opelly/Stockfish
1
0
Fork 0
mirror of https://github.com/opelly27/Stockfish.git synced 2026-05-20 15:37:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

Prefetch weights for feature transformer backprop to shared cache.

Browse Source
This commit is contained in:
Tomasz Sobczyk
2020-11-25 22:43:42 +01:00
committed by nodchip
parent 8009973381
commit e954b14196
1 changed files with 7 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/nnue/trainer/trainer_feature_transformer.h
+7 -3
View File
@@ -138,6 +138,7 @@ namespace Eval::NNUE {
for (IndexType b = offset; b < offset + count; ++b) for (IndexType b = offset; b < offset + count; ++b)
{ {
const IndexType batch_offset = kOutputDimensions * b; const IndexType batch_offset = kOutputDimensions * b;
for (IndexType c = 0; c < 2; ++c) { for (IndexType c = 0; c < 2; ++c) {
const IndexType output_offset = batch_offset + kHalfDimensions * c; const IndexType output_offset = batch_offset + kHalfDimensions * c;
@@ -459,10 +460,16 @@ namespace Eval::NNUE {
for (IndexType b = 0; b < batch_->size(); ++b) { for (IndexType b = 0; b < batch_->size(); ++b) {
const IndexType batch_offset = kOutputDimensions * b; const IndexType batch_offset = kOutputDimensions * b;
for (IndexType c = 0; c < 2; ++c) { for (IndexType c = 0; c < 2; ++c) {
const IndexType output_offset = batch_offset + kHalfDimensions * c; const IndexType output_offset = batch_offset + kHalfDimensions * c;
for (const auto& feature : (*batch_)[b].training_features[c]) { for (const auto& feature : (*batch_)[b].training_features[c]) {
const IndexType feature_index = feature.get_index(); const IndexType feature_index = feature.get_index();
const IndexType weights_offset =
kHalfDimensions * feature_index;
#if defined (USE_SSE2)
_mm_prefetch(reinterpret_cast<const char*>(&weights_[weights_offset]), _MM_HINT_T2);
#endif
// We assign each bucket a continuous range of bits at least // We assign each bucket a continuous range of bits at least
// of cache line size to prevent false sharing. // of cache line size to prevent false sharing.
@@ -479,9 +486,6 @@ namespace Eval::NNUE {
// (even a different cache line) // (even a different cache line)
observed_features.set(feature_index); observed_features.set(feature_index);
const IndexType weights_offset =
kHalfDimensions * feature_index;
const auto scale = static_cast<LearnFloatType>( const auto scale = static_cast<LearnFloatType>(
effective_learning_rate / feature.get_count()); effective_learning_rate / feature.get_count());