X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=src%2Fnnue%2Flayers%2Faffine_transform.h;h=fc65c34339ff35bb44f69dee3a85285de69eb51b;hb=8a912951de6d4bff78d3ff5258213a0c7e6f494e;hp=322e32402500029033caa5ddbb7123965adc45d7;hpb=dd63b98fb06e050aa961fbad6fd1f9316f2b17df;p=stockfish diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h index 322e3240..fc65c343 100644 --- a/src/nnue/layers/affine_transform.h +++ b/src/nnue/layers/affine_transform.h @@ -1,6 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2020 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 @@ -21,241 +21,296 @@ #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED #define NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED +#include #include + #include "../nnue_common.h" +#include "simd.h" + +/* + This file contains the definition for a fully connected layer (aka affine transform). + + - 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 + - inputs are processed in chunks of 4, weights are respectively transposed + - accumulation happens directly to int32s +*/ + +namespace Stockfish::Eval::NNUE::Layers { + +// Fallback implementation for older/other architectures. +// Requires the input to be padded to at least 16 values. +#if !defined(USE_SSSE3) + template + static void affine_transform_non_ssse3(std::int32_t* output, const std::int8_t* weights, const std::int32_t* biases, const std::uint8_t* input) + { +# if defined(USE_SSE2) || defined(USE_NEON_DOTPROD) || defined(USE_NEON) +# if defined(USE_SSE2) + // At least a multiple of 16, with SSE2. + constexpr IndexType NumChunks = ceil_to_multiple(InputDimensions, 16) / 16; + const __m128i Zeros = _mm_setzero_si128(); + const auto inputVector = reinterpret_cast(input); + +# elif defined(USE_NEON_DOTPROD) + constexpr IndexType NumChunks = ceil_to_multiple(InputDimensions, 16) / 16; + const auto inputVector = reinterpret_cast(input); + +# elif defined(USE_NEON) + constexpr IndexType NumChunks = ceil_to_multiple(InputDimensions, 16) / 16; + const auto inputVector = reinterpret_cast(input); +# endif + + for (IndexType i = 0; i < OutputDimensions; ++i) { + const IndexType offset = i * PaddedInputDimensions; + +# if defined(USE_SSE2) + __m128i sumLo = _mm_cvtsi32_si128(biases[i]); + __m128i sumHi = Zeros; + const auto row = reinterpret_cast(&weights[offset]); + for (IndexType j = 0; j < NumChunks; ++j) { + __m128i row_j = _mm_load_si128(&row[j]); + __m128i input_j = _mm_load_si128(&inputVector[j]); + __m128i extendedRowLo = _mm_srai_epi16(_mm_unpacklo_epi8(row_j, row_j), 8); + __m128i extendedRowHi = _mm_srai_epi16(_mm_unpackhi_epi8(row_j, row_j), 8); + __m128i extendedInputLo = _mm_unpacklo_epi8(input_j, Zeros); + __m128i extendedInputHi = _mm_unpackhi_epi8(input_j, Zeros); + __m128i productLo = _mm_madd_epi16(extendedRowLo, extendedInputLo); + __m128i productHi = _mm_madd_epi16(extendedRowHi, extendedInputHi); + sumLo = _mm_add_epi32(sumLo, productLo); + sumHi = _mm_add_epi32(sumHi, productHi); + } + __m128i sum = _mm_add_epi32(sumLo, sumHi); + __m128i sumHigh_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2)); + sum = _mm_add_epi32(sum, sumHigh_64); + __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2)); + sum = _mm_add_epi32(sum, sum_second_32); + output[i] = _mm_cvtsi128_si32(sum); + +# elif defined(USE_NEON_DOTPROD) + int32x4_t sum = {biases[i]}; + const auto row = reinterpret_cast(&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(&weights[offset]); + for (IndexType j = 0; j < NumChunks; ++j) { + int16x8_t product = vmull_s8(inputVector[j * 2], row[j * 2]); + product = vmlal_s8(product, inputVector[j * 2 + 1], row[j * 2 + 1]); + sum = vpadalq_s16(sum, product); + } + output[i] = sum[0] + sum[1] + sum[2] + sum[3]; -namespace Eval::NNUE::Layers { +# endif + } +# else + std::memcpy(output, biases, sizeof(std::int32_t) * OutputDimensions); + + // Traverse weights in transpose order to take advantage of input sparsity + for (IndexType i = 0; i < InputDimensions; ++i) + if (input[i]) { + const std::int8_t* w = &weights[i]; + const int in = input[i]; + for (IndexType j = 0; j < OutputDimensions; ++j) + output[j] += w[j * PaddedInputDimensions] * in; + } +# endif + } +#endif - // Affine transformation layer - template + template class AffineTransform { public: // Input/output type - using InputType = typename PreviousLayer::OutputType; + using InputType = std::uint8_t; using OutputType = std::int32_t; - static_assert(std::is_same::value, ""); // Number of input/output dimensions - static constexpr IndexType kInputDimensions = - PreviousLayer::kOutputDimensions; - static constexpr IndexType kOutputDimensions = OutputDimensions; - static constexpr IndexType kPaddedInputDimensions = - CeilToMultiple(kInputDimensions, kMaxSimdWidth); + static constexpr IndexType InputDimensions = InDims; + static constexpr IndexType OutputDimensions = OutDims; - // Size of forward propagation buffer used in this layer - static constexpr std::size_t kSelfBufferSize = - CeilToMultiple(kOutputDimensions * sizeof(OutputType), kCacheLineSize); + static constexpr IndexType PaddedInputDimensions = + ceil_to_multiple(InputDimensions, MaxSimdWidth); + static constexpr IndexType PaddedOutputDimensions = + ceil_to_multiple(OutputDimensions, MaxSimdWidth); - // Size of the forward propagation buffer used from the input layer to this layer - static constexpr std::size_t kBufferSize = - PreviousLayer::kBufferSize + kSelfBufferSize; + using OutputBuffer = OutputType[PaddedOutputDimensions]; // Hash value embedded in the evaluation file - static constexpr std::uint32_t GetHashValue() { - std::uint32_t hash_value = 0xCC03DAE4u; - hash_value += kOutputDimensions; - hash_value ^= PreviousLayer::GetHashValue() >> 1; - hash_value ^= PreviousLayer::GetHashValue() << 31; - return hash_value; + static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) { + std::uint32_t hashValue = 0xCC03DAE4u; + hashValue += OutputDimensions; + hashValue ^= prevHash >> 1; + hashValue ^= prevHash << 31; + return hashValue; + } + + static constexpr IndexType get_weight_index_scrambled(IndexType i) + { + return + (i / 4) % (PaddedInputDimensions / 4) * OutputDimensions * 4 + + i / PaddedInputDimensions * 4 + + i % 4; + } + + static constexpr IndexType get_weight_index(IndexType i) + { +#if defined (USE_SSSE3) + return get_weight_index_scrambled(i); +#else + return i; +#endif } - // Read network parameters - bool ReadParameters(std::istream& stream) { - if (!previous_layer_.ReadParameters(stream)) return false; - stream.read(reinterpret_cast(biases_), - kOutputDimensions * sizeof(BiasType)); - stream.read(reinterpret_cast(weights_), - kOutputDimensions * kPaddedInputDimensions * - sizeof(WeightType)); + // Read network parameters + bool read_parameters(std::istream& stream) { + read_little_endian(stream, biases, OutputDimensions); + for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i) + weights[get_weight_index(i)] = read_little_endian(stream); + return !stream.fail(); } + // Write network parameters + bool write_parameters(std::ostream& stream) const { + write_little_endian(stream, biases, OutputDimensions); + + for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i) + write_little_endian(stream, weights[get_weight_index(i)]); + + return !stream.fail(); + } // Forward propagation - const OutputType* Propagate( - const TransformedFeatureType* transformed_features, char* buffer) const { - const auto input = previous_layer_.Propagate( - transformed_features, buffer + kSelfBufferSize); - const auto output = reinterpret_cast(buffer); - - #if defined(USE_AVX512) - constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2); - const auto input_vector = reinterpret_cast(input); - #if !defined(USE_VNNI) - const __m512i kOnes = _mm512_set1_epi16(1); - #endif - - #elif defined(USE_AVX2) - constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; - const __m256i kOnes = _mm256_set1_epi16(1); - const auto input_vector = reinterpret_cast(input); - - #elif defined(USE_SSE2) - constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; - #ifndef USE_SSSE3 - const __m128i kZeros = _mm_setzero_si128(); - #else - const __m128i kOnes = _mm_set1_epi16(1); - #endif - const auto input_vector = reinterpret_cast(input); - - #elif defined(USE_MMX) - constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; - const __m64 kZeros = _mm_setzero_si64(); - const auto input_vector = reinterpret_cast(input); - - #elif defined(USE_NEON) - constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; - const auto input_vector = reinterpret_cast(input); - #endif - - for (IndexType i = 0; i < kOutputDimensions; ++i) { - const IndexType offset = i * kPaddedInputDimensions; - - #if defined(USE_AVX512) - __m512i sum = _mm512_setzero_si512(); - const auto row = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { - #if defined(USE_VNNI) - sum = _mm512_dpbusd_epi32(sum, _mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j])); - #else - __m512i product = _mm512_maddubs_epi16(_mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j])); - product = _mm512_madd_epi16(product, kOnes); - sum = _mm512_add_epi32(sum, product); - #endif + void propagate( + const InputType* input, OutputType* output) const { + +#if defined (USE_SSSE3) + + if constexpr (OutputDimensions > 1) + { + +#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_hadd Simd::m256_hadd +#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_hadd Simd::m128_hadd +#endif + + static constexpr IndexType OutputSimdWidth = sizeof(vec_t) / sizeof(OutputType); + + static_assert(OutputDimensions % OutputSimdWidth == 0); + + constexpr IndexType NumChunks = ceil_to_multiple(InputDimensions, 8) / 4; + constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth; + + const auto input32 = reinterpret_cast(input); + const vec_t* biasvec = reinterpret_cast(biases); + vec_t acc[NumRegs]; + for (IndexType k = 0; k < NumRegs; ++k) + acc[k] = biasvec[k]; + + for (IndexType i = 0; i < NumChunks; i += 2) + { + const vec_t in0 = vec_set_32(input32[i + 0]); + const vec_t in1 = vec_set_32(input32[i + 1]); + const auto col0 = reinterpret_cast(&weights[(i + 0) * OutputDimensions * 4]); + const auto col1 = reinterpret_cast(&weights[(i + 1) * OutputDimensions * 4]); + for (IndexType k = 0; k < NumRegs; ++k) + vec_add_dpbusd_32x2(acc[k], in0, col0[k], in1, col1[k]); } - // Note: Changing kMaxSimdWidth from 32 to 64 breaks loading existing networks. - // As a result kPaddedInputDimensions may not be an even multiple of 64(512bit) - // and we have to do one more 256bit chunk. - if (kPaddedInputDimensions != kNumChunks * kSimdWidth * 2) + vec_t* outptr = reinterpret_cast(output); + for (IndexType k = 0; k < NumRegs; ++k) + outptr[k] = acc[k]; + +# undef vec_setzero +# undef vec_set_32 +# undef vec_add_dpbusd_32 +# undef vec_add_dpbusd_32x2 +# undef vec_hadd + + } + else if constexpr (OutputDimensions == 1) + { + +// We cannot use AVX512 for the last layer because there's only 32 inputs and the buffer is not padded to 64 elements. +#if 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_hadd Simd::m256_hadd +#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_hadd Simd::m128_hadd +#endif + + const auto inputVector = reinterpret_cast(input); + + static constexpr IndexType InputSimdWidth = sizeof(vec_t) / sizeof(InputType); + + static_assert(PaddedInputDimensions % InputSimdWidth == 0); + + constexpr IndexType NumChunks = PaddedInputDimensions / InputSimdWidth; + vec_t sum0 = vec_setzero(); + const auto row0 = reinterpret_cast(&weights[0]); + + for (int j = 0; j < int(NumChunks); ++j) { - const auto iv256 = reinterpret_cast(&input_vector[kNumChunks]); - const auto row256 = reinterpret_cast(&row[kNumChunks]); - #if defined(USE_VNNI) - __m256i product256 = _mm256_dpbusd_epi32( - _mm512_castsi512_si256(sum), _mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0])); - sum = _mm512_inserti32x8(sum, product256, 0); - #else - __m256i product256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0])); - sum = _mm512_add_epi32(sum, _mm512_cvtepi16_epi32(product256)); - #endif + const vec_t in = inputVector[j]; + vec_add_dpbusd_32(sum0, in, row0[j]); } - output[i] = _mm512_reduce_add_epi32(sum) + biases_[i]; - - #elif defined(USE_AVX2) - __m256i sum = _mm256_setzero_si256(); - const auto row = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { - __m256i product = _mm256_maddubs_epi16(_mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j])); - product = _mm256_madd_epi16(product, kOnes); - sum = _mm256_add_epi32(sum, product); - } - __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1)); - sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC)); - sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB)); - output[i] = _mm_cvtsi128_si32(sum128) + biases_[i]; - - #elif defined(USE_SSSE3) - __m128i sum = _mm_setzero_si128(); - const auto row = reinterpret_cast(&weights_[offset]); - for (int j = 0; j < (int)kNumChunks - 1; j += 2) { - __m128i product0 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j]), _mm_load_si128(&row[j])); - product0 = _mm_madd_epi16(product0, kOnes); - sum = _mm_add_epi32(sum, product0); - __m128i product1 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j+1]), _mm_load_si128(&row[j+1])); - product1 = _mm_madd_epi16(product1, kOnes); - sum = _mm_add_epi32(sum, product1); - } - if (kNumChunks & 0x1) { - __m128i product = _mm_maddubs_epi16(_mm_load_si128(&input_vector[kNumChunks-1]), _mm_load_si128(&row[kNumChunks-1])); - product = _mm_madd_epi16(product, kOnes); - sum = _mm_add_epi32(sum, product); - } - sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC - sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB - output[i] = _mm_cvtsi128_si32(sum) + biases_[i]; - - #elif defined(USE_SSE2) - __m128i sum_lo = _mm_cvtsi32_si128(biases_[i]); - __m128i sum_hi = kZeros; - const auto row = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { - __m128i row_j = _mm_load_si128(&row[j]); - __m128i input_j = _mm_load_si128(&input_vector[j]); - __m128i row_signs = _mm_cmpgt_epi8(kZeros, row_j); - __m128i extended_row_lo = _mm_unpacklo_epi8(row_j, row_signs); - __m128i extended_row_hi = _mm_unpackhi_epi8(row_j, row_signs); - __m128i extended_input_lo = _mm_unpacklo_epi8(input_j, kZeros); - __m128i extended_input_hi = _mm_unpackhi_epi8(input_j, kZeros); - __m128i product_lo = _mm_madd_epi16(extended_row_lo, extended_input_lo); - __m128i product_hi = _mm_madd_epi16(extended_row_hi, extended_input_hi); - sum_lo = _mm_add_epi32(sum_lo, product_lo); - sum_hi = _mm_add_epi32(sum_hi, product_hi); - } - __m128i sum = _mm_add_epi32(sum_lo, sum_hi); - __m128i sum_high_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2)); - sum = _mm_add_epi32(sum, sum_high_64); - __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2)); - sum = _mm_add_epi32(sum, sum_second_32); - output[i] = _mm_cvtsi128_si32(sum); - - #elif defined(USE_MMX) - __m64 sum_lo = _mm_cvtsi32_si64(biases_[i]); - __m64 sum_hi = kZeros; - const auto row = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { - __m64 row_j = row[j]; - __m64 input_j = input_vector[j]; - __m64 row_signs = _mm_cmpgt_pi8(kZeros, row_j); - __m64 extended_row_lo = _mm_unpacklo_pi8(row_j, row_signs); - __m64 extended_row_hi = _mm_unpackhi_pi8(row_j, row_signs); - __m64 extended_input_lo = _mm_unpacklo_pi8(input_j, kZeros); - __m64 extended_input_hi = _mm_unpackhi_pi8(input_j, kZeros); - __m64 product_lo = _mm_madd_pi16(extended_row_lo, extended_input_lo); - __m64 product_hi = _mm_madd_pi16(extended_row_hi, extended_input_hi); - sum_lo = _mm_add_pi32(sum_lo, product_lo); - sum_hi = _mm_add_pi32(sum_hi, product_hi); - } - __m64 sum = _mm_add_pi32(sum_lo, sum_hi); - sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum)); - output[i] = _mm_cvtsi64_si32(sum); - - #elif defined(USE_NEON) - int32x4_t sum = {biases_[i]}; - const auto row = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { - int16x8_t product = vmull_s8(input_vector[j * 2], row[j * 2]); - product = vmlal_s8(product, input_vector[j * 2 + 1], row[j * 2 + 1]); - sum = vpadalq_s16(sum, product); - } - output[i] = sum[0] + sum[1] + sum[2] + sum[3]; + output[0] = vec_hadd(sum0, biases[0]); - #else - OutputType sum = biases_[i]; - for (IndexType j = 0; j < kInputDimensions; ++j) { - sum += weights_[offset + j] * input[j]; - } - output[i] = sum; - #endif +# undef vec_setzero +# undef vec_set_32 +# undef vec_add_dpbusd_32 +# undef vec_add_dpbusd_32x2 +# undef vec_hadd } - #if defined(USE_MMX) - _mm_empty(); - #endif - return output; +#else + // Use old implementation for the other architectures. + affine_transform_non_ssse3< + InputDimensions, + PaddedInputDimensions, + OutputDimensions>(output, weights, biases, input); +#endif } private: using BiasType = OutputType; using WeightType = std::int8_t; - PreviousLayer previous_layer_; - - alignas(kCacheLineSize) BiasType biases_[kOutputDimensions]; - alignas(kCacheLineSize) - WeightType weights_[kOutputDimensions * kPaddedInputDimensions]; + alignas(CacheLineSize) BiasType biases[OutputDimensions]; + alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions]; }; -} // namespace Eval::NNUE::Layers +} // namespace Stockfish::Eval::NNUE::Layers #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED