X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fnnue%2Flayers%2Faffine_transform.h;h=94d0b5a9494644e574cd111104943d18667c9196;hp=b585bc87819d23c808ce66a472c4ffba59e47072;hb=701b2427bd84d112376ce858b66befc5b66c4bb2;hpb=84f3e867903f62480c33243dd0ecbffd342796fc diff --git a/src/nnue/layers/affine_transform.h b/src/nnue/layers/affine_transform.h index b585bc87..94d0b5a9 100644 --- a/src/nnue/layers/affine_transform.h +++ b/src/nnue/layers/affine_transform.h @@ -62,11 +62,10 @@ 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) + weights_[i] = read_little_endian(stream); return !stream.fail(); } @@ -79,19 +78,32 @@ namespace Eval::NNUE::Layers { #if defined(USE_AVX512) constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2); - const __m512i kOnes = _mm512_set1_epi16(1); const auto input_vector = reinterpret_cast(input); + #if !defined(USE_VNNI) + const __m512i kOnes = _mm512_set1_epi16(1); + #endif #elif defined(USE_AVX2) constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; - const __m256i kOnes = _mm256_set1_epi16(1); const auto input_vector = reinterpret_cast(input); + #if !defined(USE_VNNI) + const __m256i kOnes = _mm256_set1_epi16(1); + #endif - #elif defined(USE_SSSE3) + #elif 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) + constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; + const __m64 kZeros = _mm_setzero_si64(); + const auto input_vector = reinterpret_cast(input); + #elif defined(USE_NEON) constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth; const auto input_vector = reinterpret_cast(input); @@ -104,79 +116,115 @@ namespace Eval::NNUE::Layers { __m512i sum = _mm512_setzero_si512(); const auto row = reinterpret_cast(&weights_[offset]); for (IndexType j = 0; j < kNumChunks; ++j) { - - #if defined(__MINGW32__) || defined(__MINGW64__) - __m512i product = _mm512_maddubs_epi16(_mm512_loadu_si512(&input_vector[j]), _mm512_load_si512(&row[j])); + #if defined(USE_VNNI) + sum = _mm512_dpbusd_epi32(sum, _mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j])); #else - __m512i product = _mm512_maddubs_epi16(_mm512_load_si512(&input_vector[j]), _mm512_load_si512(&row[j])); - #endif - + __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); + #endif } - output[i] = _mm512_reduce_add_epi32(sum) + biases_[i]; // 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 iv_256 = reinterpret_cast(input); - const auto row_256 = reinterpret_cast(&weights_[offset]); - int j = kNumChunks * 2; - - #if defined(__MINGW32__) || defined(__MINGW64__) // See HACK comment below in AVX2. - __m256i sum256 = _mm256_maddubs_epi16(_mm256_loadu_si256(&iv_256[j]), _mm256_load_si256(&row_256[j])); + const auto iv256 = reinterpret_cast(&input_vector[kNumChunks]); + const auto row256 = reinterpret_cast(&row[kNumChunks]); + #if defined(USE_VNNI) + __m256i product256 = _mm256_dpbusd_epi32( + _mm512_castsi512_si256(sum), _mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0])); + sum = _mm512_inserti32x8(sum, product256, 0); #else - __m256i sum256 = _mm256_maddubs_epi16(_mm256_load_si256(&iv_256[j]), _mm256_load_si256(&row_256[j])); + __m256i product256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0])); + sum = _mm512_add_epi32(sum, _mm512_cvtepi16_epi32(product256)); #endif - - sum256 = _mm256_madd_epi16(sum256, _mm256_set1_epi16(1)); - sum256 = _mm256_hadd_epi32(sum256, sum256); - sum256 = _mm256_hadd_epi32(sum256, sum256); - const __m128i lo = _mm256_extracti128_si256(sum256, 0); - const __m128i hi = _mm256_extracti128_si256(sum256, 1); - output[i] += _mm_cvtsi128_si32(lo) + _mm_cvtsi128_si32(hi); } + 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( - - #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 + #if defined(USE_VNNI) + sum = _mm256_dpbusd_epi32(sum, _mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j])); #else - _mm256_load_si256 - #endif - - (&input_vector[j]), _mm256_load_si256(&row[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); + #endif } - sum = _mm256_hadd_epi32(sum, sum); - sum = _mm256_hadd_epi32(sum, sum); - const __m128i lo = _mm256_extracti128_si256(sum, 0); - const __m128i hi = _mm256_extracti128_si256(sum, 1); - output[i] = _mm_cvtsi128_si32(lo) + _mm_cvtsi128_si32(hi) + biases_[i]; + __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_cvtsi32_si128(biases_[i]); + __m128i sum = _mm_setzero_si128(); const auto row = reinterpret_cast(&weights_[offset]); - for (IndexType j = 0; j < kNumChunks; ++j) { - __m128i product = _mm_maddubs_epi16( - _mm_load_si128(&input_vector[j]), _mm_load_si128(&row[j])); + 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_hadd_epi32(sum, sum); - sum = _mm_hadd_epi32(sum, sum); + 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) + __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_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); + __m128i product_hi = _mm_madd_epi16(extended_row_hi, extended_input_hi); + sum_lo = _mm_add_epi32(sum_lo, product_lo); + sum_hi = _mm_add_epi32(sum_hi, product_hi); + } + __m128i sum = _mm_add_epi32(sum_lo, sum_hi); + __m128i sum_high_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2)); + sum = _mm_add_epi32(sum, sum_high_64); + __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2)); + sum = _mm_add_epi32(sum, sum_second_32); output[i] = _mm_cvtsi128_si32(sum); + #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_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); + __m64 product_hi = _mm_madd_pi16(extended_row_hi, extended_input_hi); + sum_lo = _mm_add_pi32(sum_lo, product_lo); + sum_hi = _mm_add_pi32(sum_hi, product_hi); + } + __m64 sum = _mm_add_pi32(sum_lo, sum_hi); + sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum)); + output[i] = _mm_cvtsi64_si32(sum); + #elif defined(USE_NEON) int32x4_t sum = {biases_[i]}; const auto row = reinterpret_cast(&weights_[offset]); @@ -196,6 +244,9 @@ namespace Eval::NNUE::Layers { #endif } + #if defined(USE_MMX) + _mm_empty(); + #endif return output; }