git.sesse.net Git - stockfish/blob - src/nnue/nnue_feature_transformer.h

   1 /*
   2   Stockfish, a UCI chess playing engine derived from Glaurung 2.1
   3   Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
   4
   5   Stockfish is free software: you can redistribute it and/or modify
   6   it under the terms of the GNU General Public License as published by
   7   the Free Software Foundation, either version 3 of the License, or
   8   (at your option) any later version.
   9
  10   Stockfish is distributed in the hope that it will be useful,
  11   but WITHOUT ANY WARRANTY; without even the implied warranty of
  12   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13   GNU General Public License for more details.
  14
  15   You should have received a copy of the GNU General Public License
  16   along with this program.  If not, see <http://www.gnu.org/licenses/>.
  17 */
  18
  19 // A class that converts the input features of the NNUE evaluation function
  20
  21 #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED
  22 #define NNUE_FEATURE_TRANSFORMER_H_INCLUDED
  23
  24 #include "nnue_common.h"
  25 #include "nnue_architecture.h"
  26 #include "features/index_list.h"
  27
  28 #include <cstring> // std::memset()
  29
  30 namespace Eval::NNUE {
  31
  32   // Input feature converter
  33   class FeatureTransformer {
  34
  35    private:
  36     // Number of output dimensions for one side
  37     static constexpr IndexType kHalfDimensions = kTransformedFeatureDimensions;
  38
  39    public:
  40     // Output type
  41     using OutputType = TransformedFeatureType;
  42
  43     // Number of input/output dimensions
  44     static constexpr IndexType kInputDimensions = RawFeatures::kDimensions;
  45     static constexpr IndexType kOutputDimensions = kHalfDimensions * 2;
  46
  47     // Size of forward propagation buffer
  48     static constexpr std::size_t kBufferSize =
  49         kOutputDimensions * sizeof(OutputType);
  50
  51     // Hash value embedded in the evaluation file
  52     static constexpr std::uint32_t GetHashValue() {
  53       return RawFeatures::kHashValue ^ kOutputDimensions;
  54     }
  55
  56     // Read network parameters
  57     bool ReadParameters(std::istream& stream) {
  58       stream.read(reinterpret_cast<char*>(biases_),
  59                   kHalfDimensions * sizeof(BiasType));
  60       stream.read(reinterpret_cast<char*>(weights_),
  61                   kHalfDimensions * kInputDimensions * sizeof(WeightType));
  62       return !stream.fail();
  63     }
  64
  65     // Proceed with the difference calculation if possible
  66     bool UpdateAccumulatorIfPossible(const Position& pos) const {
  67       const auto now = pos.state();
  68       if (now->accumulator.computed_accumulation) {
  69         return true;
  70       }
  71       const auto prev = now->previous;
  72       if (prev && prev->accumulator.computed_accumulation) {
  73         UpdateAccumulator(pos);
  74         return true;
  75       }
  76       return false;
  77     }
  78
  79     // Convert input features
  80     void Transform(const Position& pos, OutputType* output, bool refresh) const {
  81       if (refresh || !UpdateAccumulatorIfPossible(pos)) {
  82         RefreshAccumulator(pos);
  83       }
  84       const auto& accumulation = pos.state()->accumulator.accumulation;
  85
  86   #if defined(USE_AVX2)
  87       constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
  88       constexpr int kControl = 0b11011000;
  89       const __m256i kZero = _mm256_setzero_si256();
  90
  91   #elif defined(USE_SSSE3)
  92       constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
  93
  94   #ifdef USE_SSE41
  95       const __m128i kZero = _mm_setzero_si128();
  96   #else
  97       const __m128i k0x80s = _mm_set1_epi8(-128);
  98   #endif
  99
 100   #elif defined(USE_NEON)
 101       constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 102       const int8x8_t kZero = {0};
 103   #endif
 104
 105       const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
 106       for (IndexType p = 0; p < 2; ++p) {
 107         const IndexType offset = kHalfDimensions * p;
 108
 109   #if defined(USE_AVX2)
 110         auto out = reinterpret_cast<__m256i*>(&output[offset]);
 111         for (IndexType j = 0; j < kNumChunks; ++j) {
 112           __m256i sum0 =
 113             _mm256_load_si256(&reinterpret_cast<const __m256i*>(
 114               accumulation[perspectives[p]][0])[j * 2 + 0]);
 115           __m256i sum1 =
 116             _mm256_load_si256(&reinterpret_cast<const __m256i*>(
 117               accumulation[perspectives[p]][0])[j * 2 + 1]);
 118           _mm256_store_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
 119               _mm256_packs_epi16(sum0, sum1), kZero), kControl));
 120         }
 121
 122   #elif defined(USE_SSSE3)
 123         auto out = reinterpret_cast<__m128i*>(&output[offset]);
 124         for (IndexType j = 0; j < kNumChunks; ++j) {
 125           __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
 126               accumulation[perspectives[p]][0])[j * 2 + 0]);
 127           __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
 128               accumulation[perspectives[p]][0])[j * 2 + 1]);
 129       const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
 130
 131           _mm_store_si128(&out[j],
 132
 133   #ifdef USE_SSE41
 134             _mm_max_epi8(packedbytes, kZero)
 135   #else
 136             _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
 137   #endif
 138
 139           );
 140         }
 141
 142   #elif defined(USE_NEON)
 143         const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
 144         for (IndexType j = 0; j < kNumChunks; ++j) {
 145           int16x8_t sum = reinterpret_cast<const int16x8_t*>(
 146               accumulation[perspectives[p]][0])[j];
 147           out[j] = vmax_s8(vqmovn_s16(sum), kZero);
 148         }
 149
 150   #else
 151         for (IndexType j = 0; j < kHalfDimensions; ++j) {
 152           BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
 153           output[offset + j] = static_cast<OutputType>(
 154               std::max<int>(0, std::min<int>(127, sum)));
 155         }
 156   #endif
 157
 158       }
 159     }
 160
 161    private:
 162     // Calculate cumulative value without using difference calculation
 163     void RefreshAccumulator(const Position& pos) const {
 164       auto& accumulator = pos.state()->accumulator;
 165       IndexType i = 0;
 166       Features::IndexList active_indices[2];
 167       RawFeatures::AppendActiveIndices(pos, kRefreshTriggers[i],
 168                                        active_indices);
 169       for (Color perspective : { WHITE, BLACK }) {
 170         std::memcpy(accumulator.accumulation[perspective][i], biases_,
 171                    kHalfDimensions * sizeof(BiasType));
 172         for (const auto index : active_indices[perspective]) {
 173           const IndexType offset = kHalfDimensions * index;
 174
 175   #if defined(USE_AVX2)
 176           auto accumulation = reinterpret_cast<__m256i*>(
 177               &accumulator.accumulation[perspective][i][0]);
 178           auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 179           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 180           for (IndexType j = 0; j < kNumChunks; ++j) {
 181             accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
 182           }
 183
 184   #elif defined(USE_SSE2)
 185           auto accumulation = reinterpret_cast<__m128i*>(
 186               &accumulator.accumulation[perspective][i][0]);
 187           auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 188           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 189           for (IndexType j = 0; j < kNumChunks; ++j) {
 190             accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
 191           }
 192
 193   #elif defined(USE_NEON)
 194           auto accumulation = reinterpret_cast<int16x8_t*>(
 195               &accumulator.accumulation[perspective][i][0]);
 196           auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 197           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 198           for (IndexType j = 0; j < kNumChunks; ++j) {
 199             accumulation[j] = vaddq_s16(accumulation[j], column[j]);
 200           }
 201
 202   #else
 203           for (IndexType j = 0; j < kHalfDimensions; ++j) {
 204             accumulator.accumulation[perspective][i][j] += weights_[offset + j];
 205           }
 206   #endif
 207
 208         }
 209       }
 210
 211       accumulator.computed_accumulation = true;
 212       accumulator.computed_score = false;
 213     }
 214
 215     // Calculate cumulative value using difference calculation
 216     void UpdateAccumulator(const Position& pos) const {
 217       const auto prev_accumulator = pos.state()->previous->accumulator;
 218       auto& accumulator = pos.state()->accumulator;
 219       IndexType i = 0;
 220       Features::IndexList removed_indices[2], added_indices[2];
 221       bool reset[2];
 222       RawFeatures::AppendChangedIndices(pos, kRefreshTriggers[i],
 223                                         removed_indices, added_indices, reset);
 224       for (Color perspective : { WHITE, BLACK }) {
 225
 226   #if defined(USE_AVX2)
 227         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 228         auto accumulation = reinterpret_cast<__m256i*>(
 229             &accumulator.accumulation[perspective][i][0]);
 230
 231   #elif defined(USE_SSE2)
 232         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 233         auto accumulation = reinterpret_cast<__m128i*>(
 234             &accumulator.accumulation[perspective][i][0]);
 235
 236   #elif defined(USE_NEON)
 237         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 238         auto accumulation = reinterpret_cast<int16x8_t*>(
 239             &accumulator.accumulation[perspective][i][0]);
 240   #endif
 241
 242         if (reset[perspective]) {
 243           std::memcpy(accumulator.accumulation[perspective][i], biases_,
 244                       kHalfDimensions * sizeof(BiasType));
 245         } else {
 246           std::memcpy(accumulator.accumulation[perspective][i],
 247                       prev_accumulator.accumulation[perspective][i],
 248                       kHalfDimensions * sizeof(BiasType));
 249           // Difference calculation for the deactivated features
 250           for (const auto index : removed_indices[perspective]) {
 251             const IndexType offset = kHalfDimensions * index;
 252
 253   #if defined(USE_AVX2)
 254             auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 255             for (IndexType j = 0; j < kNumChunks; ++j) {
 256               accumulation[j] = _mm256_sub_epi16(accumulation[j], column[j]);
 257             }
 258
 259   #elif defined(USE_SSE2)
 260             auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 261             for (IndexType j = 0; j < kNumChunks; ++j) {
 262               accumulation[j] = _mm_sub_epi16(accumulation[j], column[j]);
 263             }
 264
 265   #elif defined(USE_NEON)
 266             auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 267             for (IndexType j = 0; j < kNumChunks; ++j) {
 268               accumulation[j] = vsubq_s16(accumulation[j], column[j]);
 269             }
 270
 271   #else
 272             for (IndexType j = 0; j < kHalfDimensions; ++j) {
 273               accumulator.accumulation[perspective][i][j] -=
 274                   weights_[offset + j];
 275             }
 276   #endif
 277
 278           }
 279         }
 280         { // Difference calculation for the activated features
 281           for (const auto index : added_indices[perspective]) {
 282             const IndexType offset = kHalfDimensions * index;
 283
 284   #if defined(USE_AVX2)
 285             auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 286             for (IndexType j = 0; j < kNumChunks; ++j) {
 287               accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
 288             }
 289
 290   #elif defined(USE_SSE2)
 291             auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 292             for (IndexType j = 0; j < kNumChunks; ++j) {
 293               accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
 294             }
 295
 296   #elif defined(USE_NEON)
 297             auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 298             for (IndexType j = 0; j < kNumChunks; ++j) {
 299               accumulation[j] = vaddq_s16(accumulation[j], column[j]);
 300             }
 301
 302   #else
 303             for (IndexType j = 0; j < kHalfDimensions; ++j) {
 304               accumulator.accumulation[perspective][i][j] +=
 305                   weights_[offset + j];
 306             }
 307   #endif
 308
 309           }
 310         }
 311       }
 312
 313       accumulator.computed_accumulation = true;
 314       accumulator.computed_score = false;
 315     }
 316
 317     using BiasType = std::int16_t;
 318     using WeightType = std::int16_t;
 319
 320     alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
 321     alignas(kCacheLineSize)
 322         WeightType weights_[kHalfDimensions * kInputDimensions];
 323   };
 324
 325 }  // namespace Eval::NNUE
 326
 327 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED