/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
- Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file)
+ Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
Stockfish is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
#include <algorithm>
#include <type_traits>
#include "../nnue_common.h"
-#include "../../simd.h"
+#include "simd.h"
/*
This file contains the definition for a fully connected layer (aka affine transform).
Two approaches are employed, depending on the sizes of the transform.
- Approach 1:
+ Approach 1 (a specialization for large inputs):
- used when the PaddedInputDimensions >= 128
- uses AVX512 if possible
- processes inputs in batches of 2*InputSimdWidth
depends on the architecture (the amount of registers)
- accumulate + hadd is used
- Approach 2:
+ Approach 2 (a specialization for small inputs):
- used when the PaddedInputDimensions < 128
- - does not use AVX512
- expected use-case is for when PaddedInputDimensions == 32 and InputDimensions <= 32.
- that's why AVX512 is hard to implement
- expected use-case is small 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_DOTPROD)
+ constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
+ const auto inputVector = reinterpret_cast<const int8x16_t*>(input);
+
# elif defined(USE_NEON)
- static_assert(PaddedInputDimensions % 16 == 0);
- constexpr IndexType NumChunks = PaddedInputDimensions / 16;
+ constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
const auto inputVector = reinterpret_cast<const int8x8_t*>(input);
# endif
sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum));
output[i] = _mm_cvtsi64_si32(sum);
+# elif defined(USE_NEON_DOTPROD)
+ int32x4_t sum = {biases[i]};
+ const auto row = reinterpret_cast<const int8x16_t*>(&weights[offset]);
+ for (IndexType j = 0; j < NumChunks; ++j) {
+ sum = vdotq_s32(sum, inputVector[j], row[j]);
+ }
+ output[i] = vaddvq_s32(sum);
+
# elif defined(USE_NEON)
int32x4_t sum = {biases[i]};
const auto row = reinterpret_cast<const int8x8_t*>(&weights[offset]);
}
#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)>> {
+#if defined (USE_AVX512)
+ constexpr IndexType LargeInputSize = 2 * 64;
+#else
+ constexpr IndexType LargeInputSize = std::numeric_limits<IndexType>::max();
+#endif
+
+ // A specialization for large inputs
+ template <IndexType InDims, IndexType OutDims>
+ class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) >= LargeInputSize)>> {
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.");
+ static_assert(PaddedInputDimensions >= LargeInputSize, "Something went wrong. This specialization (for large inputs) should not have been chosen.");
#if defined (USE_AVX512)
- static constexpr const IndexType InputSimdWidth = 64;
- static constexpr const IndexType MaxNumOutputRegs = 16;
+ static constexpr IndexType InputSimdWidth = 64;
+ static constexpr IndexType MaxNumOutputRegs = 16;
#elif defined (USE_AVX2)
- static constexpr const IndexType InputSimdWidth = 32;
- static constexpr const IndexType MaxNumOutputRegs = 8;
+ static constexpr IndexType InputSimdWidth = 32;
+ static constexpr IndexType MaxNumOutputRegs = 8;
#elif defined (USE_SSSE3)
- static constexpr const IndexType InputSimdWidth = 16;
- static constexpr const IndexType MaxNumOutputRegs = 8;
+ static constexpr IndexType InputSimdWidth = 16;
+ static constexpr IndexType MaxNumOutputRegs = 8;
+#elif defined (USE_NEON_DOTPROD)
+ static constexpr IndexType InputSimdWidth = 16;
+ static constexpr IndexType MaxNumOutputRegs = 8;
+#elif defined (USE_NEON)
+ static constexpr IndexType InputSimdWidth = 8;
+ static constexpr IndexType MaxNumOutputRegs = 8;
#else
// The fallback implementation will not have permuted weights.
// We define these to avoid a lot of ifdefs later.
- static constexpr const IndexType InputSimdWidth = 1;
- static constexpr const IndexType MaxNumOutputRegs = 1;
+ static constexpr IndexType InputSimdWidth = 1;
+ static constexpr IndexType MaxNumOutputRegs = 1;
#endif
// A big block is a region in the weight matrix of the size [PaddedInputDimensions, NumOutputRegs].
// A small block is a region of size [InputSimdWidth, 1]
- static constexpr const IndexType NumOutputRegs = std::min(MaxNumOutputRegs, OutputDimensions);
- static constexpr const IndexType SmallBlockSize = InputSimdWidth;
- static constexpr const IndexType BigBlockSize = NumOutputRegs * PaddedInputDimensions;
- static constexpr const IndexType NumSmallBlocksInBigBlock = BigBlockSize / SmallBlockSize;
- static constexpr const IndexType NumSmallBlocksPerOutput = PaddedInputDimensions / SmallBlockSize;
- static constexpr const IndexType NumBigBlocks = OutputDimensions / NumOutputRegs;
+ static constexpr IndexType NumOutputRegs = std::min(MaxNumOutputRegs, OutputDimensions);
+ static constexpr IndexType SmallBlockSize = InputSimdWidth;
+ static constexpr IndexType BigBlockSize = NumOutputRegs * PaddedInputDimensions;
+ static constexpr IndexType NumSmallBlocksInBigBlock = BigBlockSize / SmallBlockSize;
+ static constexpr IndexType NumSmallBlocksPerOutput = PaddedInputDimensions / SmallBlockSize;
+ static constexpr IndexType NumBigBlocks = OutputDimensions / NumOutputRegs;
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;
}
// 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);
+ read_little_endian<BiasType>(stream, biases, OutputDimensions);
- 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);
return !stream.fail();
// 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]);
+ write_little_endian<BiasType>(stream, biases, OutputDimensions);
- 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)]);
return !stream.fail();
// 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 vec_t = __m512i;
- #define vec_setzero _mm512_setzero_si512
- #define vec_set_32 _mm512_set1_epi32
- #define vec_add_dpbusd_32 Simd::m512_add_dpbusd_epi32
+ using acc_vec_t = __m512i;
+ using bias_vec_t = __m128i;
+ using weight_vec_t = __m512i;
+ using in_vec_t = __m512i;
+ #define vec_zero _mm512_setzero_si512()
#define vec_add_dpbusd_32x2 Simd::m512_add_dpbusd_epi32x2
#define vec_hadd Simd::m512_hadd
#define vec_haddx4 Simd::m512_haddx4
#elif defined (USE_AVX2)
- using vec_t = __m256i;
- #define vec_setzero _mm256_setzero_si256
- #define vec_set_32 _mm256_set1_epi32
- #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
+ using acc_vec_t = __m256i;
+ using bias_vec_t = __m128i;
+ using weight_vec_t = __m256i;
+ using in_vec_t = __m256i;
+ #define vec_zero _mm256_setzero_si256()
#define vec_add_dpbusd_32x2 Simd::m256_add_dpbusd_epi32x2
#define vec_hadd Simd::m256_hadd
#define vec_haddx4 Simd::m256_haddx4
#elif defined (USE_SSSE3)
- using vec_t = __m128i;
- #define vec_setzero _mm_setzero_si128
- #define vec_set_32 _mm_set1_epi32
- #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
+ using acc_vec_t = __m128i;
+ using bias_vec_t = __m128i;
+ using weight_vec_t = __m128i;
+ using in_vec_t = __m128i;
+ #define vec_zero _mm_setzero_si128()
#define vec_add_dpbusd_32x2 Simd::m128_add_dpbusd_epi32x2
#define vec_hadd Simd::m128_hadd
#define vec_haddx4 Simd::m128_haddx4
+#elif defined (USE_NEON_DOTPROD)
+ using acc_vec_t = int32x4_t;
+ using bias_vec_t = int32x4_t;
+ using weight_vec_t = int8x16_t;
+ using in_vec_t = int8x16_t;
+ #define vec_zero {0}
+ #define vec_add_dpbusd_32x2 Simd::dotprod_m128_add_dpbusd_epi32x2
+ #define vec_hadd Simd::neon_m128_hadd
+ #define vec_haddx4 Simd::neon_m128_haddx4
+#elif defined (USE_NEON)
+ using acc_vec_t = int32x4_t;
+ using bias_vec_t = int32x4_t;
+ using weight_vec_t = int8x8_t;
+ using in_vec_t = int8x8_t;
+ #define vec_zero {0}
+ #define vec_add_dpbusd_32x2 Simd::neon_m128_add_dpbusd_epi32x2
+ #define vec_hadd Simd::neon_m128_hadd
+ #define vec_haddx4 Simd::neon_m128_haddx4
#endif
-#if defined (USE_SSSE3)
- const vec_t* invec = reinterpret_cast<const vec_t*>(input);
-
+#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)
{
- vec_t acc[NumOutputRegs] = { vec_setzero() };
+ acc_vec_t acc[NumOutputRegs] = { vec_zero };
// Each big block has NumOutputRegs small blocks in each "row", one per register.
// We process two small blocks at a time to save on one addition without VNNI.
for (IndexType smallBlock = 0; smallBlock < NumSmallBlocksPerOutput; smallBlock += 2)
{
- const vec_t* weightvec =
- reinterpret_cast<const vec_t*>(
+ const weight_vec_t* weightvec =
+ reinterpret_cast<const weight_vec_t*>(
weights
+ bigBlock * BigBlockSize
+ smallBlock * SmallBlockSize * NumOutputRegs);
- const vec_t in0 = invec[smallBlock + 0];
- const vec_t in1 = invec[smallBlock + 1];
+ const in_vec_t in0 = invec[smallBlock + 0];
+ const in_vec_t in1 = invec[smallBlock + 1];
for (IndexType k = 0; k < NumOutputRegs; ++k)
vec_add_dpbusd_32x2(acc[k], in0, weightvec[k], in1, weightvec[k + NumOutputRegs]);
// Horizontally add all accumulators.
if constexpr (NumOutputRegs % 4 == 0)
{
- __m128i* outputvec = reinterpret_cast<__m128i*>(output);
- const __m128i* biasvec = reinterpret_cast<const __m128i*>(biases);
+ bias_vec_t* outputvec = reinterpret_cast<bias_vec_t*>(output);
+ const bias_vec_t* biasvec = reinterpret_cast<const bias_vec_t*>(biases);
for (IndexType k = 0; k < NumOutputRegs; k += 4)
{
}
}
-# undef vec_setzero
-# undef vec_set_32
-# undef vec_add_dpbusd_32
+# undef vec_zero
# undef vec_add_dpbusd_32x2
# undef vec_hadd
# undef vec_haddx4
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)>> {
+ // A specialization for small inputs
+ template <IndexType InDims, IndexType OutDims>
+ class AffineTransform<InDims, OutDims, std::enable_if_t<(ceil_to_multiple<IndexType>(InDims, MaxSimdWidth) < LargeInputSize)>> {
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);
- static_assert(PaddedInputDimensions < 128, "Something went wrong. This specialization should not have been chosen.");
-
-#if defined (USE_SSSE3)
- static constexpr const IndexType OutputSimdWidth = SimdWidth / 4;
- static constexpr const IndexType InputSimdWidth = SimdWidth;
-#endif
+ static constexpr IndexType PaddedInputDimensions =
+ ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
+ static constexpr IndexType PaddedOutputDimensions =
+ ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
- // Size of forward propagation buffer used in this layer
- static constexpr std::size_t SelfBufferSize =
- ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);
+ using OutputBuffer = OutputType[PaddedOutputDimensions];
- // Size of the forward propagation buffer used from the input layer to this layer
- static constexpr std::size_t BufferSize =
- PreviousLayer::BufferSize + SelfBufferSize;
+ static_assert(PaddedInputDimensions < LargeInputSize, "Something went wrong. This specialization (for small inputs) should not have been chosen.");
// 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;
}
// 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)
+ read_little_endian<BiasType>(stream, biases, OutputDimensions);
+ for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
return !stream.fail();
// 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]);
+ write_little_endian<BiasType>(stream, biases, OutputDimensions);
- 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)]);
return !stream.fail();
}
// 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 defined (USE_AVX512)
+ using vec_t = __m512i;
+ #define vec_setzero _mm512_setzero_si512
+ #define vec_set_32 _mm512_set1_epi32
+ #define vec_add_dpbusd_32 Simd::m512_add_dpbusd_epi32
+ #define vec_add_dpbusd_32x2 Simd::m512_add_dpbusd_epi32x2
+ #define vec_hadd Simd::m512_hadd
+#elif defined (USE_AVX2)
using vec_t = __m256i;
#define vec_setzero _mm256_setzero_si256
#define vec_set_32 _mm256_set1_epi32
#define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
#define vec_add_dpbusd_32x2 Simd::m256_add_dpbusd_epi32x2
- #define vec_add_dpbusd_32x4 Simd::m256_add_dpbusd_epi32x4
#define vec_hadd Simd::m256_hadd
- #define vec_haddx4 Simd::m256_haddx4
#elif defined (USE_SSSE3)
using vec_t = __m128i;
#define vec_setzero _mm_setzero_si128
#define vec_set_32 _mm_set1_epi32
#define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
#define vec_add_dpbusd_32x2 Simd::m128_add_dpbusd_epi32x2
- #define vec_add_dpbusd_32x4 Simd::m128_add_dpbusd_epi32x4
#define vec_hadd Simd::m128_hadd
- #define vec_haddx4 Simd::m128_haddx4
#endif
#if defined (USE_SSSE3)
const auto inputVector = reinterpret_cast<const vec_t*>(input);
- static_assert(InputDimensions % 8 == 0);
+ static constexpr IndexType OutputSimdWidth = sizeof(vec_t) / sizeof(OutputType);
+
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);
# undef vec_set_32
# undef vec_add_dpbusd_32
# undef vec_add_dpbusd_32x2
-# undef vec_add_dpbusd_32x4
# undef vec_hadd
-# undef vec_haddx4
#else
// Use old implementation for the other architectures.
affine_transform_non_ssse3<
using BiasType = OutputType;
using WeightType = std::int8_t;
- PreviousLayer previousLayer;
-
alignas(CacheLineSize) BiasType biases[OutputDimensions];
alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
};