opelly/Stockfish
1
0
Fork 0
mirror of https://github.com/opelly27/Stockfish.git synced 2026-05-20 06:17:49 +00:00
Code Issues Packages Projects Releases Wiki Activity

Store references instead of copying the results of intermediate autograd computations.

Browse Source
This commit is contained in:
Tomasz Sobczyk
2020-11-30 21:43:24 +01:00
committed by nodchip
parent 6cd0b03098
commit 4eb0e77a2a
1 changed files with 9 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/learn/learn.cpp
+9 -9
View File
@@ -199,7 +199,7 @@ namespace Learner
// differentiation of the loss function. While it works it has it's caveats. // differentiation of the loss function. While it works it has it's caveats.
// To work fast enough it requires memoization and reference semantics. // To work fast enough it requires memoization and reference semantics.
// Memoization is mostly opaque to the user and is only per eval basis. // Memoization is mostly opaque to the user and is only per eval basis.
// As for reference semantics, we cannot copy every node, // As for reference semantics, we cannot copy every node,
// because we need a way to reuse computation. // because we need a way to reuse computation.
// But we can't really use shared_ptr because of the overhead. That means // But we can't really use shared_ptr because of the overhead. That means
// that we have to ensure all parts of a loss expression are not destroyed // that we have to ensure all parts of a loss expression are not destroyed
@@ -321,7 +321,7 @@ namespace Learner
// The model captures only up to 240 plies, so limit input (and rescale) // The model captures only up to 240 plies, so limit input (and rescale)
static thread_local auto m_ = std::forward<PlyT>(ply_) / 64.0; static thread_local auto m_ = std::forward<PlyT>(ply_) / 64.0;
static thread_local auto a_ = (((as[0] * m_ + as[1]) * m_ + as[2]) * m_) + as[3]; static thread_local auto a_ = (((as[0] * m_ + as[1]) * m_ + as[2]) * m_) + as[3];
static thread_local auto b_ = (((bs[0] * m_ + bs[1]) * m_ + bs[2]) * m_) + bs[3]; static thread_local auto b_ = (((bs[0] * m_ + bs[1]) * m_ + bs[2]) * m_) + bs[3];
@@ -392,11 +392,11 @@ namespace Learner
{ {
using namespace Learner::Autograd::UnivariateStatic; using namespace Learner::Autograd::UnivariateStatic;
static thread_local auto q_ = expected_perf_(VariableParameter<double, 0>{}); static thread_local auto& q_ = expected_perf_(VariableParameter<double, 0>{});
static thread_local auto p_ = expected_perf_(scale_score_(ConstantParameter<double, 1>{})); static thread_local auto& p_ = expected_perf_(scale_score_(ConstantParameter<double, 1>{}));
static thread_local auto t_ = (ConstantParameter<double, 2>{} + 1.0) * 0.5; static thread_local auto t_ = (ConstantParameter<double, 2>{} + 1.0) * 0.5;
static thread_local auto lambda_ = ConstantParameter<double, 3>{}; static thread_local auto lambda_ = ConstantParameter<double, 3>{};
static thread_local auto loss_ = cross_entropy_(q_, p_, t_, lambda_); static thread_local auto& loss_ = cross_entropy_(q_, p_, t_, lambda_);
auto args = std::tuple( auto args = std::tuple(
(double)shallow, (double)shallow,
@@ -415,14 +415,14 @@ namespace Learner
static thread_local auto ply_ = ConstantParameter<double, 4>{}; static thread_local auto ply_ = ConstantParameter<double, 4>{};
static thread_local auto shallow_ = VariableParameter<double, 0>{}; static thread_local auto shallow_ = VariableParameter<double, 0>{};
static thread_local auto q_ = expected_perf_use_wdl_(shallow_, ply_); static thread_local auto& q_ = expected_perf_use_wdl_(shallow_, ply_);
// We could do just this but MSVC crashes with an internal compiler error :( // We could do just this but MSVC crashes with an internal compiler error :(
// static thread_local auto scaled_teacher_ = scale_score_(ConstantParameter<double, 1>{}); // static thread_local auto& scaled_teacher_ = scale_score_(ConstantParameter<double, 1>{});
// static thread_local auto p_ = expected_perf_use_wdl_(scaled_teacher_, ply_); // static thread_local auto& p_ = expected_perf_use_wdl_(scaled_teacher_, ply_);
static thread_local auto p_ = ConstantParameter<double, 1>{}; static thread_local auto p_ = ConstantParameter<double, 1>{};
static thread_local auto t_ = (ConstantParameter<double, 2>{} + 1.0) * 0.5; static thread_local auto t_ = (ConstantParameter<double, 2>{} + 1.0) * 0.5;
static thread_local auto lambda_ = ConstantParameter<double, 3>{}; static thread_local auto lambda_ = ConstantParameter<double, 3>{};
static thread_local auto loss_ = cross_entropy_(q_, p_, t_, lambda_); static thread_local auto& loss_ = cross_entropy_(q_, p_, t_, lambda_);
auto args = std::tuple( auto args = std::tuple(
(double)shallow, (double)shallow,