Skip to content

Update architecture to "SFNNv4". Update network to nn-6877cd24400e.nnue. #3927

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
Sopel97 wants to merge 1 commit into from
Closed

Update architecture to "SFNNv4". Update network to nn-6877cd24400e.nnue. #3927

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion src/evaluate.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,7 +39,7 @@ namespace Eval {
// The default net name MUST follow the format nn-[SHA256 first 12 digits].nnue
// for the build process (profile-build and fishtest) to work. Do not change the
// name of the macro, as it is used in the Makefile.
#define EvalFileDefaultName "nn-ac07bd334b62.nnue"
#define EvalFileDefaultName "nn-6877cd24400e.nnue"

namespace NNUE {

Expand Down
17 changes: 3 additions & 14 deletions src/nnue/evaluate_nnue.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -148,22 +148,18 @@ namespace Stockfish::Eval::NNUE {
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
TransformedFeatureType transformedFeaturesUnaligned[
FeatureTransformer::BufferSize + alignment / sizeof(TransformedFeatureType)];
char bufferUnaligned[Network::BufferSize + alignment];

auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
auto* buffer = align_ptr_up<alignment>(&bufferUnaligned[0]);
#else
alignas(alignment)
TransformedFeatureType transformedFeatures[FeatureTransformer::BufferSize];
alignas(alignment) char buffer[Network::BufferSize];
#endif

ASSERT_ALIGNED(transformedFeatures, alignment);
ASSERT_ALIGNED(buffer, alignment);

const std::size_t bucket = (pos.count<ALL_PIECES>() - 1) / 4;
const auto psqt = featureTransformer->transform(pos, transformedFeatures, bucket);
const auto positional = network[bucket]->propagate(transformedFeatures, buffer)[0];
const auto positional = network[bucket]->propagate(transformedFeatures);

// Give more value to positional evaluation when adjusted flag is set
if (adjusted)
Expand All @@ -190,27 +186,20 @@ namespace Stockfish::Eval::NNUE {
#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
TransformedFeatureType transformedFeaturesUnaligned[
FeatureTransformer::BufferSize + alignment / sizeof(TransformedFeatureType)];
char bufferUnaligned[Network::BufferSize + alignment];

auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
auto* buffer = align_ptr_up<alignment>(&bufferUnaligned[0]);
#else
alignas(alignment)
TransformedFeatureType transformedFeatures[FeatureTransformer::BufferSize];
alignas(alignment) char buffer[Network::BufferSize];
#endif

ASSERT_ALIGNED(transformedFeatures, alignment);
ASSERT_ALIGNED(buffer, alignment);

NnueEvalTrace t{};
t.correctBucket = (pos.count<ALL_PIECES>() - 1) / 4;
for (std::size_t bucket = 0; bucket < LayerStacks; ++bucket) {
const auto psqt = featureTransformer->transform(pos, transformedFeatures, bucket);
const auto output = network[bucket]->propagate(transformedFeatures, buffer);

int materialist = psqt;
int positional = output[0];
const auto materialist = featureTransformer->transform(pos, transformedFeatures, bucket);
const auto positional = network[bucket]->propagate(transformedFeatures);

t.psqt[bucket] = static_cast<Value>( materialist / OutputScale );
t.positional[bucket] = static_cast<Value>( positional / OutputScale );
Expand Down
91 changes: 32 additions & 59 deletions src/nnue/layers/affine_transform.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,19 +63,17 @@ namespace Stockfish::Eval::NNUE::Layers {
{
# if defined(USE_SSE2)
// At least a multiple of 16, with SSE2.
static_assert(PaddedInputDimensions % 16 == 0);
constexpr IndexType NumChunks = PaddedInputDimensions / 16;
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
const __m128i Zeros = _mm_setzero_si128();
const auto inputVector = reinterpret_cast<const __m128i*>(input);

# elif defined(USE_MMX)
static_assert(InputDimensions % 8 == 0);
constexpr IndexType NumChunks = InputDimensions / 8;
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 8;
const __m64 Zeros = _mm_setzero_si64();
const auto inputVector = reinterpret_cast<const __m64*>(input);

# elif defined(USE_NEON)
constexpr IndexType NumChunks = (InputDimensions + 15) / 16;
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
const auto inputVector = reinterpret_cast<const int8x8_t*>(input);
# endif

Expand Down Expand Up @@ -150,24 +148,27 @@ namespace Stockfish::Eval::NNUE::Layers {
}
#endif

template <typename PreviousLayer, IndexType OutDims, typename Enabled = void>
template <IndexType InDims, IndexType OutDims, typename Enabled = void>
class AffineTransform;

// A specialization for large inputs.
template <typename PreviousLayer, IndexType OutDims>
class AffineTransform<PreviousLayer, OutDims, std::enable_if_t<(PreviousLayer::OutputDimensions >= 2*64-1)>> {
template <IndexType InDims, IndexType OutDims>
class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) >= 2*64)>> {
public:
// Input/output type
using InputType = typename PreviousLayer::OutputType;
using InputType = std::uint8_t;
using OutputType = std::int32_t;
static_assert(std::is_same<InputType, std::uint8_t>::value, "");

// Number of input/output dimensions
static constexpr IndexType InputDimensions = PreviousLayer::OutputDimensions;
static constexpr IndexType InputDimensions = InDims;
static constexpr IndexType OutputDimensions = OutDims;

static constexpr IndexType PaddedInputDimensions =
ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
static constexpr IndexType PaddedOutputDimensions =
ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);

using OutputBuffer = OutputType[PaddedOutputDimensions];

static_assert(PaddedInputDimensions >= 128, "Something went wrong. This specialization should not have been chosen.");

Expand Down Expand Up @@ -202,20 +203,12 @@ namespace Stockfish::Eval::NNUE::Layers {

static_assert(OutputDimensions % NumOutputRegs == 0);

// Size of forward propagation buffer used in this layer
static constexpr std::size_t SelfBufferSize =
ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);

// Size of the forward propagation buffer used from the input layer to this layer
static constexpr std::size_t BufferSize =
PreviousLayer::BufferSize + SelfBufferSize;

// Hash value embedded in the evaluation file
static constexpr std::uint32_t get_hash_value() {
static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
std::uint32_t hashValue = 0xCC03DAE4u;
hashValue += OutputDimensions;
hashValue ^= PreviousLayer::get_hash_value() >> 1;
hashValue ^= PreviousLayer::get_hash_value() << 31;
hashValue ^= prevHash >> 1;
hashValue ^= prevHash << 31;
return hashValue;
}

Expand All @@ -242,7 +235,6 @@ namespace Stockfish::Eval::NNUE::Layers {

// Read network parameters
bool read_parameters(std::istream& stream) {
if (!previousLayer.read_parameters(stream)) return false;
for (std::size_t i = 0; i < OutputDimensions; ++i)
biases[i] = read_little_endian<BiasType>(stream);

Expand All @@ -254,7 +246,6 @@ namespace Stockfish::Eval::NNUE::Layers {

// Write network parameters
bool write_parameters(std::ostream& stream) const {
if (!previousLayer.write_parameters(stream)) return false;
for (std::size_t i = 0; i < OutputDimensions; ++i)
write_little_endian<BiasType>(stream, biases[i]);

Expand All @@ -266,10 +257,7 @@ namespace Stockfish::Eval::NNUE::Layers {

// Forward propagation
const OutputType* propagate(
const TransformedFeatureType* transformedFeatures, char* buffer) const {
const auto input = previousLayer.propagate(
transformedFeatures, buffer + SelfBufferSize);
OutputType* output = reinterpret_cast<OutputType*>(buffer);
const InputType* input, OutputType* output) const {

#if defined (USE_AVX512)
using acc_vec_t = __m512i;
Expand Down Expand Up @@ -312,7 +300,6 @@ namespace Stockfish::Eval::NNUE::Layers {
#if defined (USE_SSSE3) || defined (USE_NEON)
const in_vec_t* invec = reinterpret_cast<const in_vec_t*>(input);


// Perform accumulation to registers for each big block
for (IndexType bigBlock = 0; bigBlock < NumBigBlocks; ++bigBlock)
{
Expand Down Expand Up @@ -377,26 +364,28 @@ namespace Stockfish::Eval::NNUE::Layers {
using BiasType = OutputType;
using WeightType = std::int8_t;

PreviousLayer previousLayer;

alignas(CacheLineSize) BiasType biases[OutputDimensions];
alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
};

template <typename PreviousLayer, IndexType OutDims>
class AffineTransform<PreviousLayer, OutDims, std::enable_if_t<(PreviousLayer::OutputDimensions < 2*64-1)>> {
template <IndexType InDims, IndexType OutDims>
class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) < 2*64)>> {
public:
// Input/output type
using InputType = typename PreviousLayer::OutputType;
// Input/output type
using InputType = std::uint8_t;
using OutputType = std::int32_t;
static_assert(std::is_same<InputType, std::uint8_t>::value, "");

// Number of input/output dimensions
static constexpr IndexType InputDimensions =
PreviousLayer::OutputDimensions;
static constexpr IndexType InputDimensions = InDims;
static constexpr IndexType OutputDimensions = OutDims;

static constexpr IndexType PaddedInputDimensions =
ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
static constexpr IndexType PaddedOutputDimensions =
ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);

using OutputBuffer = OutputType[PaddedOutputDimensions];

static_assert(PaddedInputDimensions < 128, "Something went wrong. This specialization should not have been chosen.");

Expand All @@ -405,20 +394,12 @@ namespace Stockfish::Eval::NNUE::Layers {
static constexpr const IndexType InputSimdWidth = SimdWidth;
#endif

// Size of forward propagation buffer used in this layer
static constexpr std::size_t SelfBufferSize =
ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);

// Size of the forward propagation buffer used from the input layer to this layer
static constexpr std::size_t BufferSize =
PreviousLayer::BufferSize + SelfBufferSize;

// Hash value embedded in the evaluation file
static constexpr std::uint32_t get_hash_value() {
static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
std::uint32_t hashValue = 0xCC03DAE4u;
hashValue += OutputDimensions;
hashValue ^= PreviousLayer::get_hash_value() >> 1;
hashValue ^= PreviousLayer::get_hash_value() << 31;
hashValue ^= prevHash >> 1;
hashValue ^= prevHash << 31;
return hashValue;
}

Expand All @@ -441,7 +422,6 @@ namespace Stockfish::Eval::NNUE::Layers {

// Read network parameters
bool read_parameters(std::istream& stream) {
if (!previousLayer.read_parameters(stream)) return false;
for (std::size_t i = 0; i < OutputDimensions; ++i)
biases[i] = read_little_endian<BiasType>(stream);
for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
Expand All @@ -452,7 +432,6 @@ namespace Stockfish::Eval::NNUE::Layers {

// Write network parameters
bool write_parameters(std::ostream& stream) const {
if (!previousLayer.write_parameters(stream)) return false;
for (std::size_t i = 0; i < OutputDimensions; ++i)
write_little_endian<BiasType>(stream, biases[i]);

Expand All @@ -463,10 +442,7 @@ namespace Stockfish::Eval::NNUE::Layers {
}
// Forward propagation
const OutputType* propagate(
const TransformedFeatureType* transformedFeatures, char* buffer) const {
const auto input = previousLayer.propagate(
transformedFeatures, buffer + SelfBufferSize);
const auto output = reinterpret_cast<OutputType*>(buffer);
const InputType* input, OutputType* output) const {

#if defined (USE_AVX2)
using vec_t = __m256i;
Expand All @@ -491,12 +467,11 @@ namespace Stockfish::Eval::NNUE::Layers {
#if defined (USE_SSSE3)
const auto inputVector = reinterpret_cast<const vec_t*>(input);

static_assert(InputDimensions % 8 == 0);
static_assert(OutputDimensions % OutputSimdWidth == 0 || OutputDimensions == 1);

if constexpr (OutputDimensions % OutputSimdWidth == 0)
{
constexpr IndexType NumChunks = InputDimensions / 4;
constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 4;
constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth;

const auto input32 = reinterpret_cast<const std::int32_t*>(input);
Expand Down Expand Up @@ -555,8 +530,6 @@ namespace Stockfish::Eval::NNUE::Layers {
using BiasType = OutputType;
using WeightType = std::int8_t;

PreviousLayer previousLayer;

alignas(CacheLineSize) BiasType biases[OutputDimensions];
alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
};
Expand Down
35 changes: 11 additions & 24 deletions src/nnue/layers/clipped_relu.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,51 +26,41 @@
namespace Stockfish::Eval::NNUE::Layers {

// Clipped ReLU
template <typename PreviousLayer>
template <IndexType InDims>
class ClippedReLU {
public:
// Input/output type
using InputType = typename PreviousLayer::OutputType;
using InputType = std::int32_t;
using OutputType = std::uint8_t;
static_assert(std::is_same<InputType, std::int32_t>::value, "");

// Number of input/output dimensions
static constexpr IndexType InputDimensions = PreviousLayer::OutputDimensions;
static constexpr IndexType InputDimensions = InDims;
static constexpr IndexType OutputDimensions = InputDimensions;
static constexpr IndexType PaddedOutputDimensions =
ceil_to_multiple<IndexType>(OutputDimensions, 32);

// Size of forward propagation buffer used in this layer
static constexpr std::size_t SelfBufferSize =
ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);

// Size of the forward propagation buffer used from the input layer to this layer
static constexpr std::size_t BufferSize =
PreviousLayer::BufferSize + SelfBufferSize;
using OutputBuffer = OutputType[PaddedOutputDimensions];

// Hash value embedded in the evaluation file
static constexpr std::uint32_t get_hash_value() {
static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
std::uint32_t hashValue = 0x538D24C7u;
hashValue += PreviousLayer::get_hash_value();
hashValue += prevHash;
return hashValue;
}

// Read network parameters
bool read_parameters(std::istream& stream) {
return previousLayer.read_parameters(stream);
bool read_parameters(std::istream&) {
return true;
}

// Write network parameters
bool write_parameters(std::ostream& stream) const {
return previousLayer.write_parameters(stream);
bool write_parameters(std::ostream&) const {
return true;
}

// Forward propagation
const OutputType* propagate(
const TransformedFeatureType* transformedFeatures, char* buffer) const {
const auto input = previousLayer.propagate(
transformedFeatures, buffer + SelfBufferSize);
const auto output = reinterpret_cast<OutputType*>(buffer);
const InputType* input, OutputType* output) const {

#if defined(USE_AVX2)
if constexpr (InputDimensions % SimdWidth == 0) {
Expand Down Expand Up @@ -191,9 +181,6 @@ namespace Stockfish::Eval::NNUE::Layers {

return output;
}

private:
PreviousLayer previousLayer;
};

} // namespace Stockfish::Eval::NNUE::Layers
Expand Down
Loading