X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fnnue%2Flayers%2Faffine_transform.h;h=adf152eea5b8894fcdf26cdfebffcdbd602b5d7f;hp=985ee71a4193e571f9ecdddfc144ca4c2c571aea;hb=303713b560e356a902c1830bce205716cef54a44;hpb=21df37d7fd4dcc9b4a9c319382cc43685c0259c8 diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h index 985ee71a..adf152ee 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) 2004-2021 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 @@ -41,6 +41,11 @@ namespace Eval::NNUE::Layers { static constexpr IndexType kOutputDimensions = OutputDimensions; static constexpr IndexType kPaddedInputDimensions = CeilToMultiple(kInputDimensions, kMaxSimdWidth); +#if defined (USE_AVX512) + static constexpr const IndexType kOutputSimdWidth = kSimdWidth / 2; +#elif defined (USE_SSSE3) + static constexpr const IndexType kOutputSimdWidth = kSimdWidth / 4; +#endif // Size of forward propagation buffer used in this layer static constexpr std::size_t kSelfBufferSize = @@ -62,11 +67,61 @@ namespace Eval::NNUE::Layers { // 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)); + for (std::size_t i = 0; i < kOutputDimensions; ++i) + biases_[i] = read_little_endian(stream); + for (std::size_t i = 0; i < kOutputDimensions * kPaddedInputDimensions; ++i) +#if !defined (USE_SSSE3) + weights_[i] = read_little_endian(stream); +#else + weights_[ + (i / 4) % (kPaddedInputDimensions / 4) * kOutputDimensions * 4 + + i / kPaddedInputDimensions * 4 + + i % 4 + ] = read_little_endian(stream); + + // Determine if eights of weight and input products can be summed using 16bits + // without saturation. We assume worst case combinations of 0 and 127 for all inputs. + if (kOutputDimensions > 1 && !stream.fail()) + { + canSaturate16.count = 0; +#if !defined(USE_VNNI) + for (IndexType i = 0; i < kPaddedInputDimensions; i += 16) + for (IndexType j = 0; j < kOutputDimensions; ++j) + for (int x = 0; x < 2; ++x) + { + WeightType* w = &weights_[i * kOutputDimensions + j * 4 + x * 2]; + int sum[2] = {0, 0}; + for (int k = 0; k < 8; ++k) + { + IndexType idx = k / 2 * kOutputDimensions * 4 + k % 2; + sum[w[idx] < 0] += w[idx]; + } + for (int sign : {-1, 1}) + while (sign * sum[sign == -1] > 258) + { + int maxK = 0, maxW = 0; + for (int k = 0; k < 8; ++k) + { + IndexType idx = k / 2 * kOutputDimensions * 4 + k % 2; + if (maxW < sign * w[idx]) + maxK = k, maxW = sign * w[idx]; + } + + IndexType idx = maxK / 2 * kOutputDimensions * 4 + maxK % 2; + sum[sign == -1] -= w[idx]; + canSaturate16.add(j, i + maxK / 2 * 4 + maxK % 2 + x * 2, w[idx]); + w[idx] = 0; + } + } + + // Non functional optimization for faster more linear access + std::sort(canSaturate16.ids, canSaturate16.ids + canSaturate16.count, + [](const typename CanSaturate::Entry& e1, const typename CanSaturate::Entry& e2) + { return e1.in == e2.in ? e1.out < e2.out : e1.in < e2.in; }); +#endif + } +#endif + return !stream.fail(); } @@ -75,105 +130,246 @@ namespace Eval::NNUE::Layers { const TransformedFeatureType* transformed_features, char* buffer) const { const auto input = previous_layer_.Propagate( transformed_features, buffer + kSelfBufferSize); + +#if defined (USE_AVX512) + + [[maybe_unused]] const __m512i kOnes512 = _mm512_set1_epi16(1); + + [[maybe_unused]] auto m512_hadd = [](__m512i sum, int bias) -> int { + return _mm512_reduce_add_epi32(sum) + bias; + }; + + [[maybe_unused]] auto m512_add_dpbusd_epi32 = [=](__m512i& acc, __m512i a, __m512i b) { +#if defined (USE_VNNI) + acc = _mm512_dpbusd_epi32(acc, a, b); +#else + __m512i product0 = _mm512_maddubs_epi16(a, b); + product0 = _mm512_madd_epi16(product0, kOnes512); + acc = _mm512_add_epi32(acc, product0); +#endif + }; + + [[maybe_unused]] auto m512_add_dpbusd_epi32x4 = [=](__m512i& acc, __m512i a0, __m512i b0, __m512i a1, __m512i b1, + __m512i a2, __m512i b2, __m512i a3, __m512i b3) { +#if defined (USE_VNNI) + acc = _mm512_dpbusd_epi32(acc, a0, b0); + acc = _mm512_dpbusd_epi32(acc, a1, b1); + acc = _mm512_dpbusd_epi32(acc, a2, b2); + acc = _mm512_dpbusd_epi32(acc, a3, b3); +#else + __m512i product0 = _mm512_maddubs_epi16(a0, b0); + __m512i product1 = _mm512_maddubs_epi16(a1, b1); + __m512i product2 = _mm512_maddubs_epi16(a2, b2); + __m512i product3 = _mm512_maddubs_epi16(a3, b3); + product0 = _mm512_add_epi16(product0, product1); + product2 = _mm512_add_epi16(product2, product3); + product0 = _mm512_add_epi16(product0, product2); + product0 = _mm512_madd_epi16(product0, kOnes512); + acc = _mm512_add_epi32(acc, product0); +#endif + }; + +#endif +#if defined (USE_AVX2) + + [[maybe_unused]] const __m256i kOnes256 = _mm256_set1_epi16(1); + + [[maybe_unused]] auto m256_hadd = [](__m256i sum, int bias) -> int { + __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)); + return _mm_cvtsi128_si32(sum128) + bias; + }; + + [[maybe_unused]] auto m256_add_dpbusd_epi32 = [=](__m256i& acc, __m256i a, __m256i b) { +#if defined (USE_VNNI) + acc = _mm256_dpbusd_epi32(acc, a, b); +#else + __m256i product0 = _mm256_maddubs_epi16(a, b); + product0 = _mm256_madd_epi16(product0, kOnes256); + acc = _mm256_add_epi32(acc, product0); +#endif + }; + + [[maybe_unused]] auto m256_add_dpbusd_epi32x4 = [=](__m256i& acc, __m256i a0, __m256i b0, __m256i a1, __m256i b1, + __m256i a2, __m256i b2, __m256i a3, __m256i b3) { +#if defined (USE_VNNI) + acc = _mm256_dpbusd_epi32(acc, a0, b0); + acc = _mm256_dpbusd_epi32(acc, a1, b1); + acc = _mm256_dpbusd_epi32(acc, a2, b2); + acc = _mm256_dpbusd_epi32(acc, a3, b3); +#else + __m256i product0 = _mm256_maddubs_epi16(a0, b0); + __m256i product1 = _mm256_maddubs_epi16(a1, b1); + __m256i product2 = _mm256_maddubs_epi16(a2, b2); + __m256i product3 = _mm256_maddubs_epi16(a3, b3); + product0 = _mm256_add_epi16(product0, product1); + product2 = _mm256_add_epi16(product2, product3); + product0 = _mm256_add_epi16(product0, product2); + product0 = _mm256_madd_epi16(product0, kOnes256); + acc = _mm256_add_epi32(acc, product0); +#endif + }; + +#endif +#if defined (USE_SSSE3) + + [[maybe_unused]] const __m128i kOnes128 = _mm_set1_epi16(1); + + [[maybe_unused]] auto m128_hadd = [](__m128i sum, int bias) -> int { + 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 + return _mm_cvtsi128_si32(sum) + bias; + }; + + [[maybe_unused]] auto m128_add_dpbusd_epi32 = [=](__m128i& acc, __m128i a, __m128i b) { + __m128i product0 = _mm_maddubs_epi16(a, b); + product0 = _mm_madd_epi16(product0, kOnes128); + acc = _mm_add_epi32(acc, product0); + }; + + [[maybe_unused]] auto m128_add_dpbusd_epi32x4 = [=](__m128i& acc, __m128i a0, __m128i b0, __m128i a1, __m128i b1, + __m128i a2, __m128i b2, __m128i a3, __m128i b3) { + __m128i product0 = _mm_maddubs_epi16(a0, b0); + __m128i product1 = _mm_maddubs_epi16(a1, b1); + __m128i product2 = _mm_maddubs_epi16(a2, b2); + __m128i product3 = _mm_maddubs_epi16(a3, b3); + product0 = _mm_adds_epi16(product0, product1); + product2 = _mm_adds_epi16(product2, product3); + product0 = _mm_adds_epi16(product0, product2); + product0 = _mm_madd_epi16(product0, kOnes128); + acc = _mm_add_epi32(acc, product0); + }; + +#endif + +#if defined (USE_AVX512) + using vec_t = __m512i; + #define vec_setzero _mm512_setzero_si512 + #define vec_set_32 _mm512_set1_epi32 + auto& vec_add_dpbusd_32 = m512_add_dpbusd_epi32; + auto& vec_add_dpbusd_32x4 = m512_add_dpbusd_epi32x4; + auto& vec_hadd = m512_hadd; +#elif defined (USE_AVX2) + using vec_t = __m256i; + #define vec_setzero _mm256_setzero_si256 + #define vec_set_32 _mm256_set1_epi32 + auto& vec_add_dpbusd_32 = m256_add_dpbusd_epi32; + auto& vec_add_dpbusd_32x4 = m256_add_dpbusd_epi32x4; + auto& vec_hadd = m256_hadd; +#elif defined (USE_SSSE3) + using vec_t = __m128i; + #define vec_setzero _mm_setzero_si128 + #define vec_set_32 _mm_set1_epi32 + auto& vec_add_dpbusd_32 = m128_add_dpbusd_epi32; + auto& vec_add_dpbusd_32x4 = m128_add_dpbusd_epi32x4; + auto& vec_hadd = m128_hadd; +#endif + +#if defined (USE_SSSE3) + const auto output = reinterpret_cast(buffer); + const auto input_vector = reinterpret_cast(input); + + static_assert(kOutputDimensions % kOutputSimdWidth == 0 || kOutputDimensions == 1); + + // kOutputDimensions is either 1 or a multiple of kSimdWidth + // because then it is also an input dimension. + if constexpr (kOutputDimensions % kOutputSimdWidth == 0) + { + constexpr IndexType kNumChunks = kPaddedInputDimensions / 4; + + const auto input32 = reinterpret_cast(input); + vec_t* outptr = reinterpret_cast(output); + std::memcpy(output, biases_, kOutputDimensions * sizeof(OutputType)); + + for (int i = 0; i < (int)kNumChunks - 3; i += 4) + { + const vec_t in0 = vec_set_32(input32[i + 0]); + const vec_t in1 = vec_set_32(input32[i + 1]); + const vec_t in2 = vec_set_32(input32[i + 2]); + const vec_t in3 = vec_set_32(input32[i + 3]); + const auto col0 = reinterpret_cast(&weights_[(i + 0) * kOutputDimensions * 4]); + const auto col1 = reinterpret_cast(&weights_[(i + 1) * kOutputDimensions * 4]); + const auto col2 = reinterpret_cast(&weights_[(i + 2) * kOutputDimensions * 4]); + const auto col3 = reinterpret_cast(&weights_[(i + 3) * kOutputDimensions * 4]); + for (int j = 0; j * kOutputSimdWidth < kOutputDimensions; ++j) + vec_add_dpbusd_32x4(outptr[j], in0, col0[j], in1, col1[j], in2, col2[j], in3, col3[j]); + } + for (int i = 0; i < canSaturate16.count; ++i) + output[canSaturate16.ids[i].out] += input[canSaturate16.ids[i].in] * canSaturate16.ids[i].w; + } + else if constexpr (kOutputDimensions == 1) + { +#if defined (USE_AVX512) + if constexpr (kPaddedInputDimensions % (kSimdWidth * 2) != 0) + { + constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; + const auto input_vector256 = reinterpret_cast(input); + + __m256i sum0 = _mm256_setzero_si256(); + const auto row0 = reinterpret_cast(&weights_[0]); + + for (int j = 0; j < (int)kNumChunks; ++j) + { + const __m256i in = input_vector256[j]; + m256_add_dpbusd_epi32(sum0, in, row0[j]); + } + output[0] = m256_hadd(sum0, biases_[0]); + } + else +#endif + { +#if defined (USE_AVX512) + constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2); +#else + constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; +#endif + vec_t sum0 = vec_setzero(); + const auto row0 = reinterpret_cast(&weights_[0]); + + for (int j = 0; j < (int)kNumChunks; ++j) + { + const vec_t in = input_vector[j]; + vec_add_dpbusd_32(sum0, in, row0[j]); + } + output[0] = vec_hadd(sum0, biases_[0]); + } + } - #if defined(USE_AVX512) - constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2); - const __m512i kOnes = _mm512_set1_epi16(1); - const auto input_vector = reinterpret_cast(input); +#else - #elif defined(USE_AVX2) - constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; - const __m256i kOnes = _mm256_set1_epi16(1); - const auto input_vector = reinterpret_cast(input); +// Use old implementation for the other architectures. + + auto output = reinterpret_cast(buffer); - #elif defined(USE_SSE2) +#if 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) +#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) +#elif defined(USE_NEON) constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; const auto input_vector = reinterpret_cast(input); - #endif +#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) { - __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); - } - - // 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) - { - const auto iv256 = reinterpret_cast(&input_vector[kNumChunks]); - const auto row256 = reinterpret_cast(&row[kNumChunks]); - __m256i product256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0])); - product256 = _mm256_madd_epi16(product256, _mm256_set1_epi16(1)); - sum = _mm512_add_epi32(sum, _mm512_zextsi256_si512(product256)); - } - 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) +#if 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_row_lo = _mm_srai_epi16(_mm_unpacklo_epi8(row_j, row_j), 8); + __m128i extended_row_hi = _mm_srai_epi16(_mm_unpackhi_epi8(row_j, row_j), 8); __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); @@ -188,16 +384,15 @@ namespace Eval::NNUE::Layers { sum = _mm_add_epi32(sum, sum_second_32); output[i] = _mm_cvtsi128_si32(sum); - #elif defined(USE_MMX) +#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_row_lo = _mm_srai_pi16(_mm_unpacklo_pi8(row_j, row_j), 8); + __m64 extended_row_hi = _mm_srai_pi16(_mm_unpackhi_pi8(row_j, row_j), 8); __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); @@ -209,7 +404,7 @@ namespace Eval::NNUE::Layers { sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum)); output[i] = _mm_cvtsi64_si32(sum); - #elif defined(USE_NEON) +#elif defined(USE_NEON) int32x4_t sum = {biases_[i]}; const auto row = reinterpret_cast(&weights_[offset]); for (IndexType j = 0; j < kNumChunks; ++j) { @@ -219,18 +414,21 @@ namespace Eval::NNUE::Layers { } output[i] = sum[0] + sum[1] + sum[2] + sum[3]; - #else +#else OutputType sum = biases_[i]; for (IndexType j = 0; j < kInputDimensions; ++j) { sum += weights_[offset + j] * input[j]; } output[i] = sum; - #endif +#endif } - #if defined(USE_MMX) +#if defined(USE_MMX) _mm_empty(); - #endif +#endif + +#endif + return output; } @@ -241,8 +439,24 @@ namespace Eval::NNUE::Layers { PreviousLayer previous_layer_; alignas(kCacheLineSize) BiasType biases_[kOutputDimensions]; - alignas(kCacheLineSize) - WeightType weights_[kOutputDimensions * kPaddedInputDimensions]; + alignas(kCacheLineSize) WeightType weights_[kOutputDimensions * kPaddedInputDimensions]; +#if defined (USE_SSSE3) + struct CanSaturate { + int count; + struct Entry { + uint16_t out; + uint16_t in; + int8_t w; + } ids[kPaddedInputDimensions * kOutputDimensions * 3 / 4]; + + void add(int i, int j, int8_t w) { + ids[count].out = i; + ids[count].in = j; + ids[count].w = w; + ++count; + } + } canSaturate16; +#endif }; } // namespace Eval::NNUE::Layers