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

added perfomance to readme and modified instructions

Browse Source
This commit is contained in:
Shin
2016-02-18 20:34:57 +01:00
parent 1a55a720eb
commit c45fd63591
2 changed files with 12 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
README.md
View File

@@ -14,10 +14,11 @@ Perfomace
i5-5300U & 8GB ram i5-5300U & 8GB ram
| date | feedforward_perf | recurrent_perf | backpropagation_perf | | date | feedforward_perf | recurrent_perf | backpropagation_perf |
-------------------- | ---------------- | -------------- | -------------------- | ------------------------ | ---------------- | -------------- | -------------------- |
| FANN | 12.6 | | | | FANN | 12.6 | | |
-------------------- | ---------------- | -------------- | -------------------- | ------------------------ | ---------------- | -------------- | -------------------- |
| 2016/02/07 initial | 8.27 sec | 7.15 sec | 6.00 sec | | 2016/02/07 initial | 8.27 sec | 7.15 sec | 6.00 sec |
| 2016/02/17 AVX | 5.53 sec | 4.68 sec | 4.63 sec | | 2016/02/17 AVX | 5.53 sec | 4.68 sec | 4.63 sec |
| 2016/02/17 weights | 5.53 sec | 4.68 sec | 3.02 sec | | 2016/02/17 weights | 5.53 sec | 4.68 sec | 3.02 sec |
| 2016/02/18 neuron ref. | 5.53 sec | 4.68 sec | 1.02 sec |

7
src/NeuralNetwork/BasisFunction/Linear.cpp
View File

@@ -19,12 +19,12 @@ float NeuralNetwork::BasisFunction::Linear::operator()(const std::vector<float>
partialSolution.avx=_mm256_setzero_ps(); partialSolution.avx=_mm256_setzero_ps();
for(size_t k=0;k<alignedPrev;k+=8) { for(size_t k=0;k<alignedPrev;k+=8) {
//TODO: asignement!! -- possible speedup //TODO: assignement!! -- possible speedup
partialSolution.avx=_mm256_add_ps(partialSolution.avx,_mm256_mul_ps(_mm256_loadu_ps(weightsData+k),_mm256_loadu_ps(inputData+k))); partialSolution.avx=_mm256_fmadd_ps(_mm256_loadu_ps(weightsData+k),_mm256_loadu_ps(inputData+k),partialSolution.avx);
} }
for(size_t k=alignedPrev;k<inputSize;k++) { for(size_t k=alignedPrev;k<inputSize;k++) {
partialSolution.avx=_mm256_add_ps(partialSolution.avx,_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_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_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));
@@ -32,6 +32,7 @@ float NeuralNetwork::BasisFunction::Linear::operator()(const std::vector<float>
partialSolution.avx = _mm256_hadd_ps(partialSolution.avx, partialSolution.avx); partialSolution.avx = _mm256_hadd_ps(partialSolution.avx, partialSolution.avx);
return partialSolution.f[0]; return partialSolution.f[0];
#elif USE_SSE #elif USE_SSE
std::size_t alignedPrev=inputSize-inputSize%4; std::size_t alignedPrev=inputSize-inputSize%4;