X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fnnue%2Fnnue_feature_transformer.h;h=2af80f0779250763036e7513383d79718d91fc46;hp=59a965ac769450f11c4fde517ac09d731c0749f6;hb=HEAD;hpb=d558f8a673b56b32ab6da8050f41b9e02fe1758b diff --git a/src/nnue/nnue_feature_transformer.h b/src/nnue/nnue_feature_transformer.h index 59a965ac..2af80f07 100644 --- a/src/nnue/nnue_feature_transformer.h +++ b/src/nnue/nnue_feature_transformer.h @@ -1,6 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (C) 2004-2021 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 @@ -21,595 +21,696 @@ #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED #define NNUE_FEATURE_TRANSFORMER_H_INCLUDED -#include "nnue_common.h" -#include "nnue_architecture.h" +#include +#include +#include +#include +#include +#include -#include // std::memset() +#include "../position.h" +#include "../types.h" +#include "nnue_accumulator.h" +#include "nnue_architecture.h" +#include "nnue_common.h" namespace Stockfish::Eval::NNUE { - using BiasType = std::int16_t; - using WeightType = std::int16_t; - using PSQTWeightType = std::int32_t; - - // If vector instructions are enabled, we update and refresh the - // accumulator tile by tile such that each tile fits in the CPU's - // vector registers. - #define VECTOR - - static_assert(PSQTBuckets % 8 == 0, - "Per feature PSQT values cannot be processed at granularity lower than 8 at a time."); - - #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() - #define NumRegistersSIMD 32 - - #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() - #define NumRegistersSIMD 16 - - #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() - #define NumRegistersSIMD (Is64Bit ? 16 : 8) - - #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() - #define NumRegistersSIMD 8 - - #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} - #define NumRegistersSIMD 16 - - #else - #undef VECTOR - - #endif - - - #ifdef VECTOR - - // Compute optimal SIMD register count for feature transformer accumulation. - - // We use __m* types as template arguments, which causes GCC to emit warnings - // about losing some attribute information. This is irrelevant to us as we - // only take their size, so the following pragma are harmless. - #pragma GCC diagnostic push - #pragma GCC diagnostic ignored "-Wignored-attributes" - - template - static constexpr int BestRegisterCount() - { - #define RegisterSize sizeof(SIMDRegisterType) - #define LaneSize sizeof(LaneType) - - static_assert(RegisterSize >= LaneSize); - static_assert(MaxRegisters <= NumRegistersSIMD); - static_assert(MaxRegisters > 0); - static_assert(NumRegistersSIMD > 0); - static_assert(RegisterSize % LaneSize == 0); - static_assert((NumLanes * LaneSize) % RegisterSize == 0); - - const int ideal = (NumLanes * LaneSize) / RegisterSize; - if (ideal <= MaxRegisters) - return ideal; - - // Look for the largest divisor of the ideal register count that is smaller than MaxRegisters - for (int divisor = MaxRegisters; divisor > 1; --divisor) - if (ideal % divisor == 0) - return divisor; - - return 1; - } - - static constexpr int NumRegs = BestRegisterCount(); - static constexpr int NumPsqtRegs = BestRegisterCount(); - - #pragma GCC diagnostic pop - - #endif - - - - // Input feature converter - class FeatureTransformer { +using BiasType = std::int16_t; +using WeightType = std::int16_t; +using PSQTWeightType = std::int32_t; + +// If vector instructions are enabled, we update and refresh the +// accumulator tile by tile such that each tile fits in the CPU's +// vector registers. +#define VECTOR + +static_assert(PSQTBuckets % 8 == 0, + "Per feature PSQT values cannot be processed at granularity lower than 8 at a time."); + +#ifdef USE_AVX512 +using vec_t = __m512i; +using psqt_vec_t = __m256i; + #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_mul_16(a, b) _mm512_mullo_epi16(a, b) + #define vec_zero() _mm512_setzero_epi32() + #define vec_set_16(a) _mm512_set1_epi16(a) + #define vec_max_16(a, b) _mm512_max_epi16(a, b) + #define vec_min_16(a, b) _mm512_min_epi16(a, b) +inline vec_t vec_msb_pack_16(vec_t a, vec_t b) { + vec_t compacted = _mm512_packs_epi16(_mm512_srli_epi16(a, 7), _mm512_srli_epi16(b, 7)); + return _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), compacted); +} + #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() + #define NumRegistersSIMD 16 + #define MaxChunkSize 64 + +#elif USE_AVX2 +using vec_t = __m256i; +using psqt_vec_t = __m256i; + #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_mul_16(a, b) _mm256_mullo_epi16(a, b) + #define vec_zero() _mm256_setzero_si256() + #define vec_set_16(a) _mm256_set1_epi16(a) + #define vec_max_16(a, b) _mm256_max_epi16(a, b) + #define vec_min_16(a, b) _mm256_min_epi16(a, b) +inline vec_t vec_msb_pack_16(vec_t a, vec_t b) { + vec_t compacted = _mm256_packs_epi16(_mm256_srli_epi16(a, 7), _mm256_srli_epi16(b, 7)); + return _mm256_permute4x64_epi64(compacted, 0b11011000); +} + #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() + #define NumRegistersSIMD 16 + #define MaxChunkSize 32 + +#elif USE_SSE2 +using vec_t = __m128i; +using psqt_vec_t = __m128i; + #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_mul_16(a, b) _mm_mullo_epi16(a, b) + #define vec_zero() _mm_setzero_si128() + #define vec_set_16(a) _mm_set1_epi16(a) + #define vec_max_16(a, b) _mm_max_epi16(a, b) + #define vec_min_16(a, b) _mm_min_epi16(a, b) + #define vec_msb_pack_16(a, b) _mm_packs_epi16(_mm_srli_epi16(a, 7), _mm_srli_epi16(b, 7)) + #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() + #define NumRegistersSIMD (Is64Bit ? 16 : 8) + #define MaxChunkSize 16 + +#elif USE_NEON +using vec_t = int16x8_t; +using psqt_vec_t = int32x4_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_mul_16(a, b) vmulq_s16(a, b) + #define vec_zero() \ + vec_t { 0 } + #define vec_set_16(a) vdupq_n_s16(a) + #define vec_max_16(a, b) vmaxq_s16(a, b) + #define vec_min_16(a, b) vminq_s16(a, b) +inline vec_t vec_msb_pack_16(vec_t a, vec_t b) { + const int8x8_t shifta = vshrn_n_s16(a, 7); + const int8x8_t shiftb = vshrn_n_s16(b, 7); + const int8x16_t compacted = vcombine_s8(shifta, shiftb); + return *reinterpret_cast(&compacted); +} + #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 } + #define NumRegistersSIMD 16 + #define MaxChunkSize 16 + +#else + #undef VECTOR + +#endif + + +#ifdef VECTOR + + // Compute optimal SIMD register count for feature transformer accumulation. + + // We use __m* types as template arguments, which causes GCC to emit warnings + // about losing some attribute information. This is irrelevant to us as we + // only take their size, so the following pragma are harmless. + #if defined(__GNUC__) + #pragma GCC diagnostic push + #pragma GCC diagnostic ignored "-Wignored-attributes" + #endif + +template +static constexpr int BestRegisterCount() { + #define RegisterSize sizeof(SIMDRegisterType) + #define LaneSize sizeof(LaneType) + + static_assert(RegisterSize >= LaneSize); + static_assert(MaxRegisters <= NumRegistersSIMD); + static_assert(MaxRegisters > 0); + static_assert(NumRegistersSIMD > 0); + static_assert(RegisterSize % LaneSize == 0); + static_assert((NumLanes * LaneSize) % RegisterSize == 0); + + const int ideal = (NumLanes * LaneSize) / RegisterSize; + if (ideal <= MaxRegisters) + return ideal; + + // Look for the largest divisor of the ideal register count that is smaller than MaxRegisters + for (int divisor = MaxRegisters; divisor > 1; --divisor) + if (ideal % divisor == 0) + return divisor; + + return 1; +} + +static constexpr int NumRegs = + BestRegisterCount(); +static constexpr int NumPsqtRegs = + BestRegisterCount(); + #if defined(__GNUC__) + #pragma GCC diagnostic pop + #endif +#endif + + +// Input feature converter +class FeatureTransformer { private: // Number of output dimensions for one side static constexpr IndexType HalfDimensions = TransformedFeatureDimensions; - #ifdef VECTOR - static constexpr IndexType TileHeight = NumRegs * sizeof(vec_t) / 2; +#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 +#endif public: // Output type using OutputType = TransformedFeatureType; // Number of input/output dimensions - static constexpr IndexType InputDimensions = FeatureSet::Dimensions; - static constexpr IndexType OutputDimensions = HalfDimensions * 2; + static constexpr IndexType InputDimensions = FeatureSet::Dimensions; + static constexpr IndexType OutputDimensions = HalfDimensions; // Size of forward propagation buffer - static constexpr std::size_t BufferSize = - OutputDimensions * sizeof(OutputType); + static constexpr std::size_t BufferSize = OutputDimensions * sizeof(OutputType); // Hash value embedded in the evaluation file static constexpr std::uint32_t get_hash_value() { - return FeatureSet::HashValue ^ OutputDimensions; + return FeatureSet::HashValue ^ (OutputDimensions * 2); } // Read network parameters bool read_parameters(std::istream& stream) { - read_little_endian(stream, biases , HalfDimensions ); - read_little_endian(stream, weights , HalfDimensions * InputDimensions); - read_little_endian(stream, psqtWeights, PSQTBuckets * InputDimensions); + read_leb_128(stream, biases, HalfDimensions); + read_leb_128(stream, weights, HalfDimensions * InputDimensions); + read_leb_128(stream, psqtWeights, PSQTBuckets * InputDimensions); - return !stream.fail(); + return !stream.fail(); } // Write network parameters bool write_parameters(std::ostream& stream) const { - write_little_endian(stream, biases , HalfDimensions ); - write_little_endian(stream, weights , HalfDimensions * InputDimensions); - write_little_endian(stream, psqtWeights, PSQTBuckets * InputDimensions); + write_leb_128(stream, biases, HalfDimensions); + write_leb_128(stream, weights, HalfDimensions * InputDimensions); + write_leb_128(stream, psqtWeights, PSQTBuckets * InputDimensions); - return !stream.fail(); + return !stream.fail(); } // Convert input features 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[perspectives[0]][bucket] - - psqtAccumulation[perspectives[1]][bucket] - ) / 2; - - - #if defined(USE_AVX512) - - 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(); - - for (IndexType p = 0; p < 2; ++p) - { - const IndexType offset = HalfDimensions * p; - auto out = reinterpret_cast<__m512i*>(&output[offset]); - for (IndexType j = 0; j < NumChunks; ++j) - { - __m512i sum0 = _mm512_load_si512(&reinterpret_cast - (accumulation[perspectives[p]])[j * 2 + 0]); - __m512i sum1 = _mm512_load_si512(&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))); - } - } - return psqt; - - #elif defined(USE_AVX2) - - constexpr IndexType NumChunks = HalfDimensions / SimdWidth; - constexpr int Control = 0b11011000; - const __m256i Zero = _mm256_setzero_si256(); - - for (IndexType p = 0; p < 2; ++p) - { - const IndexType offset = HalfDimensions * p; - auto out = reinterpret_cast<__m256i*>(&output[offset]); - for (IndexType j = 0; j < NumChunks; ++j) - { - __m256i sum0 = _mm256_load_si256(&reinterpret_cast - (accumulation[perspectives[p]])[j * 2 + 0]); - __m256i sum1 = _mm256_load_si256(&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)); - } - } - return psqt; - - #elif defined(USE_SSE2) - - #ifdef USE_SSE41 - constexpr IndexType NumChunks = HalfDimensions / SimdWidth; - const __m128i Zero = _mm_setzero_si128(); - #else - constexpr IndexType NumChunks = HalfDimensions / SimdWidth; - const __m128i k0x80s = _mm_set1_epi8(-128); - #endif - - for (IndexType p = 0; p < 2; ++p) - { - const IndexType offset = HalfDimensions * p; - auto out = reinterpret_cast<__m128i*>(&output[offset]); - for (IndexType j = 0; j < NumChunks; ++j) - { - __m128i sum0 = _mm_load_si128(&reinterpret_cast - (accumulation[perspectives[p]])[j * 2 + 0]); - __m128i sum1 = _mm_load_si128(&reinterpret_cast - (accumulation[perspectives[p]])[j * 2 + 1]); - const __m128i packedbytes = _mm_packs_epi16(sum0, sum1); - - #ifdef USE_SSE41 - _mm_store_si128(&out[j], _mm_max_epi8(packedbytes, Zero)); - #else - _mm_store_si128(&out[j], _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)); - #endif - } - } - return psqt; - - #elif defined(USE_MMX) - - constexpr IndexType NumChunks = HalfDimensions / SimdWidth; - const __m64 k0x80s = _mm_set1_pi8(-128); - - for (IndexType p = 0; p < 2; ++p) - { - const IndexType offset = HalfDimensions * p; - auto out = reinterpret_cast<__m64*>(&output[offset]); - for (IndexType j = 0; j < NumChunks; ++j) - { - __m64 sum0 = *(&reinterpret_cast(accumulation[perspectives[p]])[j * 2 + 0]); - __m64 sum1 = *(&reinterpret_cast(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); - } - } - _mm_empty(); - return psqt; - - #elif defined(USE_NEON) - - constexpr IndexType NumChunks = HalfDimensions / (SimdWidth / 2); - const int8x8_t Zero = {0}; - - for (IndexType p = 0; p < 2; ++p) - { - const IndexType offset = HalfDimensions * p; - const auto out = reinterpret_cast(&output[offset]); - for (IndexType j = 0; j < NumChunks; ++j) - { - int16x8_t sum = reinterpret_cast(accumulation[perspectives[p]])[j]; - out[j] = vmax_s8(vqmovn_s16(sum), Zero); - } - } - return psqt; - - #else - - for (IndexType p = 0; p < 2; ++p) - { - const IndexType offset = HalfDimensions * p; - for (IndexType j = 0; j < HalfDimensions; ++j) - { - BiasType sum = accumulation[perspectives[p]][j]; - output[offset + j] = static_cast(std::max(0, std::min(127, sum))); - } - } - return psqt; - - #endif - - } // end of function transform() + update_accumulator(pos); + update_accumulator(pos); + 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[perspectives[0]][bucket] - psqtAccumulation[perspectives[1]][bucket]) + / 2; - private: - 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 = FeatureSet::refresh_cost(pos); - while (st->previous && !st->accumulator.computed[perspective]) - { - // 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.computed[perspective]) - { - if (next == nullptr) - return; - - // 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. - 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) - FeatureSet::append_changed_indices( - ksq, st2, perspective, removed[1], added[1]); - - // Mark the accumulators as computed. - next->accumulator.computed[perspective] = true; - pos.state()->accumulator.computed[perspective] = true; - // 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) + for (IndexType p = 0; p < 2; ++p) { - // Load accumulator - auto accTile = reinterpret_cast( - &st->accumulator.accumulation[perspective][j * TileHeight]); - for (IndexType k = 0; k < NumRegs; ++k) - acc[k] = vec_load(&accTile[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 = (HalfDimensions / 2) * p; + +#if defined(VECTOR) + + constexpr IndexType OutputChunkSize = MaxChunkSize; + static_assert((HalfDimensions / 2) % OutputChunkSize == 0); + constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize; + + vec_t Zero = vec_zero(); + vec_t One = vec_set_16(127); + + const vec_t* in0 = reinterpret_cast(&(accumulation[perspectives[p]][0])); + const vec_t* in1 = + reinterpret_cast(&(accumulation[perspectives[p]][HalfDimensions / 2])); + vec_t* out = reinterpret_cast(output + offset); + + for (IndexType j = 0; j < NumOutputChunks; j += 1) { - 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]); + const vec_t sum0a = vec_max_16(vec_min_16(in0[j * 2 + 0], One), Zero); + const vec_t sum0b = vec_max_16(vec_min_16(in0[j * 2 + 1], One), Zero); + const vec_t sum1a = vec_max_16(vec_min_16(in1[j * 2 + 0], One), Zero); + const vec_t sum1b = vec_max_16(vec_min_16(in1[j * 2 + 1], One), Zero); + + const vec_t pa = vec_mul_16(sum0a, sum1a); + const vec_t pb = vec_mul_16(sum0b, sum1b); + + out[j] = vec_msb_pack_16(pa, pb); } - // Difference calculation for the activated features - for (const auto index : added[i]) +#else + + for (IndexType j = 0; j < HalfDimensions / 2; ++j) { - 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]); + BiasType sum0 = accumulation[static_cast(perspectives[p])][j + 0]; + BiasType sum1 = + accumulation[static_cast(perspectives[p])][j + HalfDimensions / 2]; + sum0 = std::clamp(sum0, 0, 127); + sum1 = std::clamp(sum1, 0, 127); + output[offset + j] = static_cast(unsigned(sum0 * sum1) / 128); } - // Store accumulator - accTile = reinterpret_cast( - &states_to_update[i]->accumulator.accumulation[perspective][j * TileHeight]); - for (IndexType k = 0; k < NumRegs; ++k) - vec_store(&accTile[k], acc[k]); - } +#endif } - for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j) + return psqt; + } // end of function transform() + + void hint_common_access(const Position& pos) const { + hint_common_access_for_perspective(pos); + hint_common_access_for_perspective(pos); + } + + private: + template + [[nodiscard]] std::pair + try_find_computed_accumulator(const Position& pos) const { + // 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 = FeatureSet::refresh_cost(pos); + while (st->previous && !st->accumulator.computed[Perspective]) { - // 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]) + // 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; + } + return {st, next}; + } + + // NOTE: The parameter states_to_update is an array of position states, ending with nullptr. + // All states must be sequential, that is states_to_update[i] must either be reachable + // by repeatedly applying ->previous from states_to_update[i+1] or + // states_to_update[i] == nullptr. + // computed_st must be reachable by repeatedly applying ->previous on + // states_to_update[0], if not nullptr. + template + void update_accumulator_incremental(const Position& pos, + StateInfo* computed_st, + StateInfo* states_to_update[N]) const { + static_assert(N > 0); + assert(states_to_update[N - 1] == nullptr); + +#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 + + if (states_to_update[0] == nullptr) + return; + + // Update incrementally going back through states_to_update. + + // Gather all features to be updated. + const Square ksq = pos.square(Perspective); + + // 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. + FeatureSet::IndexList removed[N - 1], added[N - 1]; + + { + int i = + N + - 2; // last potential state to update. Skip last element because it must be nullptr. + while (states_to_update[i] == nullptr) + --i; + + StateInfo* st2 = states_to_update[i]; + + for (; i >= 0; --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]); + states_to_update[i]->accumulator.computed[Perspective] = true; + + const StateInfo* end_state = i == 0 ? computed_st : states_to_update[i - 1]; + + for (; st2 != end_state; st2 = st2->previous) + FeatureSet::append_changed_indices(ksq, st2->dirtyPiece, + removed[i], added[i]); } + } - // Difference calculation for the activated features - for (const auto index : added[i]) + StateInfo* st = computed_st; + + // Now update the accumulators listed in states_to_update[], where the last element is a sentinel. +#ifdef VECTOR + + if (states_to_update[1] == nullptr && (removed[0].size() == 1 || removed[0].size() == 2) + && added[0].size() == 1) + { + assert(states_to_update[0]); + + auto accIn = + reinterpret_cast(&st->accumulator.accumulation[Perspective][0]); + auto accOut = reinterpret_cast( + &states_to_update[0]->accumulator.accumulation[Perspective][0]); + + const IndexType offsetR0 = HalfDimensions * removed[0][0]; + auto columnR0 = reinterpret_cast(&weights[offsetR0]); + const IndexType offsetA = HalfDimensions * added[0][0]; + auto columnA = reinterpret_cast(&weights[offsetA]); + + if (removed[0].size() == 1) { - 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]); + for (IndexType k = 0; k < HalfDimensions * sizeof(std::int16_t) / sizeof(vec_t); + ++k) + accOut[k] = vec_add_16(vec_sub_16(accIn[k], columnR0[k]), columnA[k]); } + else + { + const IndexType offsetR1 = HalfDimensions * removed[0][1]; + auto columnR1 = reinterpret_cast(&weights[offsetR1]); - // 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]); - } + for (IndexType k = 0; k < HalfDimensions * sizeof(std::int16_t) / sizeof(vec_t); + ++k) + accOut[k] = vec_sub_16(vec_add_16(accIn[k], columnA[k]), + vec_add_16(columnR0[k], columnR1[k])); + } + + auto accPsqtIn = reinterpret_cast( + &st->accumulator.psqtAccumulation[Perspective][0]); + auto accPsqtOut = reinterpret_cast( + &states_to_update[0]->accumulator.psqtAccumulation[Perspective][0]); + + const IndexType offsetPsqtR0 = PSQTBuckets * removed[0][0]; + auto columnPsqtR0 = reinterpret_cast(&psqtWeights[offsetPsqtR0]); + const IndexType offsetPsqtA = PSQTBuckets * added[0][0]; + auto columnPsqtA = reinterpret_cast(&psqtWeights[offsetPsqtA]); + + if (removed[0].size() == 1) + { + for (std::size_t k = 0; k < PSQTBuckets * sizeof(std::int32_t) / sizeof(psqt_vec_t); + ++k) + accPsqtOut[k] = vec_add_psqt_32(vec_sub_psqt_32(accPsqtIn[k], columnPsqtR0[k]), + columnPsqtA[k]); + } + else + { + const IndexType offsetPsqtR1 = PSQTBuckets * removed[0][1]; + auto columnPsqtR1 = reinterpret_cast(&psqtWeights[offsetPsqtR1]); + + for (std::size_t k = 0; k < PSQTBuckets * sizeof(std::int32_t) / sizeof(psqt_vec_t); + ++k) + accPsqtOut[k] = + vec_sub_psqt_32(vec_add_psqt_32(accPsqtIn[k], columnPsqtA[k]), + vec_add_psqt_32(columnPsqtR0[k], columnPsqtR1[k])); + } } + else + { + for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j) + { + // Load accumulator + auto accTileIn = reinterpret_cast( + &st->accumulator.accumulation[Perspective][j * TileHeight]); + for (IndexType k = 0; k < NumRegs; ++k) + acc[k] = vec_load(&accTileIn[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 = 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 = 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 + auto accTileOut = reinterpret_cast( + &states_to_update[i]->accumulator.accumulation[Perspective][j * TileHeight]); + for (IndexType k = 0; k < NumRegs; ++k) + vec_store(&accTileOut[k], acc[k]); + } + } - #else + for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j) + { + // Load accumulator + auto accTilePsqtIn = reinterpret_cast( + &st->accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]); + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + psqt[k] = vec_load_psqt(&accTilePsqtIn[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 + auto accTilePsqtOut = reinterpret_cast( + &states_to_update[i] + ->accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]); + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + vec_store_psqt(&accTilePsqtOut[k], psqt[k]); + } + } + } +#else for (IndexType i = 0; states_to_update[i]; ++i) { - std::memcpy(states_to_update[i]->accumulator.accumulation[perspective], - st->accumulator.accumulation[perspective], - HalfDimensions * sizeof(BiasType)); + std::memcpy(states_to_update[i]->accumulator.accumulation[Perspective], + st->accumulator.accumulation[Perspective], + HalfDimensions * sizeof(BiasType)); - for (std::size_t k = 0; k < PSQTBuckets; ++k) - states_to_update[i]->accumulator.psqtAccumulation[perspective][k] = st->accumulator.psqtAccumulation[perspective][k]; + 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]; + st = states_to_update[i]; - // Difference calculation for the deactivated features - for (const auto index : removed[i]) - { - const IndexType offset = HalfDimensions * index; + // Difference calculation for the deactivated features + for (const auto index : removed[i]) + { + const IndexType offset = HalfDimensions * index; - for (IndexType j = 0; j < HalfDimensions; ++j) - st->accumulator.accumulation[perspective][j] -= weights[offset + j]; + for (IndexType j = 0; j < HalfDimensions; ++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]; - } + for (std::size_t k = 0; k < PSQTBuckets; ++k) + st->accumulator.psqtAccumulation[Perspective][k] -= + psqtWeights[index * PSQTBuckets + k]; + } - // Difference calculation for the activated features - for (const auto index : added[i]) - { - const IndexType offset = HalfDimensions * index; + // Difference calculation for the activated features + for (const auto index : added[i]) + { + const IndexType offset = HalfDimensions * index; - for (IndexType j = 0; j < HalfDimensions; ++j) - st->accumulator.accumulation[perspective][j] += weights[offset + j]; + for (IndexType j = 0; j < HalfDimensions; ++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]; - } + for (std::size_t k = 0; k < PSQTBuckets; ++k) + st->accumulator.psqtAccumulation[Perspective][k] += + psqtWeights[index * PSQTBuckets + k]; + } } - #endif - } - else - { - // Refresh the accumulator - auto& accumulator = pos.state()->accumulator; - accumulator.computed[perspective] = true; - IndexList active; - FeatureSet::append_active_indices(pos, perspective, active); +#endif + } - #ifdef VECTOR + template + void update_accumulator_refresh(const Position& pos) 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[NumRegs]; + psqt_vec_t psqt[NumPsqtRegs]; +#endif + + // Refresh the accumulator + // Could be extracted to a separate function because it's done in 2 places, + // but it's unclear if compilers would correctly handle register allocation. + auto& accumulator = pos.state()->accumulator; + accumulator.computed[Perspective] = true; + FeatureSet::IndexList active; + FeatureSet::append_active_indices(pos, active); + +#ifdef VECTOR for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j) { - auto biasesTile = reinterpret_cast( - &biases[j * TileHeight]); - for (IndexType k = 0; k < NumRegs; ++k) - acc[k] = biasesTile[k]; - - for (const auto index : active) - { - const IndexType offset = HalfDimensions * index + j * TileHeight; - auto column = reinterpret_cast(&weights[offset]); - - for (unsigned k = 0; k < NumRegs; ++k) - acc[k] = vec_add_16(acc[k], column[k]); - } - - auto accTile = reinterpret_cast( - &accumulator.accumulation[perspective][j * TileHeight]); - for (unsigned k = 0; k < NumRegs; k++) - vec_store(&accTile[k], acc[k]); + auto biasesTile = reinterpret_cast(&biases[j * TileHeight]); + for (IndexType k = 0; k < NumRegs; ++k) + acc[k] = biasesTile[k]; + + for (const auto index : active) + { + const IndexType offset = HalfDimensions * index + j * TileHeight; + auto column = reinterpret_cast(&weights[offset]); + + for (unsigned k = 0; k < NumRegs; ++k) + acc[k] = vec_add_16(acc[k], column[k]); + } + + auto accTile = + reinterpret_cast(&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 (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 (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]); - } + 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]); + 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[perspective], biases, - HalfDimensions * sizeof(BiasType)); +#else + std::memcpy(accumulator.accumulation[Perspective], biases, + HalfDimensions * sizeof(BiasType)); for (std::size_t k = 0; k < PSQTBuckets; ++k) - accumulator.psqtAccumulation[perspective][k] = 0; + accumulator.psqtAccumulation[Perspective][k] = 0; for (const auto index : active) { - const IndexType offset = HalfDimensions * index; + const IndexType offset = HalfDimensions * index; + + for (IndexType j = 0; j < HalfDimensions; ++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 + } + + template + void hint_common_access_for_perspective(const Position& pos) const { + + // Works like update_accumulator, but performs less work. + // Updates ONLY the accumulator for pos. - for (IndexType j = 0; j < HalfDimensions; ++j) - accumulator.accumulation[perspective][j] += weights[offset + j]; + // Look for a usable accumulator of an earlier position. We keep track + // of the estimated gain in terms of features to be added/subtracted. + // Fast early exit. + if (pos.state()->accumulator.computed[Perspective]) + return; - for (std::size_t k = 0; k < PSQTBuckets; ++k) - accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k]; + auto [oldest_st, _] = try_find_computed_accumulator(pos); + + if (oldest_st->accumulator.computed[Perspective]) + { + // Only update current position accumulator to minimize work. + StateInfo* states_to_update[2] = {pos.state(), nullptr}; + update_accumulator_incremental(pos, oldest_st, states_to_update); + } + else + { + update_accumulator_refresh(pos); } - #endif - } + } + + template + void update_accumulator(const Position& pos) const { - #if defined(USE_MMX) - _mm_empty(); - #endif + auto [oldest_st, next] = try_find_computed_accumulator(pos); + + if (oldest_st->accumulator.computed[Perspective]) + { + if (next == nullptr) + return; + + // Now update the accumulators listed in states_to_update[], where the last element is a sentinel. + // Currently we update 2 accumulators. + // 1. for the current position + // 2. the next accumulator after the computed one + // The heuristic may change in the future. + StateInfo* states_to_update[3] = {next, next == pos.state() ? nullptr : pos.state(), + nullptr}; + + update_accumulator_incremental(pos, oldest_st, states_to_update); + } + else + { + update_accumulator_refresh(pos); + } } alignas(CacheLineSize) BiasType biases[HalfDimensions]; alignas(CacheLineSize) WeightType weights[HalfDimensions * InputDimensions]; alignas(CacheLineSize) PSQTWeightType psqtWeights[InputDimensions * PSQTBuckets]; - }; +}; } // namespace Stockfish::Eval::NNUE -#endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED +#endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED