X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fnnue%2Fnnue_feature_transformer.h;h=2b6259c328111e0e143d61d534eb80ab934b92a0;hp=1cfebbe4cbe80425f65aa3e3012594494d615294;hb=fc27d158c012341593518a05abf51903ecbcb495;hpb=651ec3b31ee68db50f38ccd8fcdedbd6673cd9ed diff --git a/src/nnue/nnue_feature_transformer.h b/src/nnue/nnue_feature_transformer.h index 1cfebbe4..2b6259c3 100644 --- a/src/nnue/nnue_feature_transformer.h +++ b/src/nnue/nnue_feature_transformer.h @@ -50,37 +50,42 @@ namespace Eval::NNUE { // Hash value embedded in the evaluation file static constexpr std::uint32_t GetHashValue() { + return RawFeatures::kHashValue ^ kOutputDimensions; } // Read network parameters bool ReadParameters(std::istream& stream) { - stream.read(reinterpret_cast(biases_), - kHalfDimensions * sizeof(BiasType)); - stream.read(reinterpret_cast(weights_), - kHalfDimensions * kInputDimensions * sizeof(WeightType)); + + for (std::size_t i = 0; i < kHalfDimensions; ++i) + biases_[i] = read_little_endian(stream); + for (std::size_t i = 0; i < kHalfDimensions * kInputDimensions; ++i) + weights_[i] = read_little_endian(stream); return !stream.fail(); } // Proceed with the difference calculation if possible bool UpdateAccumulatorIfPossible(const Position& pos) const { + const auto now = pos.state(); - if (now->accumulator.computed_accumulation) { + if (now->accumulator.computed_accumulation) return true; - } + const auto prev = now->previous; if (prev && prev->accumulator.computed_accumulation) { UpdateAccumulator(pos); return true; } + return false; } // Convert input features - void Transform(const Position& pos, OutputType* output, bool refresh) const { - if (refresh || !UpdateAccumulatorIfPossible(pos)) { + void Transform(const Position& pos, OutputType* output) const { + + if (!UpdateAccumulatorIfPossible(pos)) RefreshAccumulator(pos); - } + const auto& accumulation = pos.state()->accumulator.accumulation; #if defined(USE_AVX2) @@ -88,7 +93,7 @@ namespace Eval::NNUE { constexpr int kControl = 0b11011000; const __m256i kZero = _mm256_setzero_si256(); - #elif defined(USE_SSSE3) + #elif defined(USE_SSE2) constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth; #ifdef USE_SSE41 @@ -97,6 +102,10 @@ namespace Eval::NNUE { const __m128i k0x80s = _mm_set1_epi8(-128); #endif + #elif defined(USE_MMX) + constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth; + const __m64 k0x80s = _mm_set1_pi8(-128); + #elif defined(USE_NEON) constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2); const int8x8_t kZero = {0}; @@ -109,41 +118,15 @@ namespace Eval::NNUE { #if defined(USE_AVX2) auto out = reinterpret_cast<__m256i*>(&output[offset]); for (IndexType j = 0; j < kNumChunks; ++j) { - __m256i sum0 = - - #if defined(__MINGW32__) || defined(__MINGW64__) - // HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary - // compiled with g++ in MSYS2 crashes here because the output memory is not aligned - // even though alignas is specified. - _mm256_loadu_si256 - #else - _mm256_load_si256 - #endif - - (&reinterpret_cast( - accumulation[perspectives[p]][0])[j * 2 + 0]); - __m256i sum1 = - - #if defined(__MINGW32__) || defined(__MINGW64__) - _mm256_loadu_si256 - #else - _mm256_load_si256 - #endif - - (&reinterpret_cast( - accumulation[perspectives[p]][0])[j * 2 + 1]); - - #if defined(__MINGW32__) || defined(__MINGW64__) - _mm256_storeu_si256 - #else - _mm256_store_si256 - #endif - - (&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8( + __m256i sum0 = _mm256_loadA_si256( + &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 0]); + __m256i sum1 = _mm256_loadA_si256( + &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 1]); + _mm256_storeA_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8( _mm256_packs_epi16(sum0, sum1), kZero), kControl)); } - #elif defined(USE_SSSE3) + #elif defined(USE_SSE2) auto out = reinterpret_cast<__m128i*>(&output[offset]); for (IndexType j = 0; j < kNumChunks; ++j) { __m128i sum0 = _mm_load_si128(&reinterpret_cast( @@ -163,6 +146,17 @@ namespace Eval::NNUE { ); } + #elif defined(USE_MMX) + auto out = reinterpret_cast<__m64*>(&output[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) { + __m64 sum0 = *(&reinterpret_cast( + accumulation[perspectives[p]][0])[j * 2 + 0]); + __m64 sum1 = *(&reinterpret_cast( + accumulation[perspectives[p]][0])[j * 2 + 1]); + const __m64 packedbytes = _mm_packs_pi16(sum0, sum1); + out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s); + } + #elif defined(USE_NEON) const auto out = reinterpret_cast(&output[offset]); for (IndexType j = 0; j < kNumChunks; ++j) { @@ -180,11 +174,15 @@ namespace Eval::NNUE { #endif } + #if defined(USE_MMX) + _mm_empty(); + #endif } private: // Calculate cumulative value without using difference calculation void RefreshAccumulator(const Position& pos) const { + auto& accumulator = pos.state()->accumulator; IndexType i = 0; Features::IndexList active_indices[2]; @@ -195,53 +193,63 @@ namespace Eval::NNUE { kHalfDimensions * sizeof(BiasType)); for (const auto index : active_indices[perspective]) { const IndexType offset = kHalfDimensions * index; + #if defined(USE_AVX512) + auto accumulation = reinterpret_cast<__m512i*>( + &accumulator.accumulation[perspective][i][0]); + auto column = reinterpret_cast(&weights_[offset]); + constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth; + for (IndexType j = 0; j < kNumChunks; ++j) + _mm512_storeA_si512(&accumulation[j], _mm512_add_epi16(_mm512_loadA_si512(&accumulation[j]), column[j])); - #if defined(USE_AVX2) + #elif defined(USE_AVX2) auto accumulation = reinterpret_cast<__m256i*>( &accumulator.accumulation[perspective][i][0]); auto column = reinterpret_cast(&weights_[offset]); constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2); - for (IndexType j = 0; j < kNumChunks; ++j) { - #if defined(__MINGW32__) || defined(__MINGW64__) - _mm256_storeu_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadu_si256(&accumulation[j]), column[j])); - #else - accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]); - #endif - } + for (IndexType j = 0; j < kNumChunks; ++j) + _mm256_storeA_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadA_si256(&accumulation[j]), column[j])); #elif defined(USE_SSE2) auto accumulation = reinterpret_cast<__m128i*>( &accumulator.accumulation[perspective][i][0]); auto column = reinterpret_cast(&weights_[offset]); constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < kNumChunks; ++j) accumulation[j] = _mm_add_epi16(accumulation[j], column[j]); - } + + #elif defined(USE_MMX) + auto accumulation = reinterpret_cast<__m64*>( + &accumulator.accumulation[perspective][i][0]); + auto column = reinterpret_cast(&weights_[offset]); + constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2); + for (IndexType j = 0; j < kNumChunks; ++j) + accumulation[j] = _mm_add_pi16(accumulation[j], column[j]); #elif defined(USE_NEON) auto accumulation = reinterpret_cast( &accumulator.accumulation[perspective][i][0]); auto column = reinterpret_cast(&weights_[offset]); constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < kNumChunks; ++j) accumulation[j] = vaddq_s16(accumulation[j], column[j]); - } #else - for (IndexType j = 0; j < kHalfDimensions; ++j) { + for (IndexType j = 0; j < kHalfDimensions; ++j) accumulator.accumulation[perspective][i][j] += weights_[offset + j]; - } #endif } } + #if defined(USE_MMX) + _mm_empty(); + #endif accumulator.computed_accumulation = true; - accumulator.computed_score = false; } // Calculate cumulative value using difference calculation void UpdateAccumulator(const Position& pos) const { + const auto prev_accumulator = pos.state()->previous->accumulator; auto& accumulator = pos.state()->accumulator; IndexType i = 0; @@ -261,6 +269,11 @@ namespace Eval::NNUE { auto accumulation = reinterpret_cast<__m128i*>( &accumulator.accumulation[perspective][i][0]); + #elif defined(USE_MMX) + constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2); + auto accumulation = reinterpret_cast<__m64*>( + &accumulator.accumulation[perspective][i][0]); + #elif defined(USE_NEON) constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2); auto accumulation = reinterpret_cast( @@ -280,27 +293,27 @@ namespace Eval::NNUE { #if defined(USE_AVX2) auto column = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < kNumChunks; ++j) accumulation[j] = _mm256_sub_epi16(accumulation[j], column[j]); - } #elif defined(USE_SSE2) auto column = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < kNumChunks; ++j) accumulation[j] = _mm_sub_epi16(accumulation[j], column[j]); - } + + #elif defined(USE_MMX) + auto column = reinterpret_cast(&weights_[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) + accumulation[j] = _mm_sub_pi16(accumulation[j], column[j]); #elif defined(USE_NEON) auto column = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < kNumChunks; ++j) accumulation[j] = vsubq_s16(accumulation[j], column[j]); - } #else - for (IndexType j = 0; j < kHalfDimensions; ++j) { - accumulator.accumulation[perspective][i][j] -= - weights_[offset + j]; - } + for (IndexType j = 0; j < kHalfDimensions; ++j) + accumulator.accumulation[perspective][i][j] -= weights_[offset + j]; #endif } @@ -311,35 +324,37 @@ namespace Eval::NNUE { #if defined(USE_AVX2) auto column = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < kNumChunks; ++j) accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]); - } #elif defined(USE_SSE2) auto column = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < kNumChunks; ++j) accumulation[j] = _mm_add_epi16(accumulation[j], column[j]); - } + + #elif defined(USE_MMX) + auto column = reinterpret_cast(&weights_[offset]); + for (IndexType j = 0; j < kNumChunks; ++j) + accumulation[j] = _mm_add_pi16(accumulation[j], column[j]); #elif defined(USE_NEON) auto column = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < kNumChunks; ++j) accumulation[j] = vaddq_s16(accumulation[j], column[j]); - } #else - for (IndexType j = 0; j < kHalfDimensions; ++j) { - accumulator.accumulation[perspective][i][j] += - weights_[offset + j]; - } + for (IndexType j = 0; j < kHalfDimensions; ++j) + accumulator.accumulation[perspective][i][j] += weights_[offset + j]; #endif } } } + #if defined(USE_MMX) + _mm_empty(); + #endif accumulator.computed_accumulation = true; - accumulator.computed_score = false; } using BiasType = std::int16_t;