X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=src%2Fnnue%2Fnnue_feature_transformer.h;h=741d97cf1190c539bdcf2766ae6852bdefc52a12;hb=f193778446acc6e60d7f0f99c6eb01489f89e962;hp=de4b49374f91768dce1d04f1dc2dd2615b69bf3f;hpb=fbbd4adc3c01460faa3cc8f91771ab9b0ef718ca;p=stockfish diff --git a/src/nnue/nnue_feature_transformer.h b/src/nnue/nnue_feature_transformer.h index de4b4937..741d97cf 100644 --- a/src/nnue/nnue_feature_transformer.h +++ b/src/nnue/nnue_feature_transformer.h @@ -23,7 +23,8 @@ #include "nnue_common.h" #include "nnue_architecture.h" -#include "features/index_list.h" + +#include "../misc.h" #include // std::memset() @@ -34,45 +35,82 @@ namespace Stockfish::Eval::NNUE { // vector registers. #define VECTOR + static_assert(PSQTBuckets == 8, "Assumed by the current choice of constants."); + #ifdef USE_AVX512 typedef __m512i vec_t; + typedef __m256i psqt_vec_t; #define vec_load(a) _mm512_load_si512(a) #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) + #define vec_load_psqt(a) _mm256_load_si256(a) + #define vec_store_psqt(a,b) _mm256_store_si256(a,b) + #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b) + #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b) + #define vec_zero_psqt() _mm256_setzero_si256() static constexpr IndexType NumRegs = 8; // only 8 are needed + static constexpr IndexType NumPsqtRegs = 1; #elif USE_AVX2 typedef __m256i vec_t; + typedef __m256i psqt_vec_t; #define vec_load(a) _mm256_load_si256(a) #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) + #define vec_load_psqt(a) _mm256_load_si256(a) + #define vec_store_psqt(a,b) _mm256_store_si256(a,b) + #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b) + #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b) + #define vec_zero_psqt() _mm256_setzero_si256() static constexpr IndexType NumRegs = 16; + static constexpr IndexType NumPsqtRegs = 1; #elif USE_SSE2 typedef __m128i vec_t; + typedef __m128i psqt_vec_t; #define vec_load(a) (*(a)) #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) + #define vec_load_psqt(a) (*(a)) + #define vec_store_psqt(a,b) *(a)=(b) + #define vec_add_psqt_32(a,b) _mm_add_epi32(a,b) + #define vec_sub_psqt_32(a,b) _mm_sub_epi32(a,b) + #define vec_zero_psqt() _mm_setzero_si128() static constexpr IndexType NumRegs = Is64Bit ? 16 : 8; + static constexpr IndexType NumPsqtRegs = 2; #elif USE_MMX typedef __m64 vec_t; + typedef __m64 psqt_vec_t; #define vec_load(a) (*(a)) #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) + #define vec_load_psqt(a) (*(a)) + #define vec_store_psqt(a,b) *(a)=(b) + #define vec_add_psqt_32(a,b) _mm_add_pi32(a,b) + #define vec_sub_psqt_32(a,b) _mm_sub_pi32(a,b) + #define vec_zero_psqt() _mm_setzero_si64() static constexpr IndexType NumRegs = 8; + static constexpr IndexType NumPsqtRegs = 4; #elif USE_NEON typedef int16x8_t vec_t; + typedef int32x4_t psqt_vec_t; #define vec_load(a) (*(a)) #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) + #define vec_load_psqt(a) (*(a)) + #define vec_store_psqt(a,b) *(a)=(b) + #define vec_add_psqt_32(a,b) vaddq_s32(a,b) + #define vec_sub_psqt_32(a,b) vsubq_s32(a,b) + #define vec_zero_psqt() psqt_vec_t{0} static constexpr IndexType NumRegs = 16; + static constexpr IndexType NumPsqtRegs = 2; #else #undef VECTOR @@ -88,7 +126,9 @@ namespace Stockfish::Eval::NNUE { #ifdef VECTOR static constexpr IndexType TileHeight = NumRegs * sizeof(vec_t) / 2; + static constexpr IndexType PsqtTileHeight = NumPsqtRegs * sizeof(psqt_vec_t) / 4; static_assert(HalfDimensions % TileHeight == 0, "TileHeight must divide HalfDimensions"); + static_assert(PSQTBuckets % PsqtTileHeight == 0, "PsqtTileHeight must divide PSQTBuckets"); #endif public: @@ -96,7 +136,7 @@ namespace Stockfish::Eval::NNUE { using OutputType = TransformedFeatureType; // Number of input/output dimensions - static constexpr IndexType InputDimensions = RawFeatures::Dimensions; + static constexpr IndexType InputDimensions = FeatureSet::Dimensions; static constexpr IndexType OutputDimensions = HalfDimensions * 2; // Size of forward propagation buffer @@ -105,7 +145,7 @@ namespace Stockfish::Eval::NNUE { // Hash value embedded in the evaluation file static constexpr std::uint32_t get_hash_value() { - return RawFeatures::HashValue ^ OutputDimensions; + return FeatureSet::HashValue ^ OutputDimensions; } // Read network parameters @@ -114,15 +154,33 @@ namespace Stockfish::Eval::NNUE { biases[i] = read_little_endian(stream); for (std::size_t i = 0; i < HalfDimensions * InputDimensions; ++i) weights[i] = read_little_endian(stream); + for (std::size_t i = 0; i < PSQTBuckets * InputDimensions; ++i) + psqtWeights[i] = read_little_endian(stream); + return !stream.fail(); + } + + // Write network parameters + bool write_parameters(std::ostream& stream) const { + for (std::size_t i = 0; i < HalfDimensions; ++i) + write_little_endian(stream, biases[i]); + for (std::size_t i = 0; i < HalfDimensions * InputDimensions; ++i) + write_little_endian(stream, weights[i]); return !stream.fail(); } // Convert input features - void transform(const Position& pos, OutputType* output) const { + std::int32_t transform(const Position& pos, OutputType* output, int bucket) const { update_accumulator(pos, WHITE); update_accumulator(pos, BLACK); + const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()}; const auto& accumulation = pos.state()->accumulator.accumulation; + const auto& psqtAccumulation = pos.state()->accumulator.psqtAccumulation; + + const auto psqt = ( + psqtAccumulation[static_cast(perspectives[0])][bucket] + - psqtAccumulation[static_cast(perspectives[1])][bucket] + ) / 2; #if defined(USE_AVX512) constexpr IndexType NumChunks = HalfDimensions / (SimdWidth * 2); @@ -153,7 +211,6 @@ namespace Stockfish::Eval::NNUE { 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 = HalfDimensions * p; @@ -161,9 +218,9 @@ namespace Stockfish::Eval::NNUE { auto out = reinterpret_cast<__m512i*>(&output[offset]); for (IndexType j = 0; j < NumChunks; ++j) { __m512i sum0 = _mm512_load_si512( - &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 0]); + &reinterpret_cast(accumulation[perspectives[p]])[j * 2 + 0]); __m512i sum1 = _mm512_load_si512( - &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 1]); + &reinterpret_cast(accumulation[perspectives[p]])[j * 2 + 1]); _mm512_store_si512(&out[j], _mm512_permutexvar_epi64(Control, _mm512_max_epi8(_mm512_packs_epi16(sum0, sum1), Zero))); } @@ -172,9 +229,9 @@ namespace Stockfish::Eval::NNUE { auto out = reinterpret_cast<__m256i*>(&output[offset]); for (IndexType j = 0; j < NumChunks; ++j) { __m256i sum0 = _mm256_load_si256( - &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 0]); + &reinterpret_cast(accumulation[perspectives[p]])[j * 2 + 0]); __m256i sum1 = _mm256_load_si256( - &reinterpret_cast(accumulation[perspectives[p]][0])[j * 2 + 1]); + &reinterpret_cast(accumulation[perspectives[p]])[j * 2 + 1]); _mm256_store_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8( _mm256_packs_epi16(sum0, sum1), Zero), Control)); } @@ -183,9 +240,9 @@ namespace Stockfish::Eval::NNUE { auto out = reinterpret_cast<__m128i*>(&output[offset]); for (IndexType j = 0; j < NumChunks; ++j) { __m128i sum0 = _mm_load_si128(&reinterpret_cast( - accumulation[perspectives[p]][0])[j * 2 + 0]); + accumulation[perspectives[p]])[j * 2 + 0]); __m128i sum1 = _mm_load_si128(&reinterpret_cast( - accumulation[perspectives[p]][0])[j * 2 + 1]); + accumulation[perspectives[p]])[j * 2 + 1]); const __m128i packedbytes = _mm_packs_epi16(sum0, sum1); _mm_store_si128(&out[j], @@ -203,9 +260,9 @@ namespace Stockfish::Eval::NNUE { auto out = reinterpret_cast<__m64*>(&output[offset]); for (IndexType j = 0; j < NumChunks; ++j) { __m64 sum0 = *(&reinterpret_cast( - accumulation[perspectives[p]][0])[j * 2 + 0]); + accumulation[perspectives[p]])[j * 2 + 0]); __m64 sum1 = *(&reinterpret_cast( - accumulation[perspectives[p]][0])[j * 2 + 1]); + accumulation[perspectives[p]])[j * 2 + 1]); const __m64 packedbytes = _mm_packs_pi16(sum0, sum1); out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s); } @@ -214,13 +271,13 @@ namespace Stockfish::Eval::NNUE { const auto out = reinterpret_cast(&output[offset]); for (IndexType j = 0; j < NumChunks; ++j) { int16x8_t sum = reinterpret_cast( - accumulation[perspectives[p]][0])[j]; + accumulation[perspectives[p]])[j]; out[j] = vmax_s8(vqmovn_s16(sum), Zero); } #else for (IndexType j = 0; j < HalfDimensions; ++j) { - BiasType sum = accumulation[static_cast(perspectives[p])][0][j]; + BiasType sum = accumulation[static_cast(perspectives[p])][j]; output[offset + j] = static_cast( std::max(0, std::min(127, sum))); } @@ -230,37 +287,42 @@ namespace Stockfish::Eval::NNUE { #if defined(USE_MMX) _mm_empty(); #endif + + return psqt; } private: - void update_accumulator(const Position& pos, const Color c) const { + void update_accumulator(const Position& pos, const Color perspective) const { + + // The size must be enough to contain the largest possible update. + // That might depend on the feature set and generally relies on the + // feature set's update cost calculation to be correct and never + // allow updates with more added/removed features than MaxActiveDimensions. + using IndexList = ValueList; #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[NumRegs]; + psqt_vec_t psqt[NumPsqtRegs]; #endif // Look for a usable accumulator of an earlier position. We keep track // of the estimated gain in terms of features to be added/subtracted. StateInfo *st = pos.state(), *next = nullptr; - int gain = pos.count() - 2; - while (st->accumulator.state[c] == EMPTY) + int gain = FeatureSet::refresh_cost(pos); + while (st->accumulator.state[perspective] == EMPTY) { - auto& dp = st->dirtyPiece; - // 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 FriendlyKingMoved refresh trigger is being used."); - if ( dp.piece[0] == make_piece(c, KING) - || (gain -= dp.dirty_num + 1) < 0) + // This governs when a full feature refresh is needed and how many + // updates are better than just one full refresh. + if ( FeatureSet::requires_refresh(st, perspective) + || (gain -= FeatureSet::update_cost(st) + 1) < 0) break; next = st; st = st->previous; } - if (st->accumulator.state[c] == COMPUTED) + if (st->accumulator.state[perspective] == COMPUTED) { if (next == nullptr) return; @@ -268,34 +330,32 @@ namespace Stockfish::Eval::NNUE { // Update incrementally in two steps. First, we update the "next" // accumulator. Then, we update the current accumulator (pos.state()). - // Gather all features to be updated. This code assumes HalfKP features - // only and doesn't support refresh triggers. - static_assert(std::is_same_v>, - RawFeatures>); - Features::IndexList removed[2], added[2]; - Features::HalfKP::append_changed_indices(pos, - next->dirtyPiece, c, &removed[0], &added[0]); + // Gather all features to be updated. + const Square ksq = pos.square(perspective); + IndexList removed[2], added[2]; + FeatureSet::append_changed_indices( + ksq, next, perspective, removed[0], added[0]); for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous) - Features::HalfKP::append_changed_indices(pos, - st2->dirtyPiece, c, &removed[1], &added[1]); + FeatureSet::append_changed_indices( + ksq, st2, perspective, removed[1], added[1]); // Mark the accumulators as computed. - next->accumulator.state[c] = COMPUTED; - pos.state()->accumulator.state[c] = COMPUTED; + next->accumulator.state[perspective] = COMPUTED; + pos.state()->accumulator.state[perspective] = COMPUTED; - // Now update the accumulators listed in info[], where the last element is a sentinel. - StateInfo *info[3] = + // Now update the accumulators listed in states_to_update[], where the last element is a sentinel. + StateInfo *states_to_update[3] = { next, next == pos.state() ? nullptr : pos.state(), nullptr }; #ifdef VECTOR for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j) { // Load accumulator auto accTile = reinterpret_cast( - &st->accumulator.accumulation[c][0][j * TileHeight]); + &st->accumulator.accumulation[perspective][j * TileHeight]); for (IndexType k = 0; k < NumRegs; ++k) acc[k] = vec_load(&accTile[k]); - for (IndexType i = 0; info[i]; ++i) + for (IndexType i = 0; states_to_update[i]; ++i) { // Difference calculation for the deactivated features for (const auto index : removed[i]) @@ -317,19 +377,59 @@ namespace Stockfish::Eval::NNUE { // Store accumulator accTile = reinterpret_cast( - &info[i]->accumulator.accumulation[c][0][j * TileHeight]); + &states_to_update[i]->accumulator.accumulation[perspective][j * TileHeight]); for (IndexType k = 0; k < NumRegs; ++k) vec_store(&accTile[k], acc[k]); } } + for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j) + { + // Load accumulator + auto accTilePsqt = reinterpret_cast( + &st->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]); + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + psqt[k] = vec_load_psqt(&accTilePsqt[k]); + + for (IndexType i = 0; states_to_update[i]; ++i) + { + // Difference calculation for the deactivated features + for (const auto index : removed[i]) + { + const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight; + auto columnPsqt = reinterpret_cast(&psqtWeights[offset]); + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + psqt[k] = vec_sub_psqt_32(psqt[k], columnPsqt[k]); + } + + // Difference calculation for the activated features + for (const auto index : added[i]) + { + const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight; + auto columnPsqt = reinterpret_cast(&psqtWeights[offset]); + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]); + } + + // Store accumulator + accTilePsqt = reinterpret_cast( + &states_to_update[i]->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]); + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + vec_store_psqt(&accTilePsqt[k], psqt[k]); + } + } + #else - for (IndexType i = 0; info[i]; ++i) + for (IndexType i = 0; states_to_update[i]; ++i) { - std::memcpy(info[i]->accumulator.accumulation[c][0], - st->accumulator.accumulation[c][0], + std::memcpy(states_to_update[i]->accumulator.accumulation[perspective], + st->accumulator.accumulation[perspective], HalfDimensions * sizeof(BiasType)); - st = info[i]; + + for (std::size_t k = 0; k < PSQTBuckets; ++k) + states_to_update[i]->accumulator.psqtAccumulation[perspective][k] = st->accumulator.psqtAccumulation[perspective][k]; + + st = states_to_update[i]; // Difference calculation for the deactivated features for (const auto index : removed[i]) @@ -337,7 +437,10 @@ namespace Stockfish::Eval::NNUE { const IndexType offset = HalfDimensions * index; for (IndexType j = 0; j < HalfDimensions; ++j) - st->accumulator.accumulation[c][0][j] -= weights[offset + j]; + st->accumulator.accumulation[perspective][j] -= weights[offset + j]; + + for (std::size_t k = 0; k < PSQTBuckets; ++k) + st->accumulator.psqtAccumulation[perspective][k] -= psqtWeights[index * PSQTBuckets + k]; } // Difference calculation for the activated features @@ -346,7 +449,10 @@ namespace Stockfish::Eval::NNUE { const IndexType offset = HalfDimensions * index; for (IndexType j = 0; j < HalfDimensions; ++j) - st->accumulator.accumulation[c][0][j] += weights[offset + j]; + st->accumulator.accumulation[perspective][j] += weights[offset + j]; + + for (std::size_t k = 0; k < PSQTBuckets; ++k) + st->accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k]; } } #endif @@ -355,9 +461,9 @@ namespace Stockfish::Eval::NNUE { { // Refresh the accumulator auto& accumulator = pos.state()->accumulator; - accumulator.state[c] = COMPUTED; - Features::IndexList active; - Features::HalfKP::append_active_indices(pos, c, &active); + accumulator.state[perspective] = COMPUTED; + IndexList active; + FeatureSet::append_active_indices(pos, perspective, active); #ifdef VECTOR for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j) @@ -377,21 +483,47 @@ namespace Stockfish::Eval::NNUE { } auto accTile = reinterpret_cast( - &accumulator.accumulation[c][0][j * TileHeight]); + &accumulator.accumulation[perspective][j * TileHeight]); for (unsigned k = 0; k < NumRegs; k++) vec_store(&accTile[k], acc[k]); } + for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j) + { + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + psqt[k] = vec_zero_psqt(); + + for (const auto index : active) + { + const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight; + auto columnPsqt = reinterpret_cast(&psqtWeights[offset]); + + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]); + } + + auto accTilePsqt = reinterpret_cast( + &accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]); + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + vec_store_psqt(&accTilePsqt[k], psqt[k]); + } + #else - std::memcpy(accumulator.accumulation[c][0], biases, + std::memcpy(accumulator.accumulation[perspective], biases, HalfDimensions * sizeof(BiasType)); + for (std::size_t k = 0; k < PSQTBuckets; ++k) + accumulator.psqtAccumulation[perspective][k] = 0; + for (const auto index : active) { const IndexType offset = HalfDimensions * index; for (IndexType j = 0; j < HalfDimensions; ++j) - accumulator.accumulation[c][0][j] += weights[offset + j]; + accumulator.accumulation[perspective][j] += weights[offset + j]; + + for (std::size_t k = 0; k < PSQTBuckets; ++k) + accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k]; } #endif } @@ -403,10 +535,11 @@ namespace Stockfish::Eval::NNUE { using BiasType = std::int16_t; using WeightType = std::int16_t; + using PSQTWeightType = std::int32_t; alignas(CacheLineSize) BiasType biases[HalfDimensions]; - alignas(CacheLineSize) - WeightType weights[HalfDimensions * InputDimensions]; + alignas(CacheLineSize) WeightType weights[HalfDimensions * InputDimensions]; + alignas(CacheLineSize) PSQTWeightType psqtWeights[InputDimensions * PSQTBuckets]; }; } // namespace Stockfish::Eval::NNUE