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
  54       return RawFeatures::kHashValue ^ kOutputDimensions;
  55     }
  56
  57     // Read network parameters
  58     bool ReadParameters(std::istream& stream) {
  59
  60       for (std::size_t i = 0; i < kHalfDimensions; ++i)
  61         biases_[i] = read_little_endian<BiasType>(stream);
  62       for (std::size_t i = 0; i < kHalfDimensions * kInputDimensions; ++i)
  63         weights_[i] = read_little_endian<WeightType>(stream);
  64       return !stream.fail();
  65     }
  66
  67     // Proceed with the difference calculation if possible
  68     bool UpdateAccumulatorIfPossible(const Position& pos) const {
  69
  70       const auto now = pos.state();
  71       if (now->accumulator.computed_accumulation)
  72         return true;
  73
  74       const auto prev = now->previous;
  75       if (prev && prev->accumulator.computed_accumulation) {
  76         UpdateAccumulator(pos);
  77         return true;
  78       }
  79
  80       return false;
  81     }
  82
  83     // Convert input features
  84     void Transform(const Position& pos, OutputType* output) const {
  85
  86       if (!UpdateAccumulatorIfPossible(pos))
  87         RefreshAccumulator(pos);
  88
  89       const auto& accumulation = pos.state()->accumulator.accumulation;
  90
  91   #if defined(USE_AVX2)
  92       constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
  93       constexpr int kControl = 0b11011000;
  94       const __m256i kZero = _mm256_setzero_si256();
  95
  96   #elif defined(USE_SSE2)
  97       constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
  98
  99   #ifdef USE_SSE41
 100       const __m128i kZero = _mm_setzero_si128();
 101   #else
 102       const __m128i k0x80s = _mm_set1_epi8(-128);
 103   #endif
 104
 105   #elif defined(USE_MMX)
 106       constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
 107       const __m64 k0x80s = _mm_set1_pi8(-128);
 108
 109   #elif defined(USE_NEON)
 110       constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 111       const int8x8_t kZero = {0};
 112   #endif
 113
 114       const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
 115       for (IndexType p = 0; p < 2; ++p) {
 116         const IndexType offset = kHalfDimensions * p;
 117
 118   #if defined(USE_AVX2)
 119         auto out = reinterpret_cast<__m256i*>(&output[offset]);
 120         for (IndexType j = 0; j < kNumChunks; ++j) {
 121           __m256i sum0 = _mm256_loadA_si256(
 122               &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 0]);
 123           __m256i sum1 = _mm256_loadA_si256(
 124             &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 1]);
 125           _mm256_storeA_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
 126               _mm256_packs_epi16(sum0, sum1), kZero), kControl));
 127         }
 128
 129   #elif defined(USE_SSE2)
 130         auto out = reinterpret_cast<__m128i*>(&output[offset]);
 131         for (IndexType j = 0; j < kNumChunks; ++j) {
 132           __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
 133               accumulation[perspectives[p]][0])[j * 2 + 0]);
 134           __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
 135               accumulation[perspectives[p]][0])[j * 2 + 1]);
 136       const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
 137
 138           _mm_store_si128(&out[j],
 139
 140   #ifdef USE_SSE41
 141             _mm_max_epi8(packedbytes, kZero)
 142   #else
 143             _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
 144   #endif
 145
 146           );
 147         }
 148
 149   #elif defined(USE_MMX)
 150         auto out = reinterpret_cast<__m64*>(&output[offset]);
 151         for (IndexType j = 0; j < kNumChunks; ++j) {
 152           __m64 sum0 = *(&reinterpret_cast<const __m64*>(
 153               accumulation[perspectives[p]][0])[j * 2 + 0]);
 154           __m64 sum1 = *(&reinterpret_cast<const __m64*>(
 155               accumulation[perspectives[p]][0])[j * 2 + 1]);
 156           const __m64 packedbytes = _mm_packs_pi16(sum0, sum1);
 157           out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
 158         }
 159
 160   #elif defined(USE_NEON)
 161         const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
 162         for (IndexType j = 0; j < kNumChunks; ++j) {
 163           int16x8_t sum = reinterpret_cast<const int16x8_t*>(
 164               accumulation[perspectives[p]][0])[j];
 165           out[j] = vmax_s8(vqmovn_s16(sum), kZero);
 166         }
 167
 168   #else
 169         for (IndexType j = 0; j < kHalfDimensions; ++j) {
 170           BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
 171           output[offset + j] = static_cast<OutputType>(
 172               std::max<int>(0, std::min<int>(127, sum)));
 173         }
 174   #endif
 175
 176       }
 177   #if defined(USE_MMX)
 178       _mm_empty();
 179   #endif
 180     }
 181
 182    private:
 183     // Calculate cumulative value without using difference calculation
 184     void RefreshAccumulator(const Position& pos) const {
 185
 186       auto& accumulator = pos.state()->accumulator;
 187       IndexType i = 0;
 188       Features::IndexList active_indices[2];
 189       RawFeatures::AppendActiveIndices(pos, kRefreshTriggers[i],
 190                                        active_indices);
 191       for (Color perspective : { WHITE, BLACK }) {
 192         std::memcpy(accumulator.accumulation[perspective][i], biases_,
 193                    kHalfDimensions * sizeof(BiasType));
 194         for (const auto index : active_indices[perspective]) {
 195           const IndexType offset = kHalfDimensions * index;
 196   #if defined(USE_AVX512)
 197           auto accumulation = reinterpret_cast<__m512i*>(
 198               &accumulator.accumulation[perspective][i][0]);
 199           auto column = reinterpret_cast<const __m512i*>(&weights_[offset]);
 200           constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
 201           for (IndexType j = 0; j < kNumChunks; ++j)
 202             _mm512_storeA_si512(&accumulation[j], _mm512_add_epi16(_mm512_loadA_si512(&accumulation[j]), column[j]));
 203
 204   #elif defined(USE_AVX2)
 205           auto accumulation = reinterpret_cast<__m256i*>(
 206               &accumulator.accumulation[perspective][i][0]);
 207           auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 208           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 209           for (IndexType j = 0; j < kNumChunks; ++j)
 210             _mm256_storeA_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadA_si256(&accumulation[j]), column[j]));
 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   #elif defined(USE_MMX)
 221           auto accumulation = reinterpret_cast<__m64*>(
 222               &accumulator.accumulation[perspective][i][0]);
 223           auto column = reinterpret_cast<const __m64*>(&weights_[offset]);
 224           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 225           for (IndexType j = 0; j < kNumChunks; ++j)
 226             accumulation[j] = _mm_add_pi16(accumulation[j], column[j]);
 227
 228   #elif defined(USE_NEON)
 229           auto accumulation = reinterpret_cast<int16x8_t*>(
 230               &accumulator.accumulation[perspective][i][0]);
 231           auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 232           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 233           for (IndexType j = 0; j < kNumChunks; ++j)
 234             accumulation[j] = vaddq_s16(accumulation[j], column[j]);
 235
 236   #else
 237           for (IndexType j = 0; j < kHalfDimensions; ++j)
 238             accumulator.accumulation[perspective][i][j] += weights_[offset + j];
 239   #endif
 240
 241         }
 242       }
 243   #if defined(USE_MMX)
 244       _mm_empty();
 245   #endif
 246
 247       accumulator.computed_accumulation = true;
 248     }
 249
 250     // Calculate cumulative value using difference calculation
 251     void UpdateAccumulator(const Position& pos) const {
 252
 253       const auto prev_accumulator = pos.state()->previous->accumulator;
 254       auto& accumulator = pos.state()->accumulator;
 255       IndexType i = 0;
 256       Features::IndexList removed_indices[2], added_indices[2];
 257       bool reset[2];
 258       RawFeatures::AppendChangedIndices(pos, kRefreshTriggers[i],
 259                                         removed_indices, added_indices, reset);
 260       for (Color perspective : { WHITE, BLACK }) {
 261
 262   #if defined(USE_AVX2)
 263         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 264         auto accumulation = reinterpret_cast<__m256i*>(
 265             &accumulator.accumulation[perspective][i][0]);
 266
 267   #elif defined(USE_SSE2)
 268         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 269         auto accumulation = reinterpret_cast<__m128i*>(
 270             &accumulator.accumulation[perspective][i][0]);
 271
 272   #elif defined(USE_MMX)
 273         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 274         auto accumulation = reinterpret_cast<__m64*>(
 275             &accumulator.accumulation[perspective][i][0]);
 276
 277   #elif defined(USE_NEON)
 278         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 279         auto accumulation = reinterpret_cast<int16x8_t*>(
 280             &accumulator.accumulation[perspective][i][0]);
 281   #endif
 282
 283         if (reset[perspective]) {
 284           std::memcpy(accumulator.accumulation[perspective][i], biases_,
 285                       kHalfDimensions * sizeof(BiasType));
 286         } else {
 287           std::memcpy(accumulator.accumulation[perspective][i],
 288                       prev_accumulator.accumulation[perspective][i],
 289                       kHalfDimensions * sizeof(BiasType));
 290           // Difference calculation for the deactivated features
 291           for (const auto index : removed_indices[perspective]) {
 292             const IndexType offset = kHalfDimensions * index;
 293
 294   #if defined(USE_AVX2)
 295             auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 296             for (IndexType j = 0; j < kNumChunks; ++j)
 297               accumulation[j] = _mm256_sub_epi16(accumulation[j], column[j]);
 298
 299   #elif defined(USE_SSE2)
 300             auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 301             for (IndexType j = 0; j < kNumChunks; ++j)
 302               accumulation[j] = _mm_sub_epi16(accumulation[j], column[j]);
 303
 304   #elif defined(USE_MMX)
 305             auto column = reinterpret_cast<const __m64*>(&weights_[offset]);
 306             for (IndexType j = 0; j < kNumChunks; ++j)
 307               accumulation[j] = _mm_sub_pi16(accumulation[j], column[j]);
 308
 309   #elif defined(USE_NEON)
 310             auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 311             for (IndexType j = 0; j < kNumChunks; ++j)
 312               accumulation[j] = vsubq_s16(accumulation[j], column[j]);
 313
 314   #else
 315             for (IndexType j = 0; j < kHalfDimensions; ++j)
 316               accumulator.accumulation[perspective][i][j] -= weights_[offset + j];
 317   #endif
 318
 319           }
 320         }
 321         { // Difference calculation for the activated features
 322           for (const auto index : added_indices[perspective]) {
 323             const IndexType offset = kHalfDimensions * index;
 324
 325   #if defined(USE_AVX2)
 326             auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 327             for (IndexType j = 0; j < kNumChunks; ++j)
 328               accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
 329
 330   #elif defined(USE_SSE2)
 331             auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 332             for (IndexType j = 0; j < kNumChunks; ++j)
 333               accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
 334
 335   #elif defined(USE_MMX)
 336             auto column = reinterpret_cast<const __m64*>(&weights_[offset]);
 337             for (IndexType j = 0; j < kNumChunks; ++j)
 338               accumulation[j] = _mm_add_pi16(accumulation[j], column[j]);
 339
 340   #elif defined(USE_NEON)
 341             auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 342             for (IndexType j = 0; j < kNumChunks; ++j)
 343               accumulation[j] = vaddq_s16(accumulation[j], column[j]);
 344
 345   #else
 346             for (IndexType j = 0; j < kHalfDimensions; ++j)
 347               accumulator.accumulation[perspective][i][j] += weights_[offset + j];
 348   #endif
 349
 350           }
 351         }
 352       }
 353   #if defined(USE_MMX)
 354       _mm_empty();
 355   #endif
 356
 357       accumulator.computed_accumulation = true;
 358     }
 359
 360     using BiasType = std::int16_t;
 361     using WeightType = std::int16_t;
 362
 363     alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
 364     alignas(kCacheLineSize)
 365         WeightType weights_[kHalfDimensions * kInputDimensions];
 366   };
 367
 368 }  // namespace Eval::NNUE
 369
 370 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED