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
 114   #if defined(__MINGW32__) || defined(__MINGW64__)
 115             // HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary
 116             //       compiled with g++ in MSYS2 crashes here because the output memory is not aligned
 117             //       even though alignas is specified.
 118             _mm256_loadu_si256
 119   #else
 120             _mm256_load_si256
 121   #endif
 122
 123             (&reinterpret_cast<const __m256i*>(
 124               accumulation[perspectives[p]][0])[j * 2 + 0]);
 125           __m256i sum1 =
 126
 127   #if defined(__MINGW32__) || defined(__MINGW64__)
 128             _mm256_loadu_si256
 129   #else
 130             _mm256_load_si256
 131   #endif
 132
 133             (&reinterpret_cast<const __m256i*>(
 134               accumulation[perspectives[p]][0])[j * 2 + 1]);
 135
 136   #if defined(__MINGW32__) || defined(__MINGW64__)
 137           _mm256_storeu_si256
 138   #else
 139           _mm256_store_si256
 140   #endif
 141
 142           (&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
 143               _mm256_packs_epi16(sum0, sum1), kZero), kControl));
 144         }
 145
 146   #elif defined(USE_SSSE3)
 147         auto out = reinterpret_cast<__m128i*>(&output[offset]);
 148         for (IndexType j = 0; j < kNumChunks; ++j) {
 149           __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
 150               accumulation[perspectives[p]][0])[j * 2 + 0]);
 151           __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
 152               accumulation[perspectives[p]][0])[j * 2 + 1]);
 153       const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
 154
 155           _mm_store_si128(&out[j],
 156
 157   #ifdef USE_SSE41
 158             _mm_max_epi8(packedbytes, kZero)
 159   #else
 160             _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
 161   #endif
 162
 163           );
 164         }
 165
 166   #elif defined(USE_NEON)
 167         const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
 168         for (IndexType j = 0; j < kNumChunks; ++j) {
 169           int16x8_t sum = reinterpret_cast<const int16x8_t*>(
 170               accumulation[perspectives[p]][0])[j];
 171           out[j] = vmax_s8(vqmovn_s16(sum), kZero);
 172         }
 173
 174   #else
 175         for (IndexType j = 0; j < kHalfDimensions; ++j) {
 176           BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
 177           output[offset + j] = static_cast<OutputType>(
 178               std::max<int>(0, std::min<int>(127, sum)));
 179         }
 180   #endif
 181
 182       }
 183     }
 184
 185    private:
 186     // Calculate cumulative value without using difference calculation
 187     void RefreshAccumulator(const Position& pos) const {
 188       auto& accumulator = pos.state()->accumulator;
 189       IndexType i = 0;
 190       Features::IndexList active_indices[2];
 191       RawFeatures::AppendActiveIndices(pos, kRefreshTriggers[i],
 192                                        active_indices);
 193       for (Color perspective : { WHITE, BLACK }) {
 194         std::memcpy(accumulator.accumulation[perspective][i], biases_,
 195                    kHalfDimensions * sizeof(BiasType));
 196         for (const auto index : active_indices[perspective]) {
 197           const IndexType offset = kHalfDimensions * index;
 198
 199   #if defined(USE_AVX2)
 200           auto accumulation = reinterpret_cast<__m256i*>(
 201               &accumulator.accumulation[perspective][i][0]);
 202           auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 203           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 204           for (IndexType j = 0; j < kNumChunks; ++j) {
 205   #if defined(__MINGW32__) || defined(__MINGW64__)
 206             _mm256_storeu_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadu_si256(&accumulation[j]), column[j]));
 207   #else
 208             accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
 209   #endif
 210           }
 211
 212   #elif defined(USE_SSE2)
 213           auto accumulation = reinterpret_cast<__m128i*>(
 214               &accumulator.accumulation[perspective][i][0]);
 215           auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 216           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 217           for (IndexType j = 0; j < kNumChunks; ++j) {
 218             accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
 219           }
 220
 221   #elif defined(USE_NEON)
 222           auto accumulation = reinterpret_cast<int16x8_t*>(
 223               &accumulator.accumulation[perspective][i][0]);
 224           auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 225           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 226           for (IndexType j = 0; j < kNumChunks; ++j) {
 227             accumulation[j] = vaddq_s16(accumulation[j], column[j]);
 228           }
 229
 230   #else
 231           for (IndexType j = 0; j < kHalfDimensions; ++j) {
 232             accumulator.accumulation[perspective][i][j] += weights_[offset + j];
 233           }
 234   #endif
 235
 236         }
 237       }
 238
 239       accumulator.computed_accumulation = true;
 240       accumulator.computed_score = false;
 241     }
 242
 243     // Calculate cumulative value using difference calculation
 244     void UpdateAccumulator(const Position& pos) const {
 245       const auto prev_accumulator = pos.state()->previous->accumulator;
 246       auto& accumulator = pos.state()->accumulator;
 247       IndexType i = 0;
 248       Features::IndexList removed_indices[2], added_indices[2];
 249       bool reset[2];
 250       RawFeatures::AppendChangedIndices(pos, kRefreshTriggers[i],
 251                                         removed_indices, added_indices, reset);
 252       for (Color perspective : { WHITE, BLACK }) {
 253
 254   #if defined(USE_AVX2)
 255         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 256         auto accumulation = reinterpret_cast<__m256i*>(
 257             &accumulator.accumulation[perspective][i][0]);
 258
 259   #elif defined(USE_SSE2)
 260         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 261         auto accumulation = reinterpret_cast<__m128i*>(
 262             &accumulator.accumulation[perspective][i][0]);
 263
 264   #elif defined(USE_NEON)
 265         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 266         auto accumulation = reinterpret_cast<int16x8_t*>(
 267             &accumulator.accumulation[perspective][i][0]);
 268   #endif
 269
 270         if (reset[perspective]) {
 271           std::memcpy(accumulator.accumulation[perspective][i], biases_,
 272                       kHalfDimensions * sizeof(BiasType));
 273         } else {
 274           std::memcpy(accumulator.accumulation[perspective][i],
 275                       prev_accumulator.accumulation[perspective][i],
 276                       kHalfDimensions * sizeof(BiasType));
 277           // Difference calculation for the deactivated features
 278           for (const auto index : removed_indices[perspective]) {
 279             const IndexType offset = kHalfDimensions * index;
 280
 281   #if defined(USE_AVX2)
 282             auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 283             for (IndexType j = 0; j < kNumChunks; ++j) {
 284               accumulation[j] = _mm256_sub_epi16(accumulation[j], column[j]);
 285             }
 286
 287   #elif defined(USE_SSE2)
 288             auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 289             for (IndexType j = 0; j < kNumChunks; ++j) {
 290               accumulation[j] = _mm_sub_epi16(accumulation[j], column[j]);
 291             }
 292
 293   #elif defined(USE_NEON)
 294             auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 295             for (IndexType j = 0; j < kNumChunks; ++j) {
 296               accumulation[j] = vsubq_s16(accumulation[j], column[j]);
 297             }
 298
 299   #else
 300             for (IndexType j = 0; j < kHalfDimensions; ++j) {
 301               accumulator.accumulation[perspective][i][j] -=
 302                   weights_[offset + j];
 303             }
 304   #endif
 305
 306           }
 307         }
 308         { // Difference calculation for the activated features
 309           for (const auto index : added_indices[perspective]) {
 310             const IndexType offset = kHalfDimensions * index;
 311
 312   #if defined(USE_AVX2)
 313             auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 314             for (IndexType j = 0; j < kNumChunks; ++j) {
 315               accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
 316             }
 317
 318   #elif defined(USE_SSE2)
 319             auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 320             for (IndexType j = 0; j < kNumChunks; ++j) {
 321               accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
 322             }
 323
 324   #elif defined(USE_NEON)
 325             auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 326             for (IndexType j = 0; j < kNumChunks; ++j) {
 327               accumulation[j] = vaddq_s16(accumulation[j], column[j]);
 328             }
 329
 330   #else
 331             for (IndexType j = 0; j < kHalfDimensions; ++j) {
 332               accumulator.accumulation[perspective][i][j] +=
 333                   weights_[offset + j];
 334             }
 335   #endif
 336
 337           }
 338         }
 339       }
 340
 341       accumulator.computed_accumulation = true;
 342       accumulator.computed_score = false;
 343     }
 344
 345     using BiasType = std::int16_t;
 346     using WeightType = std::int16_t;
 347
 348     alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
 349     alignas(kCacheLineSize)
 350         WeightType weights_[kHalfDimensions * kInputDimensions];
 351   };
 352
 353 }  // namespace Eval::NNUE
 354
 355 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED