Created
May 16, 2018 06:45
-
-
Save daramkun/7e87e9b789a160df7705d9b6b5efe4e8 to your computer and use it in GitHub Desktop.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| #include <Windows.h> | |
| #include <memory> | |
| #include <mmintrin.h> | |
| #include <immintrin.h> | |
| #include <xmmintrin.h> | |
| #include <functional> | |
| #include <cstdio> | |
| class HeapMemory | |
| { | |
| public: | |
| HeapMemory ( size_t dataSize ) | |
| : _dataSize ( dataSize ) | |
| { | |
| _data = new BYTE [ dataSize ]; | |
| } | |
| HeapMemory ( size_t stride, size_t height ) | |
| : HeapMemory ( stride * height ) | |
| { } | |
| ~HeapMemory () | |
| { | |
| if ( _data ) | |
| delete [] _data; | |
| } | |
| public: | |
| inline BYTE * getData () const noexcept { return _data; } | |
| inline size_t getDataSize () const noexcept { return _dataSize; } | |
| private: | |
| BYTE * _data; | |
| size_t _dataSize; | |
| }; | |
| inline double getTime () noexcept | |
| { | |
| LARGE_INTEGER performanceFrequency, getTime; | |
| QueryPerformanceFrequency ( &performanceFrequency ); | |
| QueryPerformanceCounter ( &getTime ); | |
| return ( getTime.QuadPart / ( double ) performanceFrequency.QuadPart ); | |
| } | |
| using MemoryCopy = std::function<void* ( void *, void *, size_t )>; | |
| #define SystemMemoryCopy memcpy | |
| inline void * SSEMemoryCopy ( void * dst, const void * src, size_t size ) | |
| { | |
| const unsigned int step = 64; | |
| BYTE * bdst = reinterpret_cast< BYTE* >( dst ); | |
| const BYTE * bsrc = reinterpret_cast< const BYTE* >( src ); | |
| while ( size > step ) | |
| { | |
| //_mm_prefetch ( ( ( CHAR* ) src ) + 640, _MM_HINT_NTA ); | |
| __m128i A = _mm_load_si128 ( ( __m128i* )( bsrc + 0 ) ); | |
| __m128i B = _mm_load_si128 ( ( __m128i* )( bsrc + 16 ) ); | |
| __m128i C = _mm_load_si128 ( ( __m128i* )( bsrc + 32 ) ); | |
| __m128i D = _mm_load_si128 ( ( __m128i* )( bsrc + 48 ) ); | |
| // destination address must be 16-byte aligned! | |
| _mm_store_si128 ( ( __m128i* )( bdst + 0 ), A ); | |
| _mm_store_si128 ( ( __m128i* )( bdst + 16 ), B ); | |
| _mm_store_si128 ( ( __m128i* )( bdst + 32 ), C ); | |
| _mm_store_si128 ( ( __m128i* )( bdst + 48 ), D ); | |
| bsrc += step; | |
| bdst += step; | |
| size -= step; | |
| } | |
| memcpy ( bdst, bsrc, size ); | |
| _mm_mfence (); | |
| return dst; | |
| } | |
| inline void * AVX2MemoryCopy ( void * dst, const void * src, size_t size ) | |
| { | |
| //const unsigned long step = 64; | |
| BYTE * bdst = reinterpret_cast< BYTE* >( dst ); | |
| const BYTE * bsrc = reinterpret_cast< const BYTE* >( src ); | |
| while ( size > 64 ) | |
| { | |
| //_mm_prefetch ( ( ( CHAR* ) src ) + 640, _MM_HINT_NTA ); | |
| __m256i A = _mm256_load_si256 ( ( __m256i* )( bsrc + 0 ) ); | |
| __m256i B = _mm256_load_si256 ( ( __m256i* )( bsrc + 32 ) ); | |
| // destination address must be 32-byte aligned! | |
| _mm256_store_si256 ( ( __m256i* )( bdst + 0 ), A ); | |
| _mm256_store_si256 ( ( __m256i* )( bdst + 32 ), B ); | |
| bsrc += 64; | |
| bdst += 64; | |
| size -= 64; | |
| } | |
| memcpy ( bdst, bsrc, size ); | |
| _mm_mfence (); | |
| return dst; | |
| } | |
| inline void Test ( MemoryCopy memcpy, const char * funcName, int width, int height, int fps, int count ) noexcept | |
| { | |
| printf ( "==== %s Test (%dx%d@%dfpsx%d) ====\n", funcName, width, height, fps, count ); | |
| double startTime = getTime (); | |
| HeapMemory src ( width, height ); | |
| for ( int c = 0; c < count; ++c ) | |
| { | |
| for ( int i = 0; i < fps; ++i ) | |
| { | |
| HeapMemory dst ( width, height ); | |
| memcpy ( dst.getData (), src.getData (), dst.getDataSize () ); | |
| } | |
| } | |
| double endTime = getTime (); | |
| printf ( "Elapsed time> %lf\n", endTime - startTime ); | |
| printf ( "Average time> %lf\n", ( endTime - startTime ) / ( fps * count ) ); | |
| } | |
| int main ( void ) | |
| { | |
| SetPriorityClass ( GetCurrentProcess (), HIGH_PRIORITY_CLASS ); | |
| SetProcessAffinityMask ( GetCurrentProcess (), 0 ); | |
| SetThreadPriority ( GetCurrentThread (), THREAD_PRIORITY_HIGHEST ); | |
| Test ( SystemMemoryCopy, "memcpy ()", 1920, 1080, 60, 10 ); | |
| Test ( SSEMemoryCopy, "SSEmemcpy ()", 1920, 1080, 60, 10 ); | |
| Test ( AVX2MemoryCopy, "AVX2memcpy ()", 1920, 1080, 60, 10 ); | |
| return 0; | |
| } |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
memcpy is fastest.
Test from AMD Ryzen 5 2600X + DDR4 ESSENCORE Klevv 19200 16GBx2