HCha/Stockfish
1
0
Fork 0
mirror of https://github.com/HChaZZY/Stockfish.git synced 2025-12-24 19:16:49 +08:00
Code Issues Packages Projects Releases Wiki Activity

Parallelize input slice trainer backprop.

Browse Source
This commit is contained in:
Tomasz Sobczyk
2020-10-27 18:41:17 +01:00
committed by nodchip
parent 941897ff2c
commit b5714c4084
1 changed files with 19 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
src/nnue/trainer/trainer_input_slice.h
View File

@@ -236,17 +236,29 @@ namespace Eval::NNUE {
const LearnFloatType* gradients,
LearnFloatType learning_rate) {
for (IndexType b = 0; b < batch_size_; ++b) {
const IndexType input_offset = kInputDimensions * b;
const IndexType output_offset = kOutputDimensions * b;
for (IndexType i = 0; i < kInputDimensions; ++i) {
if ((int)i < (int)Offset || i >= Offset + kOutputDimensions) {
thread_pool.for_each_index_with_workers(
0, batch_size_,
[&](Thread&, int b) {
const IndexType input_offset = kInputDimensions * b;
const IndexType output_offset = kOutputDimensions * b;
IndexType i = 0;
for (; i < Offset; ++i) {
gradients_[input_offset + i] = static_cast<LearnFloatType>(0.0);
} else {
}
for (; i < Offset + kOutputDimensions; ++i) {
gradients_[input_offset + i] = gradients[output_offset + i - Offset];
}
for (; i < kInputDimensions; ++i)
{
gradients_[input_offset + i] = static_cast<LearnFloatType>(0.0);
}
}
}
);
thread_pool.wait_for_workers_finished();
shared_input_trainer_->backpropagate(thread_pool, gradients_.data(), learning_rate);
}