()\n\t\t{\n\t\t\tI* const result = p;\n\t\t\tp = nullptr;\n\t\t\treturn result;\n\t\t}\n\n\t\t// Detach without Release() and place to the specified address, set this pointer to nullptr\n\t\ttemplate\n\t\tvoid detach( Other** pp )\n\t\t{\n\t\t\t// If the argument points to a non-empty object, release the old instance: would leak memory otherwise.\n\t\t\tif( nullptr != *pp )\n\t\t\t\t( *pp )->Release();\n\t\t\t( *pp ) = detach();\n\t\t}\n\n\t\t// Set and AddRef()\n\t\tvoid assign( I* raw )\n\t\t{\n\t\t\trelease();\n\t\t\tattach( raw );\n\t\t\tcallAddRef();\n\t\t}\n\n\t\tvoid swap( CComPtr& that )\n\t\t{\n\t\t\tstd::swap( p, that.p );\n\t\t}\n\n\t\t// Set and AddRef()\n\t\tCComPtr( I* raw ) : p( raw )\n\t\t{\n\t\t\tcallAddRef();\n\t\t}\n\n\t\t// Set and AddRef()\n\t\tCComPtr( const CComPtr& that ) : CComPtr( that.p ) { }\n\t\t// Move constructor\n\t\tCComPtr( CComPtr&& that ) : p( that.p ) { that.p = nullptr; }\n\n\t\t// Set and AddRef()\n\t\tvoid operator=( I* raw )\n\t\t{\n\t\t\tassign( raw );\n\t\t}\n\n\t\t// Set and AddRef()\n\t\tvoid operator=( const CComPtr& that )\n\t\t{\n\t\t\tassign( that.p );\n\t\t}\n\n\t\t// Move assignment operator, destroys the other one\n\t\tvoid operator=( CComPtr&& that )\n\t\t{\n\t\t\tattach( that.detach() );\n\t\t}\n\n\t\toperator I*( ) const { return p; }\n\t\tI* operator -> () const { return p; }\n\t\tI** operator &() { return &p; }\n\n\t\toperator bool() const { return nullptr != p; }\n\t};\n}"
},
{
"path": "ComLightLib/comLightClient.h",
"content": "#pragma once\n#include \"comLightCommon.h\"\n#include \"client/CComPtr.hpp\"\n#include \"utils/typeTraits.hpp\"\n\nnamespace ComLight\n{\n\tnamespace details\n\t{\n\t\ttemplate\n\t\tinline constexpr void** castDoublePointerToVoid( T** pp )\n\t\t{\n\t\t\tstatic_assert( pointersAssignable(), \"IID_PPV_ARGS macro should be used with IUnknown interfaces\" );\n\t\t\treturn reinterpret_cast( pp );\n\t\t}\n\t}\n}\n\n#ifdef IID_PPV_ARGS\n#undef IID_PPV_ARGS\n#endif\n\n#define IID_PPV_ARGS( pp ) decltype( **pp )::iid, ::ComLight::details::castDoublePointerToVoid( pp )"
},
{
"path": "ComLightLib/comLightCommon.h",
"content": "#pragma once\n#include \"hresult.h\"\n\n#ifdef _MSC_VER\n#include \n#else\n#include \"pal/guiddef.h\"\nusing LPCTSTR = const char*;\n#endif\n\n#include \"unknwn.h\""
},
{
"path": "ComLightLib/comLightServer.h",
"content": "#pragma once\n#include \"comLightCommon.h\"\n#include \"client/CComPtr.hpp\"\n\n#include \"server/ObjectRoot.hpp\"\n#include \"server/interfaceMap.h\"\n#include \"server/Object.hpp\"\n#include \"server/freeThreadedMarshaller.h\"\n\n#ifdef _MSC_VER\n// On Windows, it's controlled by library.def module definition file. There's __declspec(dllexport), but it adds underscore, I don't like that.\n#define DLLEXPORT extern \"C\"\n#else\n#define DLLEXPORT extern \"C\" __attribute__((visibility(\"default\")))\n#endif"
},
{
"path": "ComLightLib/hresult.h",
"content": "#pragma once\n#include \n#ifdef _MSC_VER\n#include \n#include \n#else\n#include \"pal/hresult.h\"\n#endif\n\n#define CHECK( hr ) { const HRESULT __hr = ( hr ); if( FAILED( __hr ) ) return __hr; }\n\n#ifndef _MSC_VER\ninline constexpr HRESULT HRESULT_FROM_WIN32( int c )\n{\n\treturn c < 0 ? c : ( ( 0xFFFF & c ) | 0x80070000 );\n}\n\nconstexpr HRESULT OLE_E_BLANK = _HRESULT_TYPEDEF_( 0x80040007 );\nconstexpr HRESULT E_BOUNDS = _HRESULT_TYPEDEF_( 0x8000000BL ); \n\nconstexpr int ERROR_HANDLE_EOF = 38;\nconstexpr int ERROR_ALREADY_INITIALIZED = 1247;\n#endif\n\nconstexpr HRESULT E_EOF = HRESULT_FROM_WIN32( ERROR_HANDLE_EOF );\nconstexpr HRESULT E_ALREADY_INITIALIZED = HRESULT_FROM_WIN32( ERROR_ALREADY_INITIALIZED );"
},
{
"path": "ComLightLib/pal/guiddef.h",
"content": "#pragma once\n#include \n#include \n#ifndef GUID_DEFINED\n#define GUID_DEFINED\n#endif\n\nstruct GUID\n{\n\tuint32_t Data1;\n\tuint16_t Data2;\n\tuint16_t Data3;\n\tstd::array Data4;\n\n\tconstexpr inline bool operator==( const GUID& that ) const\n\t{\n\t\treturn Data1 == that.Data1 && Data2 == that.Data2 && Data3 == that.Data3 && Data4 == that.Data4;\n\t}\n};\n\nusing REFIID = const GUID&;"
},
{
"path": "ComLightLib/pal/hresult.h",
"content": "#pragma once\n#include \nusing HRESULT = int32_t;\n#define _HRESULT_TYPEDEF_(_sc) ((HRESULT)_sc)\n#define SEVERITY_ERROR 1\n#define FACILITY_CONTROL 10\n\ninline constexpr HRESULT MAKE_SCODE( uint32_t sev, uint32_t fac, uint32_t code )\n{\n\treturn (HRESULT)( ( (uint32_t)( sev ) << 31 ) | ( (unsigned long)( fac ) << 16 ) | ( (unsigned long)( code ) ) );\n};\n\n// ==== Copy-pasted from coreclr-master\\src\\pal\\inc\\rt\\palrt.h ====\n#define S_OK _HRESULT_TYPEDEF_(0x00000000L)\n#define S_FALSE _HRESULT_TYPEDEF_(0x00000001L)\n\n#define E_NOTIMPL _HRESULT_TYPEDEF_(0x80004001L)\n#define E_NOINTERFACE _HRESULT_TYPEDEF_(0x80004002L)\n#define E_UNEXPECTED _HRESULT_TYPEDEF_(0x8000FFFFL)\n#define E_OUTOFMEMORY _HRESULT_TYPEDEF_(0x8007000EL)\n#define E_INVALIDARG _HRESULT_TYPEDEF_(0x80070057L)\n#define E_POINTER _HRESULT_TYPEDEF_(0x80004003L)\n#define E_HANDLE _HRESULT_TYPEDEF_(0x80070006L)\n#define E_ABORT _HRESULT_TYPEDEF_(0x80004004L)\n#define E_FAIL _HRESULT_TYPEDEF_(0x80004005L)\n#define E_ACCESSDENIED _HRESULT_TYPEDEF_(0x80070005L)\n#define E_PENDING _HRESULT_TYPEDEF_(0x8000000AL)\n\n#define DISP_E_PARAMNOTFOUND _HRESULT_TYPEDEF_(0x80020004L)\n#define DISP_E_TYPEMISMATCH _HRESULT_TYPEDEF_(0x80020005L)\n#define DISP_E_BADVARTYPE _HRESULT_TYPEDEF_(0x80020008L)\n#define DISP_E_OVERFLOW _HRESULT_TYPEDEF_(0x8002000AL)\n#define DISP_E_DIVBYZERO _HRESULT_TYPEDEF_(0x80020012L)\n\n#define CLASS_E_CLASSNOTAVAILABLE _HRESULT_TYPEDEF_(0x80040111L)\n#define CLASS_E_NOAGGREGATION _HRESULT_TYPEDEF_(0x80040110L)\n\n#define CO_E_CLASSSTRING _HRESULT_TYPEDEF_(0x800401F3L)\n\n#define MK_E_SYNTAX _HRESULT_TYPEDEF_(0x800401E4L)\n\n#define STG_E_INVALIDFUNCTION _HRESULT_TYPEDEF_(0x80030001L)\n#define STG_E_FILENOTFOUND _HRESULT_TYPEDEF_(0x80030002L)\n#define STG_E_PATHNOTFOUND _HRESULT_TYPEDEF_(0x80030003L)\n#define STG_E_WRITEFAULT _HRESULT_TYPEDEF_(0x8003001DL)\n#define STG_E_FILEALREADYEXISTS _HRESULT_TYPEDEF_(0x80030050L)\n#define STG_E_ABNORMALAPIEXIT _HRESULT_TYPEDEF_(0x800300FAL)\n\n#define NTE_BAD_UID _HRESULT_TYPEDEF_(0x80090001L)\n#define NTE_BAD_HASH _HRESULT_TYPEDEF_(0x80090002L)\n#define NTE_BAD_KEY _HRESULT_TYPEDEF_(0x80090003L)\n#define NTE_BAD_LEN _HRESULT_TYPEDEF_(0x80090004L)\n#define NTE_BAD_DATA _HRESULT_TYPEDEF_(0x80090005L)\n#define NTE_BAD_SIGNATURE _HRESULT_TYPEDEF_(0x80090006L)\n#define NTE_BAD_VER _HRESULT_TYPEDEF_(0x80090007L)\n#define NTE_BAD_ALGID _HRESULT_TYPEDEF_(0x80090008L)\n#define NTE_BAD_FLAGS _HRESULT_TYPEDEF_(0x80090009L)\n#define NTE_BAD_TYPE _HRESULT_TYPEDEF_(0x8009000AL)\n#define NTE_BAD_KEY_STATE _HRESULT_TYPEDEF_(0x8009000BL)\n#define NTE_BAD_HASH_STATE _HRESULT_TYPEDEF_(0x8009000CL)\n#define NTE_NO_KEY _HRESULT_TYPEDEF_(0x8009000DL)\n#define NTE_NO_MEMORY _HRESULT_TYPEDEF_(0x8009000EL)\n#define NTE_SIGNATURE_FILE_BAD _HRESULT_TYPEDEF_(0x8009001CL)\n#define NTE_FAIL _HRESULT_TYPEDEF_(0x80090020L)\n\n#define CRYPT_E_HASH_VALUE _HRESULT_TYPEDEF_(0x80091007L)\n\n#define TYPE_E_SIZETOOBIG _HRESULT_TYPEDEF_(0x800288C5L)\n#define TYPE_E_DUPLICATEID _HRESULT_TYPEDEF_(0x800288C6L)\n\n#define STD_CTL_SCODE(n) MAKE_SCODE(SEVERITY_ERROR, FACILITY_CONTROL, n)\n#define CTL_E_OVERFLOW STD_CTL_SCODE(6)\n#define CTL_E_OUTOFMEMORY STD_CTL_SCODE(7)\n#define CTL_E_DIVISIONBYZERO STD_CTL_SCODE(11)\n#define CTL_E_OUTOFSTACKSPACE STD_CTL_SCODE(28)\n#define CTL_E_FILENOTFOUND STD_CTL_SCODE(53)\n#define CTL_E_DEVICEIOERROR STD_CTL_SCODE(57)\n#define CTL_E_PERMISSIONDENIED STD_CTL_SCODE(70)\n#define CTL_E_PATHFILEACCESSERROR STD_CTL_SCODE(75)\n#define CTL_E_PATHNOTFOUND STD_CTL_SCODE(76)\n\n#define INET_E_CANNOT_CONNECT _HRESULT_TYPEDEF_(0x800C0004L)\n#define INET_E_RESOURCE_NOT_FOUND _HRESULT_TYPEDEF_(0x800C0005L)\n#define INET_E_OBJECT_NOT_FOUND _HRESULT_TYPEDEF_(0x800C0006L)\n#define INET_E_DATA_NOT_AVAILABLE _HRESULT_TYPEDEF_(0x800C0007L)\n#define INET_E_DOWNLOAD_FAILURE _HRESULT_TYPEDEF_(0x800C0008L)\n#define INET_E_CONNECTION_TIMEOUT _HRESULT_TYPEDEF_(0x800C000BL)\n#define INET_E_UNKNOWN_PROTOCOL _HRESULT_TYPEDEF_(0x800C000DL)\n\n#define DBG_PRINTEXCEPTION_C _HRESULT_TYPEDEF_(0x40010006L)\n// ==== Done pasting ====\n\ninline constexpr bool SUCCEEDED( HRESULT hr )\n{\n\treturn hr >= 0;\n}\n\ninline constexpr bool FAILED( HRESULT hr )\n{\n\treturn hr < 0;\n}"
},
{
"path": "ComLightLib/server/Object.hpp",
"content": "#pragma once\n#include \n#include \"../comLightClient.h\"\n#include \"../utils/typeTraits.hpp\"\n#include \"../Exception.hpp\"\n\nnamespace ComLight\n{\n\tnamespace details\n\t{\n\t\tGENERATE_HAS_MEMBER( implQueryInterface );\n\t\tGENERATE_HAS_MEMBER( implAddRef );\n\t\tGENERATE_HAS_MEMBER( implRelease );\n\t}\n\n\t// Outer class of objects, implements IUnknown methods, also the class factory. The type argument must be your class implementing your interfaces, inherited from ObjectRoot\n\ttemplate\n\tclass Object : public T\n\t{\n\tpublic:\n\t\tObject() = default;\n\n\t\ttemplate\n\t\tObject( Args&& ... args ) : T{ std::forward( args )... } {};\n\n\t\tinline virtual ~Object() override { }\n\n\t\t// Implement IUnknown methods\n\t\tHRESULT COMLIGHTCALL QueryInterface( REFIID riid, void** ppvObject ) override\n\t\t{\n\t\t\tstatic_assert( details::has_member_implQueryInterface::value, \"Your object class must inherit from ComLight::ObjectRoot\" );\n\n\t\t\tif( nullptr == ppvObject )\n\t\t\t\treturn E_POINTER;\n\n\t\t\tif( T::implQueryInterface( riid, ppvObject ) )\n\t\t\t\treturn S_OK;\n\t\t\tif( T::queryExtraInterfaces( riid, ppvObject ) )\n\t\t\t\treturn S_OK;\n\n\t\t\tif( riid == IUnknown::iid() )\n\t\t\t{\n\t\t\t\tComLight::IUnknown* unk = T::getUnknown();\n\t\t\t\tunk->AddRef();\n\t\t\t\t*ppvObject = unk;\n\t\t\t\treturn S_OK;\n\t\t\t}\n\n\t\t\treturn E_NOINTERFACE;\n\t\t}\n\n\t\tuint32_t COMLIGHTCALL AddRef() override\n\t\t{\n\t\t\tstatic_assert( details::has_member_implAddRef::value, \"Your object class must inherit from ComLight::ObjectRoot\" );\n\t\t\treturn T::implAddRef();\n\t\t}\n\n\t\tuint32_t COMLIGHTCALL Release() override\n\t\t{\n\t\t\tstatic_assert( details::has_member_implRelease::value, \"Your object class must inherit from ComLight::ObjectRoot\" );\n\t\t\tconst uint32_t ret = T::implRelease();\n\t\t\tif( 0 == ret )\n\t\t\t{\n\t\t\t\tT::FinalRelease();\n\t\t\t\tdelete this;\n\t\t\t}\n\t\t\treturn ret;\n\t\t}\n\n\t\t// Create a new object on the heap, store in smart pointer\n\t\tstatic inline HRESULT create( CComPtr>& result )\n\t\t{\n\t\t\tCComPtr> ptr;\n\t\t\ttry\n\t\t\t{\n\t\t\t\tptr = new Object();\t// The RefCounter constructor creates it with ref.counter 0. But then CComPtr constructor calls AddRef so we have RC=1 after this line.\n\n\t\t\t\tHRESULT hr = ptr->internalFinalConstruct();\n\t\t\t\tif( FAILED( hr ) )\n\t\t\t\t\treturn hr;\n\n\t\t\t\thr = ptr->FinalConstruct();\n\t\t\t\tif( FAILED( hr ) )\n\t\t\t\t\treturn hr;\n\n\t\t\t\tptr.swap( result );\n\t\t\t\treturn S_OK;\n\t\t\t}\n\t\t\tcatch( const Exception& ex )\n\t\t\t{\n\t\t\t\treturn ex.code();\n\t\t\t}\n\t\t}\n\n\t\t// Create a new object on the heap, store in smart pointer\n\t\ttemplate\n\t\tstatic inline HRESULT create( CComPtr>& result, Args&& ... args )\n\t\t{\n\t\t\tCComPtr> ptr;\n\t\t\ttry\n\t\t\t{\n\t\t\t\tptr = new Object( std::forward( args )... );\n\n\t\t\t\tHRESULT hr = ptr->internalFinalConstruct();\n\t\t\t\tif( FAILED( hr ) )\n\t\t\t\t\treturn hr;\n\n\t\t\t\thr = ptr->FinalConstruct();\n\t\t\t\tif( FAILED( hr ) )\n\t\t\t\t\treturn hr;\n\n\t\t\t\tptr.swap( result );\n\t\t\t\treturn S_OK;\n\t\t\t}\n\t\t\tcatch( const Exception& ex )\n\t\t\t{\n\t\t\t\treturn ex.code();\n\t\t\t}\n\t\t\tcatch( HRESULT hr )\n\t\t\t{\n\t\t\t\treturn hr;\n\t\t\t}\n\t\t}\n\n\t\t// Create a new object on the heap, return one of it's interfaces. The caller is assumed to take ownership of the new object.\n\t\ttemplate\n\t\tstatic inline HRESULT create( I** pp )\n\t\t{\n\t\t\tif( pp == nullptr )\n\t\t\t\treturn E_POINTER;\n\n\t\t\tstatic_assert( details::pointersAssignable(), \"Object::create can't cast object to the requested interface\" );\n\t\t\tCComPtr> ptr;\n\t\t\tCHECK( create( ptr ) );\n\t\t\tptr.detach( pp );\n\t\t\treturn S_OK;\n\t\t}\n\t};\n}"
},
{
"path": "ComLightLib/server/ObjectRoot.hpp",
"content": "#pragma once\n#include \"RefCounter.hpp\"\n#include \"../comLightCommon.h\"\n#include \"../utils/typeTraits.hpp\"\n\nnamespace ComLight\n{\n\t// Base class of objects, implements reference counting, also a few lifetime methods.\n\t// The template argument is the interface you want clients to get when they ask for IID_IUnknown. By convention, that pointer defines object's identity.\n\ttemplate\n\tclass ObjectRoot : public RefCounter, public I\n\t{\n\tprotected:\n\n\t\tinline HRESULT internalFinalConstruct()\n\t\t{\n\t\t\treturn S_FALSE;\n\t\t}\n\n\t\tinline HRESULT FinalConstruct()\n\t\t{\n\t\t\treturn S_FALSE;\n\t\t}\n\n\t\tinline void FinalRelease() { }\n\n\t\tIUnknown* getUnknown()\n\t\t{\n\t\t\tstatic_assert( details::pointersAssignable(), \"The interface doesn't derive from IUnknown\" );\n\t\t\treturn static_cast( this );\n\t\t}\n\n\t\tbool queryExtraInterfaces( REFIID riid, void **ppvObject ) const\n\t\t{\n\t\t\treturn false;\n\t\t}\n\n\t\t// Implement query interface with 2 entries, IUnknown and I.\n\t\tbool implQueryInterface( REFIID riid, void** ppvObject )\n\t\t{\n\t\t\tif( riid == I::iid() || riid == IUnknown::iid() )\n\t\t\t{\n\t\t\t\tI* const result = this;\n\t\t\t\tresult->AddRef();\n\t\t\t\t*ppvObject = result;\n\t\t\t\treturn true;\n\t\t\t}\n\t\t\treturn false;\n\t\t}\n\t};\n}"
},
{
"path": "ComLightLib/server/RefCounter.hpp",
"content": "#pragma once\n#include \n#include \n#include \n\nnamespace ComLight\n{\n\t// Very base class of objects, implements reference counting.\n\tclass RefCounter\n\t{\n\t\tstd::atomic_uint referenceCounter;\n\n\tpublic:\n\n\t\tRefCounter() : referenceCounter( 0 ) { }\n\n\t\tinline virtual ~RefCounter() { }\n\n\t\tRefCounter( const RefCounter &that ) = delete;\n\t\tRefCounter( RefCounter &&that ) = delete;\n\n\tprotected:\n\n\t\tuint32_t implAddRef()\n\t\t{\n\t\t\treturn ++referenceCounter;\n\t\t}\n\n\t\tuint32_t implRelease()\n\t\t{\n\t\t\t// Might be a good idea to use locks, at least in debug builds. They're much slower than atomics, but with locks it's possible to detect when 2 threads call release at the same time, for object with counter = 1.\n\t\t\t// It's a memory management bug, but it would be nice if debug builds would handle that case gracefully.\n\t\t\tconst uint32_t rc = --referenceCounter;\n\t\t\tassert( rc != UINT_MAX );\n\t\t\treturn rc;\n\t\t}\n\t};\n}"
},
{
"path": "ComLightLib/server/freeThreadedMarshaller.cpp",
"content": "#include \"freeThreadedMarshaller.h\"\n#ifdef _MSC_VER\n#include \n\nHRESULT ComLight::details::createFreeThreadedMarshaller( IUnknown* pUnkOuter, IUnknown** ppUnkMarshal )\n{\n\treturn ::CoCreateFreeThreadedMarshaler( (LPUNKNOWN)pUnkOuter, (LPUNKNOWN *)ppUnkMarshal );\n}\n\nbool ComLight::details::queryMarshallerInterface( REFIID riid, void **ppvObject, IUnknown* marshaller )\n{\n\tif( riid != IID_IMarshal || nullptr == marshaller )\n\t\treturn false;\n\tconst HRESULT hr = marshaller->QueryInterface( IID_IMarshal, ppvObject );\n\treturn SUCCEEDED( hr ) ? true : false;\n}\n#endif"
},
{
"path": "ComLightLib/server/freeThreadedMarshaller.h",
"content": "#pragma once\n#ifdef _MSC_VER\n#include \"../comLightCommon.h\"\n\nnamespace ComLight\n{\n\tnamespace details\n\t{\n\t\tHRESULT createFreeThreadedMarshaller( IUnknown* pUnkOuter, IUnknown** ppUnkMarshal );\n\t\tbool queryMarshallerInterface( REFIID riid, void **ppvObject, IUnknown* marshaller );\n\t}\n}\n\n#define DECLARE_FREE_THREADED_MARSHALLER() \\\nprivate: \\\nComLight::CComPtr m_freeThreadedMarshaller; \\\nprotected: \\\nHRESULT internalFinalConstruct() \\\n{ \\\n\treturn ComLight::details::createFreeThreadedMarshaller( getUnknown(), &m_freeThreadedMarshaller ); \\\n} \\\nbool queryExtraInterfaces( REFIID riid, void **ppvObject ) const \\\n{ \\\n\treturn ComLight::details::queryMarshallerInterface( riid, ppvObject, m_freeThreadedMarshaller ); \\\n}\n\n#else\n#define DECLARE_FREE_THREADED_MARSHALLER()\n#endif"
},
{
"path": "ComLightLib/server/interfaceMap.h",
"content": "#pragma once\n#include \"../utils/typeTraits.hpp\"\n\n// Unlike ATL, the interface map is optional for ComLight.\n// If you won't declare a map, the object will support 2 interfaces: IUnknown, and whatever template argument was passed to ObjectRoot class.\n#define BEGIN_COM_MAP() \\\nprotected: \\\nbool implQueryInterface( REFIID iid, void** ppvObject ) {\n\n#define END_COM_MAP() return false; }\n\nnamespace ComLight\n{\n\tnamespace details\n\t{\n\t\ttemplate\n\t\tinline bool tryReturnInterface( REFIID iid, C* pThis, void** ppvResult )\n\t\t{\n\t\t\tstatic_assert( pointersAssignable(), \"Trying to implement an interface that doesn't derive from IUnknown\" );\n\t\t\tstatic_assert( pointersAssignable(), \"Declared support for an interface, but the class doesn't implement it\" );\n\t\t\tif( I::iid() != iid )\n\t\t\t\treturn false;\n\t\t\tI* const result = pThis;\n\t\t\tresult->AddRef();\n\t\t\t*ppvResult = result;\n\t\t\treturn true;\n\t\t}\n\t}\n}\n\n#define COM_INTERFACE_ENTRY( I ) if( ComLight::details::tryReturnInterface( iid, this, ppvObject ) ) return true;"
},
{
"path": "ComLightLib/streams.h",
"content": "#pragma once\n#include \n#include \"comLightCommon.h\"\n\n// COM interfaces to marshal streams across the interop.\nnamespace ComLight\n{\n\tenum struct eSeekOrigin : uint8_t\n\t{\n\t\tBegin = 0,\n\t\tCurrent = 1,\n\t\tEnd = 2\n\t};\n\n\tnamespace details\n\t{\n\t\ttemplate\n\t\tinline size_t sizeofVector( const std::vector& vec )\n\t\t{\n\t\t\treturn sizeof( E ) * vec.size();\n\t\t}\n\t}\n\n\t// COM interface for readonly stream. You'll get these interfaces what you use [ReadStream] attribute in C#.\n\tstruct DECLSPEC_NOVTABLE iReadStream : public IUnknown\n\t{\n\t\tDEFINE_INTERFACE_ID( \"006af6db-734e-4595-8c94-19304b2389ac\" );\n\n\t\tvirtual HRESULT COMLIGHTCALL read( void* lpBuffer, int nNumberOfBytesToRead, int &lpNumberOfBytesRead ) = 0;\n\t\tvirtual HRESULT COMLIGHTCALL seek( int64_t offset, eSeekOrigin origin ) = 0;\n\t\tvirtual HRESULT COMLIGHTCALL getPosition( int64_t& position ) = 0;\n\t\tvirtual HRESULT COMLIGHTCALL getLength( int64_t& length ) = 0;\n\n\t\ttemplate\n\t\tinline HRESULT read( std::vector& vec )\n\t\t{\n\t\t\tconst int cb = (int)details::sizeofVector( vec );\n\t\t\tint cbRead = 0;\n\t\t\tCHECK( read( vec.data(), cb, cbRead ) );\n\t\t\tif( cbRead >= cb )\n\t\t\t\treturn S_OK;\n\t\t\treturn E_EOF;\n\t\t}\n\t};\n\n\t// COM interface for readonly stream. You'll get these interfaces what you use [WriteStream] attribute in C#.\n\tstruct DECLSPEC_NOVTABLE iWriteStream : public IUnknown\n\t{\n\t\tDEFINE_INTERFACE_ID( \"d7c3eb39-9170-43b9-ba98-2ea1f2fed8a8\" );\n\n\t\tvirtual HRESULT COMLIGHTCALL write( const void* lpBuffer, int nNumberOfBytesToWrite ) = 0;\n\t\tvirtual HRESULT COMLIGHTCALL flush() = 0;\n\n\t\ttemplate\n\t\tinline HRESULT write( const std::vector& vec )\n\t\t{\n\t\t\tconst int cb = (int)details::sizeofVector( vec );\n\t\t\treturn write( vec.data(), cb );\n\t\t}\n\t};\n}"
},
{
"path": "ComLightLib/unknwn.h",
"content": "#pragma once\n#include \n\n// Calling conventions\n#ifdef _MSC_VER\n#define COMLIGHTCALL __stdcall\n#define DECLSPEC_NOVTABLE __declspec(novtable)\n#elif defined(__GNUC__) || defined(__clang__)\n#if defined(__i386__)\n#define COMLIGHTCALL __attribute__((stdcall))\n#else\n#define COMLIGHTCALL\n#endif\n#define DECLSPEC_NOVTABLE\n#else\n#error Unsupported C++ compiler\n#endif\n\n#include \"utils/guid_parse.hpp\"\n\n#define DEFINE_INTERFACE_ID( guidString ) static constexpr GUID iid() { return ::ComLight::make_guid( guidString ); }\n\nnamespace ComLight\n{\n\t// This thing is binary compatible with IUnknown from Windows SDK. See DesktopClient demo project, it uses normal COM interop in .NET framework 4.7 to call my implementation.\n\tstruct DECLSPEC_NOVTABLE IUnknown\n\t{\n\t\tDEFINE_INTERFACE_ID( \"00000000-0000-0000-c000-000000000046\" );\n\n\t\tvirtual HRESULT COMLIGHTCALL QueryInterface( REFIID riid, void **ppvObject ) = 0;\n\n\t\tvirtual uint32_t COMLIGHTCALL AddRef() = 0;\n\n\t\tvirtual uint32_t COMLIGHTCALL Release() = 0;\n\t};\n}"
},
{
"path": "ComLightLib/utils/guid_parse.hpp",
"content": "// https://github.com/tobias-loew/constexpr-GUID-cpp-11\n\n//-------------------------------------------------------------------------------------------------------\n// constexpr GUID parsing\n// Written by Alexander Bessonov\n// Written by Tobias Loew\n//\n// Licensed under the MIT license.\n//-------------------------------------------------------------------------------------------------------\n\n#pragma once\n#include \n#include \n#include \n#include \n\n#if !defined(GUID_DEFINED)\n#define GUID_DEFINED\nstruct GUID {\n\tuint32_t Data1;\n\tuint16_t Data2;\n\tuint16_t Data3;\n\tuint8_t Data4[ 8 ];\n};\n#endif\n\nnamespace ComLight\n{\n\tnamespace details\n\t{\n\t\tconstexpr const size_t short_guid_form_length = 36;\t// XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX\n\t\tconstexpr const size_t long_guid_form_length = 38;\t// {XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX}\n\n\t\tconstexpr uint8_t parse_hex_digit( const char c )\n\t\t{\n\t\t\tusing namespace std::string_literals;\n\t\t\treturn\n\t\t\t\t( '0' <= c && c <= '9' )\n\t\t\t\t? c - '0'\n\t\t\t\t: ( 'a' <= c && c <= 'f' )\n\t\t\t\t? 10 + c - 'a'\n\t\t\t\t: ( 'A' <= c && c <= 'F' )\n\t\t\t\t? 10 + c - 'A'\n\t\t\t\t:\n\t\t\t\tthrow std::domain_error{ \"invalid character in GUID\"s };\n\t\t}\n\n\t\tconstexpr uint8_t parse_hex_uint8_t( const char *ptr )\n\t\t{\n\t\t\treturn ( parse_hex_digit( ptr[ 0 ] ) << 4 ) + parse_hex_digit( ptr[ 1 ] );\n\t\t}\n\n\t\tconstexpr uint16_t parse_hex_uint16_t( const char *ptr )\n\t\t{\n\t\t\treturn ( parse_hex_uint8_t( ptr ) << 8 ) + parse_hex_uint8_t( ptr + 2 );\n\t\t}\n\n\t\tconstexpr uint32_t parse_hex_uint32_t( const char *ptr )\n\t\t{\n\t\t\treturn ( parse_hex_uint16_t( ptr ) << 16 ) + parse_hex_uint16_t( ptr + 4 );\n\t\t}\n\n\t\tconstexpr GUID parse_guid( const char *begin )\n\t\t{\n\t\t\treturn GUID{\n\t\t\t\tparse_hex_uint32_t( begin ),\n\t\t\t\tparse_hex_uint16_t( begin + 8 + 1 ),\n\t\t\t\tparse_hex_uint16_t( begin + 8 + 1 + 4 + 1 ),\n\t\t\t\t{\n\t\t\t\t\tparse_hex_uint8_t( begin + 8 + 1 + 4 + 1 + 4 + 1 ),\n\t\t\t\t\tparse_hex_uint8_t( begin + 8 + 1 + 4 + 1 + 4 + 1 + 2 ),\n\t\t\t\t\tparse_hex_uint8_t( begin + 8 + 1 + 4 + 1 + 4 + 1 + 2 + 2 + 1 ),\n\t\t\t\t\tparse_hex_uint8_t( begin + 8 + 1 + 4 + 1 + 4 + 1 + 2 + 2 + 1 + 2 ),\n\t\t\t\t\tparse_hex_uint8_t( begin + 8 + 1 + 4 + 1 + 4 + 1 + 2 + 2 + 1 + 2 + 2 ),\n\t\t\t\t\tparse_hex_uint8_t( begin + 8 + 1 + 4 + 1 + 4 + 1 + 2 + 2 + 1 + 2 + 2 + 2 ),\n\t\t\t\t\tparse_hex_uint8_t( begin + 8 + 1 + 4 + 1 + 4 + 1 + 2 + 2 + 1 + 2 + 2 + 2 + 2 ),\n\t\t\t\t\tparse_hex_uint8_t( begin + 8 + 1 + 4 + 1 + 4 + 1 + 2 + 2 + 1 + 2 + 2 + 2 + 2 + 2 )\n\t\t\t\t}\n\t\t\t};\n\t\t}\n\n\t\tconstexpr GUID make_guid_helper( const char *str, size_t N )\n\t\t{\n\t\t\tusing namespace std::string_literals;\n\t\t\tusing namespace details;\n\n\t\t\treturn ( !( N == long_guid_form_length || N == short_guid_form_length ) )\n\t\t\t\t? throw std::domain_error{ \"String GUID of the form {XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX} or XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX is expected\"s }\n\t\t\t\t: ( N == long_guid_form_length && ( str[ 0 ] != '{' || str[ long_guid_form_length - 1 ] != '}' ) )\n\t\t\t\t? throw std::domain_error{ \"Missing opening or closing brace\"s }\n\n\t\t\t: parse_guid( str + ( N == long_guid_form_length ? 1 : 0 ) );\n\t\t}\n\n\n\t\ttemplate\n\t\tconstexpr GUID make_guid( const char( &str )[ N ] )\n\t\t{\n\t\t\treturn make_guid_helper( str, N - 1 );\n\t\t}\n\t}\n\tusing details::make_guid;\n}"
},
{
"path": "ComLightLib/utils/typeTraits.hpp",
"content": "#pragma once\n#include \n\nnamespace ComLight\n{\n\tnamespace details\n\t{\n\t\ttemplate\n\t\tconstexpr bool pointersAssignable()\n\t\t{\n\t\t\t// See this for why `&` is required: https://stackoverflow.com/a/52429468/126995\n\t\t\treturn std::is_assignable::value;\n\t\t}\n\t}\n}\n\n// https://en.wikibooks.org/wiki/More_C++_Idioms/Member_Detector\n#define GENERATE_HAS_MEMBER(member) \\\n \\\ntemplate < class T > \\\nclass HasMember_##member \\\n{ \\\nprivate: \\\n using Yes = char[2]; \\\n using No = char[1]; \\\n \\\n struct Fallback { int member; }; \\\n struct Derived : T, Fallback { }; \\\n \\\n template < class U > \\\n static No& test ( decltype(U::member)* ); \\\n template < typename U > \\\n static Yes& test ( U* ); \\\n \\\npublic: \\\n static constexpr bool RESULT = sizeof(test(nullptr)) == sizeof(Yes); \\\n}; \\\n \\\ntemplate < class T > \\\nstruct has_member_##member \\\n: public std::integral_constant::RESULT> \\\n{ \\\n};\n"
},
{
"path": "ComputeShaders/ComputeShaders.cpp",
"content": "void fnComputeShaders()\n{\n}"
},
{
"path": "ComputeShaders/ComputeShaders.vcxproj",
"content": "\n\n \n \n Release\n x64\n \n \n Debug\n x64\n \n \n Release\n x64\n \n \n Debug\n x64\n \n \n \n 16.0\n Win32Proj\n {1c39d386-96d0-47a1-bbfa-68bbdb24439c}\n ComputeShaders\n 10.0\n \n \n \n StaticLibrary\n false\n v143\n true\n Unicode\n \n \n StaticLibrary\n false\n v143\n true\n Unicode\n \n \n StaticLibrary\n false\n v143\n true\n Unicode\n \n \n StaticLibrary\n false\n v143\n true\n Unicode\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n true\n $(Platform)\\$(Configuration)\\\n \n \n $(Platform)\\$(Configuration)\\\n true\n \n \n \n Level3\n true\n true\n true\n NDEBUG;_LIB;%(PreprocessorDefinitions)\n true\n \n \n \n \n true\n true\n true\n \n \n \n \n Level3\n true\n true\n true\n NDEBUG;_LIB;%(PreprocessorDefinitions)\n true\n \n \n \n \n true\n true\n true\n \n \n \n \n Level3\n true\n true\n true\n NDEBUG;_LIB;%(PreprocessorDefinitions)\n true\n \n \n \n \n true\n true\n true\n \n \n 5.0\n Compute\n \n \n \n \n Level3\n true\n true\n true\n NDEBUG;_LIB;%(PreprocessorDefinitions)\n true\n \n \n \n \n true\n true\n true\n \n \n 5.0\n Compute\n true\n true\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n"
},
{
"path": "ComputeShaders/ComputeShaders.vcxproj.filters",
"content": "\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n"
},
{
"path": "ComputeShaders/Readme.txt",
"content": "This project compiles all the compute shaders which implement the model.\n\nMany shaders come in 2 versions, something.hlsl and something64.hlsl\n\nThe version with the `64` suffix is used on AMD GPUs, the version without suffix is used on nVidia and Intel GPUs.\n\nNot all of these shaders are actually used for anything.\nSome of them are implementing binary compatibility for the reference CPU version, and not used unless messing with the `constexpr` flags in MlContext C++ class.\nSuch shaders often require FP64 support, which is an optional feature in D3D11.\nCompressShaders tool detects such shaders by looking at the SFI0 chunk in the binary, and outputs a bitmap of the FP64 shaders.\nThis way, missing FP64 hardware support shouldn’t break the library."
},
{
"path": "ComputeShaders/add.hlsl",
"content": "inline float compute( float a, float b )\n{\n\treturn a + b;\n}\n\n#include \"componentwiseBinaryOp.hlsli\""
},
{
"path": "ComputeShaders/addInPlace.hlsl",
"content": "#ifndef THREADS\n#define THREADS 512\n#endif\n\nBuffer arg0: register( t0 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 size: packoffset( c0 );\n\tuint4 strides: packoffset( c1 );\n\tuint4 argStrides: packoffset( c3 );\n}\n\ninline uint rowOffset( uint3 idx, uint4 strides )\n{\n\treturn idx[ 0 ] * strides[ 1 ] + idx[ 1 ] * strides[ 2 ] + idx[ 2 ] * strides[ 3 ];\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint rdi = rowOffset( group, strides );\n\tuint rsi = rowOffset( group, argStrides );\n\n\tconst uint rdiEnd = rdi + size[ 0 ] * strides[ 0 ];\n\trdi += thread * strides[ 0 ];\n\trsi += thread * argStrides[ 0 ];\n\n\tconst uint rdiInc = THREADS * strides[ 0 ];\n\tconst uint rsiInc = THREADS * argStrides[ 0 ];\n\n\tfor( ; rdi < rdiEnd; rdi += rdiInc, rsi += rsiInc )\n\t{\n\t\tfloat f = result[ rdi ];\n\t\tf += arg0[ rsi ];\n\t\tresult[ rdi ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/addRepeat.hlsl",
"content": "// Compute tensor = tensor + repeat( pattern, tensor ) in 1 shot, without VRAM allocations\n// Dispatch [ nb[ 1 ], nb[ 2 ], nb[ 3 ] ] thread groups of this shader, where nb is size of the destination tensor\nRWBuffer tensor: register( u0 );\nBuffer pattern: register( t0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 tensorSize: packoffset( c0 );\n\tuint4 tensorStrides: packoffset( c1 );\n\tuint4 patternSize: packoffset( c2 );\n\tuint4 patternStrides: packoffset( c3 );\n}\n\n#ifndef THREADS\n#define THREADS 256\n#endif\n\n#include \"repeatUtils.hlsli\"\n\ninline void computeSimple( uint idx, float add )\n{\n\tfloat f = tensor[ idx ];\n\tf += add;\n\ttensor[ idx ] = f;\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint3 it = tensorIteratorState( group, thread, tensorSize, tensorStrides );\n\tuint rsi = rowOffset( group % patternSize.yzw, patternStrides );\n\n\tif( patternSize[ 0 ] == 1 )\n\t{\n\t\t// The pattern only has 1 column - broadcasting over the row\n\t\tconst float p = pattern[ rsi ];\n\t\tROW_LOOP( it )\n\t\t\tcomputeSimple( it.x, p );\n\t}\n\telse if( patternSize[ 0 ] <= THREADS )\n\t{\n\t\t// pattern size doesn't exceed thread group size: load pattern value outside of the loop\n\t\tconst uint threadsPerGroup = THREADS - ( THREADS % patternSize[ 0 ] );\n\t\tif( thread >= threadsPerGroup )\n\t\t\treturn;\n\n\t\tconst float p = pattern[ rsi + ( thread % patternSize[ 0 ] ) * patternStrides[ 0 ] ];\n\t\tROW_LOOP_EX( it, threadsPerGroup, tensorStrides )\n\t\t\tcomputeSimple( it.x, p );\n\t}\n\telse\n\t{\n\t\t// Pattern rows are larger than the thread group, need to stream from both buffers\n\t\tconst uint rsiInc = THREADS * patternStrides[ 0 ];\n\t\tconst uint rsiDec = patternSize[ 0 ] * patternStrides[ 0 ];\n\t\tconst uint rsiEnd = rsi + rsiDec;\n\t\trsi += thread * patternStrides[ 0 ];\n\n\t\tROW_LOOP( it )\n\t\t{\n\t\t\tfloat f = tensor[ it.x ];\n\t\t\tfloat p = pattern[ rsi ];\n\t\t\trsi += rsiInc;\n\t\t\tif( rsi >= rsiEnd )\n\t\t\t\trsi -= rsiDec;\n\t\t\tf += p;\n\t\t\ttensor[ it.x ] = f;\n\t\t}\n\t}\n}"
},
{
"path": "ComputeShaders/addRepeat64.hlsl",
"content": "#define THREADS 64\n#include \"addRepeat.hlsl\""
},
{
"path": "ComputeShaders/addRepeatEx.hlsl",
"content": "// An equivalent of \"addRepeat.hlsl\" followed by \"addInPlace.hlsl\".\n// Merging into a single shader saves some global memory bandwidth and reduces CPU overhead wasted binding resources and dispatching shaders\nRWBuffer tensor: register( u0 );\nBuffer pattern: register( t0 );\nBuffer finalAdd: register( t1 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 tensorSize: packoffset( c0 );\n\tuint4 tensorStrides: packoffset( c1 );\n\tuint4 patternSize: packoffset( c2 );\n\tuint4 patternStrides: packoffset( c3 );\n\t// uint4 finalSize: packoffset( c4 );\n\tuint4 finalStrides: packoffset( c5 );\n}\n\n#ifndef THREADS\n#define THREADS 256\n#endif\n\n#include \"repeatUtils.hlsli\"\n\n// The micro-kernel of the shader, computes tensor[ rsi.x ] += pattern + finalAdd[ rsi.y ]\ninline void add2( uint2 rsi, float pattern )\n{\n\tfloat f = tensor[ rsi.x ];\n\tf += pattern;\n\tf += finalAdd[ rsi.y ];\n\ttensor[ rsi.x ] = f;\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint2 stridesX = uint2( tensorStrides.x, finalStrides.x );\n\tuint2 rsi;\n\trsi.x = rowOffset( group, tensorStrides );\n\trsi.y = rowOffset( group, finalStrides );\n\tconst uint rsiEnd = rsi.x + tensorSize.x * stridesX.x;\n\trsi += stridesX * thread;\n\n\tuint pat = rowOffset( group % patternSize.yzw, patternStrides );\n\n\tif( patternSize.x == 1 )\n\t{\n\t\t// The pattern only has 1 column, broadcasting over the row\n\t\tconst uint2 rsiInc = stridesX * THREADS;\n\t\tconst float p = pattern[ pat ];\n\t\tfor( ; rsi.x < rsiEnd; rsi += rsiInc )\n\t\t\tadd2( rsi, p );\n\t}\n\telse if( patternSize.x <= THREADS )\n\t{\n\t\t// pattern size doesn't exceed thread group size, load outside of the loop\n\t\tconst uint threadsPerGroup = THREADS - ( THREADS % patternSize.x );\n\t\tif( thread >= threadsPerGroup )\n\t\t\treturn;\n\n\t\tconst uint2 rsiInc = stridesX * threadsPerGroup;\n\t\tpat += ( thread % patternSize.x ) * patternStrides.x;\n\t\tconst float p = pattern[ pat ];\n\t\tfor( ; rsi.x < rsiEnd; rsi += rsiInc )\n\t\t\tadd2( rsi, p );\n\t}\n\telse\n\t{\n\t\t// Pattern rows are longer than the thread group, need to stream from both buffers\n\t\tuint3 rsi3;\n\t\trsi3.xy = rsi;\n\t\trsi3.z = pat + thread * patternStrides.x;\n\n\t\tconst uint3 rsiInc = uint3( stridesX, patternStrides.x ) * THREADS;\n\t\twhile( rsi3.x < rsiEnd )\n\t\t{\n\t\t\tadd2( rsi3.xy, pattern[ rsi3.z ] );\n\n\t\t\trsi3 += rsiInc;\n\t\t\tif( rsi3.z >= patternSize.x )\n\t\t\t\trsi3.z -= patternSize.x;\n\t\t}\n\t}\n}"
},
{
"path": "ComputeShaders/addRepeatGelu.hlsl",
"content": "// Compute tensor = GELU( tensor + repeat( pattern, tensor ) ) in 1 shot, without VRAM allocations\n// Dispatch [ nb[ 1 ], nb[ 2 ], nb[ 3 ] ] thread groups of this shader, where nb is size of the destination tensor\nRWBuffer tensor: register( u0 );\nBuffer pattern: register( t0 );\nBuffer lookupTable: register( t1 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 tensorSize: packoffset( c0 );\n\tuint4 tensorStrides: packoffset( c1 );\n\tuint4 patternSize: packoffset( c2 );\n\tuint4 patternStrides: packoffset( c3 );\n}\n\n#ifndef THREADS\n#define THREADS 1024\n#endif\n\n#include \"repeatUtils.hlsli\"\n#include \"miscUtils.hlsli\"\n\ninline float gelu( float x )\n{\n#if 1\n\tconst uint index = fp16Rounded( x );\n\tconst uint res16 = lookupTable[ index ];\n\treturn f16tof32( res16 );\n#else\n\t// This version is much slower, at least on AMD, despite saving these VRAM loads.\n\tconst float GELU_COEF_A = 0.044715;\n\tconst float SQRT_2_OVER_PI = 0.79788456080286535587989211986876;\n\treturn 0.5 * x * ( 1.0 + tanh( SQRT_2_OVER_PI * x * ( 1.0 + GELU_COEF_A * x * x ) ) );\n#endif\n}\n\ninline void computeSimple( uint idx, float add )\n{\n\tfloat f = tensor[ idx ];\n\tf += add;\n\tf = gelu( f );\n\ttensor[ idx ] = f;\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint3 it = tensorIteratorState( group, thread, tensorSize, tensorStrides );\n\tuint rsi = rowOffset( group % patternSize.yzw, patternStrides );\n\n\tif( patternSize[ 0 ] == 1 )\n\t{\n\t\t// The pattern only has 1 column - broadcasting over the row\n\t\tconst float p = pattern[ rsi ];\n\t\tROW_LOOP( it )\n\t\t\tcomputeSimple( it.x, p );\n\t}\n\telse if( patternSize[ 0 ] <= THREADS )\n\t{\n\t\t// pattern size doesn't exceed thread group size: load pattern value outside of the loop\n\t\tconst uint threadsPerGroup = THREADS - ( THREADS % patternSize[ 0 ] );\n\t\tif( thread >= threadsPerGroup )\n\t\t\treturn;\n\n\t\tconst float p = pattern[ rsi + ( thread % patternSize[ 0 ] ) * patternStrides[ 0 ] ];\n\t\tROW_LOOP_EX( it, threadsPerGroup, tensorStrides )\n\t\t\tcomputeSimple( it.x, p );\n\t}\n\telse\n\t{\n\t\t// Pattern rows are larger than the thread group, need to stream from both buffers\n\t\tconst uint rsiInc = THREADS * patternStrides[ 0 ];\n\t\tconst uint rsiDec = patternSize[ 0 ] * patternStrides[ 0 ];\n\t\tconst uint rsiEnd = rsi + rsiDec;\n\t\trsi += thread * patternStrides[ 0 ];\n\n\t\tROW_LOOP( it )\n\t\t{\n\t\t\tfloat f = tensor[ it.x ];\n\t\t\tfloat p = pattern[ rsi ];\n\t\t\trsi += rsiInc;\n\t\t\tif( rsi >= rsiEnd )\n\t\t\t\trsi -= rsiDec;\n\t\t\tf += p;\n\t\t\tf = gelu( f );\n\t\t\ttensor[ it.x ] = f;\n\t\t}\n\t}\n}"
},
{
"path": "ComputeShaders/addRepeatGelu64.hlsl",
"content": "#define THREADS 64\n#include \"addRepeatGelu.hlsl\""
},
{
"path": "ComputeShaders/addRepeatScale.hlsl",
"content": "// Compute tensor = ( tensor + repeat( pattern, tensor ) ) * scale in 1 shot, without VRAM allocations\n// Dispatch [ nb[ 1 ], nb[ 2 ], nb[ 3 ] ] thread groups of this shader, where nb is size of the destination tensor\nRWBuffer tensor: register( u0 );\nBuffer pattern: register( t0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 tensorSize: packoffset( c0 );\n\tuint4 tensorStrides: packoffset( c1 );\n\tuint4 patternSize: packoffset( c2 );\n\tuint4 patternStrides: packoffset( c3 );\n\tfloat scalingMul : packoffset( c4.x );\n}\n\n#ifndef THREADS\n#define THREADS 512\n#endif\n\n#include \"repeatUtils.hlsli\"\n\ninline void computeSimple( uint idx, float add )\n{\n\tfloat f = tensor[ idx ];\n\tf += add;\n\tf *= scalingMul;\n\ttensor[ idx ] = f;\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint3 it = tensorIteratorState( group, thread, tensorSize, tensorStrides );\n\tuint rsi = rowOffset( group % patternSize.yzw, patternStrides );\n\n\tif( patternSize[ 0 ] == 1 )\n\t{\n\t\t// The pattern only has 1 column - broadcasting over the row\n\t\tconst float p = pattern[ rsi ];\n\t\tROW_LOOP( it )\n\t\t\tcomputeSimple( it.x, p );\n\t}\n\telse if( patternSize[ 0 ] <= THREADS )\n\t{\n\t\t// pattern size doesn't exceed thread group size: load pattern value outside of the loop\n\t\tconst uint threadsPerGroup = THREADS - ( THREADS % patternSize[ 0 ] );\n\t\tif( thread >= threadsPerGroup )\n\t\t\treturn;\n\n\t\tconst float p = pattern[ rsi + ( thread % patternSize[ 0 ] ) * patternStrides[ 0 ] ];\n\t\tROW_LOOP_EX( it, threadsPerGroup, tensorStrides )\n\t\t\tcomputeSimple( it.x, p );\n\t}\n\telse\n\t{\n\t\t// Pattern rows are larger than the thread group, need to stream from both buffers\n\t\tconst uint rsiInc = THREADS * patternStrides[ 0 ];\n\t\tconst uint rsiDec = patternSize[ 0 ] * patternStrides[ 0 ];\n\t\tconst uint rsiEnd = rsi + rsiDec;\n\t\trsi += thread * patternStrides[ 0 ];\n\n\t\tROW_LOOP( it )\n\t\t{\n\t\t\tfloat f = tensor[ it.x ];\n\t\t\tfloat p = pattern[ rsi ];\n\t\t\trsi += rsiInc;\n\t\t\tif( rsi >= rsiEnd )\n\t\t\t\trsi -= rsiDec;\n\t\t\tf += p;\n\t\t\tf *= scalingMul;\n\t\t\ttensor[ it.x ] = f;\n\t\t}\n\t}\n}"
},
{
"path": "ComputeShaders/addRows.hlsl",
"content": "#ifndef THREADS\n#define THREADS 256\n#endif\n\n// dec.tokenEmbedding tensor\nBuffer tokenEmbedding: register( t0 );\n// dec.positionalEmbedding tensor\nBuffer positionalEmbedding: register( t1 );\n// R32_UINT buffer with the input tokens\nBuffer embd: register( t2 );\n// Output tensor\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint rowLength: packoffset( c0.x );\n\tuint pastTokensCount: packoffset( c0.y );\n\tuint outputRowStride: packoffset( c0.z );\n\tuint2 embStrides: packoffset( c1.x );\n\tuint2 posStrides: packoffset( c1.z );\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint row = group.x;\n\tconst uint rowTok = embd[ row ];\n\tconst uint rowPos = row + pastTokensCount;\n\n\tuint rdi = row * outputRowStride;\n\tconst uint rdiEnd = rdi + rowLength;\n\trdi += thread;\n\n\tuint rsiTok = rowTok * embStrides.y;\n\trsiTok += thread * embStrides.x;\n\n\tuint rsiPos = rowPos * posStrides.y;\n\trsiPos += thread * posStrides.x;\n\n\tfor( ; rdi < rdiEnd; rdi += THREADS, rsiTok += THREADS * embStrides.x, rsiPos += THREADS * posStrides.x )\n\t{\n\t\tfloat a = tokenEmbedding[ rsiTok ];\n\t\tfloat b = positionalEmbedding[ rsiPos ];\n\t\tresult[ rdi ] = a + b;\n\t}\n}"
},
{
"path": "ComputeShaders/componentwiseBinaryOp.hlsli",
"content": "Buffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tuint4 src1_elements: packoffset( c2 );\n\tuint4 src1_strides: packoffset( c3 );\n\tuint4 result_elements: packoffset( c4 );\n\tuint4 result_strides: packoffset( c5 );\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint j = group.x;\n\tconst uint nb1 = result_strides[ 1 ];\n\tconst uint nb01 = src0_strides[ 1 ];\n\n\tconst uint nb10 = src1_strides[ 0 ];\n\tconst uint nb11 = src1_strides[ 1 ];\n\tconst uint nc = src0_elements[ 0 ];\n\n\tuint rsi0 = j * nb01;\n\tuint rsi1 = j * nb11;\n\tuint rdi = j * nb1;\n\tconst uint rsi0End = rsi0 + nc;\n\n\trsi0 += thread;\n\trsi1 += thread * nb10;\n\trdi += thread;\n\n\tconst uint rsi1Inc = 32 * nb10;\n\tfor( ; rsi0 < rsi0End; rsi0 += 32, rsi1 += rsi1Inc, rdi += 32 )\n\t{\n\t\tconst float a = arg0[ rsi0 ];\n\t\tconst float b = arg1[ rsi1 ];\n\t\tconst float res = compute( a, b );\n\t\tresult[ rdi ] = res;\n\t}\n}"
},
{
"path": "ComputeShaders/convolutionMain.hlsl",
"content": "// ggml_compute_forward_conv_1d_1s_f16_f32, GGML_TASK_COMPUTE implementation\n// Dispatch [ ne10, ne02, 1 ] thread groups\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tuint4 src1_elements: packoffset( c2 );\n\tuint4 result_elements: packoffset( c4 );\n\tuint4 result_strides: packoffset( c5 );\n}\n\n#include \"groupReduce.hlsli\"\n\ninline void computeDotProduct( uint s0, uint s1, uint len, uint thread, inout float acc )\n{\n\tfloat curr = 0;\n\tconst uint completeVectors = len / 32;\n\tuint i;\n\tfor( i = 0; i < completeVectors; i++, s0 += 32, s1 += 32 )\n\t\tcurr = mad( arg0[ s0 + thread ], arg1[ s1 + thread ], curr );\n\n\thorizontalSumCompatNew( thread, curr );\n\n\tif( 0 == thread )\n\t{\n\t\tconst uint rem = len % 32;\n\t\tif( 0 != rem )\n\t\t{\n\t\t\tdouble f64 = curr;\n\t\t\tfor( i = 0; i < rem; i++ )\n\t\t\t{\n\t\t\t\tprecise float a = arg0[ s0 + i ];\n\t\t\t\tprecise float b = arg1[ s1 + i ];\n\t\t\t\tprecise float prod = a * b;\n\t\t\t\tf64 += prod;\n\t\t\t}\n\t\t\tcurr = (float)f64;\n\t\t}\n\t\tacc += curr;\n\t}\n}\n\n#include \"miscUtils.hlsli\"\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint i1 = group.y;\n\tconst uint i0 = group.x;\n\n\tconst uint ne00 = src0_elements[ 0 ];\n\tconst uint nk = ne00;\n\tconst int nh = (int)( nk / 2 );\n\n\tconst uint ne01 = src0_elements[ 1 ];\n\tconst int ew0 = roundUp32( ne01 );\n\n\tfloat res = 0;\n\tfor( int k = -nh; k <= nh; k++ )\n\t{\n\t\tconst uint source0 = i1 * ew0 * ne00 + uint( nh + k ) * ew0;\n\t\tconst uint source1 = uint( i0 + nh + k ) * ew0;\n\t\tcomputeDotProduct( source0, source1, ew0, thread, res );\n\t}\n\n\tif( 0 != thread )\n\t\treturn;\n\n\tconst uint nb1 = result_strides[ 1 ];\n\tconst uint rdi = i1 * nb1 + i0;\n\tresult[ rdi ] = res;\n}"
},
{
"path": "ComputeShaders/convolutionMain2.hlsl",
"content": "// ggml_compute_forward_conv_1d_2s_f16_f32, GGML_TASK_COMPUTE implementation\n// Dispatch [ ne10 / 2, ne02, 1 ] thread groups\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tuint4 src1_elements: packoffset( c2 );\n\tuint4 result_elements: packoffset( c4 );\n\tuint4 result_strides: packoffset( c5 );\n}\n\n#include \"groupReduce.hlsli\"\n\ninline void computeDotProduct( uint s0, uint s1, uint len, uint thread, inout float acc )\n{\n\tfloat curr = 0;\n\tconst uint s0End = s0 + len;\n\ts0 += thread;\n\ts1 += thread;\n\tfor( ; s0 < s0End; s0 += 32, s1 += 32 )\n\t\tcurr = mad( arg0[ s0 ], arg1[ s1 ], curr );\n\n\thorizontalSumCompatNew( thread, curr );\n\tif( 0 == thread )\n\t\tacc += curr;\n}\n\n#include \"miscUtils.hlsli\"\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint ne00 = src0_elements[ 0 ];\n\tconst uint ne01 = src0_elements[ 1 ];\n\tconst int ew0 = roundUp32( ne01 );\n\n\tfloat res = 0;\n\tuint s0 = group.y * ew0 * ne00;\n\tuint s1 = group.x * 2 * ew0;\n\t// The original implementation did following:\n\t// int nh = (int)( nk / 2 );\n\t// for( int k = -nh; k <= nh; k++ )\n\t// What we doing instead:\n\t// for( uint len = ( nk / 2 ) * 2 + 1, i = 0; i < len; i++ )\n\t// len = ( nk / 2 ) * 2 + 1 is equal to ( nk | 1 )\n\tconst uint s0End = s0 + ( ne00 | 1u ) * ew0;\n\tfor( ; s0 < s0End; s0 += ew0, s1 += ew0 )\n\t\tcomputeDotProduct( s0, s1, ew0, thread, res );\n\n\tif( 0 != thread )\n\t\treturn;\n\n\tconst uint nb1 = result_strides[ 1 ];\n\tconst uint rdi = group.y * nb1 + group.x;\n\tresult[ rdi ] = res;\n}"
},
{
"path": "ComputeShaders/convolutionMain2Fixed.hlsl",
"content": "// Optimized version of convolutionMain2.hlsl for kernel size = 3\n// Dispatch [ ( ( ne10 / 2 ) + TILE_Y - 1 ) / TILE_Y, ne02, 1 ] thread groups of this shader\n#ifndef TILE_Y\nstatic const uint TILE_Y = 8;\n#endif\n#ifndef THREADS\nstatic const uint THREADS = 64;\n#endif\n\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tuint4 src1_elements: packoffset( c2 );\n\tuint4 result_elements: packoffset( c4 );\n\tuint4 result_strides: packoffset( c5 );\n}\n\n// The accumulators we're after\ngroupshared float resTemp[ TILE_Y ][ THREADS ];\n\n// Multiply + accumulate the specified row\ninline void accumulate( float a0, float a1, const uint resultRow, const uint thread )\n{\n\tfloat acc = resTemp[ resultRow ][ thread ];\n\tacc = mad( a0, a1, acc );\n\tresTemp[ resultRow ][ thread ] = acc;\n}\n\ninline void convolutionTile( const uint s0, uint s1, const uint thread, const uint stride, const uint height )\n{\n\t// Load 3 rows from arg0\n\tconst float3 a0 = float3( arg0[ s0 ], arg0[ s0 + stride ], arg0[ s0 + stride * 2 ] );\n\n\t// Row 0\n\tfloat a1 = arg1[ s1 ];\n\taccumulate( a0[ 0 ], a1, 0, thread );\n\ts1 += stride;\n\n\tfor( uint i = 1; i < height; i++ )\n\t{\n\t\t// Row i*2-1\n\t\t// Even-indexed rows only contribute to a single output rows, after muiltiplied by kernel row #1\n\t\ta1 = arg1[ s1 ];\n\t\taccumulate( a0[ 1 ], a1, i - 1, thread );\n\t\ts1 += stride;\n\n\t\t// Row i*2, contributes to 2 output rows corresponding to kernel rows #0 and #2\n\t\ta1 = arg1[ s1 ];\n\t\taccumulate( a0[ 2 ], a1, i - 1, thread );\n\t\taccumulate( a0[ 0 ], a1, i, thread );\n\t\ts1 += stride;\n\t}\n\n\t// Row height*2 - 1\n\ta1 = arg1[ s1 ];\n\taccumulate( a0[ 1 ], a1, height - 1, thread );\n\ts1 += stride;\n\n\t// Row height*2\n\ta1 = arg1[ s1 ];\n\taccumulate( a0[ 2 ], a1, height - 1, thread );\n}\n\n#include \"miscUtils.hlsli\"\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint i;\n\t// Zero out the accumulators\n\tfor( i = 0; i < TILE_Y; i++ )\n\t\tresTemp[ i ][ thread ] = 0.0;\n\tGroupMemoryBarrierWithGroupSync();\n\n\tconst uint i1 = group.y;\n\tconst uint i0 = group.x * TILE_Y * 2;\n\tconst uint height = min( TILE_Y, ( src1_elements.x / 2 ) - group.x * TILE_Y );\n\n\tconst uint ne00 = src0_elements[ 0 ];\n\tconst uint ne01 = src0_elements[ 1 ];\n\tconst int ew0 = roundUp32( ne01 );\n\n\tuint s0 = i1 * ew0 * ne00;\n\tconst uint s0End = s0 + ew0;\n\tuint s1 = i0 * ew0;\n\ts0 += thread;\n\ts1 += thread;\n\tfor( ; s0 < s0End; s0 += THREADS, s1 += THREADS )\n\t\tconvolutionTile( s0, s1, thread, ew0, height );\n\n\tGroupMemoryBarrierWithGroupSync();\n\n\t// Now we need horizontal sums of these shared accumulators, i.e. reduce [height][THREADS] shared array into [height][1] column\n\tfor( i = THREADS / 2; i > 0; i /= 2 )\n\t{\n\t\tif( thread < i )\n\t\t{\n\t\t\tfor( uint j = 0; j < height; j++ )\n\t\t\t{\n\t\t\t\tfloat sum = resTemp[ j ][ thread ];\n\t\t\t\tsum += resTemp[ j ][ thread + i ];\n\t\t\t\tresTemp[ j ][ thread ] = sum;\n\t\t\t}\n\t\t}\n\t\tGroupMemoryBarrierWithGroupSync();\n\t}\n\n\t// And finally, store that column to global memory\n\tif( thread >= height )\n\t\treturn;\n\tconst uint nb1 = result_strides[ 1 ];\n\tconst uint rdi = i1 * nb1 + group.x * TILE_Y + thread;\n\tresult[ rdi ] = resTemp[ thread ][ 0 ];\n}"
},
{
"path": "ComputeShaders/convolutionPrep1.hlsl",
"content": "// ggml_compute_forward_conv_1d_1s_f16_f32, prepare kernel data (src0)\n// Dispatch [ ne01, ne02, 1 ] thread groups\nBuffer arg0: register( t0 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n}\n\ninline uint roundUp32( uint x )\n{\n\treturn ( x + 31 ) & ( ~31u );\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint nb01 = src0_strides[ 1 ];\n\tconst uint nb02 = src0_strides[ 2 ];\n\n\tconst uint ne00 = src0_elements[ 0 ];\n\tconst uint ne01 = src0_elements[ 1 ];\n\tconst uint ew0 = roundUp32( ne01 );\n\n\tconst uint i02 = group.y;\n\tconst uint i01 = group.x;\n\n\tuint rsi = i02 * nb02 + i01 * nb01;\n\tconst uint rsiEnd = rsi + ne00;\n\tuint rdi = i02 * ew0 * ne00 + i01;\n\trsi += thread;\n\trdi += thread * ew0;\n\tconst uint rdiInc = 32 * ew0;\n\n\tfor( ; rsi < rsiEnd; rsi += 32, rdi += rdiInc )\n\t\tresult[ rdi ] = arg0[ rsi ];\n}"
},
{
"path": "ComputeShaders/convolutionPrep2.hlsl",
"content": "// ggml_compute_forward_conv_1d_1s_f16_f32, prepare source data (src1)\n// Dispatch [ ne11, 1, 1 ] thread groups\nBuffer arg1: register( t0 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src1_elements: packoffset( c2 );\n\tuint4 src1_strides: packoffset( c3 );\n}\n\n#include \"miscUtils.hlsli\"\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint i11 = group.x;\n\n\tconst uint ne00 = src0_elements[ 0 ];\n\tconst uint ne01 = src0_elements[ 1 ];\n\tconst uint ne10 = src1_elements[ 0 ];\n\tconst uint nb11 = src1_strides[ 1 ];\n\n\tconst uint nk = ne00;\n\tconst uint nh = nk / 2;\n\tconst int ew0 = roundUp32( ne01 );\n\n\tuint rsi = i11 * nb11;\n\tuint rdi = nh * ew0 + i11;\n\tconst uint rdiInc = ew0 * 32;\n\tconst uint rsiEnd = rsi + ne10;\n\n\trsi += thread;\n\trdi += thread * ew0;\n\n\tfor( ; rsi < rsiEnd; rsi += 32, rdi += rdiInc )\n\t{\n\t\tfloat f = arg1[ rsi ];\n\t\tf = adjustFp16( f );\n\t\tresult[ rdi ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/copyConvert.hlsl",
"content": "// ggml_compute_forward_dup_f32 when we only need to convert types, but not reshape the tensor\n// Dispatch [ ne01, ne02, ne03 ] thread groups of this shader\nBuffer arg0: register( t0 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tbool downcastFp32 : packoffset( c2.x );\n}\n\n#include \"miscUtils.hlsli\"\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint nb00 = src0_strides[ 0 ];\n\tconst uint nb01 = src0_strides[ 1 ];\n\tconst uint nb02 = src0_strides[ 2 ];\n\tconst uint nb03 = src0_strides[ 3 ];\n\n\tconst uint ne00 = src0_elements[ 0 ];\n\tconst uint ne01 = src0_elements[ 1 ];\n\tconst uint ne02 = src0_elements[ 2 ];\n\tconst uint ne03 = src0_elements[ 3 ];\n\n\tconst uint i01 = group.x;\n\tconst uint i02 = group.y;\n\tconst uint i03 = group.z;\n\n\tconst uint rs = ne00 * nb00;\n\t//const uint id = i01 + i02 * ne02 + i03 * ne01 * ne02;\n\tconst uint id = ( i03 * ne01 + i02 ) * ne02 + i01;\n\n\tuint rsi = i01 * nb01 + i02 * nb02 + i03 * nb03;\n\tuint rdi = id * rs;\n\n\tconst uint rsiEnd = rsi + rs;\n\trsi += thread;\n\trdi += thread;\n\tfor( ; rsi < rsiEnd; rsi += 32, rdi += 32 )\n\t{\n\t\tfloat f = arg0[ rsi ];\n\t\t[branch]\n\t\tif( downcastFp32 )\n\t\t\tf = adjustFp16( f );\n\t\tresult[ rdi ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/copyTranspose.hlsl",
"content": "// ggml_compute_forward_dup_f32 when we actually need to reshape the tensor\n// Dispatch [ ne01, ne02, ne03 ] thread groups of this shader\nBuffer arg0: register( t0 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tbool downcastFp32 : packoffset( c2.x );\n}\n\n#include \"miscUtils.hlsli\"\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint nb00 = src0_strides[ 0 ];\n\tconst uint nb01 = src0_strides[ 1 ];\n\tconst uint nb02 = src0_strides[ 2 ];\n\tconst uint nb03 = src0_strides[ 3 ];\n\n\tconst uint ne00 = src0_elements[ 0 ];\n\tconst uint ne01 = src0_elements[ 1 ];\n\tconst uint ne02 = src0_elements[ 2 ];\n\tconst uint ne03 = src0_elements[ 3 ];\n\n\tconst uint i01 = group.x;\n\tconst uint i02 = group.y;\n\tconst uint i03 = group.z;\n\n\t// We need following integer: i01*ne00 + i02*ne00*ne01 + i03*ne00*ne01*ne02\n\t// We want to minimize count of integer multiplications\n\t// Also, DXBC assembly features `imad` instruction which computes a*b+c for integers, the actual hardware hopefully has an equivalent\n\t// i03*ne00*ne01*ne02 + i02*ne00*ne01 + i01*ne00\n\t// ( i03*ne01*ne02 + i02*ne01 + i01 ) * ne00\n\t// ( ( i03*ne02 + i02) * ne01 + i01 ) * ne00\n\tuint rdi = ( ( i03 * ne02 + i02 ) * ne01 + i01 ) * ne00;\n\n\tconst uint rdiEnd = rdi + ne00;\n\n\tuint rsi = i01 * nb01 + i02 * nb02 + i03 * nb03;\n\tconst uint rsiInc = 32 * nb00;\n\n\trdi += thread;\n\trsi += thread * nb00;\n\n\tfor( ; rdi < rdiEnd; rdi += 32, rsi += rsiInc )\n\t{\n\t\tfloat f = arg0[ rsi ];\n\t\t[branch]\n\t\tif( downcastFp32 )\n\t\t\tf = adjustFp16( f );\n\t\tresult[ rdi ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/dbgFindNaN.hlsl",
"content": "// When reset = TRUE, write zero to the output buffer\n// When reset = FALSE, test input tensor for NaN, when found at least 1 NaN element, write 1 to the output buffer\n\n// FP32 or FP16 tensor to test for NAN\nBuffer tensor: register( t0 );\n// A buffer with a single element for the output boolean. Zero means there were no NAN values in the tensor.\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint elements: packoffset( c0.x );\n\tbool reset : packoffset( c0.y );\n}\n\n// Thread group index is 16 bits per coordinate:\n// https://learn.microsoft.com/en-us/windows/win32/api/d3d11/nf-d3d11-id3d11devicecontext-dispatch\n// We want this shader to support buffers up to 2 GB.\n#ifndef THREADS\nstatic const uint THREADS = 512;\n#endif\n#ifndef ITERATIONS\nstatic const uint ITERATIONS = 128;\n#endif\n\nstatic const uint itemsPerGroup = THREADS * ITERATIONS;\n\ninline bool isNaN( float x )\n{\n\t// https://sakibsaikia.github.io/graphics/2022/01/04/Nan-Checks-In-HLSL.html\n\treturn ( asuint( x ) & 0x7fffffff ) > 0x7f800000;\n}\n\ngroupshared uint reductionBuffer;\n\n[numthreads( THREADS, 1, 1 )]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tif( reset )\n\t{\n\t\tif( 0 == thread )\n\t\t\tresult[ 0 ] = 0;\n\t\treturn;\n\t}\n\n\tuint rsi = group.x * itemsPerGroup;\n\tconst uint rsiEnd = min( rsi + itemsPerGroup, elements );\n\n\t// The main loop updates a local variable. There're THREADS instances of that variable for the group of threads.\n\tbool foundNan = false;\n\tfor( rsi += thread; rsi < rsiEnd; rsi += THREADS )\n\t{\n\t\tconst float val = tensor[ rsi ];\n\t\tif( !isNaN( val ) )\n\t\t\tcontinue;\n\t\tfoundNan = true;\n\t\tbreak;\n\t}\n\n\t// Reduce THREADS booleans to a single one, using group shared memory atomics\n\tif( 0 == thread )\n\t\treductionBuffer = 0;\n\tGroupMemoryBarrierWithGroupSync();\n\n\tif( foundNan )\n\t\tInterlockedOr( reductionBuffer, 1u );\n\n\tGroupMemoryBarrierWithGroupSync();\n\n\t// When found, update output value with global memory atomic\n\tif( 0 != thread )\n\t\treturn;\n\tif( 0 == reductionBuffer )\n\t\treturn;\n\n\tInterlockedOr( result[ 0 ], 1u );\n}"
},
{
"path": "ComputeShaders/diagMaskInf.hlsl",
"content": "// ggml_compute_forward_diag_mask_inf_f32\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 elements: packoffset( c0 );\n\tuint4 strides: packoffset( c1 );\n\tuint n_past : packoffset( c2.x );\n}\n\nstatic const float negativeInfinity = asfloat( 0xff800000 );\n\n[numthreads( 32, 1, 1 )]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint k = group.y;\n\tconst uint j = group.x;\n\n\t// Start of the row\n\tuint rdi = k * strides[ 2 ] + j * strides[ 1 ];\n\t// End of the row\n\tconst uint rdiEnd = rdi + elements[ 0 ] * strides[ 0 ];\n\t// First index to write in this thread\n\trdi += ( n_past + j + thread + 1 ) * strides[ 0 ];\n\t// Index increment\n\tconst uint rdiInc = 32 * strides[ 0 ];\n\n\tfor( ; rdi < rdiEnd; rdi += rdiInc )\n\t\tresult[ rdi ] = negativeInfinity;\n}"
},
{
"path": "ComputeShaders/flashAttention.hlsl",
"content": "// Ported from ggml_compute_forward_flash_attn_f16\n// Dispatch with [ neq1*neq2*neq3, 1, 1 ] thread groups\n\n#include \"flashAttentionCommon.hlsli\"\nBuffer lookupTable: register( t3 );\n#include \"groupReduce.hlsli\"\n\ninline void computeDotProduct( Buffer buff0, Buffer buff1, uint s0, uint s1, const uint len, const uint thread, inout float acc )\n{\n\tacc = 0;\n\tconst uint s0End = s0 + len;\n\ts0 += thread;\n\ts1 += thread;\n\tfor( ; s0 < s0End; s0 += 32, s1 += 32 )\n\t\tacc = mad( buff0[ s0 ], buff1[ s1 ], acc );\n\n\thorizontalSum( thread, acc );\n}\n\ninline void computeDotProduct( Buffer buff0, RWBuffer buff1, uint s0, uint s1, const uint len, const uint thread, inout float acc )\n{\n\tacc = 0;\n\tconst uint s0End = s0 + len;\n\ts0 += thread;\n\ts1 += thread;\n\tfor( ; s0 < s0End; s0 += 32, s1 += 32 )\n\t\tacc = mad( buff0[ s0 ], buff1[ s1 ], acc );\n\n\thorizontalSum( thread, acc );\n}\n\nvoid scaleTempVector( uint i, const uint length, const uint thread, const float multiplier, bool round )\n{\n\tconst uint end = i + length;\n\tfor( i += thread; i < end; i += 32 )\n\t{\n\t\tfloat f = temp[ i ];\n\t\tf *= multiplier;\n\t\tif( round )\n\t\t\tf = roundToFp16( f );\n\t\ttemp[ i ] = f;\n\t}\n}\n\n#include \"miscUtils.hlsli\"\n\n// Transform temp[ i ] = exp( temp[ i ] - tempMax ), and return the sum of these values\ninline float applySoftMax( uint i, const uint length, const uint thread, const float tempMax )\n{\n\t// Transform the values, and compute per-thread sum\n\tconst uint end = i + length;\n\tfloat sum = 0;\n\tfor( i += thread; i < end; i += 32 )\n\t{\n\t\tfloat f = temp[ i ];\n\t\t[branch]\n\t\tif( f != negativeInfinity )\n\t\t{\n\t\t\tf -= tempMax;\n\t\t\tconst uint index = fp16Rounded( f );\n\t\t\tconst uint res16 = lookupTable[ index ];\n\t\t\tf = f16tof32( res16 );\n\t\t}\n\t\telse\n\t\t\tf = 0;\n\n\t\ttemp[ i ] = f;\n\t\tsum += f;\n\t}\n\n\t// Reduce per-thread sum to the global one, over all threads of the group\n\thorizontalSumBroadcast( thread, sum );\n\treturn sum;\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint neq0 = q_elements[ 0 ];\n\tconst uint neq1 = q_elements[ 1 ];\n\tconst uint neq2 = q_elements[ 2 ];\n\tconst uint neq3 = q_elements[ 3 ];\n\n\tconst uint nek0 = k_elements[ 0 ];\n\tconst uint nek1 = k_elements[ 1 ];\n\n\tconst uint nev1 = v_elements[ 1 ];\n\n\tconst uint ne0 = res_elements[ 0 ];\n\tconst uint ne1 = res_elements[ 1 ];\n\n\tconst uint nbk0 = k_strides[ 0 ];\n\tconst uint nbk1 = k_strides[ 1 ];\n\tconst uint nbk2 = k_strides[ 2 ];\n\tconst uint nbk3 = k_strides[ 3 ];\n\n\tconst uint nbq0 = q_strides[ 0 ];\n\tconst uint nbq1 = q_strides[ 1 ];\n\tconst uint nbq2 = q_strides[ 2 ];\n\tconst uint nbq3 = q_strides[ 3 ];\n\n\tconst uint nbv0 = v_strides[ 0 ];\n\tconst uint nbv1 = v_strides[ 1 ];\n\tconst uint nbv2 = v_strides[ 2 ];\n\tconst uint nbv3 = v_strides[ 3 ];\n\n\tconst uint nb0 = res_strides[ 0 ];\n\tconst uint nb1 = res_strides[ 1 ];\n\tconst uint nb2 = res_strides[ 2 ];\n\tconst uint nb3 = res_strides[ 3 ];\n\n\tconst uint D = neq0;\n\tconst uint N = neq1;\n\tconst uint P = nek1 - N;\n\tconst uint M = nek1;\n\n\tconst uint ir = group.x;\n\tconst uint iq3 = ir / ( neq2 * neq1 );\n\tconst uint iq2 = ( ir - iq3 * neq2 * neq1 ) / neq1;\n\tconst uint iq1 = ( ir - iq3 * neq2 * neq1 - iq2 * neq1 );\n\n\tconst uint tempIndex = ir * tempBufferStride;\n\n\tuint ic;\n\tfloat tvm = negativeInfinity;\n\tconst uint s1 = iq1 * nbq1 + iq2 * nbq2 + iq3 * nbq3;\n\tuint s0 = iq2 * nbk2 + iq3 * nbk3;\n\tfor( ic = 0; ic < nek1; ic++, s0 += nbk1 )\n\t{\n\t\tif( masked )\n\t\t{\n\t\t\tif( ic > P + iq1 )\n\t\t\t{\n\t\t\t\tif( 0 == thread )\n\t\t\t\t\ttemp[ tempIndex + ic ] = negativeInfinity;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t}\n\n\t\tfloat dp;\n\t\tcomputeDotProduct( k, q, s0, s1, neq0, thread, dp );\n\t\tif( 0 == thread )\n\t\t{\n\t\t\tdp *= scale;\n\t\t\ttemp[ tempIndex + ic ] = dp;\n\t\t\ttvm = max( tvm, dp );\n\t\t}\n\t}\n\n\tif( 0 == thread )\n\t\tsharedAccumulators[ 0 ] = tvm;\n\tGroupMemoryBarrierWithGroupSync();\n\ttvm = sharedAccumulators[ 0 ];\n\n\t// Softmax\n\t{\n\t\tfloat sum = applySoftMax( tempIndex, M, thread, tvm );\n\t\tscaleTempVector( tempIndex, M, thread, 1.0 / sum, true );\n\t}\n\n\ts0 = iq2 * nbv2 + iq3 * nbv3;\n\tuint rdi = iq1 * nb1 + iq2 * nb2 + iq3 * nb3;\n\tfor( ic = 0; ic < nev1; ic++, s0 += nbv1, rdi += nb0 )\n\t{\n\t\tfloat dp;\n\t\tcomputeDotProduct( v, temp, s0, tempIndex, nek1, thread, dp );\n\t\tif( 0 == thread )\n\t\t\tresult[ rdi ] = dp;\n\t}\n}"
},
{
"path": "ComputeShaders/flashAttentionCommon.hlsli",
"content": "// Ported from ggml_compute_forward_flash_attn_f16\n// Dispatch with [ neq1*neq2*neq3, 1, 1 ] thread groups\nBuffer q: register( t0 );\nBuffer k: register( t1 );\nBuffer v: register( t2 );\n\nRWBuffer result: register( u0 );\n// This temporary buffer should fit tempBufferStride * neq1 * neq2 * neq3 elements, FP32 precision\nRWBuffer temp: register( u1 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 q_elements: packoffset( c0 );\n\tuint4 q_strides: packoffset( c1 );\n\tuint4 k_elements: packoffset( c2 );\n\tuint4 k_strides: packoffset( c3 );\n\tuint4 v_elements: packoffset( c4 );\n\tuint4 v_strides: packoffset( c5 );\n\tuint4 res_elements: packoffset( c6 );\n\tuint4 res_strides: packoffset( c7 );\n\n\tbool masked : packoffset( c8.x );\n\t// 1.0 / sqrt( (double) D )\n\tfloat scale : packoffset( c8.y );\n\t// This number is required to be >= nek1, and ideally rounded up to either 32 (L2 line) or 128 (L1 line) bytes\n\tuint tempBufferStride: packoffset( c8.z );\n}\n\nstatic const float negativeInfinity = asfloat( 0xff800000 );\n\n// Convert FP32 number to FP16 using rounding to nearest, then upcast back to FP32\ninline float roundToFp16( const float src )\n{\n\tconst uint trunc16 = f32tof16( src );\n\tconst float trunc32 = f16tof32( trunc16 );\n\n\tconst uint truncExp = ( trunc16 >> 10 ) & 0x1F;\n\tif( truncExp != 0x1F )\n\t{\n\t\tconst uint next16 = trunc16 + 1;\n\t\tconst float next32 = f16tof32( next16 );\n\n\t\tconst float errTrunc = abs( src - trunc32 );\n\t\tconst float errNext = abs( src - next32 );\n\n\t\tif( errTrunc < errNext )\n\t\t{\n\t\t\t// Truncated was closer to the source\n\t\t\treturn trunc32;\n\t\t}\n\t\telse if( errTrunc > errNext )\n\t\t{\n\t\t\t// Truncated + 1 was closer to the source\n\t\t\treturn next32;\n\t\t}\n\t\telse\n\t\t{\n\t\t\t// Exactly half, doing banker's rounding to nearest even\n\t\t\treturn ( 0 == ( trunc16 & 1 ) ) ? trunc32 : next32;\n\t\t}\n\t}\n\telse\n\t{\n\t\t// INF or NAN\n\t\treturn trunc32;\n\t}\n}"
},
{
"path": "ComputeShaders/flashAttentionCompat1.hlsl",
"content": "// Dispatch with [ neq1*neq2*neq3, 1, 1 ] thread groups\n#include \"flashAttentionCommon.hlsli\"\n#include \"groupReduce.hlsli\"\n\ninline void computeDotProduct( Buffer buff0, Buffer buff1, uint s0, uint s1, const uint len, const uint thread, inout float acc )\n{\n\tacc = 0;\n\t/*\n\tconst uint s0End = s0 + len;\n\ts0 += thread;\n\ts1 += thread;\n\tfor( ; s0 < s0End; s0 += 32, s1 += 32 )\n\t\tacc = mad( buff0[ s0 ], buff1[ s1 ], acc );\n\thorizontalSumCompatNew( thread, acc );\n\t*/\n\n\tconst uint completeVectors = len / 32;\n\tuint i;\n\tfor( i = 0; i < completeVectors; i++, s0 += 32, s1 += 32 )\n\t\tacc = mad( buff0[ s0 + thread ], buff1[ s1 + thread ], acc );\n\n\thorizontalSumCompatNew( thread, acc );\n\n\tif( 0 == thread )\n\t{\n\t\tconst uint rem = len % 32;\n\t\tfor( i = 0; i < rem; i++ )\n\t\t{\n\t\t\tprecise float a = buff0[ s0 + i ];\n\t\t\tprecise float b = buff1[ s1 + i ];\n\t\t\tprecise float prod = a * b;\n\t\t\tacc += prod;\n\t\t}\n\t}\n}\n\nvoid scaleTempVector( uint i, const uint length, const uint thread, const float multiplier )\n{\n\tconst uint end = i + length;\n\tfor( i += thread; i < end; i += 32 )\n\t{\n\t\tfloat f = temp[ i ];\n\t\tf *= multiplier;\n\t\ttemp[ i ] = f;\n\t}\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint neq0 = q_elements[ 0 ];\n\tconst uint neq1 = q_elements[ 1 ];\n\tconst uint neq2 = q_elements[ 2 ];\n\tconst uint neq3 = q_elements[ 3 ];\n\n\tconst uint nek0 = k_elements[ 0 ];\n\tconst uint nek1 = k_elements[ 1 ];\n\n\tconst uint nev1 = v_elements[ 1 ];\n\n\tconst uint ne0 = res_elements[ 0 ];\n\tconst uint ne1 = res_elements[ 1 ];\n\n\tconst uint nbk0 = k_strides[ 0 ];\n\tconst uint nbk1 = k_strides[ 1 ];\n\tconst uint nbk2 = k_strides[ 2 ];\n\tconst uint nbk3 = k_strides[ 3 ];\n\n\tconst uint nbq0 = q_strides[ 0 ];\n\tconst uint nbq1 = q_strides[ 1 ];\n\tconst uint nbq2 = q_strides[ 2 ];\n\tconst uint nbq3 = q_strides[ 3 ];\n\n\tconst uint nbv0 = v_strides[ 0 ];\n\tconst uint nbv1 = v_strides[ 1 ];\n\tconst uint nbv2 = v_strides[ 2 ];\n\tconst uint nbv3 = v_strides[ 3 ];\n\n\tconst uint nb0 = res_strides[ 0 ];\n\tconst uint nb1 = res_strides[ 1 ];\n\tconst uint nb2 = res_strides[ 2 ];\n\tconst uint nb3 = res_strides[ 3 ];\n\n\tconst uint D = neq0;\n\tconst uint N = neq1;\n\tconst uint P = nek1 - N;\n\t// const uint M = P + N;\n\tconst uint M = nek1;\n\n\tconst uint ir = group.x;\n\tconst uint iq3 = ir / ( neq2 * neq1 );\n\tconst uint iq2 = ( ir - iq3 * neq2 * neq1 ) / neq1;\n\tconst uint iq1 = ( ir - iq3 * neq2 * neq1 - iq2 * neq1 );\n\n\tconst uint tempIndex = ir * tempBufferStride;\n\n\tuint ic;\n\tfor( ic = 0; ic < nek1; ic++ )\n\t{\n\t\t// k indices\n\t\tconst uint ik3 = iq3;\n\t\tconst uint ik2 = iq2;\n\t\tconst uint ik1 = ic;\n\n\t\t// S indices\n\t\tconst uint i1 = ik1;\n\n\t\tif( masked )\n\t\t{\n\t\t\tif( ic > P + iq1 )\n\t\t\t{\n\t\t\t\tif( 0 == thread )\n\t\t\t\t\ttemp[ tempIndex + ic ] = negativeInfinity;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t}\n\n\t\tconst uint s0 = ik1 * nbk1 + ik2 * nbk2 + ik3 * nbk3;\n\t\tconst uint s1 = iq1 * nbq1 + iq2 * nbq2 + iq3 * nbq3;\n\t\tfloat dp;\n\t\tcomputeDotProduct( k, q, s0, s1, neq0, thread, dp );\n\t\tif( 0 == thread )\n\t\t\ttemp[ tempIndex + ic ] = dp * scale;\n\t}\n}"
},
{
"path": "ComputeShaders/flashAttentionCompat2.hlsl",
"content": "// Dispatch with [ ( neq1*neq2*neq3 + 31 ) / 32, 1, 1 ] thread groups\n#include \"flashAttentionCommon.hlsli\"\nBuffer lookupTable: register( t3 );\n\nvoid scaleTempVector( uint i, const uint length, const float multiplier )\n{\n\tconst uint end = i + length;\n\tfor( ; i < end; i++ )\n\t{\n\t\tfloat f = temp[ i ];\n\t\tf *= multiplier;\n\t\t// Rounding in this shader causes numerical errors on my GeForce 1080 Ti GPU, driver 527.56\n\t\t// f = roundToFp16( f );\n\t\ttemp[ i ] = f;\n\t}\n}\n\ninline float computeTempVectorMax( uint i, const uint length )\n{\n\t// Compute per-thread maximum\n\tconst uint end = i + length;\n\tfloat ax = negativeInfinity;\n\tfor( ; i < end; i++ )\n\t\tax = max( ax, temp[ i ] );\n\treturn ax;\n}\n\n#include \"miscUtils.hlsli\"\n#include \"fp64Utils.hlsli\"\n\n// Transform temp[ i ] = exp( temp[ i ] - tempMax ), and return the sum of these values\ninline double applySoftMax( uint i, const uint length, const float tempMax )\n{\n\t// Transform the values, and compute per-thread sum\n\tconst uint end = i + length;\n\tdouble sum = 0;\n\tfor( ; i < end; i++ )\n\t{\n\t\tfloat f = temp[ i ];\n\t\t[branch]\n\t\tif( f != negativeInfinity )\n\t\t{\n\t\t\tf -= tempMax;\n\t\t\tconst uint index = fp16Rounded( f );\n\t\t\tconst uint res16 = lookupTable[ index ];\n\t\t\tf = f16tof32( res16 );\n\t\t\tsum += f;\n\t\t}\n\t\telse\n\t\t\tf = 0;\n\n\t\ttemp[ i ] = f;\n\t}\n\treturn sum;\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 dtid: SV_DispatchThreadID )\n{\n\tconst uint neq0 = q_elements[ 0 ];\n\tconst uint neq1 = q_elements[ 1 ];\n\tconst uint neq2 = q_elements[ 2 ];\n\tconst uint neq3 = q_elements[ 3 ];\n\n\tconst uint nek0 = k_elements[ 0 ];\n\tconst uint nek1 = k_elements[ 1 ];\n\n\tconst uint nev1 = v_elements[ 1 ];\n\n\tconst uint ne0 = res_elements[ 0 ];\n\tconst uint ne1 = res_elements[ 1 ];\n\n\tconst uint nbk0 = k_strides[ 0 ];\n\tconst uint nbk1 = k_strides[ 1 ];\n\tconst uint nbk2 = k_strides[ 2 ];\n\tconst uint nbk3 = k_strides[ 3 ];\n\n\tconst uint nbq0 = q_strides[ 0 ];\n\tconst uint nbq1 = q_strides[ 1 ];\n\tconst uint nbq2 = q_strides[ 2 ];\n\tconst uint nbq3 = q_strides[ 3 ];\n\n\tconst uint nbv0 = v_strides[ 0 ];\n\tconst uint nbv1 = v_strides[ 1 ];\n\tconst uint nbv2 = v_strides[ 2 ];\n\tconst uint nbv3 = v_strides[ 3 ];\n\n\tconst uint nb0 = res_strides[ 0 ];\n\tconst uint nb1 = res_strides[ 1 ];\n\tconst uint nb2 = res_strides[ 2 ];\n\tconst uint nb3 = res_strides[ 3 ];\n\n\tconst uint D = neq0;\n\tconst uint N = neq1;\n\tconst uint P = nek1 - N;\n\t// const uint M = P + N;\n\tconst uint M = nek1;\n\n\tconst uint ir = dtid.x;\n\tif( ir >= neq1 * neq2 * neq3 )\n\t\treturn;\n\n\tconst uint iq3 = ir / ( neq2 * neq1 );\n\tconst uint iq2 = ( ir - iq3 * neq2 * neq1 ) / neq1;\n\tconst uint iq1 = ( ir - iq3 * neq2 * neq1 - iq2 * neq1 );\n\n\tconst uint tempIndex = ir * tempBufferStride;\n\n\t// Softmax\n\tfloat tvm = computeTempVectorMax( tempIndex, M );\n\tdouble sum = applySoftMax( tempIndex, M, tvm );\n\n\tscaleTempVector( tempIndex, M, (float)( 1.0 / sum ) );\n}"
},
{
"path": "ComputeShaders/flashAttentionCompat3.hlsl",
"content": "// Dispatch with [ neq1*neq2*neq3, 1, 1 ] thread groups\n#include \"flashAttentionCommon.hlsli\"\n#include \"groupReduce.hlsli\"\n#include \"miscUtils.hlsli\"\n\ninline void roundTempVector( uint i, const uint len, const uint thread )\n{\n\tconst uint iEnd = i + len;\n\tfor( i += thread; i < iEnd; i += 32 )\n\t{\n\t\tfloat f = temp[ i ];\n\t\tf = roundToFp16( f );\n\t\ttemp[ i ] = f;\n\t}\n}\n\ninline void computeDotProduct( Buffer buff0, RWBuffer buff1, uint s0, uint s1, const uint len, const uint thread, inout float acc )\n{\n\tacc = 0;\n/*\tconst uint s0End = s0 + len;\n\ts0 += thread;\n\ts1 += thread;\n\tfor( ; s0 < s0End; s0 += 32, s1 += 32 )\n\t\tacc = mad( buff0[ s0 ], buff1[ s1 ], acc );\n\n\thorizontalSumCompatNew( thread, acc ); */\n\tconst uint completeVectors = len / 32;\n\tuint i;\n\tfor( i = 0; i < completeVectors; i++, s0 += 32, s1 += 32 )\n\t\tacc = mad( buff0[ s0 + thread ], buff1[ s1 + thread ], acc );\n\n\thorizontalSumCompatNew( thread, acc );\n\n\tif( 0 == thread )\n\t{\n\t\tconst uint rem = len % 32;\n\t\tif( 0 != rem )\n\t\t{\n\t\t\tdouble f64 = acc;\n\t\t\tfor( i = 0; i < rem; i++ )\n\t\t\t{\n\t\t\t\tprecise float a = buff0[ s0 + i ];\n\t\t\t\tprecise float b = buff1[ s1 + i ];\n\t\t\t\tprecise float prod = a * b;\n\t\t\t\tf64 += prod;\n\t\t\t}\n\t\t\tacc = (float)f64;\n\t\t}\n\t}\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint neq0 = q_elements[ 0 ];\n\tconst uint neq1 = q_elements[ 1 ];\n\tconst uint neq2 = q_elements[ 2 ];\n\tconst uint neq3 = q_elements[ 3 ];\n\n\tconst uint nek0 = k_elements[ 0 ];\n\tconst uint nek1 = k_elements[ 1 ];\n\n\tconst uint nev1 = v_elements[ 1 ];\n\n\tconst uint ne0 = res_elements[ 0 ];\n\tconst uint ne1 = res_elements[ 1 ];\n\n\tconst uint nbk0 = k_strides[ 0 ];\n\tconst uint nbk1 = k_strides[ 1 ];\n\tconst uint nbk2 = k_strides[ 2 ];\n\tconst uint nbk3 = k_strides[ 3 ];\n\n\tconst uint nbq0 = q_strides[ 0 ];\n\tconst uint nbq1 = q_strides[ 1 ];\n\tconst uint nbq2 = q_strides[ 2 ];\n\tconst uint nbq3 = q_strides[ 3 ];\n\n\tconst uint nbv0 = v_strides[ 0 ];\n\tconst uint nbv1 = v_strides[ 1 ];\n\tconst uint nbv2 = v_strides[ 2 ];\n\tconst uint nbv3 = v_strides[ 3 ];\n\n\tconst uint nb0 = res_strides[ 0 ];\n\tconst uint nb1 = res_strides[ 1 ];\n\tconst uint nb2 = res_strides[ 2 ];\n\tconst uint nb3 = res_strides[ 3 ];\n\n\tconst uint D = neq0;\n\tconst uint N = neq1;\n\tconst uint P = nek1 - N;\n\t// const uint M = P + N;\n\tconst uint M = nek1;\n\n\tconst uint ir = group.x;\n\tconst uint iq3 = ir / ( neq2 * neq1 );\n\tconst uint iq2 = ( ir - iq3 * neq2 * neq1 ) / neq1;\n\tconst uint iq1 = ( ir - iq3 * neq2 * neq1 - iq2 * neq1 );\n\n\tconst uint tempIndex = ir * tempBufferStride;\n\n\troundTempVector( tempIndex, nek1, thread );\n\tAllMemoryBarrierWithGroupSync();\n\n\tuint rdi = iq1 * nb1 + iq2 * nb2 + iq3 * nb3;\n\tfor( uint ic = 0; ic < nev1; ic++, rdi += nb0 )\n\t{\n\t\t// dst indices\n\t\tconst uint i1 = iq1;\n\t\tconst uint i2 = iq2;\n\t\tconst uint i3 = iq3;\n\n\t\tconst uint s0 = ic * nbv1 + i2 * nbv2 + i3 * nbv3;\n\t\tfloat dp;\n\t\tcomputeDotProduct( v, temp, s0, tempIndex, nek1, thread, dp );\n\t\tif( 0 == thread )\n\t\t\tresult[ rdi ] = dp;\n\t}\n}"
},
{
"path": "ComputeShaders/fmaRepeat1.hlsl",
"content": "// Implementation of fmaRepeat() when both source arguments have same size and strides\n// Dispatch [ nb[ 1 ], nb[ 2 ], nb[ 3 ] ] thread groups of this shader, where nb is size of the destination tensor\nRWBuffer tensor: register( u0 );\nBuffer patternMul: register( t0 );\nBuffer patternAdd: register( t1 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 tensorSize: packoffset( c0 );\n\tuint4 tensorStrides: packoffset( c1 );\n\tuint4 patternSize: packoffset( c2 );\n\tuint4 patternStrides: packoffset( c3 );\n}\n\n#ifndef THREADS\n#define THREADS 512\n#endif\n\n#include \"repeatUtils.hlsli\"\n\ninline void computeSimple( uint idx, float mul, float add )\n{\n\tprecise float f = tensor[ idx ];\n\tf *= mul;\n\tf += add;\n\ttensor[ idx ] = f;\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint3 it = tensorIteratorState( group, thread, tensorSize, tensorStrides );\n\tuint rsi = rowOffset( group % patternSize.yzw, patternStrides );\n\n\tif( patternSize[ 0 ] == 1 )\n\t{\n\t\t// The pattern only has 1 column - broadcasting over the row\n\t\tconst float pMul = patternMul[ rsi ];\n\t\tconst float pAdd = patternAdd[ rsi ];\n\t\tROW_LOOP( it )\n\t\t\tcomputeSimple( it.x, pMul, pAdd );\n\t}\n\telse if( patternSize[ 0 ] <= THREADS )\n\t{\n\t\t// pattern size doesn't exceed thread group size: load pattern value outside of the loop\n\t\tconst uint threadsPerGroup = THREADS - ( THREADS % patternSize[ 0 ] );\n\t\tif( thread >= threadsPerGroup )\n\t\t\treturn;\n\n\t\trsi += ( thread % patternSize[ 0 ] ) * patternStrides[ 0 ];\n\t\tconst float pMul = patternMul[ rsi ];\n\t\tconst float pAdd = patternAdd[ rsi ];\n\t\tROW_LOOP_EX( it, threadsPerGroup, tensorStrides )\n\t\t\tcomputeSimple( it.x, pMul, pAdd );\n\t}\n\telse\n\t{\n\t\t// Pattern rows are larger than the thread group, need to stream from both buffers\n\t\tconst uint rsiInc = THREADS * patternStrides[ 0 ];\n\t\tconst uint rsiDec = patternSize[ 0 ] * patternStrides[ 0 ];\n\t\tconst uint rsiEnd = rsi + rsiDec;\n\t\trsi += thread * patternStrides[ 0 ];\n\n\t\tROW_LOOP( it )\n\t\t{\n\t\t\tprecise float f = tensor[ it.x ];\n\t\t\tfloat mul = patternMul[ rsi ];\n\t\t\tfloat add = patternAdd[ rsi ];\n\t\t\trsi += rsiInc;\n\t\t\tif( rsi >= rsiEnd )\n\t\t\t\trsi -= rsiDec;\n\t\t\tf *= mul;\n\t\t\tf += add;\n\t\t\ttensor[ it.x ] = f;\n\t\t}\n\t}\n}"
},
{
"path": "ComputeShaders/fmaRepeat164.hlsl",
"content": "#define THREADS 64\n#include \"fmaRepeat1.hlsl\""
},
{
"path": "ComputeShaders/fmaRepeat2.hlsl",
"content": "// Implementation of fmaRepeat() when source arguments have different shape or VRAM layout\n// Dispatch [ nb[ 1 ], nb[ 2 ], nb[ 3 ] ] thread groups of this shader, where nb is size of the destination tensor\nRWBuffer tensor: register( u0 );\nBuffer patternMul: register( t0 );\nBuffer patternAdd: register( t1 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 tensorSize: packoffset( c0 );\n\tuint4 tensorStrides: packoffset( c1 );\n\tuint4 patternSizeMul: packoffset( c2 );\n\tuint4 patternStridesMul: packoffset( c3 );\n\tuint4 patternSizeAdd: packoffset( c4 );\n\tuint4 patternStridesAdd: packoffset( c5 );\n}\n\n#ifndef THREADS\n#define THREADS 32\n#endif\n\n#include \"repeatUtils.hlsli\"\n\ninline float loadPattern( Buffer buffer, uint rowStart, uint i, uint4 size, uint4 stride )\n{\n\ti %= size.x;\n\treturn buffer[ i * stride.x + rowStart ];\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint3 it = tensorIteratorState( group, thread, tensorSize, tensorStrides );\n\tconst uint rsiMul = rowOffset( group % patternSizeMul.yzw, patternStridesMul );\n\tconst uint rsiAdd = rowOffset( group % patternSizeAdd.yzw, patternStridesAdd );\n\n\tfor( uint i = thread; it.x < it.z; it.x += it.y, i++ )\n\t{\n\t\tprecise float f = tensor[ it.x ];\n\t\tfloat mul = loadPattern( patternMul, rsiMul, i, patternSizeMul, patternStridesMul );\n\t\tfloat add = loadPattern( patternAdd, rsiAdd, i, patternSizeAdd, patternStridesAdd );\n\t\tf *= mul;\n\t\tf += add;\n\t\ttensor[ it.x ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/fp64Utils.hlsli",
"content": "// TODO: compile another version of these shader, and use it on GPUs with ExtendedDoublesShaderInstructions flag, will become slightly faster\n// https://learn.microsoft.com/en-us/windows/win32/api/d3d11/ns-d3d11-d3d11_feature_data_d3d11_options\n#ifndef ExtendedDoublesShaderInstructions\n#define ExtendedDoublesShaderInstructions 0\n#endif\n\n// Compute num/den in FP64 precision\ninline double div64( double num, double den )\n{\n#if ExtendedDoublesShaderInstructions\n\treturn num / den;\n#else\n\t// https://en.wikipedia.org/wiki/Division_algorithm#Newton%E2%80%93Raphson_division\n\tdouble x = 1.0f / (float)den;\n\tx += x * ( 1.0 - den * x );\n\tx += x * ( 1.0 - den * x );\n\treturn num * x;\n#endif\n}\n\n// Compute sqrt(x) in FP64 precision\ninline double sqrt64( double x )\n{\n\tdouble root = sqrt( (float)x );\n\troot = 0.5 * ( root + div64( x, root ) );\n\troot = 0.5 * ( root + div64( x, root ) );\n\treturn root;\n}"
},
{
"path": "ComputeShaders/groupReduce.hlsli",
"content": "groupshared float sharedAccumulators[ 32 ];\n\n// Compute horisontal sum of the numbers. The result is only correct on the thread #0 of the group.\nvoid horizontalSum( const uint thread, inout float sum )\n{\n\tsharedAccumulators[ thread ] = sum;\n\tfor( uint i = 16; i > 1; i /= 2 )\n\t{\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tif( thread < i )\n\t\t{\n\t\t\tsum += sharedAccumulators[ thread + i ];\n\t\t\tsharedAccumulators[ thread ] = sum;\n\t\t}\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\tif( 0 == thread )\n\t\tsum += sharedAccumulators[ 1 ];\n}\n\n// Compute horisontal sum of the numbers, and broadcast to all threads of the group.\nvoid horizontalSumBroadcast( const uint thread, inout float sum )\n{\n\thorizontalSum( thread, sum );\n\tif( 0 == thread )\n\t\tsharedAccumulators[ 0 ] = sum;\n\tGroupMemoryBarrierWithGroupSync();\n\tsum = sharedAccumulators[ 0 ];\n}\n\n// Compute horisontal sum of the numbers, in the order equal to the CPU-running dot product implementation.\n// The result is only correct on the thread #0 of the group.\nvoid horizontalSumCompat( const uint thread, inout float sum )\n{\n\tsharedAccumulators[ thread ] = sum;\n\tGroupMemoryBarrierWithGroupSync();\n\n\tif( 0 == ( thread & 8 ) )\n\t{\n\t\t// This runs on threads [ 0 .. 7 ] and [ 16 .. 23 ]\n\t\t// sum01 = _mm256_add_ps( sum0, sum1 );\n\t\t// sum23 = _mm256_add_ps( sum2, sum3 );\n\t\tsum += sharedAccumulators[ thread + 8 ];\n\t\tsharedAccumulators[ thread ] = sum;\n\t}\n\n\tGroupMemoryBarrierWithGroupSync();\n\tif( thread < 8 )\n\t{\n\t\t// This runs on threads [ 0 .. 7 ]\n\t\t// sum0123 = _mm256_add_ps( sum01, sum23 );\n\t\tsum += sharedAccumulators[ thread + 16 ];\n\t\tsharedAccumulators[ thread ] = sum;\n\t}\n\n\tGroupMemoryBarrierWithGroupSync();\n\tif( thread < 4 )\n\t{\n\t\t// const __m128 r4 = _mm_add_ps( _mm256_castps256_ps128( sum0123 ), _mm256_extractf128_ps( sum0123, 1 ) );\n\t\tsum += sharedAccumulators[ thread + 4 ];\n\t\tsharedAccumulators[ thread ] = sum;\n\t}\n\n\tGroupMemoryBarrierWithGroupSync();\n\tif( thread < 2 )\n\t{\n\t\t// const __m128 r2 = _mm_add_ps( r4, _mm_movehl_ps( r4, r4 ) );\n\t\tsum += sharedAccumulators[ thread + 2 ];\n\t\tsharedAccumulators[ thread ] = sum;\n\t}\n\n\tGroupMemoryBarrierWithGroupSync();\n\tif( 0 == thread )\n\t{\n\t\t// const __m128 r1 = _mm_add_ss( r2, _mm_movehdup_ps( r2 ) );\n\t\tsum += sharedAccumulators[ 1 ];\n\t}\n}\n\n// Compute horisontal sum of the numbers, in yet another creative summation order recently implemented in the upstream\nvoid horizontalSumCompatNew( const uint thread, inout float sum )\n{\n\t// GGML_F32x8_REDUCE\n\tsharedAccumulators[ thread ] = sum;\n\tGroupMemoryBarrierWithGroupSync();\n\n\tif( 0 == ( thread & 8 ) )\n\t{\n\t\t// Runs on threads [ 0 .. 7 ] and [ 16 .. 23 ]\n\t\tsum += sharedAccumulators[ thread | 8 ];\n\t\tsharedAccumulators[ thread ] = sum;\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\n\tif( thread < 8 )\n\t{\n\t\t// Runs on threads [ 0 .. 7 ]\n\t\tsum += sharedAccumulators[ thread | 0x10 ];\n\t\tsharedAccumulators[ thread ] = sum;\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\n\tif( thread < 4 )\n\t{\n\t\t// Runs on threads [ 0 .. 3 ]\n\t\tsum += sharedAccumulators[ thread | 4 ];\n\t\tsharedAccumulators[ thread ] = sum;\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\n\tif( thread < 4 && 0 == ( thread & 1 ) )\n\t{\n\t\t// Runs on threads [ 0, 2 ]\n\t\tsum += sharedAccumulators[ thread | 1 ];\n\t\tsharedAccumulators[ thread ] = sum;\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\n\tif( 0 == thread )\n\t\tsum += sharedAccumulators[ 2 ];\n}\n\n\n// Compute horizontal maximum of the numbers, and broadcast to all threads of the group.\nvoid horizontalMaxBroadcast( const uint thread, inout float ax )\n{\n\tsharedAccumulators[ thread ] = ax;\n\tfor( uint i = 16; i > 0; i /= 2 )\n\t{\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tif( thread < i )\n\t\t{\n\t\t\tax = max( ax, sharedAccumulators[ thread + i ] );\n\t\t\tsharedAccumulators[ thread ] = ax;\n\t\t}\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\tax = sharedAccumulators[ 0 ];\n}"
},
{
"path": "ComputeShaders/groupReduce64.hlsli",
"content": "groupshared float sharedAccumulators[ 64 ];\n\n// Compute horisontal sum of the numbers. The result is only correct on the thread #0 of the group.\nvoid horizontalSum( const uint thread, inout float sum )\n{\n\tsharedAccumulators[ thread ] = sum;\n\tfor( uint i = 32; i > 1; i /= 2 )\n\t{\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tif( thread < i )\n\t\t{\n\t\t\tsum += sharedAccumulators[ thread + i ];\n\t\t\tsharedAccumulators[ thread ] = sum;\n\t\t}\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\tif( 0 == thread )\n\t\tsum += sharedAccumulators[ 1 ];\n}\n\n// Compute horisontal sum of the numbers, and broadcast to all threads of the group.\nvoid horizontalSumBroadcast( const uint thread, inout float sum )\n{\n\thorizontalSum( thread, sum );\n\tif( 0 == thread )\n\t\tsharedAccumulators[ 0 ] = sum;\n\tGroupMemoryBarrierWithGroupSync();\n\tsum = sharedAccumulators[ 0 ];\n}\n\n// Compute horizontal maximum of the numbers, and broadcast to all threads of the group.\nvoid horizontalMaxBroadcast( const uint thread, inout float ax )\n{\n\tsharedAccumulators[ thread ] = ax;\n\tfor( uint i = 32; i > 0; i /= 2 )\n\t{\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tif( thread < i )\n\t\t{\n\t\t\tax = max( ax, sharedAccumulators[ thread + i ] );\n\t\t\tsharedAccumulators[ thread ] = ax;\n\t\t}\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\tax = sharedAccumulators[ 0 ];\n}"
},
{
"path": "ComputeShaders/matReshapePanels.hlsl",
"content": "// This shader reshapes a matrix into the shape expected by mulMatTiledEx.hlsl and mulMatByRowTiledEx.hlsl compute shaders\n// It's called in runtime, also while loading models from disk.\n// So far, it's only used when running on AMD GPUs.\n#ifndef TILE_SIZE\nstatic const uint TILE_SIZE = 32;\n#endif\n\n// Input tensor\nBuffer source: register( t0 );\n// Output tensor\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 arg0Size: packoffset( c0 );\n\tuint4 arg0Strides: packoffset( c1 );\n\t// Count of elements per panel\n\tuint panelSize : packoffset( c2.y );\n\t// Layer strides of the output matrix\n\tuint2 layerStrides: packoffset( c2.z );\n}\n\ninline uint hadd( uint2 v2 ) { return v2.x + v2.y; }\n\ngroupshared float tileBuffer[ TILE_SIZE ][ TILE_SIZE ];\n\n[ numthreads( TILE_SIZE, 1, 1 ) ]\nvoid main( const uint3 group: SV_GroupID, const uint thread : SV_GroupIndex )\n{\n\tuint rdi = hadd( group.yz * layerStrides );\n\trdi += group.x * panelSize;\n\trdi += thread;\n\n\tuint rsi = hadd( group.yz * arg0Strides.zw );\n\tconst uint baseY = group.x * TILE_SIZE;\n\tconst uint dispatchThread = baseY + thread;\n\t// Reshaping into a column major horizontal panel, height = TILE_SIZE, width = width of the source matrix\n\tuint width = arg0Size.x;\n\t// Usually TILE_SIZE; can be less for the last panel on the matrix when we need to generate zeros instead of loading these numbers\n\tconst uint height = min( TILE_SIZE, arg0Size.y - baseY );\n\n\tif( arg0Strides.x == 1 )\n\t{\n\t\t// The input matrix is row major, can improve performance with coalesced loads and group shared buffer.\n\t\trsi += baseY * arg0Strides.y;\n\n\t\tconst uint widthCompleteTiles = width / TILE_SIZE;\n\n\t\tif( height < TILE_SIZE )\n\t\t{\n\t\t\t// This thread group was dispatched for the last panel of the matrix, it doesn't have enough rows\n\t\t\t// Write zeros to the corresponding elements of the groupshared buffer\n\t\t\tfor( uint j = height; j < TILE_SIZE; j++ )\n\t\t\t\ttileBuffer[ thread ][ j ] = 0.0;\n\t\t}\n\n\t\tfor( uint i = 0; i < widthCompleteTiles; i++, rsi += TILE_SIZE )\n\t\t{\n\t\t\t// Load [ TILE_SIZE ] * [ TILE_SIZE ] block with fully coalesced loads, store to group shared buffer in transposed order\n\t\t\tuint rsiTile = rsi + thread;\n\t\t\tuint j;\n\t\t\tfor( j = 0; j < height; j++, rsiTile += arg0Strides.y )\n\t\t\t{\n\t\t\t\t// Each iteration of the loop loads a row of [ TILE_SIZE ] elements from the corresponding row of the source tensor\n\t\t\t\t// Fully coalesced load\n\t\t\t\tfloat f = source[ rsiTile ];\n\t\t\t\t// Random store but the local memory's fast, this works rather well in practice\n\t\t\t\ttileBuffer[ thread ][ j ] = f;\n\t\t\t}\n\n\t\t\tGroupMemoryBarrierWithGroupSync();\n\n\t\t\t// Copy from group shared buffer to output tensor\n\t\t\tfor( j = 0; j < TILE_SIZE; j++, rdi += TILE_SIZE )\n\t\t\t{\n\t\t\t\t// Fully coalesced loads and stores\n\t\t\t\tfloat f = tileBuffer[ j ][ thread ];\n\t\t\t\tresult[ rdi ] = f;\n\t\t\t}\n\n\t\t\tGroupMemoryBarrierWithGroupSync();\n\t\t}\n\n\t\twidth %= TILE_SIZE;\n\t\tif( 0 == width )\n\t\t\treturn;\n\t\trsi += thread * arg0Strides.y;\n\t}\n\telse\n\t\trsi += dispatchThread * arg0Strides.y;\n\n\tfor( uint i = 0; i < width; i++ )\n\t{\n\t\tfloat f;\n\t\t[branch]\n\t\tif( thread < height )\n\t\t\tf = source[ rsi ];\n\t\telse\n\t\t\tf = 0.0;\n\t\trsi += arg0Strides.x;\n\n\t\tresult[ rdi ] = f;\n\t\trdi += TILE_SIZE;\n\t}\n}"
},
{
"path": "ComputeShaders/miscUtils.hlsli",
"content": "// When GPUs are converting FP32 to FP16, they always truncate towards 0, documented there:\n// https://learn.microsoft.com/en-us/windows/win32/direct3d10/d3d10-graphics-programming-guide-resources-data-conversion#conververting-from-a-higher-range-representation-to-a-lower-range-representation\n// Whisper code uses _mm_cvtps_ph( x, 0 ), the 0 stands for \"Round to nearest even\": https://www.felixcloutier.com/x86/vcvtps2ph\n// This function adjusts FP32 value making it so that truncation towards 0 results in the value equal to what CPU is doing\ninline float adjustFp16( const float src )\n{\n\tconst uint trunc16 = f32tof16( src );\n\tconst float trunc32 = f16tof32( trunc16 );\n\n\tconst uint truncExp = ( trunc16 >> 10 ) & 0x1F;\n\tif( truncExp != 0x1F )\n\t{\n\t\tconst uint next16 = trunc16 + 1;\n\t\tconst float next32 = f16tof32( next16 );\n\n\t\tconst float errTrunc = abs( src - trunc32 );\n\t\tconst float errNext = abs( src - next32 );\n\n\t\tif( errTrunc < errNext )\n\t\t{\n\t\t\t// Truncated was closer to the source\n\t\t\treturn src;\n\t\t}\n\t\telse if( errTrunc > errNext )\n\t\t{\n\t\t\t// Truncated + 1 was closer to the source\n\t\t\treturn next32;\n\t\t}\n\t\telse\n\t\t{\n\t\t\t// Exactly half, doing banker's rounding to nearest even\n\t\t\treturn ( 0 == ( trunc16 & 1 ) ) ? src : next32;\n\t\t}\n\t}\n\telse\n\t{\n\t\t// INF or NAN\n\t\treturn src;\n\t}\n}\n\n// Convert FP32 number to FP16, using rounding to nearest\ninline uint fp16Rounded( const float src )\n{\n\tconst uint trunc16 = f32tof16( src );\n\tconst float trunc32 = f16tof32( trunc16 );\n\n\tconst uint truncExp = ( trunc16 >> 10 ) & 0x1F;\n\tif( truncExp != 0x1F )\n\t{\n\t\tconst uint next16 = trunc16 + 1;\n\t\tconst float next32 = f16tof32( next16 );\n\n\t\tconst float errTrunc = abs( src - trunc32 );\n\t\tconst float errNext = abs( src - next32 );\n\n\t\tif( errTrunc < errNext )\n\t\t{\n\t\t\t// Truncated was closer to the source\n\t\t\treturn trunc16;\n\t\t}\n\t\telse if( errTrunc > errNext )\n\t\t{\n\t\t\t// Truncated + 1 was closer to the source\n\t\t\treturn next16;\n\t\t}\n\t\telse\n\t\t{\n\t\t\t// Exactly half, doing banker's rounding to nearest even\n\t\t\treturn ( 0 == ( trunc16 & 1 ) ) ? trunc16 : next16;\n\t\t}\n\t}\n\telse\n\t{\n\t\t// INF or NAN\n\t\treturn trunc16;\n\t}\n}\n\n// Round up the number to be a multiple of 32\ninline uint roundUp32( uint x )\n{\n\treturn ( x + 31 ) & ( ~31u );\n}"
},
{
"path": "ComputeShaders/mulMatByRow.hlsl",
"content": "// Matrix * row product, like [ E0, E1, E2, E3 ] * [ E0, 1, E2, E3 ] = [ E1, 1, E2, E3 ]\n// Dispatch [ E1, E2, E3 ] groups of this shader\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 arg0Size: packoffset( c0 );\n\tuint4 arg0Strides: packoffset( c1 );\n\tuint4 arg1Size: packoffset( c2 );\n\tuint4 arg1Strides: packoffset( c3 );\n\tuint4 resultSize: packoffset( c4 );\n\tuint4 resultStrides: packoffset( c5 );\n}\n\n#include \"groupReduce.hlsli\"\n\ninline uint hadd( uint3 vec )\n{\n\treturn vec.x + vec.y + vec.z;\n}\ninline uint hadd( uint2 vec )\n{\n\treturn vec.x + vec.y;\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint s0 = hadd( group * arg0Strides.yzw );\n\tuint s1 = hadd( group.yz * arg1Strides.zw );\n\tconst uint s0End = s0 + arg0Size.x * arg0Strides.x;\n\tconst uint s0Inc = 32 * arg0Strides.x;\n\tconst uint s1Inc = 32 * arg1Strides.x;\n\n\ts0 += thread * arg0Strides.x;\n\ts1 += thread * arg1Strides.x;\n\tfloat dp = 0;\n\tfor( ; s0 < s0End; s0 += s0Inc, s1 += s1Inc )\n\t\tdp = mad( arg0[ s0 ], arg1[ s1 ], dp );\n\n\thorizontalSum( thread, dp );\n\tif( 0 != thread )\n\t\treturn;\n\n\tconst uint rdi = group.x + hadd( group.yz * resultStrides.zw );\n\tresult[ rdi ] = dp;\n}"
},
{
"path": "ComputeShaders/mulMatByRow64.hlsl",
"content": "// Matrix * row product, like [ E0, E1, E2, E3 ] * [ E0, 1, E2, E3 ] = [ E1, 1, E2, E3 ]\n// Dispatch [ E1, E2, E3 ] groups of this shader\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 arg0Size: packoffset( c0 );\n\tuint4 arg0Strides: packoffset( c1 );\n\tuint4 arg1Size: packoffset( c2 );\n\tuint4 arg1Strides: packoffset( c3 );\n\tuint4 resultSize: packoffset( c4 );\n\tuint4 resultStrides: packoffset( c5 );\n}\n\ninline uint hadd( uint3 vec )\n{\n\treturn vec.x + vec.y + vec.z;\n}\ninline uint hadd( uint2 vec )\n{\n\treturn vec.x + vec.y;\n}\n\n// No idea why, but that particular configuration appears to be the fastest one on Ryzen 7 5700G iGPU\n// Not by much, though: when trying a few numbers I saw 1.30 - 1.42 seconds for this compute shader\nstatic const uint THREADS = 64;\nstatic const uint REDUCTION_BUFFER = 32;\ngroupshared float sharedAccumulators[ REDUCTION_BUFFER ];\n\n// Compute horisontal sum of the numbers. The result is only correct on the thread #0 of the group.\nvoid horizontalSum( const uint thread, inout float sum )\n{\n\tif( THREADS > REDUCTION_BUFFER )\n\t{\n\t\tfor( uint t = REDUCTION_BUFFER; t < THREADS; t += REDUCTION_BUFFER )\n\t\t{\n\t\t\t// Threads [ t .. t + REDUCTION_BUFFER ] store into the buffer\n\t\t\tif( thread >= t && thread < t + REDUCTION_BUFFER )\n\t\t\t\tsharedAccumulators[ thread - t ] = sum;\n\n\t\t\tGroupMemoryBarrierWithGroupSync();\n\n\t\t\t// Threads [ 0 .. REDUCTION_BUFFER ] increment their local sum with the value loaded from the buffer\n\t\t\tif( thread < REDUCTION_BUFFER )\n\t\t\t\tsum += sharedAccumulators[ thread ];\n\t\t}\n\t}\n\n\tif( thread < REDUCTION_BUFFER )\n\t\tsharedAccumulators[ thread ] = sum;\n\n\tfor( uint i = REDUCTION_BUFFER / 2; i > 1; i /= 2 )\n\t{\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tif( thread < i )\n\t\t{\n\t\t\tsum += sharedAccumulators[ thread + i ];\n\t\t\tsharedAccumulators[ thread ] = sum;\n\t\t}\n\t}\n\n\tGroupMemoryBarrierWithGroupSync();\n\tif( 0 == thread )\n\t\tsum += sharedAccumulators[ 1 ];\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint s0 = hadd( group * arg0Strides.yzw );\n\tuint s1 = hadd( group.yz * arg1Strides.zw );\n\tconst uint s0End = s0 + arg0Size.x * arg0Strides.x;\n\tconst uint s0Inc = THREADS * arg0Strides.x;\n\tconst uint s1Inc = THREADS * arg1Strides.x;\n\n\ts0 += thread * arg0Strides.x;\n\ts1 += thread * arg1Strides.x;\n\tfloat dp = 0;\n\tfor( ; s0 < s0End; s0 += s0Inc, s1 += s1Inc )\n\t\tdp = mad( arg0[ s0 ], arg1[ s1 ], dp );\n\n\thorizontalSum( thread, dp );\n\tif( 0 != thread )\n\t\treturn;\n\n\tconst uint rdi = group.x + hadd( group.yz * resultStrides.zw );\n\tresult[ rdi ] = dp;\n}"
},
{
"path": "ComputeShaders/mulMatByRowTiled.hlsl",
"content": "// Matrix * row product, like [ E0, E1, E2, E3 ] * [ E0, 1, E2, E3 ] = [ E1, 1, E2, E3 ]\n// Dispatch [ ( E1 + TILE_Y - 1 ) / TILE_Y, E2, E3 ] thread groups of this shader\n// This one here is the second most expensive shader in the model, after matrix*matrix product.\n// Optimized heavily, as a result the readability ain't great.\n\n#ifndef TILE_Y\nstatic const uint TILE_Y = 64;\n#endif\n#ifndef THREADS_X\nstatic const uint THREADS_X = 32;\n#endif\n#ifndef THREADS_Y\nstatic const uint THREADS_Y = 16;\n#endif\n\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 arg0Size: packoffset( c0 );\n\tuint4 arg0Strides: packoffset( c1 );\n\tuint4 arg1Size: packoffset( c2 );\n\tuint4 arg1Strides: packoffset( c3 );\n\tuint4 resultSize: packoffset( c4 );\n\tuint4 resultStrides: packoffset( c5 );\n}\n\ninline uint hadd( uint2 vec )\n{\n\treturn vec.x + vec.y;\n}\n\n// Count of FP32 accumulators we need in every thread of the shader\nstatic const uint heightScalars = TILE_Y / THREADS_Y;\n// The local accumulators are float4 vectors, compute count of these vectors\nstatic const uint heightVectors = ( heightScalars + 3 ) / 4;\n\ngroupshared float4 reductionBuffer[ heightVectors ][ THREADS_Y ][ THREADS_X ];\n\n[numthreads( THREADS_X, THREADS_Y, 1 )]\nvoid main( uint3 group: SV_GroupID, uint3 thread : SV_GroupThreadID )\n{\n\tuint i;\n\t// Despite inside GPU cores, the shared memory is still much slower than registers\n\t// For this reason, this shader accumulates numbers in local variables. Only uses groupshared buffer for the final reduction.\n\tfloat4 acc[ heightVectors ];\n\t// Zero out the accumulators\n\t[unroll]\n\tfor( i = 0; i < heightVectors; i++ )\n\t\tacc[ i ] = 0.0;\n\n\t// Count of rows to compute in this thread group\n\tconst uint height = min( TILE_Y, arg0Size.y - group.x * TILE_Y );\n\n\tuint s0 = hadd( group.yz * arg0Strides.zw ); //< arg0 layer for the thread group\n\ts0 += group.x * TILE_Y * arg0Strides.y; //< arg0 first row for the thread group\n\ts0 += hadd( arg0Strides.xy * thread.xy ); //< arg0 load index for the thread\n\n\tuint s1 = hadd( group.yz * arg1Strides.zw ); //< arg1 layer for the thread group\n\ts1 += thread.x * arg1Strides.x; //< arg1 load index for the thread\n\n\tconst uint completeTiles = arg0Size.x / THREADS_X;\n\t// Each iteration of that loop loads THREADS_X elements from arg1,\n\t// a block of [ THREADS_X, height ] elements from arg0,\n\t// and accumulates these dot products in the local variables\n\tfor( uint t = 0; t < completeTiles; t++, s0 += THREADS_X * arg0Strides.x, s1 += THREADS_X * arg1Strides.x )\n\t{\n\t\t// Load THREADS_X elements from arg1\n\t\tconst float v1 = arg1[ s1 ];\n\n\t\tuint rsi = s0;\n\t\t[unroll]\n\t\tfor( i = 0; i < heightVectors; i++ )\n\t\t{\n\t\t\tfloat4 v0 = 0.0;\n\t\t\t// Load up to 4*THREADS_X elements from arg0\n\t\t\t[unroll]\n\t\t\tfor( uint j = 0; j < 4; j++, rsi += arg0Strides.y * THREADS_Y )\n\t\t\t{\n\t\t\t\tconst uint y = ( i * 4 + j ) * THREADS_Y + thread.y;\n\t\t\t\t[branch]\n\t\t\t\tif( y < height )\n\t\t\t\t\tv0[ j ] = arg0[ rsi ];\n\t\t\t}\n\t\t\t// Multiply + accumulate\n\t\t\tacc[ i ] = mad( v0, v1, acc[ i ] );\n\t\t}\n\t}\n\n\tconst uint rem = arg0Size.x % THREADS_X;\n\tif( thread.x < rem )\n\t{\n\t\t// E0 ain't a multiple of THREADS_X, we have a remainder\n\n\t\t// Load `rem` elements from arg1\n\t\tconst float v1 = arg1[ s1 ];\n\n\t\t[unroll]\n\t\tfor( i = 0; i < heightVectors; i++ )\n\t\t{\n\t\t\tfloat4 v0 = 0.0;\n\t\t\t// Load up to 4*rem elements from arg0\n\t\t\t[unroll]\n\t\t\tfor( uint j = 0; j < 4; j++, s0 += arg0Strides.y * THREADS_Y )\n\t\t\t{\n\t\t\t\tconst uint y = ( i * 4 + j ) * THREADS_Y + thread.y;\n\t\t\t\t[branch]\n\t\t\t\tif( y < height )\n\t\t\t\t\tv0[ j ] = arg0[ s0 ];\n\t\t\t}\n\t\t\t// Multiply + accumulate\n\t\t\tacc[ i ] = mad( v0, v1, acc[ i ] );\n\t\t}\n\t}\n\n\t// Now we need horizontal sum of these accumulators, reducing [height][THREADS_X] of them into [height][1] column\n\t// First, store local variables into the shared memory.\n\t[ unroll ]\n\tfor( i = 0; i < heightVectors; i++ )\n\t\treductionBuffer[ i ][ thread.y ][ thread.x ] = acc[ i ];\n\tGroupMemoryBarrierWithGroupSync();\n\n\t// Run reduction using that shared memory buffer\n\tfor( i = THREADS_X / 2; i > 1; i /= 2 )\n\t{\n\t\tif( thread.x < i )\n\t\t{\n\t\t\t[unroll]\n\t\t\tfor( uint iv = 0; iv < heightVectors; iv++ )\n\t\t\t{\n\t\t\t\tfloat4 that = reductionBuffer[ iv ][ thread.y ][ thread.x + i ];\n\t\t\t\tfloat4 tmp = acc[ iv ];\n\t\t\t\ttmp += that;\n\t\t\t\treductionBuffer[ iv ][ thread.y ][ thread.x ] = tmp;\n\t\t\t\tacc[ iv ] = tmp;\n\t\t\t}\n\t\t}\n\t\tGroupMemoryBarrierWithGroupSync();\n\t}\n\n\t// And finally, store that column to global memory.\n\t// Only running that code on the threads of the group with thread.x = 0, to save a few loads from the groupshared buffer\n\t// This allows to use registers instead, faster to access\n\tif( thread.x != 0 )\n\t\treturn;\n\n\tuint rdi = hadd( group.yz * resultStrides.zw );\n\trdi += ( group.x * TILE_Y + thread.y ) * resultStrides.x;\n\tconst uint rdiInc = THREADS_Y * resultStrides.x;\n\n\t[unroll]\n\tfor( i = 0; i < heightVectors; i++ )\n\t{\n\t\t// The previous loop had \"i > 1\" continue condition, it didn't complete the last step of the reduction\n\t\t// The following line is doing that last reduction step\n\t\tconst float4 resultVec = acc[ i ] + reductionBuffer[ i ][ thread.y ][ 1 ];\n\n\t\t// Conditionally store these 4 floats to the output tensor\n\t\t[unroll]\n\t\tfor( uint j = 0; j < 4; j++, rdi += rdiInc )\n\t\t{\n\t\t\tconst uint y = ( i * 4 + j ) * THREADS_Y + thread.y;\n\t\t\t[branch]\n\t\t\tif( y < height )\n\t\t\t\tresult[ rdi ] = resultVec[ j ];\n\t\t}\n\t}\n}"
},
{
"path": "ComputeShaders/mulMatByRowTiledEx.hlsl",
"content": "// matrix*row vector product, needs first argument reshaped into a sequence of horizontal column major panels\n#ifndef TILE_SIZE\nstatic const uint TILE_SIZE = 32;\n#endif\n#ifndef THREADS_Y\nstatic const uint THREADS_Y = 8;\n#endif\n\n// First tensor, reshaped into dense column major horizontal panels of size [ width, TILE_SIZE ]\nBuffer arg0: register( t0 );\n// Second tensor, reshaped into dense column major horizontal panels of size [ width, TILE_SIZE ]\nBuffer arg1: register( t1 );\n// FP32 output tensor, row major and continuous\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 arg0Size: packoffset( c0 );\n\tuint arg0panel: packoffset( c1.y );\n\tuint2 arg0LayerStrides: packoffset( c1.z );\n\t// uint4 arg1Size: packoffset( c2 );\n\tuint4 arg1Strides: packoffset( c3 );\n\tuint4 resultSize: packoffset( c4 );\n\tuint4 resultStrides: packoffset( c5 );\n}\n\ninline uint hadd4( const uint4 v )\n{\n\tconst uint2 v2 = v.xy + v.zw;\n\treturn v2.x + v2.y;\n}\n\ninline float hadd4( const float4 v )\n{\n\tconst float2 v2 = v.xy + v.zw;\n\treturn v2.x + v2.y;\n}\n\ngroupshared float reductionBuffer[ THREADS_Y ][ TILE_SIZE ];\n\n[numthreads( TILE_SIZE, THREADS_Y, 1 )]\nvoid main( const uint3 group: SV_GroupID, const uint3 thread : SV_GroupThreadID )\n{\n\tconst uint2 layer = group.yz;\n\t// Source offsets for the complete thread group\n\tuint2 rsi;\n\trsi.x = group.x * arg0panel + layer.x * arg0LayerStrides.x + layer.y * arg0LayerStrides.y;\n\trsi.y = layer.x * arg1Strides.z + layer.y * arg1Strides.w;\n\t// Apply source offsets for this particular thread\n\trsi.x += thread.y * TILE_SIZE + thread.x;\n\trsi.y += thread.y * arg1Strides.x;\n\n\tconst uint2 rsiInc = uint2( THREADS_Y * TILE_SIZE, THREADS_Y * arg1Strides.x );\n\n\tconst uint completeTiles = arg0Size.x / ( THREADS_Y * 4 );\n\tuint i;\n\tfloat4 acc = 0.0;\n\tfor( i = 0; i < completeTiles; i++ )\n\t{\n\t\t// Each iteration of this loop consumes THREADS_Y*4 columns from the arg0 panel, and THREADS_Y*4 values from arg1\n\t\tfloat4 v0, v1;\n\t\t[unroll]\n\t\tfor( uint j = 0; j < 4; j++, rsi += rsiInc )\n\t\t{\n\t\t\t// Load [ TILE_SIZE, THREADS_Y ] block from the first source tensor\n\t\t\tv0[ j ] = arg0[ rsi.x ];\n\t\t\t// Broadcast [ THREADS_Y ] row from the second source tensor\n\t\t\tv1[ j ] = arg1[ rsi.y ];\n\t\t}\n\n\t\t// Now we have [ TILE_SIZE, THREADS_Y * 4 ] block from the first source tensor in the v0 vector,\n\t\t// and [ THREADS_Y * 4 ] row from the second one in the v1 vector\n\t\t// Multiply and accumulate.\n\t\tacc = mad( v0, v1, acc );\n\t}\n\n\t// Handle the remainder columns, if any.\n\t// When present, their count is in [ 1 .. THREADS_Y * 4 - 1 ] interval\n\tconst uint rem = arg0Size.x % ( THREADS_Y * 4 );\n\tif( rem != 0 )\n\t{\n\t\tfloat4 v0 = 0.0, v1 = 0.0;\n\t\t[unroll]\n\t\tfor( uint j = 0; j < 4; j++, rsi += rsiInc )\n\t\t{\n\t\t\tconst uint x = ( j * THREADS_Y ) + thread.y;\n\t\t\tif( x < rem )\n\t\t\t{\n\t\t\t\tv0[ j ] = arg0[ rsi.x ];\n\t\t\t\tv1[ j ] = arg1[ rsi.y ];\n\t\t\t}\n\t\t}\n\t\tacc = mad( v0, v1, acc );\n\t}\n\n\t// We now have [ TILE_SIZE, THREADS_Y * 4 ] block in the local variables of this thread group\n\t// The group however only outputs [ TILE_SIZE ] elements max, need a reduction\n\tfloat acc1 = hadd4( acc );\n\treductionBuffer[ thread.y ][ thread.x ] = acc1;\n\tGroupMemoryBarrierWithGroupSync();\n\n\tfor( i = THREADS_Y / 2; i > 1; i /= 2 )\n\t{\n\t\tif( thread.y < i )\n\t\t{\n\t\t\tacc1 += reductionBuffer[ thread.y + i ][ thread.x ];\n\t\t\treductionBuffer[ thread.y ][ thread.x ] = acc1;\n\t\t}\n\t\tGroupMemoryBarrierWithGroupSync();\n\t}\n\n\tif( thread.y != 0 )\n\t\treturn;\n\n\tconst uint resultPos = group.x * TILE_SIZE;\n\tconst uint outputSize = min( TILE_SIZE, resultSize.x - resultPos );\n\tif( thread.x >= outputSize )\n\t\treturn;\n\n\tconst uint4 resultPos4 = uint4( resultPos + thread.x, 0, layer );\n\tconst uint rdi = hadd4( resultPos4 * resultStrides );\n\tresult[ rdi ] = acc1 + reductionBuffer[ 1 ][ thread.x ];\n}"
},
{
"path": "ComputeShaders/mulMatByScalar.hlsl",
"content": "// Matrix * scalar product, like [ 1, E1, E2, E3 ] * [ 1, 1, E2, E3 ] = [ E1, 1, E2, E3 ]\n// Dispatch [ E2, E3, 1 ] thread groups of this shader\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 arg0Size: packoffset( c0 );\n\tuint4 arg0Strides: packoffset( c1 );\n\tuint4 arg1Size: packoffset( c2 );\n\tuint4 arg1Strides: packoffset( c3 );\n\tuint4 resultSize: packoffset( c4 );\n\tuint4 resultStrides: packoffset( c5 );\n}\n\ninline uint hadd( uint2 vec )\n{\n\treturn vec.x + vec.y;\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst float scalarValue = arg1[ hadd( group.xy * arg1Strides.zw ) ];\n\n\tuint s0 = hadd( group.xy * arg0Strides.zw );\n\tconst uint s0Inc = 32 * arg0Strides.y;\n\ts0 += thread * arg0Strides.y;\n\n\tuint rdi = hadd( group.xy * resultStrides.zw );\n\tconst uint rdiEnd = rdi + arg0Size.y;\n\trdi += thread;\n\n\tfor( ; rdi < rdiEnd; rdi += 32, s0 += s0Inc )\n\t{\n\t\tfloat f = arg0[ s0 ];\n\t\tf *= scalarValue;\n\t\tresult[ rdi ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/mulMatDotMain.hlsl",
"content": "// GGML_TASK_COMPUTE step for matrix*matrix product, where nb01 >= nb00;\n// Dispatch with [ ne11, ne01*ne02*ne03 ] thread groups\n// Each thread group computes a single dot product\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tuint4 src1_elements: packoffset( c2 );\n\tuint4 result_elements: packoffset( c4 );\n\tuint4 result_strides: packoffset( c5 );\n}\n\ninline uint product( uint3 vec )\n{\n\treturn vec.x * vec.y * vec.z;\n}\n\ninline uint product( uint4 vec )\n{\n\tuint2 tmp = vec.xy * vec.zw;\n\treturn tmp.x * tmp.y;\n}\n\ninline float dotProductInner( uint i0, uint i1, uint length, uint thread )\n{\n\tfloat res = 0;\n\tfor( uint i = thread; i < length; i += 32 )\n\t\tres = mad( arg0[ i0 + i ], arg1[ i1 + i ], res );\n\treturn res;\n}\n\n#include \"groupReduce.hlsli\"\n\n[numthreads( 32, 1, 1 )]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint ne00 = src0_elements.x;\n\tconst uint ne01 = src0_elements.y;\n\tconst uint ne02 = src0_elements.z;\n\tconst uint ne03 = src0_elements.w;\n\n\tconst uint ne10 = src1_elements.x;\n\tconst uint ne11 = src1_elements.y;\n\tconst uint ne12 = src1_elements.z;\n\tconst uint ne13 = src1_elements.w;\n\n\tconst int nb00 = src0_strides.x;\n\tconst int nb01 = src0_strides.y;\n\tconst int nb02 = src0_strides.z;\n\tconst int nb03 = src0_strides.w;\n\n\t// total rows in src0\n\t// const int nr = ne01*ne02*ne03;\n\tconst uint nr = product( src0_elements.yzw );\n\n\tconst uint ir = group.y;\n\n\t// src0 indices\n\tconst uint i03 = ir / ( ne02 * ne01 );\n\tconst uint i02 = ( ir - i03 * ne02 * ne01 ) / ne01;\n\tconst uint i01 = ( ir - i03 * ne02 * ne01 - i02 * ne01 );\n\n\tconst uint i13 = i03;\n\tconst uint i12 = i02;\n\n\tconst uint i0 = i01;\n\tconst uint i2 = i02;\n\tconst uint i3 = i03;\n\n\t// src0_row = (ggml_fp16_t *) ((char *) src0->data + (i01*nb01 + i02*nb02 + i03*nb03));\n\t// src1_col = wdata + ( i13 * ne12 * ne11 + i12 * ne11 + 0 ) * ne00;\n\tconst uint src0_row = i01 * nb01 + i02 * nb02 + i03 * nb03;\n\tconst uint src1_col = ( i13 * ne12 * ne11 + i12 * ne11 ) * ne00;\n\n\tconst uint ic = group.x;\n\tfloat curr = dotProductInner( src0_row, src1_col + ic * ne00, ne00, thread );\n\thorizontalSumCompatNew( thread, curr );\n\n\tif( 0 != thread )\n\t\treturn;\n\n\tconst uint nb0 = result_strides.x;\n\tconst uint nb1 = result_strides.y;\n\tconst uint nb2 = result_strides.z;\n\tconst uint nb3 = result_strides.w;\n\n\tconst uint ne0 = result_elements.x;\n\t// float * dst_col = (float *) ((char *) dst->data + (i0*nb0 + 0*nb1 + i2*nb2 + i3*nb3));\n\tconst uint dst_col = i0 * nb0 + i2 * nb2 + i3 * nb3;\n\tresult[ dst_col + ic * ne0 ] = curr;\n}"
},
{
"path": "ComputeShaders/mulMatDotReshape.hlsl",
"content": "// GGML_TASK_INIT step for matrix*matrix product, where nb01 >= nb00;\n// Dispatch with [ ne11, ne12 ] groups\nBuffer arg0: register( t0 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n}\n\n#include \"miscUtils.hlsli\"\n\n// Each thread group of this shader copies a single rows of the matrix\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint i12 = group.y;\n\tconst uint i11 = group.x;\n\tconst uint ne10 = src0_elements.x;\n\tconst uint ne11 = src0_elements.y;\n\tconst uint nb12 = src0_strides.z;\n\tconst uint nb11 = src0_strides.y;\n\n\tuint rdi = i11 * ne10 + i12 * ne10 * ne11;\n\tconst uint rdiEnd = rdi + ne10;\n\tuint rsi = i12 * nb12 + i11 * nb11;\n\trdi += thread;\n\trsi += thread;\n\n\tfor( ; rdi < rdiEnd; rdi += 32, rsi += 32 )\n\t\tresult[ rdi ] = adjustFp16( arg0[ rsi ] );\n}"
},
{
"path": "ComputeShaders/mulMatMadMain.hlsl",
"content": "// GGML_TASK_COMPUTE step for matrix*matrix product, where nb01 < nb00\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer resultTensor: register( u0 );\nRWBuffer tempBuffer: register( u1 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 aSize: packoffset( c0 );\n\tuint4 aStride: packoffset( c1 );\n\tuint4 bSize: packoffset( c2 );\n\tuint4 bStride: packoffset( c3 );\n\tuint4 resSize: packoffset( c4 );\n\tbool resultFp16 : packoffset( c5.x );\n\tuint ne: packoffset( c5.y );\n}\n\n#include \"miscUtils.hlsli\"\n\n// tempBuffer[ rdi .. ] = 0.0\ninline void writeTempZeros( uint rdi, const uint len, const uint thread )\n{\n\tconst uint rdiEnd = rdi + len;\n\tfor( rdi += thread; rdi < rdiEnd; rdi += 32 )\n\t\ttempBuffer[ rdi ] = 0.0;\n}\n\n// tempBuffer[ rdi .. ] += mul * arg0[ rsi .. ]\ninline void vectorMad( uint rsi, uint rdi, const uint len, const float mul, const uint thread )\n{\n\tconst uint rsiEnd = rsi + len;\n\trsi += thread;\n\trdi += thread;\n\tfor( ; rsi < rsiEnd; rsi += 32, rdi += 32 )\n\t{\n\t\tfloat f = tempBuffer[ rdi ];\n\t\tf = mad( mul, arg0[ rsi ], f );\n\t\t[branch]\n\t\tif( resultFp16 )\n\t\t\tf = adjustFp16( f );\n\t\ttempBuffer[ rdi ] = f;\n\t}\n}\n\n// resultTensor[ rdi .. ] = tempBuffer[ rsi .. ]\ninline void copyRow( uint rsi, uint rdi, const uint len, const uint thread )\n{\n\tconst uint rsiEnd = rsi + len;\n\trsi += thread;\n\trdi += thread;\n\tfor( ; rsi < rsiEnd; rsi += 32, rdi += 32 )\n\t{\n\t\tfloat f = tempBuffer[ rsi ];\n\t\tresultTensor[ rdi ] = f;\n\t}\n}\n\n// resultTensor[ rdi .. ] += tempBuffer[ rsi .. ]\ninline void addRow( uint rsi, uint rdi, const uint len, const uint thread )\n{\n\tconst uint rsiEnd = rsi + len;\n\trsi += thread;\n\trdi += thread;\n\tfor( ; rsi < rsiEnd; rsi += 32, rdi += 32 )\n\t{\n\t\tfloat f = resultTensor[ rdi ];\n\t\tf += tempBuffer[ rsi ];\n\t\tresultTensor[ rdi ] = f;\n\t}\n}\n\n[numthreads( 32, 1, 1 )]\nvoid main( const uint3 group: SV_GroupID, const uint thread : SV_GroupIndex )\n{\n\tconst uint i1 = group[ 0 ];\n\tconst uint i2 = group[ 1 ];\n\tconst uint i3 = group[ 2 ];\n\n\tconst uint ne00 = aSize[ 0 ];\n\tconst uint ne01 = aSize[ 1 ];\n\tconst uint ne02 = aSize[ 2 ];\n\tconst uint ne03 = aSize[ 3 ];\n\n\tconst uint ne10 = bSize[ 0 ];\n\tconst uint ne11 = bSize[ 1 ];\n\tconst uint ne12 = bSize[ 2 ];\n\tconst uint ne13 = bSize[ 3 ];\n\n\tconst uint ne0 = resSize[ 0 ];\n\tconst uint ne1 = resSize[ 1 ];\n\tconst uint ne2 = resSize[ 2 ];\n\tconst uint ne3 = resSize[ 3 ];\n\n\tconst uint nb00 = aStride[ 0 ];\n\tconst uint nb01 = aStride[ 1 ];\n\tconst uint nb02 = aStride[ 2 ];\n\tconst uint nb03 = aStride[ 3 ];\n\n\tconst uint nb10 = bStride[ 0 ];\n\tconst uint nb11 = bStride[ 1 ];\n\tconst uint nb12 = bStride[ 2 ];\n\tconst uint nb13 = bStride[ 3 ];\n\n\t// dst_row = wdata + wo + i3*ne2*ne1*ne0 + i2*ne1*ne0 + i1*ne0;\n\tconst uint tempRowThread0 = i3 * ne2 * ne1 * ne0 + i2 * ne1 * ne0 + i1 * ne0;\n\n\t// Faking 4 CPU threads trying to achieve bitwise compatibility with the CPU version\n\tconst uint nth = 4;\n\n\t// GGML_TASK_COMPUTE\n\t{\n\t\t// src0_col = src0->data + ( i00 * nb00 + i02 * nb02 + i03 * nb03 );\n\t\tconst uint aBase = i2 * nb02 + i3 * nb03;\n\t\t// src1_val = * (float *) ((char *) src1->data + (i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13));\n\t\tconst uint bBase = i1 * nb11 + i2 * nb12 + i3 * nb13;\n\n\t\t// total columns in src1\n\t\tconst uint nc = ne10;\n\t\t// columns per thread\n\t\tconst uint dc = ( nc + nth - 1 ) / nth;\n\n\t\tuint tempRow = tempRowThread0;\n\t\tfor( uint ith = 0; ith < nth; ith++, tempRow += ne )\n\t\t{\n\t\t\twriteTempZeros( tempRow, ne01, thread );\n\n\t\t\t// column range for this thread\n\t\t\tconst uint ic0 = dc * ith;\n\t\t\tconst uint ic1 = min( ic0 + dc, nc );\n\n\t\t\tfor( uint ic = ic0; ic < ic1; ic++ )\n\t\t\t{\n\t\t\t\tconst uint idxA = aBase + ic * aStride[ 0 ];\n\t\t\t\tconst uint idxB = bBase + ic * bStride[ 0 ];\n\t\t\t\tconst float bValue = arg1[ idxB ];\n\t\t\t\tvectorMad( idxA, tempRow, ne01, bValue, thread );\n\t\t\t}\n\t\t}\n\t}\n\n\t// GGML_TASK_FINALIZE\n\t{\n\t\tconst uint rdi = tempRowThread0;\n\t\t// const uint rdi = i1 * resSize[ 0 ] + i2 * resSize[ 0 ] * resSize[ 1 ] + i3 * resSize[ 0 ] * resSize[ 1 ] * resSize[ 2 ];\n\t\t// const uint rdi = ( ( i3 * resSize[ 2 ] + i2 ) * resSize[ 1 ] + i1 ) * resSize[ 0 ];\n\n\t\tuint tempRow = tempRowThread0;\n\t\tcopyRow( tempRow, rdi, ne01, thread );\n\n\t\ttempRow += ne;\n\t\tfor( uint ith = 1; ith < nth; ith++, tempRow += ne )\n\t\t\taddRow( tempRow, rdi, ne01, thread );\n\t}\n}"
},
{
"path": "ComputeShaders/mulMatTiled.hlsl",
"content": "// This compute shader implements matrix*matrix product, using tiling and many other tricks to improve the performance\n// This one here is _the_ most expensive shader in the model. Optimized heavily, as a result the readability ain't great.\n\n#ifndef TILE_SIZE\nstatic const uint TILE_SIZE = 32;\n#endif\n#ifndef THREADS_Y\nstatic const uint THREADS_Y = 8;\n#endif\n// The above values have a following constraint: TILE_SIZE = THREADS_Y * N * 4 where N is an integer\n\n#ifndef STREAM_SECOND_MATRIX\n// Funfact: enabling this on 1080Ti ruins the performance, by a factor of 3.5\n#define STREAM_SECOND_MATRIX 0\n#endif\n\n#ifndef LOAD_ORDER\n\n// Load with coalesced loads from global memory whenever possible, store into groupshared buffer with random stores\n// #define LOAD_ORDER bool2( ( 1 == arg0Strides[ 0 ] ) || ( 1 != arg0Strides[ 1 ] ), ( 1 == arg1Strides[ 0 ] ) || ( 1 != arg1Strides[ 1 ] ) )\n\n// Load with random loads from global memory, store into groupshared buffer with coalesced stores\n// On my AMD iGPU inside Ryzen 7 5700G, there's whopping 15% performance win with that tactics, from 6.67 to 5.66 seconds for this shader.\n// My nVidia GPU does about the same\n#define LOAD_ORDER bool2( false, true )\n\n#endif\n\nBuffer arg0: register( t0 );\nBuffer arg1: register( t1 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 arg0Size: packoffset( c0 );\n\tuint4 arg0Strides: packoffset( c1 );\n\tuint4 arg1Strides: packoffset( c3 );\n\tuint4 resultSize: packoffset( c4 );\n\tuint4 resultStrides: packoffset( c5 );\n}\n\ngroupshared float tile0[ TILE_SIZE ][ TILE_SIZE ];\n#if !STREAM_SECOND_MATRIX\ngroupshared float tile1[ TILE_SIZE ][ TILE_SIZE ];\n#endif\n\n// Count of FP32 accumulators we need in every thread of the shader\nstatic const uint heightScalars = TILE_SIZE / THREADS_Y;\n// The local accumulators are float4 vectors, compute count of these vectors\nstatic const uint heightVectors = ( heightScalars + 3 ) / 4;\n\n#if STREAM_SECOND_MATRIX\nvoid multiplyTiles( uint rsi, const uint3 thread, const uint w, const uint h, inout float4 acc[ heightVectors ] )\n{\n\tuint4 rsi4 = ( THREADS_Y * arg1Strides.y ) * uint4( 0, 1, 2, 3 ) + rsi;\n\t[unroll]\n\tfor( uint iv = 0; iv < heightVectors; iv++, rsi4 += THREADS_Y * 4 * arg1Strides.y )\n\t{\n\t\tfloat4 r = 0;\n\t\tuint4 rsiRow = rsi4;\n\t\tfor( uint j = 0; j < w; j++, rsiRow += arg1Strides.x )\n\t\t{\n\t\t\t// One TILE_SIZE * 4 bytes coalesced load, broadcasted into THREADS_Y copies\n\t\t\tconst float s0 = tile0[ j ][ thread.x ];\n\t\t\tfloat4 s1 = 0.0;\n\t\t\t[unroll]\n\t\t\tfor( uint k = 0; k < 4; k++ )\n\t\t\t{\n\t\t\t\tconst uint i = ( iv * 4 + k ) * THREADS_Y + thread.y;\n\t\t\t\tif( i < h )\n\t\t\t\t\ts1[ k ] = arg1[ rsiRow[ k ] ];\n\t\t\t}\n\t\t\t// Multiply and accumulate\n\t\t\tr = mad( s0, s1, r );\n\t\t}\n\t\t// Accumulate into the output tile\n\t\tacc[ iv ] += r;\n\t}\n}\n#else\n// Compute resTemp += tile0 * tile1, for TILE_SIZE^2 square matrices\n// The group size is TILE_SIZE*THREADS_Y threads in this shader\nvoid multiplyTiles( const uint3 thread, inout float4 acc[ heightVectors ] )\n{\n\t[unroll]\n\tfor( uint iv = 0; iv < heightVectors; iv++ )\n\t{\n\t\tfloat4 r = 0;\n\t\tfor( uint j = 0; j < TILE_SIZE; j++ )\n\t\t{\n\t\t\t// One TILE_SIZE * 4 bytes coalesced load, broadcasted into THREADS_Y copies\n\t\t\tconst float s0 = tile0[ j ][ thread.x ];\n\t\t\tfloat4 s1;\n\t\t\t[unroll]\n\t\t\tfor( uint k = 0; k < 4; k++ )\n\t\t\t{\n\t\t\t\tconst uint i = ( iv * 4 + k ) * THREADS_Y + thread.y;\n\t\t\t\t// THREADS_Y broadcasts, each one is 4 bytes broadcasted into TILE_SIZE copies\n\t\t\t\ts1[ k ] = tile1[ i ][ j ];\n\t\t\t}\n\t\t\t// Multiply and accumulate\n\t\t\tr = mad( s0, s1, r );\n\t\t}\n\t\t// Accumulate into the output tile\n\t\tacc[ iv ] += r;\n\t}\n}\n#endif\n\n// Note we transposed these tiles while loading\nvoid loadTile0( uint rsi, const uint3 thread, const uint w, const uint h, const bool rowMajor )\n{\n\tuint i;\n\tif( rowMajor )\n\t{\n\t\trsi += arg0Strides.y * thread.y;\n\t\tfor( i = thread.y; i < h; i += THREADS_Y, rsi += arg0Strides.y * THREADS_Y )\n\t\t{\n\t\t\tif( thread.x < w )\n\t\t\t\ttile0[ thread.x ][ i ] = arg0[ rsi + thread.x * arg0Strides.x ];\n\t\t\telse\n\t\t\t\ttile0[ thread.x ][ i ] = 0.0;\n\t\t}\n\t}\n\telse\n\t{\n\t\t// Unlike width which is smaller for the last tile, the height is always the same, and all these tiles are zero-initialized\n\t\tif( thread.x >= h )\n\t\t\treturn;\n\n\t\trsi += arg0Strides.x * thread.y;\n\t\tfor( i = thread.y; i < w; i += THREADS_Y, rsi += arg0Strides.x * THREADS_Y )\n\t\t\ttile0[ i ][ thread.x ] = arg0[ rsi + thread.x * arg0Strides.y ];\n\n\t\tif( i >= TILE_SIZE )\n\t\t\treturn;\n\t\tfor( ; i < TILE_SIZE; i += THREADS_Y )\n\t\t\ttile0[ i ][ thread.x ] = 0.0;\n\t}\n}\n\n#if !STREAM_SECOND_MATRIX\nvoid loadTile1( uint rsi, const uint3 thread, const uint w, const uint h, const bool rowMajor )\n{\n\tuint i;\n\tif( rowMajor )\n\t{\n\t\trsi += thread.y * arg1Strides.y;\n\n\t\tfor( i = thread.y; i < h; i += THREADS_Y, rsi += arg1Strides.y * THREADS_Y )\n\t\t{\n\t\t\tif( thread.x < w )\n\t\t\t\ttile1[ i ][ thread.x ] = arg1[ rsi + thread.x * arg1Strides.x ];\n\t\t\telse\n\t\t\t\ttile1[ i ][ thread.x ] = 0.0;\n\t\t}\n\t}\n\telse\n\t{\n\t\t// Unlike width which is smaller for the last tile, the height is always the same, and all these tiles are zero-initialized\n\t\tif( thread.x >= h )\n\t\t\treturn;\n\n\t\trsi += thread.y * arg1Strides.x;\n\t\tfor( i = thread.y; i < w; i += THREADS_Y, rsi += arg1Strides.x * THREADS_Y )\n\t\t\ttile1[ thread.x ][ i ] = arg1[ rsi + thread.x * arg0Strides.y ];\n\t\tif( i >= TILE_SIZE )\n\t\t\treturn;\n\t\tfor( ; i < TILE_SIZE; i += THREADS_Y )\n\t\t\ttile1[ thread.x ][ i ] = 0.0;\n\t}\n}\n#endif\n\nvoid storeTile( const uint3 thread, const uint4 pos, const uint2 size, in float4 acc[ heightVectors ] )\n{\n\tif( thread.x >= size.x )\n\t\treturn;\n\n\tconst uint4 prod4 = pos * resultStrides;\n\tconst uint2 prod2 = prod4.xy + prod4.zw;\n\tuint rdi = prod2.x + prod2.y;\n\trdi += resultStrides.y * thread.y;\n\trdi += resultStrides.x * thread.x;\n\n\tconst uint4 offsets = THREADS_Y * uint4( 0, 1, 2, 3 );\t//< a compile-time constant vector\n\tuint4 rdi4 = resultStrides.y * offsets + rdi;\n\n\t[unroll]\n\tfor( uint iv = 0; iv < heightVectors; iv++, rdi4 += resultStrides.y * THREADS_Y * 4 )\n\t{\n\t\tconst float4 source = acc[ iv ];\n\t\t[unroll]\n\t\tfor( uint k = 0; k < 4; k++ )\n\t\t{\n\t\t\tconst uint i = ( iv * 4 + k ) * THREADS_Y + thread.y;\n\t\t\tif( i < size.y )\n\t\t\t\tresult[ rdi4[ k ] ] = source[ k ];\n\t\t}\n\t}\n}\n\n[ numthreads( TILE_SIZE, THREADS_Y, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint3 thread : SV_GroupThreadID )\n{\n\t// Zero out these shared buffers\n\tfor( uint i = 0; i < TILE_SIZE; i += THREADS_Y )\n\t{\n\t\ttile0[ i + thread.y ][ thread.x ] = 0.0;\n#if !STREAM_SECOND_MATRIX\n\t\ttile1[ i + thread.y ][ thread.x ] = 0.0;\n#endif\n\t}\n\t// Despite inside GPU cores, the shared memory is still much slower than registers\n\t// For this reason, this shader accumulates numbers in local variables. Only uses groupshared memory for tiles of the argument matrices.\n\tfloat4 acc[ heightVectors ];\n\t// Zero out the accumulators\n\t[unroll]\n\tfor( i = 0; i < heightVectors; i++ )\n\t\tacc[ i ] = 0.0;\n\n\tconst uint2 resultPos = group.xy * TILE_SIZE;\n\tconst uint2 layer = uint2( group.z % resultSize.z, group.z / resultSize.z );\n\tuint rsi0 = resultPos.x * arg0Strides.y + layer.x * arg0Strides.z + layer.y * arg0Strides.w;\n\tuint rsi1 = resultPos.y * arg1Strides.y + layer.x * arg1Strides.z + layer.y * arg1Strides.w;\n\n\tconst uint rsi0Inc = TILE_SIZE * arg0Strides.x;\n\tconst uint rsi1Inc = TILE_SIZE * arg1Strides.x;\n\n\tconst uint completeTiles = arg0Size.x / TILE_SIZE;\n\tconst uint rsi0AndAligned = rsi0 + rsi0Inc * completeTiles;\n\t// Output tile size\n\t// Normally TILE_SIZE^2, less than that for the tiles at the right and bottom edges of the output matrix\n\tconst uint2 outputSize = min( TILE_SIZE, resultSize.xy - resultPos );\n\n\tconst bool2 loadOrder = LOAD_ORDER;\n\n#if STREAM_SECOND_MATRIX\n\trsi1 += thread.y * arg1Strides.y;\n#endif\n\tfor( ; rsi0 < rsi0AndAligned; rsi0 += rsi0Inc, rsi1 += rsi1Inc )\n\t{\n\t\tloadTile0( rsi0, thread, TILE_SIZE, outputSize.x, loadOrder.x );\n#if STREAM_SECOND_MATRIX\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tmultiplyTiles( rsi1, thread, TILE_SIZE, outputSize.y, acc );\n#else\n\t\tloadTile1( rsi1, thread, TILE_SIZE, outputSize.y, loadOrder.y );\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tmultiplyTiles( thread, acc );\n#endif\n\t\t// Need one moar barrier here.\n\t\t// Otherwise, some threads of the group are loading the next tile into tile0/tile1 groupshared buffers on the next iteration of the loop,\n\t\t// while other threads of the same group are still computing the matrix product, and getting incorrect values from that groupshared buffer.\n\t\t// The missing barrier only caused a bug on AMD, and only with \"ggml-large.bin\" model; no idea why that is.\n\t\tGroupMemoryBarrierWithGroupSync();\n\t}\n\n\tconst uint rem = arg0Size.x % TILE_SIZE;\n\tif( 0 != rem )\n\t{\n\t\tloadTile0( rsi0, thread, rem, outputSize.x, loadOrder.x );\n#if STREAM_SECOND_MATRIX\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tmultiplyTiles( rsi1, thread, rem, outputSize.y, acc );\n#else\n\t\tloadTile1( rsi1, thread, rem, outputSize.y, loadOrder.y );\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tmultiplyTiles( thread, acc );\n#endif\n\t}\n\n\tstoreTile( thread, uint4( resultPos, layer ), outputSize, acc );\n}"
},
{
"path": "ComputeShaders/mulMatTiledEx.hlsl",
"content": "// This compute shader implements yet another version of matrix*matrix product\n// For optimal VRAM access pattern, it requires both arguments to be reshaped into a sequence of horizontal column major panels.\n// The panel height is TILE_SIZE, and the last panel of the matrix needs to be padded with zeros; see matReshapePanels.hlsl shader for the reshaping.\n// So far, it's only used when running on AMD GPUs.\n#ifndef TILE_SIZE\nstatic const uint TILE_SIZE = 32;\n#endif\n#ifndef TILE_HEIGHT\nstatic const uint TILE_HEIGHT = 64;\n#endif\n#ifndef THREADS_Y\nstatic const uint THREADS_Y = 8;\n#endif\n// The above values have a following constraint: TILE_SIZE = THREADS_Y * N * 4 where N is an integer\n\n#ifndef STREAM_SECOND_MATRIX\n#define STREAM_SECOND_MATRIX 1\n#endif\n\n// First tensor, reshaped into dense column major horizontal panels of size [ width, TILE_SIZE ]\nBuffer arg0: register( t0 );\n// Second tensor, reshaped into dense column major horizontal panels of size [ width, TILE_SIZE ]\nBuffer arg1: register( t1 );\n// FP32 output tensor, row major and continuous\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 arg0Size: packoffset( c0 );\n\tuint arg0panel: packoffset( c1.y );\n\tuint2 arg0LayerStrides: packoffset( c1.z );\n\n\t// uint4 arg1Size: packoffset( c2 );\n\tuint arg1panel: packoffset( c3.y );\n\tuint2 arg1LayerStrides: packoffset( c3.z );\n\n\tuint4 resultSize: packoffset( c4 );\n\tuint4 resultStrides: packoffset( c5 );\n}\n\n// A smaller tile loaded from the first source matrix\ngroupshared float tile0[ TILE_HEIGHT ][ TILE_SIZE ];\n#if !STREAM_SECOND_MATRIX\n// A smaller tile loaded from the second source matrix\ngroupshared float tile1[ TILE_HEIGHT ][ TILE_SIZE ];\n#endif\n\n// Count of FP32 accumulators we need in every thread of the shader\nstatic const uint heightScalars = TILE_SIZE / THREADS_Y;\n// The local accumulators are float4 vectors, compute count of these vectors\nstatic const uint heightVectors = ( heightScalars + 3 ) / 4;\n\n#if STREAM_SECOND_MATRIX\nvoid multiplyTiles( const uint3 thread, uint rsi, const uint h, inout float4 acc[ heightVectors ] )\n{\n\tuint4 rsi4 = rsi + uint4( 0, THREADS_Y, THREADS_Y * 2, THREADS_Y * 3 );\n\t[unroll]\n\tfor( uint iv = 0; iv < heightVectors; iv++, rsi4 += THREADS_Y * 4 )\n\t{\n\t\tfloat4 r = 0.0;\n\t\tuint4 rsiRow = rsi4;\n\t\tfor( uint j = 0; j < h; j++, rsiRow += TILE_SIZE )\n\t\t{\n\t\t\tconst float a = tile0[ j ][ thread.x ];\n\t\t\tfloat4 b = 0.0;\n\t\t\t[unroll]\n\t\t\tfor( uint k = 0; k < 4; k++ )\n\t\t\t{\n\t\t\t\tb[ k ] = arg1[ rsiRow[ k ] ];\n\t\t\t}\n\t\t\tr = mad( a, b, r );\n\t\t}\n\t\tacc[ iv ] += r;\n\t}\n}\n#else\nvoid multiplyTiles( const uint3 thread, inout float4 acc[ heightVectors ] )\n{\n\t[unroll]\n\tfor( uint i = 0; i < heightVectors; i++ )\n\t{\n\t\tfloat4 r = 0.0;\n\t\tfor( uint j = 0; j < TILE_HEIGHT; j++ )\n\t\t{\n\t\t\tconst float a = tile0[ j ][ thread.x ];\n\t\t\tfloat4 b;\n\t\t\t[unroll]\n\t\t\tfor( uint k = 0; k < 4; k++ )\n\t\t\t{\n\t\t\t\tconst uint row = ( i * 4 + k ) * THREADS_Y + thread.y;\n\t\t\t\tb[ k ] = tile1[ j ][ row ];\n\t\t\t}\n\t\t\tr = mad( a, b, r );\n\t\t}\n\t\tacc[ i ] += r;\n\t}\n}\n#endif\n\nvoid storeTile( const uint3 thread, const uint4 pos, const uint2 size, in float4 acc[ heightVectors ] )\n{\n\tif( thread.x >= size.x )\n\t\treturn;\n\n\tconst uint4 prod4 = pos * resultStrides;\n\tconst uint2 prod2 = prod4.xy + prod4.zw;\n\tuint rdi = prod2.x + prod2.y;\n\trdi += resultStrides.y * thread.y;\n\trdi += resultStrides.x * thread.x;\n\n\tconst uint4 offsets = THREADS_Y * uint4( 0, 1, 2, 3 );\t//< a compile-time constant vector\n\tuint4 rdi4 = resultStrides.y * offsets + rdi;\n\n\t[unroll]\n\tfor( uint iv = 0; iv < heightVectors; iv++, rdi4 += resultStrides.y * THREADS_Y * 4 )\n\t{\n\t\tconst float4 source = acc[ iv ];\n\t\t[unroll]\n\t\tfor( uint k = 0; k < 4; k++ )\n\t\t{\n\t\t\tconst uint i = ( iv * 4 + k ) * THREADS_Y + thread.y;\n\t\t\tif( i < size.y )\n\t\t\t\tresult[ rdi4[ k ] ] = source[ k ];\n\t\t}\n\t}\n}\n\n[numthreads( TILE_SIZE, THREADS_Y, 1 )]\nvoid main( const uint3 group: SV_GroupID, const uint3 thread : SV_GroupThreadID )\n{\n\tuint i;\n\t// Zero all shared buffers\n\tfor( i = thread.y; i < TILE_HEIGHT; i += THREADS_Y )\n\t{\n\t\ttile0[ i ][ thread.x ] = 0.0;\n#if !STREAM_SECOND_MATRIX\n\t\ttile1[ i ][ thread.x ] = 0.0;\n#endif\n\t}\n\t// Despite inside GPU cores, the shared memory is still much slower than registers\n\t// For this reason, this shader accumulates numbers in local variables. Only uses groupshared memory for tiles of the argument matrices.\n\tfloat4 acc[ heightVectors ];\n\t// Zero out the accumulators\n\t[unroll]\n\tfor( i = 0; i < heightVectors; i++ )\n\t\tacc[ i ] = 0.0;\n\n\tconst uint2 layer = uint2( group.z % resultSize.z, group.z / resultSize.z );\n\n\tuint rsi0 = group.x * arg0panel + layer.x * arg0LayerStrides.x + layer.y * arg0LayerStrides.y;\n\tuint rsi1 = group.y * arg1panel + layer.x * arg1LayerStrides.x + layer.y * arg1LayerStrides.y;\n\n\tconst uint threadOffset = thread.y * TILE_SIZE + thread.x;\n\trsi0 += threadOffset;\n#if STREAM_SECOND_MATRIX\n\trsi1 += thread.y;\n#else\n\trsi1 += threadOffset;\n#endif\n\n\tconst uint completeTiles = arg0Size.x / TILE_HEIGHT;\n\tfor( i = 0; i < completeTiles; i++ )\n\t{\n\t\t// Load [ TILE_SIZE, TILE_HEIGHT ] block from both source tensors into these groupshared buffers\n\t\tfor( uint j = thread.y; j < TILE_HEIGHT; j += THREADS_Y )\n\t\t{\n\t\t\ttile0[ j ][ thread.x ] = arg0[ rsi0 ];\n\t\t\trsi0 += THREADS_Y * TILE_SIZE;\n#if !STREAM_SECOND_MATRIX\n\t\t\ttile1[ j ][ thread.x ] = arg1[ rsi1 ];\n\t\t\trsi1 += THREADS_Y * TILE_SIZE;\n#endif\n\t\t}\n\n\t\t// Wait for all threads in the group to complete these loads\n\t\tGroupMemoryBarrierWithGroupSync();\n\n#if STREAM_SECOND_MATRIX\n\t\tmultiplyTiles( thread, rsi1, TILE_HEIGHT, acc );\n\t\trsi1 += TILE_HEIGHT * TILE_SIZE;\n#else\n\t\t// Multiply + accumulate the elements collected in the groupshared buffers\n\t\tmultiplyTiles( thread, acc );\n#endif\n\t\tGroupMemoryBarrierWithGroupSync();\n\t}\n\n\tconst uint rem = arg0Size.x % TILE_HEIGHT;\n\tif( rem != 0 )\n\t{\n\t\t// Load [ TILE_SIZE, rem ] block from both source tensors, and zero out the padding elements\n\t\tfor( uint j = thread.y; j < TILE_HEIGHT; j += THREADS_Y )\n\t\t{\n\t\t\t[branch]\n\t\t\tif( j < rem )\n\t\t\t{\n\t\t\t\ttile0[ j ][ thread.x ] = arg0[ rsi0 ];\n\t\t\t\trsi0 += THREADS_Y * TILE_SIZE;\n#if !STREAM_SECOND_MATRIX\n\t\t\t\ttile1[ j ][ thread.x ] = arg1[ rsi1 ];\n\t\t\t\trsi1 += THREADS_Y * TILE_SIZE;\n#endif\n\t\t\t}\n\t\t\telse\n\t\t\t{\n\t\t\t\ttile0[ j ][ thread.x ] = 0.0;\n#if !STREAM_SECOND_MATRIX\n\t\t\t\ttile1[ j ][ thread.x ] = 0.0;\n#endif\n\t\t\t}\n\t\t}\n\n\t\t// Wait for all threads in the group to complete these loads\n\t\tGroupMemoryBarrierWithGroupSync();\n\n\t\t// Multiply + accumulate the elements collected in the groupshared buffers\n#if STREAM_SECOND_MATRIX\n\t\tmultiplyTiles( thread, rsi1, rem, acc );\n#else\n\t\tmultiplyTiles( thread, acc );\n#endif\n\t\tGroupMemoryBarrierWithGroupSync();\n\t}\n\n\tconst uint2 resultPos = group.xy * TILE_SIZE;\n\tconst uint2 outputSize = min( TILE_SIZE, resultSize.xy - resultPos );\n\tstoreTile( thread, uint4( resultPos, layer ), outputSize, acc );\n}"
},
{
"path": "ComputeShaders/norm.hlsl",
"content": "// Ported from ggml_compute_forward_norm_f32\n// Dispatch [ ne01, ne02, ne03 ] thread groups of this shader\nBuffer arg0: register( t0 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tuint4 result_strides: packoffset( c3 );\n}\n\nstatic const float eps = 1e-5f; // TODO: make this a parameter\n\n#include \"groupReduce.hlsli\"\n\nfloat computeVectorSum( uint i, const uint length, const uint thread )\n{\n\tfloat res = 0.0;\n\n\tconst uint iEnd = i + length;\n\ti += thread;\n\tfor( ; i < iEnd; i += 32 )\n\t\tres += arg0[ i ];\n\n\thorizontalSumBroadcast( thread, res );\n\treturn res;\n}\n\nfloat offsetAndComputeSumSquares( uint rsi, uint rdi, const float mean, const uint length, const uint thread )\n{\n\tfloat sum2 = 0.0;\n\n\tconst uint rsiEnd = rsi + length;\n\trsi += thread;\n\trdi += thread;\n\tfor( ; rsi < rsiEnd; rsi += 32, rdi += 32 )\n\t{\n\t\tfloat v = arg0[ rsi ] - mean;\n\t\tresult[ rdi ] = v;\n\t\tsum2 = mad( v, v, sum2 );\n\t}\n\n\thorizontalSumBroadcast( thread, sum2 );\n\treturn sum2;\n}\n\nvoid scaleVector( uint rdi, const float scale, const uint length, const uint thread )\n{\n\tconst uint rdiEnd = rdi + length;\n\tfor( rdi += thread; rdi < rdiEnd; rdi += 32 )\n\t{\n\t\tfloat f = result[ rdi ];\n\t\tf *= scale;\n\t\tresult[ rdi ] = f;\n\t}\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint i03 = group.z;\n\tconst uint i02 = group.y;\n\tconst uint i01 = group.x;\n\n\tconst uint nb01 = src0_strides[ 1 ];\n\tconst uint nb02 = src0_strides[ 2 ];\n\tconst uint nb03 = src0_strides[ 3 ];\n\n\tconst uint p = i01 * nb01 + i02 * nb02 + i03 * nb03;\n\n\tconst uint ne00 = src0_elements[ 0 ];\n\n\tfloat mean = computeVectorSum( p, ne00, thread );\n\tmean /= (float)(int)ne00;\n\n\tconst uint nb1 = result_strides[ 1 ];\n\tconst uint nb2 = result_strides[ 2 ];\n\tconst uint nb3 = result_strides[ 3 ];\n\tconst uint y = i01 * nb1 + i02 * nb2 + i03 * nb3;\n\n\tfloat sum2 = offsetAndComputeSumSquares( p, y, mean, ne00, thread );\n\tconst float scale = 1.0 / sqrt( sum2 / (float)(int)ne00 + eps );\n\n\tscaleVector( y, scale, ne00, thread );\n}"
},
{
"path": "ComputeShaders/normCompat.hlsl",
"content": "// Ported from ggml_compute_forward_norm_f32\n// Dispatch [ ( ne01 + 31 ) / 32, ne02, ne03 ] thread groups of this shader\nBuffer arg0: register( t0 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tuint4 result_strides: packoffset( c3 );\n}\n\nstatic const double eps = 1e-5; // TODO: make this a parameter\n\n#include \"groupReduce.hlsli\"\n\ndouble computeVectorSum( uint i, const uint length )\n{\n\tdouble res = 0.0;\n\tconst uint iEnd = i + length;\n\tfor( ; i < iEnd; i++ )\n\t\tres += arg0[ i ];\n\treturn res;\n}\n\ndouble offsetAndComputeSumSquares( uint rsi, uint rdi, const double mean, const uint length )\n{\n\tprecise double sum2 = 0.0;\n\tconst uint rsiEnd = rsi + length;\n\tfor( ; rsi < rsiEnd; rsi++, rdi++ )\n\t{\n\t\tdouble v = arg0[ rsi ];\n\t\tv -= mean;\n\t\tresult[ rdi ] = (float)v;\n\t\tdouble prod = v * v;\n\t\tsum2 += prod;\n\t}\n\treturn sum2;\n}\n\nvoid scaleVector( uint rdi, const float scale, const uint length )\n{\n\tconst uint rdiEnd = rdi + length;\n\tfor( ; rdi < rdiEnd; rdi++ )\n\t{\n\t\tfloat f = result[ rdi ];\n\t\tf *= scale;\n\t\tresult[ rdi ] = f;\n\t}\n}\n\n#include \"fp64Utils.hlsli\"\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 dtid: SV_DispatchThreadID )\n{\n\tconst uint i03 = dtid.z;\n\tconst uint i02 = dtid.y;\n\tconst uint i01 = dtid.x;\n\tif( i01 >= src0_elements[ 1 ] )\n\t\treturn;\n\n\tconst uint nb01 = src0_strides[ 1 ];\n\tconst uint nb02 = src0_strides[ 2 ];\n\tconst uint nb03 = src0_strides[ 3 ];\n\n\tconst uint p = i01 * nb01 + i02 * nb02 + i03 * nb03;\n\tconst uint ne00 = src0_elements[ 0 ];\n\n\tdouble mean = computeVectorSum( p, ne00 );\n\tmean = div64( mean, (double)(int)ne00 );\n\n\tconst uint nb1 = result_strides[ 1 ];\n\tconst uint nb2 = result_strides[ 2 ];\n\tconst uint nb3 = result_strides[ 3 ];\n\tconst uint y = i01 * nb1 + i02 * nb2 + i03 * nb3;\n\n\tconst double sum2 = offsetAndComputeSumSquares( p, y, mean, ne00 );\n\tconst float scale = (float)div64( 1.0, sqrt64( sum2 / (float)(int)ne00 + eps ) );\n\n\tscaleVector( y, scale, ne00 );\n}"
},
{
"path": "ComputeShaders/normFixed.hlsl",
"content": "// Ported from ggml_compute_forward_norm_f32\n// Dispatch [ ne01, ne02, ne03 ] thread groups of this shader\nBuffer arg0: register( t0 );\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tuint4 result_strides: packoffset( c3 );\n}\n\nstatic const float eps = 1e-5f; // TODO: make this a parameter\n\n// #include \"groupReduce.hlsli\"\n\n#ifndef THREADS\nstatic const uint THREADS = 32;\n#endif\nstatic const uint ROW_LENGTH = 1024;\ngroupshared float rowBuffer[ ROW_LENGTH ];\n\nstatic const uint REDUCTION_BUFFER = 32;\ngroupshared float sharedAccumulators[ REDUCTION_BUFFER ];\n\n// Compute horisontal sum of the numbers. The result is only correct on the thread #0 of the group.\nvoid horizontalSum( const uint thread, inout float sum )\n{\n\tif( THREADS > REDUCTION_BUFFER )\n\t{\n\t\tfor( uint t = REDUCTION_BUFFER; t < THREADS; t += REDUCTION_BUFFER )\n\t\t{\n\t\t\t// Threads [ t .. t + REDUCTION_BUFFER ] store into the buffer\n\t\t\tif( thread >= t && thread < t + REDUCTION_BUFFER )\n\t\t\t\tsharedAccumulators[ thread - t ] = sum;\n\n\t\t\tGroupMemoryBarrierWithGroupSync();\n\n\t\t\t// Threads [ 0 .. REDUCTION_BUFFER ] increment their local sum with the value loaded from the buffer\n\t\t\tif( thread < REDUCTION_BUFFER )\n\t\t\t\tsum += sharedAccumulators[ thread ];\n\t\t}\n\t}\n\n\tif( thread < REDUCTION_BUFFER )\n\t\tsharedAccumulators[ thread ] = sum;\n\n\tfor( uint i = REDUCTION_BUFFER / 2; i > 1; i /= 2 )\n\t{\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tif( thread < i )\n\t\t{\n\t\t\tsum += sharedAccumulators[ thread + i ];\n\t\t\tsharedAccumulators[ thread ] = sum;\n\t\t}\n\t}\n\n\tGroupMemoryBarrierWithGroupSync();\n\tif( 0 == thread )\n\t\tsum += sharedAccumulators[ 1 ];\n}\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint i03 = group.z;\n\tconst uint i02 = group.y;\n\tconst uint i01 = group.x;\n\tconst uint ne00 = ROW_LENGTH;\n\n\t// First pass: copy the data to local buffer, and compute sum\n\t{\n\t\tconst uint nb01 = src0_strides[ 1 ];\n\t\tconst uint nb02 = src0_strides[ 2 ];\n\t\tconst uint nb03 = src0_strides[ 3 ];\n\t\tconst uint p = i01 * nb01 + i02 * nb02 + i03 * nb03;\n\n\t\tfloat sum = 0;\n\t\tfor( uint i = thread; i < ne00; i += THREADS )\n\t\t{\n\t\t\tfloat f = arg0[ p + i ];\n\t\t\trowBuffer[ i ] = f;\n\t\t\tsum += f;\n\t\t}\n\t\thorizontalSum( thread, sum );\n\t\tif( 0 == thread )\n\t\t\tsharedAccumulators[ 0 ] = sum / (float)(int)ne00;\n\t\tGroupMemoryBarrierWithGroupSync();\n\t}\n\n\t// Second pass: offset and compute sum of squares\n\t{\n\t\tconst float mean = sharedAccumulators[ 0 ];\n\t\tfloat sum2 = 0;\n\t\tfor( uint i = thread; i < ne00; i += THREADS )\n\t\t{\n\t\t\tfloat v = rowBuffer[ i ];\n\t\t\tv -= mean;\n\t\t\trowBuffer[ i ] = v;\n\t\t\tsum2 = mad( v, v, sum2 );\n\t\t}\n\t\thorizontalSum( thread, sum2 );\n\t\tif( 0 == thread )\n\t\t\tsharedAccumulators[ 0 ] = 1.0 / sqrt( sum2 / (float)(int)ne00 + eps );\n\t\tGroupMemoryBarrierWithGroupSync();\n\t}\n\n\t// Final pass: apply the scale, and copy from group shared buffer to the destination\n\t{\n\t\tconst float scale = sharedAccumulators[ 0 ];\n\n\t\tconst uint nb1 = result_strides[ 1 ];\n\t\tconst uint nb2 = result_strides[ 2 ];\n\t\tconst uint nb3 = result_strides[ 3 ];\n\t\tconst uint y = i01 * nb1 + i02 * nb2 + i03 * nb3;\n\n\t\tfor( uint i = thread; i < ne00; i += THREADS )\n\t\t{\n\t\t\tfloat v = rowBuffer[ i ];\n\t\t\tv *= scale;\n\t\t\tresult[ y + i ] = v;\n\t\t}\n\t}\n}"
},
{
"path": "ComputeShaders/normFixed64.hlsl",
"content": "#define THREADS 64\n#include \"normFixed.hlsl\""
},
{
"path": "ComputeShaders/repeatUtils.hlsli",
"content": "inline uint rowOffset( uint3 idx, uint4 strides )\n{\n\treturn idx[ 0 ] * strides[ 1 ] + idx[ 1 ] * strides[ 2 ] + idx[ 2 ] * strides[ 3 ];\n}\n\n// Initial iterator state for a row of the output tensor\n// x = current index, y = index increment, z = end of the index\ninline uint3 tensorIteratorState( uint3 group, uint thread, uint4 size, uint4 stride )\n{\n\tuint3 res;\n\tres.x = rowOffset( group, stride );\n\tres.y = THREADS * stride[ 0 ];\n\tres.z = res.x + size[ 0 ] * stride[ 0 ];\n\tres.x += thread * stride[ 0 ];\n\treturn res;\n}\n\n// Handle a complete row of output tensor, using the iterator made by tensorIteratorState() function\n#define ROW_LOOP( ts ) for( ; ts.x < ts.z; ts.x += ts.y )\n// Same as above, using different row length\n#define ROW_LOOP_EX( ts, len, stride ) for( ; ts.x < ts.z; ts.x += len * stride[ 0 ] )"
},
{
"path": "ComputeShaders/scaleInPlace.hlsl",
"content": "RWBuffer buffer: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 src0_elements: packoffset( c0 );\n\tuint4 src0_strides: packoffset( c1 );\n\tfloat multiplier: packoffset( c2.x );\n}\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint nc0 = src0_elements[ 0 ];\n\tuint i = group.x * src0_strides[ 1 ];\n\tconst uint iEnd = i + nc0;\n\tconst float mul = multiplier;\n\tfor( i += thread; i < iEnd; i += 32 )\n\t{\n\t\tfloat f = buffer[ i ];\n\t\tf *= mul;\n\t\tbuffer[ i ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/softMax.hlsl",
"content": "// Dispatch [ nr, 1, 1 ] thread groups of this shader\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 elements: packoffset( c0 );\n\tuint4 strides: packoffset( c1 );\n\tuint nr: packoffset( c2.x );\n\tfloat inputScale: packoffset( c2.y );\n}\n\n#ifndef THREADS\nstatic const uint THREADS = 32;\n#endif\n\ngroupshared float sharedAccumulators[ THREADS ];\n\n// Compute horizontal maximum of the numbers, and broadcast to all threads of the group.\nvoid horizontalMaxBroadcast( const uint thread, inout float ax )\n{\n\tsharedAccumulators[ thread ] = ax;\n\tfor( uint i = THREADS / 2; i > 0; i /= 2 )\n\t{\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tif( thread < i )\n\t\t{\n\t\t\tax = max( ax, sharedAccumulators[ thread + i ] );\n\t\t\tsharedAccumulators[ thread ] = ax;\n\t\t}\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\tax = sharedAccumulators[ 0 ];\n}\n\n// Compute horisontal sum of the numbers. The result is only correct on the thread #0 of the group.\nvoid horizontalSum( const uint thread, inout float sum )\n{\n\tsharedAccumulators[ thread ] = sum;\n\tfor( uint i = THREADS / 2; i > 1; i /= 2 )\n\t{\n\t\tGroupMemoryBarrierWithGroupSync();\n\t\tif( thread < i )\n\t\t{\n\t\t\tsum += sharedAccumulators[ thread + i ];\n\t\t\tsharedAccumulators[ thread ] = sum;\n\t\t}\n\t}\n\tGroupMemoryBarrierWithGroupSync();\n\tif( 0 == thread )\n\t\tsum += sharedAccumulators[ 1 ];\n}\n\nstatic const float negativeInfinity = asfloat( 0xff800000 );\n\n[numthreads( THREADS, 1, 1 )]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint p = group.x * strides[ 1 ];\n\tconst uint nc = elements[ 0 ];\n\tconst uint pEnd = p + nc;\n\tuint i;\n\n\tfloat m = negativeInfinity;\n\tfor( i = p + thread; i < pEnd; i += THREADS )\n\t\tm = max( m, result[ i ] );\n\thorizontalMaxBroadcast( thread, m );\n\n\tfloat sum = 0;\n\tfor( i = p + thread; i < pEnd; i += THREADS )\n\t{\n\t\tfloat f = result[ i ];\n\n\t\t[branch]\n\t\tif( f != negativeInfinity )\n\t\t{\n\t\t\tf = ( f - m ) * inputScale;\n\t\t\t// On both Radeon Graphics and nVidia 1080Ti, computing the exponent is slightly faster than loading from the lookup table\n\t\t\tf = exp( f );\n\t\t\tsum += f;\n\t\t}\n\t\telse\n\t\t\tf = 0;\n\n\t\tresult[ i ] = f;\n\t}\n\n\thorizontalSum( thread, sum );\n\tif( 0 == thread )\n\t\tsharedAccumulators[ 0 ] = 1.0 / sum;\n\tGroupMemoryBarrierWithGroupSync();\n\tconst float scale = sharedAccumulators[ 0 ];\n\n\t// ggml_vec_scale_f32\n\tfor( i = p + thread; i < pEnd; i += THREADS )\n\t{\n\t\tfloat f = result[ i ];\n\t\tf *= scale;\n\t\tresult[ i ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/softMax64.hlsl",
"content": "#define THREADS 64\n#include \"softMax.hlsl\""
},
{
"path": "ComputeShaders/softMaxCompat.hlsl",
"content": "// ggml_compute_forward_soft_max_f32\n// Dispatch [ ( nr + 31 ) / 32, 1, 1 ] thread groups of this shader\nRWBuffer result: register( u0 );\n\n// table_exp_f16\nBuffer lookupTable: register( t0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 elements: packoffset( c0 );\n\tuint4 strides: packoffset( c1 );\n\tuint nr: packoffset( c2.x );\n}\n\n#include \"miscUtils.hlsli\"\n#include \"fp64Utils.hlsli\"\n\nstatic const float negativeInfinity = asfloat( 0xff800000 );\n\n[ numthreads( 32, 1, 1 ) ]\nvoid main( uint3 dtid: SV_DispatchThreadID )\n{\n\tif( dtid.x >= nr )\n\t\treturn;\n\n\tconst uint p = dtid.x * strides[ 1 ];\n\tconst uint nc = elements[ 0 ];\n\tconst uint pEnd = p + nc;\n\tuint i;\n\n\tfloat m = negativeInfinity;\n\tfor( i = p; i < pEnd; i++ )\n\t\tm = max( m, result[ i ] );\n\n\tdouble sum = 0;\n\tfor( i = p; i < pEnd; i++ )\n\t{\n\t\tfloat f = result[ i ];\n\n\t\t[branch]\n\t\tif( f != negativeInfinity )\n\t\t{\n\t\t\tuint s = fp16Rounded( f - m );\n\t\t\ts = lookupTable[ s ];\n\t\t\tf = f16tof32( s );\n\t\t\tsum += f;\n\t\t}\n\t\telse\n\t\t\tf = 0;\n\n\t\tresult[ i ] = f;\n\t}\n\n\tconst float scale = (float)div64( 1.0, sum );\n\t// ggml_vec_scale_f32\n\tfor( i = p; i < pEnd; i++ )\n\t{\n\t\tfloat f = result[ i ];\n\t\tf *= scale;\n\t\tresult[ i ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/softMaxFixed.hlsl",
"content": "// Special softMax shader for matrices with rows of 1500 elements.\n// Uses group shared buffer of that length to save global memory bandwidth, more than 2x faster than the original.\n// Dispatch [ nr, 1, 1 ] thread groups of this shader\nRWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint4 elements: packoffset( c0 );\n\tuint4 strides: packoffset( c1 );\n\tuint nr: packoffset( c2.x );\n\tfloat inputScale: packoffset( c2.y );\n}\n\n#include \"miscUtils.hlsli\"\n#include \"groupReduce64.hlsli\"\n\nstatic const uint THREADS = 64;\nstatic const uint ROW_LENGTH = 1500;\ngroupshared float rowBuffer[ ROW_LENGTH ];\n\nstatic const float negativeInfinity = asfloat( 0xff800000 );\n\n[ numthreads( THREADS, 1, 1 ) ]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tconst uint p = group.x * strides[ 1 ];\n\tconst uint nc = ROW_LENGTH;\n\tuint i;\n\n\tfloat m = negativeInfinity;\n\t// First pass: compute maximum, and copy the row into the group shared buffer\n\tfor( i = thread; i < nc; i += THREADS )\n\t{\n\t\tfloat f = result[ p + i ];\n\t\tm = max( m, f );\n\t\trowBuffer[ i ] = f;\n\t}\n\thorizontalMaxBroadcast( thread, m );\n\n\t// Second pass: apply initial scale, compute the exponent, and compute total sum over the row\n\tfloat sum = 0;\n\tfor( i = thread; i < nc; i += THREADS )\n\t{\n\t\tfloat f = rowBuffer[ i ];\n\n\t\t[branch]\n\t\tif( f != negativeInfinity )\n\t\t{\n\t\t\tf = ( f - m ) * inputScale;\n#if 1\n\t\t\t// At least on Radeon Graphics GPU inside Ryzen 7 5700G, computing exponent instead of loading from the buffer improves the performance\n\t\t\tf = exp( f );\n#else\n\t\t\tuint s = fp16Rounded( f );\n\t\t\ts = lookupTable[ s ];\n\t\t\tf = f16tof32( s );\n#endif\n\t\t\tsum += f;\n\t\t}\n\t\telse\n\t\t\tf = 0;\n\n\t\trowBuffer[ i ] = f;\n\t}\n\n\thorizontalSum( thread, sum );\n\tif( 0 == thread )\n\t\tsharedAccumulators[ 0 ] = 1.0 / sum;\n\tGroupMemoryBarrierWithGroupSync();\n\tconst float scale = sharedAccumulators[ 0 ];\n\n\t// Final pass: apply the final scale, and copy the row from the group shared buffer back into the global memory\n\tfor( i = thread; i < nc; i += THREADS )\n\t{\n\t\tfloat f = rowBuffer[ i ];\n\t\tf *= scale;\n\t\tresult[ p + i ] = f;\n\t}\n}"
},
{
"path": "ComputeShaders/softMaxLong.hlsl",
"content": "// This version is for the \"dec.probs\" shader tag\n// The input tensor has a size [ 51865, 3 ], a very long tensor with just 3 rows.\n// Despite the shader only runs on 3 GPU cores, large count of threads helps substantially, this shader is about 50% faster.\n#define THREADS 1024\n\n#include \"softMax.hlsl\""
},
{
"path": "ComputeShaders/zeroMemory.hlsl",
"content": "RWBuffer result: register( u0 );\n\ncbuffer Constants: register( b0 )\n{\n\tuint elements: packoffset( c0.x );\n\tbool writeNan: packoffset( c0.y );\n}\n\n// Thread group index is 16 bits per coordinate:\n// https://learn.microsoft.com/en-us/windows/win32/api/d3d11/nf-d3d11-id3d11devicecontext-dispatch\n// We want this shader to support buffers up to 2 GB.\n#ifndef THREADS\nstatic const uint THREADS = 512;\n#endif\n#ifndef ITERATIONS\nstatic const uint ITERATIONS = 128;\n#endif\n\nstatic const uint itemsPerGroup = THREADS * ITERATIONS;\n\n[numthreads( THREADS, 1, 1 )]\nvoid main( uint3 group: SV_GroupID, uint thread : SV_GroupIndex )\n{\n\tuint rdi = group.x * itemsPerGroup;\n\tconst uint rdiEnd = min( rdi + itemsPerGroup, elements );\n\t// https://www.h-schmidt.net/FloatConverter/IEEE754.html\n\tconst float pattern = writeNan ? asfloat( 0x7FFFFFFFu ) : 0.0;\n\tfor( rdi += thread; rdi < rdiEnd; rdi += THREADS )\n\t\tresult[ rdi ] = pattern;\n}"
},
{
"path": "Examples/MicrophoneCS/CaptureThread.cs",
"content": "using System.Runtime.ExceptionServices;\nusing Whisper;\n\nnamespace MicrophoneCS\n{\n\tsealed class CaptureThread: CaptureCallbacks\n\t{\n\t\tpublic CaptureThread( CommandLineArgs args, Context context, iAudioCapture source )\n\t\t{\n\t\t\tcallbacks = new TranscribeCallbacks( args );\n\t\t\tthis.context = context;\n\t\t\tthis.source = source;\n\n\t\t\tthread = new Thread( threadMain ) { Name = \"Capture Thread\" };\n\t\t\tConsole.WriteLine( \"Press any key to quit\" );\n\t\t\tthread.Start();\n\t\t}\n\n\t\tstatic void readKeyCallback( object? state )\n\t\t{\n\t\t\tCaptureThread ct = ( state as CaptureThread ) ?? throw new ApplicationException();\n\t\t\tConsole.ReadKey();\n\t\t\tct.shouldQuit = true;\n\t\t}\n\n\t\tpublic void join()\n\t\t{\n\t\t\tThreadPool.QueueUserWorkItem( readKeyCallback, this );\n\t\t\tthread.Join();\n\t\t\tedi?.Throw();\n\t\t}\n\n\t\tvolatile bool shouldQuit = false;\n\n\t\tprotected override bool shouldCancel( Context sender ) =>\n\t\t\tshouldQuit;\n\n\t\tprotected override void captureStatusChanged( Context sender, eCaptureStatus status )\n\t\t{\n\t\t\tConsole.WriteLine( $\"CaptureStatusChanged: {status}\" );\n\t\t}\n\n\t\treadonly TranscribeCallbacks callbacks;\n\t\treadonly Thread thread;\n\t\treadonly Context context;\n\t\treadonly iAudioCapture source;\n\t\tExceptionDispatchInfo? edi = null;\n\n\t\tvoid threadMain()\n\t\t{\n\t\t\ttry\n\t\t\t{\n\t\t\t\tcontext.runCapture( source, callbacks, this );\n\t\t\t}\n\t\t\tcatch( Exception ex )\n\t\t\t{\n\t\t\t\tedi = ExceptionDispatchInfo.Capture( ex );\n\t\t\t}\n\t\t}\n\t}\n}"
},
{
"path": "Examples/MicrophoneCS/CommandLineArgs.cs",
"content": "using System.Globalization;\nusing System.Reflection;\nusing Whisper;\n\nnamespace MicrophoneCS\n{\n\tsealed record class CommandLineArgs\n\t{\n\t\tpublic int n_threads = Environment.ProcessorCount;\n\t\tpublic int offset_t_ms = 0;\n\t\tpublic int offset_n = 0;\n\t\tpublic int duration_ms = 0;\n\t\tpublic int max_context = -1;\n\t\tpublic int max_len = 0;\n\n\t\tpublic float word_thold = 0.01f;\n\n\t\tpublic bool speed_up = false;\n\t\tpublic bool translate = false;\n\t\tpublic bool diarize = false;\n\t\tpublic bool output_txt = false;\n\t\tpublic bool print_special = false;\n\t\tpublic bool print_progress = false;\n\t\tpublic bool print_colors = true;\n\t\tpublic bool no_timestamps = false;\n\t\tpublic int[]? prompt = null;\n\t\tpublic int captureDeviceIndex = 0;\n\n\t\tpublic eLanguage language = eLanguage.English;\n\t\tpublic string model = string.Empty;\n\n\t\tconst bool output_wts = false;\n\t\tpublic bool listDevices = false;\n\n\t\tpublic void apply( ref Parameters p )\n\t\t{\n\t\t\tp.setFlag( eFullParamsFlags.PrintRealtime, false );\n\t\t\tp.setFlag( eFullParamsFlags.PrintProgress, print_progress );\n\t\t\tp.setFlag( eFullParamsFlags.PrintTimestamps, !no_timestamps );\n\t\t\tp.setFlag( eFullParamsFlags.PrintSpecial, print_special );\n\t\t\tp.setFlag( eFullParamsFlags.Translate, translate );\n\t\t\tp.language = language;\n\t\t\tp.cpuThreads = n_threads;\n\t\t\tif( max_context >= 0 )\n\t\t\t\tp.n_max_text_ctx = max_context;\n\t\t\tp.offset_ms = offset_t_ms;\n\t\t\tp.duration_ms = duration_ms;\n\t\t\tp.setFlag( eFullParamsFlags.TokenTimestamps, output_wts || max_len > 0 );\n\t\t\tp.thold_pt = word_thold;\n\t\t\tp.max_len = output_wts && max_len == 0 ? 60 : max_len;\n\t\t\tp.setFlag( eFullParamsFlags.SpeedupAudio, speed_up );\n\t\t}\n\n\t\tpublic eResultFlags resultFlags()\n\t\t{\n\t\t\teResultFlags flags = eResultFlags.None;\n\t\t\tbool wts = output_wts || max_len > 0;\n\t\t\tif( !no_timestamps || wts )\n\t\t\t\tflags |= eResultFlags.Timestamps;\n\t\t\tif( wts || print_colors )\n\t\t\t\tflags |= eResultFlags.Tokens;\n\t\t\treturn flags;\n\t\t}\n\n\t\tstatic eLanguage parseLanguage( string lang ) =>\n\t\t\tLibrary.languageFromCode( lang ) ?? throw new ArgumentException( $\"Unknown language code \\\"{lang}\\\"\" );\n\n\t\tpublic CommandLineArgs( string[] argv )\n\t\t{\n\t\t\tfor( int i = 0; i < argv.Length; i++ )\n\t\t\t{\n\t\t\t\tstring arg = argv[ i ];\n\t\t\t\tif( arg == \"-h\" || arg == \"--help\" )\n\t\t\t\t{\n\t\t\t\t\tprintUsage();\n\t\t\t\t\tthrow new OperationCanceledException();\n\t\t\t\t}\n\t\t\t\telse if( arg == \"-c\" || arg == \"--capture\" ) captureDeviceIndex = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-ld\" || arg == \"--list-devices\" ) listDevices = true;\n\t\t\t\telse if( arg == \"-t\" || arg == \"--threads\" ) n_threads = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-ot\" || arg == \"--offset-t\" ) offset_t_ms = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-on\" || arg == \"--offset-n\" ) offset_n = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-d\" || arg == \"--duration\" ) duration_ms = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-mc\" || arg == \"--max-context\" ) max_context = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-ml\" || arg == \"--max-len\" ) max_len = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-wt\" || arg == \"--word-thold\" ) word_thold = float.Parse( argv[ ++i ], CultureInfo.InvariantCulture );\n\t\t\t\telse if( arg == \"-su\" || arg == \"--speed-up\" ) speed_up = true;\n\t\t\t\telse if( arg == \"-tr\" || arg == \"--translate\" ) translate = true;\n\t\t\t\telse if( arg == \"-di\" || arg == \"--diarize\" ) diarize = true;\n\t\t\t\telse if( arg == \"-otxt\" || arg == \"--output-txt\" ) output_txt = true;\n\t\t\t\telse if( arg == \"-ps\" || arg == \"--print-special\" ) print_special = true;\n\t\t\t\telse if( arg == \"-nc\" || arg == \"--no-colors\" ) print_colors = false;\n\t\t\t\telse if( arg == \"-pp\" || arg == \"--print-progress\" ) print_progress = true;\n\t\t\t\telse if( arg == \"-nt\" || arg == \"--no-timestamps\" ) no_timestamps = true;\n\t\t\t\telse if( arg == \"-l\" || arg == \"--language\" ) language = parseLanguage( argv[ ++i ] );\n\t\t\t\telse if( arg == \"--prompt\" ) prompt = parsePrompt( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-m\" || arg == \"--model\" ) model = argv[ ++i ];\n\t\t\t\telse\n\t\t\t\t\tthrow new ArgumentException( $\"Unknown argument: \\\"{arg}\\\"\" );\n\t\t\t}\n\t\t\tif( listDevices )\n\t\t\t\treturn;\n\t\t\tif( string.IsNullOrWhiteSpace( model ) )\n\t\t\t\tthrow new ArgumentException( \"The model file is not provided in the arguments\" );\n\t\t\tif( !File.Exists( model ) )\n\t\t\t\tthrow new FileNotFoundException( \"Model not found\", model );\n\t\t}\n\n\t\tstatic string cstr( bool b ) => b.ToString();\n\n\t\tstatic int[]? parsePrompt( string str )\n\t\t{\n\t\t\tif( string.IsNullOrWhiteSpace( str ) )\n\t\t\t\treturn null;\n\t\t\t// TODO: expose whisper_tokenize function, as a method of iModel COM interface\n\t\t\tthrow new NotImplementedException();\n\t\t}\n\n\t\tvoid printUsage()\n\t\t{\n\t\t\tConsole.WriteLine();\n\n\t\t\tConsole.WriteLine( \"usage: {0} [options] file0.mp3 file1.wma ...\", Path.GetFileName( Assembly.GetExecutingAssembly().Location ) );\n\t\t\tConsole.WriteLine();\n\t\t\tConsole.WriteLine( \"options:\" );\n\t\t\tConsole.WriteLine( \" -h, --help [default] show this help message and exit\" );\n\t\t\tConsole.WriteLine( \" -t N, --threads N [{0,-7:D}] number of threads to use during computation\", n_threads );\n\t\t\tConsole.WriteLine( \" -ot N, --offset-t N [{0,-7:D}] time offset in milliseconds\", offset_t_ms );\n\t\t\tConsole.WriteLine( \" -on N, --offset-n N [{0,-7:D}] segment index offset\", offset_n );\n\t\t\tConsole.WriteLine( \" -d N, --duration N [{0,-7:D}] duration of audio to process in milliseconds\", duration_ms );\n\t\t\tConsole.WriteLine( \" -mc N, --max-context N [{0,-7:D}] maximum number of text context tokens to store\", max_context );\n\t\t\tConsole.WriteLine( \" -ml N, --max-len N [{0,-7:D}] maximum segment length in characters\", max_len );\n\t\t\tConsole.WriteLine( \" -wt N, --word-thold N [{0,-7:F2}] word timestamp probability threshold\", word_thold );\n\t\t\tConsole.WriteLine( \" -su, --speed-up [{0,-7}] speed up audio by x2 (reduced accuracy)\", cstr( speed_up ) );\n\t\t\tConsole.WriteLine( \" -tr, --translate [{0,-7}] translate from source language to english\", cstr( translate ) );\n\t\t\tConsole.WriteLine( \" -di, --diarize [{0,-7}] stereo audio diarization\", cstr( diarize ) );\n\t\t\tConsole.WriteLine( \" -otxt, --output-txt [{0,-7}] output result in a text file\", cstr( output_txt ) );\n\t\t\tConsole.WriteLine( \" -ps, --print-special [{0,-7}] print special tokens\", cstr( print_special ) );\n\t\t\tConsole.WriteLine( \" -nc, --no-colors [{0,-7}] do not print colors\", cstr( !print_colors ) );\n\t\t\tConsole.WriteLine( \" -nt, --no-timestamps [{0,-7}] do not print timestamps\", cstr( no_timestamps ) );\n\t\t\tConsole.WriteLine( \" -l LANG, --language LANG [{0,-7}] spoken language\", language.getCode() );\n\t\t\tConsole.WriteLine( \" --prompt PROMPT [ ] initial prompt\" );\n\t\t\tConsole.WriteLine( \" -m FNAME, --model FNAME [{0,-7}] model path\", model );\n\t\t\tConsole.WriteLine( \" -f FNAME, --file FNAME [{0,-7}] path of the input audio file\", \"\" );\n\t\t}\n\t}\n}"
},
{
"path": "Examples/MicrophoneCS/MicrophoneCS.cs",
"content": "using Whisper;\n\nnamespace MicrophoneCS\n{\n\tstatic class Program\n\t{\n\t\tstatic int Main( string[] args )\n\t\t{\n\t\t\ttry\n\t\t\t{\n\t\t\t\tCommandLineArgs cla;\n\t\t\t\ttry\n\t\t\t\t{\n\t\t\t\t\tcla = new CommandLineArgs( args );\n\t\t\t\t}\n\t\t\t\tcatch( OperationCanceledException )\n\t\t\t\t{\n\t\t\t\t\treturn 1;\n\t\t\t\t}\n\t\t\t\tconst eLoggerFlags loggerFlags = eLoggerFlags.UseStandardError | eLoggerFlags.SkipFormatMessage;\n\t\t\t\tLibrary.setLogSink( eLogLevel.Debug, loggerFlags );\n\n\t\t\t\tusing iMediaFoundation mf = Library.initMediaFoundation();\n\t\t\t\tCaptureDeviceId[] devices = mf.listCaptureDevices() ??\n\t\t\t\t\tthrow new ApplicationException( \"This computer has no audio capture devices\" );\n\n\t\t\t\tif( cla.listDevices )\n\t\t\t\t{\n\t\t\t\t\tfor( int i = 0; i < devices.Length; i++ )\n\t\t\t\t\t\tConsole.WriteLine( \"#{0}: {1}\", i, devices[ i ].displayName );\n\t\t\t\t\treturn 0;\n\t\t\t\t}\n\t\t\t\tif( cla.captureDeviceIndex < 0 || cla.captureDeviceIndex >= devices.Length )\n\t\t\t\t\tthrow new ApplicationException( $\"Capture device index is out of range; the valid range is [ 0 .. {devices.Length - 1} ]\" );\n\n\t\t\t\tsCaptureParams cp = new sCaptureParams( true );\n\t\t\t\tif( cla.diarize )\n\t\t\t\t\tcp.flags |= eCaptureFlags.Stereo;\n\t\t\t\tusing iAudioCapture captureDev = mf.openCaptureDevice( devices[ cla.captureDeviceIndex ], cp );\n\n\t\t\t\tusing iModel model = Library.loadModel( cla.model );\n\t\t\t\tusing Context context = model.createContext();\n\t\t\t\tcla.apply( ref context.parameters );\n\n\t\t\t\tCaptureThread thread = new CaptureThread( cla, context, captureDev );\n\t\t\t\tthread.join();\n\n\t\t\t\tcontext.timingsPrint();\n\t\t\t\treturn 0;\n\t\t\t}\n\t\t\tcatch( Exception ex )\n\t\t\t{\n\t\t\t\t// Console.WriteLine( ex.Message );\n\t\t\t\tConsole.WriteLine( ex.ToString() );\n\t\t\t\treturn ex.HResult;\n\t\t\t}\n\t\t}\n\t}\n}"
},
{
"path": "Examples/MicrophoneCS/MicrophoneCS.csproj",
"content": "\n\n\t\n\t\tExe\n\t\tnet6.0-windows\n\t\tenable\n\t\tenable\n\t\ttrue\n\t\tfalse\n\t\tx64\n\t\n\n\t\n\t \n\t\n\n\t\n\t\t\n\t\t\tPreserveNewest\n\t\t\n\t\n\n\t\n\t\t\n\t\n\n"
},
{
"path": "Examples/MicrophoneCS/Readme.txt",
"content": "This example builds .NET 6 console application which shows how to use audio capture API of the .NET wrapper."
},
{
"path": "Examples/MicrophoneCS/TranscribeCallbacks.cs",
"content": "using System.Globalization;\nusing Whisper;\n\nnamespace MicrophoneCS\n{\n\t/// Implementation of Callbacks abstract class, to print these segments as soon as they’re produced by the library.\n\tsealed class TranscribeCallbacks: Callbacks\n\t{\n\t\treadonly CommandLineArgs args;\n\t\treadonly eResultFlags resultFlags;\n\n\t\tpublic TranscribeCallbacks( CommandLineArgs args )\n\t\t{\n\t\t\tthis.args = args;\n\t\t\tresultFlags = args.resultFlags();\n\t\t\tConsole.OutputEncoding = System.Text.Encoding.UTF8;\n\t\t}\n\n\t\t// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]\n\t\t// Lowest is red, middle is yellow, highest is green.\n\t\treadonly string[] k_colors = new string[]\n\t\t{\n\t\t\t\"\\x1B[38;5;196m\", \"\\x1B[38;5;202m\", \"\\x1B[38;5;208m\", \"\\x1B[38;5;214m\", \"\\x1B[38;5;220m\",\n\t\t\t\"\\x1B[38;5;226m\", \"\\x1B[38;5;190m\", \"\\x1B[38;5;154m\", \"\\x1B[38;5;118m\", \"\\x1B[38;5;82m\"\n\t\t};\n\n\t\tint colorIndex( in sToken tok )\n\t\t{\n\t\t\tfloat p = tok.probability;\n\t\t\tfloat p3 = p * p * p;\n\t\t\tint col = (int)( p3 * k_colors.Length );\n\t\t\tcol = Math.Clamp( col, 0, k_colors.Length - 1 );\n\t\t\treturn col;\n\t\t}\n\n\t\tpublic static string printTime( TimeSpan ts ) =>\n\t\t\tts.ToString( \"hh':'mm':'ss'.'fff\", CultureInfo.InvariantCulture );\n\t\tpublic static string printTimeWithComma( TimeSpan ts ) =>\n\t\t\tts.ToString( \"hh':'mm':'ss','fff\", CultureInfo.InvariantCulture );\n\n\t\tprotected override void onNewSegment( Context sender, int countNew )\n\t\t{\n\t\t\tTranscribeResult res = sender.results( resultFlags );\n\t\t\tReadOnlySpan tokens = res.tokens;\n\n\t\t\tint s0 = res.segments.Length - countNew;\n\t\t\tif( s0 == 0 )\n\t\t\t\tConsole.WriteLine();\n\n\t\t\tfor( int i = s0; i < res.segments.Length; i++ )\n\t\t\t{\n\t\t\t\tsSegment seg = res.segments[ i ];\n\n\t\t\t\tif( args.no_timestamps )\n\t\t\t\t{\n\t\t\t\t\tif( args.print_colors && AnsiCodes.enabled )\n\t\t\t\t\t{\n\t\t\t\t\t\tforeach( sToken tok in res.getTokens( seg ) )\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\tif( !args.print_special && tok.hasFlag( eTokenFlags.Special ) )\n\t\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t\tConsole.Write( \"{0}{1}{2}\", k_colors[ colorIndex( tok ) ], tok.text, \"\\x1B[0m\" );\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\telse\n\t\t\t\t\t\tConsole.Write( seg.text );\n\t\t\t\t\tConsole.Out.Flush();\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\n\t\t\t\tstring speaker = \"\";\n\t\t\t\tif( args.diarize )\n\t\t\t\t{\n\t\t\t\t\tspeaker = sender.detectSpeaker( seg.time ) switch\n\t\t\t\t\t{\n\t\t\t\t\t\teSpeakerChannel.Unsure => \"(speaker ?)\",\n\t\t\t\t\t\teSpeakerChannel.Left => \"(speaker 0)\",\n\t\t\t\t\t\teSpeakerChannel.Right => \"(speaker 1)\",\n\t\t\t\t\t\t_ => \"\"\n\t\t\t\t\t};\n\t\t\t\t}\n\n\t\t\t\tif( args.print_colors && AnsiCodes.enabled )\n\t\t\t\t{\n\t\t\t\t\tConsole.Write( \"[{0} --> {1}] {2} \", printTime( seg.time.begin ), printTime( seg.time.end ), speaker );\n\t\t\t\t\tforeach( sToken tok in res.getTokens( seg ) )\n\t\t\t\t\t{\n\t\t\t\t\t\tif( !args.print_special && tok.hasFlag( eTokenFlags.Special ) )\n\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\tConsole.Write( \"{0}{1}{2}\", k_colors[ colorIndex( tok ) ], tok.text, \"\\x1B[0m\" );\n\t\t\t\t\t}\n\t\t\t\t\tConsole.WriteLine();\n\t\t\t\t}\n\t\t\t\telse\n\t\t\t\t\tConsole.WriteLine( \"[{0} --> {1}] {2} {3}\", printTime( seg.time.begin ), printTime( seg.time.end ), speaker, seg.text );\n\t\t\t}\n\t\t}\n\t}\n}"
},
{
"path": "Examples/OldMain/OldMain.vcxproj",
"content": "\n\n \n \n Debug\n x64\n \n \n Release\n x64\n \n \n \n 16.0\n Win32Proj\n {596f9770-9aeb-49d3-86ca-4200197df12b}\n OldMain\n 10.0\n \n \n \n Application\n true\n v143\n Unicode\n \n \n Application\n false\n v143\n true\n Unicode\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n $(ProjectDir);$(SolutionDir)Whisper\\Source\\;$(IncludePath)\n \n \n $(ProjectDir);$(SolutionDir)Whisper\\Source\\;$(IncludePath)\n \n \n \n Level3\n true\n _DEBUG;_CONSOLE;%(PreprocessorDefinitions)\n true\n AdvancedVectorExtensions2\n stdcpp20\n \n \n Console\n true\n \n \n \n \n Level3\n true\n true\n true\n NDEBUG;_CONSOLE;%(PreprocessorDefinitions)\n true\n AdvancedVectorExtensions2\n stdcpp20\n \n \n Console\n true\n true\n true\n UseLinkTimeCodeGeneration\n \n \n \n \n \n true\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n"
},
{
"path": "Examples/OldMain/OldMain.vcxproj.filters",
"content": "\n\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n"
},
{
"path": "Examples/OldMain/Readme.txt",
"content": "This project builds the original whisper.cpp command-line sample"
},
{
"path": "Examples/OldMain/Utils/Logger.cpp",
"content": "#include \n#include \n#include \n#include \"Logger.h\"\n\nnamespace\n{\n\tvoid logMessage( const char* lvl, const char8_t* pszFormat, std::va_list va )\n\t{\n\t\tfprintf( stderr, \"%s: \", lvl );\n\t\tvfprintf( stderr, (const char*)pszFormat, va );\n\t\tfprintf( stderr, \"\\n\" );\n\t}\n}\n\n#define LOG_MESSAGE_IMPL( lvl ) \\\n\tstd::va_list args; \\\n\tva_start( args, pszFormat ); \\\n\tlogMessage( lvl, pszFormat, args ); \\\n\tva_end( args );\n\nvoid logError( const char8_t* pszFormat, ... )\n{\n\tLOG_MESSAGE_IMPL( \"Error\" );\n}\n\nvoid logWarning( const char8_t* pszFormat, ... )\n{\n\tLOG_MESSAGE_IMPL( \"Warning\" );\n}\n\nvoid logInfo( const char8_t* pszFormat, ... )\n{\n\tLOG_MESSAGE_IMPL( \"Info\" );\n}\n\nvoid logDebug( const char8_t* pszFormat, ... )\n{\n\tLOG_MESSAGE_IMPL( \"Debug\" );\n}"
},
{
"path": "Examples/OldMain/Utils/Logger.h",
"content": "#pragma once\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n\nstruct ggml_tensor;\n\nvoid logError( const char8_t* pszFormat, ... );\nvoid logWarning( const char8_t* pszFormat, ... );\nvoid logInfo( const char8_t* pszFormat, ... );\nvoid logDebug( const char8_t* pszFormat, ... );\n\n#ifdef __cplusplus\n}\n\nnamespace Tracing\n{\n\tstruct ItemName\n\t{\n\t\tItemName( const char* str ) { }\n\t\tItemName( const char* str, uint32_t a0 ) { }\n\t\tItemName( const char* str, int a0 ) { }\n\t};\n\n\tinline void tensor( const ItemName& name, const ggml_tensor* tensor ) { }\n\tinline void delayTensor( const ItemName& name, const ggml_tensor* tensor ) { }\n\tinline void vector( const ItemName& name, const std::vector& vec ) { }\n\tinline void writeDelayedTensors() { }\n}\n#endif"
},
{
"path": "Examples/OldMain/dr_wav.h",
"content": "/*\nWAV audio loader and writer. Choice of public domain or MIT-0. See license statements at the end of this file.\ndr_wav - v0.12.16 - 2020-12-02\n\nDavid Reid - mackron@gmail.com\n\nGitHub: https://github.com/mackron/dr_libs\n*/\n\n/*\nRELEASE NOTES - VERSION 0.12\n============================\nVersion 0.12 includes breaking changes to custom chunk handling.\n\n\nChanges to Chunk Callback\n-------------------------\ndr_wav supports the ability to fire a callback when a chunk is encounted (except for WAVE and FMT chunks). The callback has been updated to include both the\ncontainer (RIFF or Wave64) and the FMT chunk which contains information about the format of the data in the wave file.\n\nPreviously, there was no direct way to determine the container, and therefore no way to discriminate against the different IDs in the chunk header (RIFF and\nWave64 containers encode chunk ID's differently). The `container` parameter can be used to know which ID to use.\n\nSometimes it can be useful to know the data format at the time the chunk callback is fired. A pointer to a `drwav_fmt` object is now passed into the chunk\ncallback which will give you information about the data format. To determine the sample format, use `drwav_fmt_get_format()`. This will return one of the\n`DR_WAVE_FORMAT_*` tokens.\n*/\n\n/*\nIntroduction\n============\nThis is a single file library. To use it, do something like the following in one .c file.\n \n ```c\n #define DR_WAV_IMPLEMENTATION\n #include \"dr_wav.h\"\n ```\n\nYou can then #include this file in other parts of the program as you would with any other header file. Do something like the following to read audio data:\n\n ```c\n drwav wav;\n if (!drwav_init_file(&wav, \"my_song.wav\", NULL)) {\n // Error opening WAV file.\n }\n\n drwav_int32* pDecodedInterleavedPCMFrames = malloc(wav.totalPCMFrameCount * wav.channels * sizeof(drwav_int32));\n size_t numberOfSamplesActuallyDecoded = drwav_read_pcm_frames_s32(&wav, wav.totalPCMFrameCount, pDecodedInterleavedPCMFrames);\n\n ...\n\n drwav_uninit(&wav);\n ```\n\nIf you just want to quickly open and read the audio data in a single operation you can do something like this:\n\n ```c\n unsigned int channels;\n unsigned int sampleRate;\n drwav_uint64 totalPCMFrameCount;\n float* pSampleData = drwav_open_file_and_read_pcm_frames_f32(\"my_song.wav\", &channels, &sampleRate, &totalPCMFrameCount, NULL);\n if (pSampleData == NULL) {\n // Error opening and reading WAV file.\n }\n\n ...\n\n drwav_free(pSampleData);\n ```\n\nThe examples above use versions of the API that convert the audio data to a consistent format (32-bit signed PCM, in this case), but you can still output the\naudio data in its internal format (see notes below for supported formats):\n\n ```c\n size_t framesRead = drwav_read_pcm_frames(&wav, wav.totalPCMFrameCount, pDecodedInterleavedPCMFrames);\n ```\n\nYou can also read the raw bytes of audio data, which could be useful if dr_wav does not have native support for a particular data format:\n\n ```c\n size_t bytesRead = drwav_read_raw(&wav, bytesToRead, pRawDataBuffer);\n ```\n\ndr_wav can also be used to output WAV files. This does not currently support compressed formats. To use this, look at `drwav_init_write()`,\n`drwav_init_file_write()`, etc. Use `drwav_write_pcm_frames()` to write samples, or `drwav_write_raw()` to write raw data in the \"data\" chunk.\n\n ```c\n drwav_data_format format;\n format.container = drwav_container_riff; // <-- drwav_container_riff = normal WAV files, drwav_container_w64 = Sony Wave64.\n format.format = DR_WAVE_FORMAT_PCM; // <-- Any of the DR_WAVE_FORMAT_* codes.\n format.channels = 2;\n format.sampleRate = 44100;\n format.bitsPerSample = 16;\n drwav_init_file_write(&wav, \"data/recording.wav\", &format, NULL);\n\n ...\n\n drwav_uint64 framesWritten = drwav_write_pcm_frames(pWav, frameCount, pSamples);\n ```\n\ndr_wav has seamless support the Sony Wave64 format. The decoder will automatically detect it and it should Just Work without any manual intervention.\n\n\nBuild Options\n=============\n#define these options before including this file.\n\n#define DR_WAV_NO_CONVERSION_API\n Disables conversion APIs such as `drwav_read_pcm_frames_f32()` and `drwav_s16_to_f32()`.\n\n#define DR_WAV_NO_STDIO\n Disables APIs that initialize a decoder from a file such as `drwav_init_file()`, `drwav_init_file_write()`, etc.\n\n\n\nNotes\n=====\n- Samples are always interleaved.\n- The default read function does not do any data conversion. Use `drwav_read_pcm_frames_f32()`, `drwav_read_pcm_frames_s32()` and `drwav_read_pcm_frames_s16()`\n to read and convert audio data to 32-bit floating point, signed 32-bit integer and signed 16-bit integer samples respectively. Tested and supported internal\n formats include the following:\n - Unsigned 8-bit PCM\n - Signed 12-bit PCM\n - Signed 16-bit PCM\n - Signed 24-bit PCM\n - Signed 32-bit PCM\n - IEEE 32-bit floating point\n - IEEE 64-bit floating point\n - A-law and u-law\n - Microsoft ADPCM\n - IMA ADPCM (DVI, format code 0x11)\n- dr_wav will try to read the WAV file as best it can, even if it's not strictly conformant to the WAV format.\n*/\n\n#ifndef dr_wav_h\n#define dr_wav_h\n\n#ifdef __cplusplus\nextern \"C\" {\n#endif\n\n#define DRWAV_STRINGIFY(x) #x\n#define DRWAV_XSTRINGIFY(x) DRWAV_STRINGIFY(x)\n\n#define DRWAV_VERSION_MAJOR 0\n#define DRWAV_VERSION_MINOR 12\n#define DRWAV_VERSION_REVISION 16\n#define DRWAV_VERSION_STRING DRWAV_XSTRINGIFY(DRWAV_VERSION_MAJOR) \".\" DRWAV_XSTRINGIFY(DRWAV_VERSION_MINOR) \".\" DRWAV_XSTRINGIFY(DRWAV_VERSION_REVISION)\n\n#include /* For size_t. */\n\n/* Sized types. */\ntypedef signed char drwav_int8;\ntypedef unsigned char drwav_uint8;\ntypedef signed short drwav_int16;\ntypedef unsigned short drwav_uint16;\ntypedef signed int drwav_int32;\ntypedef unsigned int drwav_uint32;\n#if defined(_MSC_VER)\n typedef signed __int64 drwav_int64;\n typedef unsigned __int64 drwav_uint64;\n#else\n #if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))\n #pragma GCC diagnostic push\n #pragma GCC diagnostic ignored \"-Wlong-long\"\n #if defined(__clang__)\n #pragma GCC diagnostic ignored \"-Wc++11-long-long\"\n #endif\n #endif\n typedef signed long long drwav_int64;\n typedef unsigned long long drwav_uint64;\n #if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))\n #pragma GCC diagnostic pop\n #endif\n#endif\n#if defined(__LP64__) || defined(_WIN64) || (defined(__x86_64__) && !defined(__ILP32__)) || defined(_M_X64) || defined(__ia64) || defined (_M_IA64) || defined(__aarch64__) || defined(__powerpc64__)\n typedef drwav_uint64 drwav_uintptr;\n#else\n typedef drwav_uint32 drwav_uintptr;\n#endif\ntypedef drwav_uint8 drwav_bool8;\ntypedef drwav_uint32 drwav_bool32;\n#define DRWAV_TRUE 1\n#define DRWAV_FALSE 0\n\n#if !defined(DRWAV_API)\n #if defined(DRWAV_DLL)\n #if defined(_WIN32)\n #define DRWAV_DLL_IMPORT __declspec(dllimport)\n #define DRWAV_DLL_EXPORT __declspec(dllexport)\n #define DRWAV_DLL_PRIVATE static\n #else\n #if defined(__GNUC__) && __GNUC__ >= 4\n #define DRWAV_DLL_IMPORT __attribute__((visibility(\"default\")))\n #define DRWAV_DLL_EXPORT __attribute__((visibility(\"default\")))\n #define DRWAV_DLL_PRIVATE __attribute__((visibility(\"hidden\")))\n #else\n #define DRWAV_DLL_IMPORT\n #define DRWAV_DLL_EXPORT\n #define DRWAV_DLL_PRIVATE static\n #endif\n #endif\n\n #if defined(DR_WAV_IMPLEMENTATION) || defined(DRWAV_IMPLEMENTATION)\n #define DRWAV_API DRWAV_DLL_EXPORT\n #else\n #define DRWAV_API DRWAV_DLL_IMPORT\n #endif\n #define DRWAV_PRIVATE DRWAV_DLL_PRIVATE\n #else\n #define DRWAV_API extern\n #define DRWAV_PRIVATE static\n #endif\n#endif\n\ntypedef drwav_int32 drwav_result;\n#define DRWAV_SUCCESS 0\n#define DRWAV_ERROR -1 /* A generic error. */\n#define DRWAV_INVALID_ARGS -2\n#define DRWAV_INVALID_OPERATION -3\n#define DRWAV_OUT_OF_MEMORY -4\n#define DRWAV_OUT_OF_RANGE -5\n#define DRWAV_ACCESS_DENIED -6\n#define DRWAV_DOES_NOT_EXIST -7\n#define DRWAV_ALREADY_EXISTS -8\n#define DRWAV_TOO_MANY_OPEN_FILES -9\n#define DRWAV_INVALID_FILE -10\n#define DRWAV_TOO_BIG -11\n#define DRWAV_PATH_TOO_LONG -12\n#define DRWAV_NAME_TOO_LONG -13\n#define DRWAV_NOT_DIRECTORY -14\n#define DRWAV_IS_DIRECTORY -15\n#define DRWAV_DIRECTORY_NOT_EMPTY -16\n#define DRWAV_END_OF_FILE -17\n#define DRWAV_NO_SPACE -18\n#define DRWAV_BUSY -19\n#define DRWAV_IO_ERROR -20\n#define DRWAV_INTERRUPT -21\n#define DRWAV_UNAVAILABLE -22\n#define DRWAV_ALREADY_IN_USE -23\n#define DRWAV_BAD_ADDRESS -24\n#define DRWAV_BAD_SEEK -25\n#define DRWAV_BAD_PIPE -26\n#define DRWAV_DEADLOCK -27\n#define DRWAV_TOO_MANY_LINKS -28\n#define DRWAV_NOT_IMPLEMENTED -29\n#define DRWAV_NO_MESSAGE -30\n#define DRWAV_BAD_MESSAGE -31\n#define DRWAV_NO_DATA_AVAILABLE -32\n#define DRWAV_INVALID_DATA -33\n#define DRWAV_TIMEOUT -34\n#define DRWAV_NO_NETWORK -35\n#define DRWAV_NOT_UNIQUE -36\n#define DRWAV_NOT_SOCKET -37\n#define DRWAV_NO_ADDRESS -38\n#define DRWAV_BAD_PROTOCOL -39\n#define DRWAV_PROTOCOL_UNAVAILABLE -40\n#define DRWAV_PROTOCOL_NOT_SUPPORTED -41\n#define DRWAV_PROTOCOL_FAMILY_NOT_SUPPORTED -42\n#define DRWAV_ADDRESS_FAMILY_NOT_SUPPORTED -43\n#define DRWAV_SOCKET_NOT_SUPPORTED -44\n#define DRWAV_CONNECTION_RESET -45\n#define DRWAV_ALREADY_CONNECTED -46\n#define DRWAV_NOT_CONNECTED -47\n#define DRWAV_CONNECTION_REFUSED -48\n#define DRWAV_NO_HOST -49\n#define DRWAV_IN_PROGRESS -50\n#define DRWAV_CANCELLED -51\n#define DRWAV_MEMORY_ALREADY_MAPPED -52\n#define DRWAV_AT_END -53\n\n/* Common data formats. */\n#define DR_WAVE_FORMAT_PCM 0x1\n#define DR_WAVE_FORMAT_ADPCM 0x2\n#define DR_WAVE_FORMAT_IEEE_FLOAT 0x3\n#define DR_WAVE_FORMAT_ALAW 0x6\n#define DR_WAVE_FORMAT_MULAW 0x7\n#define DR_WAVE_FORMAT_DVI_ADPCM 0x11\n#define DR_WAVE_FORMAT_EXTENSIBLE 0xFFFE\n\n/* Constants. */\n#ifndef DRWAV_MAX_SMPL_LOOPS\n#define DRWAV_MAX_SMPL_LOOPS 1\n#endif\n\n/* Flags to pass into drwav_init_ex(), etc. */\n#define DRWAV_SEQUENTIAL 0x00000001\n\nDRWAV_API void drwav_version(drwav_uint32* pMajor, drwav_uint32* pMinor, drwav_uint32* pRevision);\nDRWAV_API const char* drwav_version_string(void);\n\ntypedef enum\n{\n drwav_seek_origin_start,\n drwav_seek_origin_current\n} drwav_seek_origin;\n\ntypedef enum\n{\n drwav_container_riff,\n drwav_container_w64,\n drwav_container_rf64\n} drwav_container;\n\ntypedef struct\n{\n union\n {\n drwav_uint8 fourcc[4];\n drwav_uint8 guid[16];\n } id;\n\n /* The size in bytes of the chunk. */\n drwav_uint64 sizeInBytes;\n\n /*\n RIFF = 2 byte alignment.\n W64 = 8 byte alignment.\n */\n unsigned int paddingSize;\n} drwav_chunk_header;\n\ntypedef struct\n{\n /*\n The format tag exactly as specified in the wave file's \"fmt\" chunk. This can be used by applications\n that require support for data formats not natively supported by dr_wav.\n */\n drwav_uint16 formatTag;\n\n /* The number of channels making up the audio data. When this is set to 1 it is mono, 2 is stereo, etc. */\n drwav_uint16 channels;\n\n /* The sample rate. Usually set to something like 44100. */\n drwav_uint32 sampleRate;\n\n /* Average bytes per second. You probably don't need this, but it's left here for informational purposes. */\n drwav_uint32 avgBytesPerSec;\n\n /* Block align. This is equal to the number of channels * bytes per sample. */\n drwav_uint16 blockAlign;\n\n /* Bits per sample. */\n drwav_uint16 bitsPerSample;\n\n /* The size of the extended data. Only used internally for validation, but left here for informational purposes. */\n drwav_uint16 extendedSize;\n\n /*\n The number of valid bits per sample. When is equal to WAVE_FORMAT_EXTENSIBLE, \n is always rounded up to the nearest multiple of 8. This variable contains information about exactly how\n many bits are valid per sample. Mainly used for informational purposes.\n */\n drwav_uint16 validBitsPerSample;\n\n /* The channel mask. Not used at the moment. */\n drwav_uint32 channelMask;\n\n /* The sub-format, exactly as specified by the wave file. */\n drwav_uint8 subFormat[16];\n} drwav_fmt;\n\nDRWAV_API drwav_uint16 drwav_fmt_get_format(const drwav_fmt* pFMT);\n\n\n/*\nCallback for when data is read. Return value is the number of bytes actually read.\n\npUserData [in] The user data that was passed to drwav_init() and family.\npBufferOut [out] The output buffer.\nbytesToRead [in] The number of bytes to read.\n\nReturns the number of bytes actually read.\n\nA return value of less than bytesToRead indicates the end of the stream. Do _not_ return from this callback until\neither the entire bytesToRead is filled or you have reached the end of the stream.\n*/\ntypedef size_t (* drwav_read_proc)(void* pUserData, void* pBufferOut, size_t bytesToRead);\n\n/*\nCallback for when data is written. Returns value is the number of bytes actually written.\n\npUserData [in] The user data that was passed to drwav_init_write() and family.\npData [out] A pointer to the data to write.\nbytesToWrite [in] The number of bytes to write.\n\nReturns the number of bytes actually written.\n\nIf the return value differs from bytesToWrite, it indicates an error.\n*/\ntypedef size_t (* drwav_write_proc)(void* pUserData, const void* pData, size_t bytesToWrite);\n\n/*\nCallback for when data needs to be seeked.\n\npUserData [in] The user data that was passed to drwav_init() and family.\noffset [in] The number of bytes to move, relative to the origin. Will never be negative.\norigin [in] The origin of the seek - the current position or the start of the stream.\n\nReturns whether or not the seek was successful.\n\nWhether or not it is relative to the beginning or current position is determined by the \"origin\" parameter which will be either drwav_seek_origin_start or\ndrwav_seek_origin_current.\n*/\ntypedef drwav_bool32 (* drwav_seek_proc)(void* pUserData, int offset, drwav_seek_origin origin);\n\n/*\nCallback for when drwav_init_ex() finds a chunk.\n\npChunkUserData [in] The user data that was passed to the pChunkUserData parameter of drwav_init_ex() and family.\nonRead [in] A pointer to the function to call when reading.\nonSeek [in] A pointer to the function to call when seeking.\npReadSeekUserData [in] The user data that was passed to the pReadSeekUserData parameter of drwav_init_ex() and family.\npChunkHeader [in] A pointer to an object containing basic header information about the chunk. Use this to identify the chunk.\ncontainer [in] Whether or not the WAV file is a RIFF or Wave64 container. If you're unsure of the difference, assume RIFF.\npFMT [in] A pointer to the object containing the contents of the \"fmt\" chunk.\n\nReturns the number of bytes read + seeked.\n\nTo read data from the chunk, call onRead(), passing in pReadSeekUserData as the first parameter. Do the same for seeking with onSeek(). The return value must\nbe the total number of bytes you have read _plus_ seeked.\n\nUse the `container` argument to discriminate the fields in `pChunkHeader->id`. If the container is `drwav_container_riff` or `drwav_container_rf64` you should\nuse `id.fourcc`, otherwise you should use `id.guid`.\n\nThe `pFMT` parameter can be used to determine the data format of the wave file. Use `drwav_fmt_get_format()` to get the sample format, which will be one of the\n`DR_WAVE_FORMAT_*` identifiers. \n\nThe read pointer will be sitting on the first byte after the chunk's header. You must not attempt to read beyond the boundary of the chunk.\n*/\ntypedef drwav_uint64 (* drwav_chunk_proc)(void* pChunkUserData, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pReadSeekUserData, const drwav_chunk_header* pChunkHeader, drwav_container container, const drwav_fmt* pFMT);\n\ntypedef struct\n{\n void* pUserData;\n void* (* onMalloc)(size_t sz, void* pUserData);\n void* (* onRealloc)(void* p, size_t sz, void* pUserData);\n void (* onFree)(void* p, void* pUserData);\n} drwav_allocation_callbacks;\n\n/* Structure for internal use. Only used for loaders opened with drwav_init_memory(). */\ntypedef struct\n{\n const drwav_uint8* data;\n size_t dataSize;\n size_t currentReadPos;\n} drwav__memory_stream;\n\n/* Structure for internal use. Only used for writers opened with drwav_init_memory_write(). */\ntypedef struct\n{\n void** ppData;\n size_t* pDataSize;\n size_t dataSize;\n size_t dataCapacity;\n size_t currentWritePos;\n} drwav__memory_stream_write;\n\ntypedef struct\n{\n drwav_container container; /* RIFF, W64. */\n drwav_uint32 format; /* DR_WAVE_FORMAT_* */\n drwav_uint32 channels;\n drwav_uint32 sampleRate;\n drwav_uint32 bitsPerSample;\n} drwav_data_format;\n\n\n/* See the following for details on the 'smpl' chunk: https://sites.google.com/site/musicgapi/technical-documents/wav-file-format#smpl */\ntypedef struct\n{\n drwav_uint32 cuePointId;\n drwav_uint32 type;\n drwav_uint32 start;\n drwav_uint32 end;\n drwav_uint32 fraction;\n drwav_uint32 playCount;\n} drwav_smpl_loop;\n\n typedef struct\n{\n drwav_uint32 manufacturer;\n drwav_uint32 product;\n drwav_uint32 samplePeriod;\n drwav_uint32 midiUnityNotes;\n drwav_uint32 midiPitchFraction;\n drwav_uint32 smpteFormat;\n drwav_uint32 smpteOffset;\n drwav_uint32 numSampleLoops;\n drwav_uint32 samplerData;\n drwav_smpl_loop loops[DRWAV_MAX_SMPL_LOOPS];\n} drwav_smpl;\n\ntypedef struct\n{\n /* A pointer to the function to call when more data is needed. */\n drwav_read_proc onRead;\n\n /* A pointer to the function to call when data needs to be written. Only used when the drwav object is opened in write mode. */\n drwav_write_proc onWrite;\n\n /* A pointer to the function to call when the wav file needs to be seeked. */\n drwav_seek_proc onSeek;\n\n /* The user data to pass to callbacks. */\n void* pUserData;\n\n /* Allocation callbacks. */\n drwav_allocation_callbacks allocationCallbacks;\n\n\n /* Whether or not the WAV file is formatted as a standard RIFF file or W64. */\n drwav_container container;\n\n\n /* Structure containing format information exactly as specified by the wav file. */\n drwav_fmt fmt;\n\n /* The sample rate. Will be set to something like 44100. */\n drwav_uint32 sampleRate;\n\n /* The number of channels. This will be set to 1 for monaural streams, 2 for stereo, etc. */\n drwav_uint16 channels;\n\n /* The bits per sample. Will be set to something like 16, 24, etc. */\n drwav_uint16 bitsPerSample;\n\n /* Equal to fmt.formatTag, or the value specified by fmt.subFormat if fmt.formatTag is equal to 65534 (WAVE_FORMAT_EXTENSIBLE). */\n drwav_uint16 translatedFormatTag;\n\n /* The total number of PCM frames making up the audio data. */\n drwav_uint64 totalPCMFrameCount;\n\n\n /* The size in bytes of the data chunk. */\n drwav_uint64 dataChunkDataSize;\n \n /* The position in the stream of the first byte of the data chunk. This is used for seeking. */\n drwav_uint64 dataChunkDataPos;\n\n /* The number of bytes remaining in the data chunk. */\n drwav_uint64 bytesRemaining;\n\n\n /*\n Only used in sequential write mode. Keeps track of the desired size of the \"data\" chunk at the point of initialization time. Always\n set to 0 for non-sequential writes and when the drwav object is opened in read mode. Used for validation.\n */\n drwav_uint64 dataChunkDataSizeTargetWrite;\n\n /* Keeps track of whether or not the wav writer was initialized in sequential mode. */\n drwav_bool32 isSequentialWrite;\n\n\n /* smpl chunk. */\n drwav_smpl smpl;\n\n\n /* A hack to avoid a DRWAV_MALLOC() when opening a decoder with drwav_init_memory(). */\n drwav__memory_stream memoryStream;\n drwav__memory_stream_write memoryStreamWrite;\n\n /* Generic data for compressed formats. This data is shared across all block-compressed formats. */\n struct\n {\n drwav_uint64 iCurrentPCMFrame; /* The index of the next PCM frame that will be read by drwav_read_*(). This is used with \"totalPCMFrameCount\" to ensure we don't read excess samples at the end of the last block. */\n } compressed;\n \n /* Microsoft ADPCM specific data. */\n struct\n {\n drwav_uint32 bytesRemainingInBlock;\n drwav_uint16 predictor[2];\n drwav_int32 delta[2];\n drwav_int32 cachedFrames[4]; /* Samples are stored in this cache during decoding. */\n drwav_uint32 cachedFrameCount;\n drwav_int32 prevFrames[2][2]; /* The previous 2 samples for each channel (2 channels at most). */\n } msadpcm;\n\n /* IMA ADPCM specific data. */\n struct\n {\n drwav_uint32 bytesRemainingInBlock;\n drwav_int32 predictor[2];\n drwav_int32 stepIndex[2];\n drwav_int32 cachedFrames[16]; /* Samples are stored in this cache during decoding. */\n drwav_uint32 cachedFrameCount;\n } ima;\n} drwav;\n\n\n/*\nInitializes a pre-allocated drwav object for reading.\n\npWav [out] A pointer to the drwav object being initialized.\nonRead [in] The function to call when data needs to be read from the client.\nonSeek [in] The function to call when the read position of the client data needs to move.\nonChunk [in, optional] The function to call when a chunk is enumerated at initialized time.\npUserData, pReadSeekUserData [in, optional] A pointer to application defined data that will be passed to onRead and onSeek.\npChunkUserData [in, optional] A pointer to application defined data that will be passed to onChunk.\nflags [in, optional] A set of flags for controlling how things are loaded.\n\nReturns true if successful; false otherwise.\n\nClose the loader with drwav_uninit().\n\nThis is the lowest level function for initializing a WAV file. You can also use drwav_init_file() and drwav_init_memory()\nto open the stream from a file or from a block of memory respectively.\n\nPossible values for flags:\n DRWAV_SEQUENTIAL: Never perform a backwards seek while loading. This disables the chunk callback and will cause this function\n to return as soon as the data chunk is found. Any chunks after the data chunk will be ignored.\n\ndrwav_init() is equivalent to \"drwav_init_ex(pWav, onRead, onSeek, NULL, pUserData, NULL, 0);\".\n\nThe onChunk callback is not called for the WAVE or FMT chunks. The contents of the FMT chunk can be read from pWav->fmt\nafter the function returns.\n\nSee also: drwav_init_file(), drwav_init_memory(), drwav_uninit()\n*/\nDRWAV_API drwav_bool32 drwav_init(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_ex(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, drwav_chunk_proc onChunk, void* pReadSeekUserData, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);\n\n/*\nInitializes a pre-allocated drwav object for writing.\n\nonWrite [in] The function to call when data needs to be written.\nonSeek [in] The function to call when the write position needs to move.\npUserData [in, optional] A pointer to application defined data that will be passed to onWrite and onSeek.\n\nReturns true if successful; false otherwise.\n\nClose the writer with drwav_uninit().\n\nThis is the lowest level function for initializing a WAV file. You can also use drwav_init_file_write() and drwav_init_memory_write()\nto open the stream from a file or from a block of memory respectively.\n\nIf the total sample count is known, you can use drwav_init_write_sequential(). This avoids the need for dr_wav to perform\na post-processing step for storing the total sample count and the size of the data chunk which requires a backwards seek.\n\nSee also: drwav_init_file_write(), drwav_init_memory_write(), drwav_uninit()\n*/\nDRWAV_API drwav_bool32 drwav_init_write(drwav* pWav, const drwav_data_format* pFormat, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_write_sequential(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_write_sequential_pcm_frames(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);\n\n/*\nUtility function to determine the target size of the entire data to be written (including all headers and chunks).\n\nReturns the target size in bytes.\n\nUseful if the application needs to know the size to allocate.\n\nOnly writing to the RIFF chunk and one data chunk is currently supported.\n\nSee also: drwav_init_write(), drwav_init_file_write(), drwav_init_memory_write()\n*/\nDRWAV_API drwav_uint64 drwav_target_write_size_bytes(const drwav_data_format* pFormat, drwav_uint64 totalSampleCount);\n\n/*\nUninitializes the given drwav object.\n\nUse this only for objects initialized with drwav_init*() functions (drwav_init(), drwav_init_ex(), drwav_init_write(), drwav_init_write_sequential()).\n*/\nDRWAV_API drwav_result drwav_uninit(drwav* pWav);\n\n\n/*\nReads raw audio data.\n\nThis is the lowest level function for reading audio data. It simply reads the given number of\nbytes of the raw internal sample data.\n\nConsider using drwav_read_pcm_frames_s16(), drwav_read_pcm_frames_s32() or drwav_read_pcm_frames_f32() for\nreading sample data in a consistent format.\n\npBufferOut can be NULL in which case a seek will be performed.\n\nReturns the number of bytes actually read.\n*/\nDRWAV_API size_t drwav_read_raw(drwav* pWav, size_t bytesToRead, void* pBufferOut);\n\n/*\nReads up to the specified number of PCM frames from the WAV file.\n\nThe output data will be in the file's internal format, converted to native-endian byte order. Use\ndrwav_read_pcm_frames_s16/f32/s32() to read data in a specific format.\n\nIf the return value is less than it means the end of the file has been reached or\nyou have requested more PCM frames than can possibly fit in the output buffer.\n\nThis function will only work when sample data is of a fixed size and uncompressed. If you are\nusing a compressed format consider using drwav_read_raw() or drwav_read_pcm_frames_s16/s32/f32().\n\npBufferOut can be NULL in which case a seek will be performed.\n*/\nDRWAV_API drwav_uint64 drwav_read_pcm_frames(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut);\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_le(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut);\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_be(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut);\n\n/*\nSeeks to the given PCM frame.\n\nReturns true if successful; false otherwise.\n*/\nDRWAV_API drwav_bool32 drwav_seek_to_pcm_frame(drwav* pWav, drwav_uint64 targetFrameIndex);\n\n\n/*\nWrites raw audio data.\n\nReturns the number of bytes actually written. If this differs from bytesToWrite, it indicates an error.\n*/\nDRWAV_API size_t drwav_write_raw(drwav* pWav, size_t bytesToWrite, const void* pData);\n\n/*\nWrites PCM frames.\n\nReturns the number of PCM frames written.\n\nInput samples need to be in native-endian byte order. On big-endian architectures the input data will be converted to\nlittle-endian. Use drwav_write_raw() to write raw audio data without performing any conversion.\n*/\nDRWAV_API drwav_uint64 drwav_write_pcm_frames(drwav* pWav, drwav_uint64 framesToWrite, const void* pData);\nDRWAV_API drwav_uint64 drwav_write_pcm_frames_le(drwav* pWav, drwav_uint64 framesToWrite, const void* pData);\nDRWAV_API drwav_uint64 drwav_write_pcm_frames_be(drwav* pWav, drwav_uint64 framesToWrite, const void* pData);\n\n\n/* Conversion Utilities */\n#ifndef DR_WAV_NO_CONVERSION_API\n\n/*\nReads a chunk of audio data and converts it to signed 16-bit PCM samples.\n\npBufferOut can be NULL in which case a seek will be performed.\n\nReturns the number of PCM frames actually read.\n\nIf the return value is less than it means the end of the file has been reached.\n*/\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s16(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut);\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s16le(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut);\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s16be(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut);\n\n/* Low-level function for converting unsigned 8-bit PCM samples to signed 16-bit PCM samples. */\nDRWAV_API void drwav_u8_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n/* Low-level function for converting signed 24-bit PCM samples to signed 16-bit PCM samples. */\nDRWAV_API void drwav_s24_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n/* Low-level function for converting signed 32-bit PCM samples to signed 16-bit PCM samples. */\nDRWAV_API void drwav_s32_to_s16(drwav_int16* pOut, const drwav_int32* pIn, size_t sampleCount);\n\n/* Low-level function for converting IEEE 32-bit floating point samples to signed 16-bit PCM samples. */\nDRWAV_API void drwav_f32_to_s16(drwav_int16* pOut, const float* pIn, size_t sampleCount);\n\n/* Low-level function for converting IEEE 64-bit floating point samples to signed 16-bit PCM samples. */\nDRWAV_API void drwav_f64_to_s16(drwav_int16* pOut, const double* pIn, size_t sampleCount);\n\n/* Low-level function for converting A-law samples to signed 16-bit PCM samples. */\nDRWAV_API void drwav_alaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n/* Low-level function for converting u-law samples to signed 16-bit PCM samples. */\nDRWAV_API void drwav_mulaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n\n/*\nReads a chunk of audio data and converts it to IEEE 32-bit floating point samples.\n\npBufferOut can be NULL in which case a seek will be performed.\n\nReturns the number of PCM frames actually read.\n\nIf the return value is less than it means the end of the file has been reached.\n*/\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_f32(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut);\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_f32le(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut);\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_f32be(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut);\n\n/* Low-level function for converting unsigned 8-bit PCM samples to IEEE 32-bit floating point samples. */\nDRWAV_API void drwav_u8_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n/* Low-level function for converting signed 16-bit PCM samples to IEEE 32-bit floating point samples. */\nDRWAV_API void drwav_s16_to_f32(float* pOut, const drwav_int16* pIn, size_t sampleCount);\n\n/* Low-level function for converting signed 24-bit PCM samples to IEEE 32-bit floating point samples. */\nDRWAV_API void drwav_s24_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n/* Low-level function for converting signed 32-bit PCM samples to IEEE 32-bit floating point samples. */\nDRWAV_API void drwav_s32_to_f32(float* pOut, const drwav_int32* pIn, size_t sampleCount);\n\n/* Low-level function for converting IEEE 64-bit floating point samples to IEEE 32-bit floating point samples. */\nDRWAV_API void drwav_f64_to_f32(float* pOut, const double* pIn, size_t sampleCount);\n\n/* Low-level function for converting A-law samples to IEEE 32-bit floating point samples. */\nDRWAV_API void drwav_alaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n/* Low-level function for converting u-law samples to IEEE 32-bit floating point samples. */\nDRWAV_API void drwav_mulaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n\n/*\nReads a chunk of audio data and converts it to signed 32-bit PCM samples.\n\npBufferOut can be NULL in which case a seek will be performed.\n\nReturns the number of PCM frames actually read.\n\nIf the return value is less than it means the end of the file has been reached.\n*/\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s32(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut);\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s32le(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut);\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s32be(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut);\n\n/* Low-level function for converting unsigned 8-bit PCM samples to signed 32-bit PCM samples. */\nDRWAV_API void drwav_u8_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n/* Low-level function for converting signed 16-bit PCM samples to signed 32-bit PCM samples. */\nDRWAV_API void drwav_s16_to_s32(drwav_int32* pOut, const drwav_int16* pIn, size_t sampleCount);\n\n/* Low-level function for converting signed 24-bit PCM samples to signed 32-bit PCM samples. */\nDRWAV_API void drwav_s24_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n/* Low-level function for converting IEEE 32-bit floating point samples to signed 32-bit PCM samples. */\nDRWAV_API void drwav_f32_to_s32(drwav_int32* pOut, const float* pIn, size_t sampleCount);\n\n/* Low-level function for converting IEEE 64-bit floating point samples to signed 32-bit PCM samples. */\nDRWAV_API void drwav_f64_to_s32(drwav_int32* pOut, const double* pIn, size_t sampleCount);\n\n/* Low-level function for converting A-law samples to signed 32-bit PCM samples. */\nDRWAV_API void drwav_alaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n/* Low-level function for converting u-law samples to signed 32-bit PCM samples. */\nDRWAV_API void drwav_mulaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);\n\n#endif /* DR_WAV_NO_CONVERSION_API */\n\n\n/* High-Level Convenience Helpers */\n\n#ifndef DR_WAV_NO_STDIO\n/*\nHelper for initializing a wave file for reading using stdio.\n\nThis holds the internal FILE object until drwav_uninit() is called. Keep this in mind if you're caching drwav\nobjects because the operating system may restrict the number of file handles an application can have open at\nany given time.\n*/\nDRWAV_API drwav_bool32 drwav_init_file(drwav* pWav, const char* filename, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_file_ex(drwav* pWav, const char* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_file_w(drwav* pWav, const wchar_t* filename, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_file_ex_w(drwav* pWav, const wchar_t* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);\n\n/*\nHelper for initializing a wave file for writing using stdio.\n\nThis holds the internal FILE object until drwav_uninit() is called. Keep this in mind if you're caching drwav\nobjects because the operating system may restrict the number of file handles an application can have open at\nany given time.\n*/\nDRWAV_API drwav_bool32 drwav_init_file_write(drwav* pWav, const char* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_file_write_sequential(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_file_write_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_file_write_sequential_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks);\n#endif /* DR_WAV_NO_STDIO */\n\n/*\nHelper for initializing a loader from a pre-allocated memory buffer.\n\nThis does not create a copy of the data. It is up to the application to ensure the buffer remains valid for\nthe lifetime of the drwav object.\n\nThe buffer should contain the contents of the entire wave file, not just the sample data.\n*/\nDRWAV_API drwav_bool32 drwav_init_memory(drwav* pWav, const void* data, size_t dataSize, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_memory_ex(drwav* pWav, const void* data, size_t dataSize, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);\n\n/*\nHelper for initializing a writer which outputs data to a memory buffer.\n\ndr_wav will manage the memory allocations, however it is up to the caller to free the data with drwav_free().\n\nThe buffer will remain allocated even after drwav_uninit() is called. The buffer should not be considered valid\nuntil after drwav_uninit() has been called.\n*/\nDRWAV_API drwav_bool32 drwav_init_memory_write(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_memory_write_sequential(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_bool32 drwav_init_memory_write_sequential_pcm_frames(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks);\n\n\n#ifndef DR_WAV_NO_CONVERSION_API\n/*\nOpens and reads an entire wav file in a single operation.\n\nThe return value is a heap-allocated buffer containing the audio data. Use drwav_free() to free the buffer.\n*/\nDRWAV_API drwav_int16* drwav_open_and_read_pcm_frames_s16(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API float* drwav_open_and_read_pcm_frames_f32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_int32* drwav_open_and_read_pcm_frames_s32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\n#ifndef DR_WAV_NO_STDIO\n/*\nOpens and decodes an entire wav file in a single operation.\n\nThe return value is a heap-allocated buffer containing the audio data. Use drwav_free() to free the buffer.\n*/\nDRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API float* drwav_open_file_and_read_pcm_frames_f32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API float* drwav_open_file_and_read_pcm_frames_f32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\n#endif\n/*\nOpens and decodes an entire wav file from a block of memory in a single operation.\n\nThe return value is a heap-allocated buffer containing the audio data. Use drwav_free() to free the buffer.\n*/\nDRWAV_API drwav_int16* drwav_open_memory_and_read_pcm_frames_s16(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API float* drwav_open_memory_and_read_pcm_frames_f32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\nDRWAV_API drwav_int32* drwav_open_memory_and_read_pcm_frames_s32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);\n#endif\n\n/* Frees data that was allocated internally by dr_wav. */\nDRWAV_API void drwav_free(void* p, const drwav_allocation_callbacks* pAllocationCallbacks);\n\n/* Converts bytes from a wav stream to a sized type of native endian. */\nDRWAV_API drwav_uint16 drwav_bytes_to_u16(const drwav_uint8* data);\nDRWAV_API drwav_int16 drwav_bytes_to_s16(const drwav_uint8* data);\nDRWAV_API drwav_uint32 drwav_bytes_to_u32(const drwav_uint8* data);\nDRWAV_API drwav_int32 drwav_bytes_to_s32(const drwav_uint8* data);\nDRWAV_API drwav_uint64 drwav_bytes_to_u64(const drwav_uint8* data);\nDRWAV_API drwav_int64 drwav_bytes_to_s64(const drwav_uint8* data);\n\n/* Compares a GUID for the purpose of checking the type of a Wave64 chunk. */\nDRWAV_API drwav_bool32 drwav_guid_equal(const drwav_uint8 a[16], const drwav_uint8 b[16]);\n\n/* Compares a four-character-code for the purpose of checking the type of a RIFF chunk. */\nDRWAV_API drwav_bool32 drwav_fourcc_equal(const drwav_uint8* a, const char* b);\n\n#ifdef __cplusplus\n}\n#endif\n#endif /* dr_wav_h */\n\n\n/************************************************************************************************************************************************************\n ************************************************************************************************************************************************************\n\n IMPLEMENTATION\n\n ************************************************************************************************************************************************************\n ************************************************************************************************************************************************************/\n#if defined(DR_WAV_IMPLEMENTATION) || defined(DRWAV_IMPLEMENTATION)\n#ifndef dr_wav_c\n#define dr_wav_c\n\n#include \n#include /* For memcpy(), memset() */\n#include /* For INT_MAX */\n\n#ifndef DR_WAV_NO_STDIO\n#include \n#include \n#endif\n\n/* Standard library stuff. */\n#ifndef DRWAV_ASSERT\n#include \n#define DRWAV_ASSERT(expression) assert(expression)\n#endif\n#ifndef DRWAV_MALLOC\n#define DRWAV_MALLOC(sz) malloc((sz))\n#endif\n#ifndef DRWAV_REALLOC\n#define DRWAV_REALLOC(p, sz) realloc((p), (sz))\n#endif\n#ifndef DRWAV_FREE\n#define DRWAV_FREE(p) free((p))\n#endif\n#ifndef DRWAV_COPY_MEMORY\n#define DRWAV_COPY_MEMORY(dst, src, sz) memcpy((dst), (src), (sz))\n#endif\n#ifndef DRWAV_ZERO_MEMORY\n#define DRWAV_ZERO_MEMORY(p, sz) memset((p), 0, (sz))\n#endif\n#ifndef DRWAV_ZERO_OBJECT\n#define DRWAV_ZERO_OBJECT(p) DRWAV_ZERO_MEMORY((p), sizeof(*p))\n#endif\n\n#define drwav_countof(x) (sizeof(x) / sizeof(x[0]))\n#define drwav_align(x, a) ((((x) + (a) - 1) / (a)) * (a))\n#define drwav_min(a, b) (((a) < (b)) ? (a) : (b))\n#define drwav_max(a, b) (((a) > (b)) ? (a) : (b))\n#define drwav_clamp(x, lo, hi) (drwav_max((lo), drwav_min((hi), (x))))\n\n#define DRWAV_MAX_SIMD_VECTOR_SIZE 64 /* 64 for AVX-512 in the future. */\n\n/* CPU architecture. */\n#if defined(__x86_64__) || defined(_M_X64)\n #define DRWAV_X64\n#elif defined(__i386) || defined(_M_IX86)\n #define DRWAV_X86\n#elif defined(__arm__) || defined(_M_ARM)\n #define DRWAV_ARM\n#endif\n\n#ifdef _MSC_VER\n #define DRWAV_INLINE __forceinline\n#elif defined(__GNUC__)\n /*\n I've had a bug report where GCC is emitting warnings about functions possibly not being inlineable. This warning happens when\n the __attribute__((always_inline)) attribute is defined without an \"inline\" statement. I think therefore there must be some\n case where \"__inline__\" is not always defined, thus the compiler emitting these warnings. When using -std=c89 or -ansi on the\n command line, we cannot use the \"inline\" keyword and instead need to use \"__inline__\". In an attempt to work around this issue\n I am using \"__inline__\" only when we're compiling in strict ANSI mode.\n */\n #if defined(__STRICT_ANSI__)\n #define DRWAV_INLINE __inline__ __attribute__((always_inline))\n #else\n #define DRWAV_INLINE inline __attribute__((always_inline))\n #endif\n#elif defined(__WATCOMC__)\n #define DRWAV_INLINE __inline\n#else\n #define DRWAV_INLINE\n#endif\n\n#if defined(SIZE_MAX)\n #define DRWAV_SIZE_MAX SIZE_MAX\n#else\n #if defined(_WIN64) || defined(_LP64) || defined(__LP64__)\n #define DRWAV_SIZE_MAX ((drwav_uint64)0xFFFFFFFFFFFFFFFF)\n #else\n #define DRWAV_SIZE_MAX 0xFFFFFFFF\n #endif\n#endif\n\n#if defined(_MSC_VER) && _MSC_VER >= 1400\n #define DRWAV_HAS_BYTESWAP16_INTRINSIC\n #define DRWAV_HAS_BYTESWAP32_INTRINSIC\n #define DRWAV_HAS_BYTESWAP64_INTRINSIC\n#elif defined(__clang__)\n #if defined(__has_builtin)\n #if __has_builtin(__builtin_bswap16)\n #define DRWAV_HAS_BYTESWAP16_INTRINSIC\n #endif\n #if __has_builtin(__builtin_bswap32)\n #define DRWAV_HAS_BYTESWAP32_INTRINSIC\n #endif\n #if __has_builtin(__builtin_bswap64)\n #define DRWAV_HAS_BYTESWAP64_INTRINSIC\n #endif\n #endif\n#elif defined(__GNUC__)\n #if ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))\n #define DRWAV_HAS_BYTESWAP32_INTRINSIC\n #define DRWAV_HAS_BYTESWAP64_INTRINSIC\n #endif\n #if ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8))\n #define DRWAV_HAS_BYTESWAP16_INTRINSIC\n #endif\n#endif\n\nDRWAV_API void drwav_version(drwav_uint32* pMajor, drwav_uint32* pMinor, drwav_uint32* pRevision)\n{\n if (pMajor) {\n *pMajor = DRWAV_VERSION_MAJOR;\n }\n\n if (pMinor) {\n *pMinor = DRWAV_VERSION_MINOR;\n }\n\n if (pRevision) {\n *pRevision = DRWAV_VERSION_REVISION;\n }\n}\n\nDRWAV_API const char* drwav_version_string(void)\n{\n return DRWAV_VERSION_STRING;\n}\n\n/*\nThese limits are used for basic validation when initializing the decoder. If you exceed these limits, first of all: what on Earth are\nyou doing?! (Let me know, I'd be curious!) Second, you can adjust these by #define-ing them before the dr_wav implementation.\n*/\n#ifndef DRWAV_MAX_SAMPLE_RATE\n#define DRWAV_MAX_SAMPLE_RATE 384000\n#endif\n#ifndef DRWAV_MAX_CHANNELS\n#define DRWAV_MAX_CHANNELS 256\n#endif\n#ifndef DRWAV_MAX_BITS_PER_SAMPLE\n#define DRWAV_MAX_BITS_PER_SAMPLE 64\n#endif\n\nstatic const drwav_uint8 drwavGUID_W64_RIFF[16] = {0x72,0x69,0x66,0x66, 0x2E,0x91, 0xCF,0x11, 0xA5,0xD6, 0x28,0xDB,0x04,0xC1,0x00,0x00}; /* 66666972-912E-11CF-A5D6-28DB04C10000 */\nstatic const drwav_uint8 drwavGUID_W64_WAVE[16] = {0x77,0x61,0x76,0x65, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A}; /* 65766177-ACF3-11D3-8CD1-00C04F8EDB8A */\n/*static const drwav_uint8 drwavGUID_W64_JUNK[16] = {0x6A,0x75,0x6E,0x6B, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};*/ /* 6B6E756A-ACF3-11D3-8CD1-00C04F8EDB8A */\nstatic const drwav_uint8 drwavGUID_W64_FMT [16] = {0x66,0x6D,0x74,0x20, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A}; /* 20746D66-ACF3-11D3-8CD1-00C04F8EDB8A */\nstatic const drwav_uint8 drwavGUID_W64_FACT[16] = {0x66,0x61,0x63,0x74, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A}; /* 74636166-ACF3-11D3-8CD1-00C04F8EDB8A */\nstatic const drwav_uint8 drwavGUID_W64_DATA[16] = {0x64,0x61,0x74,0x61, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A}; /* 61746164-ACF3-11D3-8CD1-00C04F8EDB8A */\nstatic const drwav_uint8 drwavGUID_W64_SMPL[16] = {0x73,0x6D,0x70,0x6C, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A}; /* 6C706D73-ACF3-11D3-8CD1-00C04F8EDB8A */\n\nstatic DRWAV_INLINE drwav_bool32 drwav__guid_equal(const drwav_uint8 a[16], const drwav_uint8 b[16])\n{\n int i;\n for (i = 0; i < 16; i += 1) {\n if (a[i] != b[i]) {\n return DRWAV_FALSE;\n }\n }\n\n return DRWAV_TRUE;\n}\n\nstatic DRWAV_INLINE drwav_bool32 drwav__fourcc_equal(const drwav_uint8* a, const char* b)\n{\n return\n a[0] == b[0] &&\n a[1] == b[1] &&\n a[2] == b[2] &&\n a[3] == b[3];\n}\n\n\n\nstatic DRWAV_INLINE int drwav__is_little_endian(void)\n{\n#if defined(DRWAV_X86) || defined(DRWAV_X64)\n return DRWAV_TRUE;\n#elif defined(__BYTE_ORDER) && defined(__LITTLE_ENDIAN) && __BYTE_ORDER == __LITTLE_ENDIAN\n return DRWAV_TRUE;\n#else\n int n = 1;\n return (*(char*)&n) == 1;\n#endif\n}\n\nstatic DRWAV_INLINE drwav_uint16 drwav__bytes_to_u16(const drwav_uint8* data)\n{\n return (data[0] << 0) | (data[1] << 8);\n}\n\nstatic DRWAV_INLINE drwav_int16 drwav__bytes_to_s16(const drwav_uint8* data)\n{\n return (short)drwav__bytes_to_u16(data);\n}\n\nstatic DRWAV_INLINE drwav_uint32 drwav__bytes_to_u32(const drwav_uint8* data)\n{\n return (data[0] << 0) | (data[1] << 8) | (data[2] << 16) | (data[3] << 24);\n}\n\nstatic DRWAV_INLINE drwav_int32 drwav__bytes_to_s32(const drwav_uint8* data)\n{\n return (drwav_int32)drwav__bytes_to_u32(data);\n}\n\nstatic DRWAV_INLINE drwav_uint64 drwav__bytes_to_u64(const drwav_uint8* data)\n{\n return\n ((drwav_uint64)data[0] << 0) | ((drwav_uint64)data[1] << 8) | ((drwav_uint64)data[2] << 16) | ((drwav_uint64)data[3] << 24) |\n ((drwav_uint64)data[4] << 32) | ((drwav_uint64)data[5] << 40) | ((drwav_uint64)data[6] << 48) | ((drwav_uint64)data[7] << 56);\n}\n\nstatic DRWAV_INLINE drwav_int64 drwav__bytes_to_s64(const drwav_uint8* data)\n{\n return (drwav_int64)drwav__bytes_to_u64(data);\n}\n\nstatic DRWAV_INLINE void drwav__bytes_to_guid(const drwav_uint8* data, drwav_uint8* guid)\n{\n int i;\n for (i = 0; i < 16; ++i) {\n guid[i] = data[i];\n }\n}\n\n\nstatic DRWAV_INLINE drwav_uint16 drwav__bswap16(drwav_uint16 n)\n{\n#ifdef DRWAV_HAS_BYTESWAP16_INTRINSIC\n #if defined(_MSC_VER)\n return _byteswap_ushort(n);\n #elif defined(__GNUC__) || defined(__clang__)\n return __builtin_bswap16(n);\n #else\n #error \"This compiler does not support the byte swap intrinsic.\"\n #endif\n#else\n return ((n & 0xFF00) >> 8) |\n ((n & 0x00FF) << 8);\n#endif\n}\n\nstatic DRWAV_INLINE drwav_uint32 drwav__bswap32(drwav_uint32 n)\n{\n#ifdef DRWAV_HAS_BYTESWAP32_INTRINSIC\n #if defined(_MSC_VER)\n return _byteswap_ulong(n);\n #elif defined(__GNUC__) || defined(__clang__)\n #if defined(DRWAV_ARM) && (defined(__ARM_ARCH) && __ARM_ARCH >= 6) && !defined(DRWAV_64BIT) /* <-- 64-bit inline assembly has not been tested, so disabling for now. */\n /* Inline assembly optimized implementation for ARM. In my testing, GCC does not generate optimized code with __builtin_bswap32(). */\n drwav_uint32 r;\n __asm__ __volatile__ (\n #if defined(DRWAV_64BIT)\n \"rev %w[out], %w[in]\" : [out]\"=r\"(r) : [in]\"r\"(n) /* <-- This is untested. If someone in the community could test this, that would be appreciated! */\n #else\n \"rev %[out], %[in]\" : [out]\"=r\"(r) : [in]\"r\"(n)\n #endif\n );\n return r;\n #else\n return __builtin_bswap32(n);\n #endif\n #else\n #error \"This compiler does not support the byte swap intrinsic.\"\n #endif\n#else\n return ((n & 0xFF000000) >> 24) |\n ((n & 0x00FF0000) >> 8) |\n ((n & 0x0000FF00) << 8) |\n ((n & 0x000000FF) << 24);\n#endif\n}\n\nstatic DRWAV_INLINE drwav_uint64 drwav__bswap64(drwav_uint64 n)\n{\n#ifdef DRWAV_HAS_BYTESWAP64_INTRINSIC\n #if defined(_MSC_VER)\n return _byteswap_uint64(n);\n #elif defined(__GNUC__) || defined(__clang__)\n return __builtin_bswap64(n);\n #else\n #error \"This compiler does not support the byte swap intrinsic.\"\n #endif\n#else\n /* Weird \"<< 32\" bitshift is required for C89 because it doesn't support 64-bit constants. Should be optimized out by a good compiler. */\n return ((n & ((drwav_uint64)0xFF000000 << 32)) >> 56) |\n ((n & ((drwav_uint64)0x00FF0000 << 32)) >> 40) |\n ((n & ((drwav_uint64)0x0000FF00 << 32)) >> 24) |\n ((n & ((drwav_uint64)0x000000FF << 32)) >> 8) |\n ((n & ((drwav_uint64)0xFF000000 )) << 8) |\n ((n & ((drwav_uint64)0x00FF0000 )) << 24) |\n ((n & ((drwav_uint64)0x0000FF00 )) << 40) |\n ((n & ((drwav_uint64)0x000000FF )) << 56);\n#endif\n}\n\n\nstatic DRWAV_INLINE drwav_int16 drwav__bswap_s16(drwav_int16 n)\n{\n return (drwav_int16)drwav__bswap16((drwav_uint16)n);\n}\n\nstatic DRWAV_INLINE void drwav__bswap_samples_s16(drwav_int16* pSamples, drwav_uint64 sampleCount)\n{\n drwav_uint64 iSample;\n for (iSample = 0; iSample < sampleCount; iSample += 1) {\n pSamples[iSample] = drwav__bswap_s16(pSamples[iSample]);\n }\n}\n\n\nstatic DRWAV_INLINE void drwav__bswap_s24(drwav_uint8* p)\n{\n drwav_uint8 t;\n t = p[0];\n p[0] = p[2];\n p[2] = t;\n}\n\nstatic DRWAV_INLINE void drwav__bswap_samples_s24(drwav_uint8* pSamples, drwav_uint64 sampleCount)\n{\n drwav_uint64 iSample;\n for (iSample = 0; iSample < sampleCount; iSample += 1) {\n drwav_uint8* pSample = pSamples + (iSample*3);\n drwav__bswap_s24(pSample);\n }\n}\n\n\nstatic DRWAV_INLINE drwav_int32 drwav__bswap_s32(drwav_int32 n)\n{\n return (drwav_int32)drwav__bswap32((drwav_uint32)n);\n}\n\nstatic DRWAV_INLINE void drwav__bswap_samples_s32(drwav_int32* pSamples, drwav_uint64 sampleCount)\n{\n drwav_uint64 iSample;\n for (iSample = 0; iSample < sampleCount; iSample += 1) {\n pSamples[iSample] = drwav__bswap_s32(pSamples[iSample]);\n }\n}\n\n\nstatic DRWAV_INLINE float drwav__bswap_f32(float n)\n{\n union {\n drwav_uint32 i;\n float f;\n } x;\n x.f = n;\n x.i = drwav__bswap32(x.i);\n\n return x.f;\n}\n\nstatic DRWAV_INLINE void drwav__bswap_samples_f32(float* pSamples, drwav_uint64 sampleCount)\n{\n drwav_uint64 iSample;\n for (iSample = 0; iSample < sampleCount; iSample += 1) {\n pSamples[iSample] = drwav__bswap_f32(pSamples[iSample]);\n }\n}\n\n\nstatic DRWAV_INLINE double drwav__bswap_f64(double n)\n{\n union {\n drwav_uint64 i;\n double f;\n } x;\n x.f = n;\n x.i = drwav__bswap64(x.i);\n\n return x.f;\n}\n\nstatic DRWAV_INLINE void drwav__bswap_samples_f64(double* pSamples, drwav_uint64 sampleCount)\n{\n drwav_uint64 iSample;\n for (iSample = 0; iSample < sampleCount; iSample += 1) {\n pSamples[iSample] = drwav__bswap_f64(pSamples[iSample]);\n }\n}\n\n\nstatic DRWAV_INLINE void drwav__bswap_samples_pcm(void* pSamples, drwav_uint64 sampleCount, drwav_uint32 bytesPerSample)\n{\n /* Assumes integer PCM. Floating point PCM is done in drwav__bswap_samples_ieee(). */\n switch (bytesPerSample)\n {\n case 2: /* s16, s12 (loosely packed) */\n {\n drwav__bswap_samples_s16((drwav_int16*)pSamples, sampleCount);\n } break;\n case 3: /* s24 */\n {\n drwav__bswap_samples_s24((drwav_uint8*)pSamples, sampleCount);\n } break;\n case 4: /* s32 */\n {\n drwav__bswap_samples_s32((drwav_int32*)pSamples, sampleCount);\n } break;\n default:\n {\n /* Unsupported format. */\n DRWAV_ASSERT(DRWAV_FALSE);\n } break;\n }\n}\n\nstatic DRWAV_INLINE void drwav__bswap_samples_ieee(void* pSamples, drwav_uint64 sampleCount, drwav_uint32 bytesPerSample)\n{\n switch (bytesPerSample)\n {\n #if 0 /* Contributions welcome for f16 support. */\n case 2: /* f16 */\n {\n drwav__bswap_samples_f16((drwav_float16*)pSamples, sampleCount);\n } break;\n #endif\n case 4: /* f32 */\n {\n drwav__bswap_samples_f32((float*)pSamples, sampleCount);\n } break;\n case 8: /* f64 */\n {\n drwav__bswap_samples_f64((double*)pSamples, sampleCount);\n } break;\n default:\n {\n /* Unsupported format. */\n DRWAV_ASSERT(DRWAV_FALSE);\n } break;\n }\n}\n\nstatic DRWAV_INLINE void drwav__bswap_samples(void* pSamples, drwav_uint64 sampleCount, drwav_uint32 bytesPerSample, drwav_uint16 format)\n{\n switch (format)\n {\n case DR_WAVE_FORMAT_PCM:\n {\n drwav__bswap_samples_pcm(pSamples, sampleCount, bytesPerSample);\n } break;\n\n case DR_WAVE_FORMAT_IEEE_FLOAT:\n {\n drwav__bswap_samples_ieee(pSamples, sampleCount, bytesPerSample);\n } break;\n\n case DR_WAVE_FORMAT_ALAW:\n case DR_WAVE_FORMAT_MULAW:\n {\n drwav__bswap_samples_s16((drwav_int16*)pSamples, sampleCount);\n } break;\n\n case DR_WAVE_FORMAT_ADPCM:\n case DR_WAVE_FORMAT_DVI_ADPCM:\n default:\n {\n /* Unsupported format. */\n DRWAV_ASSERT(DRWAV_FALSE);\n } break;\n }\n}\n\n\nstatic void* drwav__malloc_default(size_t sz, void* pUserData)\n{\n (void)pUserData;\n return DRWAV_MALLOC(sz);\n}\n\nstatic void* drwav__realloc_default(void* p, size_t sz, void* pUserData)\n{\n (void)pUserData;\n return DRWAV_REALLOC(p, sz);\n}\n\nstatic void drwav__free_default(void* p, void* pUserData)\n{\n (void)pUserData;\n DRWAV_FREE(p);\n}\n\n\nstatic void* drwav__malloc_from_callbacks(size_t sz, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pAllocationCallbacks == NULL) {\n return NULL;\n }\n\n if (pAllocationCallbacks->onMalloc != NULL) {\n return pAllocationCallbacks->onMalloc(sz, pAllocationCallbacks->pUserData);\n }\n\n /* Try using realloc(). */\n if (pAllocationCallbacks->onRealloc != NULL) {\n return pAllocationCallbacks->onRealloc(NULL, sz, pAllocationCallbacks->pUserData);\n }\n\n return NULL;\n}\n\nstatic void* drwav__realloc_from_callbacks(void* p, size_t szNew, size_t szOld, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pAllocationCallbacks == NULL) {\n return NULL;\n }\n\n if (pAllocationCallbacks->onRealloc != NULL) {\n return pAllocationCallbacks->onRealloc(p, szNew, pAllocationCallbacks->pUserData);\n }\n\n /* Try emulating realloc() in terms of malloc()/free(). */\n if (pAllocationCallbacks->onMalloc != NULL && pAllocationCallbacks->onFree != NULL) {\n void* p2;\n\n p2 = pAllocationCallbacks->onMalloc(szNew, pAllocationCallbacks->pUserData);\n if (p2 == NULL) {\n return NULL;\n }\n\n if (p != NULL) {\n DRWAV_COPY_MEMORY(p2, p, szOld);\n pAllocationCallbacks->onFree(p, pAllocationCallbacks->pUserData);\n }\n\n return p2;\n }\n\n return NULL;\n}\n\nstatic void drwav__free_from_callbacks(void* p, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (p == NULL || pAllocationCallbacks == NULL) {\n return;\n }\n\n if (pAllocationCallbacks->onFree != NULL) {\n pAllocationCallbacks->onFree(p, pAllocationCallbacks->pUserData);\n }\n}\n\n\nstatic drwav_allocation_callbacks drwav_copy_allocation_callbacks_or_defaults(const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pAllocationCallbacks != NULL) {\n /* Copy. */\n return *pAllocationCallbacks;\n } else {\n /* Defaults. */\n drwav_allocation_callbacks allocationCallbacks;\n allocationCallbacks.pUserData = NULL;\n allocationCallbacks.onMalloc = drwav__malloc_default;\n allocationCallbacks.onRealloc = drwav__realloc_default;\n allocationCallbacks.onFree = drwav__free_default;\n return allocationCallbacks;\n }\n}\n\n\nstatic DRWAV_INLINE drwav_bool32 drwav__is_compressed_format_tag(drwav_uint16 formatTag)\n{\n return\n formatTag == DR_WAVE_FORMAT_ADPCM ||\n formatTag == DR_WAVE_FORMAT_DVI_ADPCM;\n}\n\nstatic unsigned int drwav__chunk_padding_size_riff(drwav_uint64 chunkSize)\n{\n return (unsigned int)(chunkSize % 2);\n}\n\nstatic unsigned int drwav__chunk_padding_size_w64(drwav_uint64 chunkSize)\n{\n return (unsigned int)(chunkSize % 8);\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s16__msadpcm(drwav* pWav, drwav_uint64 samplesToRead, drwav_int16* pBufferOut);\nstatic drwav_uint64 drwav_read_pcm_frames_s16__ima(drwav* pWav, drwav_uint64 samplesToRead, drwav_int16* pBufferOut);\nstatic drwav_bool32 drwav_init_write__internal(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount);\n\nstatic drwav_result drwav__read_chunk_header(drwav_read_proc onRead, void* pUserData, drwav_container container, drwav_uint64* pRunningBytesReadOut, drwav_chunk_header* pHeaderOut)\n{\n if (container == drwav_container_riff || container == drwav_container_rf64) {\n drwav_uint8 sizeInBytes[4];\n\n if (onRead(pUserData, pHeaderOut->id.fourcc, 4) != 4) {\n return DRWAV_AT_END;\n }\n\n if (onRead(pUserData, sizeInBytes, 4) != 4) {\n return DRWAV_INVALID_FILE;\n }\n\n pHeaderOut->sizeInBytes = drwav__bytes_to_u32(sizeInBytes);\n pHeaderOut->paddingSize = drwav__chunk_padding_size_riff(pHeaderOut->sizeInBytes);\n *pRunningBytesReadOut += 8;\n } else {\n drwav_uint8 sizeInBytes[8];\n\n if (onRead(pUserData, pHeaderOut->id.guid, 16) != 16) {\n return DRWAV_AT_END;\n }\n\n if (onRead(pUserData, sizeInBytes, 8) != 8) {\n return DRWAV_INVALID_FILE;\n }\n\n pHeaderOut->sizeInBytes = drwav__bytes_to_u64(sizeInBytes) - 24; /* <-- Subtract 24 because w64 includes the size of the header. */\n pHeaderOut->paddingSize = drwav__chunk_padding_size_w64(pHeaderOut->sizeInBytes);\n *pRunningBytesReadOut += 24;\n }\n\n return DRWAV_SUCCESS;\n}\n\nstatic drwav_bool32 drwav__seek_forward(drwav_seek_proc onSeek, drwav_uint64 offset, void* pUserData)\n{\n drwav_uint64 bytesRemainingToSeek = offset;\n while (bytesRemainingToSeek > 0) {\n if (bytesRemainingToSeek > 0x7FFFFFFF) {\n if (!onSeek(pUserData, 0x7FFFFFFF, drwav_seek_origin_current)) {\n return DRWAV_FALSE;\n }\n bytesRemainingToSeek -= 0x7FFFFFFF;\n } else {\n if (!onSeek(pUserData, (int)bytesRemainingToSeek, drwav_seek_origin_current)) {\n return DRWAV_FALSE;\n }\n bytesRemainingToSeek = 0;\n }\n }\n\n return DRWAV_TRUE;\n}\n\nstatic drwav_bool32 drwav__seek_from_start(drwav_seek_proc onSeek, drwav_uint64 offset, void* pUserData)\n{\n if (offset <= 0x7FFFFFFF) {\n return onSeek(pUserData, (int)offset, drwav_seek_origin_start);\n }\n\n /* Larger than 32-bit seek. */\n if (!onSeek(pUserData, 0x7FFFFFFF, drwav_seek_origin_start)) {\n return DRWAV_FALSE;\n }\n offset -= 0x7FFFFFFF;\n\n for (;;) {\n if (offset <= 0x7FFFFFFF) {\n return onSeek(pUserData, (int)offset, drwav_seek_origin_current);\n }\n\n if (!onSeek(pUserData, 0x7FFFFFFF, drwav_seek_origin_current)) {\n return DRWAV_FALSE;\n }\n offset -= 0x7FFFFFFF;\n }\n\n /* Should never get here. */\n /*return DRWAV_TRUE; */\n}\n\n\nstatic drwav_bool32 drwav__read_fmt(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, drwav_container container, drwav_uint64* pRunningBytesReadOut, drwav_fmt* fmtOut)\n{\n drwav_chunk_header header;\n drwav_uint8 fmt[16];\n\n if (drwav__read_chunk_header(onRead, pUserData, container, pRunningBytesReadOut, &header) != DRWAV_SUCCESS) {\n return DRWAV_FALSE;\n }\n\n\n /* Skip non-fmt chunks. */\n while (((container == drwav_container_riff || container == drwav_container_rf64) && !drwav__fourcc_equal(header.id.fourcc, \"fmt \")) || (container == drwav_container_w64 && !drwav__guid_equal(header.id.guid, drwavGUID_W64_FMT))) {\n if (!drwav__seek_forward(onSeek, header.sizeInBytes + header.paddingSize, pUserData)) {\n return DRWAV_FALSE;\n }\n *pRunningBytesReadOut += header.sizeInBytes + header.paddingSize;\n\n /* Try the next header. */\n if (drwav__read_chunk_header(onRead, pUserData, container, pRunningBytesReadOut, &header) != DRWAV_SUCCESS) {\n return DRWAV_FALSE;\n }\n }\n\n\n /* Validation. */\n if (container == drwav_container_riff || container == drwav_container_rf64) {\n if (!drwav__fourcc_equal(header.id.fourcc, \"fmt \")) {\n return DRWAV_FALSE;\n }\n } else {\n if (!drwav__guid_equal(header.id.guid, drwavGUID_W64_FMT)) {\n return DRWAV_FALSE;\n }\n }\n\n\n if (onRead(pUserData, fmt, sizeof(fmt)) != sizeof(fmt)) {\n return DRWAV_FALSE;\n }\n *pRunningBytesReadOut += sizeof(fmt);\n\n fmtOut->formatTag = drwav__bytes_to_u16(fmt + 0);\n fmtOut->channels = drwav__bytes_to_u16(fmt + 2);\n fmtOut->sampleRate = drwav__bytes_to_u32(fmt + 4);\n fmtOut->avgBytesPerSec = drwav__bytes_to_u32(fmt + 8);\n fmtOut->blockAlign = drwav__bytes_to_u16(fmt + 12);\n fmtOut->bitsPerSample = drwav__bytes_to_u16(fmt + 14);\n\n fmtOut->extendedSize = 0;\n fmtOut->validBitsPerSample = 0;\n fmtOut->channelMask = 0;\n memset(fmtOut->subFormat, 0, sizeof(fmtOut->subFormat));\n\n if (header.sizeInBytes > 16) {\n drwav_uint8 fmt_cbSize[2];\n int bytesReadSoFar = 0;\n\n if (onRead(pUserData, fmt_cbSize, sizeof(fmt_cbSize)) != sizeof(fmt_cbSize)) {\n return DRWAV_FALSE; /* Expecting more data. */\n }\n *pRunningBytesReadOut += sizeof(fmt_cbSize);\n\n bytesReadSoFar = 18;\n\n fmtOut->extendedSize = drwav__bytes_to_u16(fmt_cbSize);\n if (fmtOut->extendedSize > 0) {\n /* Simple validation. */\n if (fmtOut->formatTag == DR_WAVE_FORMAT_EXTENSIBLE) {\n if (fmtOut->extendedSize != 22) {\n return DRWAV_FALSE;\n }\n }\n\n if (fmtOut->formatTag == DR_WAVE_FORMAT_EXTENSIBLE) {\n drwav_uint8 fmtext[22];\n if (onRead(pUserData, fmtext, fmtOut->extendedSize) != fmtOut->extendedSize) {\n return DRWAV_FALSE; /* Expecting more data. */\n }\n\n fmtOut->validBitsPerSample = drwav__bytes_to_u16(fmtext + 0);\n fmtOut->channelMask = drwav__bytes_to_u32(fmtext + 2);\n drwav__bytes_to_guid(fmtext + 6, fmtOut->subFormat);\n } else {\n if (!onSeek(pUserData, fmtOut->extendedSize, drwav_seek_origin_current)) {\n return DRWAV_FALSE;\n }\n }\n *pRunningBytesReadOut += fmtOut->extendedSize;\n\n bytesReadSoFar += fmtOut->extendedSize;\n }\n\n /* Seek past any leftover bytes. For w64 the leftover will be defined based on the chunk size. */\n if (!onSeek(pUserData, (int)(header.sizeInBytes - bytesReadSoFar), drwav_seek_origin_current)) {\n return DRWAV_FALSE;\n }\n *pRunningBytesReadOut += (header.sizeInBytes - bytesReadSoFar);\n }\n\n if (header.paddingSize > 0) {\n if (!onSeek(pUserData, header.paddingSize, drwav_seek_origin_current)) {\n return DRWAV_FALSE;\n }\n *pRunningBytesReadOut += header.paddingSize;\n }\n\n return DRWAV_TRUE;\n}\n\n\nstatic size_t drwav__on_read(drwav_read_proc onRead, void* pUserData, void* pBufferOut, size_t bytesToRead, drwav_uint64* pCursor)\n{\n size_t bytesRead;\n\n DRWAV_ASSERT(onRead != NULL);\n DRWAV_ASSERT(pCursor != NULL);\n\n bytesRead = onRead(pUserData, pBufferOut, bytesToRead);\n *pCursor += bytesRead;\n return bytesRead;\n}\n\n#if 0\nstatic drwav_bool32 drwav__on_seek(drwav_seek_proc onSeek, void* pUserData, int offset, drwav_seek_origin origin, drwav_uint64* pCursor)\n{\n DRWAV_ASSERT(onSeek != NULL);\n DRWAV_ASSERT(pCursor != NULL);\n\n if (!onSeek(pUserData, offset, origin)) {\n return DRWAV_FALSE;\n }\n\n if (origin == drwav_seek_origin_start) {\n *pCursor = offset;\n } else {\n *pCursor += offset;\n }\n\n return DRWAV_TRUE;\n}\n#endif\n\n\n\nstatic drwav_uint32 drwav_get_bytes_per_pcm_frame(drwav* pWav)\n{\n /*\n The bytes per frame is a bit ambiguous. It can be either be based on the bits per sample, or the block align. The way I'm doing it here\n is that if the bits per sample is a multiple of 8, use floor(bitsPerSample*channels/8), otherwise fall back to the block align.\n */\n if ((pWav->bitsPerSample & 0x7) == 0) {\n /* Bits per sample is a multiple of 8. */\n return (pWav->bitsPerSample * pWav->fmt.channels) >> 3;\n } else {\n return pWav->fmt.blockAlign;\n }\n}\n\nDRWAV_API drwav_uint16 drwav_fmt_get_format(const drwav_fmt* pFMT)\n{\n if (pFMT == NULL) {\n return 0;\n }\n\n if (pFMT->formatTag != DR_WAVE_FORMAT_EXTENSIBLE) {\n return pFMT->formatTag;\n } else {\n return drwav__bytes_to_u16(pFMT->subFormat); /* Only the first two bytes are required. */\n }\n}\n\nstatic drwav_bool32 drwav_preinit(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pReadSeekUserData, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pWav == NULL || onRead == NULL || onSeek == NULL) {\n return DRWAV_FALSE;\n }\n\n DRWAV_ZERO_MEMORY(pWav, sizeof(*pWav));\n pWav->onRead = onRead;\n pWav->onSeek = onSeek;\n pWav->pUserData = pReadSeekUserData;\n pWav->allocationCallbacks = drwav_copy_allocation_callbacks_or_defaults(pAllocationCallbacks);\n\n if (pWav->allocationCallbacks.onFree == NULL || (pWav->allocationCallbacks.onMalloc == NULL && pWav->allocationCallbacks.onRealloc == NULL)) {\n return DRWAV_FALSE; /* Invalid allocation callbacks. */\n }\n\n return DRWAV_TRUE;\n}\n\nstatic drwav_bool32 drwav_init__internal(drwav* pWav, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags)\n{\n /* This function assumes drwav_preinit() has been called beforehand. */\n\n drwav_uint64 cursor; /* <-- Keeps track of the byte position so we can seek to specific locations. */\n drwav_bool32 sequential;\n drwav_uint8 riff[4];\n drwav_fmt fmt;\n unsigned short translatedFormatTag;\n drwav_bool32 foundDataChunk;\n drwav_uint64 dataChunkSize = 0; /* <-- Important! Don't explicitly set this to 0 anywhere else. Calculation of the size of the data chunk is performed in different paths depending on the container. */\n drwav_uint64 sampleCountFromFactChunk = 0; /* Same as dataChunkSize - make sure this is the only place this is initialized to 0. */\n drwav_uint64 chunkSize;\n\n cursor = 0;\n sequential = (flags & DRWAV_SEQUENTIAL) != 0;\n\n /* The first 4 bytes should be the RIFF identifier. */\n if (drwav__on_read(pWav->onRead, pWav->pUserData, riff, sizeof(riff), &cursor) != sizeof(riff)) {\n return DRWAV_FALSE;\n }\n\n /*\n The first 4 bytes can be used to identify the container. For RIFF files it will start with \"RIFF\" and for\n w64 it will start with \"riff\".\n */\n if (drwav__fourcc_equal(riff, \"RIFF\")) {\n pWav->container = drwav_container_riff;\n } else if (drwav__fourcc_equal(riff, \"riff\")) {\n int i;\n drwav_uint8 riff2[12];\n\n pWav->container = drwav_container_w64;\n\n /* Check the rest of the GUID for validity. */\n if (drwav__on_read(pWav->onRead, pWav->pUserData, riff2, sizeof(riff2), &cursor) != sizeof(riff2)) {\n return DRWAV_FALSE;\n }\n\n for (i = 0; i < 12; ++i) {\n if (riff2[i] != drwavGUID_W64_RIFF[i+4]) {\n return DRWAV_FALSE;\n }\n }\n } else if (drwav__fourcc_equal(riff, \"RF64\")) {\n pWav->container = drwav_container_rf64;\n } else {\n return DRWAV_FALSE; /* Unknown or unsupported container. */\n }\n\n\n if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rf64) {\n drwav_uint8 chunkSizeBytes[4];\n drwav_uint8 wave[4];\n\n /* RIFF/WAVE */\n if (drwav__on_read(pWav->onRead, pWav->pUserData, chunkSizeBytes, sizeof(chunkSizeBytes), &cursor) != sizeof(chunkSizeBytes)) {\n return DRWAV_FALSE;\n }\n\n if (pWav->container == drwav_container_riff) {\n if (drwav__bytes_to_u32(chunkSizeBytes) < 36) {\n return DRWAV_FALSE; /* Chunk size should always be at least 36 bytes. */\n }\n } else {\n if (drwav__bytes_to_u32(chunkSizeBytes) != 0xFFFFFFFF) {\n return DRWAV_FALSE; /* Chunk size should always be set to -1/0xFFFFFFFF for RF64. The actual size is retrieved later. */\n }\n }\n\n if (drwav__on_read(pWav->onRead, pWav->pUserData, wave, sizeof(wave), &cursor) != sizeof(wave)) {\n return DRWAV_FALSE;\n }\n\n if (!drwav__fourcc_equal(wave, \"WAVE\")) {\n return DRWAV_FALSE; /* Expecting \"WAVE\". */\n }\n } else {\n drwav_uint8 chunkSizeBytes[8];\n drwav_uint8 wave[16];\n\n /* W64 */\n if (drwav__on_read(pWav->onRead, pWav->pUserData, chunkSizeBytes, sizeof(chunkSizeBytes), &cursor) != sizeof(chunkSizeBytes)) {\n return DRWAV_FALSE;\n }\n\n if (drwav__bytes_to_u64(chunkSizeBytes) < 80) {\n return DRWAV_FALSE;\n }\n\n if (drwav__on_read(pWav->onRead, pWav->pUserData, wave, sizeof(wave), &cursor) != sizeof(wave)) {\n return DRWAV_FALSE;\n }\n\n if (!drwav__guid_equal(wave, drwavGUID_W64_WAVE)) {\n return DRWAV_FALSE;\n }\n }\n\n\n /* For RF64, the \"ds64\" chunk must come next, before the \"fmt \" chunk. */\n if (pWav->container == drwav_container_rf64) {\n drwav_uint8 sizeBytes[8];\n drwav_uint64 bytesRemainingInChunk;\n drwav_chunk_header header;\n drwav_result result = drwav__read_chunk_header(pWav->onRead, pWav->pUserData, pWav->container, &cursor, &header);\n if (result != DRWAV_SUCCESS) {\n return DRWAV_FALSE;\n }\n\n if (!drwav__fourcc_equal(header.id.fourcc, \"ds64\")) {\n return DRWAV_FALSE; /* Expecting \"ds64\". */\n }\n\n bytesRemainingInChunk = header.sizeInBytes + header.paddingSize;\n\n /* We don't care about the size of the RIFF chunk - skip it. */\n if (!drwav__seek_forward(pWav->onSeek, 8, pWav->pUserData)) {\n return DRWAV_FALSE;\n }\n bytesRemainingInChunk -= 8;\n cursor += 8;\n\n\n /* Next 8 bytes is the size of the \"data\" chunk. */\n if (drwav__on_read(pWav->onRead, pWav->pUserData, sizeBytes, sizeof(sizeBytes), &cursor) != sizeof(sizeBytes)) {\n return DRWAV_FALSE;\n }\n bytesRemainingInChunk -= 8;\n dataChunkSize = drwav__bytes_to_u64(sizeBytes);\n\n\n /* Next 8 bytes is the same count which we would usually derived from the FACT chunk if it was available. */\n if (drwav__on_read(pWav->onRead, pWav->pUserData, sizeBytes, sizeof(sizeBytes), &cursor) != sizeof(sizeBytes)) {\n return DRWAV_FALSE;\n }\n bytesRemainingInChunk -= 8;\n sampleCountFromFactChunk = drwav__bytes_to_u64(sizeBytes);\n\n\n /* Skip over everything else. */\n if (!drwav__seek_forward(pWav->onSeek, bytesRemainingInChunk, pWav->pUserData)) {\n return DRWAV_FALSE;\n }\n cursor += bytesRemainingInChunk;\n }\n\n\n /* The next bytes should be the \"fmt \" chunk. */\n if (!drwav__read_fmt(pWav->onRead, pWav->onSeek, pWav->pUserData, pWav->container, &cursor, &fmt)) {\n return DRWAV_FALSE; /* Failed to read the \"fmt \" chunk. */\n }\n\n /* Basic validation. */\n if ((fmt.sampleRate == 0 || fmt.sampleRate > DRWAV_MAX_SAMPLE_RATE) ||\n (fmt.channels == 0 || fmt.channels > DRWAV_MAX_CHANNELS) ||\n (fmt.bitsPerSample == 0 || fmt.bitsPerSample > DRWAV_MAX_BITS_PER_SAMPLE) ||\n fmt.blockAlign == 0) {\n return DRWAV_FALSE; /* Probably an invalid WAV file. */\n }\n\n\n /* Translate the internal format. */\n translatedFormatTag = fmt.formatTag;\n if (translatedFormatTag == DR_WAVE_FORMAT_EXTENSIBLE) {\n translatedFormatTag = drwav__bytes_to_u16(fmt.subFormat + 0);\n }\n\n\n /*\n We need to enumerate over each chunk for two reasons:\n 1) The \"data\" chunk may not be the next one\n 2) We may want to report each chunk back to the client\n \n In order to correctly report each chunk back to the client we will need to keep looping until the end of the file.\n */\n foundDataChunk = DRWAV_FALSE;\n\n /* The next chunk we care about is the \"data\" chunk. This is not necessarily the next chunk so we'll need to loop. */\n for (;;)\n {\n drwav_chunk_header header;\n drwav_result result = drwav__read_chunk_header(pWav->onRead, pWav->pUserData, pWav->container, &cursor, &header);\n if (result != DRWAV_SUCCESS) {\n if (!foundDataChunk) {\n return DRWAV_FALSE;\n } else {\n break; /* Probably at the end of the file. Get out of the loop. */\n }\n }\n\n /* Tell the client about this chunk. */\n if (!sequential && onChunk != NULL) {\n drwav_uint64 callbackBytesRead = onChunk(pChunkUserData, pWav->onRead, pWav->onSeek, pWav->pUserData, &header, pWav->container, &fmt);\n\n /*\n dr_wav may need to read the contents of the chunk, so we now need to seek back to the position before\n we called the callback.\n */\n if (callbackBytesRead > 0) {\n if (!drwav__seek_from_start(pWav->onSeek, cursor, pWav->pUserData)) {\n return DRWAV_FALSE;\n }\n }\n }\n \n\n if (!foundDataChunk) {\n pWav->dataChunkDataPos = cursor;\n }\n\n chunkSize = header.sizeInBytes;\n if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rf64) {\n if (drwav__fourcc_equal(header.id.fourcc, \"data\")) {\n foundDataChunk = DRWAV_TRUE;\n if (pWav->container != drwav_container_rf64) { /* The data chunk size for RF64 will always be set to 0xFFFFFFFF here. It was set to it's true value earlier. */\n dataChunkSize = chunkSize;\n }\n }\n } else {\n if (drwav__guid_equal(header.id.guid, drwavGUID_W64_DATA)) {\n foundDataChunk = DRWAV_TRUE;\n dataChunkSize = chunkSize;\n }\n }\n\n /*\n If at this point we have found the data chunk and we're running in sequential mode, we need to break out of this loop. The reason for\n this is that we would otherwise require a backwards seek which sequential mode forbids.\n */\n if (foundDataChunk && sequential) {\n break;\n }\n\n /* Optional. Get the total sample count from the FACT chunk. This is useful for compressed formats. */\n if (pWav->container == drwav_container_riff) {\n if (drwav__fourcc_equal(header.id.fourcc, \"fact\")) {\n drwav_uint32 sampleCount;\n if (drwav__on_read(pWav->onRead, pWav->pUserData, &sampleCount, 4, &cursor) != 4) {\n return DRWAV_FALSE;\n }\n chunkSize -= 4;\n\n if (!foundDataChunk) {\n pWav->dataChunkDataPos = cursor;\n }\n\n /*\n The sample count in the \"fact\" chunk is either unreliable, or I'm not understanding it properly. For now I am only enabling this\n for Microsoft ADPCM formats.\n */\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {\n sampleCountFromFactChunk = sampleCount;\n } else {\n sampleCountFromFactChunk = 0;\n }\n }\n } else if (pWav->container == drwav_container_w64) {\n if (drwav__guid_equal(header.id.guid, drwavGUID_W64_FACT)) {\n if (drwav__on_read(pWav->onRead, pWav->pUserData, &sampleCountFromFactChunk, 8, &cursor) != 8) {\n return DRWAV_FALSE;\n }\n chunkSize -= 8;\n\n if (!foundDataChunk) {\n pWav->dataChunkDataPos = cursor;\n }\n }\n } else if (pWav->container == drwav_container_rf64) {\n /* We retrieved the sample count from the ds64 chunk earlier so no need to do that here. */\n }\n\n /* \"smpl\" chunk. */\n if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rf64) {\n if (drwav__fourcc_equal(header.id.fourcc, \"smpl\")) {\n drwav_uint8 smplHeaderData[36]; /* 36 = size of the smpl header section, not including the loop data. */\n if (chunkSize >= sizeof(smplHeaderData)) {\n drwav_uint64 bytesJustRead = drwav__on_read(pWav->onRead, pWav->pUserData, smplHeaderData, sizeof(smplHeaderData), &cursor);\n chunkSize -= bytesJustRead;\n\n if (bytesJustRead == sizeof(smplHeaderData)) {\n drwav_uint32 iLoop;\n\n pWav->smpl.manufacturer = drwav__bytes_to_u32(smplHeaderData+0);\n pWav->smpl.product = drwav__bytes_to_u32(smplHeaderData+4);\n pWav->smpl.samplePeriod = drwav__bytes_to_u32(smplHeaderData+8);\n pWav->smpl.midiUnityNotes = drwav__bytes_to_u32(smplHeaderData+12);\n pWav->smpl.midiPitchFraction = drwav__bytes_to_u32(smplHeaderData+16);\n pWav->smpl.smpteFormat = drwav__bytes_to_u32(smplHeaderData+20);\n pWav->smpl.smpteOffset = drwav__bytes_to_u32(smplHeaderData+24);\n pWav->smpl.numSampleLoops = drwav__bytes_to_u32(smplHeaderData+28);\n pWav->smpl.samplerData = drwav__bytes_to_u32(smplHeaderData+32);\n\n for (iLoop = 0; iLoop < pWav->smpl.numSampleLoops && iLoop < drwav_countof(pWav->smpl.loops); ++iLoop) {\n drwav_uint8 smplLoopData[24]; /* 24 = size of a loop section in the smpl chunk. */\n bytesJustRead = drwav__on_read(pWav->onRead, pWav->pUserData, smplLoopData, sizeof(smplLoopData), &cursor);\n chunkSize -= bytesJustRead;\n\n if (bytesJustRead == sizeof(smplLoopData)) {\n pWav->smpl.loops[iLoop].cuePointId = drwav__bytes_to_u32(smplLoopData+0);\n pWav->smpl.loops[iLoop].type = drwav__bytes_to_u32(smplLoopData+4);\n pWav->smpl.loops[iLoop].start = drwav__bytes_to_u32(smplLoopData+8);\n pWav->smpl.loops[iLoop].end = drwav__bytes_to_u32(smplLoopData+12);\n pWav->smpl.loops[iLoop].fraction = drwav__bytes_to_u32(smplLoopData+16);\n pWav->smpl.loops[iLoop].playCount = drwav__bytes_to_u32(smplLoopData+20);\n } else {\n break; /* Break from the smpl loop for loop. */\n }\n }\n }\n } else {\n /* Looks like invalid data. Ignore the chunk. */\n }\n }\n } else {\n if (drwav__guid_equal(header.id.guid, drwavGUID_W64_SMPL)) {\n /*\n This path will be hit when a W64 WAV file contains a smpl chunk. I don't have a sample file to test this path, so a contribution\n is welcome to add support for this.\n */\n }\n }\n\n /* Make sure we seek past the padding. */\n chunkSize += header.paddingSize;\n if (!drwav__seek_forward(pWav->onSeek, chunkSize, pWav->pUserData)) {\n break;\n }\n cursor += chunkSize;\n\n if (!foundDataChunk) {\n pWav->dataChunkDataPos = cursor;\n }\n }\n\n /* If we haven't found a data chunk, return an error. */\n if (!foundDataChunk) {\n return DRWAV_FALSE;\n }\n\n /* We may have moved passed the data chunk. If so we need to move back. If running in sequential mode we can assume we are already sitting on the data chunk. */\n if (!sequential) {\n if (!drwav__seek_from_start(pWav->onSeek, pWav->dataChunkDataPos, pWav->pUserData)) {\n return DRWAV_FALSE;\n }\n cursor = pWav->dataChunkDataPos;\n }\n \n\n /* At this point we should be sitting on the first byte of the raw audio data. */\n\n pWav->fmt = fmt;\n pWav->sampleRate = fmt.sampleRate;\n pWav->channels = fmt.channels;\n pWav->bitsPerSample = fmt.bitsPerSample;\n pWav->bytesRemaining = dataChunkSize;\n pWav->translatedFormatTag = translatedFormatTag;\n pWav->dataChunkDataSize = dataChunkSize;\n\n if (sampleCountFromFactChunk != 0) {\n pWav->totalPCMFrameCount = sampleCountFromFactChunk;\n } else {\n pWav->totalPCMFrameCount = dataChunkSize / drwav_get_bytes_per_pcm_frame(pWav);\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {\n drwav_uint64 totalBlockHeaderSizeInBytes;\n drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;\n\n /* Make sure any trailing partial block is accounted for. */\n if ((blockCount * fmt.blockAlign) < dataChunkSize) {\n blockCount += 1;\n }\n\n /* We decode two samples per byte. There will be blockCount headers in the data chunk. This is enough to know how to calculate the total PCM frame count. */\n totalBlockHeaderSizeInBytes = blockCount * (6*fmt.channels);\n pWav->totalPCMFrameCount = ((dataChunkSize - totalBlockHeaderSizeInBytes) * 2) / fmt.channels;\n }\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {\n drwav_uint64 totalBlockHeaderSizeInBytes;\n drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;\n\n /* Make sure any trailing partial block is accounted for. */\n if ((blockCount * fmt.blockAlign) < dataChunkSize) {\n blockCount += 1;\n }\n\n /* We decode two samples per byte. There will be blockCount headers in the data chunk. This is enough to know how to calculate the total PCM frame count. */\n totalBlockHeaderSizeInBytes = blockCount * (4*fmt.channels);\n pWav->totalPCMFrameCount = ((dataChunkSize - totalBlockHeaderSizeInBytes) * 2) / fmt.channels;\n\n /* The header includes a decoded sample for each channel which acts as the initial predictor sample. */\n pWav->totalPCMFrameCount += blockCount;\n }\n }\n\n /* Some formats only support a certain number of channels. */\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM || pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {\n if (pWav->channels > 2) {\n return DRWAV_FALSE;\n }\n }\n\n#ifdef DR_WAV_LIBSNDFILE_COMPAT\n /*\n I use libsndfile as a benchmark for testing, however in the version I'm using (from the Windows installer on the libsndfile website),\n it appears the total sample count libsndfile uses for MS-ADPCM is incorrect. It would seem they are computing the total sample count\n from the number of blocks, however this results in the inclusion of extra silent samples at the end of the last block. The correct\n way to know the total sample count is to inspect the \"fact\" chunk, which should always be present for compressed formats, and should\n always include the sample count. This little block of code below is only used to emulate the libsndfile logic so I can properly run my\n correctness tests against libsndfile, and is disabled by default.\n */\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {\n drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;\n pWav->totalPCMFrameCount = (((blockCount * (fmt.blockAlign - (6*pWav->channels))) * 2)) / fmt.channels; /* x2 because two samples per byte. */\n }\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {\n drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;\n pWav->totalPCMFrameCount = (((blockCount * (fmt.blockAlign - (4*pWav->channels))) * 2) + (blockCount * pWav->channels)) / fmt.channels;\n }\n#endif\n\n return DRWAV_TRUE;\n}\n\nDRWAV_API drwav_bool32 drwav_init(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_ex(pWav, onRead, onSeek, NULL, pUserData, NULL, 0, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_ex(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, drwav_chunk_proc onChunk, void* pReadSeekUserData, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (!drwav_preinit(pWav, onRead, onSeek, pReadSeekUserData, pAllocationCallbacks)) {\n return DRWAV_FALSE;\n }\n\n return drwav_init__internal(pWav, onChunk, pChunkUserData, flags);\n}\n\n\nstatic drwav_uint32 drwav__riff_chunk_size_riff(drwav_uint64 dataChunkSize)\n{\n drwav_uint64 chunkSize = 4 + 24 + dataChunkSize + drwav__chunk_padding_size_riff(dataChunkSize); /* 4 = \"WAVE\". 24 = \"fmt \" chunk. */\n if (chunkSize > 0xFFFFFFFFUL) {\n chunkSize = 0xFFFFFFFFUL;\n }\n\n return (drwav_uint32)chunkSize; /* Safe cast due to the clamp above. */\n}\n\nstatic drwav_uint32 drwav__data_chunk_size_riff(drwav_uint64 dataChunkSize)\n{\n if (dataChunkSize <= 0xFFFFFFFFUL) {\n return (drwav_uint32)dataChunkSize;\n } else {\n return 0xFFFFFFFFUL;\n }\n}\n\nstatic drwav_uint64 drwav__riff_chunk_size_w64(drwav_uint64 dataChunkSize)\n{\n drwav_uint64 dataSubchunkPaddingSize = drwav__chunk_padding_size_w64(dataChunkSize);\n\n return 80 + 24 + dataChunkSize + dataSubchunkPaddingSize; /* +24 because W64 includes the size of the GUID and size fields. */\n}\n\nstatic drwav_uint64 drwav__data_chunk_size_w64(drwav_uint64 dataChunkSize)\n{\n return 24 + dataChunkSize; /* +24 because W64 includes the size of the GUID and size fields. */\n}\n\nstatic drwav_uint64 drwav__riff_chunk_size_rf64(drwav_uint64 dataChunkSize)\n{\n drwav_uint64 chunkSize = 4 + 36 + 24 + dataChunkSize + drwav__chunk_padding_size_riff(dataChunkSize); /* 4 = \"WAVE\". 36 = \"ds64\" chunk. 24 = \"fmt \" chunk. */\n if (chunkSize > 0xFFFFFFFFUL) {\n chunkSize = 0xFFFFFFFFUL;\n }\n\n return chunkSize;\n}\n\nstatic drwav_uint64 drwav__data_chunk_size_rf64(drwav_uint64 dataChunkSize)\n{\n return dataChunkSize;\n}\n\n\nstatic size_t drwav__write(drwav* pWav, const void* pData, size_t dataSize)\n{\n DRWAV_ASSERT(pWav != NULL);\n DRWAV_ASSERT(pWav->onWrite != NULL);\n\n /* Generic write. Assumes no byte reordering required. */\n return pWav->onWrite(pWav->pUserData, pData, dataSize);\n}\n\nstatic size_t drwav__write_u16ne_to_le(drwav* pWav, drwav_uint16 value)\n{\n DRWAV_ASSERT(pWav != NULL);\n DRWAV_ASSERT(pWav->onWrite != NULL);\n\n if (!drwav__is_little_endian()) {\n value = drwav__bswap16(value);\n }\n\n return drwav__write(pWav, &value, 2);\n}\n\nstatic size_t drwav__write_u32ne_to_le(drwav* pWav, drwav_uint32 value)\n{\n DRWAV_ASSERT(pWav != NULL);\n DRWAV_ASSERT(pWav->onWrite != NULL);\n\n if (!drwav__is_little_endian()) {\n value = drwav__bswap32(value);\n }\n\n return drwav__write(pWav, &value, 4);\n}\n\nstatic size_t drwav__write_u64ne_to_le(drwav* pWav, drwav_uint64 value)\n{\n DRWAV_ASSERT(pWav != NULL);\n DRWAV_ASSERT(pWav->onWrite != NULL);\n\n if (!drwav__is_little_endian()) {\n value = drwav__bswap64(value);\n }\n\n return drwav__write(pWav, &value, 8);\n}\n\n\nstatic drwav_bool32 drwav_preinit_write(drwav* pWav, const drwav_data_format* pFormat, drwav_bool32 isSequential, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pWav == NULL || onWrite == NULL) {\n return DRWAV_FALSE;\n }\n\n if (!isSequential && onSeek == NULL) {\n return DRWAV_FALSE; /* <-- onSeek is required when in non-sequential mode. */\n }\n\n /* Not currently supporting compressed formats. Will need to add support for the \"fact\" chunk before we enable this. */\n if (pFormat->format == DR_WAVE_FORMAT_EXTENSIBLE) {\n return DRWAV_FALSE;\n }\n if (pFormat->format == DR_WAVE_FORMAT_ADPCM || pFormat->format == DR_WAVE_FORMAT_DVI_ADPCM) {\n return DRWAV_FALSE;\n }\n\n DRWAV_ZERO_MEMORY(pWav, sizeof(*pWav));\n pWav->onWrite = onWrite;\n pWav->onSeek = onSeek;\n pWav->pUserData = pUserData;\n pWav->allocationCallbacks = drwav_copy_allocation_callbacks_or_defaults(pAllocationCallbacks);\n\n if (pWav->allocationCallbacks.onFree == NULL || (pWav->allocationCallbacks.onMalloc == NULL && pWav->allocationCallbacks.onRealloc == NULL)) {\n return DRWAV_FALSE; /* Invalid allocation callbacks. */\n }\n\n pWav->fmt.formatTag = (drwav_uint16)pFormat->format;\n pWav->fmt.channels = (drwav_uint16)pFormat->channels;\n pWav->fmt.sampleRate = pFormat->sampleRate;\n pWav->fmt.avgBytesPerSec = (drwav_uint32)((pFormat->bitsPerSample * pFormat->sampleRate * pFormat->channels) / 8);\n pWav->fmt.blockAlign = (drwav_uint16)((pFormat->channels * pFormat->bitsPerSample) / 8);\n pWav->fmt.bitsPerSample = (drwav_uint16)pFormat->bitsPerSample;\n pWav->fmt.extendedSize = 0;\n pWav->isSequentialWrite = isSequential;\n\n return DRWAV_TRUE;\n}\n\nstatic drwav_bool32 drwav_init_write__internal(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount)\n{\n /* The function assumes drwav_preinit_write() was called beforehand. */\n\n size_t runningPos = 0;\n drwav_uint64 initialDataChunkSize = 0;\n drwav_uint64 chunkSizeFMT;\n\n /*\n The initial values for the \"RIFF\" and \"data\" chunks depends on whether or not we are initializing in sequential mode or not. In\n sequential mode we set this to its final values straight away since they can be calculated from the total sample count. In non-\n sequential mode we initialize it all to zero and fill it out in drwav_uninit() using a backwards seek.\n */\n if (pWav->isSequentialWrite) {\n initialDataChunkSize = (totalSampleCount * pWav->fmt.bitsPerSample) / 8;\n\n /*\n The RIFF container has a limit on the number of samples. drwav is not allowing this. There's no practical limits for Wave64\n so for the sake of simplicity I'm not doing any validation for that.\n */\n if (pFormat->container == drwav_container_riff) {\n if (initialDataChunkSize > (0xFFFFFFFFUL - 36)) {\n return DRWAV_FALSE; /* Not enough room to store every sample. */\n }\n }\n }\n\n pWav->dataChunkDataSizeTargetWrite = initialDataChunkSize;\n\n\n /* \"RIFF\" chunk. */\n if (pFormat->container == drwav_container_riff) {\n drwav_uint32 chunkSizeRIFF = 28 + (drwav_uint32)initialDataChunkSize; /* +28 = \"WAVE\" + [sizeof \"fmt \" chunk] */\n runningPos += drwav__write(pWav, \"RIFF\", 4);\n runningPos += drwav__write_u32ne_to_le(pWav, chunkSizeRIFF);\n runningPos += drwav__write(pWav, \"WAVE\", 4);\n } else if (pFormat->container == drwav_container_w64) {\n drwav_uint64 chunkSizeRIFF = 80 + 24 + initialDataChunkSize; /* +24 because W64 includes the size of the GUID and size fields. */\n runningPos += drwav__write(pWav, drwavGUID_W64_RIFF, 16);\n runningPos += drwav__write_u64ne_to_le(pWav, chunkSizeRIFF);\n runningPos += drwav__write(pWav, drwavGUID_W64_WAVE, 16);\n } else if (pFormat->container == drwav_container_rf64) {\n runningPos += drwav__write(pWav, \"RF64\", 4);\n runningPos += drwav__write_u32ne_to_le(pWav, 0xFFFFFFFF); /* Always 0xFFFFFFFF for RF64. Set to a proper value in the \"ds64\" chunk. */\n runningPos += drwav__write(pWav, \"WAVE\", 4);\n }\n\n \n /* \"ds64\" chunk (RF64 only). */\n if (pFormat->container == drwav_container_rf64) {\n drwav_uint32 initialds64ChunkSize = 28; /* 28 = [Size of RIFF (8 bytes)] + [Size of DATA (8 bytes)] + [Sample Count (8 bytes)] + [Table Length (4 bytes)]. Table length always set to 0. */\n drwav_uint64 initialRiffChunkSize = 8 + initialds64ChunkSize + initialDataChunkSize; /* +8 for the ds64 header. */\n\n runningPos += drwav__write(pWav, \"ds64\", 4);\n runningPos += drwav__write_u32ne_to_le(pWav, initialds64ChunkSize); /* Size of ds64. */\n runningPos += drwav__write_u64ne_to_le(pWav, initialRiffChunkSize); /* Size of RIFF. Set to true value at the end. */\n runningPos += drwav__write_u64ne_to_le(pWav, initialDataChunkSize); /* Size of DATA. Set to true value at the end. */\n runningPos += drwav__write_u64ne_to_le(pWav, totalSampleCount); /* Sample count. */\n runningPos += drwav__write_u32ne_to_le(pWav, 0); /* Table length. Always set to zero in our case since we're not doing any other chunks than \"DATA\". */\n }\n\n\n /* \"fmt \" chunk. */\n if (pFormat->container == drwav_container_riff || pFormat->container == drwav_container_rf64) {\n chunkSizeFMT = 16;\n runningPos += drwav__write(pWav, \"fmt \", 4);\n runningPos += drwav__write_u32ne_to_le(pWav, (drwav_uint32)chunkSizeFMT);\n } else if (pFormat->container == drwav_container_w64) {\n chunkSizeFMT = 40;\n runningPos += drwav__write(pWav, drwavGUID_W64_FMT, 16);\n runningPos += drwav__write_u64ne_to_le(pWav, chunkSizeFMT);\n }\n\n runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.formatTag);\n runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.channels);\n runningPos += drwav__write_u32ne_to_le(pWav, pWav->fmt.sampleRate);\n runningPos += drwav__write_u32ne_to_le(pWav, pWav->fmt.avgBytesPerSec);\n runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.blockAlign);\n runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.bitsPerSample);\n\n pWav->dataChunkDataPos = runningPos;\n\n /* \"data\" chunk. */\n if (pFormat->container == drwav_container_riff) {\n drwav_uint32 chunkSizeDATA = (drwav_uint32)initialDataChunkSize;\n runningPos += drwav__write(pWav, \"data\", 4);\n runningPos += drwav__write_u32ne_to_le(pWav, chunkSizeDATA);\n } else if (pFormat->container == drwav_container_w64) {\n drwav_uint64 chunkSizeDATA = 24 + initialDataChunkSize; /* +24 because W64 includes the size of the GUID and size fields. */\n runningPos += drwav__write(pWav, drwavGUID_W64_DATA, 16);\n runningPos += drwav__write_u64ne_to_le(pWav, chunkSizeDATA);\n } else if (pFormat->container == drwav_container_rf64) {\n runningPos += drwav__write(pWav, \"data\", 4);\n runningPos += drwav__write_u32ne_to_le(pWav, 0xFFFFFFFF); /* Always set to 0xFFFFFFFF for RF64. The true size of the data chunk is specified in the ds64 chunk. */\n }\n\n /*\n The runningPos variable is incremented in the section above but is left unused which is causing some static analysis tools to detect it\n as a dead store. I'm leaving this as-is for safety just in case I want to expand this function later to include other tags and want to\n keep track of the running position for whatever reason. The line below should silence the static analysis tools.\n */\n (void)runningPos;\n\n /* Set some properties for the client's convenience. */\n pWav->container = pFormat->container;\n pWav->channels = (drwav_uint16)pFormat->channels;\n pWav->sampleRate = pFormat->sampleRate;\n pWav->bitsPerSample = (drwav_uint16)pFormat->bitsPerSample;\n pWav->translatedFormatTag = (drwav_uint16)pFormat->format;\n\n return DRWAV_TRUE;\n}\n\n\nDRWAV_API drwav_bool32 drwav_init_write(drwav* pWav, const drwav_data_format* pFormat, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (!drwav_preinit_write(pWav, pFormat, DRWAV_FALSE, onWrite, onSeek, pUserData, pAllocationCallbacks)) {\n return DRWAV_FALSE;\n }\n\n return drwav_init_write__internal(pWav, pFormat, 0); /* DRWAV_FALSE = Not Sequential */\n}\n\nDRWAV_API drwav_bool32 drwav_init_write_sequential(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (!drwav_preinit_write(pWav, pFormat, DRWAV_TRUE, onWrite, NULL, pUserData, pAllocationCallbacks)) {\n return DRWAV_FALSE;\n }\n\n return drwav_init_write__internal(pWav, pFormat, totalSampleCount); /* DRWAV_TRUE = Sequential */\n}\n\nDRWAV_API drwav_bool32 drwav_init_write_sequential_pcm_frames(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pFormat == NULL) {\n return DRWAV_FALSE;\n }\n\n return drwav_init_write_sequential(pWav, pFormat, totalPCMFrameCount*pFormat->channels, onWrite, pUserData, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_uint64 drwav_target_write_size_bytes(const drwav_data_format* pFormat, drwav_uint64 totalSampleCount)\n{\n /* Casting totalSampleCount to drwav_int64 for VC6 compatibility. No issues in practice because nobody is going to exhaust the whole 63 bits. */\n drwav_uint64 targetDataSizeBytes = (drwav_uint64)((drwav_int64)totalSampleCount * pFormat->channels * pFormat->bitsPerSample/8.0);\n drwav_uint64 riffChunkSizeBytes;\n drwav_uint64 fileSizeBytes = 0;\n\n if (pFormat->container == drwav_container_riff) {\n riffChunkSizeBytes = drwav__riff_chunk_size_riff(targetDataSizeBytes);\n fileSizeBytes = (8 + riffChunkSizeBytes); /* +8 because WAV doesn't include the size of the ChunkID and ChunkSize fields. */\n } else if (pFormat->container == drwav_container_w64) {\n riffChunkSizeBytes = drwav__riff_chunk_size_w64(targetDataSizeBytes);\n fileSizeBytes = riffChunkSizeBytes;\n } else if (pFormat->container == drwav_container_rf64) {\n riffChunkSizeBytes = drwav__riff_chunk_size_rf64(targetDataSizeBytes);\n fileSizeBytes = (8 + riffChunkSizeBytes); /* +8 because WAV doesn't include the size of the ChunkID and ChunkSize fields. */\n }\n\n return fileSizeBytes;\n}\n\n\n#ifndef DR_WAV_NO_STDIO\n\n/* drwav_result_from_errno() is only used for fopen() and wfopen() so putting it inside DR_WAV_NO_STDIO for now. If something else needs this later we can move it out. */\n#include \nstatic drwav_result drwav_result_from_errno(int e)\n{\n switch (e)\n {\n case 0: return DRWAV_SUCCESS;\n #ifdef EPERM\n case EPERM: return DRWAV_INVALID_OPERATION;\n #endif\n #ifdef ENOENT\n case ENOENT: return DRWAV_DOES_NOT_EXIST;\n #endif\n #ifdef ESRCH\n case ESRCH: return DRWAV_DOES_NOT_EXIST;\n #endif\n #ifdef EINTR\n case EINTR: return DRWAV_INTERRUPT;\n #endif\n #ifdef EIO\n case EIO: return DRWAV_IO_ERROR;\n #endif\n #ifdef ENXIO\n case ENXIO: return DRWAV_DOES_NOT_EXIST;\n #endif\n #ifdef E2BIG\n case E2BIG: return DRWAV_INVALID_ARGS;\n #endif\n #ifdef ENOEXEC\n case ENOEXEC: return DRWAV_INVALID_FILE;\n #endif\n #ifdef EBADF\n case EBADF: return DRWAV_INVALID_FILE;\n #endif\n #ifdef ECHILD\n case ECHILD: return DRWAV_ERROR;\n #endif\n #ifdef EAGAIN\n case EAGAIN: return DRWAV_UNAVAILABLE;\n #endif\n #ifdef ENOMEM\n case ENOMEM: return DRWAV_OUT_OF_MEMORY;\n #endif\n #ifdef EACCES\n case EACCES: return DRWAV_ACCESS_DENIED;\n #endif\n #ifdef EFAULT\n case EFAULT: return DRWAV_BAD_ADDRESS;\n #endif\n #ifdef ENOTBLK\n case ENOTBLK: return DRWAV_ERROR;\n #endif\n #ifdef EBUSY\n case EBUSY: return DRWAV_BUSY;\n #endif\n #ifdef EEXIST\n case EEXIST: return DRWAV_ALREADY_EXISTS;\n #endif\n #ifdef EXDEV\n case EXDEV: return DRWAV_ERROR;\n #endif\n #ifdef ENODEV\n case ENODEV: return DRWAV_DOES_NOT_EXIST;\n #endif\n #ifdef ENOTDIR\n case ENOTDIR: return DRWAV_NOT_DIRECTORY;\n #endif\n #ifdef EISDIR\n case EISDIR: return DRWAV_IS_DIRECTORY;\n #endif\n #ifdef EINVAL\n case EINVAL: return DRWAV_INVALID_ARGS;\n #endif\n #ifdef ENFILE\n case ENFILE: return DRWAV_TOO_MANY_OPEN_FILES;\n #endif\n #ifdef EMFILE\n case EMFILE: return DRWAV_TOO_MANY_OPEN_FILES;\n #endif\n #ifdef ENOTTY\n case ENOTTY: return DRWAV_INVALID_OPERATION;\n #endif\n #ifdef ETXTBSY\n case ETXTBSY: return DRWAV_BUSY;\n #endif\n #ifdef EFBIG\n case EFBIG: return DRWAV_TOO_BIG;\n #endif\n #ifdef ENOSPC\n case ENOSPC: return DRWAV_NO_SPACE;\n #endif\n #ifdef ESPIPE\n case ESPIPE: return DRWAV_BAD_SEEK;\n #endif\n #ifdef EROFS\n case EROFS: return DRWAV_ACCESS_DENIED;\n #endif\n #ifdef EMLINK\n case EMLINK: return DRWAV_TOO_MANY_LINKS;\n #endif\n #ifdef EPIPE\n case EPIPE: return DRWAV_BAD_PIPE;\n #endif\n #ifdef EDOM\n case EDOM: return DRWAV_OUT_OF_RANGE;\n #endif\n #ifdef ERANGE\n case ERANGE: return DRWAV_OUT_OF_RANGE;\n #endif\n #ifdef EDEADLK\n case EDEADLK: return DRWAV_DEADLOCK;\n #endif\n #ifdef ENAMETOOLONG\n case ENAMETOOLONG: return DRWAV_PATH_TOO_LONG;\n #endif\n #ifdef ENOLCK\n case ENOLCK: return DRWAV_ERROR;\n #endif\n #ifdef ENOSYS\n case ENOSYS: return DRWAV_NOT_IMPLEMENTED;\n #endif\n #ifdef ENOTEMPTY\n case ENOTEMPTY: return DRWAV_DIRECTORY_NOT_EMPTY;\n #endif\n #ifdef ELOOP\n case ELOOP: return DRWAV_TOO_MANY_LINKS;\n #endif\n #ifdef ENOMSG\n case ENOMSG: return DRWAV_NO_MESSAGE;\n #endif\n #ifdef EIDRM\n case EIDRM: return DRWAV_ERROR;\n #endif\n #ifdef ECHRNG\n case ECHRNG: return DRWAV_ERROR;\n #endif\n #ifdef EL2NSYNC\n case EL2NSYNC: return DRWAV_ERROR;\n #endif\n #ifdef EL3HLT\n case EL3HLT: return DRWAV_ERROR;\n #endif\n #ifdef EL3RST\n case EL3RST: return DRWAV_ERROR;\n #endif\n #ifdef ELNRNG\n case ELNRNG: return DRWAV_OUT_OF_RANGE;\n #endif\n #ifdef EUNATCH\n case EUNATCH: return DRWAV_ERROR;\n #endif\n #ifdef ENOCSI\n case ENOCSI: return DRWAV_ERROR;\n #endif\n #ifdef EL2HLT\n case EL2HLT: return DRWAV_ERROR;\n #endif\n #ifdef EBADE\n case EBADE: return DRWAV_ERROR;\n #endif\n #ifdef EBADR\n case EBADR: return DRWAV_ERROR;\n #endif\n #ifdef EXFULL\n case EXFULL: return DRWAV_ERROR;\n #endif\n #ifdef ENOANO\n case ENOANO: return DRWAV_ERROR;\n #endif\n #ifdef EBADRQC\n case EBADRQC: return DRWAV_ERROR;\n #endif\n #ifdef EBADSLT\n case EBADSLT: return DRWAV_ERROR;\n #endif\n #ifdef EBFONT\n case EBFONT: return DRWAV_INVALID_FILE;\n #endif\n #ifdef ENOSTR\n case ENOSTR: return DRWAV_ERROR;\n #endif\n #ifdef ENODATA\n case ENODATA: return DRWAV_NO_DATA_AVAILABLE;\n #endif\n #ifdef ETIME\n case ETIME: return DRWAV_TIMEOUT;\n #endif\n #ifdef ENOSR\n case ENOSR: return DRWAV_NO_DATA_AVAILABLE;\n #endif\n #ifdef ENONET\n case ENONET: return DRWAV_NO_NETWORK;\n #endif\n #ifdef ENOPKG\n case ENOPKG: return DRWAV_ERROR;\n #endif\n #ifdef EREMOTE\n case EREMOTE: return DRWAV_ERROR;\n #endif\n #ifdef ENOLINK\n case ENOLINK: return DRWAV_ERROR;\n #endif\n #ifdef EADV\n case EADV: return DRWAV_ERROR;\n #endif\n #ifdef ESRMNT\n case ESRMNT: return DRWAV_ERROR;\n #endif\n #ifdef ECOMM\n case ECOMM: return DRWAV_ERROR;\n #endif\n #ifdef EPROTO\n case EPROTO: return DRWAV_ERROR;\n #endif\n #ifdef EMULTIHOP\n case EMULTIHOP: return DRWAV_ERROR;\n #endif\n #ifdef EDOTDOT\n case EDOTDOT: return DRWAV_ERROR;\n #endif\n #ifdef EBADMSG\n case EBADMSG: return DRWAV_BAD_MESSAGE;\n #endif\n #ifdef EOVERFLOW\n case EOVERFLOW: return DRWAV_TOO_BIG;\n #endif\n #ifdef ENOTUNIQ\n case ENOTUNIQ: return DRWAV_NOT_UNIQUE;\n #endif\n #ifdef EBADFD\n case EBADFD: return DRWAV_ERROR;\n #endif\n #ifdef EREMCHG\n case EREMCHG: return DRWAV_ERROR;\n #endif\n #ifdef ELIBACC\n case ELIBACC: return DRWAV_ACCESS_DENIED;\n #endif\n #ifdef ELIBBAD\n case ELIBBAD: return DRWAV_INVALID_FILE;\n #endif\n #ifdef ELIBSCN\n case ELIBSCN: return DRWAV_INVALID_FILE;\n #endif\n #ifdef ELIBMAX\n case ELIBMAX: return DRWAV_ERROR;\n #endif\n #ifdef ELIBEXEC\n case ELIBEXEC: return DRWAV_ERROR;\n #endif\n #ifdef EILSEQ\n case EILSEQ: return DRWAV_INVALID_DATA;\n #endif\n #ifdef ERESTART\n case ERESTART: return DRWAV_ERROR;\n #endif\n #ifdef ESTRPIPE\n case ESTRPIPE: return DRWAV_ERROR;\n #endif\n #ifdef EUSERS\n case EUSERS: return DRWAV_ERROR;\n #endif\n #ifdef ENOTSOCK\n case ENOTSOCK: return DRWAV_NOT_SOCKET;\n #endif\n #ifdef EDESTADDRREQ\n case EDESTADDRREQ: return DRWAV_NO_ADDRESS;\n #endif\n #ifdef EMSGSIZE\n case EMSGSIZE: return DRWAV_TOO_BIG;\n #endif\n #ifdef EPROTOTYPE\n case EPROTOTYPE: return DRWAV_BAD_PROTOCOL;\n #endif\n #ifdef ENOPROTOOPT\n case ENOPROTOOPT: return DRWAV_PROTOCOL_UNAVAILABLE;\n #endif\n #ifdef EPROTONOSUPPORT\n case EPROTONOSUPPORT: return DRWAV_PROTOCOL_NOT_SUPPORTED;\n #endif\n #ifdef ESOCKTNOSUPPORT\n case ESOCKTNOSUPPORT: return DRWAV_SOCKET_NOT_SUPPORTED;\n #endif\n #ifdef EOPNOTSUPP\n case EOPNOTSUPP: return DRWAV_INVALID_OPERATION;\n #endif\n #ifdef EPFNOSUPPORT\n case EPFNOSUPPORT: return DRWAV_PROTOCOL_FAMILY_NOT_SUPPORTED;\n #endif\n #ifdef EAFNOSUPPORT\n case EAFNOSUPPORT: return DRWAV_ADDRESS_FAMILY_NOT_SUPPORTED;\n #endif\n #ifdef EADDRINUSE\n case EADDRINUSE: return DRWAV_ALREADY_IN_USE;\n #endif\n #ifdef EADDRNOTAVAIL\n case EADDRNOTAVAIL: return DRWAV_ERROR;\n #endif\n #ifdef ENETDOWN\n case ENETDOWN: return DRWAV_NO_NETWORK;\n #endif\n #ifdef ENETUNREACH\n case ENETUNREACH: return DRWAV_NO_NETWORK;\n #endif\n #ifdef ENETRESET\n case ENETRESET: return DRWAV_NO_NETWORK;\n #endif\n #ifdef ECONNABORTED\n case ECONNABORTED: return DRWAV_NO_NETWORK;\n #endif\n #ifdef ECONNRESET\n case ECONNRESET: return DRWAV_CONNECTION_RESET;\n #endif\n #ifdef ENOBUFS\n case ENOBUFS: return DRWAV_NO_SPACE;\n #endif\n #ifdef EISCONN\n case EISCONN: return DRWAV_ALREADY_CONNECTED;\n #endif\n #ifdef ENOTCONN\n case ENOTCONN: return DRWAV_NOT_CONNECTED;\n #endif\n #ifdef ESHUTDOWN\n case ESHUTDOWN: return DRWAV_ERROR;\n #endif\n #ifdef ETOOMANYREFS\n case ETOOMANYREFS: return DRWAV_ERROR;\n #endif\n #ifdef ETIMEDOUT\n case ETIMEDOUT: return DRWAV_TIMEOUT;\n #endif\n #ifdef ECONNREFUSED\n case ECONNREFUSED: return DRWAV_CONNECTION_REFUSED;\n #endif\n #ifdef EHOSTDOWN\n case EHOSTDOWN: return DRWAV_NO_HOST;\n #endif\n #ifdef EHOSTUNREACH\n case EHOSTUNREACH: return DRWAV_NO_HOST;\n #endif\n #ifdef EALREADY\n case EALREADY: return DRWAV_IN_PROGRESS;\n #endif\n #ifdef EINPROGRESS\n case EINPROGRESS: return DRWAV_IN_PROGRESS;\n #endif\n #ifdef ESTALE\n case ESTALE: return DRWAV_INVALID_FILE;\n #endif\n #ifdef EUCLEAN\n case EUCLEAN: return DRWAV_ERROR;\n #endif\n #ifdef ENOTNAM\n case ENOTNAM: return DRWAV_ERROR;\n #endif\n #ifdef ENAVAIL\n case ENAVAIL: return DRWAV_ERROR;\n #endif\n #ifdef EISNAM\n case EISNAM: return DRWAV_ERROR;\n #endif\n #ifdef EREMOTEIO\n case EREMOTEIO: return DRWAV_IO_ERROR;\n #endif\n #ifdef EDQUOT\n case EDQUOT: return DRWAV_NO_SPACE;\n #endif\n #ifdef ENOMEDIUM\n case ENOMEDIUM: return DRWAV_DOES_NOT_EXIST;\n #endif\n #ifdef EMEDIUMTYPE\n case EMEDIUMTYPE: return DRWAV_ERROR;\n #endif\n #ifdef ECANCELED\n case ECANCELED: return DRWAV_CANCELLED;\n #endif\n #ifdef ENOKEY\n case ENOKEY: return DRWAV_ERROR;\n #endif\n #ifdef EKEYEXPIRED\n case EKEYEXPIRED: return DRWAV_ERROR;\n #endif\n #ifdef EKEYREVOKED\n case EKEYREVOKED: return DRWAV_ERROR;\n #endif\n #ifdef EKEYREJECTED\n case EKEYREJECTED: return DRWAV_ERROR;\n #endif\n #ifdef EOWNERDEAD\n case EOWNERDEAD: return DRWAV_ERROR;\n #endif\n #ifdef ENOTRECOVERABLE\n case ENOTRECOVERABLE: return DRWAV_ERROR;\n #endif\n #ifdef ERFKILL\n case ERFKILL: return DRWAV_ERROR;\n #endif\n #ifdef EHWPOISON\n case EHWPOISON: return DRWAV_ERROR;\n #endif\n default: return DRWAV_ERROR;\n }\n}\n\nstatic drwav_result drwav_fopen(FILE** ppFile, const char* pFilePath, const char* pOpenMode)\n{\n#if _MSC_VER && _MSC_VER >= 1400\n errno_t err;\n#endif\n\n if (ppFile != NULL) {\n *ppFile = NULL; /* Safety. */\n }\n\n if (pFilePath == NULL || pOpenMode == NULL || ppFile == NULL) {\n return DRWAV_INVALID_ARGS;\n }\n\n#if _MSC_VER && _MSC_VER >= 1400\n err = fopen_s(ppFile, pFilePath, pOpenMode);\n if (err != 0) {\n return drwav_result_from_errno(err);\n }\n#else\n#if defined(_WIN32) || defined(__APPLE__)\n *ppFile = fopen(pFilePath, pOpenMode);\n#else\n #if defined(_FILE_OFFSET_BITS) && _FILE_OFFSET_BITS == 64 && defined(_LARGEFILE64_SOURCE)\n *ppFile = fopen64(pFilePath, pOpenMode);\n #else\n *ppFile = fopen(pFilePath, pOpenMode);\n #endif\n#endif\n if (*ppFile == NULL) {\n drwav_result result = drwav_result_from_errno(errno);\n if (result == DRWAV_SUCCESS) {\n result = DRWAV_ERROR; /* Just a safety check to make sure we never ever return success when pFile == NULL. */\n }\n\n return result;\n }\n#endif\n\n return DRWAV_SUCCESS;\n}\n\n/*\n_wfopen() isn't always available in all compilation environments.\n\n * Windows only.\n * MSVC seems to support it universally as far back as VC6 from what I can tell (haven't checked further back).\n * MinGW-64 (both 32- and 64-bit) seems to support it.\n * MinGW wraps it in !defined(__STRICT_ANSI__).\n * OpenWatcom wraps it in !defined(_NO_EXT_KEYS).\n\nThis can be reviewed as compatibility issues arise. The preference is to use _wfopen_s() and _wfopen() as opposed to the wcsrtombs()\nfallback, so if you notice your compiler not detecting this properly I'm happy to look at adding support.\n*/\n#if defined(_WIN32)\n #if defined(_MSC_VER) || defined(__MINGW64__) || (!defined(__STRICT_ANSI__) && !defined(_NO_EXT_KEYS))\n #define DRWAV_HAS_WFOPEN\n #endif\n#endif\n\nstatic drwav_result drwav_wfopen(FILE** ppFile, const wchar_t* pFilePath, const wchar_t* pOpenMode, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (ppFile != NULL) {\n *ppFile = NULL; /* Safety. */\n }\n\n if (pFilePath == NULL || pOpenMode == NULL || ppFile == NULL) {\n return DRWAV_INVALID_ARGS;\n }\n\n#if defined(DRWAV_HAS_WFOPEN)\n {\n /* Use _wfopen() on Windows. */\n #if defined(_MSC_VER) && _MSC_VER >= 1400\n errno_t err = _wfopen_s(ppFile, pFilePath, pOpenMode);\n if (err != 0) {\n return drwav_result_from_errno(err);\n }\n #else\n *ppFile = _wfopen(pFilePath, pOpenMode);\n if (*ppFile == NULL) {\n return drwav_result_from_errno(errno);\n }\n #endif\n (void)pAllocationCallbacks;\n }\n#else\n /*\n Use fopen() on anything other than Windows. Requires a conversion. This is annoying because fopen() is locale specific. The only real way I can\n think of to do this is with wcsrtombs(). Note that wcstombs() is apparently not thread-safe because it uses a static global mbstate_t object for\n maintaining state. I've checked this with -std=c89 and it works, but if somebody get's a compiler error I'll look into improving compatibility.\n */\n {\n mbstate_t mbs;\n size_t lenMB;\n const wchar_t* pFilePathTemp = pFilePath;\n char* pFilePathMB = NULL;\n char pOpenModeMB[32] = {0};\n\n /* Get the length first. */\n DRWAV_ZERO_OBJECT(&mbs);\n lenMB = wcsrtombs(NULL, &pFilePathTemp, 0, &mbs);\n if (lenMB == (size_t)-1) {\n return drwav_result_from_errno(errno);\n }\n\n pFilePathMB = (char*)drwav__malloc_from_callbacks(lenMB + 1, pAllocationCallbacks);\n if (pFilePathMB == NULL) {\n return DRWAV_OUT_OF_MEMORY;\n }\n\n pFilePathTemp = pFilePath;\n DRWAV_ZERO_OBJECT(&mbs);\n wcsrtombs(pFilePathMB, &pFilePathTemp, lenMB + 1, &mbs);\n\n /* The open mode should always consist of ASCII characters so we should be able to do a trivial conversion. */\n {\n size_t i = 0;\n for (;;) {\n if (pOpenMode[i] == 0) {\n pOpenModeMB[i] = '\\0';\n break;\n }\n\n pOpenModeMB[i] = (char)pOpenMode[i];\n i += 1;\n }\n }\n\n *ppFile = fopen(pFilePathMB, pOpenModeMB);\n\n drwav__free_from_callbacks(pFilePathMB, pAllocationCallbacks);\n }\n\n if (*ppFile == NULL) {\n return DRWAV_ERROR;\n }\n#endif\n\n return DRWAV_SUCCESS;\n}\n\n\nstatic size_t drwav__on_read_stdio(void* pUserData, void* pBufferOut, size_t bytesToRead)\n{\n return fread(pBufferOut, 1, bytesToRead, (FILE*)pUserData);\n}\n\nstatic size_t drwav__on_write_stdio(void* pUserData, const void* pData, size_t bytesToWrite)\n{\n return fwrite(pData, 1, bytesToWrite, (FILE*)pUserData);\n}\n\nstatic drwav_bool32 drwav__on_seek_stdio(void* pUserData, int offset, drwav_seek_origin origin)\n{\n return fseek((FILE*)pUserData, offset, (origin == drwav_seek_origin_current) ? SEEK_CUR : SEEK_SET) == 0;\n}\n\nDRWAV_API drwav_bool32 drwav_init_file(drwav* pWav, const char* filename, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_file_ex(pWav, filename, NULL, NULL, 0, pAllocationCallbacks);\n}\n\n\nstatic drwav_bool32 drwav_init_file__internal_FILE(drwav* pWav, FILE* pFile, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav_bool32 result;\n\n result = drwav_preinit(pWav, drwav__on_read_stdio, drwav__on_seek_stdio, (void*)pFile, pAllocationCallbacks);\n if (result != DRWAV_TRUE) {\n fclose(pFile);\n return result;\n }\n\n result = drwav_init__internal(pWav, onChunk, pChunkUserData, flags);\n if (result != DRWAV_TRUE) {\n fclose(pFile);\n return result;\n }\n\n return DRWAV_TRUE;\n}\n\nDRWAV_API drwav_bool32 drwav_init_file_ex(drwav* pWav, const char* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n FILE* pFile;\n if (drwav_fopen(&pFile, filename, \"rb\") != DRWAV_SUCCESS) {\n return DRWAV_FALSE;\n }\n\n /* This takes ownership of the FILE* object. */\n return drwav_init_file__internal_FILE(pWav, pFile, onChunk, pChunkUserData, flags, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_file_w(drwav* pWav, const wchar_t* filename, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_file_ex_w(pWav, filename, NULL, NULL, 0, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_file_ex_w(drwav* pWav, const wchar_t* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n FILE* pFile;\n if (drwav_wfopen(&pFile, filename, L\"rb\", pAllocationCallbacks) != DRWAV_SUCCESS) {\n return DRWAV_FALSE;\n }\n\n /* This takes ownership of the FILE* object. */\n return drwav_init_file__internal_FILE(pWav, pFile, onChunk, pChunkUserData, flags, pAllocationCallbacks);\n}\n\n\nstatic drwav_bool32 drwav_init_file_write__internal_FILE(drwav* pWav, FILE* pFile, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav_bool32 result;\n\n result = drwav_preinit_write(pWav, pFormat, isSequential, drwav__on_write_stdio, drwav__on_seek_stdio, (void*)pFile, pAllocationCallbacks);\n if (result != DRWAV_TRUE) {\n fclose(pFile);\n return result;\n }\n\n result = drwav_init_write__internal(pWav, pFormat, totalSampleCount);\n if (result != DRWAV_TRUE) {\n fclose(pFile);\n return result;\n }\n\n return DRWAV_TRUE;\n}\n\nstatic drwav_bool32 drwav_init_file_write__internal(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n FILE* pFile;\n if (drwav_fopen(&pFile, filename, \"wb\") != DRWAV_SUCCESS) {\n return DRWAV_FALSE;\n }\n\n /* This takes ownership of the FILE* object. */\n return drwav_init_file_write__internal_FILE(pWav, pFile, pFormat, totalSampleCount, isSequential, pAllocationCallbacks);\n}\n\nstatic drwav_bool32 drwav_init_file_write_w__internal(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n FILE* pFile;\n if (drwav_wfopen(&pFile, filename, L\"wb\", pAllocationCallbacks) != DRWAV_SUCCESS) {\n return DRWAV_FALSE;\n }\n\n /* This takes ownership of the FILE* object. */\n return drwav_init_file_write__internal_FILE(pWav, pFile, pFormat, totalSampleCount, isSequential, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_file_write(drwav* pWav, const char* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_file_write__internal(pWav, filename, pFormat, 0, DRWAV_FALSE, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_file_write_sequential(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_file_write__internal(pWav, filename, pFormat, totalSampleCount, DRWAV_TRUE, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pFormat == NULL) {\n return DRWAV_FALSE;\n }\n\n return drwav_init_file_write_sequential(pWav, filename, pFormat, totalPCMFrameCount*pFormat->channels, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_file_write_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_file_write_w__internal(pWav, filename, pFormat, 0, DRWAV_FALSE, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_file_write_sequential_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_file_write_w__internal(pWav, filename, pFormat, totalSampleCount, DRWAV_TRUE, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pFormat == NULL) {\n return DRWAV_FALSE;\n }\n\n return drwav_init_file_write_sequential_w(pWav, filename, pFormat, totalPCMFrameCount*pFormat->channels, pAllocationCallbacks);\n}\n#endif /* DR_WAV_NO_STDIO */\n\n\nstatic size_t drwav__on_read_memory(void* pUserData, void* pBufferOut, size_t bytesToRead)\n{\n drwav* pWav = (drwav*)pUserData;\n size_t bytesRemaining;\n\n DRWAV_ASSERT(pWav != NULL);\n DRWAV_ASSERT(pWav->memoryStream.dataSize >= pWav->memoryStream.currentReadPos);\n\n bytesRemaining = pWav->memoryStream.dataSize - pWav->memoryStream.currentReadPos;\n if (bytesToRead > bytesRemaining) {\n bytesToRead = bytesRemaining;\n }\n\n if (bytesToRead > 0) {\n DRWAV_COPY_MEMORY(pBufferOut, pWav->memoryStream.data + pWav->memoryStream.currentReadPos, bytesToRead);\n pWav->memoryStream.currentReadPos += bytesToRead;\n }\n\n return bytesToRead;\n}\n\nstatic drwav_bool32 drwav__on_seek_memory(void* pUserData, int offset, drwav_seek_origin origin)\n{\n drwav* pWav = (drwav*)pUserData;\n DRWAV_ASSERT(pWav != NULL);\n\n if (origin == drwav_seek_origin_current) {\n if (offset > 0) {\n if (pWav->memoryStream.currentReadPos + offset > pWav->memoryStream.dataSize) {\n return DRWAV_FALSE; /* Trying to seek too far forward. */\n }\n } else {\n if (pWav->memoryStream.currentReadPos < (size_t)-offset) {\n return DRWAV_FALSE; /* Trying to seek too far backwards. */\n }\n }\n\n /* This will never underflow thanks to the clamps above. */\n pWav->memoryStream.currentReadPos += offset;\n } else {\n if ((drwav_uint32)offset <= pWav->memoryStream.dataSize) {\n pWav->memoryStream.currentReadPos = offset;\n } else {\n return DRWAV_FALSE; /* Trying to seek too far forward. */\n }\n }\n \n return DRWAV_TRUE;\n}\n\nstatic size_t drwav__on_write_memory(void* pUserData, const void* pDataIn, size_t bytesToWrite)\n{\n drwav* pWav = (drwav*)pUserData;\n size_t bytesRemaining;\n\n DRWAV_ASSERT(pWav != NULL);\n DRWAV_ASSERT(pWav->memoryStreamWrite.dataCapacity >= pWav->memoryStreamWrite.currentWritePos);\n\n bytesRemaining = pWav->memoryStreamWrite.dataCapacity - pWav->memoryStreamWrite.currentWritePos;\n if (bytesRemaining < bytesToWrite) {\n /* Need to reallocate. */\n void* pNewData;\n size_t newDataCapacity = (pWav->memoryStreamWrite.dataCapacity == 0) ? 256 : pWav->memoryStreamWrite.dataCapacity * 2;\n\n /* If doubling wasn't enough, just make it the minimum required size to write the data. */\n if ((newDataCapacity - pWav->memoryStreamWrite.currentWritePos) < bytesToWrite) {\n newDataCapacity = pWav->memoryStreamWrite.currentWritePos + bytesToWrite;\n }\n\n pNewData = drwav__realloc_from_callbacks(*pWav->memoryStreamWrite.ppData, newDataCapacity, pWav->memoryStreamWrite.dataCapacity, &pWav->allocationCallbacks);\n if (pNewData == NULL) {\n return 0;\n }\n\n *pWav->memoryStreamWrite.ppData = pNewData;\n pWav->memoryStreamWrite.dataCapacity = newDataCapacity;\n }\n\n DRWAV_COPY_MEMORY(((drwav_uint8*)(*pWav->memoryStreamWrite.ppData)) + pWav->memoryStreamWrite.currentWritePos, pDataIn, bytesToWrite);\n\n pWav->memoryStreamWrite.currentWritePos += bytesToWrite;\n if (pWav->memoryStreamWrite.dataSize < pWav->memoryStreamWrite.currentWritePos) {\n pWav->memoryStreamWrite.dataSize = pWav->memoryStreamWrite.currentWritePos;\n }\n\n *pWav->memoryStreamWrite.pDataSize = pWav->memoryStreamWrite.dataSize;\n\n return bytesToWrite;\n}\n\nstatic drwav_bool32 drwav__on_seek_memory_write(void* pUserData, int offset, drwav_seek_origin origin)\n{\n drwav* pWav = (drwav*)pUserData;\n DRWAV_ASSERT(pWav != NULL);\n\n if (origin == drwav_seek_origin_current) {\n if (offset > 0) {\n if (pWav->memoryStreamWrite.currentWritePos + offset > pWav->memoryStreamWrite.dataSize) {\n offset = (int)(pWav->memoryStreamWrite.dataSize - pWav->memoryStreamWrite.currentWritePos); /* Trying to seek too far forward. */\n }\n } else {\n if (pWav->memoryStreamWrite.currentWritePos < (size_t)-offset) {\n offset = -(int)pWav->memoryStreamWrite.currentWritePos; /* Trying to seek too far backwards. */\n }\n }\n\n /* This will never underflow thanks to the clamps above. */\n pWav->memoryStreamWrite.currentWritePos += offset;\n } else {\n if ((drwav_uint32)offset <= pWav->memoryStreamWrite.dataSize) {\n pWav->memoryStreamWrite.currentWritePos = offset;\n } else {\n pWav->memoryStreamWrite.currentWritePos = pWav->memoryStreamWrite.dataSize; /* Trying to seek too far forward. */\n }\n }\n \n return DRWAV_TRUE;\n}\n\nDRWAV_API drwav_bool32 drwav_init_memory(drwav* pWav, const void* data, size_t dataSize, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_memory_ex(pWav, data, dataSize, NULL, NULL, 0, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_memory_ex(drwav* pWav, const void* data, size_t dataSize, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (data == NULL || dataSize == 0) {\n return DRWAV_FALSE;\n }\n\n if (!drwav_preinit(pWav, drwav__on_read_memory, drwav__on_seek_memory, pWav, pAllocationCallbacks)) {\n return DRWAV_FALSE;\n }\n\n pWav->memoryStream.data = (const drwav_uint8*)data;\n pWav->memoryStream.dataSize = dataSize;\n pWav->memoryStream.currentReadPos = 0;\n\n return drwav_init__internal(pWav, onChunk, pChunkUserData, flags);\n}\n\n\nstatic drwav_bool32 drwav_init_memory_write__internal(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (ppData == NULL || pDataSize == NULL) {\n return DRWAV_FALSE;\n }\n\n *ppData = NULL; /* Important because we're using realloc()! */\n *pDataSize = 0;\n\n if (!drwav_preinit_write(pWav, pFormat, isSequential, drwav__on_write_memory, drwav__on_seek_memory_write, pWav, pAllocationCallbacks)) {\n return DRWAV_FALSE;\n }\n\n pWav->memoryStreamWrite.ppData = ppData;\n pWav->memoryStreamWrite.pDataSize = pDataSize;\n pWav->memoryStreamWrite.dataSize = 0;\n pWav->memoryStreamWrite.dataCapacity = 0;\n pWav->memoryStreamWrite.currentWritePos = 0;\n\n return drwav_init_write__internal(pWav, pFormat, totalSampleCount);\n}\n\nDRWAV_API drwav_bool32 drwav_init_memory_write(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_memory_write__internal(pWav, ppData, pDataSize, pFormat, 0, DRWAV_FALSE, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_memory_write_sequential(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n return drwav_init_memory_write__internal(pWav, ppData, pDataSize, pFormat, totalSampleCount, DRWAV_TRUE, pAllocationCallbacks);\n}\n\nDRWAV_API drwav_bool32 drwav_init_memory_write_sequential_pcm_frames(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pFormat == NULL) {\n return DRWAV_FALSE;\n }\n\n return drwav_init_memory_write_sequential(pWav, ppData, pDataSize, pFormat, totalPCMFrameCount*pFormat->channels, pAllocationCallbacks);\n}\n\n\n\nDRWAV_API drwav_result drwav_uninit(drwav* pWav)\n{\n drwav_result result = DRWAV_SUCCESS;\n\n if (pWav == NULL) {\n return DRWAV_INVALID_ARGS;\n }\n\n /*\n If the drwav object was opened in write mode we'll need to finalize a few things:\n - Make sure the \"data\" chunk is aligned to 16-bits for RIFF containers, or 64 bits for W64 containers.\n - Set the size of the \"data\" chunk.\n */\n if (pWav->onWrite != NULL) {\n drwav_uint32 paddingSize = 0;\n\n /* Padding. Do not adjust pWav->dataChunkDataSize - this should not include the padding. */\n if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rf64) {\n paddingSize = drwav__chunk_padding_size_riff(pWav->dataChunkDataSize);\n } else {\n paddingSize = drwav__chunk_padding_size_w64(pWav->dataChunkDataSize);\n }\n \n if (paddingSize > 0) {\n drwav_uint64 paddingData = 0;\n drwav__write(pWav, &paddingData, paddingSize); /* Byte order does not matter for this. */\n }\n\n /*\n Chunk sizes. When using sequential mode, these will have been filled in at initialization time. We only need\n to do this when using non-sequential mode.\n */\n if (pWav->onSeek && !pWav->isSequentialWrite) {\n if (pWav->container == drwav_container_riff) {\n /* The \"RIFF\" chunk size. */\n if (pWav->onSeek(pWav->pUserData, 4, drwav_seek_origin_start)) {\n drwav_uint32 riffChunkSize = drwav__riff_chunk_size_riff(pWav->dataChunkDataSize);\n drwav__write_u32ne_to_le(pWav, riffChunkSize);\n }\n\n /* the \"data\" chunk size. */\n if (pWav->onSeek(pWav->pUserData, (int)pWav->dataChunkDataPos + 4, drwav_seek_origin_start)) {\n drwav_uint32 dataChunkSize = drwav__data_chunk_size_riff(pWav->dataChunkDataSize);\n drwav__write_u32ne_to_le(pWav, dataChunkSize);\n }\n } else if (pWav->container == drwav_container_w64) {\n /* The \"RIFF\" chunk size. */\n if (pWav->onSeek(pWav->pUserData, 16, drwav_seek_origin_start)) {\n drwav_uint64 riffChunkSize = drwav__riff_chunk_size_w64(pWav->dataChunkDataSize);\n drwav__write_u64ne_to_le(pWav, riffChunkSize);\n }\n\n /* The \"data\" chunk size. */\n if (pWav->onSeek(pWav->pUserData, (int)pWav->dataChunkDataPos + 16, drwav_seek_origin_start)) {\n drwav_uint64 dataChunkSize = drwav__data_chunk_size_w64(pWav->dataChunkDataSize);\n drwav__write_u64ne_to_le(pWav, dataChunkSize);\n }\n } else if (pWav->container == drwav_container_rf64) {\n /* We only need to update the ds64 chunk. The \"RIFF\" and \"data\" chunks always have their sizes set to 0xFFFFFFFF for RF64. */\n int ds64BodyPos = 12 + 8;\n\n /* The \"RIFF\" chunk size. */\n if (pWav->onSeek(pWav->pUserData, ds64BodyPos + 0, drwav_seek_origin_start)) {\n drwav_uint64 riffChunkSize = drwav__riff_chunk_size_rf64(pWav->dataChunkDataSize);\n drwav__write_u64ne_to_le(pWav, riffChunkSize);\n }\n\n /* The \"data\" chunk size. */\n if (pWav->onSeek(pWav->pUserData, ds64BodyPos + 8, drwav_seek_origin_start)) {\n drwav_uint64 dataChunkSize = drwav__data_chunk_size_rf64(pWav->dataChunkDataSize);\n drwav__write_u64ne_to_le(pWav, dataChunkSize);\n }\n }\n }\n\n /* Validation for sequential mode. */\n if (pWav->isSequentialWrite) {\n if (pWav->dataChunkDataSize != pWav->dataChunkDataSizeTargetWrite) {\n result = DRWAV_INVALID_FILE;\n }\n }\n }\n\n#ifndef DR_WAV_NO_STDIO\n /*\n If we opened the file with drwav_open_file() we will want to close the file handle. We can know whether or not drwav_open_file()\n was used by looking at the onRead and onSeek callbacks.\n */\n if (pWav->onRead == drwav__on_read_stdio || pWav->onWrite == drwav__on_write_stdio) {\n fclose((FILE*)pWav->pUserData);\n }\n#endif\n\n return result;\n}\n\n\n\nDRWAV_API size_t drwav_read_raw(drwav* pWav, size_t bytesToRead, void* pBufferOut)\n{\n size_t bytesRead;\n\n if (pWav == NULL || bytesToRead == 0) {\n return 0;\n }\n\n if (bytesToRead > pWav->bytesRemaining) {\n bytesToRead = (size_t)pWav->bytesRemaining;\n }\n\n if (pBufferOut != NULL) {\n bytesRead = pWav->onRead(pWav->pUserData, pBufferOut, bytesToRead);\n } else {\n /* We need to seek. If we fail, we need to read-and-discard to make sure we get a good byte count. */\n bytesRead = 0;\n while (bytesRead < bytesToRead) {\n size_t bytesToSeek = (bytesToRead - bytesRead);\n if (bytesToSeek > 0x7FFFFFFF) {\n bytesToSeek = 0x7FFFFFFF;\n }\n\n if (pWav->onSeek(pWav->pUserData, (int)bytesToSeek, drwav_seek_origin_current) == DRWAV_FALSE) {\n break;\n }\n\n bytesRead += bytesToSeek;\n }\n\n /* When we get here we may need to read-and-discard some data. */\n while (bytesRead < bytesToRead) {\n drwav_uint8 buffer[4096];\n size_t bytesSeeked;\n size_t bytesToSeek = (bytesToRead - bytesRead);\n if (bytesToSeek > sizeof(buffer)) {\n bytesToSeek = sizeof(buffer);\n }\n\n bytesSeeked = pWav->onRead(pWav->pUserData, buffer, bytesToSeek);\n bytesRead += bytesSeeked;\n\n if (bytesSeeked < bytesToSeek) {\n break; /* Reached the end. */\n }\n }\n }\n\n pWav->bytesRemaining -= bytesRead;\n return bytesRead;\n}\n\n\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_le(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut)\n{\n drwav_uint32 bytesPerFrame;\n drwav_uint64 bytesToRead; /* Intentionally uint64 instead of size_t so we can do a check that we're not reading too much on 32-bit builds. */\n\n if (pWav == NULL || framesToRead == 0) {\n return 0;\n }\n\n /* Cannot use this function for compressed formats. */\n if (drwav__is_compressed_format_tag(pWav->translatedFormatTag)) {\n return 0;\n }\n\n bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n /* Don't try to read more samples than can potentially fit in the output buffer. */\n bytesToRead = framesToRead * bytesPerFrame;\n if (bytesToRead > DRWAV_SIZE_MAX) {\n bytesToRead = (DRWAV_SIZE_MAX / bytesPerFrame) * bytesPerFrame; /* Round the number of bytes to read to a clean frame boundary. */\n }\n\n /*\n Doing an explicit check here just to make it clear that we don't want to be attempt to read anything if there's no bytes to read. There\n *could* be a time where it evaluates to 0 due to overflowing.\n */\n if (bytesToRead == 0) {\n return 0;\n }\n\n return drwav_read_raw(pWav, (size_t)bytesToRead, pBufferOut) / bytesPerFrame;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_be(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut)\n{\n drwav_uint64 framesRead = drwav_read_pcm_frames_le(pWav, framesToRead, pBufferOut);\n\n if (pBufferOut != NULL) {\n drwav__bswap_samples(pBufferOut, framesRead*pWav->channels, drwav_get_bytes_per_pcm_frame(pWav)/pWav->channels, pWav->translatedFormatTag);\n }\n\n return framesRead;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut)\n{\n if (drwav__is_little_endian()) {\n return drwav_read_pcm_frames_le(pWav, framesToRead, pBufferOut);\n } else {\n return drwav_read_pcm_frames_be(pWav, framesToRead, pBufferOut);\n }\n}\n\n\n\nDRWAV_API drwav_bool32 drwav_seek_to_first_pcm_frame(drwav* pWav)\n{\n if (pWav->onWrite != NULL) {\n return DRWAV_FALSE; /* No seeking in write mode. */\n }\n\n if (!pWav->onSeek(pWav->pUserData, (int)pWav->dataChunkDataPos, drwav_seek_origin_start)) {\n return DRWAV_FALSE;\n }\n\n if (drwav__is_compressed_format_tag(pWav->translatedFormatTag)) {\n pWav->compressed.iCurrentPCMFrame = 0;\n\n /* Cached data needs to be cleared for compressed formats. */\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {\n DRWAV_ZERO_OBJECT(&pWav->msadpcm);\n } else if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {\n DRWAV_ZERO_OBJECT(&pWav->ima);\n } else {\n DRWAV_ASSERT(DRWAV_FALSE); /* If this assertion is triggered it means I've implemented a new compressed format but forgot to add a branch for it here. */\n }\n }\n \n pWav->bytesRemaining = pWav->dataChunkDataSize;\n return DRWAV_TRUE;\n}\n\nDRWAV_API drwav_bool32 drwav_seek_to_pcm_frame(drwav* pWav, drwav_uint64 targetFrameIndex)\n{\n /* Seeking should be compatible with wave files > 2GB. */\n\n if (pWav == NULL || pWav->onSeek == NULL) {\n return DRWAV_FALSE;\n }\n\n /* No seeking in write mode. */\n if (pWav->onWrite != NULL) {\n return DRWAV_FALSE;\n }\n\n /* If there are no samples, just return DRWAV_TRUE without doing anything. */\n if (pWav->totalPCMFrameCount == 0) {\n return DRWAV_TRUE;\n }\n\n /* Make sure the sample is clamped. */\n if (targetFrameIndex >= pWav->totalPCMFrameCount) {\n targetFrameIndex = pWav->totalPCMFrameCount - 1;\n }\n\n /*\n For compressed formats we just use a slow generic seek. If we are seeking forward we just seek forward. If we are going backwards we need\n to seek back to the start.\n */\n if (drwav__is_compressed_format_tag(pWav->translatedFormatTag)) {\n /* TODO: This can be optimized. */\n \n /*\n If we're seeking forward it's simple - just keep reading samples until we hit the sample we're requesting. If we're seeking backwards,\n we first need to seek back to the start and then just do the same thing as a forward seek.\n */\n if (targetFrameIndex < pWav->compressed.iCurrentPCMFrame) {\n if (!drwav_seek_to_first_pcm_frame(pWav)) {\n return DRWAV_FALSE;\n }\n }\n\n if (targetFrameIndex > pWav->compressed.iCurrentPCMFrame) {\n drwav_uint64 offsetInFrames = targetFrameIndex - pWav->compressed.iCurrentPCMFrame;\n\n drwav_int16 devnull[2048];\n while (offsetInFrames > 0) {\n drwav_uint64 framesRead = 0;\n drwav_uint64 framesToRead = offsetInFrames;\n if (framesToRead > drwav_countof(devnull)/pWav->channels) {\n framesToRead = drwav_countof(devnull)/pWav->channels;\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {\n framesRead = drwav_read_pcm_frames_s16__msadpcm(pWav, framesToRead, devnull);\n } else if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {\n framesRead = drwav_read_pcm_frames_s16__ima(pWav, framesToRead, devnull);\n } else {\n DRWAV_ASSERT(DRWAV_FALSE); /* If this assertion is triggered it means I've implemented a new compressed format but forgot to add a branch for it here. */\n }\n\n if (framesRead != framesToRead) {\n return DRWAV_FALSE;\n }\n\n offsetInFrames -= framesRead;\n }\n }\n } else {\n drwav_uint64 totalSizeInBytes;\n drwav_uint64 currentBytePos;\n drwav_uint64 targetBytePos;\n drwav_uint64 offset;\n\n totalSizeInBytes = pWav->totalPCMFrameCount * drwav_get_bytes_per_pcm_frame(pWav);\n DRWAV_ASSERT(totalSizeInBytes >= pWav->bytesRemaining);\n\n currentBytePos = totalSizeInBytes - pWav->bytesRemaining;\n targetBytePos = targetFrameIndex * drwav_get_bytes_per_pcm_frame(pWav);\n\n if (currentBytePos < targetBytePos) {\n /* Offset forwards. */\n offset = (targetBytePos - currentBytePos);\n } else {\n /* Offset backwards. */\n if (!drwav_seek_to_first_pcm_frame(pWav)) {\n return DRWAV_FALSE;\n }\n offset = targetBytePos;\n }\n\n while (offset > 0) {\n int offset32 = ((offset > INT_MAX) ? INT_MAX : (int)offset);\n if (!pWav->onSeek(pWav->pUserData, offset32, drwav_seek_origin_current)) {\n return DRWAV_FALSE;\n }\n\n pWav->bytesRemaining -= offset32;\n offset -= offset32;\n }\n }\n\n return DRWAV_TRUE;\n}\n\n\nDRWAV_API size_t drwav_write_raw(drwav* pWav, size_t bytesToWrite, const void* pData)\n{\n size_t bytesWritten;\n\n if (pWav == NULL || bytesToWrite == 0 || pData == NULL) {\n return 0;\n }\n\n bytesWritten = pWav->onWrite(pWav->pUserData, pData, bytesToWrite);\n pWav->dataChunkDataSize += bytesWritten;\n\n return bytesWritten;\n}\n\n\nDRWAV_API drwav_uint64 drwav_write_pcm_frames_le(drwav* pWav, drwav_uint64 framesToWrite, const void* pData)\n{\n drwav_uint64 bytesToWrite;\n drwav_uint64 bytesWritten;\n const drwav_uint8* pRunningData;\n\n if (pWav == NULL || framesToWrite == 0 || pData == NULL) {\n return 0;\n }\n\n bytesToWrite = ((framesToWrite * pWav->channels * pWav->bitsPerSample) / 8);\n if (bytesToWrite > DRWAV_SIZE_MAX) {\n return 0;\n }\n\n bytesWritten = 0;\n pRunningData = (const drwav_uint8*)pData;\n\n while (bytesToWrite > 0) {\n size_t bytesJustWritten;\n drwav_uint64 bytesToWriteThisIteration;\n\n bytesToWriteThisIteration = bytesToWrite;\n DRWAV_ASSERT(bytesToWriteThisIteration <= DRWAV_SIZE_MAX); /* <-- This is checked above. */\n\n bytesJustWritten = drwav_write_raw(pWav, (size_t)bytesToWriteThisIteration, pRunningData);\n if (bytesJustWritten == 0) {\n break;\n }\n\n bytesToWrite -= bytesJustWritten;\n bytesWritten += bytesJustWritten;\n pRunningData += bytesJustWritten;\n }\n\n return (bytesWritten * 8) / pWav->bitsPerSample / pWav->channels;\n}\n\nDRWAV_API drwav_uint64 drwav_write_pcm_frames_be(drwav* pWav, drwav_uint64 framesToWrite, const void* pData)\n{\n drwav_uint64 bytesToWrite;\n drwav_uint64 bytesWritten;\n drwav_uint32 bytesPerSample;\n const drwav_uint8* pRunningData;\n\n if (pWav == NULL || framesToWrite == 0 || pData == NULL) {\n return 0;\n }\n\n bytesToWrite = ((framesToWrite * pWav->channels * pWav->bitsPerSample) / 8);\n if (bytesToWrite > DRWAV_SIZE_MAX) {\n return 0;\n }\n\n bytesWritten = 0;\n pRunningData = (const drwav_uint8*)pData;\n\n bytesPerSample = drwav_get_bytes_per_pcm_frame(pWav) / pWav->channels;\n \n while (bytesToWrite > 0) {\n drwav_uint8 temp[4096];\n drwav_uint32 sampleCount;\n size_t bytesJustWritten;\n drwav_uint64 bytesToWriteThisIteration;\n\n bytesToWriteThisIteration = bytesToWrite;\n DRWAV_ASSERT(bytesToWriteThisIteration <= DRWAV_SIZE_MAX); /* <-- This is checked above. */\n\n /*\n WAV files are always little-endian. We need to byte swap on big-endian architectures. Since our input buffer is read-only we need\n to use an intermediary buffer for the conversion.\n */\n sampleCount = sizeof(temp)/bytesPerSample;\n\n if (bytesToWriteThisIteration > ((drwav_uint64)sampleCount)*bytesPerSample) {\n bytesToWriteThisIteration = ((drwav_uint64)sampleCount)*bytesPerSample;\n }\n\n DRWAV_COPY_MEMORY(temp, pRunningData, (size_t)bytesToWriteThisIteration);\n drwav__bswap_samples(temp, sampleCount, bytesPerSample, pWav->translatedFormatTag);\n\n bytesJustWritten = drwav_write_raw(pWav, (size_t)bytesToWriteThisIteration, temp);\n if (bytesJustWritten == 0) {\n break;\n }\n\n bytesToWrite -= bytesJustWritten;\n bytesWritten += bytesJustWritten;\n pRunningData += bytesJustWritten;\n }\n\n return (bytesWritten * 8) / pWav->bitsPerSample / pWav->channels;\n}\n\nDRWAV_API drwav_uint64 drwav_write_pcm_frames(drwav* pWav, drwav_uint64 framesToWrite, const void* pData)\n{\n if (drwav__is_little_endian()) {\n return drwav_write_pcm_frames_le(pWav, framesToWrite, pData);\n } else {\n return drwav_write_pcm_frames_be(pWav, framesToWrite, pData);\n }\n}\n\n\nstatic drwav_uint64 drwav_read_pcm_frames_s16__msadpcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)\n{\n drwav_uint64 totalFramesRead = 0;\n\n DRWAV_ASSERT(pWav != NULL);\n DRWAV_ASSERT(framesToRead > 0);\n\n /* TODO: Lots of room for optimization here. */\n\n while (framesToRead > 0 && pWav->compressed.iCurrentPCMFrame < pWav->totalPCMFrameCount) {\n /* If there are no cached frames we need to load a new block. */\n if (pWav->msadpcm.cachedFrameCount == 0 && pWav->msadpcm.bytesRemainingInBlock == 0) {\n if (pWav->channels == 1) {\n /* Mono. */\n drwav_uint8 header[7];\n if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {\n return totalFramesRead;\n }\n pWav->msadpcm.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);\n\n pWav->msadpcm.predictor[0] = header[0];\n pWav->msadpcm.delta[0] = drwav__bytes_to_s16(header + 1);\n pWav->msadpcm.prevFrames[0][1] = (drwav_int32)drwav__bytes_to_s16(header + 3);\n pWav->msadpcm.prevFrames[0][0] = (drwav_int32)drwav__bytes_to_s16(header + 5);\n pWav->msadpcm.cachedFrames[2] = pWav->msadpcm.prevFrames[0][0];\n pWav->msadpcm.cachedFrames[3] = pWav->msadpcm.prevFrames[0][1];\n pWav->msadpcm.cachedFrameCount = 2;\n } else {\n /* Stereo. */\n drwav_uint8 header[14];\n if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {\n return totalFramesRead;\n }\n pWav->msadpcm.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);\n\n pWav->msadpcm.predictor[0] = header[0];\n pWav->msadpcm.predictor[1] = header[1];\n pWav->msadpcm.delta[0] = drwav__bytes_to_s16(header + 2);\n pWav->msadpcm.delta[1] = drwav__bytes_to_s16(header + 4);\n pWav->msadpcm.prevFrames[0][1] = (drwav_int32)drwav__bytes_to_s16(header + 6);\n pWav->msadpcm.prevFrames[1][1] = (drwav_int32)drwav__bytes_to_s16(header + 8);\n pWav->msadpcm.prevFrames[0][0] = (drwav_int32)drwav__bytes_to_s16(header + 10);\n pWav->msadpcm.prevFrames[1][0] = (drwav_int32)drwav__bytes_to_s16(header + 12);\n\n pWav->msadpcm.cachedFrames[0] = pWav->msadpcm.prevFrames[0][0];\n pWav->msadpcm.cachedFrames[1] = pWav->msadpcm.prevFrames[1][0];\n pWav->msadpcm.cachedFrames[2] = pWav->msadpcm.prevFrames[0][1];\n pWav->msadpcm.cachedFrames[3] = pWav->msadpcm.prevFrames[1][1];\n pWav->msadpcm.cachedFrameCount = 2;\n }\n }\n\n /* Output anything that's cached. */\n while (framesToRead > 0 && pWav->msadpcm.cachedFrameCount > 0 && pWav->compressed.iCurrentPCMFrame < pWav->totalPCMFrameCount) {\n if (pBufferOut != NULL) {\n drwav_uint32 iSample = 0;\n for (iSample = 0; iSample < pWav->channels; iSample += 1) {\n pBufferOut[iSample] = (drwav_int16)pWav->msadpcm.cachedFrames[(drwav_countof(pWav->msadpcm.cachedFrames) - (pWav->msadpcm.cachedFrameCount*pWav->channels)) + iSample];\n }\n\n pBufferOut += pWav->channels;\n }\n\n framesToRead -= 1;\n totalFramesRead += 1;\n pWav->compressed.iCurrentPCMFrame += 1;\n pWav->msadpcm.cachedFrameCount -= 1;\n }\n\n if (framesToRead == 0) {\n return totalFramesRead;\n }\n\n\n /*\n If there's nothing left in the cache, just go ahead and load more. If there's nothing left to load in the current block we just continue to the next\n loop iteration which will trigger the loading of a new block.\n */\n if (pWav->msadpcm.cachedFrameCount == 0) {\n if (pWav->msadpcm.bytesRemainingInBlock == 0) {\n continue;\n } else {\n static drwav_int32 adaptationTable[] = { \n 230, 230, 230, 230, 307, 409, 512, 614, \n 768, 614, 512, 409, 307, 230, 230, 230 \n };\n static drwav_int32 coeff1Table[] = { 256, 512, 0, 192, 240, 460, 392 };\n static drwav_int32 coeff2Table[] = { 0, -256, 0, 64, 0, -208, -232 };\n\n drwav_uint8 nibbles;\n drwav_int32 nibble0;\n drwav_int32 nibble1;\n\n if (pWav->onRead(pWav->pUserData, &nibbles, 1) != 1) {\n return totalFramesRead;\n }\n pWav->msadpcm.bytesRemainingInBlock -= 1;\n\n /* TODO: Optimize away these if statements. */\n nibble0 = ((nibbles & 0xF0) >> 4); if ((nibbles & 0x80)) { nibble0 |= 0xFFFFFFF0UL; }\n nibble1 = ((nibbles & 0x0F) >> 0); if ((nibbles & 0x08)) { nibble1 |= 0xFFFFFFF0UL; }\n\n if (pWav->channels == 1) {\n /* Mono. */\n drwav_int32 newSample0;\n drwav_int32 newSample1;\n\n newSample0 = ((pWav->msadpcm.prevFrames[0][1] * coeff1Table[pWav->msadpcm.predictor[0]]) + (pWav->msadpcm.prevFrames[0][0] * coeff2Table[pWav->msadpcm.predictor[0]])) >> 8;\n newSample0 += nibble0 * pWav->msadpcm.delta[0];\n newSample0 = drwav_clamp(newSample0, -32768, 32767);\n\n pWav->msadpcm.delta[0] = (adaptationTable[((nibbles & 0xF0) >> 4)] * pWav->msadpcm.delta[0]) >> 8;\n if (pWav->msadpcm.delta[0] < 16) {\n pWav->msadpcm.delta[0] = 16;\n }\n\n pWav->msadpcm.prevFrames[0][0] = pWav->msadpcm.prevFrames[0][1];\n pWav->msadpcm.prevFrames[0][1] = newSample0;\n\n\n newSample1 = ((pWav->msadpcm.prevFrames[0][1] * coeff1Table[pWav->msadpcm.predictor[0]]) + (pWav->msadpcm.prevFrames[0][0] * coeff2Table[pWav->msadpcm.predictor[0]])) >> 8;\n newSample1 += nibble1 * pWav->msadpcm.delta[0];\n newSample1 = drwav_clamp(newSample1, -32768, 32767);\n\n pWav->msadpcm.delta[0] = (adaptationTable[((nibbles & 0x0F) >> 0)] * pWav->msadpcm.delta[0]) >> 8;\n if (pWav->msadpcm.delta[0] < 16) {\n pWav->msadpcm.delta[0] = 16;\n }\n\n pWav->msadpcm.prevFrames[0][0] = pWav->msadpcm.prevFrames[0][1];\n pWav->msadpcm.prevFrames[0][1] = newSample1;\n\n\n pWav->msadpcm.cachedFrames[2] = newSample0;\n pWav->msadpcm.cachedFrames[3] = newSample1;\n pWav->msadpcm.cachedFrameCount = 2;\n } else {\n /* Stereo. */\n drwav_int32 newSample0;\n drwav_int32 newSample1;\n\n /* Left. */\n newSample0 = ((pWav->msadpcm.prevFrames[0][1] * coeff1Table[pWav->msadpcm.predictor[0]]) + (pWav->msadpcm.prevFrames[0][0] * coeff2Table[pWav->msadpcm.predictor[0]])) >> 8;\n newSample0 += nibble0 * pWav->msadpcm.delta[0];\n newSample0 = drwav_clamp(newSample0, -32768, 32767);\n\n pWav->msadpcm.delta[0] = (adaptationTable[((nibbles & 0xF0) >> 4)] * pWav->msadpcm.delta[0]) >> 8;\n if (pWav->msadpcm.delta[0] < 16) {\n pWav->msadpcm.delta[0] = 16;\n }\n\n pWav->msadpcm.prevFrames[0][0] = pWav->msadpcm.prevFrames[0][1];\n pWav->msadpcm.prevFrames[0][1] = newSample0;\n\n\n /* Right. */\n newSample1 = ((pWav->msadpcm.prevFrames[1][1] * coeff1Table[pWav->msadpcm.predictor[1]]) + (pWav->msadpcm.prevFrames[1][0] * coeff2Table[pWav->msadpcm.predictor[1]])) >> 8;\n newSample1 += nibble1 * pWav->msadpcm.delta[1];\n newSample1 = drwav_clamp(newSample1, -32768, 32767);\n\n pWav->msadpcm.delta[1] = (adaptationTable[((nibbles & 0x0F) >> 0)] * pWav->msadpcm.delta[1]) >> 8;\n if (pWav->msadpcm.delta[1] < 16) {\n pWav->msadpcm.delta[1] = 16;\n }\n\n pWav->msadpcm.prevFrames[1][0] = pWav->msadpcm.prevFrames[1][1];\n pWav->msadpcm.prevFrames[1][1] = newSample1;\n\n pWav->msadpcm.cachedFrames[2] = newSample0;\n pWav->msadpcm.cachedFrames[3] = newSample1;\n pWav->msadpcm.cachedFrameCount = 1;\n }\n }\n }\n }\n\n return totalFramesRead;\n}\n\n\nstatic drwav_uint64 drwav_read_pcm_frames_s16__ima(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)\n{\n drwav_uint64 totalFramesRead = 0;\n drwav_uint32 iChannel;\n\n static drwav_int32 indexTable[16] = {\n -1, -1, -1, -1, 2, 4, 6, 8,\n -1, -1, -1, -1, 2, 4, 6, 8\n };\n\n static drwav_int32 stepTable[89] = {\n 7, 8, 9, 10, 11, 12, 13, 14, 16, 17, \n 19, 21, 23, 25, 28, 31, 34, 37, 41, 45, \n 50, 55, 60, 66, 73, 80, 88, 97, 107, 118, \n 130, 143, 157, 173, 190, 209, 230, 253, 279, 307,\n 337, 371, 408, 449, 494, 544, 598, 658, 724, 796,\n 876, 963, 1060, 1166, 1282, 1411, 1552, 1707, 1878, 2066, \n 2272, 2499, 2749, 3024, 3327, 3660, 4026, 4428, 4871, 5358,\n 5894, 6484, 7132, 7845, 8630, 9493, 10442, 11487, 12635, 13899, \n 15289, 16818, 18500, 20350, 22385, 24623, 27086, 29794, 32767 \n };\n\n DRWAV_ASSERT(pWav != NULL);\n DRWAV_ASSERT(framesToRead > 0);\n\n /* TODO: Lots of room for optimization here. */\n\n while (framesToRead > 0 && pWav->compressed.iCurrentPCMFrame < pWav->totalPCMFrameCount) {\n /* If there are no cached samples we need to load a new block. */\n if (pWav->ima.cachedFrameCount == 0 && pWav->ima.bytesRemainingInBlock == 0) {\n if (pWav->channels == 1) {\n /* Mono. */\n drwav_uint8 header[4];\n if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {\n return totalFramesRead;\n }\n pWav->ima.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);\n\n if (header[2] >= drwav_countof(stepTable)) {\n pWav->onSeek(pWav->pUserData, pWav->ima.bytesRemainingInBlock, drwav_seek_origin_current);\n pWav->ima.bytesRemainingInBlock = 0;\n return totalFramesRead; /* Invalid data. */\n }\n\n pWav->ima.predictor[0] = drwav__bytes_to_s16(header + 0);\n pWav->ima.stepIndex[0] = header[2];\n pWav->ima.cachedFrames[drwav_countof(pWav->ima.cachedFrames) - 1] = pWav->ima.predictor[0];\n pWav->ima.cachedFrameCount = 1;\n } else {\n /* Stereo. */\n drwav_uint8 header[8];\n if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {\n return totalFramesRead;\n }\n pWav->ima.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);\n\n if (header[2] >= drwav_countof(stepTable) || header[6] >= drwav_countof(stepTable)) {\n pWav->onSeek(pWav->pUserData, pWav->ima.bytesRemainingInBlock, drwav_seek_origin_current);\n pWav->ima.bytesRemainingInBlock = 0;\n return totalFramesRead; /* Invalid data. */\n }\n\n pWav->ima.predictor[0] = drwav__bytes_to_s16(header + 0);\n pWav->ima.stepIndex[0] = header[2];\n pWav->ima.predictor[1] = drwav__bytes_to_s16(header + 4);\n pWav->ima.stepIndex[1] = header[6];\n\n pWav->ima.cachedFrames[drwav_countof(pWav->ima.cachedFrames) - 2] = pWav->ima.predictor[0];\n pWav->ima.cachedFrames[drwav_countof(pWav->ima.cachedFrames) - 1] = pWav->ima.predictor[1];\n pWav->ima.cachedFrameCount = 1;\n }\n }\n\n /* Output anything that's cached. */\n while (framesToRead > 0 && pWav->ima.cachedFrameCount > 0 && pWav->compressed.iCurrentPCMFrame < pWav->totalPCMFrameCount) {\n if (pBufferOut != NULL) {\n drwav_uint32 iSample;\n for (iSample = 0; iSample < pWav->channels; iSample += 1) {\n pBufferOut[iSample] = (drwav_int16)pWav->ima.cachedFrames[(drwav_countof(pWav->ima.cachedFrames) - (pWav->ima.cachedFrameCount*pWav->channels)) + iSample];\n }\n pBufferOut += pWav->channels;\n }\n\n framesToRead -= 1;\n totalFramesRead += 1;\n pWav->compressed.iCurrentPCMFrame += 1;\n pWav->ima.cachedFrameCount -= 1;\n }\n\n if (framesToRead == 0) {\n return totalFramesRead;\n }\n\n /*\n If there's nothing left in the cache, just go ahead and load more. If there's nothing left to load in the current block we just continue to the next\n loop iteration which will trigger the loading of a new block.\n */\n if (pWav->ima.cachedFrameCount == 0) {\n if (pWav->ima.bytesRemainingInBlock == 0) {\n continue;\n } else {\n /*\n From what I can tell with stereo streams, it looks like every 4 bytes (8 samples) is for one channel. So it goes 4 bytes for the\n left channel, 4 bytes for the right channel.\n */\n pWav->ima.cachedFrameCount = 8;\n for (iChannel = 0; iChannel < pWav->channels; ++iChannel) {\n drwav_uint32 iByte;\n drwav_uint8 nibbles[4];\n if (pWav->onRead(pWav->pUserData, &nibbles, 4) != 4) {\n pWav->ima.cachedFrameCount = 0;\n return totalFramesRead;\n }\n pWav->ima.bytesRemainingInBlock -= 4;\n\n for (iByte = 0; iByte < 4; ++iByte) {\n drwav_uint8 nibble0 = ((nibbles[iByte] & 0x0F) >> 0);\n drwav_uint8 nibble1 = ((nibbles[iByte] & 0xF0) >> 4);\n\n drwav_int32 step = stepTable[pWav->ima.stepIndex[iChannel]];\n drwav_int32 predictor = pWav->ima.predictor[iChannel];\n\n drwav_int32 diff = step >> 3;\n if (nibble0 & 1) diff += step >> 2;\n if (nibble0 & 2) diff += step >> 1;\n if (nibble0 & 4) diff += step;\n if (nibble0 & 8) diff = -diff;\n\n predictor = drwav_clamp(predictor + diff, -32768, 32767);\n pWav->ima.predictor[iChannel] = predictor;\n pWav->ima.stepIndex[iChannel] = drwav_clamp(pWav->ima.stepIndex[iChannel] + indexTable[nibble0], 0, (drwav_int32)drwav_countof(stepTable)-1);\n pWav->ima.cachedFrames[(drwav_countof(pWav->ima.cachedFrames) - (pWav->ima.cachedFrameCount*pWav->channels)) + (iByte*2+0)*pWav->channels + iChannel] = predictor;\n\n\n step = stepTable[pWav->ima.stepIndex[iChannel]];\n predictor = pWav->ima.predictor[iChannel];\n\n diff = step >> 3;\n if (nibble1 & 1) diff += step >> 2;\n if (nibble1 & 2) diff += step >> 1;\n if (nibble1 & 4) diff += step;\n if (nibble1 & 8) diff = -diff;\n\n predictor = drwav_clamp(predictor + diff, -32768, 32767);\n pWav->ima.predictor[iChannel] = predictor;\n pWav->ima.stepIndex[iChannel] = drwav_clamp(pWav->ima.stepIndex[iChannel] + indexTable[nibble1], 0, (drwav_int32)drwav_countof(stepTable)-1);\n pWav->ima.cachedFrames[(drwav_countof(pWav->ima.cachedFrames) - (pWav->ima.cachedFrameCount*pWav->channels)) + (iByte*2+1)*pWav->channels + iChannel] = predictor;\n }\n }\n }\n }\n }\n\n return totalFramesRead;\n}\n\n\n#ifndef DR_WAV_NO_CONVERSION_API\nstatic unsigned short g_drwavAlawTable[256] = {\n 0xEA80, 0xEB80, 0xE880, 0xE980, 0xEE80, 0xEF80, 0xEC80, 0xED80, 0xE280, 0xE380, 0xE080, 0xE180, 0xE680, 0xE780, 0xE480, 0xE580, \n 0xF540, 0xF5C0, 0xF440, 0xF4C0, 0xF740, 0xF7C0, 0xF640, 0xF6C0, 0xF140, 0xF1C0, 0xF040, 0xF0C0, 0xF340, 0xF3C0, 0xF240, 0xF2C0, \n 0xAA00, 0xAE00, 0xA200, 0xA600, 0xBA00, 0xBE00, 0xB200, 0xB600, 0x8A00, 0x8E00, 0x8200, 0x8600, 0x9A00, 0x9E00, 0x9200, 0x9600, \n 0xD500, 0xD700, 0xD100, 0xD300, 0xDD00, 0xDF00, 0xD900, 0xDB00, 0xC500, 0xC700, 0xC100, 0xC300, 0xCD00, 0xCF00, 0xC900, 0xCB00, \n 0xFEA8, 0xFEB8, 0xFE88, 0xFE98, 0xFEE8, 0xFEF8, 0xFEC8, 0xFED8, 0xFE28, 0xFE38, 0xFE08, 0xFE18, 0xFE68, 0xFE78, 0xFE48, 0xFE58, \n 0xFFA8, 0xFFB8, 0xFF88, 0xFF98, 0xFFE8, 0xFFF8, 0xFFC8, 0xFFD8, 0xFF28, 0xFF38, 0xFF08, 0xFF18, 0xFF68, 0xFF78, 0xFF48, 0xFF58, \n 0xFAA0, 0xFAE0, 0xFA20, 0xFA60, 0xFBA0, 0xFBE0, 0xFB20, 0xFB60, 0xF8A0, 0xF8E0, 0xF820, 0xF860, 0xF9A0, 0xF9E0, 0xF920, 0xF960, \n 0xFD50, 0xFD70, 0xFD10, 0xFD30, 0xFDD0, 0xFDF0, 0xFD90, 0xFDB0, 0xFC50, 0xFC70, 0xFC10, 0xFC30, 0xFCD0, 0xFCF0, 0xFC90, 0xFCB0, \n 0x1580, 0x1480, 0x1780, 0x1680, 0x1180, 0x1080, 0x1380, 0x1280, 0x1D80, 0x1C80, 0x1F80, 0x1E80, 0x1980, 0x1880, 0x1B80, 0x1A80, \n 0x0AC0, 0x0A40, 0x0BC0, 0x0B40, 0x08C0, 0x0840, 0x09C0, 0x0940, 0x0EC0, 0x0E40, 0x0FC0, 0x0F40, 0x0CC0, 0x0C40, 0x0DC0, 0x0D40, \n 0x5600, 0x5200, 0x5E00, 0x5A00, 0x4600, 0x4200, 0x4E00, 0x4A00, 0x7600, 0x7200, 0x7E00, 0x7A00, 0x6600, 0x6200, 0x6E00, 0x6A00, \n 0x2B00, 0x2900, 0x2F00, 0x2D00, 0x2300, 0x2100, 0x2700, 0x2500, 0x3B00, 0x3900, 0x3F00, 0x3D00, 0x3300, 0x3100, 0x3700, 0x3500, \n 0x0158, 0x0148, 0x0178, 0x0168, 0x0118, 0x0108, 0x0138, 0x0128, 0x01D8, 0x01C8, 0x01F8, 0x01E8, 0x0198, 0x0188, 0x01B8, 0x01A8, \n 0x0058, 0x0048, 0x0078, 0x0068, 0x0018, 0x0008, 0x0038, 0x0028, 0x00D8, 0x00C8, 0x00F8, 0x00E8, 0x0098, 0x0088, 0x00B8, 0x00A8, \n 0x0560, 0x0520, 0x05E0, 0x05A0, 0x0460, 0x0420, 0x04E0, 0x04A0, 0x0760, 0x0720, 0x07E0, 0x07A0, 0x0660, 0x0620, 0x06E0, 0x06A0, \n 0x02B0, 0x0290, 0x02F0, 0x02D0, 0x0230, 0x0210, 0x0270, 0x0250, 0x03B0, 0x0390, 0x03F0, 0x03D0, 0x0330, 0x0310, 0x0370, 0x0350\n};\n\nstatic unsigned short g_drwavMulawTable[256] = {\n 0x8284, 0x8684, 0x8A84, 0x8E84, 0x9284, 0x9684, 0x9A84, 0x9E84, 0xA284, 0xA684, 0xAA84, 0xAE84, 0xB284, 0xB684, 0xBA84, 0xBE84, \n 0xC184, 0xC384, 0xC584, 0xC784, 0xC984, 0xCB84, 0xCD84, 0xCF84, 0xD184, 0xD384, 0xD584, 0xD784, 0xD984, 0xDB84, 0xDD84, 0xDF84, \n 0xE104, 0xE204, 0xE304, 0xE404, 0xE504, 0xE604, 0xE704, 0xE804, 0xE904, 0xEA04, 0xEB04, 0xEC04, 0xED04, 0xEE04, 0xEF04, 0xF004, \n 0xF0C4, 0xF144, 0xF1C4, 0xF244, 0xF2C4, 0xF344, 0xF3C4, 0xF444, 0xF4C4, 0xF544, 0xF5C4, 0xF644, 0xF6C4, 0xF744, 0xF7C4, 0xF844, \n 0xF8A4, 0xF8E4, 0xF924, 0xF964, 0xF9A4, 0xF9E4, 0xFA24, 0xFA64, 0xFAA4, 0xFAE4, 0xFB24, 0xFB64, 0xFBA4, 0xFBE4, 0xFC24, 0xFC64, \n 0xFC94, 0xFCB4, 0xFCD4, 0xFCF4, 0xFD14, 0xFD34, 0xFD54, 0xFD74, 0xFD94, 0xFDB4, 0xFDD4, 0xFDF4, 0xFE14, 0xFE34, 0xFE54, 0xFE74, \n 0xFE8C, 0xFE9C, 0xFEAC, 0xFEBC, 0xFECC, 0xFEDC, 0xFEEC, 0xFEFC, 0xFF0C, 0xFF1C, 0xFF2C, 0xFF3C, 0xFF4C, 0xFF5C, 0xFF6C, 0xFF7C, \n 0xFF88, 0xFF90, 0xFF98, 0xFFA0, 0xFFA8, 0xFFB0, 0xFFB8, 0xFFC0, 0xFFC8, 0xFFD0, 0xFFD8, 0xFFE0, 0xFFE8, 0xFFF0, 0xFFF8, 0x0000, \n 0x7D7C, 0x797C, 0x757C, 0x717C, 0x6D7C, 0x697C, 0x657C, 0x617C, 0x5D7C, 0x597C, 0x557C, 0x517C, 0x4D7C, 0x497C, 0x457C, 0x417C, \n 0x3E7C, 0x3C7C, 0x3A7C, 0x387C, 0x367C, 0x347C, 0x327C, 0x307C, 0x2E7C, 0x2C7C, 0x2A7C, 0x287C, 0x267C, 0x247C, 0x227C, 0x207C, \n 0x1EFC, 0x1DFC, 0x1CFC, 0x1BFC, 0x1AFC, 0x19FC, 0x18FC, 0x17FC, 0x16FC, 0x15FC, 0x14FC, 0x13FC, 0x12FC, 0x11FC, 0x10FC, 0x0FFC, \n 0x0F3C, 0x0EBC, 0x0E3C, 0x0DBC, 0x0D3C, 0x0CBC, 0x0C3C, 0x0BBC, 0x0B3C, 0x0ABC, 0x0A3C, 0x09BC, 0x093C, 0x08BC, 0x083C, 0x07BC, \n 0x075C, 0x071C, 0x06DC, 0x069C, 0x065C, 0x061C, 0x05DC, 0x059C, 0x055C, 0x051C, 0x04DC, 0x049C, 0x045C, 0x041C, 0x03DC, 0x039C, \n 0x036C, 0x034C, 0x032C, 0x030C, 0x02EC, 0x02CC, 0x02AC, 0x028C, 0x026C, 0x024C, 0x022C, 0x020C, 0x01EC, 0x01CC, 0x01AC, 0x018C, \n 0x0174, 0x0164, 0x0154, 0x0144, 0x0134, 0x0124, 0x0114, 0x0104, 0x00F4, 0x00E4, 0x00D4, 0x00C4, 0x00B4, 0x00A4, 0x0094, 0x0084, \n 0x0078, 0x0070, 0x0068, 0x0060, 0x0058, 0x0050, 0x0048, 0x0040, 0x0038, 0x0030, 0x0028, 0x0020, 0x0018, 0x0010, 0x0008, 0x0000\n};\n\nstatic DRWAV_INLINE drwav_int16 drwav__alaw_to_s16(drwav_uint8 sampleIn)\n{\n return (short)g_drwavAlawTable[sampleIn];\n}\n\nstatic DRWAV_INLINE drwav_int16 drwav__mulaw_to_s16(drwav_uint8 sampleIn)\n{\n return (short)g_drwavMulawTable[sampleIn];\n}\n\n\n\nstatic void drwav__pcm_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)\n{\n unsigned int i;\n\n /* Special case for 8-bit sample data because it's treated as unsigned. */\n if (bytesPerSample == 1) {\n drwav_u8_to_s16(pOut, pIn, totalSampleCount);\n return;\n }\n\n\n /* Slightly more optimal implementation for common formats. */\n if (bytesPerSample == 2) {\n for (i = 0; i < totalSampleCount; ++i) {\n *pOut++ = ((const drwav_int16*)pIn)[i];\n }\n return;\n }\n if (bytesPerSample == 3) {\n drwav_s24_to_s16(pOut, pIn, totalSampleCount);\n return;\n }\n if (bytesPerSample == 4) {\n drwav_s32_to_s16(pOut, (const drwav_int32*)pIn, totalSampleCount);\n return;\n }\n\n\n /* Anything more than 64 bits per sample is not supported. */\n if (bytesPerSample > 8) {\n DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));\n return;\n }\n\n\n /* Generic, slow converter. */\n for (i = 0; i < totalSampleCount; ++i) {\n drwav_uint64 sample = 0;\n unsigned int shift = (8 - bytesPerSample) * 8;\n\n unsigned int j;\n for (j = 0; j < bytesPerSample; j += 1) {\n DRWAV_ASSERT(j < 8);\n sample |= (drwav_uint64)(pIn[j]) << shift;\n shift += 8;\n }\n\n pIn += j;\n *pOut++ = (drwav_int16)((drwav_int64)sample >> 48);\n }\n}\n\nstatic void drwav__ieee_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)\n{\n if (bytesPerSample == 4) {\n drwav_f32_to_s16(pOut, (const float*)pIn, totalSampleCount);\n return;\n } else if (bytesPerSample == 8) {\n drwav_f64_to_s16(pOut, (const double*)pIn, totalSampleCount);\n return;\n } else {\n /* Only supporting 32- and 64-bit float. Output silence in all other cases. Contributions welcome for 16-bit float. */\n DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));\n return;\n }\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s16__pcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)\n{\n drwav_uint32 bytesPerFrame;\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n\n /* Fast path. */\n if ((pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM && pWav->bitsPerSample == 16) || pBufferOut == NULL) {\n return drwav_read_pcm_frames(pWav, framesToRead, pBufferOut);\n }\n \n bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n \n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav__pcm_to_s16(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s16__ieee(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n drwav_uint32 bytesPerFrame;\n\n if (pBufferOut == NULL) {\n return drwav_read_pcm_frames(pWav, framesToRead, NULL);\n }\n\n bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n \n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav__ieee_to_s16(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s16__alaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n drwav_uint32 bytesPerFrame;\n\n if (pBufferOut == NULL) {\n return drwav_read_pcm_frames(pWav, framesToRead, NULL);\n }\n\n bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n \n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav_alaw_to_s16(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s16__mulaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n drwav_uint32 bytesPerFrame;\n\n if (pBufferOut == NULL) {\n return drwav_read_pcm_frames(pWav, framesToRead, NULL);\n }\n\n bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav_mulaw_to_s16(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s16(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)\n{\n if (pWav == NULL || framesToRead == 0) {\n return 0;\n }\n\n if (pBufferOut == NULL) {\n return drwav_read_pcm_frames(pWav, framesToRead, NULL);\n }\n\n /* Don't try to read more samples than can potentially fit in the output buffer. */\n if (framesToRead * pWav->channels * sizeof(drwav_int16) > DRWAV_SIZE_MAX) {\n framesToRead = DRWAV_SIZE_MAX / sizeof(drwav_int16) / pWav->channels;\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM) {\n return drwav_read_pcm_frames_s16__pcm(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT) {\n return drwav_read_pcm_frames_s16__ieee(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW) {\n return drwav_read_pcm_frames_s16__alaw(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {\n return drwav_read_pcm_frames_s16__mulaw(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {\n return drwav_read_pcm_frames_s16__msadpcm(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {\n return drwav_read_pcm_frames_s16__ima(pWav, framesToRead, pBufferOut);\n }\n\n return 0;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s16le(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)\n{\n drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, framesToRead, pBufferOut);\n if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_FALSE) {\n drwav__bswap_samples_s16(pBufferOut, framesRead*pWav->channels);\n }\n\n return framesRead;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s16be(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)\n{\n drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, framesToRead, pBufferOut);\n if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_TRUE) {\n drwav__bswap_samples_s16(pBufferOut, framesRead*pWav->channels);\n }\n\n return framesRead;\n}\n\n\nDRWAV_API void drwav_u8_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n int r;\n size_t i;\n for (i = 0; i < sampleCount; ++i) {\n int x = pIn[i];\n r = x << 8;\n r = r - 32768;\n pOut[i] = (short)r;\n }\n}\n\nDRWAV_API void drwav_s24_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n int r;\n size_t i;\n for (i = 0; i < sampleCount; ++i) {\n int x = ((int)(((unsigned int)(((const drwav_uint8*)pIn)[i*3+0]) << 8) | ((unsigned int)(((const drwav_uint8*)pIn)[i*3+1]) << 16) | ((unsigned int)(((const drwav_uint8*)pIn)[i*3+2])) << 24)) >> 8;\n r = x >> 8;\n pOut[i] = (short)r;\n }\n}\n\nDRWAV_API void drwav_s32_to_s16(drwav_int16* pOut, const drwav_int32* pIn, size_t sampleCount)\n{\n int r;\n size_t i;\n for (i = 0; i < sampleCount; ++i) {\n int x = pIn[i];\n r = x >> 16;\n pOut[i] = (short)r;\n }\n}\n\nDRWAV_API void drwav_f32_to_s16(drwav_int16* pOut, const float* pIn, size_t sampleCount)\n{\n int r;\n size_t i;\n for (i = 0; i < sampleCount; ++i) {\n float x = pIn[i];\n float c;\n c = ((x < -1) ? -1 : ((x > 1) ? 1 : x));\n c = c + 1;\n r = (int)(c * 32767.5f);\n r = r - 32768;\n pOut[i] = (short)r;\n }\n}\n\nDRWAV_API void drwav_f64_to_s16(drwav_int16* pOut, const double* pIn, size_t sampleCount)\n{\n int r;\n size_t i;\n for (i = 0; i < sampleCount; ++i) {\n double x = pIn[i];\n double c;\n c = ((x < -1) ? -1 : ((x > 1) ? 1 : x));\n c = c + 1;\n r = (int)(c * 32767.5);\n r = r - 32768;\n pOut[i] = (short)r;\n }\n}\n\nDRWAV_API void drwav_alaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n for (i = 0; i < sampleCount; ++i) {\n pOut[i] = drwav__alaw_to_s16(pIn[i]);\n }\n}\n\nDRWAV_API void drwav_mulaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n for (i = 0; i < sampleCount; ++i) {\n pOut[i] = drwav__mulaw_to_s16(pIn[i]);\n }\n}\n\n\n\nstatic void drwav__pcm_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount, unsigned int bytesPerSample)\n{\n unsigned int i;\n\n /* Special case for 8-bit sample data because it's treated as unsigned. */\n if (bytesPerSample == 1) {\n drwav_u8_to_f32(pOut, pIn, sampleCount);\n return;\n }\n\n /* Slightly more optimal implementation for common formats. */\n if (bytesPerSample == 2) {\n drwav_s16_to_f32(pOut, (const drwav_int16*)pIn, sampleCount);\n return;\n }\n if (bytesPerSample == 3) {\n drwav_s24_to_f32(pOut, pIn, sampleCount);\n return;\n }\n if (bytesPerSample == 4) {\n drwav_s32_to_f32(pOut, (const drwav_int32*)pIn, sampleCount);\n return;\n }\n\n\n /* Anything more than 64 bits per sample is not supported. */\n if (bytesPerSample > 8) {\n DRWAV_ZERO_MEMORY(pOut, sampleCount * sizeof(*pOut));\n return;\n }\n\n\n /* Generic, slow converter. */\n for (i = 0; i < sampleCount; ++i) {\n drwav_uint64 sample = 0;\n unsigned int shift = (8 - bytesPerSample) * 8;\n\n unsigned int j;\n for (j = 0; j < bytesPerSample; j += 1) {\n DRWAV_ASSERT(j < 8);\n sample |= (drwav_uint64)(pIn[j]) << shift;\n shift += 8;\n }\n\n pIn += j;\n *pOut++ = (float)((drwav_int64)sample / 9223372036854775807.0);\n }\n}\n\nstatic void drwav__ieee_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount, unsigned int bytesPerSample)\n{\n if (bytesPerSample == 4) {\n unsigned int i;\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = ((const float*)pIn)[i];\n }\n return;\n } else if (bytesPerSample == 8) {\n drwav_f64_to_f32(pOut, (const double*)pIn, sampleCount);\n return;\n } else {\n /* Only supporting 32- and 64-bit float. Output silence in all other cases. Contributions welcome for 16-bit float. */\n DRWAV_ZERO_MEMORY(pOut, sampleCount * sizeof(*pOut));\n return;\n }\n}\n\n\nstatic drwav_uint64 drwav_read_pcm_frames_f32__pcm(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n\n drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav__pcm_to_f32(pBufferOut, sampleData, (size_t)framesRead*pWav->channels, bytesPerFrame/pWav->channels);\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_f32__msadpcm(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)\n{\n /*\n We're just going to borrow the implementation from the drwav_read_s16() since ADPCM is a little bit more complicated than other formats and I don't\n want to duplicate that code.\n */\n drwav_uint64 totalFramesRead = 0;\n drwav_int16 samples16[2048];\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels), samples16);\n if (framesRead == 0) {\n break;\n }\n\n drwav_s16_to_f32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels)); /* <-- Safe cast because we're clamping to 2048. */\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_f32__ima(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)\n{\n /*\n We're just going to borrow the implementation from the drwav_read_s16() since IMA-ADPCM is a little bit more complicated than other formats and I don't\n want to duplicate that code.\n */\n drwav_uint64 totalFramesRead = 0;\n drwav_int16 samples16[2048];\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels), samples16);\n if (framesRead == 0) {\n break;\n }\n\n drwav_s16_to_f32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels)); /* <-- Safe cast because we're clamping to 2048. */\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_f32__ieee(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n drwav_uint32 bytesPerFrame;\n\n /* Fast path. */\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT && pWav->bitsPerSample == 32) {\n return drwav_read_pcm_frames(pWav, framesToRead, pBufferOut);\n }\n \n bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav__ieee_to_f32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_f32__alaw(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav_alaw_to_f32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_f32__mulaw(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n\n drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav_mulaw_to_f32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_f32(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)\n{\n if (pWav == NULL || framesToRead == 0) {\n return 0;\n }\n\n if (pBufferOut == NULL) {\n return drwav_read_pcm_frames(pWav, framesToRead, NULL);\n }\n\n /* Don't try to read more samples than can potentially fit in the output buffer. */\n if (framesToRead * pWav->channels * sizeof(float) > DRWAV_SIZE_MAX) {\n framesToRead = DRWAV_SIZE_MAX / sizeof(float) / pWav->channels;\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM) {\n return drwav_read_pcm_frames_f32__pcm(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {\n return drwav_read_pcm_frames_f32__msadpcm(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT) {\n return drwav_read_pcm_frames_f32__ieee(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW) {\n return drwav_read_pcm_frames_f32__alaw(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {\n return drwav_read_pcm_frames_f32__mulaw(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {\n return drwav_read_pcm_frames_f32__ima(pWav, framesToRead, pBufferOut);\n }\n\n return 0;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_f32le(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)\n{\n drwav_uint64 framesRead = drwav_read_pcm_frames_f32(pWav, framesToRead, pBufferOut);\n if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_FALSE) {\n drwav__bswap_samples_f32(pBufferOut, framesRead*pWav->channels);\n }\n\n return framesRead;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_f32be(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)\n{\n drwav_uint64 framesRead = drwav_read_pcm_frames_f32(pWav, framesToRead, pBufferOut);\n if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_TRUE) {\n drwav__bswap_samples_f32(pBufferOut, framesRead*pWav->channels);\n }\n\n return framesRead;\n}\n\n\nDRWAV_API void drwav_u8_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n#ifdef DR_WAV_LIBSNDFILE_COMPAT\n /*\n It appears libsndfile uses slightly different logic for the u8 -> f32 conversion to dr_wav, which in my opinion is incorrect. It appears\n libsndfile performs the conversion something like \"f32 = (u8 / 256) * 2 - 1\", however I think it should be \"f32 = (u8 / 255) * 2 - 1\" (note\n the divisor of 256 vs 255). I use libsndfile as a benchmark for testing, so I'm therefore leaving this block here just for my automated\n correctness testing. This is disabled by default.\n */\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = (pIn[i] / 256.0f) * 2 - 1;\n }\n#else\n for (i = 0; i < sampleCount; ++i) {\n float x = pIn[i];\n x = x * 0.00784313725490196078f; /* 0..255 to 0..2 */\n x = x - 1; /* 0..2 to -1..1 */\n\n *pOut++ = x;\n }\n#endif\n}\n\nDRWAV_API void drwav_s16_to_f32(float* pOut, const drwav_int16* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = pIn[i] * 0.000030517578125f;\n }\n}\n\nDRWAV_API void drwav_s24_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n double x;\n drwav_uint32 a = ((drwav_uint32)(pIn[i*3+0]) << 8);\n drwav_uint32 b = ((drwav_uint32)(pIn[i*3+1]) << 16);\n drwav_uint32 c = ((drwav_uint32)(pIn[i*3+2]) << 24);\n\n x = (double)((drwav_int32)(a | b | c) >> 8);\n *pOut++ = (float)(x * 0.00000011920928955078125);\n }\n}\n\nDRWAV_API void drwav_s32_to_f32(float* pOut, const drwav_int32* pIn, size_t sampleCount)\n{\n size_t i;\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = (float)(pIn[i] / 2147483648.0);\n }\n}\n\nDRWAV_API void drwav_f64_to_f32(float* pOut, const double* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = (float)pIn[i];\n }\n}\n\nDRWAV_API void drwav_alaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = drwav__alaw_to_s16(pIn[i]) / 32768.0f;\n }\n}\n\nDRWAV_API void drwav_mulaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = drwav__mulaw_to_s16(pIn[i]) / 32768.0f;\n }\n}\n\n\n\nstatic void drwav__pcm_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)\n{\n unsigned int i;\n\n /* Special case for 8-bit sample data because it's treated as unsigned. */\n if (bytesPerSample == 1) {\n drwav_u8_to_s32(pOut, pIn, totalSampleCount);\n return;\n }\n\n /* Slightly more optimal implementation for common formats. */\n if (bytesPerSample == 2) {\n drwav_s16_to_s32(pOut, (const drwav_int16*)pIn, totalSampleCount);\n return;\n }\n if (bytesPerSample == 3) {\n drwav_s24_to_s32(pOut, pIn, totalSampleCount);\n return;\n }\n if (bytesPerSample == 4) {\n for (i = 0; i < totalSampleCount; ++i) {\n *pOut++ = ((const drwav_int32*)pIn)[i];\n }\n return;\n }\n\n\n /* Anything more than 64 bits per sample is not supported. */\n if (bytesPerSample > 8) {\n DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));\n return;\n }\n\n\n /* Generic, slow converter. */\n for (i = 0; i < totalSampleCount; ++i) {\n drwav_uint64 sample = 0;\n unsigned int shift = (8 - bytesPerSample) * 8;\n\n unsigned int j;\n for (j = 0; j < bytesPerSample; j += 1) {\n DRWAV_ASSERT(j < 8);\n sample |= (drwav_uint64)(pIn[j]) << shift;\n shift += 8;\n }\n\n pIn += j;\n *pOut++ = (drwav_int32)((drwav_int64)sample >> 32);\n }\n}\n\nstatic void drwav__ieee_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)\n{\n if (bytesPerSample == 4) {\n drwav_f32_to_s32(pOut, (const float*)pIn, totalSampleCount);\n return;\n } else if (bytesPerSample == 8) {\n drwav_f64_to_s32(pOut, (const double*)pIn, totalSampleCount);\n return;\n } else {\n /* Only supporting 32- and 64-bit float. Output silence in all other cases. Contributions welcome for 16-bit float. */\n DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));\n return;\n }\n}\n\n\nstatic drwav_uint64 drwav_read_pcm_frames_s32__pcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n drwav_uint32 bytesPerFrame;\n\n /* Fast path. */\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM && pWav->bitsPerSample == 32) {\n return drwav_read_pcm_frames(pWav, framesToRead, pBufferOut);\n }\n \n bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav__pcm_to_s32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s32__msadpcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)\n{\n /*\n We're just going to borrow the implementation from the drwav_read_s16() since ADPCM is a little bit more complicated than other formats and I don't\n want to duplicate that code.\n */\n drwav_uint64 totalFramesRead = 0;\n drwav_int16 samples16[2048];\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels), samples16);\n if (framesRead == 0) {\n break;\n }\n\n drwav_s16_to_s32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels)); /* <-- Safe cast because we're clamping to 2048. */\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s32__ima(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)\n{\n /*\n We're just going to borrow the implementation from the drwav_read_s16() since IMA-ADPCM is a little bit more complicated than other formats and I don't\n want to duplicate that code.\n */\n drwav_uint64 totalFramesRead = 0;\n drwav_int16 samples16[2048];\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels), samples16);\n if (framesRead == 0) {\n break;\n }\n\n drwav_s16_to_s32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels)); /* <-- Safe cast because we're clamping to 2048. */\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s32__ieee(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n\n drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav__ieee_to_s32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s32__alaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n\n drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav_alaw_to_s32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nstatic drwav_uint64 drwav_read_pcm_frames_s32__mulaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)\n{\n drwav_uint64 totalFramesRead;\n drwav_uint8 sampleData[4096];\n\n drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);\n if (bytesPerFrame == 0) {\n return 0;\n }\n\n totalFramesRead = 0;\n\n while (framesToRead > 0) {\n drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);\n if (framesRead == 0) {\n break;\n }\n\n drwav_mulaw_to_s32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));\n\n pBufferOut += framesRead*pWav->channels;\n framesToRead -= framesRead;\n totalFramesRead += framesRead;\n }\n\n return totalFramesRead;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s32(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)\n{\n if (pWav == NULL || framesToRead == 0) {\n return 0;\n }\n\n if (pBufferOut == NULL) {\n return drwav_read_pcm_frames(pWav, framesToRead, NULL);\n }\n\n /* Don't try to read more samples than can potentially fit in the output buffer. */\n if (framesToRead * pWav->channels * sizeof(drwav_int32) > DRWAV_SIZE_MAX) {\n framesToRead = DRWAV_SIZE_MAX / sizeof(drwav_int32) / pWav->channels;\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM) {\n return drwav_read_pcm_frames_s32__pcm(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {\n return drwav_read_pcm_frames_s32__msadpcm(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT) {\n return drwav_read_pcm_frames_s32__ieee(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW) {\n return drwav_read_pcm_frames_s32__alaw(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {\n return drwav_read_pcm_frames_s32__mulaw(pWav, framesToRead, pBufferOut);\n }\n\n if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {\n return drwav_read_pcm_frames_s32__ima(pWav, framesToRead, pBufferOut);\n }\n\n return 0;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s32le(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)\n{\n drwav_uint64 framesRead = drwav_read_pcm_frames_s32(pWav, framesToRead, pBufferOut);\n if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_FALSE) {\n drwav__bswap_samples_s32(pBufferOut, framesRead*pWav->channels);\n }\n\n return framesRead;\n}\n\nDRWAV_API drwav_uint64 drwav_read_pcm_frames_s32be(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)\n{\n drwav_uint64 framesRead = drwav_read_pcm_frames_s32(pWav, framesToRead, pBufferOut);\n if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_TRUE) {\n drwav__bswap_samples_s32(pBufferOut, framesRead*pWav->channels);\n }\n\n return framesRead;\n}\n\n\nDRWAV_API void drwav_u8_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = ((int)pIn[i] - 128) << 24;\n }\n}\n\nDRWAV_API void drwav_s16_to_s32(drwav_int32* pOut, const drwav_int16* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = pIn[i] << 16;\n }\n}\n\nDRWAV_API void drwav_s24_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n unsigned int s0 = pIn[i*3 + 0];\n unsigned int s1 = pIn[i*3 + 1];\n unsigned int s2 = pIn[i*3 + 2];\n\n drwav_int32 sample32 = (drwav_int32)((s0 << 8) | (s1 << 16) | (s2 << 24));\n *pOut++ = sample32;\n }\n}\n\nDRWAV_API void drwav_f32_to_s32(drwav_int32* pOut, const float* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = (drwav_int32)(2147483648.0 * pIn[i]);\n }\n}\n\nDRWAV_API void drwav_f64_to_s32(drwav_int32* pOut, const double* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = (drwav_int32)(2147483648.0 * pIn[i]);\n }\n}\n\nDRWAV_API void drwav_alaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i = 0; i < sampleCount; ++i) {\n *pOut++ = ((drwav_int32)drwav__alaw_to_s16(pIn[i])) << 16;\n }\n}\n\nDRWAV_API void drwav_mulaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)\n{\n size_t i;\n\n if (pOut == NULL || pIn == NULL) {\n return;\n }\n\n for (i= 0; i < sampleCount; ++i) {\n *pOut++ = ((drwav_int32)drwav__mulaw_to_s16(pIn[i])) << 16;\n }\n}\n\n\n\nstatic drwav_int16* drwav__read_pcm_frames_and_close_s16(drwav* pWav, unsigned int* channels, unsigned int* sampleRate, drwav_uint64* totalFrameCount)\n{\n drwav_uint64 sampleDataSize;\n drwav_int16* pSampleData;\n drwav_uint64 framesRead;\n\n DRWAV_ASSERT(pWav != NULL);\n\n sampleDataSize = pWav->totalPCMFrameCount * pWav->channels * sizeof(drwav_int16);\n if (sampleDataSize > DRWAV_SIZE_MAX) {\n drwav_uninit(pWav);\n return NULL; /* File's too big. */\n }\n\n pSampleData = (drwav_int16*)drwav__malloc_from_callbacks((size_t)sampleDataSize, &pWav->allocationCallbacks); /* <-- Safe cast due to the check above. */\n if (pSampleData == NULL) {\n drwav_uninit(pWav);\n return NULL; /* Failed to allocate memory. */\n }\n\n framesRead = drwav_read_pcm_frames_s16(pWav, (size_t)pWav->totalPCMFrameCount, pSampleData);\n if (framesRead != pWav->totalPCMFrameCount) {\n drwav__free_from_callbacks(pSampleData, &pWav->allocationCallbacks);\n drwav_uninit(pWav);\n return NULL; /* There was an error reading the samples. */\n }\n\n drwav_uninit(pWav);\n\n if (sampleRate) {\n *sampleRate = pWav->sampleRate;\n }\n if (channels) {\n *channels = pWav->channels;\n }\n if (totalFrameCount) {\n *totalFrameCount = pWav->totalPCMFrameCount;\n }\n\n return pSampleData;\n}\n\nstatic float* drwav__read_pcm_frames_and_close_f32(drwav* pWav, unsigned int* channels, unsigned int* sampleRate, drwav_uint64* totalFrameCount)\n{\n drwav_uint64 sampleDataSize;\n float* pSampleData;\n drwav_uint64 framesRead;\n\n DRWAV_ASSERT(pWav != NULL);\n\n sampleDataSize = pWav->totalPCMFrameCount * pWav->channels * sizeof(float);\n if (sampleDataSize > DRWAV_SIZE_MAX) {\n drwav_uninit(pWav);\n return NULL; /* File's too big. */\n }\n\n pSampleData = (float*)drwav__malloc_from_callbacks((size_t)sampleDataSize, &pWav->allocationCallbacks); /* <-- Safe cast due to the check above. */\n if (pSampleData == NULL) {\n drwav_uninit(pWav);\n return NULL; /* Failed to allocate memory. */\n }\n\n framesRead = drwav_read_pcm_frames_f32(pWav, (size_t)pWav->totalPCMFrameCount, pSampleData);\n if (framesRead != pWav->totalPCMFrameCount) {\n drwav__free_from_callbacks(pSampleData, &pWav->allocationCallbacks);\n drwav_uninit(pWav);\n return NULL; /* There was an error reading the samples. */\n }\n\n drwav_uninit(pWav);\n\n if (sampleRate) {\n *sampleRate = pWav->sampleRate;\n }\n if (channels) {\n *channels = pWav->channels;\n }\n if (totalFrameCount) {\n *totalFrameCount = pWav->totalPCMFrameCount;\n }\n\n return pSampleData;\n}\n\nstatic drwav_int32* drwav__read_pcm_frames_and_close_s32(drwav* pWav, unsigned int* channels, unsigned int* sampleRate, drwav_uint64* totalFrameCount)\n{\n drwav_uint64 sampleDataSize;\n drwav_int32* pSampleData;\n drwav_uint64 framesRead;\n\n DRWAV_ASSERT(pWav != NULL);\n\n sampleDataSize = pWav->totalPCMFrameCount * pWav->channels * sizeof(drwav_int32);\n if (sampleDataSize > DRWAV_SIZE_MAX) {\n drwav_uninit(pWav);\n return NULL; /* File's too big. */\n }\n\n pSampleData = (drwav_int32*)drwav__malloc_from_callbacks((size_t)sampleDataSize, &pWav->allocationCallbacks); /* <-- Safe cast due to the check above. */\n if (pSampleData == NULL) {\n drwav_uninit(pWav);\n return NULL; /* Failed to allocate memory. */\n }\n\n framesRead = drwav_read_pcm_frames_s32(pWav, (size_t)pWav->totalPCMFrameCount, pSampleData);\n if (framesRead != pWav->totalPCMFrameCount) {\n drwav__free_from_callbacks(pSampleData, &pWav->allocationCallbacks);\n drwav_uninit(pWav);\n return NULL; /* There was an error reading the samples. */\n }\n\n drwav_uninit(pWav);\n\n if (sampleRate) {\n *sampleRate = pWav->sampleRate;\n }\n if (channels) {\n *channels = pWav->channels;\n }\n if (totalFrameCount) {\n *totalFrameCount = pWav->totalPCMFrameCount;\n }\n\n return pSampleData;\n}\n\n\n\nDRWAV_API drwav_int16* drwav_open_and_read_pcm_frames_s16(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init(&wav, onRead, onSeek, pUserData, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\nDRWAV_API float* drwav_open_and_read_pcm_frames_f32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init(&wav, onRead, onSeek, pUserData, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\nDRWAV_API drwav_int32* drwav_open_and_read_pcm_frames_s32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init(&wav, onRead, onSeek, pUserData, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\n#ifndef DR_WAV_NO_STDIO\nDRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init_file(&wav, filename, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\nDRWAV_API float* drwav_open_file_and_read_pcm_frames_f32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init_file(&wav, filename, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\nDRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init_file(&wav, filename, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\n\nDRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init_file_w(&wav, filename, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\nDRWAV_API float* drwav_open_file_and_read_pcm_frames_f32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init_file_w(&wav, filename, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\nDRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init_file_w(&wav, filename, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n#endif\n\nDRWAV_API drwav_int16* drwav_open_memory_and_read_pcm_frames_s16(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init_memory(&wav, data, dataSize, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\nDRWAV_API float* drwav_open_memory_and_read_pcm_frames_f32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init_memory(&wav, data, dataSize, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n\nDRWAV_API drwav_int32* drwav_open_memory_and_read_pcm_frames_s32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n drwav wav;\n\n if (channelsOut) {\n *channelsOut = 0;\n }\n if (sampleRateOut) {\n *sampleRateOut = 0;\n }\n if (totalFrameCountOut) {\n *totalFrameCountOut = 0;\n }\n\n if (!drwav_init_memory(&wav, data, dataSize, pAllocationCallbacks)) {\n return NULL;\n }\n\n return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);\n}\n#endif /* DR_WAV_NO_CONVERSION_API */\n\n\nDRWAV_API void drwav_free(void* p, const drwav_allocation_callbacks* pAllocationCallbacks)\n{\n if (pAllocationCallbacks != NULL) {\n drwav__free_from_callbacks(p, pAllocationCallbacks);\n } else {\n drwav__free_default(p, NULL);\n }\n}\n\nDRWAV_API drwav_uint16 drwav_bytes_to_u16(const drwav_uint8* data)\n{\n return drwav__bytes_to_u16(data);\n}\n\nDRWAV_API drwav_int16 drwav_bytes_to_s16(const drwav_uint8* data)\n{\n return drwav__bytes_to_s16(data);\n}\n\nDRWAV_API drwav_uint32 drwav_bytes_to_u32(const drwav_uint8* data)\n{\n return drwav__bytes_to_u32(data);\n}\n\nDRWAV_API drwav_int32 drwav_bytes_to_s32(const drwav_uint8* data)\n{\n return drwav__bytes_to_s32(data);\n}\n\nDRWAV_API drwav_uint64 drwav_bytes_to_u64(const drwav_uint8* data)\n{\n return drwav__bytes_to_u64(data);\n}\n\nDRWAV_API drwav_int64 drwav_bytes_to_s64(const drwav_uint8* data)\n{\n return drwav__bytes_to_s64(data);\n}\n\n\nDRWAV_API drwav_bool32 drwav_guid_equal(const drwav_uint8 a[16], const drwav_uint8 b[16])\n{\n return drwav__guid_equal(a, b);\n}\n\nDRWAV_API drwav_bool32 drwav_fourcc_equal(const drwav_uint8* a, const char* b)\n{\n return drwav__fourcc_equal(a, b);\n}\n\n#endif /* dr_wav_c */\n#endif /* DR_WAV_IMPLEMENTATION */\n\n/*\nRELEASE NOTES - v0.11.0\n=======================\nVersion 0.11.0 has breaking API changes.\n\nImproved Client-Defined Memory Allocation\n-----------------------------------------\nThe main change with this release is the addition of a more flexible way of implementing custom memory allocation routines. The\nexisting system of DRWAV_MALLOC, DRWAV_REALLOC and DRWAV_FREE are still in place and will be used by default when no custom\nallocation callbacks are specified.\n\nTo use the new system, you pass in a pointer to a drwav_allocation_callbacks object to drwav_init() and family, like this:\n\n void* my_malloc(size_t sz, void* pUserData)\n {\n return malloc(sz);\n }\n void* my_realloc(void* p, size_t sz, void* pUserData)\n {\n return realloc(p, sz);\n }\n void my_free(void* p, void* pUserData)\n {\n free(p);\n }\n\n ...\n\n drwav_allocation_callbacks allocationCallbacks;\n allocationCallbacks.pUserData = &myData;\n allocationCallbacks.onMalloc = my_malloc;\n allocationCallbacks.onRealloc = my_realloc;\n allocationCallbacks.onFree = my_free;\n drwav_init_file(&wav, \"my_file.wav\", &allocationCallbacks);\n\nThe advantage of this new system is that it allows you to specify user data which will be passed in to the allocation routines.\n\nPassing in null for the allocation callbacks object will cause dr_wav to use defaults which is the same as DRWAV_MALLOC,\nDRWAV_REALLOC and DRWAV_FREE and the equivalent of how it worked in previous versions.\n\nEvery API that opens a drwav object now takes this extra parameter. These include the following:\n\n drwav_init()\n drwav_init_ex()\n drwav_init_file()\n drwav_init_file_ex()\n drwav_init_file_w()\n drwav_init_file_w_ex()\n drwav_init_memory()\n drwav_init_memory_ex()\n drwav_init_write()\n drwav_init_write_sequential()\n drwav_init_write_sequential_pcm_frames()\n drwav_init_file_write()\n drwav_init_file_write_sequential()\n drwav_init_file_write_sequential_pcm_frames()\n drwav_init_file_write_w()\n drwav_init_file_write_sequential_w()\n drwav_init_file_write_sequential_pcm_frames_w()\n drwav_init_memory_write()\n drwav_init_memory_write_sequential()\n drwav_init_memory_write_sequential_pcm_frames()\n drwav_open_and_read_pcm_frames_s16()\n drwav_open_and_read_pcm_frames_f32()\n drwav_open_and_read_pcm_frames_s32()\n drwav_open_file_and_read_pcm_frames_s16()\n drwav_open_file_and_read_pcm_frames_f32()\n drwav_open_file_and_read_pcm_frames_s32()\n drwav_open_file_and_read_pcm_frames_s16_w()\n drwav_open_file_and_read_pcm_frames_f32_w()\n drwav_open_file_and_read_pcm_frames_s32_w()\n drwav_open_memory_and_read_pcm_frames_s16()\n drwav_open_memory_and_read_pcm_frames_f32()\n drwav_open_memory_and_read_pcm_frames_s32()\n\nEndian Improvements\n-------------------\nPreviously, the following APIs returned little-endian audio data. These now return native-endian data. This improves compatibility\non big-endian architectures.\n\n drwav_read_pcm_frames()\n drwav_read_pcm_frames_s16()\n drwav_read_pcm_frames_s32()\n drwav_read_pcm_frames_f32()\n drwav_open_and_read_pcm_frames_s16()\n drwav_open_and_read_pcm_frames_s32()\n drwav_open_and_read_pcm_frames_f32()\n drwav_open_file_and_read_pcm_frames_s16()\n drwav_open_file_and_read_pcm_frames_s32()\n drwav_open_file_and_read_pcm_frames_f32()\n drwav_open_file_and_read_pcm_frames_s16_w()\n drwav_open_file_and_read_pcm_frames_s32_w()\n drwav_open_file_and_read_pcm_frames_f32_w()\n drwav_open_memory_and_read_pcm_frames_s16()\n drwav_open_memory_and_read_pcm_frames_s32()\n drwav_open_memory_and_read_pcm_frames_f32()\n\nAPIs have been added to give you explicit control over whether or not audio data is read or written in big- or little-endian byte\norder:\n\n drwav_read_pcm_frames_le()\n drwav_read_pcm_frames_be()\n drwav_read_pcm_frames_s16le()\n drwav_read_pcm_frames_s16be()\n drwav_read_pcm_frames_f32le()\n drwav_read_pcm_frames_f32be()\n drwav_read_pcm_frames_s32le()\n drwav_read_pcm_frames_s32be()\n drwav_write_pcm_frames_le()\n drwav_write_pcm_frames_be()\n\nRemoved APIs\n------------\nThe following APIs were deprecated in version 0.10.0 and have now been removed:\n\n drwav_open()\n drwav_open_ex()\n drwav_open_write()\n drwav_open_write_sequential()\n drwav_open_file()\n drwav_open_file_ex()\n drwav_open_file_write()\n drwav_open_file_write_sequential()\n drwav_open_memory()\n drwav_open_memory_ex()\n drwav_open_memory_write()\n drwav_open_memory_write_sequential()\n drwav_close()\n\n\n\nRELEASE NOTES - v0.10.0\n=======================\nVersion 0.10.0 has breaking API changes. There are no significant bug fixes in this release, so if you are affected you do\nnot need to upgrade.\n\nRemoved APIs\n------------\nThe following APIs were deprecated in version 0.9.0 and have been completely removed in version 0.10.0:\n\n drwav_read()\n drwav_read_s16()\n drwav_read_f32()\n drwav_read_s32()\n drwav_seek_to_sample()\n drwav_write()\n drwav_open_and_read_s16()\n drwav_open_and_read_f32()\n drwav_open_and_read_s32()\n drwav_open_file_and_read_s16()\n drwav_open_file_and_read_f32()\n drwav_open_file_and_read_s32()\n drwav_open_memory_and_read_s16()\n drwav_open_memory_and_read_f32()\n drwav_open_memory_and_read_s32()\n drwav::totalSampleCount\n\nSee release notes for version 0.9.0 at the bottom of this file for replacement APIs.\n\nDeprecated APIs\n---------------\nThe following APIs have been deprecated. There is a confusing and completely arbitrary difference between drwav_init*() and\ndrwav_open*(), where drwav_init*() initializes a pre-allocated drwav object, whereas drwav_open*() will first allocated a\ndrwav object on the heap and then initialize it. drwav_open*() has been deprecated which means you must now use a pre-\nallocated drwav object with drwav_init*(). If you need the previous functionality, you can just do a malloc() followed by\na called to one of the drwav_init*() APIs.\n\n drwav_open()\n drwav_open_ex()\n drwav_open_write()\n drwav_open_write_sequential()\n drwav_open_file()\n drwav_open_file_ex()\n drwav_open_file_write()\n drwav_open_file_write_sequential()\n drwav_open_memory()\n drwav_open_memory_ex()\n drwav_open_memory_write()\n drwav_open_memory_write_sequential()\n drwav_close()\n\nThese APIs will be removed completely in a future version. The rationale for this change is to remove confusion between the\ntwo different ways to initialize a drwav object.\n*/\n\n/*\nREVISION HISTORY\n================\nv0.12.16 - 2020-12-02\n - Fix a bug when trying to read more bytes than can fit in a size_t.\n\nv0.12.15 - 2020-11-21\n - Fix compilation with OpenWatcom.\n\nv0.12.14 - 2020-11-13\n - Minor code clean up.\n\nv0.12.13 - 2020-11-01\n - Improve compiler support for older versions of GCC.\n\nv0.12.12 - 2020-09-28\n - Add support for RF64.\n - Fix a bug in writing mode where the size of the RIFF chunk incorrectly includes the header section.\n\nv0.12.11 - 2020-09-08\n - Fix a compilation error on older compilers.\n\nv0.12.10 - 2020-08-24\n - Fix a bug when seeking with ADPCM formats.\n\nv0.12.9 - 2020-08-02\n - Simplify sized types.\n\nv0.12.8 - 2020-07-25\n - Fix a compilation warning.\n\nv0.12.7 - 2020-07-15\n - Fix some bugs on big-endian architectures.\n - Fix an error in s24 to f32 conversion.\n\nv0.12.6 - 2020-06-23\n - Change drwav_read_*() to allow NULL to be passed in as the output buffer which is equivalent to a forward seek.\n - Fix a buffer overflow when trying to decode invalid IMA-ADPCM files.\n - Add include guard for the implementation section.\n\nv0.12.5 - 2020-05-27\n - Minor documentation fix.\n\nv0.12.4 - 2020-05-16\n - Replace assert() with DRWAV_ASSERT().\n - Add compile-time and run-time version querying.\n - DRWAV_VERSION_MINOR\n - DRWAV_VERSION_MAJOR\n - DRWAV_VERSION_REVISION\n - DRWAV_VERSION_STRING\n - drwav_version()\n - drwav_version_string()\n\nv0.12.3 - 2020-04-30\n - Fix compilation errors with VC6.\n\nv0.12.2 - 2020-04-21\n - Fix a bug where drwav_init_file() does not close the file handle after attempting to load an erroneous file.\n\nv0.12.1 - 2020-04-13\n - Fix some pedantic warnings.\n\nv0.12.0 - 2020-04-04\n - API CHANGE: Add container and format parameters to the chunk callback.\n - Minor documentation updates.\n\nv0.11.5 - 2020-03-07\n - Fix compilation error with Visual Studio .NET 2003.\n\nv0.11.4 - 2020-01-29\n - Fix some static analysis warnings.\n - Fix a bug when reading f32 samples from an A-law encoded stream.\n\nv0.11.3 - 2020-01-12\n - Minor changes to some f32 format conversion routines.\n - Minor bug fix for ADPCM conversion when end of file is reached.\n\nv0.11.2 - 2019-12-02\n - Fix a possible crash when using custom memory allocators without a custom realloc() implementation.\n - Fix an integer overflow bug.\n - Fix a null pointer dereference bug.\n - Add limits to sample rate, channels and bits per sample to tighten up some validation.\n\nv0.11.1 - 2019-10-07\n - Internal code clean up.\n\nv0.11.0 - 2019-10-06\n - API CHANGE: Add support for user defined memory allocation routines. This system allows the program to specify their own memory allocation\n routines with a user data pointer for client-specific contextual data. This adds an extra parameter to the end of the following APIs:\n - drwav_init()\n - drwav_init_ex()\n - drwav_init_file()\n - drwav_init_file_ex()\n - drwav_init_file_w()\n - drwav_init_file_w_ex()\n - drwav_init_memory()\n - drwav_init_memory_ex()\n - drwav_init_write()\n - drwav_init_write_sequential()\n - drwav_init_write_sequential_pcm_frames()\n - drwav_init_file_write()\n - drwav_init_file_write_sequential()\n - drwav_init_file_write_sequential_pcm_frames()\n - drwav_init_file_write_w()\n - drwav_init_file_write_sequential_w()\n - drwav_init_file_write_sequential_pcm_frames_w()\n - drwav_init_memory_write()\n - drwav_init_memory_write_sequential()\n - drwav_init_memory_write_sequential_pcm_frames()\n - drwav_open_and_read_pcm_frames_s16()\n - drwav_open_and_read_pcm_frames_f32()\n - drwav_open_and_read_pcm_frames_s32()\n - drwav_open_file_and_read_pcm_frames_s16()\n - drwav_open_file_and_read_pcm_frames_f32()\n - drwav_open_file_and_read_pcm_frames_s32()\n - drwav_open_file_and_read_pcm_frames_s16_w()\n - drwav_open_file_and_read_pcm_frames_f32_w()\n - drwav_open_file_and_read_pcm_frames_s32_w()\n - drwav_open_memory_and_read_pcm_frames_s16()\n - drwav_open_memory_and_read_pcm_frames_f32()\n - drwav_open_memory_and_read_pcm_frames_s32()\n Set this extra parameter to NULL to use defaults which is the same as the previous behaviour. Setting this NULL will use\n DRWAV_MALLOC, DRWAV_REALLOC and DRWAV_FREE.\n - Add support for reading and writing PCM frames in an explicit endianness. New APIs:\n - drwav_read_pcm_frames_le()\n - drwav_read_pcm_frames_be()\n - drwav_read_pcm_frames_s16le()\n - drwav_read_pcm_frames_s16be()\n - drwav_read_pcm_frames_f32le()\n - drwav_read_pcm_frames_f32be()\n - drwav_read_pcm_frames_s32le()\n - drwav_read_pcm_frames_s32be()\n - drwav_write_pcm_frames_le()\n - drwav_write_pcm_frames_be()\n - Remove deprecated APIs.\n - API CHANGE: The following APIs now return native-endian data. Previously they returned little-endian data.\n - drwav_read_pcm_frames()\n - drwav_read_pcm_frames_s16()\n - drwav_read_pcm_frames_s32()\n - drwav_read_pcm_frames_f32()\n - drwav_open_and_read_pcm_frames_s16()\n - drwav_open_and_read_pcm_frames_s32()\n - drwav_open_and_read_pcm_frames_f32()\n - drwav_open_file_and_read_pcm_frames_s16()\n - drwav_open_file_and_read_pcm_frames_s32()\n - drwav_open_file_and_read_pcm_frames_f32()\n - drwav_open_file_and_read_pcm_frames_s16_w()\n - drwav_open_file_and_read_pcm_frames_s32_w()\n - drwav_open_file_and_read_pcm_frames_f32_w()\n - drwav_open_memory_and_read_pcm_frames_s16()\n - drwav_open_memory_and_read_pcm_frames_s32()\n - drwav_open_memory_and_read_pcm_frames_f32()\n\nv0.10.1 - 2019-08-31\n - Correctly handle partial trailing ADPCM blocks.\n\nv0.10.0 - 2019-08-04\n - Remove deprecated APIs.\n - Add wchar_t variants for file loading APIs:\n drwav_init_file_w()\n drwav_init_file_ex_w()\n drwav_init_file_write_w()\n drwav_init_file_write_sequential_w()\n - Add drwav_target_write_size_bytes() which calculates the total size in bytes of a WAV file given a format and sample count.\n - Add APIs for specifying the PCM frame count instead of the sample count when opening in sequential write mode:\n drwav_init_write_sequential_pcm_frames()\n drwav_init_file_write_sequential_pcm_frames()\n drwav_init_file_write_sequential_pcm_frames_w()\n drwav_init_memory_write_sequential_pcm_frames()\n - Deprecate drwav_open*() and drwav_close():\n drwav_open()\n drwav_open_ex()\n drwav_open_write()\n drwav_open_write_sequential()\n drwav_open_file()\n drwav_open_file_ex()\n drwav_open_file_write()\n drwav_open_file_write_sequential()\n drwav_open_memory()\n drwav_open_memory_ex()\n drwav_open_memory_write()\n drwav_open_memory_write_sequential()\n drwav_close()\n - Minor documentation updates.\n\nv0.9.2 - 2019-05-21\n - Fix warnings.\n\nv0.9.1 - 2019-05-05\n - Add support for C89.\n - Change license to choice of public domain or MIT-0.\n\nv0.9.0 - 2018-12-16\n - API CHANGE: Add new reading APIs for reading by PCM frames instead of samples. Old APIs have been deprecated and\n will be removed in v0.10.0. Deprecated APIs and their replacements:\n drwav_read() -> drwav_read_pcm_frames()\n drwav_read_s16() -> drwav_read_pcm_frames_s16()\n drwav_read_f32() -> drwav_read_pcm_frames_f32()\n drwav_read_s32() -> drwav_read_pcm_frames_s32()\n drwav_seek_to_sample() -> drwav_seek_to_pcm_frame()\n drwav_write() -> drwav_write_pcm_frames()\n drwav_open_and_read_s16() -> drwav_open_and_read_pcm_frames_s16()\n drwav_open_and_read_f32() -> drwav_open_and_read_pcm_frames_f32()\n drwav_open_and_read_s32() -> drwav_open_and_read_pcm_frames_s32()\n drwav_open_file_and_read_s16() -> drwav_open_file_and_read_pcm_frames_s16()\n drwav_open_file_and_read_f32() -> drwav_open_file_and_read_pcm_frames_f32()\n drwav_open_file_and_read_s32() -> drwav_open_file_and_read_pcm_frames_s32()\n drwav_open_memory_and_read_s16() -> drwav_open_memory_and_read_pcm_frames_s16()\n drwav_open_memory_and_read_f32() -> drwav_open_memory_and_read_pcm_frames_f32()\n drwav_open_memory_and_read_s32() -> drwav_open_memory_and_read_pcm_frames_s32()\n drwav::totalSampleCount -> drwav::totalPCMFrameCount\n - API CHANGE: Rename drwav_open_and_read_file_*() to drwav_open_file_and_read_*().\n - API CHANGE: Rename drwav_open_and_read_memory_*() to drwav_open_memory_and_read_*().\n - Add built-in support for smpl chunks.\n - Add support for firing a callback for each chunk in the file at initialization time.\n - This is enabled through the drwav_init_ex(), etc. family of APIs.\n - Handle invalid FMT chunks more robustly.\n\nv0.8.5 - 2018-09-11\n - Const correctness.\n - Fix a potential stack overflow.\n\nv0.8.4 - 2018-08-07\n - Improve 64-bit detection.\n\nv0.8.3 - 2018-08-05\n - Fix C++ build on older versions of GCC.\n\nv0.8.2 - 2018-08-02\n - Fix some big-endian bugs.\n\nv0.8.1 - 2018-06-29\n - Add support for sequential writing APIs.\n - Disable seeking in write mode.\n - Fix bugs with Wave64.\n - Fix typos.\n\nv0.8 - 2018-04-27\n - Bug fix.\n - Start using major.minor.revision versioning.\n\nv0.7f - 2018-02-05\n - Restrict ADPCM formats to a maximum of 2 channels.\n\nv0.7e - 2018-02-02\n - Fix a crash.\n\nv0.7d - 2018-02-01\n - Fix a crash.\n\nv0.7c - 2018-02-01\n - Set drwav.bytesPerSample to 0 for all compressed formats.\n - Fix a crash when reading 16-bit floating point WAV files. In this case dr_wav will output silence for\n all format conversion reading APIs (*_s16, *_s32, *_f32 APIs).\n - Fix some divide-by-zero errors.\n\nv0.7b - 2018-01-22\n - Fix errors with seeking of compressed formats.\n - Fix compilation error when DR_WAV_NO_CONVERSION_API\n\nv0.7a - 2017-11-17\n - Fix some GCC warnings.\n\nv0.7 - 2017-11-04\n - Add writing APIs.\n\nv0.6 - 2017-08-16\n - API CHANGE: Rename dr_* types to drwav_*.\n - Add support for custom implementations of malloc(), realloc(), etc.\n - Add support for Microsoft ADPCM.\n - Add support for IMA ADPCM (DVI, format code 0x11).\n - Optimizations to drwav_read_s16().\n - Bug fixes.\n\nv0.5g - 2017-07-16\n - Change underlying type for booleans to unsigned.\n\nv0.5f - 2017-04-04\n - Fix a minor bug with drwav_open_and_read_s16() and family.\n\nv0.5e - 2016-12-29\n - Added support for reading samples as signed 16-bit integers. Use the _s16() family of APIs for this.\n - Minor fixes to documentation.\n\nv0.5d - 2016-12-28\n - Use drwav_int* and drwav_uint* sized types to improve compiler support.\n\nv0.5c - 2016-11-11\n - Properly handle JUNK chunks that come before the FMT chunk.\n\nv0.5b - 2016-10-23\n - A minor change to drwav_bool8 and drwav_bool32 types.\n\nv0.5a - 2016-10-11\n - Fixed a bug with drwav_open_and_read() and family due to incorrect argument ordering.\n - Improve A-law and mu-law efficiency.\n\nv0.5 - 2016-09-29\n - API CHANGE. Swap the order of \"channels\" and \"sampleRate\" parameters in drwav_open_and_read*(). Rationale for this is to\n keep it consistent with dr_audio and dr_flac.\n\nv0.4b - 2016-09-18\n - Fixed a typo in documentation.\n\nv0.4a - 2016-09-18\n - Fixed a typo.\n - Change date format to ISO 8601 (YYYY-MM-DD)\n\nv0.4 - 2016-07-13\n - API CHANGE. Make onSeek consistent with dr_flac.\n - API CHANGE. Rename drwav_seek() to drwav_seek_to_sample() for clarity and consistency with dr_flac.\n - Added support for Sony Wave64.\n\nv0.3a - 2016-05-28\n - API CHANGE. Return drwav_bool32 instead of int in onSeek callback.\n - Fixed a memory leak.\n\nv0.3 - 2016-05-22\n - Lots of API changes for consistency.\n\nv0.2a - 2016-05-16\n - Fixed Linux/GCC build.\n\nv0.2 - 2016-05-11\n - Added support for reading data as signed 32-bit PCM for consistency with dr_flac.\n\nv0.1a - 2016-05-07\n - Fixed a bug in drwav_open_file() where the file handle would not be closed if the loader failed to initialize.\n\nv0.1 - 2016-05-04\n - Initial versioned release.\n*/\n\n/*\nThis software is available as a choice of the following licenses. Choose\nwhichever you prefer.\n\n===============================================================================\nALTERNATIVE 1 - Public Domain (www.unlicense.org)\n===============================================================================\nThis is free and unencumbered software released into the public domain.\n\nAnyone is free to copy, modify, publish, use, compile, sell, or distribute this\nsoftware, either in source code form or as a compiled binary, for any purpose,\ncommercial or non-commercial, and by any means.\n\nIn jurisdictions that recognize copyright laws, the author or authors of this\nsoftware dedicate any and all copyright interest in the software to the public\ndomain. We make this dedication for the benefit of the public at large and to\nthe detriment of our heirs and successors. We intend this dedication to be an\novert act of relinquishment in perpetuity of all present and future rights to\nthis software under copyright law.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN\nACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION\nWITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.\n\nFor more information, please refer to \n\n===============================================================================\nALTERNATIVE 2 - MIT No Attribution\n===============================================================================\nCopyright 2020 David Reid\n\nPermission is hereby granted, free of charge, to any person obtaining a copy of\nthis software and associated documentation files (the \"Software\"), to deal in\nthe Software without restriction, including without limitation the rights to\nuse, copy, modify, merge, publish, distribute, sublicense, and/or sell copies\nof the Software, and to permit persons to whom the Software is furnished to do\nso.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\nSOFTWARE.\n*/\n"
},
{
"path": "Examples/OldMain/main.cpp",
"content": "#include \"whisper.h\"\n\n// third-party utilities\n// use your favorite implementations\n#define DR_WAV_IMPLEMENTATION\n#include \"dr_wav.h\"\n\n#include \n#include \n#include \n#include \n#include \n#include \n\n// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]\n// Lowest is red, middle is yellow, highest is green.\nconst std::vector k_colors = {\n \"\\033[38;5;196m\", \"\\033[38;5;202m\", \"\\033[38;5;208m\", \"\\033[38;5;214m\", \"\\033[38;5;220m\",\n \"\\033[38;5;226m\", \"\\033[38;5;190m\", \"\\033[38;5;154m\", \"\\033[38;5;118m\", \"\\033[38;5;82m\",\n};\n\n// 500 -> 00:05.000\n// 6000 -> 01:00.000\nstd::string to_timestamp(int64_t t, bool comma = false) {\n int64_t msec = t * 10;\n int64_t hr = msec / (1000 * 60 * 60);\n msec = msec - hr * (1000 * 60 * 60);\n int64_t min = msec / (1000 * 60);\n msec = msec - min * (1000 * 60);\n int64_t sec = msec / 1000;\n msec = msec - sec * 1000;\n\n char buf[32];\n snprintf(buf, sizeof(buf), \"%02d:%02d:%02d%s%03d\", (int) hr, (int) min, (int) sec, comma ? \",\" : \".\", (int) msec);\n\n return std::string(buf);\n}\n\nint timestamp_to_sample(int64_t t, int n_samples) {\n return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));\n}\n\n// helper function to replace substrings\nvoid replace_all(std::string & s, const std::string & search, const std::string & replace) {\n for (size_t pos = 0; ; pos += replace.length()) {\n pos = s.find(search, pos);\n if (pos == std::string::npos) break;\n s.erase(pos, search.length());\n s.insert(pos, replace);\n }\n}\n\n// command-line parameters\nstruct whisper_params {\n int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());\n int32_t n_processors = 1;\n int32_t offset_t_ms = 0;\n int32_t offset_n = 0;\n int32_t duration_ms = 0;\n int32_t max_context = -1;\n int32_t max_len = 0;\n\n float word_thold = 0.01f;\n\n bool speed_up = false;\n bool translate = false;\n bool diarize = false;\n bool output_txt = false;\n bool output_vtt = false;\n bool output_srt = false;\n bool output_wts = false;\n bool print_special = false;\n bool print_colors = false;\n bool print_progress = false;\n bool no_timestamps = false;\n\n std::string language = \"en\";\n std::string prompt;\n std::string model = \"models/ggml-base.en.bin\";\n\n std::vector fname_inp = {};\n};\n\nvoid whisper_print_usage(int argc, char ** argv, const whisper_params & params);\n\nbool whisper_params_parse(int argc, char ** argv, whisper_params & params) {\n for (int i = 1; i < argc; i++) {\n std::string arg = argv[i];\n\n if (arg[0] != '-') {\n params.fname_inp.push_back(arg);\n continue;\n }\n\n if (arg == \"-h\" || arg == \"--help\") {\n whisper_print_usage(argc, argv, params);\n exit(0);\n }\n else if (arg == \"-t\" || arg == \"--threads\") { params.n_threads = std::stoi(argv[++i]); }\n else if (arg == \"-p\" || arg == \"--processors\") { params.n_processors = std::stoi(argv[++i]); }\n else if (arg == \"-ot\" || arg == \"--offset-t\") { params.offset_t_ms = std::stoi(argv[++i]); }\n else if (arg == \"-on\" || arg == \"--offset-n\") { params.offset_n = std::stoi(argv[++i]); }\n else if (arg == \"-d\" || arg == \"--duration\") { params.duration_ms = std::stoi(argv[++i]); }\n else if (arg == \"-mc\" || arg == \"--max-context\") { params.max_context = std::stoi(argv[++i]); }\n else if (arg == \"-ml\" || arg == \"--max-len\") { params.max_len = std::stoi(argv[++i]); }\n else if (arg == \"-wt\" || arg == \"--word-thold\") { params.word_thold = std::stof(argv[++i]); }\n else if (arg == \"-su\" || arg == \"--speed-up\") { params.speed_up = true; }\n else if (arg == \"-tr\" || arg == \"--translate\") { params.translate = true; }\n else if (arg == \"-di\" || arg == \"--diarize\") { params.diarize = true; }\n else if (arg == \"-otxt\" || arg == \"--output-txt\") { params.output_txt = true; }\n else if (arg == \"-ovtt\" || arg == \"--output-vtt\") { params.output_vtt = true; }\n else if (arg == \"-osrt\" || arg == \"--output-srt\") { params.output_srt = true; }\n else if (arg == \"-owts\" || arg == \"--output-words\") { params.output_wts = true; }\n else if (arg == \"-ps\" || arg == \"--print-special\") { params.print_special = true; }\n else if (arg == \"-pc\" || arg == \"--print-colors\") { params.print_colors = true; }\n else if (arg == \"-pp\" || arg == \"--print-progress\") { params.print_progress = true; }\n else if (arg == \"-nt\" || arg == \"--no-timestamps\") { params.no_timestamps = true; }\n else if (arg == \"-l\" || arg == \"--language\") { params.language = argv[++i]; }\n else if ( arg == \"--prompt\") { params.prompt = argv[++i]; }\n else if (arg == \"-m\" || arg == \"--model\") { params.model = argv[++i]; }\n else if (arg == \"-f\" || arg == \"--file\") { params.fname_inp.emplace_back(argv[++i]); }\n else {\n fprintf(stderr, \"error: unknown argument: %s\\n\", arg.c_str());\n whisper_print_usage(argc, argv, params);\n exit(0);\n }\n }\n\n return true;\n}\n\nvoid whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {\n fprintf(stderr, \"\\n\");\n fprintf(stderr, \"usage: %s [options] file0.wav file1.wav ...\\n\", argv[0]);\n fprintf(stderr, \"\\n\");\n fprintf(stderr, \"options:\\n\");\n fprintf(stderr, \" -h, --help [default] show this help message and exit\\n\");\n fprintf(stderr, \" -t N, --threads N [%-7d] number of threads to use during computation\\n\", params.n_threads);\n fprintf(stderr, \" -p N, --processors N [%-7d] number of processors to use during computation\\n\", params.n_processors);\n fprintf(stderr, \" -ot N, --offset-t N [%-7d] time offset in milliseconds\\n\", params.offset_t_ms);\n fprintf(stderr, \" -on N, --offset-n N [%-7d] segment index offset\\n\", params.offset_n);\n fprintf(stderr, \" -d N, --duration N [%-7d] duration of audio to process in milliseconds\\n\", params.duration_ms);\n fprintf(stderr, \" -mc N, --max-context N [%-7d] maximum number of text context tokens to store\\n\", params.max_context);\n fprintf(stderr, \" -ml N, --max-len N [%-7d] maximum segment length in characters\\n\", params.max_len);\n fprintf(stderr, \" -wt N, --word-thold N [%-7.2f] word timestamp probability threshold\\n\", params.word_thold);\n fprintf(stderr, \" -su, --speed-up [%-7s] speed up audio by x2 (reduced accuracy)\\n\", params.speed_up ? \"true\" : \"false\");\n fprintf(stderr, \" -tr, --translate [%-7s] translate from source language to english\\n\", params.translate ? \"true\" : \"false\");\n fprintf(stderr, \" -di, --diarize [%-7s] stereo audio diarization\\n\", params.diarize ? \"true\" : \"false\");\n fprintf(stderr, \" -otxt, --output-txt [%-7s] output result in a text file\\n\", params.output_txt ? \"true\" : \"false\");\n fprintf(stderr, \" -ovtt, --output-vtt [%-7s] output result in a vtt file\\n\", params.output_vtt ? \"true\" : \"false\");\n fprintf(stderr, \" -osrt, --output-srt [%-7s] output result in a srt file\\n\", params.output_srt ? \"true\" : \"false\");\n fprintf(stderr, \" -owts, --output-words [%-7s] output script for generating karaoke video\\n\", params.output_wts ? \"true\" : \"false\");\n fprintf(stderr, \" -ps, --print-special [%-7s] print special tokens\\n\", params.print_special ? \"true\" : \"false\");\n fprintf(stderr, \" -pc, --print-colors [%-7s] print colors\\n\", params.print_colors ? \"true\" : \"false\");\n fprintf(stderr, \" -pp, --print-progress [%-7s] print progress\\n\", params.print_progress ? \"true\" : \"false\");\n fprintf(stderr, \" -nt, --no-timestamps [%-7s] do not print timestamps\\n\", params.no_timestamps ? \"false\" : \"true\");\n fprintf(stderr, \" -l LANG, --language LANG [%-7s] spoken language ('auto' for auto-detect)\\n\", params.language.c_str());\n fprintf(stderr, \" --prompt PROMPT [%-7s] initial prompt\\n\", params.prompt.c_str());\n fprintf(stderr, \" -m FNAME, --model FNAME [%-7s] model path\\n\", params.model.c_str());\n fprintf(stderr, \" -f FNAME, --file FNAME [%-7s] input WAV file path\\n\", \"\");\n fprintf(stderr, \"\\n\");\n}\n\nstruct whisper_print_user_data {\n const whisper_params * params;\n\n const std::vector> * pcmf32s;\n};\n\nvoid whisper_print_segment_callback(struct whisper_context * ctx, int n_new, void * user_data) {\n const auto & params = *((whisper_print_user_data *) user_data)->params;\n const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;\n\n const int n_segments = whisper_full_n_segments(ctx);\n\n // print the last n_new segments\n const int s0 = n_segments - n_new;\n if (s0 == 0) {\n printf(\"\\n\");\n }\n\n for (int i = s0; i < n_segments; i++) {\n if (params.no_timestamps) {\n if (params.print_colors) {\n for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {\n if (params.print_special == false) {\n const whisper_token id = whisper_full_get_token_id(ctx, i, j);\n if (id >= whisper_token_eot(ctx)) {\n continue;\n }\n }\n\n const char * text = whisper_full_get_token_text(ctx, i, j);\n const float p = whisper_full_get_token_p (ctx, i, j);\n\n const int col = std::max(0, std::min((int) k_colors.size(), (int) (std::pow(p, 3)*float(k_colors.size()))));\n\n printf(\"%s%s%s\", k_colors[col].c_str(), text, \"\\033[0m\");\n }\n } else {\n const char * text = whisper_full_get_segment_text(ctx, i);\n printf(\"%s\", text);\n }\n fflush(stdout);\n } else {\n const int64_t t0 = whisper_full_get_segment_t0(ctx, i);\n const int64_t t1 = whisper_full_get_segment_t1(ctx, i);\n\n std::string speaker;\n\n if (params.diarize && pcmf32s.size() == 2) {\n const int64_t n_samples = pcmf32s[0].size();\n\n const int64_t is0 = timestamp_to_sample(t0, n_samples);\n const int64_t is1 = timestamp_to_sample(t1, n_samples);\n\n double energy0 = 0.0f;\n double energy1 = 0.0f;\n\n for (int64_t j = is0; j < is1; j++) {\n energy0 += fabs(pcmf32s[0][j]);\n energy1 += fabs(pcmf32s[1][j]);\n }\n\n if (energy0 > 1.1*energy1) {\n speaker = \"(speaker 0)\";\n } else if (energy1 > 1.1*energy0) {\n speaker = \"(speaker 1)\";\n } else {\n speaker = \"(speaker ?)\";\n }\n\n //printf(\"is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, %s\\n\", is0, is1, energy0, energy1, speaker.c_str());\n }\n\n if (params.print_colors) {\n printf(\"[%s --> %s] \", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());\n for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {\n if (params.print_special == false) {\n const whisper_token id = whisper_full_get_token_id(ctx, i, j);\n if (id >= whisper_token_eot(ctx)) {\n continue;\n }\n }\n\n const char * text = whisper_full_get_token_text(ctx, i, j);\n const float p = whisper_full_get_token_p (ctx, i, j);\n\n const int col = std::max(0, std::min((int) k_colors.size(), (int) (std::pow(p, 3)*float(k_colors.size()))));\n\n printf(\"%s%s%s%s\", speaker.c_str(), k_colors[col].c_str(), text, \"\\033[0m\");\n }\n printf(\"\\n\");\n } else {\n const char * text = whisper_full_get_segment_text(ctx, i);\n\n printf(\"[%s --> %s] %s%s\\n\", to_timestamp(t0).c_str(), to_timestamp(t1).c_str(), speaker.c_str(), text);\n }\n }\n }\n}\n\nbool output_txt(struct whisper_context * ctx, const char * fname) {\n std::ofstream fout(fname);\n if (!fout.is_open()) {\n fprintf(stderr, \"%s: failed to open '%s' for writing\\n\", __func__, fname);\n return false;\n }\n\n fprintf(stderr, \"%s: saving output to '%s'\\n\", __func__, fname);\n\n const int n_segments = whisper_full_n_segments(ctx);\n for (int i = 0; i < n_segments; ++i) {\n const char * text = whisper_full_get_segment_text(ctx, i);\n fout << text << \"\\n\";\n }\n\n return true;\n}\n\nbool output_vtt(struct whisper_context * ctx, const char * fname) {\n std::ofstream fout(fname);\n if (!fout.is_open()) {\n fprintf(stderr, \"%s: failed to open '%s' for writing\\n\", __func__, fname);\n return false;\n }\n\n fprintf(stderr, \"%s: saving output to '%s'\\n\", __func__, fname);\n\n fout << \"WEBVTT\\n\\n\";\n\n const int n_segments = whisper_full_n_segments(ctx);\n for (int i = 0; i < n_segments; ++i) {\n const char * text = whisper_full_get_segment_text(ctx, i);\n const int64_t t0 = whisper_full_get_segment_t0(ctx, i);\n const int64_t t1 = whisper_full_get_segment_t1(ctx, i);\n\n fout << to_timestamp(t0) << \" --> \" << to_timestamp(t1) << \"\\n\";\n fout << text << \"\\n\\n\";\n }\n\n return true;\n}\n\nbool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params) {\n std::ofstream fout(fname);\n if (!fout.is_open()) {\n fprintf(stderr, \"%s: failed to open '%s' for writing\\n\", __func__, fname);\n return false;\n }\n\n fprintf(stderr, \"%s: saving output to '%s'\\n\", __func__, fname);\n\n const int n_segments = whisper_full_n_segments(ctx);\n for (int i = 0; i < n_segments; ++i) {\n const char * text = whisper_full_get_segment_text(ctx, i);\n const int64_t t0 = whisper_full_get_segment_t0(ctx, i);\n const int64_t t1 = whisper_full_get_segment_t1(ctx, i);\n\n fout << i + 1 + params.offset_n << \"\\n\";\n fout << to_timestamp(t0, true) << \" --> \" << to_timestamp(t1, true) << \"\\n\";\n fout << text << \"\\n\\n\";\n }\n\n return true;\n}\n\n// karaoke video generation\n// outputs a bash script that uses ffmpeg to generate a video with the subtitles\n// TODO: font parameter adjustments\nbool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & /*params*/, float t_sec) {\n std::ofstream fout(fname);\n\n fprintf(stderr, \"%s: saving output to '%s'\\n\", __func__, fname);\n\n // TODO: become parameter\n static const char * font = \"/System/Library/Fonts/Supplemental/Courier New Bold.ttf\";\n\n fout << \"#!/bin/bash\" << \"\\n\";\n fout << \"\\n\";\n\n fout << \"ffmpeg -i \" << fname_inp << \" -f lavfi -i color=size=1200x120:duration=\" << t_sec << \":rate=25:color=black -vf \\\"\";\n\n for (int i = 0; i < whisper_full_n_segments(ctx); i++) {\n const int64_t t0 = whisper_full_get_segment_t0(ctx, i);\n const int64_t t1 = whisper_full_get_segment_t1(ctx, i);\n\n const int n = whisper_full_n_tokens(ctx, i);\n\n std::vector tokens(n);\n for (int j = 0; j < n; ++j) {\n tokens[j] = whisper_full_get_token_data(ctx, i, j);\n }\n\n if (i > 0) {\n fout << \",\";\n }\n\n // background text\n fout << \"drawtext=fontfile='\" << font << \"':fontsize=24:fontcolor=gray:x=(w-text_w)/2:y=h/2:text='':enable='between(t,\" << t0/100.0 << \",\" << t0/100.0 << \")'\";\n\n bool is_first = true;\n\n for (int j = 0; j < n; ++j) {\n const auto & token = tokens[j];\n\n if (tokens[j].id >= whisper_token_eot(ctx)) {\n continue;\n }\n\n std::string txt_bg;\n std::string txt_fg; // highlight token\n std::string txt_ul; // underline\n\n txt_bg = \"> \";\n txt_fg = \"> \";\n txt_ul = \"\\\\ \\\\ \";\n\n {\n for (int k = 0; k < n; ++k) {\n const auto & token2 = tokens[k];\n\n if (tokens[k].id >= whisper_token_eot(ctx)) {\n continue;\n }\n\n const std::string txt = whisper_token_to_str(ctx, token2.id);\n\n txt_bg += txt;\n\n if (k == j) {\n for (int l = 0; l < (int) txt.size(); ++l) {\n txt_fg += txt[l];\n txt_ul += \"_\";\n }\n txt_fg += \"|\";\n } else {\n for (int l = 0; l < (int) txt.size(); ++l) {\n txt_fg += \"\\\\ \";\n txt_ul += \"\\\\ \";\n }\n }\n }\n\n ::replace_all(txt_bg, \"'\", \"\\u2019\");\n ::replace_all(txt_bg, \"\\\"\", \"\\\\\\\"\");\n ::replace_all(txt_fg, \"'\", \"\\u2019\");\n ::replace_all(txt_fg, \"\\\"\", \"\\\\\\\"\");\n }\n\n if (is_first) {\n // background text\n fout << \",drawtext=fontfile='\" << font << \"':fontsize=24:fontcolor=gray:x=(w-text_w)/2:y=h/2:text='\" << txt_bg << \"':enable='between(t,\" << t0/100.0 << \",\" << t1/100.0 << \")'\";\n is_first = false;\n }\n\n // foreground text\n fout << \",drawtext=fontfile='\" << font << \"':fontsize=24:fontcolor=lightgreen:x=(w-text_w)/2+8:y=h/2:text='\" << txt_fg << \"':enable='between(t,\" << token.t0/100.0 << \",\" << token.t1/100.0 << \")'\";\n\n // underline\n fout << \",drawtext=fontfile='\" << font << \"':fontsize=24:fontcolor=lightgreen:x=(w-text_w)/2+8:y=h/2+16:text='\" << txt_ul << \"':enable='between(t,\" << token.t0/100.0 << \",\" << token.t1/100.0 << \")'\";\n }\n }\n\n fout << \"\\\" -c:v libx264 -pix_fmt yuv420p -y \" << fname_inp << \".mp4\" << \"\\n\";\n\n fout << \"\\n\\n\";\n fout << \"echo \\\"Your video has been saved to \" << fname_inp << \".mp4\\\"\" << \"\\n\";\n fout << \"\\n\";\n fout << \"echo \\\" ffplay \" << fname_inp << \".mp4\\\"\\n\";\n fout << \"\\n\";\n\n fout.close();\n\n fprintf(stderr, \"%s: run 'source %s' to generate karaoke video\\n\", __func__, fname);\n\n return true;\n}\n\nint main(int argc, char ** argv) {\n whisper_params params;\n\n if (whisper_params_parse(argc, argv, params) == false) {\n return 1;\n }\n\n if (params.fname_inp.empty()) {\n fprintf(stderr, \"error: no input files specified\\n\");\n whisper_print_usage(argc, argv, params);\n return 2;\n }\n\n if (params.language != \"auto\" && whisper_lang_id(params.language.c_str()) == -1) {\n fprintf(stderr, \"error: unknown language '%s'\\n\", params.language.c_str());\n whisper_print_usage(argc, argv, params);\n exit(0);\n }\n\n // whisper init\n\n struct whisper_context * ctx = whisper_init(params.model.c_str());\n\n if (ctx == nullptr) {\n fprintf(stderr, \"error: failed to initialize whisper context\\n\");\n return 3;\n }\n\n // initial prompt\n std::vector prompt_tokens;\n\n if (!params.prompt.empty()) {\n prompt_tokens.resize(1024);\n prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));\n\n fprintf(stderr, \"\\n\");\n fprintf(stderr, \"initial prompt: '%s'\\n\", params.prompt.c_str());\n fprintf(stderr, \"initial tokens: [ \");\n for (int i = 0; i < (int) prompt_tokens.size(); ++i) {\n fprintf(stderr, \"%d \", prompt_tokens[i]);\n }\n fprintf(stderr, \"]\\n\");\n }\n\n for (int f = 0; f < (int) params.fname_inp.size(); ++f) {\n const auto fname_inp = params.fname_inp[f];\n\n std::vector pcmf32; // mono-channel F32 PCM\n std::vector> pcmf32s; // stereo-channel F32 PCM\n\n // WAV input\n {\n drwav wav;\n std::vector wav_data; // used for pipe input from stdin\n\n if (fname_inp == \"-\") {\n {\n uint8_t buf[1024];\n while (true)\n {\n const size_t n = fread(buf, 1, sizeof(buf), stdin);\n if (n == 0) {\n break;\n }\n wav_data.insert(wav_data.end(), buf, buf + n);\n }\n }\n\n if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {\n fprintf(stderr, \"error: failed to open WAV file from stdin\\n\");\n return 4;\n }\n\n fprintf(stderr, \"%s: read %zu bytes from stdin\\n\", __func__, wav_data.size());\n }\n else if (drwav_init_file(&wav, fname_inp.c_str(), nullptr) == false) {\n fprintf(stderr, \"error: failed to open '%s' as WAV file\\n\", fname_inp.c_str());\n return 5;\n }\n\n if (wav.channels != 1 && wav.channels != 2) {\n fprintf(stderr, \"%s: WAV file '%s' must be mono or stereo\\n\", argv[0], fname_inp.c_str());\n return 6;\n }\n\n if (params.diarize && wav.channels != 2 && params.no_timestamps == false) {\n fprintf(stderr, \"%s: WAV file '%s' must be stereo for diarization and timestamps have to be enabled\\n\", argv[0], fname_inp.c_str());\n return 6;\n }\n\n if (wav.sampleRate != WHISPER_SAMPLE_RATE) {\n fprintf(stderr, \"%s: WAV file '%s' must be 16 kHz\\n\", argv[0], fname_inp.c_str());\n return 8;\n }\n\n if (wav.bitsPerSample != 16) {\n fprintf(stderr, \"%s: WAV file '%s' must be 16-bit\\n\", argv[0], fname_inp.c_str());\n return 9;\n }\n\n const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);\n\n std::vector pcm16;\n pcm16.resize(n*wav.channels);\n drwav_read_pcm_frames_s16(&wav, n, pcm16.data());\n drwav_uninit(&wav);\n\n // convert to mono, float\n pcmf32.resize(n);\n if (wav.channels == 1) {\n for (uint64_t i = 0; i < n; i++) {\n pcmf32[i] = float(pcm16[i])/32768.0f;\n }\n } else {\n for (uint64_t i = 0; i < n; i++) {\n pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;\n }\n }\n\n if (params.diarize) {\n // convert to stereo, float\n pcmf32s.resize(2);\n\n pcmf32s[0].resize(n);\n pcmf32s[1].resize(n);\n for (uint64_t i = 0; i < n; i++) {\n pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;\n pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;\n }\n }\n }\n\n // print system information\n {\n fprintf(stderr, \"\\n\");\n fprintf(stderr, \"system_info: n_threads = %d / %d | %s\\n\",\n params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());\n }\n\n // print some info about the processing\n {\n fprintf(stderr, \"\\n\");\n if (!whisper_is_multilingual(ctx)) {\n if (params.language != \"en\" || params.translate) {\n params.language = \"en\";\n params.translate = false;\n fprintf(stderr, \"%s: WARNING: model is not multilingual, ignoring language and translation options\\n\", __func__);\n }\n }\n fprintf(stderr, \"%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d ...\\n\",\n __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,\n params.n_threads, params.n_processors,\n params.language.c_str(),\n params.translate ? \"translate\" : \"transcribe\",\n params.no_timestamps ? 0 : 1);\n\n fprintf(stderr, \"\\n\");\n }\n\n // run the inference\n {\n whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);\n\n wparams.print_realtime = false;\n wparams.print_progress = params.print_progress;\n wparams.print_timestamps = !params.no_timestamps;\n wparams.print_special = params.print_special;\n wparams.translate = params.translate;\n wparams.language = params.language.c_str();\n wparams.n_threads = params.n_threads;\n wparams.n_max_text_ctx = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;\n wparams.offset_ms = params.offset_t_ms;\n wparams.duration_ms = params.duration_ms;\n\n wparams.token_timestamps = params.output_wts || params.max_len > 0;\n wparams.thold_pt = params.word_thold;\n wparams.max_len = params.output_wts && params.max_len == 0 ? 60 : params.max_len;\n\n wparams.speed_up = params.speed_up;\n\n wparams.prompt_tokens = prompt_tokens.empty() ? nullptr : prompt_tokens.data();\n wparams.prompt_n_tokens = prompt_tokens.empty() ? 0 : prompt_tokens.size();\n\n whisper_print_user_data user_data = { ¶ms, &pcmf32s };\n\n // this callback is called on each new segment\n if (!wparams.print_realtime) {\n wparams.new_segment_callback = whisper_print_segment_callback;\n wparams.new_segment_callback_user_data = &user_data;\n }\n\n // example for abort mechanism\n // in this example, we do not abort the processing, but we could if the flag is set to true\n // the callback is called before every encoder run - if it returns false, the processing is aborted\n {\n static bool is_aborted = false; // NOTE: this should be atomic to avoid data race\n\n wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {\n bool is_aborted = *(bool*)user_data;\n return !is_aborted;\n };\n wparams.encoder_begin_callback_user_data = &is_aborted;\n }\n\n if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {\n fprintf(stderr, \"%s: failed to process audio\\n\", argv[0]);\n return 10;\n }\n }\n\n // output stuff\n {\n printf(\"\\n\");\n\n // output to text file\n if (params.output_txt) {\n const auto fname_txt = fname_inp + \".txt\";\n output_txt(ctx, fname_txt.c_str());\n }\n\n // output to VTT file\n if (params.output_vtt) {\n const auto fname_vtt = fname_inp + \".vtt\";\n output_vtt(ctx, fname_vtt.c_str());\n }\n\n // output to SRT file\n if (params.output_srt) {\n const auto fname_srt = fname_inp + \".srt\";\n output_srt(ctx, fname_srt.c_str(), params);\n }\n\n // output to WTS file\n if (params.output_wts) {\n const auto fname_wts = fname_inp + \".wts\";\n output_wts(ctx, fname_wts.c_str(), fname_inp.c_str(), params, float(pcmf32.size() + 1000)/WHISPER_SAMPLE_RATE);\n }\n }\n }\n\n whisper_print_timings(ctx);\n whisper_free(ctx);\n\n return 0;\n}\n"
},
{
"path": "Examples/TranscribeCS/AnsiCodes.cs",
"content": "using System.Runtime.InteropServices;\n\n/// Utility class to setup console coloring with ANSI codes.\n/// The feature requires Windows 10 or newer\nstatic class AnsiCodes\n{\n\tconst string dll = \"kernel32.dll\";\n\n\t[DllImport( dll, SetLastError = true )]\n\tstatic extern IntPtr GetStdHandle( int nStdHandle );\n\n\tconst int STD_OUTPUT_HANDLE = -11;\n\n\t[Flags]\n\tenum ConsoleModes: uint\n\t{\n\t\t// Input flags\n\t\tENABLE_PROCESSED_INPUT = 0x0001,\n\t\tENABLE_LINE_INPUT = 0x0002,\n\t\tENABLE_ECHO_INPUT = 0x0004,\n\t\tENABLE_WINDOW_INPUT = 0x0008,\n\t\tENABLE_MOUSE_INPUT = 0x0010,\n\t\tENABLE_INSERT_MODE = 0x0020,\n\t\tENABLE_QUICK_EDIT_MODE = 0x0040,\n\t\tENABLE_EXTENDED_FLAGS = 0x0080,\n\t\tENABLE_AUTO_POSITION = 0x0100,\n\t\tENABLE_VIRTUAL_TERMINAL_INPUT = 0x0200,\n\n\t\t// Output flags\n\t\tENABLE_PROCESSED_OUTPUT = 0x0001,\n\t\tENABLE_WRAP_AT_EOL_OUTPUT = 0x0002,\n\t\tENABLE_VIRTUAL_TERMINAL_PROCESSING = 0x0004,\n\t\tDISABLE_NEWLINE_AUTO_RETURN = 0x0008,\n\t\tENABLE_LVB_GRID_WORLDWIDE = 0x0010\n\t}\n\n\t[DllImport( dll, SetLastError = true )]\n\tstatic extern bool GetConsoleMode( IntPtr hConsoleHandle, out ConsoleModes mode );\n\n\t[DllImport( dll, SetLastError = true )]\n\tstatic extern bool SetConsoleMode( IntPtr hConsoleHandle, ConsoleModes mode );\n\n\tstatic AnsiCodes()\n\t{\n\t\tIntPtr h = GetStdHandle( STD_OUTPUT_HANDLE );\n\t\tIntPtr INVALID_HANDLE_VALUE = (IntPtr)( -1 );\n\t\tif( h == INVALID_HANDLE_VALUE )\n\t\t\treturn;\n\n\t\tif( !GetConsoleMode( h, out ConsoleModes mode ) )\n\t\t\treturn;\n\n\t\tif( mode.HasFlag( ConsoleModes.ENABLE_VIRTUAL_TERMINAL_PROCESSING ) )\n\t\t{\n\t\t\tenabled = true;\n\t\t\treturn;\n\t\t}\n\n\t\tmode |= ConsoleModes.ENABLE_VIRTUAL_TERMINAL_PROCESSING;\n\t\tif( SetConsoleMode( h, mode ) )\n\t\t{\n\t\t\tenabled = true;\n\t\t\treturn;\n\t\t}\n\t}\n\n\tpublic static readonly bool enabled = false;\n}"
},
{
"path": "Examples/TranscribeCS/CommandLineArgs.cs",
"content": "using System.Globalization;\nusing System.Reflection;\nusing Whisper;\n\nnamespace TranscribeCS\n{\n\tsealed record class CommandLineArgs\n\t{\n\t\tpublic int n_threads = Environment.ProcessorCount;\n\t\tpublic int offset_t_ms = 0;\n\t\tpublic int offset_n = 0;\n\t\tpublic int duration_ms = 0;\n\t\tpublic int max_context = -1;\n\t\tpublic int max_len = 0;\n\n\t\tpublic float word_thold = 0.01f;\n\n\t\tpublic bool speed_up = false;\n\t\tpublic bool translate = false;\n\t\tpublic bool diarize = false;\n\t\tpublic bool output_txt = false;\n\t\tpublic bool output_vtt = false;\n\t\tpublic bool output_srt = false;\n\t\tpublic bool print_special = false;\n\t\tpublic bool print_progress = false;\n\t\tpublic bool print_colors = true;\n\t\tpublic bool no_timestamps = false;\n\t\tpublic string? prompt = null;\n\n\t\tpublic eLanguage language = eLanguage.English;\n\t\tpublic string model = string.Empty;\n\t\tpublic readonly List fileNames = new List();\n\n\t\tconst bool output_wts = false;\n\t\tpublic void apply( ref Parameters p )\n\t\t{\n\t\t\tp.setFlag( eFullParamsFlags.PrintRealtime, false );\n\t\t\tp.setFlag( eFullParamsFlags.PrintProgress, print_progress );\n\t\t\tp.setFlag( eFullParamsFlags.PrintTimestamps, !no_timestamps );\n\t\t\tp.setFlag( eFullParamsFlags.PrintSpecial, print_special );\n\t\t\tp.setFlag( eFullParamsFlags.Translate, translate );\n\t\t\tp.language = language;\n\t\t\tp.cpuThreads = n_threads;\n\t\t\tif( max_context >= 0 )\n\t\t\t\tp.n_max_text_ctx = max_context;\n\t\t\tp.offset_ms = offset_t_ms;\n\t\t\tp.duration_ms = duration_ms;\n\t\t\tp.setFlag( eFullParamsFlags.TokenTimestamps, output_wts || max_len > 0 );\n\t\t\tp.thold_pt = word_thold;\n\t\t\tp.max_len = output_wts && max_len == 0 ? 60 : max_len;\n\t\t\tp.setFlag( eFullParamsFlags.SpeedupAudio, speed_up );\n\t\t}\n\n\t\tpublic eResultFlags resultFlags()\n\t\t{\n\t\t\teResultFlags flags = eResultFlags.None;\n\t\t\tbool wts = output_wts || max_len > 0;\n\t\t\tif( !no_timestamps || wts )\n\t\t\t\tflags |= eResultFlags.Timestamps;\n\t\t\tif( wts || print_colors )\n\t\t\t\tflags |= eResultFlags.Tokens;\n\t\t\treturn flags;\n\t\t}\n\n\t\tstatic eLanguage parseLanguage( string lang ) =>\n\t\t\tLibrary.languageFromCode( lang ) ?? throw new ArgumentException( $\"Unknown language code \\\"{lang}\\\"\" );\n\n\t\tpublic CommandLineArgs( string[] argv )\n\t\t{\n\t\t\tfor( int i = 0; i < argv.Length; i++ )\n\t\t\t{\n\t\t\t\tstring arg = argv[ i ];\n\t\t\t\tif( arg[ 0 ] != '-' )\n\t\t\t\t{\n\t\t\t\t\tfileNames.Add( arg );\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\t\t\t\tif( arg == \"-h\" || arg == \"--help\" )\n\t\t\t\t{\n\t\t\t\t\tprintUsage();\n\t\t\t\t\tthrow new OperationCanceledException();\n\t\t\t\t}\n\t\t\t\telse if( arg == \"-t\" || arg == \"--threads\" ) n_threads = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-ot\" || arg == \"--offset-t\" ) offset_t_ms = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-on\" || arg == \"--offset-n\" ) offset_n = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-d\" || arg == \"--duration\" ) duration_ms = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-mc\" || arg == \"--max-context\" ) max_context = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-ml\" || arg == \"--max-len\" ) max_len = int.Parse( argv[ ++i ] );\n\t\t\t\telse if( arg == \"-wt\" || arg == \"--word-thold\" ) word_thold = float.Parse( argv[ ++i ], CultureInfo.InvariantCulture );\n\t\t\t\telse if( arg == \"-su\" || arg == \"--speed-up\" ) speed_up = true;\n\t\t\t\telse if( arg == \"-tr\" || arg == \"--translate\" ) translate = true;\n\t\t\t\telse if( arg == \"-di\" || arg == \"--diarize\" ) diarize = true;\n\t\t\t\telse if( arg == \"-otxt\" || arg == \"--output-txt\" ) output_txt = true;\n\t\t\t\telse if( arg == \"-ovtt\" || arg == \"--output-vtt\" ) output_vtt = true;\n\t\t\t\telse if( arg == \"-osrt\" || arg == \"--output-srt\" ) output_srt = true;\n\t\t\t\telse if( arg == \"-ps\" || arg == \"--print-special\" ) print_special = true;\n\t\t\t\telse if( arg == \"-nc\" || arg == \"--no-colors\" ) print_colors = false;\n\t\t\t\telse if( arg == \"-pp\" || arg == \"--print-progress\" ) print_progress = true;\n\t\t\t\telse if( arg == \"-nt\" || arg == \"--no-timestamps\" ) no_timestamps = true;\n\t\t\t\telse if( arg == \"-l\" || arg == \"--language\" ) language = parseLanguage( argv[ ++i ] );\n\t\t\t\telse if( arg == \"--prompt\" ) prompt = argv[ ++i ];\n\t\t\t\telse if( arg == \"-m\" || arg == \"--model\" ) model = argv[ ++i ];\n\t\t\t\telse if( arg == \"-f\" || arg == \"--file\" ) fileNames.Add( argv[ ++i ] );\n\t\t\t\telse\n\t\t\t\t\tthrow new ArgumentException( $\"Unknown argument: \\\"{arg}\\\"\" );\n\t\t\t}\n\t\t\tif( string.IsNullOrWhiteSpace( model ) )\n\t\t\t\tthrow new ArgumentException( \"The model file is not provided in the arguments\" );\n\t\t\tif( !File.Exists( model ) )\n\t\t\t\tthrow new FileNotFoundException( \"Model not found\", model );\n\t\t\tif( fileNames.Count <= 0 )\n\t\t\t\tthrow new ArgumentException( \"Please supply at least 1 input audio file to process\" );\n\t\t}\n\n\t\tstatic string cstr( bool b ) => b.ToString();\n\n\t\tvoid printUsage()\n\t\t{\n\t\t\tConsole.WriteLine();\n\n\t\t\tConsole.WriteLine( \"usage: {0} [options] file0.mp3 file1.wma ...\", Path.GetFileName( Assembly.GetExecutingAssembly().Location ) );\n\t\t\tConsole.WriteLine();\n\t\t\tConsole.WriteLine( \"options:\" );\n\t\t\tConsole.WriteLine( \" -h, --help [default] show this help message and exit\" );\n\t\t\tConsole.WriteLine( \" -t N, --threads N [{0,-7:D}] number of threads to use during computation\", n_threads );\n\t\t\tConsole.WriteLine( \" -ot N, --offset-t N [{0,-7:D}] time offset in milliseconds\", offset_t_ms );\n\t\t\tConsole.WriteLine( \" -on N, --offset-n N [{0,-7:D}] segment index offset\", offset_n );\n\t\t\tConsole.WriteLine( \" -d N, --duration N [{0,-7:D}] duration of audio to process in milliseconds\", duration_ms );\n\t\t\tConsole.WriteLine( \" -mc N, --max-context N [{0,-7:D}] maximum number of text context tokens to store\", max_context );\n\t\t\tConsole.WriteLine( \" -ml N, --max-len N [{0,-7:D}] maximum segment length in characters\", max_len );\n\t\t\tConsole.WriteLine( \" -wt N, --word-thold N [{0,-7:F2}] word timestamp probability threshold\", word_thold );\n\t\t\tConsole.WriteLine( \" -su, --speed-up [{0,-7}] speed up audio by x2 (reduced accuracy)\", cstr( speed_up ) );\n\t\t\tConsole.WriteLine( \" -tr, --translate [{0,-7}] translate from source language to english\", cstr( translate ) );\n\t\t\tConsole.WriteLine( \" -di, --diarize [{0,-7}] stereo audio diarization\", cstr( diarize ) );\n\t\t\tConsole.WriteLine( \" -otxt, --output-txt [{0,-7}] output result in a text file\", cstr( output_txt ) );\n\t\t\tConsole.WriteLine( \" -ovtt, --output-vtt [{0,-7}] output result in a vtt file\", cstr( output_vtt ) );\n\t\t\tConsole.WriteLine( \" -osrt, --output-srt [{0,-7}] output result in a srt file\", cstr( output_srt ) );\n\t\t\tConsole.WriteLine( \" -ps, --print-special [{0,-7}] print special tokens\", cstr( print_special ) );\n\t\t\tConsole.WriteLine( \" -nc, --no-colors [{0,-7}] do not print colors\", cstr( !print_colors ) );\n\t\t\tConsole.WriteLine( \" -nt, --no-timestamps [{0,-7}] do not print timestamps\", cstr( no_timestamps ) );\n\t\t\tConsole.WriteLine( \" -l LANG, --language LANG [{0,-7}] spoken language\", language.getCode() );\n\t\t\tConsole.WriteLine( \" --prompt PROMPT [ ] initial prompt\" );\n\t\t\tConsole.WriteLine( \" -m FNAME, --model FNAME [{0,-7}] model path\", model );\n\t\t\tConsole.WriteLine( \" -f FNAME, --file FNAME [{0,-7}] path of the input audio file\", \"\" );\n\t\t}\n\t}\n}"
},
{
"path": "Examples/TranscribeCS/Readme.txt",
"content": "This example builds .NET 6 console application which shows how to transcribe or translate audio files with the .NET wrapper."
},
{
"path": "Examples/TranscribeCS/Transcribe.cs",
"content": "using System.Globalization;\nusing Whisper;\n\nnamespace TranscribeCS\n{\n\t/// Implementation of Callbacks abstract class, to print these segments as soon as they’re produced by the library.\n\tsealed class Transcribe: Callbacks\n\t{\n\t\treadonly CommandLineArgs args;\n\t\treadonly eResultFlags resultFlags;\n\n\t\tpublic Transcribe( CommandLineArgs args )\n\t\t{\n\t\t\tthis.args = args;\n\t\t\tresultFlags = args.resultFlags();\n\t\t\tConsole.OutputEncoding = System.Text.Encoding.UTF8;\n\t\t}\n\n\t\t// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]\n\t\t// Lowest is red, middle is yellow, highest is green.\n\t\treadonly string[] k_colors = new string[]\n\t\t{\n\t\t\t\"\\x1B[38;5;196m\", \"\\x1B[38;5;202m\", \"\\x1B[38;5;208m\", \"\\x1B[38;5;214m\", \"\\x1B[38;5;220m\",\n\t\t\t\"\\x1B[38;5;226m\", \"\\x1B[38;5;190m\", \"\\x1B[38;5;154m\", \"\\x1B[38;5;118m\", \"\\x1B[38;5;82m\"\n\t\t};\n\n\t\tint colorIndex( in sToken tok )\n\t\t{\n\t\t\tfloat p = tok.probability;\n\t\t\tfloat p3 = p * p * p;\n\t\t\tint col = (int)( p3 * k_colors.Length );\n\t\t\tcol = Math.Clamp( col, 0, k_colors.Length - 1 );\n\t\t\treturn col;\n\t\t}\n\n\t\tpublic static string printTime( TimeSpan ts ) =>\n\t\t\tts.ToString( \"hh':'mm':'ss'.'fff\", CultureInfo.InvariantCulture );\n\t\tpublic static string printTimeWithComma( TimeSpan ts ) =>\n\t\t\tts.ToString( \"hh':'mm':'ss','fff\", CultureInfo.InvariantCulture );\n\n\t\tprotected override void onNewSegment( Context sender, int countNew )\n\t\t{\n\t\t\tTranscribeResult res = sender.results( resultFlags );\n\t\t\tReadOnlySpan tokens = res.tokens;\n\n\t\t\tint s0 = res.segments.Length - countNew;\n\t\t\tif( s0 == 0 )\n\t\t\t\tConsole.WriteLine();\n\n\t\t\tfor( int i = s0; i < res.segments.Length; i++ )\n\t\t\t{\n\t\t\t\tsSegment seg = res.segments[ i ];\n\n\t\t\t\tif( args.no_timestamps )\n\t\t\t\t{\n\t\t\t\t\tif( args.print_colors && AnsiCodes.enabled )\n\t\t\t\t\t{\n\t\t\t\t\t\tforeach( sToken tok in res.getTokens( seg ) )\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\tif( !args.print_special && tok.hasFlag( eTokenFlags.Special ) )\n\t\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\t\tConsole.Write( \"{0}{1}{2}\", k_colors[ colorIndex( tok ) ], tok.text, \"\\x1B[0m\" );\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\telse\n\t\t\t\t\t\tConsole.Write( seg.text );\n\t\t\t\t\tConsole.Out.Flush();\n\t\t\t\t\tcontinue;\n\t\t\t\t}\n\n\t\t\t\tstring speaker = \"\";\n\t\t\t\tif( args.diarize )\n\t\t\t\t{\n\t\t\t\t\tspeaker = sender.detectSpeaker( seg.time ) switch\n\t\t\t\t\t{\n\t\t\t\t\t\teSpeakerChannel.Unsure => \"(speaker ?)\",\n\t\t\t\t\t\teSpeakerChannel.Left => \"(speaker 0)\",\n\t\t\t\t\t\teSpeakerChannel.Right => \"(speaker 1)\",\n\t\t\t\t\t\t_ => \"\"\n\t\t\t\t\t};\n\t\t\t\t}\n\n\t\t\t\tif( args.print_colors && AnsiCodes.enabled )\n\t\t\t\t{\n\t\t\t\t\tConsole.Write( \"[{0} --> {1}] {2} \", printTime( seg.time.begin ), printTime( seg.time.end ), speaker );\n\t\t\t\t\tforeach( sToken tok in res.getTokens( seg ) )\n\t\t\t\t\t{\n\t\t\t\t\t\tif( !args.print_special && tok.hasFlag( eTokenFlags.Special ) )\n\t\t\t\t\t\t\tcontinue;\n\t\t\t\t\t\tConsole.Write( \"{0}{1}{2}\", k_colors[ colorIndex( tok ) ], tok.text, \"\\x1B[0m\" );\n\t\t\t\t\t}\n\t\t\t\t\tConsole.WriteLine();\n\t\t\t\t}\n\t\t\t\telse\n\t\t\t\t\tConsole.WriteLine( \"[{0} --> {1}] {2} {3}\", printTime( seg.time.begin ), printTime( seg.time.end ), speaker, seg.text );\n\t\t\t}\n\t\t}\n\t}\n}"
},
{
"path": "Examples/TranscribeCS/TranscribeCS.cs",
"content": "namespace TranscribeCS;\nusing Whisper;\n\nenum eFileOpenMode: byte\n{\n\t/// Decode chunks of audio directly from the file, as needed\n\tStreamFile,\n\n\t/// Decode the complete file into FP32 PCM buffer, transcribe from there\n\tBufferPCM,\n\n\t/// Load the complete input file into a buffer, decode chunks of audio from that memory buffer as needed\n\tBufferFile\n}\n\nstatic class Program\n{\n\tstatic readonly eFileOpenMode openMode = eFileOpenMode.StreamFile;\n\t// static readonly eFileOpenMode openMode = eFileOpenMode.BufferPCM;\n\t// static readonly eFileOpenMode openMode = eFileOpenMode.BufferFile;\n\n\tstatic int Main( string[] args )\n\t{\n\t\ttry\n\t\t{\n\t\t\t// dbgListGPUs();\n\n\t\t\tCommandLineArgs cla;\n\t\t\ttry\n\t\t\t{\n\t\t\t\tcla = new CommandLineArgs( args );\n\t\t\t}\n\t\t\tcatch( OperationCanceledException )\n\t\t\t{\n\t\t\t\treturn 1;\n\t\t\t}\n\t\t\tconst eLoggerFlags loggerFlags = eLoggerFlags.UseStandardError | eLoggerFlags.SkipFormatMessage;\n\t\t\tLibrary.setLogSink( eLogLevel.Debug, loggerFlags );\n\n\t\t\tusing iModel model = Library.loadModel( cla.model );\n\t\t\tint[]? prompt = null;\n\t\t\tif( !string.IsNullOrEmpty( cla.prompt ) )\n\t\t\t\tprompt = model.tokenize( cla.prompt );\n\n\t\t\tusing Context context = model.createContext();\n\t\t\tcla.apply( ref context.parameters );\n\t\t\t// When there're multiple input files, assuming they're independent clips\n\t\t\tcontext.parameters.setFlag( eFullParamsFlags.NoContext, true );\n\t\t\tusing iMediaFoundation mf = Library.initMediaFoundation();\n\t\t\tTranscribe transcribe = new Transcribe( cla );\n\n\t\t\tforeach( string audioFile in cla.fileNames )\n\t\t\t{\n\t\t\t\tif( openMode == eFileOpenMode.StreamFile )\n\t\t\t\t{\n\t\t\t\t\tusing iAudioReader reader = mf.openAudioFile( audioFile, cla.diarize );\n\t\t\t\t\tcontext.runFull( reader, transcribe, null, prompt );\n\t\t\t\t}\n\t\t\t\telse if( openMode == eFileOpenMode.BufferPCM )\n\t\t\t\t{\n\t\t\t\t\tusing iAudioBuffer buffer = mf.loadAudioFile( audioFile, cla.diarize );\n\t\t\t\t\tcontext.runFull( buffer, transcribe, prompt );\n\t\t\t\t}\n\t\t\t\telse if( openMode == eFileOpenMode.BufferFile )\n\t\t\t\t{\n\t\t\t\t\tbyte[] buffer = File.ReadAllBytes( audioFile );\n\t\t\t\t\tusing iAudioReader reader = mf.loadAudioFileData( buffer, cla.diarize );\n\t\t\t\t\tcontext.runFull( reader, transcribe, null, prompt );\n\t\t\t\t}\n\n\t\t\t\t// When asked to, produce these text files\n\t\t\t\tif( cla.output_txt )\n\t\t\t\t\twriteTextFile( context, audioFile );\n\t\t\t\tif( cla.output_srt )\n\t\t\t\t\twriteSubRip( context, audioFile, cla );\n\t\t\t\tif( cla.output_vtt )\n\t\t\t\t\twriteWebVTT( context, audioFile );\n\t\t\t}\n\n\t\t\tcontext.timingsPrint();\n\t\t\treturn 0;\n\t\t}\n\t\tcatch( Exception ex )\n\t\t{\n\t\t\tConsole.WriteLine( ex.Message );\n\t\t\treturn ex.HResult;\n\t\t}\n\t}\n\n\tstatic void writeTextFile( Context context, string audioPath )\n\t{\n\t\tusing var stream = File.CreateText( Path.ChangeExtension( audioPath, \".txt\" ) );\n\t\tforeach( sSegment seg in context.results().segments )\n\t\t\tstream.WriteLine( seg.text );\n\t}\n\n\tstatic void writeSubRip( Context context, string audioPath, CommandLineArgs cliArgs )\n\t{\n\t\tusing var stream = File.CreateText( Path.ChangeExtension( audioPath, \".srt\" ) );\n\t\tvar segments = context.results( eResultFlags.Timestamps ).segments;\n\n\t\tfor( int i = 0; i < segments.Length; i++ )\n\t\t{\n\t\t\tstream.WriteLine( i + 1 + cliArgs.offset_n );\n\t\t\tsSegment seg = segments[ i ];\n\t\t\tstring begin = Transcribe.printTimeWithComma( seg.time.begin );\n\t\t\tstring end = Transcribe.printTimeWithComma( seg.time.end );\n\t\t\tstream.WriteLine( \"{0} --> {1}\", begin, end );\n\t\t\tstream.WriteLine( seg.text );\n\t\t\tstream.WriteLine();\n\t\t}\n\t}\n\n\tstatic void writeWebVTT( Context context, string audioPath )\n\t{\n\t\tusing var stream = File.CreateText( Path.ChangeExtension( audioPath, \".vtt\" ) );\n\t\tstream.WriteLine( \"WEBVTT\" );\n\t\tstream.WriteLine();\n\n\t\tforeach( sSegment seg in context.results( eResultFlags.Timestamps ).segments )\n\t\t{\n\t\t\tstring begin = Transcribe.printTime( seg.time.begin );\n\t\t\tstring end = Transcribe.printTime( seg.time.end );\n\t\t\tstream.WriteLine( \"{0} --> {1}\", begin, end );\n\t\t\tstream.WriteLine( seg.text );\n\t\t\tstream.WriteLine();\n\t\t}\n\t}\n\n\tstatic void dbgListGPUs()\n\t{\n\t\tstring[] list = Library.listGraphicAdapters();\n\t\tConsole.WriteLine( \" Graphics Adapters:\\n{0}\", string.Join( \"\\n\", list ) );\n\t}\n}"
},
{
"path": "Examples/TranscribeCS/TranscribeCS.csproj",
"content": "\n\t\n\t\tExe\n\t\tnet6.0-windows\n\t\tenable\n\t\tenable\n\t\ttrue\n\t\tfalse\n\t\tx64\n\t\n\t\n\t\t\n\t\t\tPreserveNewest\n\t\t\n\t\n\t\n\t\t\n\t\n"
},
{
"path": "Examples/WhisperDesktop/AppState.cpp",
"content": "#include \"stdafx.h\"\n#include \"AppState.h\"\n#include \"Utils/miscUtils.h\"\n#include \n#pragma comment(lib, \"Comctl32.lib\")\n#include \"CircleIndicator.h\"\n\nnamespace\n{\n\tstatic const HKEY regKeyRoot = HKEY_CURRENT_USER;\n\tconst LPCTSTR regKey = LR\"(SOFTWARE\\const.me\\WhisperDesktop)\";\n\tconst LPCTSTR regValPath = L\"modelPath\";\n\tconst LPCTSTR regValImpl = L\"modelImpl\";\n\tconst LPCTSTR regValLang = L\"language\";\n\tconst LPCTSTR regValLastScreen = L\"screen\";\n\tconst LPCTSTR regValGpuFlags = L\"gpuFlags\";\n\n\tstatic HRESULT readString( CRegKey& k, LPCTSTR name, CString& rdi )\n\t{\n\t\tULONG nChars = 0;\n\t\tLSTATUS lss = k.QueryStringValue( name, nullptr, &nChars );\n\t\tif( lss != ERROR_SUCCESS )\n\t\t\treturn HRESULT_FROM_WIN32( lss );\n\t\tif( nChars == 0 )\n\t\t{\n\t\t\trdi = L\"\";\n\t\t\treturn S_FALSE;\n\t\t}\n\n\t\tlss = k.QueryStringValue( name, rdi.GetBufferSetLength( nChars ), &nChars );\n\t\trdi.ReleaseBuffer();\n\t\tif( lss != ERROR_SUCCESS )\n\t\t\treturn HRESULT_FROM_WIN32( lss );\n\n\t\treturn S_OK;\n\t}\n\n\tusing Whisper::eModelImplementation;\n}\n\nHRESULT AppState::startup()\n{\n\tHRESULT hr = CoInitializeEx( nullptr, COINIT_MULTITHREADED );\n\tif( FAILED( hr ) )\n\t{\n\t\treportFatalError( \"CoInitializeEx failed\", hr );\n\t\treturn hr;\n\t}\n\tcoInit = true;\n\n\tLSTATUS lss = registryKey.Create( regKeyRoot, regKey );\n\tif( lss != ERROR_SUCCESS )\n\t{\n\t\thr = HRESULT_FROM_WIN32( lss );\n\t\treportFatalError( \"Unable to open the registry key\", hr );\n\t\treturn hr;\n\t}\n\n\tINITCOMMONCONTROLSEX init;\n\tinit.dwSize = sizeof( init );\n\tinit.dwICC = ICC_LINK_CLASS | ICC_PROGRESS_CLASS | ICC_STANDARD_CLASSES | ICC_TAB_CLASSES;\n\tconst BOOL icc = InitCommonControlsEx( &init );\n\tif( !icc )\n\t{\n\t\treportFatalError( \"InitCommonControlsEx failed\", HRESULT_FROM_WIN32( GetLastError() ) );\n\t\treturn E_FAIL;\n\t}\n\n\thr = initMediaFoundation( &mediaFoundation );\n\tif( FAILED( hr ) )\n\t{\n\t\treportFatalError( \"Unable to initialize Media Foundation runtime\", hr );\n\t\treturn hr;\n\t}\n\n\thr = console.initialize();\n\tif( FAILED( hr ) )\n\t{\n\t\treportFatalError( \"Unable to initialize logging\", hr );\n\t\treturn hr;\n\t}\n\n\thr = CircleIndicator::registerClass();\n\tif( FAILED( hr ) )\n\t{\n\t\treportFatalError( \"Unable to register custom controls\", hr );\n\t\treturn hr;\n\t}\n\tappIcon.LoadIcon( IDI_WHISPERDESKTOP );\n\treturn S_OK;\n}\n\nAppState::~AppState()\n{\n\tif( coInit )\n\t{\n\t\tCoUninitialize();\n\t\tcoInit = false;\n\t}\n}\n\nHRESULT AppState::findModelSource()\n{\n\tCHECK( readString( registryKey, regValPath, source.path ) );\n\n\t{\n\t\tCAtlFile file;\n\t\tCHECK( file.Create( source.path, GENERIC_READ, FILE_SHARE_READ, OPEN_EXISTING ) );\n\t\tULONGLONG len;\n\t\tCHECK( file.GetSize( len ) );\n\t\tsource.sizeInBytes = len;\n\t}\n\n\tCString impl;\n\tCHECK( readString( registryKey, regValImpl, impl ) );\n\tCHECK( implParse( impl, source.impl ) );\n\tsource.found = true;\n\treturn S_OK;\n}\n\nHRESULT AppState::saveModelSource()\n{\n\tLSTATUS lss = registryKey.SetStringValue( regValPath, source.path );\n\tif( lss != ERROR_SUCCESS )\n\t\treturn HRESULT_FROM_WIN32( lss );\n\n\tLPCTSTR impl = implString( source.impl );\n\tif( nullptr == impl )\n\t\treturn E_INVALIDARG;\n\tlss = registryKey.SetStringValue( regValImpl, impl );\n\tif( lss != ERROR_SUCCESS )\n\t\treturn HRESULT_FROM_WIN32( lss );\n\n\treturn S_OK;\n}\n\nuint32_t AppState::languageRead()\n{\n\tDWORD dw;\n\tLSTATUS lss = registryKey.QueryDWORDValue( regValLang, dw );\n\tif( lss == ERROR_SUCCESS )\n\t\treturn dw;\n\treturn UINT_MAX;\n}\n\nvoid AppState::languageWrite( uint32_t key )\n{\n\tregistryKey.SetDWORDValue( regValLang, key );\n}\n\nCString AppState::stringLoad( LPCTSTR name )\n{\n\tCString res;\n\treadString( registryKey, name, res );\n\treturn res;\n}\nvoid AppState::stringStore( LPCTSTR name, LPCTSTR value )\n{\n\tif( nullptr != value )\n\t\tregistryKey.SetStringValue( name, value );\n\telse\n\t\tregistryKey.DeleteValue( name );\n}\n\nuint32_t AppState::dwordLoad( LPCTSTR name, uint32_t fallback )\n{\n\tDWORD dw;\n\tLSTATUS lss = registryKey.QueryDWORDValue( name, dw );\n\tif( lss == ERROR_SUCCESS )\n\t\treturn dw;\n\treturn fallback;\n}\nvoid AppState::dwordStore( LPCTSTR name, uint32_t value )\n{\n\tregistryKey.SetDWORDValue( name, value );\n}\n\nvoid AppState::lastScreenSave( HRESULT code )\n{\n\tdwordStore( regValLastScreen, (uint32_t)code );\n}\n\nHRESULT AppState::lastScreenLoad()\n{\n\treturn (HRESULT)dwordLoad( regValLastScreen, SCREEN_TRANSCRIBE );\n}\n\nvoid AppState::setupIcon( CWindow* wnd )\n{\n\tHICON ic = appIcon;\n\tif( nullptr != ic )\n\t{\n\t\twnd->SendMessage( WM_SETICON, ICON_SMALL, (LPARAM)ic );\n\t\twnd->SendMessage( WM_SETICON, ICON_BIG, (LPARAM)ic );\n\t}\n}\n\nuint32_t AppState::gpuFlagsLoad()\n{\n\treturn dwordLoad( regValGpuFlags, 0 );\n}\n\nvoid AppState::gpuFlagsStore( uint32_t flags )\n{\n\tif( 0 == flags )\n\t\tregistryKey.DeleteValue( regValGpuFlags );\n\telse\n\t\tdwordStore( regValGpuFlags, flags );\n}\n\nbool AppState::boolLoad( LPCTSTR name )\n{\n\treturn dwordLoad( name, 0 ) != 0;\n}\n\nvoid AppState::boolStore( LPCTSTR name, bool val )\n{\n\tif( val )\n\t\tdwordStore( name, 1 );\n\telse\n\t\tregistryKey.DeleteValue( name );\n}"
},
{
"path": "Examples/WhisperDesktop/AppState.h",
"content": "#pragma once\n#include \"Utils/DebugConsole.h\"\n\nclass AppState\n{\n\tbool coInit = false;\n\tCRegKey registryKey;\n\tCIcon appIcon;\npublic:\n\n\tstruct ModelSource\n\t{\n\t\tCString path;\n\t\tbool found = false;\n\t\tWhisper::eModelImplementation impl = (Whisper::eModelImplementation)0;\n\t\tuint64_t sizeInBytes = 0;\n\t};\n\tModelSource source;\n\n\tDebugConsole console;\n\tCComPtr mediaFoundation;\n\tCComPtr model;\n\n\t~AppState();\n\n\t// Setup the initial things\n\tHRESULT startup();\n\n\tHRESULT findModelSource();\n\n\tHRESULT saveModelSource();\n\n\tuint32_t languageRead();\n\tvoid languageWrite( uint32_t key );\n\n\tCString stringLoad( LPCTSTR name );\n\tvoid stringStore( LPCTSTR name, LPCTSTR value );\n\tuint32_t dwordLoad( LPCTSTR name, uint32_t fallback );\n\tvoid dwordStore( LPCTSTR name, uint32_t value );\n\tbool boolLoad( LPCTSTR name );\n\tvoid boolStore( LPCTSTR name, bool val );\n\n\tbool automaticallyLoadModel = true;\n\n\tvoid lastScreenSave( HRESULT code );\n\tHRESULT lastScreenLoad();\n\n\tvoid setupIcon( CWindow* wnd );\n\n\tuint32_t gpuFlagsLoad();\n\tvoid gpuFlagsStore( uint32_t flags );\n};\n\nconstexpr HRESULT SCREEN_MODEL = 1;\nconstexpr HRESULT SCREEN_TRANSCRIBE = 2;\nconstexpr HRESULT SCREEN_CAPTURE = 3;"
},
{
"path": "Examples/WhisperDesktop/CaptureDlg.cpp",
"content": "#include \"stdafx.h\"\n#include \"CaptureDlg.h\"\n\nHRESULT CaptureDlg::show()\n{\n\tauto res = DoModal( nullptr );\n\tif( res == -1 )\n\t\treturn HRESULT_FROM_WIN32( GetLastError() );\n\tswitch( res )\n\t{\n\tcase IDC_BACK:\n\t\treturn SCREEN_MODEL;\n\tcase IDC_TRANSCRIBE:\n\t\treturn SCREEN_TRANSCRIBE;\n\t}\n\treturn S_OK;\n}\n\nstatic const LPCTSTR regValDevice = L\"captureDevice\";\nstatic const LPCTSTR regValOutPath = L\"captureTextFile\";\nstatic const LPCTSTR regValOutFormat = L\"captureTextFlags\";\n\nenum struct CaptureDlg::eTextFlags : uint32_t\n{\n\tSave = 1,\n\tAppend = 2,\n\tTimestamps = 4,\n};\n\nLRESULT CaptureDlg::OnInitDialog( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled )\n{\n\t// First DDX call, hooks up variables to controls.\n\tDoDataExchange( false );\n\n\tlanguageSelector.initialize( m_hWnd, IDC_LANGUAGE, appState );\n\tcbTranslate.initialize( m_hWnd, IDC_TRANSLATE, appState );\n\tcbConsole.initialize( m_hWnd, IDC_CONSOLE, appState );\n\n\tpendingState.initialize(\n\t\t// Controls to disable while pending, re-enable afterwards\n\t\t{\n\t\t\tlanguageSelector, GetDlgItem( IDC_TRANSLATE ),\n\t\t\tcbCaptureDevice,\n\t\t\tcheckSave, checkAppend, checkTimestamps, transcribeOutputPath, transcribeOutputBrowse,\n\t\t\tGetDlgItem( IDC_DEV_REFRESH ),\n\t\t\tGetDlgItem( IDC_BACK ),\n\t\t\tGetDlgItem( IDC_TRANSCRIBE ),\n\t\t\tGetDlgItem( IDCANCEL ),\n\t\t},\n\t\t// Controls to show while pending, hide afterwards\n\t\t{\n\t\t\tvoiceActivity, GetDlgItem( IDC_VOICE_ACTIVITY_LBL ),\n\t\t\ttranscribeActivity, GetDlgItem( IDC_TRANS_LBL ),\n\t\t\tstalled, GetDlgItem( IDC_STALL_LBL ),\n\t\t\tprogressBar,\n\t\t} );\n\n\tstalled.setActiveColor( flipRgb( 0xffcc33 ) );\n\n\tHRESULT hr = work.create( this );\n\tif( FAILED( hr ) )\n\t{\n\t\treportError( m_hWnd, L\"CreateThreadpoolWork failed\", nullptr, hr );\n\t\tEndDialog( IDCANCEL );\n\t}\n\n\tlistDevices();\n\tselectDevice( appState.stringLoad( regValDevice ) );\n\n\tconstexpr uint32_t defaultFlags = (uint32_t)eTextFlags::Append;\n\tuint32_t flags = appState.dwordLoad( regValOutFormat, defaultFlags );\n\tif( flags & (uint32_t)eTextFlags::Save )\n\t\tcheckSave.SetCheck( BST_CHECKED );\n\tif( flags & (uint32_t)eTextFlags::Append )\n\t\tcheckAppend.SetCheck( BST_CHECKED );\n\tif( flags & (uint32_t)eTextFlags::Timestamps )\n\t\tcheckTimestamps.SetCheck( BST_CHECKED );\n\n\ttranscribeOutputPath.SetWindowText( appState.stringLoad( regValOutPath ) );\n\tonSaveTextCheckbox();\n\n\tappState.lastScreenSave( SCREEN_CAPTURE );\n\tappState.setupIcon( this );\n\tATLVERIFY( CenterWindow() );\n\treturn 0;\n}\n\nHRESULT __stdcall CaptureDlg::listDevicesCallback( int len, const Whisper::sCaptureDevice* buffer, void* pv ) noexcept\n{\n\tstd::vector& devices = *( std::vector * )pv;\n\tdevices.resize( len );\n\tfor( int i = 0; i < len; i++ )\n\t{\n\t\tdevices[ i ].displayName = buffer[ i ].displayName;\n\t\tdevices[ i ].endpoint = buffer[ i ].endpoint;\n\t}\n\treturn S_OK;\n}\n\nbool CaptureDlg::listDevices()\n{\n\tappState.mediaFoundation->listCaptureDevices( &listDevicesCallback, &devices );\n\tcbCaptureDevice.ResetContent();\n\tfor( const auto& dev : devices )\n\t\tcbCaptureDevice.AddString( dev.displayName );\n\treturn !devices.empty();\n}\n\nvoid CaptureDlg::onDeviceRefresh()\n{\n\t// Save the current selection\n\tconst int curSel = cbCaptureDevice.GetCurSel();\n\tCString str;\n\tif( curSel >= 0 && curSel < (int)devices.size() )\n\t\tstr = std::move( devices[ curSel ].endpoint );\n\n\t// Refresh\n\tlistDevices();\n\n\t// Restore the selection\n\tselectDevice( str );\n\n\tconst size_t len = devices.size();\n\tif( len == 0 )\n\t{\n\t\tMessageBox( L\"No capture devices found on this computer.\\nIf you have a USB microphone, connect it to this PC,\\nand press refresh button.\",\n\t\t\tL\"Capture Devices\", MB_OK | MB_ICONWARNING );\n\t}\n\telse\n\t{\n\t\tconst char* suffix = ( len != 1 ) ? \"s\" : \"\";\n\t\tstr.Format( L\"Detected %zu audio capture device%S.\", len, suffix );\n\t\tMessageBox( str, L\"Capture Devices\", MB_OK | MB_ICONINFORMATION );\n\t}\n}\n\nbool CaptureDlg::selectDevice( LPCTSTR endpoint )\n{\n\tif( nullptr != endpoint && 0 != *endpoint )\n\t{\n\t\tfor( size_t i = 0; i < devices.size(); i++ )\n\t\t{\n\t\t\tif( devices[ i ].endpoint == endpoint )\n\t\t\t{\n\t\t\t\tcbCaptureDevice.SetCurSel( (int)i );\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\n\tif( !devices.empty() )\n\t\tcbCaptureDevice.SetCurSel( 0 );\n\treturn false;\n}\n\nvoid CaptureDlg::onSaveTextCheckbox()\n{\n\tconst BOOL enabled = ( checkSave.GetCheck() == BST_CHECKED );\n\tstd::array controls = { checkAppend, checkTimestamps, transcribeOutputPath, transcribeOutputBrowse };\n\tfor( HWND w : controls )\n\t\t::EnableWindow( w, enabled );\n}\n\nvoid CaptureDlg::onBrowseResult()\n{\n\tLPCTSTR title = L\"Output Text File\";\n\tLPCTSTR outputFilters = L\"Text files (*.txt)\\0*.txt\\0\\0\";\n\tCString path;\n\ttranscribeOutputPath.GetWindowText( path );\n\tif( !getSaveFileName( m_hWnd, title, outputFilters, path ) )\n\t\treturn;\n\n\tLPCTSTR ext = PathFindExtension( path );\n\tif( 0 == *ext )\n\t{\n\t\twchar_t* const buffer = path.GetBufferSetLength( path.GetLength() + 5 );\n\t\tPathRenameExtension( buffer, L\".txt\" );\n\t\tpath.ReleaseBuffer();\n\t}\n\n\ttranscribeOutputPath.SetWindowText( path );\n}\n\nCaptureDlg::eTextFlags CaptureDlg::getOutputFlags()\n{\n\tuint32_t flags = 0;\n\tif( checkSave.GetCheck() == BST_CHECKED )\n\t\tflags |= (uint32_t)eTextFlags::Save;\n\tif( checkAppend.GetCheck() == BST_CHECKED )\n\t\tflags |= (uint32_t)eTextFlags::Append;\n\tif( checkTimestamps.GetCheck() == BST_CHECKED )\n\t\tflags |= (uint32_t)eTextFlags::Timestamps;\n\treturn (eTextFlags)flags;\n}\n\nvoid CaptureDlg::setPending( bool nowPending )\n{\n\tpendingState.setPending( nowPending );\n\tif( nowPending )\n\t{\n\t\tprogressBar.SetMarquee( TRUE, 0 );\n\t\tbtnRunCapture.SetWindowText( L\"Stop\" );\n\t}\n\telse\n\t{\n\t\tprogressBar.SetMarquee( FALSE, 0 );\n\t\tbtnRunCapture.SetWindowText( L\"Capture\" );\n\t\tbtnRunCapture.EnableWindow( TRUE );\n\t\tcaptureRunning = false;\n\t}\n}\n\nvoid CaptureDlg::onRunCapture()\n{\n\tif( captureRunning )\n\t{\n\t\tthreadState.stopRequested = true;\n\t\tbtnRunCapture.EnableWindow( FALSE );\n\t\treturn;\n\t}\n\n\tint dev = cbCaptureDevice.GetCurSel();\n\tif( dev < 0 || dev >= (int)devices.size() )\n\t{\n\t\tshowError( L\"Please select a capture device\", S_FALSE );\n\t\treturn;\n\t}\n\tthreadState.endpoint = devices[ dev ].endpoint;\n\tthreadState.language = languageSelector.selectedLanguage();\n\tthreadState.translate = cbTranslate.checked();\n\tif( isInvalidTranslate( m_hWnd, threadState.language, threadState.translate ) )\n\t\treturn;\n\n\tthreadState.flags = getOutputFlags();\n\tif( (uint32_t)threadState.flags & (uint32_t)eTextFlags::Save )\n\t{\n\t\ttranscribeOutputPath.GetWindowText( threadState.textOutputPath );\n\t\tif( threadState.textOutputPath.GetLength() <= 0 )\n\t\t{\n\t\t\tshowError( L\"Please specify the output text file\", S_FALSE );\n\t\t\treturn;\n\t\t}\n\t\tappState.stringStore( regValOutPath, threadState.textOutputPath );\n\t}\n\telse\n\t\tcbConsole.ensureChecked();\n\n\tlanguageSelector.saveSelection( appState );\n\tcbTranslate.saveSelection( appState );\n\tappState.stringStore( regValDevice, threadState.endpoint );\n\tappState.dwordStore( regValOutFormat, (uint32_t)threadState.flags );\n\n\tcaptureRunning = true;\n\tthreadState.errorMessage = L\"\";\n\tthreadState.stopRequested = false;\n\tthreadState.captureParams.minDuration = 7;\n\tthreadState.captureParams.maxDuration = 11;\n\tsetPending( true );\n\twork.post();\n}\n\nvoid __declspec( noinline ) CaptureDlg::getThreadError()\n{\n\tgetLastError( threadState.errorMessage );\n}\n\n#define CHECK_EX( hr ) { const HRESULT __hr = ( hr ); if( FAILED( __hr ) ) { getThreadError(); return __hr; } }\n\nstatic HRESULT appendDate( CString& str, const SYSTEMTIME& time )\n{\n\tconstexpr DWORD dateFlags = DATE_LONGDATE;\n\tint cc = GetDateFormatEx( LOCALE_NAME_USER_DEFAULT, dateFlags, &time, nullptr, nullptr, 0, nullptr );\n\tif( 0 == cc )\n\t\treturn getLastHr();\n\n\tconst int oldLength = str.GetLength();\n\twchar_t* const buffer = str.GetBufferSetLength( oldLength + cc );\n\tcc = GetDateFormatEx( LOCALE_NAME_USER_DEFAULT, dateFlags, &time, nullptr, buffer + oldLength, cc, nullptr );\n\tif( 0 != cc )\n\t{\n\t\tstr.ReleaseBuffer();\n\t\treturn S_OK;\n\t}\n\tHRESULT hr = getLastHr();\n\tstr.ReleaseBuffer();\n\treturn hr;\n}\n\nstatic HRESULT appendTime( CString& str, const SYSTEMTIME& time )\n{\n\tconstexpr DWORD timeFlags = 0;\n\tint cc = GetTimeFormatEx( LOCALE_NAME_USER_DEFAULT, timeFlags, &time, nullptr, nullptr, 0 );\n\tif( 0 == cc )\n\t\treturn getLastHr();\n\n\tconst int oldLength = str.GetLength();\n\twchar_t* const buffer = str.GetBufferSetLength( oldLength + cc );\n\tcc = GetTimeFormatEx( LOCALE_NAME_USER_DEFAULT, timeFlags, &time, nullptr, buffer + oldLength, cc );\n\tif( 0 != cc )\n\t{\n\t\tstr.ReleaseBuffer();\n\t\treturn S_OK;\n\t}\n\tHRESULT hr = getLastHr();\n\tstr.ReleaseBuffer();\n\treturn hr;\n}\n\nstatic HRESULT printDateTime( CAtlFile& file )\n{\n\tSYSTEMTIME time;\n\tGetLocalTime( &time );\n\n\tCString str;\n\tstr = L\"==== Captured on \";\n\tCHECK( appendDate( str, time ) );\n\tstr += L\", \";\n\tCHECK( appendTime( str, time ) );\n\tstr += L\" ====\\r\\n\";\n\n\tCStringA u8;\n\tmakeUtf8( u8, str );\n\treturn file.Write( cstr( u8 ), (DWORD)u8.GetLength() );\n}\n\ninline HRESULT CaptureDlg::runCapture()\n{\n\tclearLastError();\n\tusing namespace Whisper;\n\tCComPtr capture;\n\tCHECK_EX( appState.mediaFoundation->openCaptureDevice( threadState.endpoint, threadState.captureParams, &capture ) );\n\n\tHRESULT hr;\n\tCAtlFile file;\n\tconst uint32_t flags = (uint32_t)threadState.flags;\n\tif( flags & (uint32_t)eTextFlags::Save )\n\t{\n\t\tconst bool append = 0 != ( flags & (uint32_t)eTextFlags::Append );\n\t\tconst DWORD creation = append ? OPEN_ALWAYS : CREATE_ALWAYS;\n\t\thr = file.Create( threadState.textOutputPath, GENERIC_WRITE, FILE_SHARE_READ, creation );\n\t\tif( FAILED( hr ) )\n\t\t{\n\t\t\tthreadState.errorMessage = L\"Unable to create the output text file\";\n\t\t\treturn hr;\n\t\t}\n\t\tif( append )\n\t\t{\n\t\t\tULONGLONG size;\n\t\t\tCHECK( file.GetSize( size ) );\n\t\t\tif( size == 0 )\n\t\t\t\tCHECK( writeUtf8Bom( file ) )\n\t\t\telse\n\t\t\t\tCHECK( file.Seek( 0, SEEK_END ) );\n\t\t}\n\t\telse\n\t\t{\n\t\t\tCHECK( writeUtf8Bom( file ) );\n\t\t}\n\n\t\tif( flags & (uint32_t)eTextFlags::Timestamps )\n\t\t\tCHECK( printDateTime( file ) );\n\n\t\tthreadState.file = &file;\n\t}\n\telse\n\t\tthreadState.file = nullptr;\n\n\tCComPtr context;\n\tCHECK_EX( appState.model->createContext( &context ) );\n\n\tsFullParams fullParams;\n\tCHECK_EX( context->fullDefaultParams( eSamplingStrategy::Greedy, &fullParams ) );\n\tfullParams.language = threadState.language;\n\tfullParams.setFlag( eFullParamsFlags::Translate, threadState.translate );\n\tfullParams.resetFlag( eFullParamsFlags::PrintRealtime );\n\tfullParams.new_segment_callback = &newSegmentCallback;\n\tfullParams.new_segment_callback_user_data = this;\n\n\tsCaptureCallbacks callbacks;\n\tcallbacks.shouldCancel = &cbCancel;\n\tcallbacks.captureStatus = &cbStatus;\n\tcallbacks.pv = this;\n\n\tCHECK_EX( context->runCapture( fullParams, callbacks, capture ) );\n\tthreadState.file = nullptr;\n\n\tcontext->timingsPrint();\n\treturn S_OK;\n}\n\nvoid __stdcall CaptureDlg::poolCallback() noexcept\n{\n\tconst HRESULT hr = runCapture();\n\tPostMessage( WM_CALLBACK_COMPLETION, hr );\n}\n\nvoid CaptureDlg::showError( LPCTSTR text, HRESULT hr )\n{\n\treportError( m_hWnd, text, L\"Capture failed\", hr );\n}\n\nLRESULT CaptureDlg::onThreadQuit( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled )\n{\n\tsetPending( false );\n\n\tconst HRESULT hr = (HRESULT)wParam;\n\tif( FAILED( hr ) )\n\t{\n\t\tLPCTSTR failMessage = L\"Capture failed\";\n\n\t\tif( threadState.errorMessage.GetLength() > 0 )\n\t\t{\n\t\t\tCString tmp = failMessage;\n\t\t\ttmp += L\"\\n\";\n\t\t\ttmp += threadState.errorMessage;\n\t\t\tshowError( tmp, hr );\n\t\t}\n\t\telse\n\t\t\tshowError( failMessage, hr );\n\n\t\treturn 0;\n\t}\n\telse\n\t{\n\t\tif( (uint32_t)threadState.flags & (uint32_t)eTextFlags::Save )\n\t\t\tShellExecute( NULL, L\"open\", threadState.textOutputPath, NULL, NULL, SW_SHOW );\n\t}\n\n\treturn 0;\n}\n\nLRESULT CaptureDlg::onThreadStatus( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled )\n{\n\tusing namespace Whisper;\n\tconst uint8_t newStatus = (uint8_t)wParam;\n\t// Update the GUI\n\tvoiceActivity.setActive( 0 != ( newStatus & (uint8_t)eCaptureStatus::Voice ) );\n\ttranscribeActivity.setActive( 0 != ( newStatus & (uint8_t)eCaptureStatus::Transcribing ) );\n\tstalled.setActive( 0 != ( newStatus & (uint8_t)eCaptureStatus::Stalled ) );\n\treturn 0;\n}\n\nHRESULT __stdcall CaptureDlg::cbCancel( void* pv ) noexcept\n{\n\tconst bool stopRequested = ( (CaptureDlg*)pv )->threadState.stopRequested;\n\treturn stopRequested ? S_FALSE : S_OK;\n}\n\nHRESULT __stdcall CaptureDlg::cbStatus( void* pv, Whisper::eCaptureStatus status ) noexcept\n{\n\tCaptureDlg& dialog = *(CaptureDlg*)pv;\n\tif( dialog.PostMessage( WM_CALLBACK_STATUS, (uint8_t)status ) )\n\t\treturn S_OK;\n\treturn getLastHr();\n}\n\nHRESULT __cdecl CaptureDlg::newSegmentCallback( Whisper::iContext* ctx, uint32_t n_new, void* user_data ) noexcept\n{\n\tusing namespace Whisper;\n\tCComPtr result;\n\tconst eResultFlags flags = eResultFlags::Timestamps | eResultFlags::Tokens;\n\tCHECK( ctx->getResults( flags, &result ) );\n\tCHECK( logNewSegments( result, n_new ) );\n\n\tCaptureDlg& dialog = *(CaptureDlg*)user_data;\n\treturn dialog.appendTextFile( result, n_new );\n}\n\nHRESULT CaptureDlg::appendTextFile( Whisper::iTranscribeResult* results, uint32_t newSegments )\n{\n\tif( nullptr == threadState.file || 0 == newSegments )\n\t\treturn S_OK;\n\n\tusing namespace Whisper;\n\tsTranscribeLength length;\n\tCHECK( results->getSize( length ) );\n\n\tconst size_t len = length.countSegments;\n\tsize_t i = len - newSegments;\n\n\tconst sSegment* const segments = results->getSegments();\n\tCStringA str;\n\tfor( ; i < len; i++ )\n\t{\n\t\tconst sSegment& seg = segments[ i ];\n\t\tif( 0 != ( (uint32_t)threadState.flags & (uint32_t)eTextFlags::Timestamps ) )\n\t\t{\n\t\t\tstr = \"[\";\n\t\t\tprintTime( str, seg.time.begin );\n\t\t\tstr += \" --> \";\n\t\t\tprintTime( str, seg.time.end );\n\t\t\tstr += \"] \";\n\t\t}\n\t\telse\n\t\t\tstr = \"\";\n\n\t\tstr += seg.text;\n\t\tstr += \"\\r\\n\";\n\n\t\tCHECK( threadState.file->Write( cstr( str ), (DWORD)str.GetLength() ) );\n\t}\n\n\tCHECK( threadState.file->Flush() );\n\treturn S_OK;\n}"
},
{
"path": "Examples/WhisperDesktop/CaptureDlg.h",
"content": "#pragma once\n#include \"AppState.h\"\n#include \"Utils/WTL/atlddx.h\"\n#include \"Utils/miscUtils.h\"\n#include \"Utils/LanguageDropdown.h\"\n#include \"Utils/TranslateCheckbox.h\"\n#include \"Utils/PendingState.h\"\n#include \"CircleIndicator.h\"\n\nclass CaptureDlg :\n\tpublic CDialogImpl,\n\tpublic CWinDataExchange,\n\tpublic iThreadPoolCallback\n{\n\tAppState& appState;\n\npublic:\n\tstatic constexpr UINT IDD = IDD_CAPTURE_DIALOG;\n\tstatic constexpr UINT WM_CALLBACK_COMPLETION = WM_APP + 1;\n\tstatic constexpr UINT WM_CALLBACK_STATUS = WM_APP + 2;\n\n\tCaptureDlg( AppState& app ) : appState( app ) { }\n\n\tHRESULT show();\n\n\tBEGIN_MSG_MAP( CaptureDlg )\n\t\tMESSAGE_HANDLER( WM_INITDIALOG, OnInitDialog )\n\t\tON_BUTTON_CLICK( IDC_CONSOLE, cbConsole.click )\n\t\tON_BUTTON_CLICK( IDC_DEV_REFRESH, onDeviceRefresh );\n\t\tON_BUTTON_CLICK( IDC_BROWSE_RESULT, onBrowseResult );\n\t\tON_BUTTON_CLICK( IDC_SAVE_TEXT, onSaveTextCheckbox );\n\t\tON_BUTTON_CLICK( IDC_RUN_CAPTURE, onRunCapture );\n\n\t\tON_BUTTON_CLICK( IDCANCEL, onClose )\n\t\tON_BUTTON_CLICK( IDC_BACK, onBack )\n\t\tON_BUTTON_CLICK( IDC_TRANSCRIBE, onTranscribe );\n\n\t\tMESSAGE_HANDLER( WM_CALLBACK_COMPLETION, onThreadQuit );\n\t\tMESSAGE_HANDLER( WM_CALLBACK_STATUS, onThreadStatus );\n\tEND_MSG_MAP()\n\n\tBEGIN_DDX_MAP( CaptureDlg )\n\t\tDDX_CONTROL_HANDLE( IDC_DEVICE, cbCaptureDevice )\n\t\tDDX_CONTROL_HANDLE( IDC_RUN_CAPTURE, btnRunCapture );\n\t\tDDX_CONTROL_HANDLE( IDC_TRANSCRIBE_PROGRESS, progressBar );\n\t\tDDX_CONTROL_HANDLE( IDC_SAVE_TEXT, checkSave )\n\t\tDDX_CONTROL_HANDLE( IDC_SAVE_APPEND, checkAppend )\n\t\tDDX_CONTROL_HANDLE( IDC_SAVE_TIMESTAMPS, checkTimestamps )\n\t\tDDX_CONTROL_HANDLE( IDC_PATH_RESULT, transcribeOutputPath )\n\t\tDDX_CONTROL_HANDLE( IDC_BROWSE_RESULT, transcribeOutputBrowse );\n\n\t\tDDX_CONTROL( IDC_VOICE_ACTIVITY, voiceActivity );\n\t\tDDX_CONTROL( IDC_TRANS_STATUS, transcribeActivity );\n\t\tDDX_CONTROL( IDC_STALL_STATUS, stalled );\n\t\t\n\tEND_DDX_MAP()\n\nprivate:\n\tPendingState pendingState;\n\tvoid setPending( bool nowPending );\n\n\tLRESULT OnInitDialog( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled );\n\n\tvoid onClose()\n\t{\n\t\tATLVERIFY( EndDialog( IDCANCEL ) );\n\t}\n\tvoid onBack()\n\t{\n\t\tATLVERIFY( EndDialog( IDC_BACK ) );\n\t}\n\tvoid onTranscribe()\n\t{\n\t\tATLVERIFY( EndDialog( IDC_TRANSCRIBE ) );\n\t}\n\n\t// List capture devices, and populate the combobox\n\tbool listDevices();\n\tvoid onDeviceRefresh();\n\tbool selectDevice( LPCTSTR endpoint );\n\n\tstatic HRESULT __stdcall listDevicesCallback( int len, const Whisper::sCaptureDevice* buffer, void* pv ) noexcept;\n\tConsoleCheckbox cbConsole;\n\tLanguageDropdown languageSelector;\n\tTranslateCheckbox cbTranslate;\n\tCComboBox cbCaptureDevice;\n\n\tvoid onBrowseResult();\n\n\tenum struct eTextFlags : uint32_t;\n\tCButton\tcheckSave, checkAppend, checkTimestamps;\n\tCEdit transcribeOutputPath;\n\tCButton transcribeOutputBrowse;\n\tvoid onSaveTextCheckbox();\n\teTextFlags getOutputFlags();\n\n\tCButton btnRunCapture;\n\tCProgressBarCtrl progressBar;\n\tThreadPoolWork work;\n\n\tstruct sCaptureDevice\n\t{\n\t\tCString displayName;\n\t\tCString endpoint;\n\t};\n\tstd::vector devices;\n\n\tvoid showError( LPCTSTR text, HRESULT hr );\n\n\tCircleIndicator voiceActivity;\n\tCircleIndicator transcribeActivity;\n\tCircleIndicator stalled;\n\n\tstruct sThreadState\n\t{\n\t\tvolatile bool stopRequested;\n\t\tbool translate;\n\t\teTextFlags flags;\n\t\tCAtlFile* file;\n\t\tuint32_t language;\n\t\tWhisper::sCaptureParams captureParams;\n\t\tCString endpoint;\n\t\tCString textOutputPath;\n\t\tCString errorMessage;\n\t};\n\tsThreadState threadState;\n\tbool captureRunning = false;\n\n\tvoid getThreadError();\n\tvoid onRunCapture();\n\tHRESULT runCapture();\n\tvoid __stdcall poolCallback() noexcept override final;\n\n\tLRESULT onThreadQuit( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled );\n\tLRESULT onThreadStatus( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled );\n\n\tstatic HRESULT __stdcall cbCancel( void* pv ) noexcept;\n\tstatic HRESULT __stdcall cbStatus( void* pv, Whisper::eCaptureStatus status ) noexcept;\n\n\tstatic HRESULT __cdecl newSegmentCallback( Whisper::iContext* ctx, uint32_t n_new, void* user_data ) noexcept;\n\n\tHRESULT appendTextFile( Whisper::iTranscribeResult* results, uint32_t newSegments );\n};"
},
{
"path": "Examples/WhisperDesktop/CircleIndicator.cpp",
"content": "#include \"stdafx.h\"\n#include \"CircleIndicator.h\"\n#include \n#include \"AppState.h\"\n\nnamespace\n{\n\tstatic const LPCTSTR windowClassName = L\"CircleIndicator\";\n\n\t// Font with these symbols, shipped with Windows since forever:\n\t// https://learn.microsoft.com/en-us/typography/font-list/segoe-ui-symbol\n\tstatic const LPCTSTR fontName = L\"Segoe UI Symbol\";\n\n\t// Outlined circle\n\tstatic const LPCTSTR circleOutline = L\"⚪\";\n\t// Filled circle\n\tstatic const LPCTSTR circleFilled = L\"⚫\";\n\n\t// Font size for that symbol font, in points\n\tconstexpr int fontSizePoints = 14;\n\n\t// Default active color for the indicator\n\tconstexpr uint32_t defaultActiveColor = 0x7FFF7F;\n}\n\nCircleIndicator::CircleIndicator() :\n\tactiveColor( defaultActiveColor )\n{ }\n\nATL::CWndClassInfo& CircleIndicator::GetWndClassInfo()\n{\n\t// Use custom class style with CS_PARENTDC, and COLOR_3DFACE for the background\n\tstatic ATL::CWndClassInfo wc = \n\t{\n\t\t{ sizeof( WNDCLASSEX ),\n\t\tCS_HREDRAW | CS_VREDRAW | CS_PARENTDC,\n\t\tStartWindowProc,\n\t\t0, 0, NULL, NULL, NULL, (HBRUSH)( COLOR_3DFACE + 1 ), NULL, windowClassName, NULL },\n\t\tNULL, NULL, IDC_ARROW, TRUE, 0, _T( \"\" )\n\t};\n\treturn wc;\n}\n\n// Class registration\nHRESULT CircleIndicator::registerClass()\n{\n\tWNDPROC pUnusedWndSuperProc = nullptr;\n\tATOM a = GetWndClassInfo().Register( &pUnusedWndSuperProc );\n\tif( 0 != a )\n\t\treturn S_OK;\n\treturn getLastHr();\n}\n\nHRESULT CircleIndicator::createFont( int height )\n{\n\tLOGFONT logFont;\n\tmemset( &logFont, 0, sizeof( logFont ) );\n\tlogFont.lfHeight = height;\n\tlogFont.lfCharSet = ANSI_CHARSET;\n\tlogFont.lfOutPrecision = OUT_TT_PRECIS;\n\tlogFont.lfClipPrecision = CLIP_DEFAULT_PRECIS;\n\twcsncpy_s( logFont.lfFaceName, fontName, _TRUNCATE );\n\tfont.CreateFontIndirect( &logFont );\n\tif( font )\n\t\treturn S_OK;\n\treturn E_FAIL;\n}\n\nvoid CircleIndicator::onDestroy()\n{\n\tif( font )\n\t\tfont.DeleteObject();\n}\n\nvoid CircleIndicator::onPaint( CDCHandle dc )\n{\n\tCRect rectInt32;\n\tGetClientRect( &rectInt32 );\n\n\tCPaintDC pdc( m_hWnd );\n\n\tconst int logPixels = pdc.GetDeviceCaps( LOGPIXELSY );\n\tint fontSize = -MulDiv( fontSizePoints, logPixels, 72 );\n\tif( !font || fontHeight != fontSize )\n\t{\n\t\tif( font )\n\t\t\tfont.DeleteObject();\n\t\tHRESULT hr = createFont( fontSize );\n\t\tif( FAILED( hr ) )\n\t\t\treturn;\n\t\tfontHeight = fontSize;\n\t}\n\n\tpdc.SetBkColor( GetSysColor( COLOR_3DFACE ) );\n\tpdc.SelectFont( font );\n\tpdc.SetBkMode( OPAQUE );\n\tconstexpr UINT textFormat = DT_CENTER | DT_VCENTER;\n\n\tif( isActive )\n\t{\n\t\tpdc.SetTextColor( activeColor );\n\t\tpdc.DrawText( circleFilled, 1, rectInt32, textFormat );\n\t\tpdc.SetBkMode( TRANSPARENT );\n\t}\n\n\tpdc.SetTextColor( 0 );\n\tpdc.DrawText( circleOutline, 1, rectInt32, textFormat );\n}\n\nvoid CircleIndicator::setActive( bool nowActive )\n{\n\tif( nowActive == isActive )\n\t\treturn;\n\n\t// Repaint the control\n\tisActive = nowActive;\n\tInvalidateRect( nullptr );\n}"
},
{
"path": "Examples/WhisperDesktop/CircleIndicator.h",
"content": "#pragma once\n#include \"Utils/miscUtils.h\"\n#include \"Utils/WTL/atlcrack.h\"\n\n// This control renders a black circle, and in the active state, the circle is filled with a bright color.\nclass CircleIndicator: public CWindowImpl\n{\npublic:\n\tstatic ATL::CWndClassInfo& GetWndClassInfo();\n\n\tBEGIN_MSG_MAP( CircleIndicator )\n\t\tMSG_WM_PAINT( onPaint )\n\t\tMSG_WM_DESTROY( onDestroy )\n\tEND_MSG_MAP()\n\n\t// Class registration\n\tstatic HRESULT registerClass();\n\n\tvoid setActive( bool nowActive );\n\n\tvoid setActiveColor( uint32_t col )\n\t{\n\t\tactiveColor = col;\n\t}\n\tCircleIndicator();\n\nprivate:\n\tbool isActive = false;\n\tuint32_t activeColor;\n\tint fontHeight = 0;\n\tCFont font;\n\tHRESULT createFont( int height );\n\n\tvoid onDestroy();\n\tvoid onPaint( CDCHandle dc );\n};"
},
{
"path": "Examples/WhisperDesktop/LoadModelDlg.cpp",
"content": "#include \"stdafx.h\"\n#include \"LoadModelDlg.h\"\n#include \"Utils/miscUtils.h\"\n#include \"Utils/logger.h\"\n#include \"ModelAdvancedDlg.h\"\n\nconstexpr int progressMaxInteger = 1024 * 8;\n\nHRESULT LoadModelDlg::show()\n{\n\tauto res = DoModal( nullptr );\n\tif( res == -1 )\n\t\treturn HRESULT_FROM_WIN32( GetLastError() );\n\tif( res == IDOK )\n\t{\n\t\tHRESULT hr = appState.lastScreenLoad();\n\t\tswitch( hr )\n\t\t{\n\t\tcase SCREEN_TRANSCRIBE:\n\t\tcase SCREEN_CAPTURE:\n\t\t\treturn hr;\n\t\tdefault:\n\t\t\treturn SCREEN_TRANSCRIBE;\n\t\t}\n\t}\n\treturn S_OK;\n}\n\nLRESULT LoadModelDlg::OnInitDialog( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled )\n{\n\t// First DDX call, hooks up variables to controls.\n\tDoDataExchange( false );\n\n\tcbConsole.initialize( m_hWnd, IDC_CONSOLE, appState );\n\timplPopulateCombobox( cbModelType, appState.source.impl );\n\tmodelPath.SetWindowTextW( appState.source.path );\n\n\tHRESULT hr = work.create( this );\n\tif( FAILED( hr ) )\n\t{\n\t\tCString text = L\"CreateThreadpoolWork failed\\n\";\n\t\ttext += formatErrorMessage( hr );\n\t\t::MessageBox( m_hWnd, text, L\"Unable to load the model\", MB_OK | MB_ICONWARNING );\n\t\treturn TRUE;\n\t}\n\n\teditorsWindows.reserve( 6 );\n\teditorsWindows = { modelPath, cbModelType, GetDlgItem( IDC_BROWSE ), GetDlgItem( IDC_MODEL_ADV ), GetDlgItem( IDOK ), GetDlgItem( IDCANCEL ) };\n\tpendingWindows.reserve( 2 );\n\tpendingWindows = { GetDlgItem( IDC_PENDING_TEXT ), progressBar };\n\n\tprogressBar.SetRange32( 0, progressMaxInteger );\n\tprogressBar.SetStep( 1 );\n\n\tappState.setupIcon( this );\n\tATLVERIFY( CenterWindow() );\n\tif( !appState.source.found || !appState.automaticallyLoadModel )\n\t\treturn 0;\n\n\t// AppState.findModelSource() method has located model parameters in registry;\n\t// Post a notification identical to the \"OK\" button click event.\n\tPostMessage( WM_COMMAND, IDOK, (LPARAM)( GetDlgItem( IDOK ).m_hWnd ) );\n\n\treturn 0;\n}\n\nLRESULT LoadModelDlg::OnBrowse( UINT, INT, HWND, BOOL& bHandled )\n{\n\tbHandled = TRUE;\n\n\tCString path;\n\tmodelPath.GetWindowText( path );\n\tif( !getOpenFileName( m_hWnd, L\"Select a GGML Model File\", L\"Binary files (*.bin)\\0*.bin\\0\\0\", path ) )\n\t\treturn 0;\n\n\tmodelPath.SetWindowText( path );\n\tappState.source.path = path;\n\treturn 0;\n}\n\nLRESULT LoadModelDlg::validationError( LPCTSTR message )\n{\n\treportError( m_hWnd, message, L\"Unable to load the model\" );\n\treturn 0;\n}\n\nLRESULT LoadModelDlg::validationError( LPCTSTR message, HRESULT hr )\n{\n\treportError( m_hWnd, message, L\"Unable to load the model\", hr );\n\treturn 0;\n}\n\nvoid LoadModelDlg::setPending( bool nowPending )\n{\n\tconst BOOL enable = nowPending ? FALSE : TRUE;\n\tfor( HWND w : editorsWindows )\n\t\t::EnableWindow( w, enable );\n\n\tconst int show = nowPending ? SW_NORMAL : SW_HIDE;\n\tfor( HWND w : pendingWindows )\n\t\t::ShowWindow( w, show );\n\n\tif( nowPending )\n\t\tprogressBar.SetMarquee( TRUE, 0 );\n\telse\n\t\tprogressBar.SetMarquee( FALSE, 0 );\n}\n\nLRESULT LoadModelDlg::OnOk( UINT, INT, HWND, BOOL& bHandled )\n{\n\tmodelPath.GetWindowText( path );\n\tif( path.GetLength() <= 0 )\n\t\treturn validationError( L\"Please select a model GGML file\" );\n\n\t{\n\t\tCAtlFile file;\n\t\tHRESULT hr = file.Create( path, GENERIC_READ, FILE_SHARE_READ, OPEN_EXISTING );\n\t\tif( FAILED( hr ) )\n\t\t\treturn validationError( L\"Unable to open the model file\", hr );\n\n\t\tULONGLONG cb = 0;\n\t\tfile.GetSize( cb );\n\t\tappState.source.sizeInBytes = cb;\n\t}\n\n\timpl = implGetValue( cbModelType );\n\tif( impl == (Whisper::eModelImplementation)0 )\n\t\treturn validationError( L\"Please select a model type\" );\n\n\tsetPending( true );\n\twork.post();\n\treturn 0;\n}\n\nstatic HRESULT loadModel( const wchar_t* path, Whisper::sModelSetup setup, const Whisper::sLoadModelCallbacks* callbacks, Whisper::iModel** pp )\n{\n\tconstexpr bool dbgTestClone = false;\n\n\tif constexpr( dbgTestClone )\n\t\tsetup.flags |= (uint32_t)Whisper::eGpuModelFlags::Cloneable;\n\n\tCComPtr res;\n\tCHECK( Whisper::loadModel( path, setup, callbacks, &res ) );\n\n\tif constexpr( dbgTestClone )\n\t{\n\t\tCComPtr clone;\n\t\tCHECK( res->clone( &clone ) );\n\t\t*pp = clone.Detach();\n\t\treturn S_OK;\n\t}\n\telse\n\t{\n\t\t*pp = res.Detach();\n\t\treturn S_OK;\n\t}\n}\n\nvoid __stdcall LoadModelDlg::poolCallback() noexcept\n{\n\tCComPtr model;\n\tclearLastError();\n\tloadError = L\"\";\n\tWhisper::sLoadModelCallbacks lmcb;\n\tlmcb.cancel = nullptr;\n\tlmcb.progress = &LoadModelDlg::progressCallback;\n\tlmcb.pv = this;\n\tWhisper::sModelSetup setup;\n\tsetup.impl = impl;\n\tsetup.flags = appState.gpuFlagsLoad();\n\tCString adapter = appState.stringLoad( L\"gpu\" );\n\tsetup.adapter = adapter;\n\tHRESULT hr = ::loadModel( path, setup, &lmcb, &model );\n\tif( SUCCEEDED( hr ) )\n\t\tappState.model = model;\n\telse\n\t\tgetLastError( loadError );\n\n\tthis->PostMessage( WM_CALLBACK_STATUS, (WPARAM)hr );\n}\n\nHRESULT __stdcall LoadModelDlg::progressCallback( double val, void* pv ) noexcept\n{\n\tLoadModelDlg& dialog = *(LoadModelDlg*)pv;\n\tconstexpr double mul = progressMaxInteger;\n\tint pos = lround( mul * val );\n\tdialog.progressBar.PostMessage( PBM_SETPOS, pos, 0 );\n\treturn S_OK;\n}\n\nLRESULT LoadModelDlg::OnCallbackStatus( UINT, WPARAM wParam, LPARAM, BOOL& bHandled )\n{\n\tsetPending( false );\n\n\tbHandled = TRUE;\n\tconst HRESULT hr = (HRESULT)wParam;\n\tif( FAILED( hr ) )\n\t{\n\t\tLPCTSTR failMessage = L\"Error loading the model\";\n\t\tif( loadError.GetLength() > 0 )\n\t\t{\n\t\t\tCString tmp = failMessage;\n\t\t\ttmp += L\"\\n\";\n\t\t\ttmp += loadError;\n\t\t\treturn validationError( tmp, hr );\n\t\t}\n\t\telse\n\t\t\treturn validationError( failMessage, hr );\n\t}\n\n\tappState.source.path = path;\n\tappState.source.impl = impl;\n\tappState.saveModelSource();\n\n\tEndDialog( IDOK );\n\treturn 0;\n}\n\nLRESULT LoadModelDlg::OnHyperlink( int idCtrl, LPNMHDR pnmh, BOOL& bHandled )\n{\n\tconst UINT code = pnmh->code;\n\tswitch( code )\n\t{\n\tcase NM_CLICK:\n\tcase NM_RETURN:\n\t\tbreak;\n\tdefault:\n\t\treturn 0;\n\t}\n\n\tPNMLINK pNMLink = (PNMLINK)pnmh;\n\tLPCTSTR url = pNMLink->item.szUrl;\n\tShellExecute( NULL, L\"open\", url, NULL, NULL, SW_SHOW );\n\tbHandled = TRUE;\n\treturn 0;\n}\n\nvoid LoadModelDlg::onModelAdvanced()\n{\n\tModelAdvancedDlg dlg{ appState };\n\tdlg.show( m_hWnd );\n}"
},
{
"path": "Examples/WhisperDesktop/LoadModelDlg.h",
"content": "#pragma once\n#include \"AppState.h\"\n#include \"Utils/WTL/atlddx.h\"\n#include \"Utils/miscUtils.h\"\n\nclass LoadModelDlg:\n\tpublic CDialogImpl,\n\tpublic CWinDataExchange,\n\tpublic iThreadPoolCallback\n{\n\tAppState& appState;\npublic:\n\tstatic constexpr UINT IDD = IDD_OPEN_MODEL;\n\tstatic constexpr UINT WM_CALLBACK_STATUS = WM_APP + 1;\n\n\tLoadModelDlg( AppState& app ) : appState( app ) { }\n\n\tHRESULT show();\n\n\tBEGIN_MSG_MAP( LoadModelDlg )\n\t\tMESSAGE_HANDLER( WM_INITDIALOG, OnInitDialog )\n\t\tON_BUTTON_CLICK( IDC_CONSOLE, cbConsole.click )\n\t\tCOMMAND_ID_HANDLER( IDCANCEL, OnCommand )\n\t\tCOMMAND_ID_HANDLER( IDOK, OnOk )\n\t\tCOMMAND_ID_HANDLER( IDC_BROWSE, OnBrowse )\n\t\tMESSAGE_HANDLER( WM_CALLBACK_STATUS, OnCallbackStatus )\n\t\tNOTIFY_ID_HANDLER( IDC_HYPERLINKS, OnHyperlink )\n\t\tON_BUTTON_CLICK( IDC_MODEL_ADV, onModelAdvanced )\n\tEND_MSG_MAP()\n\n\tBEGIN_DDX_MAP( LoadModelDlg )\n\t\tDDX_CONTROL_HANDLE( IDC_PATH, modelPath )\n\t\tDDX_CONTROL_HANDLE( IDC_MODEL_TYPE, cbModelType )\n\t\tDDX_CONTROL_HANDLE( IDC_PROGRESS, progressBar );\n\tEND_DDX_MAP()\n\t\nprivate:\n\tstd::vector editorsWindows;\n\tstd::vector pendingWindows;\n\tvoid setPending( bool nowPending );\n\n\tLRESULT OnInitDialog( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled );\n\tLRESULT OnCallbackStatus( UINT, WPARAM wParam, LPARAM, BOOL& bHandled );\n\n\tLRESULT OnCommand( UINT, INT nIdentifier, HWND, BOOL& bHandled )\n\t{\n\t\tATLVERIFY( EndDialog( nIdentifier ) );\n\t\treturn 0;\n\t}\n\tLRESULT OnBrowse( UINT, INT, HWND, BOOL& bHandled );\n\tLRESULT OnOk( UINT, INT, HWND, BOOL& bHandled );\n\n\tConsoleCheckbox cbConsole;\n\tCComboBox cbModelType;\n\tCEdit modelPath;\n\tCProgressBarCtrl progressBar;\n\n\tLRESULT validationError( LPCTSTR message );\n\tLRESULT validationError( LPCTSTR message, HRESULT hr );\n\n\tThreadPoolWork work;\n\tCString path;\n\tWhisper::eModelImplementation impl;\n\tCString loadError;\n\tvoid __stdcall poolCallback() noexcept override final;\n\n\tLRESULT OnHyperlink( int idCtrl, LPNMHDR pnmh, BOOL& bHandled );\n\n\tstatic HRESULT __stdcall progressCallback( double val, void* pv ) noexcept;\n\n\tvoid onModelAdvanced();\n};"
},
{
"path": "Examples/WhisperDesktop/ModelAdvancedDlg.cpp",
"content": "#include \"stdafx.h\"\n#include \"ModelAdvancedDlg.h\"\nusing Whisper::eGpuModelFlags;\n\nstatic void __stdcall addGpu( const wchar_t* name, void* pv )\n{\n\tCComboBox& cb = *(CComboBox*)pv;\n\tcb.AddString( name );\n}\n\nLRESULT ModelAdvancedDlg::onInitDialog( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled )\n{\n\tcbWave = GetDlgItem( IDC_WAVE );\n\tcbReshapedMatMul = GetDlgItem( IDC_RESHAPED_MAT_MUL );\n\tcbAdapter = GetDlgItem( IDC_GPU );\n\tconst uint32_t flags = appState.gpuFlagsLoad();\n\n\t// Setup the \"Compute shaders\" combobox\n\tcbWave.AddString( L\"Wave64 shaders on AMD\" );\n\tcbWave.AddString( L\"Wave32, always\" );\n\tcbWave.AddString( L\"Wave64, always\" );\n\tint i = 0;\n\tif( 0 != ( flags & (uint32_t)eGpuModelFlags::Wave32 ) )\n\t\ti = 1;\n\telse if( 0 != ( flags & (uint32_t)eGpuModelFlags::Wave64 ) )\n\t\ti = 2;\n\tcbWave.SetCurSel( i );\n\n\t// Setup the \"Reshaped multiply\" combobox\n\tcbReshapedMatMul.AddString( L\"Reshape on AMD\" );\n\tcbReshapedMatMul.AddString( L\"Don’t reshape tensors\" );\n\tcbReshapedMatMul.AddString( L\"Reshape some tensors\" );\n\ti = 0;\n\tif( 0 != ( flags & (uint32_t)eGpuModelFlags::NoReshapedMatMul ) )\n\t\ti = 1;\n\telse if( 0 != ( flags & (uint32_t)eGpuModelFlags::UseReshapedMatMul ) )\n\t\ti = 2;\n\tcbReshapedMatMul.SetCurSel( i );\n\n\tcbAdapter.AddString( L\"Use default\" );\n\tHRESULT hr = Whisper::listGPUs( &addGpu, &cbAdapter );\n\tif( FAILED( hr ) )\n\t{\n\t\treportError( m_hWnd, L\"Unable to enumerate GPUs\", L\"\", hr );\n\t\treturn 0;\n\t}\n\tconst CString setting = appState.stringLoad( L\"gpu\" );\n\tif( setting.GetLength() <= 0 )\n\t\tcbAdapter.SetCurSel( 0 );\n\telse\n\t\tcbAdapter.SetCurSel( cbAdapter.FindStringExact( 1, setting ) );\n\n\treturn 0;\n}\n\nbool ModelAdvancedDlg::show( HWND owner )\n{\n\tauto res = DoModal( owner );\n\treturn res == IDOK;\n}\n\nvoid ModelAdvancedDlg::onOk()\n{\n\t// Gather values from these comboboxes\n\tuint32_t flags = 0;\n\n\tint i = cbWave.GetCurSel();\n\tif( 1 == i )\n\t\tflags |= (uint32_t)eGpuModelFlags::Wave32;\n\telse if( 2 == i )\n\t\tflags |= (uint32_t)eGpuModelFlags::Wave64;\n\n\ti = cbReshapedMatMul.GetCurSel();\n\tif( 1 == i )\n\t\tflags |= (uint32_t)eGpuModelFlags::NoReshapedMatMul;\n\telse if( 2 == i )\n\t\tflags |= (uint32_t)eGpuModelFlags::UseReshapedMatMul;\n\n\t// Save to registry\n\tappState.gpuFlagsStore( flags );\n\n\tint gpuIndex = cbAdapter.GetCurSel();\n\tif( gpuIndex <= 0 )\n\t\tappState.stringStore( L\"gpu\", nullptr );\n\telse\n\t{\n\t\tCString gpu;\n\t\tcbAdapter.GetLBText( gpuIndex, gpu );\n\t\tappState.stringStore( L\"gpu\", gpu );\n\t}\n\n\tEndDialog( IDOK );\n}"
},
{
"path": "Examples/WhisperDesktop/ModelAdvancedDlg.h",
"content": "#pragma once\n#include \"AppState.h\"\n#include \"Utils/WTL/atlddx.h\"\n#include \"Utils/miscUtils.h\"\n\nclass ModelAdvancedDlg :\n\tpublic CDialogImpl\n{\n\tCComboBox cbWave, cbReshapedMatMul, cbAdapter;\n\tAppState& appState;\n\npublic:\n\tstatic constexpr UINT IDD = IDD_MODEL_ADV;\n\n\tModelAdvancedDlg( AppState& app ) : appState( app ) { }\n\n\tBEGIN_MSG_MAP( ModelAdvancedDlg )\n\t\tMESSAGE_HANDLER( WM_INITDIALOG, onInitDialog )\n\t\tON_BUTTON_CLICK( IDOK, onOk )\n\t\tON_BUTTON_CLICK( IDCANCEL, onCancel )\n\tEND_MSG_MAP()\n\n\tbool show( HWND owner );\n\nprivate:\n\n\tLRESULT onInitDialog( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled );\n\n\tvoid onOk();\n\n\tvoid onCancel()\n\t{\n\t\tEndDialog( IDCANCEL );\n\t}\n};"
},
{
"path": "Examples/WhisperDesktop/Readme.txt",
"content": "This example shows how to consume the DLL from a C++ GUI application.\n\nThe GUI is implemented with ATL and WTL libraries."
},
{
"path": "Examples/WhisperDesktop/Resource.h",
"content": "//{{NO_DEPENDENCIES}}\n// Microsoft Visual C++ generated include file.\n// Used by WhisperDesktop.rc\n//\n#define IDC_MYICON 2\n#define IDD_WHISPERDESKTOP_DIALOG 102\n#define IDM_ABOUT 104\n#define IDI_WHISPERDESKTOP 107\n#define IDI_SMALL 108\n#define IDR_MAINFRAME 128\n#define IDD_OPEN_MODEL 129\n#define IDD_MAIN_DIALOG 130\n#define IDD_TRANSCRIBE_DIALOG 130\n#define IDD_CAPTURE_DIALOG 131\n#define IDD_MODEL_ADV 132\n#define IDC_PATH 1000\n#define IDC_BROWSE 1001\n#define IDC_MODEL_TYPE 1002\n#define IDC_PATH_RESULT 1002\n#define IDC_PROGRESS 1003\n#define IDC_BROWSE_RESULT 1003\n#define IDC_SYSLINK1 1004\n#define IDC_HYPERLINKS 1004\n#define IDC_TRANSCRIBE_PROGRESS 1004\n#define IDC_PENDING_TEXT 1005\n#define IDC_MODEL_DESC 1006\n#define IDC_LANGUAGE 1007\n#define IDC_OUTPUT_FORMAT 1008\n#define IDC_PATH_MEDIA 1009\n#define IDC_DEVICE 1009\n#define IDC_BROWSE_MEDIA 1010\n#define IDC_TRANSCRIBE 1011\n#define IDC_BACK 1012\n#define IDC_CHECK1 1013\n#define IDC_CONSOLE 1013\n#define IDC_CAPTURE 1014\n#define IDC_DEV_REFRESH 1015\n#define IDC_SAVE_TEXT 1016\n#define IDC_SAVE_APPEND 1017\n#define IDC_SAVE_TIMESTAMPS 1018\n#define IDC_RUN_CAPTURE 1019\n#define IDC_VOICE_ACTIVITY 1020\n#define IDC_VOICE_ACTIVITY_LBL 1021\n#define IDC_TRANS_STATUS 1022\n#define IDC_TRANS_LBL 1023\n#define IDC_STALL_STATUS 1024\n#define IDC_STALL_LBL 1025\n#define IDC_TRANSLATE 1026\n#define IDC_MODEL_ADV 1027\n#define IDC_WAVE 1028\n#define IDC_RESHAPED_MAT_MUL 1029\n#define IDC_CHECK2 1029\n#define IDC_USE_INPUT_FOLDER 1029\n#define IDC_RESHAPED_MAT_MUL2 1030\n#define IDC_GPU 1030\n#define IDC_STATIC -1\n\n// Next default values for new objects\n// \n#ifdef APSTUDIO_INVOKED\n#ifndef APSTUDIO_READONLY_SYMBOLS\n#define _APS_NO_MFC 1\n#define _APS_NEXT_RESOURCE_VALUE 131\n#define _APS_NEXT_COMMAND_VALUE 32771\n#define _APS_NEXT_CONTROL_VALUE 1030\n#define _APS_NEXT_SYMED_VALUE 110\n#endif\n#endif\n"
},
{
"path": "Examples/WhisperDesktop/TranscribeDlg.cpp",
"content": "#include \"stdafx.h\"\n#include \"TranscribeDlg.h\"\n#include \"Utils/logger.h\"\n\nHRESULT TranscribeDlg::show()\n{\n\tauto res = DoModal( nullptr );\n\tif( res == -1 )\n\t\treturn HRESULT_FROM_WIN32( GetLastError() );\n\tswitch( res )\n\t{\n\tcase IDC_BACK:\n\t\treturn SCREEN_MODEL;\n\tcase IDC_CAPTURE:\n\t\treturn SCREEN_CAPTURE;\n\t}\n\treturn S_OK;\n}\n\nconstexpr int progressMaxInteger = 1024 * 8;\n\nstatic const LPCTSTR regValInput = L\"sourceMedia\";\nstatic const LPCTSTR regValOutFormat = L\"resultFormat\";\nstatic const LPCTSTR regValOutPath = L\"resultPath\";\nstatic const LPCTSTR regValUseInputFolder = L\"useInputFolder\";\n\nLRESULT TranscribeDlg::OnInitDialog( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled )\n{\n\t// First DDX call, hooks up variables to controls.\n\tDoDataExchange( false );\n\tprintModelDescription();\n\tlanguageSelector.initialize( m_hWnd, IDC_LANGUAGE, appState );\n\tcbConsole.initialize( m_hWnd, IDC_CONSOLE, appState );\n\tcbTranslate.initialize( m_hWnd, IDC_TRANSLATE, appState );\n\tpopulateOutputFormats();\n\n\tpendingState.initialize(\n\t\t{\n\t\t\tlanguageSelector, GetDlgItem( IDC_TRANSLATE ),\n\t\t\tsourceMediaPath, GetDlgItem( IDC_BROWSE_MEDIA ),\n\t\t\ttranscribeOutFormat, useInputFolder,\n\t\t\ttranscribeOutputPath, GetDlgItem( IDC_BROWSE_RESULT ),\n\t\t\tGetDlgItem( IDCANCEL ),\n\t\t\tGetDlgItem( IDC_BACK ),\n\t\t\tGetDlgItem( IDC_CAPTURE )\n\t\t},\n\t\t{\n\t\t\tprogressBar, GetDlgItem( IDC_PENDING_TEXT )\n\t\t} );\n\n\tHRESULT hr = work.create( this );\n\tif( FAILED( hr ) )\n\t{\n\t\treportError( m_hWnd, L\"CreateThreadpoolWork failed\", nullptr, hr );\n\t\tEndDialog( IDCANCEL );\n\t}\n\n\tprogressBar.SetRange32( 0, progressMaxInteger );\n\tprogressBar.SetStep( 1 );\n\n\tsourceMediaPath.SetWindowText( appState.stringLoad( regValInput ) );\n\ttranscribeOutFormat.SetCurSel( (int)appState.dwordLoad( regValOutFormat, 0 ) );\n\ttranscribeOutputPath.SetWindowText( appState.stringLoad( regValOutPath ) );\n\tif( appState.boolLoad( regValUseInputFolder ) )\n\t\tuseInputFolder.SetCheck( BST_CHECKED );\n\tBOOL unused;\n\tonOutFormatChange( 0, 0, nullptr, unused );\n\n\tappState.lastScreenSave( SCREEN_TRANSCRIBE );\n\tappState.setupIcon( this );\n\tATLVERIFY( CenterWindow() );\n\treturn 0;\n}\n\nvoid TranscribeDlg::printModelDescription()\n{\n\tCString text;\n\tif( S_OK == appState.model->isMultilingual() )\n\t\ttext = L\"Multilingual\";\n\telse\n\t\ttext = L\"Single-language\";\n\ttext += L\" model \\\"\";\n\tLPCTSTR path = appState.source.path;\n\tpath = ::PathFindFileName( path );\n\ttext += path;\n\ttext += L\"\\\", \";\n\tconst int64_t cb = appState.source.sizeInBytes;\n\tif( cb < 1 << 30 )\n\t{\n\t\tconstexpr double mul = 1.0 / ( 1 << 20 );\n\t\tdouble mb = (double)cb * mul;\n\t\ttext.AppendFormat( L\"%.1f MB\", mb );\n\t}\n\telse\n\t{\n\t\tconstexpr double mul = 1.0 / ( 1 << 30 );\n\t\tdouble gb = (double)cb * mul;\n\t\ttext.AppendFormat( L\"%.2f GB\", gb );\n\t}\n\ttext += L\" on disk, \";\n\ttext += implString( appState.source.impl );\n\ttext += L\" implementation\";\n\n\tmodelDesc.SetWindowText( text );\n}\n\n// Populate the \"Output Format\" combobox\nvoid TranscribeDlg::populateOutputFormats()\n{\n\ttranscribeOutFormat.AddString( L\"None\" );\n\ttranscribeOutFormat.AddString( L\"Text file\" );\n\ttranscribeOutFormat.AddString( L\"Text with timestamps\" );\n\ttranscribeOutFormat.AddString( L\"SubRip subtitles\" );\n\ttranscribeOutFormat.AddString( L\"WebVTT subtitles\" );\n}\n\n// The enum values should match 0-based indices of the combobox items\nenum struct TranscribeDlg::eOutputFormat : uint8_t\n{\n\tNone = 0,\n\tText = 1,\n\tTextTimestamps = 2,\n\tSubRip = 3,\n\tWebVTT = 4,\n};\n\nenum struct TranscribeDlg::eVisualState : uint8_t\n{\n\tIdle = 0,\n\tRunning = 1,\n\tStopping = 2\n};\n\n// CBN_SELCHANGE notification for IDC_OUTPUT_FORMAT combobox\nLRESULT TranscribeDlg::onOutFormatChange( UINT, INT, HWND, BOOL& bHandled )\n{\n\tBOOL enabled = transcribeOutFormat.GetCurSel() != 0;\n\tuseInputFolder.EnableWindow( enabled );\n\n\tif( isChecked( useInputFolder ) && enabled )\n\t{\n\t\tenabled = FALSE;\n\t\tsetOutputPath();\n\t}\n\ttranscribeOutputPath.EnableWindow( enabled );\n\ttranscribeOutputBrowse.EnableWindow( enabled );\n\n\treturn 0;\n}\n\n// EN_CHANGE notification for IDC_PATH_MEDIA edit box\nLRESULT TranscribeDlg::onInputChange( UINT, INT, HWND, BOOL& )\n{\n\tif( !useInputFolder.IsWindowEnabled() )\n\t\treturn 0;\n\tif( !isChecked( useInputFolder ) )\n\t\treturn 0;\n\tsetOutputPath();\n\treturn 0;\n}\n\nvoid TranscribeDlg::onBrowseMedia()\n{\n\tLPCTSTR title = L\"Input audio file to transcribe\";\n\tLPCTSTR filters = L\"Multimedia Files\\0*.wav;*.wave;*.mp3;*.wma;*.mp4;*.mpeg4;*.mkv;*.m4a\\0\\0\";\n\n\tCString path;\n\tsourceMediaPath.GetWindowText( path );\n\tif( !getOpenFileName( m_hWnd, title, filters, path ) )\n\t\treturn;\n\tsourceMediaPath.SetWindowText( path );\n\tif( useInputFolder.IsWindowEnabled() && useInputFolder.GetCheck() == BST_CHECKED )\n\t\tsetOutputPath( path );\n}\n\nstatic const LPCTSTR outputFilters = L\"Text files (*.txt)\\0*.txt\\0Text with timestamps (*.txt)\\0*.txt\\0SubRip subtitles (*.srt)\\0*.srt\\0WebVTT subtitles (*.vtt)\\0*.vtt\\0\\0\";\nstatic const std::array outputExtensions =\n{\n\tL\".txt\", L\".txt\", L\".srt\", L\".vtt\"\n};\n\nvoid TranscribeDlg::setOutputPath( const CString& input )\n{\n\tconst int format = transcribeOutFormat.GetCurSel() - 1;\n\tif( format < 0 || format >= outputExtensions.size() )\n\t\treturn;\n\tconst LPCTSTR ext = outputExtensions[ format ];\n\tCString path = input;\n\tpath.Trim();\n\tconst bool renamed = PathRenameExtension( path.GetBufferSetLength( path.GetLength() + 4 ), ext );\n\tpath.ReleaseBuffer();\n\tif( !renamed )\n\t\treturn;\n\ttranscribeOutputPath.SetWindowText( path );\n}\n\nvoid TranscribeDlg::setOutputPath()\n{\n\tCString path;\n\tif( !sourceMediaPath.GetWindowText( path ) )\n\t\treturn;\n\tif( path.GetLength() <= 0 )\n\t\treturn;\n\tsetOutputPath( path );\n}\n\nvoid TranscribeDlg::onInputFolderCheck()\n{\n\tconst bool checked = isChecked( useInputFolder );\n\n\tBOOL enableOutput = checked ? FALSE : TRUE;\n\ttranscribeOutputPath.EnableWindow( enableOutput );\n\ttranscribeOutputBrowse.EnableWindow( enableOutput );\n\n\tif( !checked )\n\t\treturn;\n\tsetOutputPath();\n}\n\nvoid TranscribeDlg::onBrowseOutput()\n{\n\tconst DWORD origFilterIndex = (DWORD)transcribeOutFormat.GetCurSel() - 1;\n\n\tLPCTSTR title = L\"Output Text File\";\n\tCString path;\n\ttranscribeOutputPath.GetWindowText( path );\n\tDWORD filterIndex = origFilterIndex;\n\tif( !getSaveFileName( m_hWnd, title, outputFilters, path, &filterIndex ) )\n\t\treturn;\n\n\tLPCTSTR ext = PathFindExtension( path );\n\tif( 0 == *ext && filterIndex < outputExtensions.size() )\n\t{\n\t\twchar_t* const buffer = path.GetBufferSetLength( path.GetLength() + 5 );\n\t\tPathRenameExtension( buffer, outputExtensions[ filterIndex ] );\n\t\tpath.ReleaseBuffer();\n\t}\n\n\ttranscribeOutputPath.SetWindowText( path );\n\tif( filterIndex != origFilterIndex )\n\t\ttranscribeOutFormat.SetCurSel( filterIndex + 1 );\n}\n\nvoid TranscribeDlg::setPending( bool nowPending )\n{\n\tpendingState.setPending( nowPending );\n}\n\nvoid TranscribeDlg::transcribeError( LPCTSTR text, HRESULT hr )\n{\n\treportError( m_hWnd, text, L\"Unable to transcribe audio\", hr );\n}\n\nvoid TranscribeDlg::onTranscribe()\n{\n\tswitch( transcribeArgs.visualState )\n\t{\n\tcase eVisualState::Running:\n\t\ttranscribeArgs.visualState = eVisualState::Stopping;\n\t\ttranscribeButton.EnableWindow( FALSE );\n\t\treturn;\n\tcase eVisualState::Stopping:\n\t\treturn;\n\t}\n\n\t// Validate input\n\tsourceMediaPath.GetWindowText( transcribeArgs.pathMedia );\n\tif( transcribeArgs.pathMedia.GetLength() <= 0 )\n\t{\n\t\ttranscribeError( L\"Please select an input audio file\" );\n\t\treturn;\n\t}\n\n\tif( !PathFileExists( transcribeArgs.pathMedia ) )\n\t{\n\t\ttranscribeError( L\"Input audio file does not exist\", HRESULT_FROM_WIN32( ERROR_FILE_NOT_FOUND ) );\n\t\treturn;\n\t}\n\n\ttranscribeArgs.language = languageSelector.selectedLanguage();\n\ttranscribeArgs.translate = cbTranslate.checked();\n\tif( isInvalidTranslate( m_hWnd, transcribeArgs.language, transcribeArgs.translate ) )\n\t\treturn;\n\n\ttranscribeArgs.format = (eOutputFormat)(uint8_t)transcribeOutFormat.GetCurSel();\n\tif( transcribeArgs.format != eOutputFormat::None )\n\t{\n\t\ttranscribeOutputPath.GetWindowText( transcribeArgs.pathOutput );\n\t\tif( transcribeArgs.pathOutput.GetLength() <= 0 )\n\t\t{\n\t\t\ttranscribeError( L\"Please select an output text file\" );\n\t\t\treturn;\n\t\t}\n\t\tif( PathFileExists( transcribeArgs.pathOutput ) )\n\t\t{\n\t\t\tconst int resp = MessageBox( L\"The output file is already there.\\nOverwrite the file?\", L\"Confirm Overwrite\", MB_ICONQUESTION | MB_YESNO );\n\t\t\tif( resp != IDYES )\n\t\t\t\treturn;\n\t\t}\n\t\tappState.stringStore( regValOutPath, transcribeArgs.pathOutput );\n\t}\n\telse\n\t\tcbConsole.ensureChecked();\n\n\tappState.dwordStore( regValOutFormat, (uint32_t)(int)transcribeArgs.format );\n\tappState.boolStore( regValUseInputFolder, isChecked( useInputFolder ) );\n\tlanguageSelector.saveSelection( appState );\n\tcbTranslate.saveSelection( appState );\n\tappState.stringStore( regValInput, transcribeArgs.pathMedia );\n\n\tsetPending( true );\n\ttranscribeArgs.visualState = eVisualState::Running;\n\ttranscribeButton.SetWindowText( L\"Stop\" );\n\twork.post();\n}\n\nvoid __stdcall TranscribeDlg::poolCallback() noexcept\n{\n\tHRESULT hr = transcribe();\n\tPostMessage( WM_CALLBACK_STATUS, (WPARAM)hr );\n}\n\nstatic void printTime( CString& rdi, int64_t ticks )\n{\n\tconst Whisper::sTimeSpan ts{ (uint64_t)ticks };\n\tconst Whisper::sTimeSpanFields fields = ts;\n\n\tif( fields.days != 0 )\n\t{\n\t\trdi.AppendFormat( L\"%i days, %i hours\", fields.days, (int)fields.hours );\n\t\treturn;\n\t}\n\tif( ( fields.hours | fields.minutes ) != 0 )\n\t{\n\t\trdi.AppendFormat( L\"%02d:%02d:%02d\", (int)fields.hours, (int)fields.minutes, (int)fields.seconds );\n\t\treturn;\n\t}\n\trdi.AppendFormat( L\"%.3f seconds\", (double)ticks / 1E7 );\n}\n\nLRESULT TranscribeDlg::onCallbackStatus( UINT, WPARAM wParam, LPARAM, BOOL& bHandled )\n{\n\tsetPending( false );\n\ttranscribeButton.SetWindowText( L\"Transcribe\" );\n\ttranscribeButton.EnableWindow( TRUE );\n\tconst bool prematurely = ( transcribeArgs.visualState == eVisualState::Stopping );\n\ttranscribeArgs.visualState = eVisualState::Idle;\n\n\tconst HRESULT hr = (HRESULT)wParam;\n\tif( FAILED( hr ) )\n\t{\n\t\tLPCTSTR failMessage = L\"Transcribe failed\";\n\n\t\tif( transcribeArgs.errorMessage.GetLength() > 0 )\n\t\t{\n\t\t\tCString tmp = failMessage;\n\t\t\ttmp += L\"\\n\";\n\t\t\ttmp += transcribeArgs.errorMessage;\n\t\t\ttranscribeError( tmp, hr );\n\t\t}\n\t\telse\n\t\t\ttranscribeError( failMessage, hr );\n\n\t\treturn 0;\n\t}\n\n\tconst int64_t elapsed = ( GetTickCount64() - transcribeArgs.startTime ) * 10'000;\n\tconst int64_t media = transcribeArgs.mediaDuration;\n\tCString message;\n\tif( prematurely )\n\t\tmessage = L\"Transcribed an initial portion of the audio\";\n\telse\n\t\tmessage = L\"Transcribed the audio\";\n\tmessage += L\"\\nMedia duration: \";\n\tprintTime( message, media );\n\tmessage += L\"\\nProcessing time: \";\n\tprintTime( message, elapsed );\n\tmessage += L\"\\nRelative processing speed: \";\n\tdouble mul = (double)media / (double)elapsed;\n\tmessage.AppendFormat( L\"%g\", mul );\n\n\tMessageBox( message, L\"Transcribe Completed\", MB_OK | MB_ICONINFORMATION );\n\treturn 0;\n}\n\nvoid TranscribeDlg::getThreadError()\n{\n\tgetLastError( transcribeArgs.errorMessage );\n}\n\n#define CHECK_EX( hr ) { const HRESULT __hr = ( hr ); if( FAILED( __hr ) ) { getThreadError(); return __hr; } }\n\nHRESULT TranscribeDlg::transcribe()\n{\n\ttranscribeArgs.startTime = GetTickCount64();\n\tclearLastError();\n\ttranscribeArgs.errorMessage = L\"\";\n\n\tusing namespace Whisper;\n\tCComPtr reader;\n\n\tCHECK_EX( appState.mediaFoundation->openAudioFile( transcribeArgs.pathMedia, false, &reader ) );\n\n\tconst eOutputFormat format = transcribeArgs.format;\n\tCAtlFile outputFile;\n\tif( format != eOutputFormat::None )\n\t\tCHECK( outputFile.Create( transcribeArgs.pathOutput, GENERIC_WRITE, 0, CREATE_ALWAYS ) );\n\n\ttranscribeArgs.resultFlags = eResultFlags::Timestamps | eResultFlags::Tokens;\n\n\tCComPtr context;\n\tCHECK_EX( appState.model->createContext( &context ) );\n\n\tsFullParams fullParams;\n\tCHECK_EX( context->fullDefaultParams( eSamplingStrategy::Greedy, &fullParams ) );\n\tfullParams.language = transcribeArgs.language;\n\tfullParams.setFlag( eFullParamsFlags::Translate, transcribeArgs.translate );\n\tfullParams.resetFlag( eFullParamsFlags::PrintRealtime );\n\n\t// Setup the callbacks\n\tfullParams.new_segment_callback = &newSegmentCallbackStatic;\n\tfullParams.new_segment_callback_user_data = this;\n\tfullParams.encoder_begin_callback = &encoderBeginCallback;\n\tfullParams.encoder_begin_callback_user_data = this;\n\n\t// Setup the progress indication sink\n\tsProgressSink progressSink{ &progressCallbackStatic, this };\n\t// Run the transcribe\n\tCHECK_EX( context->runStreamed( fullParams, progressSink, reader ) );\n\n\t// Once finished, query duration of the audio.\n\t// The duration before the processing is sometimes different, by 20 seconds for the file in that issue:\n\t// https://github.com/Const-me/Whisper/issues/4\n\tCHECK_EX( reader->getDuration( transcribeArgs.mediaDuration ) );\n\n\tcontext->timingsPrint();\n\n\tif( format == eOutputFormat::None )\n\t\treturn S_OK;\n\n\tCComPtr result;\n\tCHECK_EX( context->getResults( transcribeArgs.resultFlags, &result ) );\n\n\tsTranscribeLength len;\n\tCHECK_EX( result->getSize( len ) );\n\tconst sSegment* const segments = result->getSegments();\n\n\tswitch( format )\n\t{\n\tcase eOutputFormat::Text:\n\t\treturn writeTextFile( segments, len.countSegments, outputFile, false );\n\tcase eOutputFormat::TextTimestamps:\n\t\treturn writeTextFile( segments, len.countSegments, outputFile, true );\n\tcase eOutputFormat::SubRip:\n\t\treturn writeSubRip( segments, len.countSegments, outputFile );\n\tcase eOutputFormat::WebVTT:\n\t\treturn writeWebVTT( segments, len.countSegments, outputFile );\n\tdefault:\n\t\treturn E_FAIL;\n\t}\n}\n\n#undef CHECK_EX\n\ninline HRESULT TranscribeDlg::progressCallback( double p ) noexcept\n{\n\tconstexpr double mul = progressMaxInteger;\n\tint pos = lround( mul * p );\n\tprogressBar.PostMessage( PBM_SETPOS, pos, 0 );\n\treturn S_OK;\n}\n\nHRESULT __cdecl TranscribeDlg::progressCallbackStatic( double p, Whisper::iContext* ctx, void* pv ) noexcept\n{\n\tTranscribeDlg* dlg = (TranscribeDlg*)pv;\n\treturn dlg->progressCallback( p );\n}\n\nnamespace\n{\n\tHRESULT write( CAtlFile& file, const CStringA& line )\n\t{\n\t\tif( line.GetLength() > 0 )\n\t\t\tCHECK( file.Write( cstr( line ), (DWORD)line.GetLength() ) );\n\t\treturn S_OK;\n\t}\n\n\tconst char* skipBlank( const char* rsi )\n\t{\n\t\twhile( true )\n\t\t{\n\t\t\tconst char c = *rsi;\n\t\t\tif( c == ' ' || c == '\\t' )\n\t\t\t{\n\t\t\t\trsi++;\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\treturn rsi;\n\t\t}\n\t}\n}\n\nusing Whisper::sSegment;\n\n\nHRESULT TranscribeDlg::writeTextFile( const sSegment* const segments, const size_t length, CAtlFile& file, bool timestamps )\n{\n\tusing namespace Whisper;\n\tCHECK( writeUtf8Bom( file ) );\n\tCStringA line;\n\tfor( size_t i = 0; i < length; i++ )\n\t{\n\t\tconst sSegment& seg = segments[ i ];\n\n\t\tif( timestamps )\n\t\t{\n\t\t\tline = \"[\";\n\t\t\tprintTime( line, seg.time.begin );\n\t\t\tline += \" --> \";\n\t\t\tprintTime( line, seg.time.end );\n\t\t\tline += \"] \";\n\t\t}\n\t\telse\n\t\t\tline = \"\";\n\n\t\tline += skipBlank( seg.text );\n\t\tline += \"\\r\\n\";\n\t\tCHECK( write( file, line ) );\n\t}\n\treturn S_OK;\n}\n\nHRESULT TranscribeDlg::writeSubRip( const sSegment* const segments, const size_t length, CAtlFile& file )\n{\n\tCHECK( writeUtf8Bom( file ) );\n\tCStringA line;\n\tfor( size_t i = 0; i < length; i++ )\n\t{\n\t\tconst sSegment& seg = segments[ i ];\n\n\t\tline.Format( \"%zu\\r\\n\", i + 1 );\n\t\tprintTime( line, seg.time.begin, true );\n\t\tline += \" --> \";\n\t\tprintTime( line, seg.time.end, true );\n\t\tline += \"\\r\\n\";\n\t\tline += skipBlank( seg.text );\n\t\tline += \"\\r\\n\\r\\n\";\n\t\tCHECK( write( file, line ) );\n\t}\n\treturn S_OK;\n}\n\nHRESULT TranscribeDlg::writeWebVTT( const sSegment* const segments, const size_t length, CAtlFile& file )\n{\n\tCHECK( writeUtf8Bom( file ) );\n\tCStringA line;\n\tline = \"WEBVTT\\r\\n\\r\\n\";\n\tCHECK( write( file, line ) );\n\n\tfor( size_t i = 0; i < length; i++ )\n\t{\n\t\tconst sSegment& seg = segments[ i ];\n\t\tline = \"\";\n\n\t\tprintTime( line, seg.time.begin, false );\n\t\tline += \" --> \";\n\t\tprintTime( line, seg.time.end, false );\n\t\tline += \"\\r\\n\";\n\t\tline += skipBlank( seg.text );\n\t\tline += \"\\r\\n\\r\\n\";\n\t\tCHECK( write( file, line ) );\n\t}\n\treturn S_OK;\n}\n\ninline HRESULT TranscribeDlg::newSegmentCallback( Whisper::iContext* ctx, uint32_t n_new )\n{\n\tusing namespace Whisper;\n\tCComPtr result;\n\tCHECK( ctx->getResults( transcribeArgs.resultFlags, &result ) );\n\treturn logNewSegments( result, n_new );\n}\n\nHRESULT __cdecl TranscribeDlg::newSegmentCallbackStatic( Whisper::iContext* ctx, uint32_t n_new, void* user_data ) noexcept\n{\n\tTranscribeDlg* dlg = (TranscribeDlg*)user_data;\n\treturn dlg->newSegmentCallback( ctx, n_new );\n}\n\nHRESULT __cdecl TranscribeDlg::encoderBeginCallback( Whisper::iContext* ctx, void* user_data ) noexcept\n{\n\tTranscribeDlg* dlg = (TranscribeDlg*)user_data;\n\tconst eVisualState visualState = dlg->transcribeArgs.visualState;\n\tswitch( visualState )\n\t{\n\tcase eVisualState::Idle:\n\t\treturn E_NOT_VALID_STATE;\n\tcase eVisualState::Running:\n\t\treturn S_OK;\n\tcase eVisualState::Stopping:\n\t\treturn S_FALSE;\n\tdefault:\n\t\treturn E_UNEXPECTED;\n\t}\n}\n\nvoid TranscribeDlg::onWmClose()\n{\n\tif( GetDlgItem( IDCANCEL ).IsWindowEnabled() )\n\t{\n\t\tEndDialog( IDCANCEL );\n\t\treturn;\n\t}\n\n\tconstexpr UINT flags = MB_YESNO | MB_ICONQUESTION | MB_DEFBUTTON2;\n\tconst int res = this->MessageBox( L\"Transcribe is in progress.\\nDo you want to quit anyway?\", L\"Confirm exit\", flags );\n\tif( res != IDYES )\n\t\treturn;\n\n\t// TODO: instead of ExitProcess(), implement another callback in the DLL API, for proper cancellation of the background task\n\tExitProcess( 1 );\n}"
},
{
"path": "Examples/WhisperDesktop/TranscribeDlg.h",
"content": "#pragma once\n#include \"AppState.h\"\n#include \"Utils/WTL/atlddx.h\"\n#include \"Utils/WTL/atlcrack.h\"\n#include \"Utils/miscUtils.h\"\n#include \"Utils/LanguageDropdown.h\"\n#include \"Utils/TranslateCheckbox.h\"\n#include \"Utils/PendingState.h\"\n\nclass TranscribeDlg :\n\tpublic CDialogImpl,\n\tpublic CWinDataExchange,\n\tpublic iThreadPoolCallback\n{\n\tAppState& appState;\n\npublic:\n\tstatic constexpr UINT IDD = IDD_TRANSCRIBE_DIALOG;\n\tstatic constexpr UINT WM_CALLBACK_STATUS = WM_APP + 1;\n\n\tTranscribeDlg( AppState& app ) : appState( app ) { }\n\n\t// Show this dialog modally, without parent.\n\tHRESULT show();\n\n\tBEGIN_MSG_MAP( LoadModelDlg )\n\t\tMESSAGE_HANDLER( WM_INITDIALOG, OnInitDialog )\n\t\tON_BUTTON_CLICK( IDC_CONSOLE, cbConsole.click )\n\t\tON_BUTTON_CLICK( IDCANCEL, onClose )\n\t\tON_BUTTON_CLICK( IDC_BACK, onBack )\n\t\tON_BUTTON_CLICK( IDC_USE_INPUT_FOLDER, onInputFolderCheck )\n\t\tON_BUTTON_CLICK( IDC_BROWSE_MEDIA, onBrowseMedia )\n\t\tON_BUTTON_CLICK( IDC_BROWSE_RESULT, onBrowseOutput )\n\t\tON_BUTTON_CLICK( IDC_TRANSCRIBE, onTranscribe )\n\t\tON_BUTTON_CLICK( IDC_CAPTURE, onCapture )\n\t\tCOMMAND_HANDLER( IDC_OUTPUT_FORMAT, CBN_SELCHANGE, onOutFormatChange )\n\t\tCOMMAND_HANDLER( IDC_PATH_MEDIA, EN_CHANGE, onInputChange )\n\t\tMESSAGE_HANDLER( WM_CALLBACK_STATUS, onCallbackStatus )\n\t\tMSG_WM_CLOSE( onWmClose )\n\tEND_MSG_MAP()\n\n\tBEGIN_DDX_MAP( LoadModelDlg )\n\t\tDDX_CONTROL_HANDLE( IDC_MODEL_DESC, modelDesc )\n\t\tDDX_CONTROL_HANDLE( IDC_PATH_MEDIA, sourceMediaPath )\n\t\tDDX_CONTROL_HANDLE( IDC_OUTPUT_FORMAT, transcribeOutFormat )\n\t\tDDX_CONTROL_HANDLE( IDC_USE_INPUT_FOLDER, useInputFolder )\n\t\tDDX_CONTROL_HANDLE( IDC_PATH_RESULT, transcribeOutputPath )\n\t\tDDX_CONTROL_HANDLE( IDC_BROWSE_RESULT, transcribeOutputBrowse );\n\t\tDDX_CONTROL_HANDLE( IDC_TRANSCRIBE, transcribeButton );\n\t\tDDX_CONTROL_HANDLE( IDC_TRANSCRIBE_PROGRESS, progressBar );\n\tEND_DDX_MAP()\n\nprivate:\n\tPendingState pendingState;\n\tvoid setPending( bool nowPending );\n\tvoid transcribeError( LPCTSTR text, HRESULT hr = S_FALSE );\n\n\tLRESULT OnInitDialog( UINT nMessage, WPARAM wParam, LPARAM lParam, BOOL& bHandled );\n\n\tvoid onClose()\n\t{\n\t\tATLVERIFY( EndDialog( IDCANCEL ) );\n\t}\n\tvoid onBack()\n\t{\n\t\tATLVERIFY( EndDialog( IDC_BACK ) );\n\t}\n\n\tvoid printModelDescription();\n\tCStatic modelDesc;\n\tConsoleCheckbox cbConsole;\n\n\tLanguageDropdown languageSelector;\n\tTranslateCheckbox cbTranslate;\n\n\tCEdit sourceMediaPath;\n\tCButton useInputFolder;\n\tCEdit transcribeOutputPath;\n\tCButton transcribeOutputBrowse;\n\tCComboBox transcribeOutFormat;\n\tCButton transcribeButton;\n\tCProgressBarCtrl progressBar;\n\tvoid populateOutputFormats();\n\n\tLRESULT onOutFormatChange( UINT, INT, HWND, BOOL& bHandled );\n\tLRESULT onInputChange( UINT, INT, HWND, BOOL& );\n\tvoid onInputFolderCheck();\n\tvoid onBrowseMedia();\n\tvoid onBrowseOutput();\n\t// Despite the name, the method also handles the \"Stop\" button\n\tvoid onTranscribe();\n\tvoid onCapture()\n\t{\n\t\tEndDialog( IDC_CAPTURE );\n\t}\n\n\tThreadPoolWork work;\n\n\tenum struct eOutputFormat : uint8_t;\n\tenum struct eVisualState : uint8_t;\n\n\tstruct TranscribeArgs\n\t{\n\t\tCString pathMedia;\n\t\tCString pathOutput;\n\t\tuint32_t language;\n\t\tbool translate;\n\t\teOutputFormat format;\n\t\tWhisper::eResultFlags resultFlags;\n\t\tvolatile eVisualState visualState = (eVisualState)0;\n\t\tuint64_t startTime;\n\t\tint64_t mediaDuration;\n\t\tCString errorMessage;\n\t};\n\tTranscribeArgs transcribeArgs;\n\n\tvoid __stdcall poolCallback() noexcept override final;\n\n\tLRESULT onCallbackStatus( UINT, WPARAM wParam, LPARAM, BOOL& bHandled );\n\n\tHRESULT transcribe();\n\tvoid getThreadError();\n\t\n\tstatic HRESULT writeTextFile( const Whisper::sSegment* const segments, const size_t length, CAtlFile& file, bool timestamps );\n\tstatic HRESULT writeSubRip( const Whisper::sSegment* const segments, const size_t length, CAtlFile& file );\n\tstatic HRESULT writeWebVTT( const Whisper::sSegment* const segments, const size_t length, CAtlFile& file );\n\n\tstatic HRESULT __cdecl newSegmentCallbackStatic( Whisper::iContext* ctx, uint32_t n_new, void* user_data ) noexcept;\n\tstatic HRESULT __cdecl encoderBeginCallback( Whisper::iContext* ctx, void* user_data ) noexcept;\n\tHRESULT newSegmentCallback( Whisper::iContext* ctx, uint32_t n_new );\n\n\tstatic HRESULT __cdecl progressCallbackStatic( double p, Whisper::iContext* ctx, void* pv ) noexcept;\n\tHRESULT progressCallback( double p ) noexcept;\n\n\tvoid onWmClose();\n\t// Populate output path based on the provided input media path\n\tvoid setOutputPath( const CString& input );\n\t// Populate output path based on the input media path in the edit box\n\tvoid setOutputPath();\n};"
},
{
"path": "Examples/WhisperDesktop/Utils/DebugConsole.cpp",
"content": "// https://github.com/Const-me/vis_avs_dx/blob/master/avs_dx/DxVisuals/Interop/ConsoleLogger.cpp\n#include \"stdafx.h\"\n#include \"DebugConsole.h\"\n#include \"miscUtils.h\"\n#include \"../AppState.h\"\n#include \"logger.h\"\n#include \n\nnamespace\n{\n\tusing Whisper::eLogLevel;\n\n\tconstexpr uint16_t defaultAttributes = FOREGROUND_RED | FOREGROUND_GREEN | FOREGROUND_BLUE;\n\n\tinline uint16_t textAttributes( eLogLevel lvl )\n\t{\n\t\tswitch( lvl )\n\t\t{\n\t\tcase eLogLevel::Error:\n\t\t\treturn FOREGROUND_RED | FOREGROUND_INTENSITY;\n\t\tcase eLogLevel::Warning:\n\t\t\treturn FOREGROUND_RED | FOREGROUND_GREEN | FOREGROUND_INTENSITY;\n\t\tcase eLogLevel::Info:\n\t\t\treturn FOREGROUND_GREEN | FOREGROUND_INTENSITY;\n\t\tcase eLogLevel::Debug:\n\t\t\treturn FOREGROUND_BLUE | FOREGROUND_INTENSITY;\n\t\t}\n\t\treturn defaultAttributes;\n\t}\n\n\t// Background stuff: accumulate messages in a small buffer, in case user will want to see them in the console.\n\t// Ideally, we should accumulate them in a more efficient data structure, maybe a circular buffer.\n\t// However, we don't have that many messages per second, this simple solution that uses std::deque is probably good enough for the job.\n\tstatic constexpr uint16_t bufferSize = 64;\n\n\tusing Lock = CComCritSecLock;\n#define LOCK() Lock __lock{ critSec }\n\n\tthread_local std::string threadError;\n}\n\nHRESULT DebugConsole::Entry::print( HANDLE hConsole, CString& tempString ) const\n{\n\tif( !SetConsoleTextAttribute( hConsole, textAttributes( level ) ) )\n\t\treturn getLastHr();\n\n\tmakeUtf16( tempString, message );\n\ttempString += L\"\\r\\n\";\n\tif( !WriteConsoleW( hConsole, tempString, (DWORD)tempString.GetLength(), nullptr, nullptr ) )\n\t\treturn getLastHr();\n\treturn S_OK;\n}\n\nvoid clearLastError()\n{\n\tthreadError.clear();\n}\n\nbool getLastError( CString& rdi )\n{\n\tif( threadError.empty() )\n\t{\n\t\trdi = L\"\";\n\t\treturn false;\n\t}\n\telse\n\t{\n\t\tmakeUtf16( rdi, threadError.c_str() );\n\t\tthreadError.clear();\n\t\treturn true;\n\t}\n}\n\ninline void DebugConsole::logSink( eLogLevel lvl, const char* message )\n{\n\tLOCK();\n\n\t// Add to the buffer\n\twhile( buffer.size() >= bufferSize )\n\t\tbuffer.pop_front();\n\tbuffer.emplace_back( Entry{ lvl, message } );\n\n\t// If the console window is shown, print there, too.\n\tif( output )\n\t\tbuffer.rbegin()->print( output, tempString );\n}\n\nvoid __stdcall DebugConsole::logSinkStatic( void* context, eLogLevel lvl, const char* message )\n{\n\tif( lvl == eLogLevel::Error )\n\t\tthreadError = message;\n\n\tDebugConsole* con = (DebugConsole*)context;\n\tcon->logSink( lvl, message );\n}\n\nHRESULT DebugConsole::initialize( Whisper::eLogLevel level )\n{\n\tif( nullptr != pGlobalInstance )\n\t\treturn HRESULT_FROM_WIN32( ERROR_ALREADY_INITIALIZED );\n\tpGlobalInstance = this;\n\n\tWhisper::sLoggerSetup setup;\n\tsetup.sink = &logSinkStatic;\n\tsetup.context = this;\n\tsetup.level = level;\n\tsetup.flags = Whisper::eLoggerFlags::SkipFormatMessage;\n\treturn Whisper::setupLogger( setup );\n}\n\nDebugConsole::~DebugConsole()\n{\n\thide();\n\n\tWhisper::sLoggerSetup setup;\n\tmemset( &setup, 0, sizeof( setup ) );\n\tWhisper::setupLogger( setup );\n\n\tpGlobalInstance = nullptr;\n}\n\nDebugConsole* DebugConsole::pGlobalInstance = nullptr;\n\nvoid DebugConsole::windowClosed()\n{\n\tLOCK();\n\tif( FreeConsole() )\n\t{\n\t\t// Apparently, FreeConsole already closes that handle: https://stackoverflow.com/q/12676312/126995\n\t\toutput.Detach();\n\t}\n\toutput.Close();\n\n\tfor( CButton* b : checkboxes )\n\t{\n\t\tif( !*b )\n\t\t\tcontinue;\n\t\tif( !b->IsWindow() )\n\t\t\tcontinue;\n\t\tPostMessage( *b, BM_SETCHECK, BST_UNCHECKED, 0 );\n\t}\n}\n\nBOOL __stdcall DebugConsole::consoleHandlerRoutine( DWORD dwCtrlType )\n{\n\tswitch( dwCtrlType )\n\t{\n\tcase CTRL_CLOSE_EVENT:\n\tcase CTRL_C_EVENT:\n\tcase CTRL_BREAK_EVENT:\n\t\tpGlobalInstance->windowClosed();\n\t\treturn TRUE;\n\t}\n\treturn TRUE;\n}\n\nHRESULT DebugConsole::show()\n{\n\tHWND hWnd = GetConsoleWindow();\n\tif( nullptr != hWnd )\n\t{\n\t\tShowWindow( hWnd, SW_RESTORE );\n\t\tSetForegroundWindow( hWnd );\n\t\treturn S_FALSE;\n\t}\n\n\tif( !AllocConsole() )\n\t\treturn getLastHr();\n\n\toutput.Close();\n\toutput.Attach( GetStdHandle( STD_OUTPUT_HANDLE ) );\n\tif( !output )\n\t\treturn getLastHr();\n\n\tconstexpr UINT cp = CP_UTF8;\n\tif( IsValidCodePage( cp ) )\n\t\tSetConsoleOutputCP( cp );\n\n\t// Enable ANSI color coding\n\tDWORD mode = 0;\n\tif( !GetConsoleMode( output, &mode ) )\n\t\treturn getLastHr();\n\tif( 0 == ( mode & ENABLE_VIRTUAL_TERMINAL_PROCESSING ) )\n\t{\n\t\tmode |= ENABLE_VIRTUAL_TERMINAL_PROCESSING;\n\t\tif( !SetConsoleMode( output, mode ) )\n\t\t\treturn getLastHr();\n\t}\n\n\tSetConsoleTitle( L\"Whisper Desktop Debug Console\" );\n\n\tSetConsoleCtrlHandler( &consoleHandlerRoutine, TRUE );\n\n\t// Disable close command in the sys.menu of the new console, otherwise the whole process will quit: https://stackoverflow.com/a/12015131/126995\n\tHWND hwnd = ::GetConsoleWindow();\n\tif( hwnd != nullptr )\n\t{\n\t\tHMENU hMenu = ::GetSystemMenu( hwnd, FALSE );\n\t\tif( hMenu != NULL )\n\t\t\tDeleteMenu( hMenu, SC_CLOSE, MF_BYCOMMAND );\n\t}\n\n\t// Print old log entries\n\tfor( const auto& e : buffer )\n\t\tCHECK( e.print( output, tempString ) );\n\n\tconst CStringA msg = \"Press Control+C or Control+Break to close this window\\r\\n\";\n\tif( !SetConsoleTextAttribute( output, FOREGROUND_RED | FOREGROUND_GREEN | FOREGROUND_BLUE | FOREGROUND_INTENSITY ) )\n\t\treturn getLastHr();\n\tif( !WriteConsoleA( output, cstr( msg ), msg.GetLength(), nullptr, nullptr ) )\n\t\treturn getLastHr();\n\n\treturn S_OK;\n}\n\nHRESULT DebugConsole::hide()\n{\n\tLOCK();\n\tif( !output )\n\t\treturn S_FALSE;\n\twindowClosed();\n\treturn S_OK;\n}\n\nvoid DebugConsole::addCheckbox( CButton& cb )\n{\n\tcheckboxes.emplace( &cb );\n}\nvoid DebugConsole::removeCheckbox( CButton& cb )\n{\n\tcheckboxes.erase( &cb );\n}\n\nHRESULT ConsoleCheckbox::initialize( HWND dialog, int idc, AppState& state )\n{\n\tcontrol = GetDlgItem( dialog, idc );\n\tassert( control );\n\n\tconsole = &state.console;\n\tif( state.console.isVisible() )\n\t\tcontrol.SetCheck( BST_CHECKED );\n\n\tstate.console.addCheckbox( control );\n\treturn S_OK;\n}\n\nvoid ConsoleCheckbox::click()\n{\n\tconst int state = control.GetCheck();\n\tif( state == BST_CHECKED )\n\t\tconsole->show();\n\telse\n\t\tconsole->hide();\n}\n\nvoid ConsoleCheckbox::ensureChecked()\n{\n\tconst int state = control.GetCheck();\n\tif( state == BST_CHECKED )\n\t\treturn;\n\tcontrol.SetCheck( BST_CHECKED );\n\tconsole->show();\n}\n\nvoid DebugConsole::log( eLogLevel lvl, const char* pszFormat, va_list args )\n{\n\tLOCK();\n\t// Add to the buffer\n\twhile( buffer.size() >= bufferSize )\n\t\tbuffer.pop_front();\n\n\ttempStringA.FormatV( pszFormat, args );\n\tbuffer.emplace_back( Entry{ lvl, tempStringA } );\n\n\t// If the console window is shown, print there, too.\n\tif( output )\n\t\tbuffer.rbegin()->print( output, tempString );\n}\n\nvoid DebugConsole::logMessage( eLogLevel lvl, const char* pszFormat, va_list args )\n{\n\tif( nullptr == pGlobalInstance )\n\t\treturn;\n\tpGlobalInstance->log( lvl, pszFormat, args );\n}\n\nvoid logMessage( Whisper::eLogLevel lvl, const char8_t* pczFormat, va_list args )\n{\n\tDebugConsole::logMessage( lvl, (const char*)pczFormat, args );\n}"
},
{
"path": "Examples/WhisperDesktop/Utils/DebugConsole.h",
"content": "#pragma once\n#include \n#include \n#include \n\nclass AppState;\nclass DebugConsole\n{\n\tusing eLogLevel = Whisper::eLogLevel;\n\n\tstruct Entry\n\t{\n\t\teLogLevel level;\n\t\tCStringA message;\n\t\tHRESULT print( HANDLE hConsole, CString& tempString ) const;\n\t};\n\n\tCComAutoCriticalSection critSec;\n\tstd::deque buffer;\n\tCString tempString;\n\tCHandle output;\n\n\tinline void logSink( eLogLevel lvl, const char* message );\n\tstatic void __stdcall logSinkStatic( void* context, eLogLevel lvl, const char* message );\n\n\tstatic BOOL __stdcall consoleHandlerRoutine( DWORD dwCtrlType );\n\n\tstatic DebugConsole* pGlobalInstance;\n\tvoid windowClosed();\n\n\tstd::unordered_set checkboxes;\n\n\tCStringA tempStringA;\n\tvoid log( eLogLevel lvl, const char* pszFormat, va_list args );\n\npublic:\n\tHRESULT initialize( eLogLevel level = eLogLevel::Debug );\n\t~DebugConsole();\n\n\tHRESULT show();\n\tHRESULT hide();\n\tbool isVisible() const { return output; }\n\n\tvoid addCheckbox( CButton& cb );\n\tvoid removeCheckbox( CButton& cb );\n\n\tstatic void logMessage( eLogLevel lvl, const char* pszFormat, va_list args );\n};\n\nclass ConsoleCheckbox\n{\n\tCButton control;\n\tDebugConsole* console = nullptr;\n\npublic:\n\tHRESULT initialize( HWND dialog, int idc, AppState& state );\n\tvoid click();\n\t~ConsoleCheckbox()\n\t{\n\t\tif( nullptr != console )\n\t\t\tconsole->removeCheckbox( control );\n\t}\n\tvoid ensureChecked();\n};"
},
{
"path": "Examples/WhisperDesktop/Utils/LanguageDropdown.cpp",
"content": "#include \"stdafx.h\"\n#include \"LanguageDropdown.h\"\n#include \"miscUtils.h\"\n\nnamespace\n{\n\tinline wchar_t toUpper( wchar_t c )\n\t{\n\t\tsize_t st = (uint16_t)c;\n\t\tst = reinterpret_cast( CharUpperW( reinterpret_cast( st ) ) );\n\t\treturn (wchar_t)(uint16_t)st;\n\t}\n\n\tvoid makeTitleCase( CString& s )\n\t{\n\t\tbool cap = true;\n\t\tfor( int i = 0; i < s.GetLength(); i++ )\n\t\t{\n\t\t\twchar_t c = s[ i ];\n\t\t\tif( cap )\n\t\t\t{\n\t\t\t\tc = toUpper( c );\n\t\t\t\ts.SetAt( i, c );\n\t\t\t}\n\t\t\tcap = false;\n\t\t\tif( c == ' ' )\n\t\t\t\tcap = true;\n\t\t}\n\t}\n}\n\nint LanguageDropdown::getInitialSelection( AppState& state ) const\n{\n\tconstexpr uint32_t english = 0x6E65;\n\n\t// Load preference from the registry\n\tuint32_t id = state.languageRead();\n\tif( id == UINT_MAX )\n\t\tid = english;\n\n\tauto it = std::find( keys.begin(), keys.end(), id );\n\tif( it == keys.end() )\n\t{\n\t\tid = english;\n\t\tit = std::find( keys.begin(), keys.end(), id );\n\t\tassert( it != keys.end() );\n\t}\n\n\tptrdiff_t idx = it - keys.begin();\n\treturn (int)idx;\n}\n\nvoid LanguageDropdown::initialize( HWND owner, int idc, AppState& state )\n{\n\tm_hWnd = GetDlgItem( owner, idc );\n\tassert( nullptr != m_hWnd );\n\n\tWhisper::sLanguageList list;\n\tWhisper::getSupportedLanguages( list );\n\n\tconst size_t length = list.length;\n\tkeys.resize( length );\n\tCString utf16;\n\tfor( size_t i = 0; i < length; i++ )\n\t{\n\t\tkeys[ i ] = list.pointer[ i ].key;\n\t\tmakeUtf16( utf16, list.pointer[ i ].name );\n\t\tmakeTitleCase( utf16 );\n\t\tSendMessage( m_hWnd, CB_ADDSTRING, 0, (LPARAM)cstr( utf16 ) );\n\t}\n\n\tconst int curSel = getInitialSelection( state );\n\tSendMessage( m_hWnd, CB_SETCURSEL, curSel, 0 );\n}\n\nuint32_t LanguageDropdown::selectedLanguage()\n{\n\tconst int cs = (int)SendMessage( m_hWnd, CB_GETCURSEL, 0, 0 );\n\tif( cs < 0 || cs >= keys.size() )\n\t\treturn UINT_MAX;\n\treturn keys[ cs ];\n}\n\nvoid LanguageDropdown::saveSelection( AppState& state )\n{\n\tstate.languageWrite( selectedLanguage() );\n}"
},
{
"path": "Examples/WhisperDesktop/Utils/LanguageDropdown.h",
"content": "#pragma once\n#include \"../AppState.h\"\n\n// Dropdown list which implements language selector control\nclass LanguageDropdown\n{\n\tHWND m_hWnd = nullptr;\n\tstd::vector keys;\n\tint getInitialSelection( AppState& state ) const;\n\npublic:\n\toperator HWND() const\n\t{\n\t\treturn m_hWnd;\n\t}\n\n\t// Query language list form the native library, populate the combo box\n\t// Then load the last saved language selection from registry, and preselect an item.\n\tvoid initialize( HWND owner, int idc, AppState& state );\n\n\t// Get the ID of the currently selected language, or UINT_MAX if nothing's selected\n\tuint32_t selectedLanguage();\n\n\t// Get the ID of the currently selected language, and store in registry\n\tvoid saveSelection( AppState& state );\n};"
},
{
"path": "Examples/WhisperDesktop/Utils/PendingState.cpp",
"content": "#include \"stdafx.h\"\n#include \"PendingState.h\"\n\nvoid PendingState::initialize( std::initializer_list editors, std::initializer_list pending )\n{\n\teditorsWindows = editors;\n\twasEnabled.resize( editorsWindows.size() );\n\tpendingWindows = pending;\n}\n\nvoid PendingState::setPending( bool nowPending )\n{\n\tif( nowPending )\n\t{\n\t\tfor( size_t i = 0; i < editorsWindows.size(); i++ )\n\t\t{\n\t\t\tBOOL e = IsWindowEnabled( editorsWindows[ i ] );\n\t\t\tif( e )\n\t\t\t{\n\t\t\t\twasEnabled[ i ] = true;\n\t\t\t\tEnableWindow( editorsWindows[ i ], FALSE );\n\t\t\t}\n\t\t\telse\n\t\t\t\twasEnabled[ i ] = false;\n\t\t}\n\t}\n\telse\n\t{\n\t\tfor( size_t i = 0; i < editorsWindows.size(); i++ )\n\t\t{\n\t\t\tif( !wasEnabled[ i ] )\n\t\t\t\tcontinue;\n\t\t\tEnableWindow( editorsWindows[ i ], TRUE );\n\t\t}\n\t}\n\n\tconst int show = nowPending ? SW_NORMAL : SW_HIDE;\n\tfor( HWND w : pendingWindows )\n\t\t::ShowWindow( w, show );\n}"
},
{
"path": "Examples/WhisperDesktop/Utils/PendingState.h",
"content": "#pragma once\n\n// Utility class to switch visual state of dialog controls between idle and pending\nclass PendingState\n{\n\tstd::vector editorsWindows;\n\tstd::vector wasEnabled;\n\tstd::vector pendingWindows;\npublic:\n\tvoid initialize( std::initializer_list