jirayu_sri/stockfish
1
0
Fork 0
mirror of https://github.com/official-stockfish/Stockfish.git synced 2026-07-22 12:47:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update nnue-architecture to SFNNv15

Browse Source 
Updating nnue to a new architecture including a newly trained net using recipe https://github.com/vondele/nettest/pull/395 and trainer https://github.com/official-stockfish/nnue-pytorch/pull/480.

The recipe contains many small improvements additionally the newly relabled data. Thus this is a combined effort.

Passed STC
https://tests.stockfishchess.org/tests/view/6a429fd8f97ff95f78795110
```
LLR: 2.94 (-2.94,2.94) <0.00,2.00>
Total: 25312 W: 6725 L: 6422 D: 12165
Ptnml(0-2): 78, 2915, 6397, 3158, 108
```

Passed LTC
https://tests.stockfishchess.org/tests/view/6a434c4cf97ff95f78795235
```
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 32814 W: 8655 L: 8345 D: 15814
Ptnml(0-2): 17, 3469, 9132, 3765, 24
```

closes https://github.com/official-stockfish/Stockfish/pull/6938

Bench: 2639962

Co-authored-by: anematode <anematode@users.noreply.github.com>
Co-authored-by: Joost VandeVondele <vondele@users.noreply.github.com>
Co-authored-by: xu-shawn <xu-shawn@users.noreply.github.com>
This commit is contained in:
tony
2026-07-03 20:07:49 +02:00
committed by Joost VandeVondele
co-authored by anematode Joost VandeVondele xu-shawn
parent 6088838797
commit e33bb26ee7
2 changed files with 28 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/evaluate.h
+1 -1
View File
@@ -33,7 +33,7 @@ namespace Eval {
// for the build process (profile-build and fishtest) to work. Do not change the
// name of the macro or the location where this macro is defined, as it is used
// in the Makefile/Fishtest.
#define EvalFileDefaultName "nn-af1339a6dea3.nnue"
#define EvalFileDefaultName "nn-2ab42c9eeb3e.nnue"
namespace NNUE {
class Network;
src/nnue/nnue_architecture.h
+27 -20
View File
@@ -42,7 +42,7 @@ using PSQFeatureSet = Features::HalfKAv2_hm;
// Number of input feature dimensions after conversion
constexpr IndexType L1 = 1024;
constexpr int L2 = 31;
constexpr int L2 = 32;
constexpr int L3 = 32;
constexpr IndexType PSQTBuckets = 8;
@@ -59,12 +59,13 @@ struct NetworkArchitecture {
static constexpr int FC_0_OUTPUTS = L2;
static constexpr int FC_1_OUTPUTS = L3;
Layers::AffineTransformSparseInput<TransformedFeatureDimensions, FC_0_OUTPUTS + 1> fc_0;
Layers::SqrClippedReLU<FC_0_OUTPUTS + 1, WeightScaleBits + 1> ac_sqr_0;
Layers::ClippedReLU<FC_0_OUTPUTS + 1, WeightScaleBits + 1> ac_0;
Layers::AffineTransform<FC_0_OUTPUTS * 2, FC_1_OUTPUTS> fc_1;
Layers::ClippedReLU<FC_1_OUTPUTS, WeightScaleBits> ac_1;
Layers::AffineTransform<FC_1_OUTPUTS, 1> fc_2;
Layers::AffineTransformSparseInput<TransformedFeatureDimensions, FC_0_OUTPUTS> fc_0;
Layers::SqrClippedReLU<FC_0_OUTPUTS, WeightScaleBits + 1> ac_sqr_0;
Layers::ClippedReLU<FC_0_OUTPUTS, WeightScaleBits + 1> ac_0;
Layers::AffineTransform<FC_0_OUTPUTS * 2, FC_1_OUTPUTS> fc_1;
Layers::SqrClippedReLU<FC_1_OUTPUTS, WeightScaleBits> ac_sqr_1;
Layers::ClippedReLU<FC_1_OUTPUTS, WeightScaleBits> ac_1;
Layers::AffineTransform<FC_0_OUTPUTS * 2 + FC_1_OUTPUTS * 2, 1> fc_2;
// Hash value embedded in the evaluation file
static constexpr u32 get_hash_value() {
@@ -102,34 +103,39 @@ struct NetworkArchitecture {
struct alignas(CacheLineSize) Buffer {
alignas(CacheLineSize) typename decltype(fc_0)::OutputBuffer fc_0_out;
alignas(CacheLineSize) typename decltype(ac_sqr_0)::OutputType
ac_sqr_0_out[ceil_to_multiple<IndexType>(FC_0_OUTPUTS * 2, 32)];
concat_buffer[ceil_to_multiple<IndexType>(FC_0_OUTPUTS * 2 + FC_1_OUTPUTS * 2, 32)];
alignas(CacheLineSize) typename decltype(ac_0)::OutputBuffer ac_0_out;
alignas(CacheLineSize) typename decltype(fc_1)::OutputBuffer fc_1_out;
alignas(CacheLineSize) typename decltype(ac_1)::OutputBuffer ac_1_out;
alignas(CacheLineSize) typename decltype(fc_2)::OutputBuffer fc_2_out;
Buffer() { std::memset(ac_sqr_0_out, 0, sizeof(ac_sqr_0_out)); }
Buffer() { std::memset(concat_buffer, 0, sizeof(concat_buffer)); }
};
Buffer buffer;
fc_0.propagate(transformedFeatures, buffer.fc_0_out, nnzInfo);
ac_sqr_0.propagate(buffer.fc_0_out, buffer.ac_sqr_0_out);
ac_sqr_0.propagate(buffer.fc_0_out, buffer.concat_buffer);
ac_0.propagate(buffer.fc_0_out, buffer.ac_0_out);
std::memcpy(buffer.ac_sqr_0_out + FC_0_OUTPUTS, buffer.ac_0_out,
std::memcpy(buffer.concat_buffer + FC_0_OUTPUTS, buffer.ac_0_out,
FC_0_OUTPUTS * sizeof(typename decltype(ac_0)::OutputType));
fc_1.propagate(buffer.ac_sqr_0_out, buffer.fc_1_out);
ac_1.propagate(buffer.fc_1_out, buffer.ac_1_out);
fc_2.propagate(buffer.ac_1_out, buffer.fc_2_out);
// max value for fwdOut is (L1 + L3) * HiddenMaxVal * WeightMaxVal
// for int8 activations and weights this is (L1 + L3) * 16129 making
// fwdOut safe from overflow until (L1 + L3) > 133,144
// first layer and last layer use WeightScaleBits + 1
i32 fwdOut = buffer.fc_2_out[0] + buffer.fc_0_out[FC_0_OUTPUTS];
fc_1.propagate(buffer.concat_buffer, buffer.fc_1_out);
ac_sqr_1.propagate(buffer.fc_1_out, buffer.concat_buffer + FC_0_OUTPUTS * 2);
ac_1.propagate(buffer.fc_1_out, buffer.ac_1_out);
std::memcpy(buffer.concat_buffer + FC_0_OUTPUTS * 2 + FC_1_OUTPUTS, buffer.ac_1_out,
FC_1_OUTPUTS * sizeof(typename decltype(ac_1)::OutputType));
fc_2.propagate(buffer.concat_buffer, buffer.fc_2_out);
static_assert(FC_0_OUTPUTS >= 2);
i32 fwdOut = buffer.fc_2_out[0];
i32 skip_0 = buffer.fc_0_out[FC_0_OUTPUTS - 2] - buffer.fc_0_out[FC_0_OUTPUTS - 1];
fwdOut += skip_0;
// fwdOut is such that 1.0 is equal to HiddenOneVal*(1<<WeightScaleBits)*2 in
// quantized form, but we want 1.0 to be equal to 600*OutputScale
// to make overflow impossible we cast to i64
// to make overflow impossible we cast to int64_t
constexpr i64 multiplier = 600 * OutputScale;
constexpr i64 denominator =
static_cast<i64>(HiddenOneVal) * static_cast<i64>(1U << WeightScaleBits) * 2;
@@ -144,6 +150,7 @@ struct NetworkArchitecture {
hash_combine(h, ac_sqr_0.get_content_hash());
hash_combine(h, ac_0.get_content_hash());
hash_combine(h, fc_1.get_content_hash());
// hash_combine(h, ac_sqr_1.get_content_hash()); TODO
hash_combine(h, ac_1.get_content_hash());
hash_combine(h, fc_2.get_content_hash());
hash_combine(h, get_hash_value());