X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fnnue%2Fnnue_feature_transformer.h;fp=src%2Fnnue%2Fnnue_feature_transformer.h;h=de4b49374f91768dce1d04f1dc2dd2615b69bf3f;hp=1e0b0e6da55112dbcddf78bddeb31668564f5e66;hb=fbbd4adc3c01460faa3cc8f91771ab9b0ef718ca;hpb=a7ab92ec25c91e8413630c52cfc2db6b4ecacf0c diff --git a/src/nnue/nnue_feature_transformer.h b/src/nnue/nnue_feature_transformer.h index 1e0b0e6d..de4b4937 100644 --- a/src/nnue/nnue_feature_transformer.h +++ b/src/nnue/nnue_feature_transformer.h @@ -40,7 +40,7 @@ namespace Stockfish::Eval::NNUE { #define vec_store(a,b) _mm512_store_si512(a,b) #define vec_add_16(a,b) _mm512_add_epi16(a,b) #define vec_sub_16(a,b) _mm512_sub_epi16(a,b) - static constexpr IndexType kNumRegs = 8; // only 8 are needed + static constexpr IndexType NumRegs = 8; // only 8 are needed #elif USE_AVX2 typedef __m256i vec_t; @@ -48,7 +48,7 @@ namespace Stockfish::Eval::NNUE { #define vec_store(a,b) _mm256_store_si256(a,b) #define vec_add_16(a,b) _mm256_add_epi16(a,b) #define vec_sub_16(a,b) _mm256_sub_epi16(a,b) - static constexpr IndexType kNumRegs = 16; + static constexpr IndexType NumRegs = 16; #elif USE_SSE2 typedef __m128i vec_t; @@ -56,7 +56,7 @@ namespace Stockfish::Eval::NNUE { #define vec_store(a,b) *(a)=(b) #define vec_add_16(a,b) _mm_add_epi16(a,b) #define vec_sub_16(a,b) _mm_sub_epi16(a,b) - static constexpr IndexType kNumRegs = Is64Bit ? 16 : 8; + static constexpr IndexType NumRegs = Is64Bit ? 16 : 8; #elif USE_MMX typedef __m64 vec_t; @@ -64,7 +64,7 @@ namespace Stockfish::Eval::NNUE { #define vec_store(a,b) *(a)=(b) #define vec_add_16(a,b) _mm_add_pi16(a,b) #define vec_sub_16(a,b) _mm_sub_pi16(a,b) - static constexpr IndexType kNumRegs = 8; + static constexpr IndexType NumRegs = 8; #elif USE_NEON typedef int16x8_t vec_t; @@ -72,7 +72,7 @@ namespace Stockfish::Eval::NNUE { #define vec_store(a,b) *(a)=(b) #define vec_add_16(a,b) vaddq_s16(a,b) #define vec_sub_16(a,b) vsubq_s16(a,b) - static constexpr IndexType kNumRegs = 16; + static constexpr IndexType NumRegs = 16; #else #undef VECTOR @@ -84,11 +84,11 @@ namespace Stockfish::Eval::NNUE { private: // Number of output dimensions for one side - static constexpr IndexType kHalfDimensions = kTransformedFeatureDimensions; + static constexpr IndexType HalfDimensions = TransformedFeatureDimensions; #ifdef VECTOR - static constexpr IndexType kTileHeight = kNumRegs * sizeof(vec_t) / 2; - static_assert(kHalfDimensions % kTileHeight == 0, "kTileHeight must divide kHalfDimensions"); + static constexpr IndexType TileHeight = NumRegs * sizeof(vec_t) / 2; + static_assert(HalfDimensions % TileHeight == 0, "TileHeight must divide HalfDimensions"); #endif public: @@ -96,95 +96,92 @@ namespace Stockfish::Eval::NNUE { using OutputType = TransformedFeatureType; // Number of input/output dimensions - static constexpr IndexType kInputDimensions = RawFeatures::kDimensions; - static constexpr IndexType kOutputDimensions = kHalfDimensions * 2; + static constexpr IndexType InputDimensions = RawFeatures::Dimensions; + static constexpr IndexType OutputDimensions = HalfDimensions * 2; // Size of forward propagation buffer - static constexpr std::size_t kBufferSize = - kOutputDimensions * sizeof(OutputType); + static constexpr std::size_t BufferSize = + OutputDimensions * sizeof(OutputType); // Hash value embedded in the evaluation file - static constexpr std::uint32_t GetHashValue() { - - return RawFeatures::kHashValue ^ kOutputDimensions; + static constexpr std::uint32_t get_hash_value() { + return RawFeatures::HashValue ^ OutputDimensions; } // Read network parameters - bool ReadParameters(std::istream& stream) { - - 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); + bool read_parameters(std::istream& stream) { + for (std::size_t i = 0; i < HalfDimensions; ++i) + biases[i] = read_little_endian(stream); + for (std::size_t i = 0; i < HalfDimensions * InputDimensions; ++i) + weights[i] = read_little_endian(stream); return !stream.fail(); } // Convert input features - void Transform(const Position& pos, OutputType* output) const { - - UpdateAccumulator(pos, WHITE); - UpdateAccumulator(pos, BLACK); + void transform(const Position& pos, OutputType* output) const { + update_accumulator(pos, WHITE); + update_accumulator(pos, BLACK); const auto& accumulation = pos.state()->accumulator.accumulation; #if defined(USE_AVX512) - constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth * 2); - static_assert(kHalfDimensions % (kSimdWidth * 2) == 0); - const __m512i kControl = _mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7); - const __m512i kZero = _mm512_setzero_si512(); + constexpr IndexType NumChunks = HalfDimensions / (SimdWidth * 2); + static_assert(HalfDimensions % (SimdWidth * 2) == 0); + const __m512i Control = _mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7); + const __m512i Zero = _mm512_setzero_si512(); #elif defined(USE_AVX2) - constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth; - constexpr int kControl = 0b11011000; - const __m256i kZero = _mm256_setzero_si256(); + constexpr IndexType NumChunks = HalfDimensions / SimdWidth; + constexpr int Control = 0b11011000; + const __m256i Zero = _mm256_setzero_si256(); #elif defined(USE_SSE2) - constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth; + constexpr IndexType NumChunks = HalfDimensions / SimdWidth; #ifdef USE_SSE41 - const __m128i kZero = _mm_setzero_si128(); + const __m128i Zero = _mm_setzero_si128(); #else const __m128i k0x80s = _mm_set1_epi8(-128); #endif #elif defined(USE_MMX) - constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth; + constexpr IndexType NumChunks = HalfDimensions / SimdWidth; const __m64 k0x80s = _mm_set1_pi8(-128); #elif defined(USE_NEON) - constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2); - const int8x8_t kZero = {0}; + constexpr IndexType NumChunks = HalfDimensions / (SimdWidth / 2); + const int8x8_t Zero = {0}; #endif const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()}; for (IndexType p = 0; p < 2; ++p) { - const IndexType offset = kHalfDimensions * p; + const IndexType offset = HalfDimensions * p; #if defined(USE_AVX512) auto out = reinterpret_cast<__m512i*>(&output[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < NumChunks; ++j) { __m512i sum0 = _mm512_load_si512( &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 0]); __m512i sum1 = _mm512_load_si512( &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 1]); - _mm512_store_si512(&out[j], _mm512_permutexvar_epi64(kControl, - _mm512_max_epi8(_mm512_packs_epi16(sum0, sum1), kZero))); + _mm512_store_si512(&out[j], _mm512_permutexvar_epi64(Control, + _mm512_max_epi8(_mm512_packs_epi16(sum0, sum1), Zero))); } #elif defined(USE_AVX2) auto out = reinterpret_cast<__m256i*>(&output[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < NumChunks; ++j) { __m256i sum0 = _mm256_load_si256( &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 0]); __m256i sum1 = _mm256_load_si256( &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 1]); _mm256_store_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8( - _mm256_packs_epi16(sum0, sum1), kZero), kControl)); + _mm256_packs_epi16(sum0, sum1), Zero), Control)); } #elif defined(USE_SSE2) auto out = reinterpret_cast<__m128i*>(&output[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < NumChunks; ++j) { __m128i sum0 = _mm_load_si128(&reinterpret_cast( accumulation[perspectives[p]][0])[j * 2 + 0]); __m128i sum1 = _mm_load_si128(&reinterpret_cast( @@ -194,7 +191,7 @@ namespace Stockfish::Eval::NNUE { _mm_store_si128(&out[j], #ifdef USE_SSE41 - _mm_max_epi8(packedbytes, kZero) + _mm_max_epi8(packedbytes, Zero) #else _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s) #endif @@ -204,7 +201,7 @@ namespace Stockfish::Eval::NNUE { #elif defined(USE_MMX) auto out = reinterpret_cast<__m64*>(&output[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < NumChunks; ++j) { __m64 sum0 = *(&reinterpret_cast( accumulation[perspectives[p]][0])[j * 2 + 0]); __m64 sum1 = *(&reinterpret_cast( @@ -215,14 +212,14 @@ namespace Stockfish::Eval::NNUE { #elif defined(USE_NEON) const auto out = reinterpret_cast(&output[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { + for (IndexType j = 0; j < NumChunks; ++j) { int16x8_t sum = reinterpret_cast( accumulation[perspectives[p]][0])[j]; - out[j] = vmax_s8(vqmovn_s16(sum), kZero); + out[j] = vmax_s8(vqmovn_s16(sum), Zero); } #else - for (IndexType j = 0; j < kHalfDimensions; ++j) { + for (IndexType j = 0; j < HalfDimensions; ++j) { BiasType sum = accumulation[static_cast(perspectives[p])][0][j]; output[offset + j] = static_cast( std::max(0, std::min(127, sum))); @@ -236,12 +233,12 @@ namespace Stockfish::Eval::NNUE { } private: - void UpdateAccumulator(const Position& pos, const Color c) const { + void update_accumulator(const Position& pos, const Color c) const { #ifdef VECTOR // Gcc-10.2 unnecessarily spills AVX2 registers if this array // is defined in the VECTOR code below, once in each branch - vec_t acc[kNumRegs]; + vec_t acc[NumRegs]; #endif // Look for a usable accumulator of an earlier position. We keep track @@ -254,8 +251,8 @@ namespace Stockfish::Eval::NNUE { // The first condition tests whether an incremental update is // possible at all: if this side's king has moved, it is not possible. static_assert(std::is_same_v>, - "Current code assumes that only kFriendlyKingMoved refresh trigger is being used."); + Features::CompileTimeList>, + "Current code assumes that only FriendlyKingMoved refresh trigger is being used."); if ( dp.piece[0] == make_piece(c, KING) || (gain -= dp.dirty_num + 1) < 0) break; @@ -273,13 +270,13 @@ namespace Stockfish::Eval::NNUE { // Gather all features to be updated. This code assumes HalfKP features // only and doesn't support refresh triggers. - static_assert(std::is_same_v>, + static_assert(std::is_same_v>, RawFeatures>); Features::IndexList removed[2], added[2]; - Features::HalfKP::AppendChangedIndices(pos, + Features::HalfKP::append_changed_indices(pos, next->dirtyPiece, c, &removed[0], &added[0]); for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous) - Features::HalfKP::AppendChangedIndices(pos, + Features::HalfKP::append_changed_indices(pos, st2->dirtyPiece, c, &removed[1], &added[1]); // Mark the accumulators as computed. @@ -290,12 +287,12 @@ namespace Stockfish::Eval::NNUE { StateInfo *info[3] = { next, next == pos.state() ? nullptr : pos.state(), nullptr }; #ifdef VECTOR - for (IndexType j = 0; j < kHalfDimensions / kTileHeight; ++j) + for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j) { // Load accumulator auto accTile = reinterpret_cast( - &st->accumulator.accumulation[c][0][j * kTileHeight]); - for (IndexType k = 0; k < kNumRegs; ++k) + &st->accumulator.accumulation[c][0][j * TileHeight]); + for (IndexType k = 0; k < NumRegs; ++k) acc[k] = vec_load(&accTile[k]); for (IndexType i = 0; info[i]; ++i) @@ -303,25 +300,25 @@ namespace Stockfish::Eval::NNUE { // Difference calculation for the deactivated features for (const auto index : removed[i]) { - const IndexType offset = kHalfDimensions * index + j * kTileHeight; - auto column = reinterpret_cast(&weights_[offset]); - for (IndexType k = 0; k < kNumRegs; ++k) + const IndexType offset = HalfDimensions * index + j * TileHeight; + auto column = reinterpret_cast(&weights[offset]); + for (IndexType k = 0; k < NumRegs; ++k) acc[k] = vec_sub_16(acc[k], column[k]); } // Difference calculation for the activated features for (const auto index : added[i]) { - const IndexType offset = kHalfDimensions * index + j * kTileHeight; - auto column = reinterpret_cast(&weights_[offset]); - for (IndexType k = 0; k < kNumRegs; ++k) + const IndexType offset = HalfDimensions * index + j * TileHeight; + auto column = reinterpret_cast(&weights[offset]); + for (IndexType k = 0; k < NumRegs; ++k) acc[k] = vec_add_16(acc[k], column[k]); } // Store accumulator accTile = reinterpret_cast( - &info[i]->accumulator.accumulation[c][0][j * kTileHeight]); - for (IndexType k = 0; k < kNumRegs; ++k) + &info[i]->accumulator.accumulation[c][0][j * TileHeight]); + for (IndexType k = 0; k < NumRegs; ++k) vec_store(&accTile[k], acc[k]); } } @@ -331,25 +328,25 @@ namespace Stockfish::Eval::NNUE { { std::memcpy(info[i]->accumulator.accumulation[c][0], st->accumulator.accumulation[c][0], - kHalfDimensions * sizeof(BiasType)); + HalfDimensions * sizeof(BiasType)); st = info[i]; // Difference calculation for the deactivated features for (const auto index : removed[i]) { - const IndexType offset = kHalfDimensions * index; + const IndexType offset = HalfDimensions * index; - for (IndexType j = 0; j < kHalfDimensions; ++j) - st->accumulator.accumulation[c][0][j] -= weights_[offset + j]; + for (IndexType j = 0; j < HalfDimensions; ++j) + st->accumulator.accumulation[c][0][j] -= weights[offset + j]; } // Difference calculation for the activated features for (const auto index : added[i]) { - const IndexType offset = kHalfDimensions * index; + const IndexType offset = HalfDimensions * index; - for (IndexType j = 0; j < kHalfDimensions; ++j) - st->accumulator.accumulation[c][0][j] += weights_[offset + j]; + for (IndexType j = 0; j < HalfDimensions; ++j) + st->accumulator.accumulation[c][0][j] += weights[offset + j]; } } #endif @@ -360,41 +357,41 @@ namespace Stockfish::Eval::NNUE { auto& accumulator = pos.state()->accumulator; accumulator.state[c] = COMPUTED; Features::IndexList active; - Features::HalfKP::AppendActiveIndices(pos, c, &active); + Features::HalfKP::append_active_indices(pos, c, &active); #ifdef VECTOR - for (IndexType j = 0; j < kHalfDimensions / kTileHeight; ++j) + for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j) { auto biasesTile = reinterpret_cast( - &biases_[j * kTileHeight]); - for (IndexType k = 0; k < kNumRegs; ++k) + &biases[j * TileHeight]); + for (IndexType k = 0; k < NumRegs; ++k) acc[k] = biasesTile[k]; for (const auto index : active) { - const IndexType offset = kHalfDimensions * index + j * kTileHeight; - auto column = reinterpret_cast(&weights_[offset]); + const IndexType offset = HalfDimensions * index + j * TileHeight; + auto column = reinterpret_cast(&weights[offset]); - for (unsigned k = 0; k < kNumRegs; ++k) + for (unsigned k = 0; k < NumRegs; ++k) acc[k] = vec_add_16(acc[k], column[k]); } auto accTile = reinterpret_cast( - &accumulator.accumulation[c][0][j * kTileHeight]); - for (unsigned k = 0; k < kNumRegs; k++) + &accumulator.accumulation[c][0][j * TileHeight]); + for (unsigned k = 0; k < NumRegs; k++) vec_store(&accTile[k], acc[k]); } #else - std::memcpy(accumulator.accumulation[c][0], biases_, - kHalfDimensions * sizeof(BiasType)); + std::memcpy(accumulator.accumulation[c][0], biases, + HalfDimensions * sizeof(BiasType)); for (const auto index : active) { - const IndexType offset = kHalfDimensions * index; + const IndexType offset = HalfDimensions * index; - for (IndexType j = 0; j < kHalfDimensions; ++j) - accumulator.accumulation[c][0][j] += weights_[offset + j]; + for (IndexType j = 0; j < HalfDimensions; ++j) + accumulator.accumulation[c][0][j] += weights[offset + j]; } #endif } @@ -407,9 +404,9 @@ namespace Stockfish::Eval::NNUE { using BiasType = std::int16_t; using WeightType = std::int16_t; - alignas(kCacheLineSize) BiasType biases_[kHalfDimensions]; - alignas(kCacheLineSize) - WeightType weights_[kHalfDimensions * kInputDimensions]; + alignas(CacheLineSize) BiasType biases[HalfDimensions]; + alignas(CacheLineSize) + WeightType weights[HalfDimensions * InputDimensions]; }; } // namespace Stockfish::Eval::NNUE