shin/NeuralNetworkLib
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

AVX fix for non-FMA CPU

Browse Source
This commit is contained in:
Shin
2016-02-23 14:08:46 +01:00
parent dc2c3ef0ca
commit d9614b5796
3 changed files with 27 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
CMakeLists.txt
View File

@@ -18,7 +18,6 @@ if(CPU_SSE3_AVAILABLE)
endif(CPU_SSE3_AVAILABLE) endif(CPU_SSE3_AVAILABLE)
OPTION(USE_AVX "If AVX instruction set should be used." OFF) OPTION(USE_AVX "If AVX instruction set should be used." OFF)
OPTION(USE_SSE "If SSE instruction set should be used." OFF) OPTION(USE_SSE "If SSE instruction set should be used." OFF)
OPTION(USE_SSE2 "If SSE 2 instruction set should be used." ON) OPTION(USE_SSE2 "If SSE 2 instruction set should be used." ON)

14
cmake/CPUFeatures
View File

@@ -10,6 +10,7 @@ macro(CHECK_CPU)
MESSAGE("-- -- Found SSE2") MESSAGE("-- -- Found SSE2")
set(CPU_SSE2_AVAILABLE "true" INTERNAL BOOL "SSE2 available on host") set(CPU_SSE2_AVAILABLE "true" INTERNAL BOOL "SSE2 available on host")
ELSE () ELSE ()
MESSAGE("-- -- Not found SSE2")
ENDIF (SSE2_TRUE) ENDIF (SSE2_TRUE)
STRING(REGEX REPLACE "^.*(sse3).*$" "\\1" SSE_THERE ${CPUINFO}) STRING(REGEX REPLACE "^.*(sse3).*$" "\\1" SSE_THERE ${CPUINFO})
@@ -18,6 +19,7 @@ macro(CHECK_CPU)
MESSAGE("-- -- Found SSE3") MESSAGE("-- -- Found SSE3")
set(CPU_SSE3_AVAILABLE "true" INTERNAL BOOL "SSE3 available on host") set(CPU_SSE3_AVAILABLE "true" INTERNAL BOOL "SSE3 available on host")
ELSE () ELSE ()
MESSAGE("-- -- Not found SSE3")
ENDIF (SSE3_TRUE) ENDIF (SSE3_TRUE)
STRING(REGEX REPLACE "^.*(sse4_2).*$" "\\1" SSE_THERE ${CPUINFO}) STRING(REGEX REPLACE "^.*(sse4_2).*$" "\\1" SSE_THERE ${CPUINFO})
@@ -26,6 +28,7 @@ macro(CHECK_CPU)
MESSAGE("-- -- Found SSE4.2") MESSAGE("-- -- Found SSE4.2")
set(CPU_SSE4.2_AVAILABLE "true" INTERNAL BOOL "SSE4.2 available on host") set(CPU_SSE4.2_AVAILABLE "true" INTERNAL BOOL "SSE4.2 available on host")
ELSE () ELSE ()
MESSAGE("-- -- Not found SSE4.2")
ENDIF (SSE4.2_TRUE) ENDIF (SSE4.2_TRUE)
STRING(REGEX REPLACE "^.*(avx).*$" "\\1" SSE_THERE ${CPUINFO}) STRING(REGEX REPLACE "^.*(avx).*$" "\\1" SSE_THERE ${CPUINFO})
@@ -34,7 +37,18 @@ macro(CHECK_CPU)
MESSAGE("-- -- Found AVX") MESSAGE("-- -- Found AVX")
set(CPU_AVX_AVAILABLE "TRUE" INTERNAL BOOL "AVX available on host") set(CPU_AVX_AVAILABLE "TRUE" INTERNAL BOOL "AVX available on host")
ELSE () ELSE ()
MESSAGE("-- -- Not found AVX")
ENDIF (AVX_TRUE) ENDIF (AVX_TRUE)
STRING(REGEX REPLACE "^.*(fma).*$" "\\1" SSE_THERE ${CPUINFO})
STRING(COMPARE EQUAL "fma" "${SSE_THERE}" FMA_TRUE)
IF (FMA_TRUE)
MESSAGE("-- -- Found FMA")
set(CPU_FMA_AVAILABLE "TRUE" INTERNAL BOOL "FMA available on host")
ELSE ()
MESSAGE("-- -- Not found FMA")
ENDIF (FMA_TRUE)
else() else()
MESSAGE("Error detecting CPU features") MESSAGE("Error detecting CPU features")
endif(CMAKE_SYSTEM_NAME MATCHES "Linux") endif(CMAKE_SYSTEM_NAME MATCHES "Linux")

13
src/NeuralNetwork/BasisFunction/Linear.cpp
View File

@@ -18,13 +18,26 @@ float NeuralNetwork::BasisFunction::Linear::operator()(const std::vector<float>
partialSolution.avx=_mm256_setzero_ps(); partialSolution.avx=_mm256_setzero_ps();
#ifndef USE_FMA
__m256 tmp;
#endif
for(size_t k=0;k<alignedPrev;k+=8) { for(size_t k=0;k<alignedPrev;k+=8) {
//TODO: assignement!! -- possible speedup //TODO: assignement!! -- possible speedup
#ifdef USE_FMA
partialSolution.avx=_mm256_fmadd_ps(_mm256_loadu_ps(weightsData+k),_mm256_loadu_ps(inputData+k),partialSolution.avx); partialSolution.avx=_mm256_fmadd_ps(_mm256_loadu_ps(weightsData+k),_mm256_loadu_ps(inputData+k),partialSolution.avx);
#else
tmp=_mm256_mul_ps(_mm256_loadu_ps(weightsData+k),_mm256_loadu_ps(inputData+k));
partialSolution.avx=_mm256_add_ps(tmp,partialSolution.avx);
#endif
} }
for(size_t k=alignedPrev;k<inputSize;k++) { for(size_t k=alignedPrev;k<inputSize;k++) {
#ifdef USE_FMA
partialSolution.avx=_mm256_fmadd_ps(_mm256_set_ps(weightsData[k],0,0,0,0,0,0,0),_mm256_set_ps(inputData[k],0,0,0,0,0,0,0),partialSolution.avx); partialSolution.avx=_mm256_fmadd_ps(_mm256_set_ps(weightsData[k],0,0,0,0,0,0,0),_mm256_set_ps(inputData[k],0,0,0,0,0,0,0),partialSolution.avx);
#else
tmp=_mm256_mul_ps(_mm256_set_ps(weightsData[k],0,0,0,0,0,0,0),_mm256_set_ps(inputData[k],0,0,0,0,0,0,0));
partialSolution.avx=_mm256_add_ps(tmp,partialSolution.avx);
#endif
} }
partialSolution.avx = _mm256_add_ps(partialSolution.avx, _mm256_permute2f128_ps(partialSolution.avx , partialSolution.avx , 1)); partialSolution.avx = _mm256_add_ps(partialSolution.avx, _mm256_permute2f128_ps(partialSolution.avx , partialSolution.avx , 1));