kate/BadFish
1
0
Fork 0
mirror of https://github.com/sockspls/badfish synced 2025-05-01 01:03:09 +00:00
Code Issues Releases Wiki Activity

Clean up and simplify some nnue code.

Browse source 
Remove some unnecessary code and it's execution during inference. Also the change on line 49 in nnue_architecture.h results in a more efficient SIMD code path through ClippedReLU::propagate().

passed STC:
https://tests.stockfishchess.org/tests/view/6217d3bfda649bba32ef25d5
LLR: 2.94 (-2.94,2.94) <-2.25,0.25>
Total: 12056 W: 3281 L: 3092 D: 5683
Ptnml(0-2): 55, 1213, 3312, 1384, 64

passed STC SMP:
https://tests.stockfishchess.org/tests/view/6217f344da649bba32ef295e
LLR: 2.94 (-2.94,2.94) <-2.25,0.25>
Total: 27376 W: 7295 L: 7137 D: 12944
Ptnml(0-2): 52, 2859, 7715, 3003, 59

closes https://github.com/official-stockfish/Stockfish/pull/3944

No functional change

bench: 6820724
This commit is contained in:
mstembera 2022-02-23 18:19:36 -08:00 • committed by Joost VandeVondele
parent 27139dedac
commit 5f781d366e
6 changed files with 27 additions and 32 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
src/nnue/evaluate_nnue.cpp
View file

@ -109,7 +109,7 @@ namespace Stockfish::Eval::NNUE {
{ {
write_little_endian<std::uint32_t>(stream, Version); write_little_endian<std::uint32_t>(stream, Version);
write_little_endian<std::uint32_t>(stream, hashValue); write_little_endian<std::uint32_t>(stream, hashValue);
write_little_endian<std::uint32_t>(stream, desc.size()); write_little_endian<std::uint32_t>(stream, (std::uint32_t)desc.size());
stream.write(&desc[0], desc.size()); stream.write(&desc[0], desc.size());
return !stream.fail(); return !stream.fail();
} }
@ -157,7 +157,7 @@ namespace Stockfish::Eval::NNUE {
ASSERT_ALIGNED(transformedFeatures, alignment); ASSERT_ALIGNED(transformedFeatures, alignment);
const std::size_t bucket = (pos.count<ALL_PIECES>() - 1) / 4; const int bucket = (pos.count<ALL_PIECES>() - 1) / 4;
const auto psqt = featureTransformer->transform(pos, transformedFeatures, bucket); const auto psqt = featureTransformer->transform(pos, transformedFeatures, bucket);
const auto positional = network[bucket]->propagate(transformedFeatures); const auto positional = network[bucket]->propagate(transformedFeatures);
@ -197,7 +197,7 @@ namespace Stockfish::Eval::NNUE {
NnueEvalTrace t{}; NnueEvalTrace t{};
t.correctBucket = (pos.count<ALL_PIECES>() - 1) / 4; t.correctBucket = (pos.count<ALL_PIECES>() - 1) / 4;
for (std::size_t bucket = 0; bucket < LayerStacks; ++bucket) { for (IndexType bucket = 0; bucket < LayerStacks; ++bucket) {
const auto materialist = featureTransformer->transform(pos, transformedFeatures, bucket); const auto materialist = featureTransformer->transform(pos, transformedFeatures, bucket);
const auto positional = network[bucket]->propagate(transformedFeatures); const auto positional = network[bucket]->propagate(transformedFeatures);

16
src/nnue/layers/affine_transform.h
View file

@ -235,10 +235,10 @@ namespace Stockfish::Eval::NNUE::Layers {
// Read network parameters // Read network parameters
bool read_parameters(std::istream& stream) { bool read_parameters(std::istream& stream) {
for (std::size_t i = 0; i < OutputDimensions; ++i) for (IndexType i = 0; i < OutputDimensions; ++i)
biases[i] = read_little_endian<BiasType>(stream); biases[i] = read_little_endian<BiasType>(stream);
for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i) for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
weights[get_weight_index(i)] = read_little_endian<WeightType>(stream); weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
return !stream.fail(); return !stream.fail();
@ -246,10 +246,10 @@ namespace Stockfish::Eval::NNUE::Layers {
// Write network parameters // Write network parameters
bool write_parameters(std::ostream& stream) const { bool write_parameters(std::ostream& stream) const {
for (std::size_t i = 0; i < OutputDimensions; ++i) for (IndexType i = 0; i < OutputDimensions; ++i)
write_little_endian<BiasType>(stream, biases[i]); write_little_endian<BiasType>(stream, biases[i]);
for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i) for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
write_little_endian<WeightType>(stream, weights[get_weight_index(i)]); write_little_endian<WeightType>(stream, weights[get_weight_index(i)]);
return !stream.fail(); return !stream.fail();
@ -422,9 +422,9 @@ namespace Stockfish::Eval::NNUE::Layers {
// Read network parameters // Read network parameters
bool read_parameters(std::istream& stream) { bool read_parameters(std::istream& stream) {
for (std::size_t i = 0; i < OutputDimensions; ++i) for (IndexType i = 0; i < OutputDimensions; ++i)
biases[i] = read_little_endian<BiasType>(stream); biases[i] = read_little_endian<BiasType>(stream);
for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i) for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
weights[get_weight_index(i)] = read_little_endian<WeightType>(stream); weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
return !stream.fail(); return !stream.fail();
@ -432,10 +432,10 @@ namespace Stockfish::Eval::NNUE::Layers {
// Write network parameters // Write network parameters
bool write_parameters(std::ostream& stream) const { bool write_parameters(std::ostream& stream) const {
for (std::size_t i = 0; i < OutputDimensions; ++i) for (IndexType i = 0; i < OutputDimensions; ++i)
write_little_endian<BiasType>(stream, biases[i]); write_little_endian<BiasType>(stream, biases[i]);
for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i) for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
write_little_endian<WeightType>(stream, weights[get_weight_index(i)]); write_little_endian<WeightType>(stream, weights[get_weight_index(i)]);
return !stream.fail(); return !stream.fail();

8
src/nnue/layers/clipped_relu.h
View file

@ -171,14 +171,6 @@ namespace Stockfish::Eval::NNUE::Layers {
std::max(0, std::min(127, input[i] >> WeightScaleBits))); std::max(0, std::min(127, input[i] >> WeightScaleBits)));
} }
// Affine transform layers expect that there is at least
// ceil_to_multiple(OutputDimensions, 32) initialized values.
// We cannot do this in the affine transform because it requires
// preallocating space here.
for (IndexType i = OutputDimensions; i < PaddedOutputDimensions; ++i) {
output[i] = 0;
}
return output; return output;
} }
}; };

19
src/nnue/nnue_architecture.h
View file

@ -46,7 +46,7 @@ struct Network
static constexpr int FC_1_OUTPUTS = 32; static constexpr int FC_1_OUTPUTS = 32;
Layers::AffineTransform<TransformedFeatureDimensions, FC_0_OUTPUTS + 1> fc_0; Layers::AffineTransform<TransformedFeatureDimensions, FC_0_OUTPUTS + 1> fc_0;
Layers::ClippedReLU<FC_0_OUTPUTS> ac_0; Layers::ClippedReLU<FC_0_OUTPUTS + 1> ac_0;
Layers::AffineTransform<FC_0_OUTPUTS, FC_1_OUTPUTS> fc_1; Layers::AffineTransform<FC_0_OUTPUTS, FC_1_OUTPUTS> fc_1;
Layers::ClippedReLU<FC_1_OUTPUTS> ac_1; Layers::ClippedReLU<FC_1_OUTPUTS> ac_1;
Layers::AffineTransform<FC_1_OUTPUTS, 1> fc_2; Layers::AffineTransform<FC_1_OUTPUTS, 1> fc_2;
@ -97,14 +97,19 @@ struct Network
alignas(CacheLineSize) decltype(fc_1)::OutputBuffer fc_1_out; alignas(CacheLineSize) decltype(fc_1)::OutputBuffer fc_1_out;
alignas(CacheLineSize) decltype(ac_1)::OutputBuffer ac_1_out; alignas(CacheLineSize) decltype(ac_1)::OutputBuffer ac_1_out;
alignas(CacheLineSize) decltype(fc_2)::OutputBuffer fc_2_out; alignas(CacheLineSize) decltype(fc_2)::OutputBuffer fc_2_out;
Buffer()
{
std::memset(this, 0, sizeof(*this));
}
}; };
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN) #if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
char bufferRaw[sizeof(Buffer) + alignment]; static thread_local char bufferRaw[sizeof(Buffer) + alignment];
char* bufferRawAligned = align_ptr_up<alignment>(&bufferRaw[0]); static thread_local char* bufferRawAligned = align_ptr_up<alignment>(&bufferRaw[0]);
Buffer& buffer = *(new (bufferRawAligned) Buffer); static thread_local Buffer& buffer = *(new (bufferRawAligned) Buffer);
#else #else
alignas(alignment) Buffer buffer; alignas(alignment) static thread_local Buffer buffer;
#endif #endif
fc_0.propagate(transformedFeatures, buffer.fc_0_out); fc_0.propagate(transformedFeatures, buffer.fc_0_out);
@ -118,10 +123,6 @@ struct Network
std::int32_t fwdOut = int(buffer.fc_0_out[FC_0_OUTPUTS]) * (600*OutputScale) / (127*(1<<WeightScaleBits)); std::int32_t fwdOut = int(buffer.fc_0_out[FC_0_OUTPUTS]) * (600*OutputScale) / (127*(1<<WeightScaleBits));
std::int32_t outputValue = buffer.fc_2_out[0] + fwdOut; std::int32_t outputValue = buffer.fc_2_out[0] + fwdOut;
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
buffer.~Buffer();
#endif
return outputValue; return outputValue;
} }
}; };

4
src/nnue/nnue_common.h
View file

@ -127,11 +127,11 @@ namespace Stockfish::Eval::NNUE {
{ {
for (; i + 1 < sizeof(IntType); ++i) for (; i + 1 < sizeof(IntType); ++i)
{ {
u[i] = v; u[i] = (std::uint8_t)v;
v >>= 8; v >>= 8;
} }
} }
u[i] = v; u[i] = (std::uint8_t)v;
stream.write(reinterpret_cast<char*>(u), sizeof(IntType)); stream.write(reinterpret_cast<char*>(u), sizeof(IntType));
} }

6
src/nnue/nnue_feature_transformer.h
View file

@ -123,8 +123,10 @@ namespace Stockfish::Eval::NNUE {
// We use __m* types as template arguments, which causes GCC to emit warnings // We use __m* types as template arguments, which causes GCC to emit warnings
// about losing some attribute information. This is irrelevant to us as we // about losing some attribute information. This is irrelevant to us as we
// only take their size, so the following pragma are harmless. // only take their size, so the following pragma are harmless.
#if defined(__GNUC__)
#pragma GCC diagnostic push #pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wignored-attributes" #pragma GCC diagnostic ignored "-Wignored-attributes"
#endif
template <typename SIMDRegisterType, template <typename SIMDRegisterType,
typename LaneType, typename LaneType,
@ -156,9 +158,9 @@ namespace Stockfish::Eval::NNUE {
static constexpr int NumRegs = BestRegisterCount<vec_t, WeightType, TransformedFeatureDimensions, NumRegistersSIMD>(); static constexpr int NumRegs = BestRegisterCount<vec_t, WeightType, TransformedFeatureDimensions, NumRegistersSIMD>();
static constexpr int NumPsqtRegs = BestRegisterCount<psqt_vec_t, PSQTWeightType, PSQTBuckets, NumRegistersSIMD>(); static constexpr int NumPsqtRegs = BestRegisterCount<psqt_vec_t, PSQTWeightType, PSQTBuckets, NumRegistersSIMD>();
#if defined(__GNUC__)
#pragma GCC diagnostic pop #pragma GCC diagnostic pop
#endif
#endif #endif