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

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();
#ifndef USE_FMA
__m256 tmp;
#endif
for(size_t k=0;k<alignedPrev;k+=8) {
//TODO: assignement!! -- possible speedup
#ifdef USE_FMA
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++) {
#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);
#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));