X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fnnue%2Flayers%2Fclipped_relu.h;h=a3a0c1ede9efdb892a18674dd86ca3460b7bc426;hp=13196ec28b49d133afeb0c0f704e644b86583b8d;hb=HEAD;hpb=875183b310a8249922c2155e82cb4cecfae2097e diff --git a/src/nnue/layers/clipped_relu.h b/src/nnue/layers/clipped_relu.h index 13196ec2..a3a0c1ed 100644 --- a/src/nnue/layers/clipped_relu.h +++ b/src/nnue/layers/clipped_relu.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 @@ -21,128 +21,148 @@ #ifndef NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED #define NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED +#include +#include +#include + #include "../nnue_common.h" -namespace Eval::NNUE::Layers { +namespace Stockfish::Eval::NNUE::Layers { - // Clipped ReLU - template - class ClippedReLU { +// Clipped ReLU +template +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::value, ""); // Number of input/output dimensions - static constexpr IndexType kInputDimensions = - PreviousLayer::kOutputDimensions; - static constexpr IndexType kOutputDimensions = kInputDimensions; - - // Size of forward propagation buffer used in this layer - static constexpr std::size_t kSelfBufferSize = - CeilToMultiple(kOutputDimensions * sizeof(OutputType), kCacheLineSize); + static constexpr IndexType InputDimensions = InDims; + static constexpr IndexType OutputDimensions = InputDimensions; + static constexpr IndexType PaddedOutputDimensions = + ceil_to_multiple(OutputDimensions, 32); - // 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 = 0x538D24C7u; - hash_value += PreviousLayer::GetHashValue(); - return hash_value; + static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) { + std::uint32_t hashValue = 0x538D24C7u; + hashValue += prevHash; + return hashValue; } // Read network parameters - bool ReadParameters(std::istream& stream) { - return previous_layer_.ReadParameters(stream); - } + bool read_parameters(std::istream&) { return true; } + + // Write network parameters + bool write_parameters(std::ostream&) const { return true; } // 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_AVX2) - constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth; - const __m256i kZero = _mm256_setzero_si256(); - const __m256i kOffsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0); - const auto in = reinterpret_cast(input); - const auto out = reinterpret_cast<__m256i*>(output); - for (IndexType i = 0; i < kNumChunks; ++i) { - const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32( - _mm256_loadA_si256(&in[i * 4 + 0]), - _mm256_loadA_si256(&in[i * 4 + 1])), kWeightScaleBits); - const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32( - _mm256_loadA_si256(&in[i * 4 + 2]), - _mm256_loadA_si256(&in[i * 4 + 3])), kWeightScaleBits); - _mm256_storeA_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8( - _mm256_packs_epi16(words0, words1), kZero), kOffsets)); - } - constexpr IndexType kStart = kNumChunks * kSimdWidth; - - #elif defined(USE_SSSE3) - constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth; - - #ifdef USE_SSE41 - const __m128i kZero = _mm_setzero_si128(); - #else - const __m128i k0x80s = _mm_set1_epi8(-128); - #endif - - const auto in = reinterpret_cast(input); - const auto out = reinterpret_cast<__m128i*>(output); - for (IndexType i = 0; i < kNumChunks; ++i) { - const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32( - _mm_load_si128(&in[i * 4 + 0]), - _mm_load_si128(&in[i * 4 + 1])), kWeightScaleBits); - const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32( - _mm_load_si128(&in[i * 4 + 2]), - _mm_load_si128(&in[i * 4 + 3])), kWeightScaleBits); - const __m128i packedbytes = _mm_packs_epi16(words0, words1); - _mm_store_si128(&out[i], - - #ifdef USE_SSE41 - _mm_max_epi8(packedbytes, kZero) - #else - _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s) - #endif - - ); - } - constexpr IndexType kStart = kNumChunks * kSimdWidth; - - #elif defined(USE_NEON) - constexpr IndexType kNumChunks = kInputDimensions / (kSimdWidth / 2); - const int8x8_t kZero = {0}; - const auto in = reinterpret_cast(input); - const auto out = reinterpret_cast(output); - for (IndexType i = 0; i < kNumChunks; ++i) { - int16x8_t shifted; - const auto pack = reinterpret_cast(&shifted); - pack[0] = vqshrn_n_s32(in[i * 2 + 0], kWeightScaleBits); - pack[1] = vqshrn_n_s32(in[i * 2 + 1], kWeightScaleBits); - out[i] = vmax_s8(vqmovn_s16(shifted), kZero); - } - constexpr IndexType kStart = kNumChunks * (kSimdWidth / 2); - #else - constexpr IndexType kStart = 0; - #endif - - for (IndexType i = kStart; i < kInputDimensions; ++i) { - output[i] = static_cast( - std::max(0, std::min(127, input[i] >> kWeightScaleBits))); - } - return output; + void propagate(const InputType* input, OutputType* output) const { + +#if defined(USE_AVX2) + if constexpr (InputDimensions % SimdWidth == 0) + { + constexpr IndexType NumChunks = InputDimensions / SimdWidth; + const __m256i Zero = _mm256_setzero_si256(); + const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0); + const auto in = reinterpret_cast(input); + const auto out = reinterpret_cast<__m256i*>(output); + for (IndexType i = 0; i < NumChunks; ++i) + { + const __m256i words0 = + _mm256_srai_epi16(_mm256_packs_epi32(_mm256_load_si256(&in[i * 4 + 0]), + _mm256_load_si256(&in[i * 4 + 1])), + WeightScaleBits); + const __m256i words1 = + _mm256_srai_epi16(_mm256_packs_epi32(_mm256_load_si256(&in[i * 4 + 2]), + _mm256_load_si256(&in[i * 4 + 3])), + WeightScaleBits); + _mm256_store_si256( + &out[i], _mm256_permutevar8x32_epi32( + _mm256_max_epi8(_mm256_packs_epi16(words0, words1), Zero), Offsets)); + } + } + else + { + constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2); + const __m128i Zero = _mm_setzero_si128(); + const auto in = reinterpret_cast(input); + const auto out = reinterpret_cast<__m128i*>(output); + for (IndexType i = 0; i < NumChunks; ++i) + { + const __m128i words0 = _mm_srai_epi16( + _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])), + WeightScaleBits); + const __m128i words1 = _mm_srai_epi16( + _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])), + WeightScaleBits); + const __m128i packedbytes = _mm_packs_epi16(words0, words1); + _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero)); + } + } + constexpr IndexType Start = InputDimensions % SimdWidth == 0 + ? InputDimensions / SimdWidth * SimdWidth + : InputDimensions / (SimdWidth / 2) * (SimdWidth / 2); + +#elif defined(USE_SSE2) + constexpr IndexType NumChunks = InputDimensions / SimdWidth; + + #ifdef USE_SSE41 + const __m128i Zero = _mm_setzero_si128(); + #else + const __m128i k0x80s = _mm_set1_epi8(-128); + #endif + + const auto in = reinterpret_cast(input); + const auto out = reinterpret_cast<__m128i*>(output); + for (IndexType i = 0; i < NumChunks; ++i) + { + const __m128i words0 = _mm_srai_epi16( + _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])), + WeightScaleBits); + const __m128i words1 = _mm_srai_epi16( + _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])), + WeightScaleBits); + const __m128i packedbytes = _mm_packs_epi16(words0, words1); + _mm_store_si128(&out[i], + + #ifdef USE_SSE41 + _mm_max_epi8(packedbytes, Zero) + #else + _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s) + #endif + + ); + } + constexpr IndexType Start = NumChunks * SimdWidth; + +#elif defined(USE_NEON) + constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2); + const int8x8_t Zero = {0}; + const auto in = reinterpret_cast(input); + const auto out = reinterpret_cast(output); + for (IndexType i = 0; i < NumChunks; ++i) + { + int16x8_t shifted; + const auto pack = reinterpret_cast(&shifted); + pack[0] = vqshrn_n_s32(in[i * 2 + 0], WeightScaleBits); + pack[1] = vqshrn_n_s32(in[i * 2 + 1], WeightScaleBits); + out[i] = vmax_s8(vqmovn_s16(shifted), Zero); + } + constexpr IndexType Start = NumChunks * (SimdWidth / 2); +#else + constexpr IndexType Start = 0; +#endif + + for (IndexType i = Start; i < InputDimensions; ++i) + { + output[i] = static_cast(std::clamp(input[i] >> WeightScaleBits, 0, 127)); + } } +}; - private: - PreviousLayer previous_layer_; - }; - -} // namespace Eval::NNUE::Layers +} // namespace Stockfish::Eval::NNUE::Layers -#endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED +#endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED