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       for (std::size_t i = 0; i < kHalfDimensions; ++i)
  59         biases_[i] = read_little_endian<BiasType>(stream);
  60       for (std::size_t i = 0; i < kHalfDimensions * kInputDimensions; ++i)
  61         weights_[i] = read_little_endian<WeightType>(stream);
  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_SSE2)
  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_MMX)
 101       constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
 102       const __m64 k0x80s = _mm_set1_pi8(-128);
 103
 104   #elif defined(USE_NEON)
 105       constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 106       const int8x8_t kZero = {0};
 107   #endif
 108
 109       const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
 110       for (IndexType p = 0; p < 2; ++p) {
 111         const IndexType offset = kHalfDimensions * p;
 112
 113   #if defined(USE_AVX2)
 114         auto out = reinterpret_cast<__m256i*>(&output[offset]);
 115         for (IndexType j = 0; j < kNumChunks; ++j) {
 116           __m256i sum0 = _mm256_loadA_si256(
 117               &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 0]);
 118           __m256i sum1 = _mm256_loadA_si256(
 119             &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 1]);
 120           _mm256_storeA_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
 121               _mm256_packs_epi16(sum0, sum1), kZero), kControl));
 122         }
 123
 124   #elif defined(USE_SSE2)
 125         auto out = reinterpret_cast<__m128i*>(&output[offset]);
 126         for (IndexType j = 0; j < kNumChunks; ++j) {
 127           __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
 128               accumulation[perspectives[p]][0])[j * 2 + 0]);
 129           __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
 130               accumulation[perspectives[p]][0])[j * 2 + 1]);
 131       const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
 132
 133           _mm_store_si128(&out[j],
 134
 135   #ifdef USE_SSE41
 136             _mm_max_epi8(packedbytes, kZero)
 137   #else
 138             _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
 139   #endif
 140
 141           );
 142         }
 143
 144   #elif defined(USE_MMX)
 145         auto out = reinterpret_cast<__m64*>(&output[offset]);
 146         for (IndexType j = 0; j < kNumChunks; ++j) {
 147           __m64 sum0 = *(&reinterpret_cast<const __m64*>(
 148               accumulation[perspectives[p]][0])[j * 2 + 0]);
 149           __m64 sum1 = *(&reinterpret_cast<const __m64*>(
 150               accumulation[perspectives[p]][0])[j * 2 + 1]);
 151           const __m64 packedbytes = _mm_packs_pi16(sum0, sum1);
 152           out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
 153         }
 154
 155   #elif defined(USE_NEON)
 156         const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
 157         for (IndexType j = 0; j < kNumChunks; ++j) {
 158           int16x8_t sum = reinterpret_cast<const int16x8_t*>(
 159               accumulation[perspectives[p]][0])[j];
 160           out[j] = vmax_s8(vqmovn_s16(sum), kZero);
 161         }
 162
 163   #else
 164         for (IndexType j = 0; j < kHalfDimensions; ++j) {
 165           BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
 166           output[offset + j] = static_cast<OutputType>(
 167               std::max<int>(0, std::min<int>(127, sum)));
 168         }
 169   #endif
 170
 171       }
 172   #if defined(USE_MMX)
 173       _mm_empty();
 174   #endif
 175     }
 176
 177    private:
 178     // Calculate cumulative value without using difference calculation
 179     void RefreshAccumulator(const Position& pos) const {
 180       auto& accumulator = pos.state()->accumulator;
 181       IndexType i = 0;
 182       Features::IndexList active_indices[2];
 183       RawFeatures::AppendActiveIndices(pos, kRefreshTriggers[i],
 184                                        active_indices);
 185       for (Color perspective : { WHITE, BLACK }) {
 186         std::memcpy(accumulator.accumulation[perspective][i], biases_,
 187                    kHalfDimensions * sizeof(BiasType));
 188         for (const auto index : active_indices[perspective]) {
 189           const IndexType offset = kHalfDimensions * index;
 190   #if defined(USE_AVX512)
 191           auto accumulation = reinterpret_cast<__m512i*>(
 192               &accumulator.accumulation[perspective][i][0]);
 193           auto column = reinterpret_cast<const __m512i*>(&weights_[offset]);
 194           constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
 195           for (IndexType j = 0; j < kNumChunks; ++j)
 196             _mm512_storeA_si512(&accumulation[j], _mm512_add_epi16(_mm512_loadA_si512(&accumulation[j]), column[j]));
 197
 198   #elif defined(USE_AVX2)
 199           auto accumulation = reinterpret_cast<__m256i*>(
 200               &accumulator.accumulation[perspective][i][0]);
 201           auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 202           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 203           for (IndexType j = 0; j < kNumChunks; ++j)
 204             _mm256_storeA_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadA_si256(&accumulation[j]), column[j]));
 205
 206   #elif defined(USE_SSE2)
 207           auto accumulation = reinterpret_cast<__m128i*>(
 208               &accumulator.accumulation[perspective][i][0]);
 209           auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 210           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 211           for (IndexType j = 0; j < kNumChunks; ++j)
 212             accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
 213
 214   #elif defined(USE_MMX)
 215           auto accumulation = reinterpret_cast<__m64*>(
 216               &accumulator.accumulation[perspective][i][0]);
 217           auto column = reinterpret_cast<const __m64*>(&weights_[offset]);
 218           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 219           for (IndexType j = 0; j < kNumChunks; ++j) {
 220             accumulation[j] = _mm_add_pi16(accumulation[j], column[j]);
 221           }
 222
 223   #elif defined(USE_NEON)
 224           auto accumulation = reinterpret_cast<int16x8_t*>(
 225               &accumulator.accumulation[perspective][i][0]);
 226           auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 227           constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 228           for (IndexType j = 0; j < kNumChunks; ++j)
 229             accumulation[j] = vaddq_s16(accumulation[j], column[j]);
 230
 231   #else
 232           for (IndexType j = 0; j < kHalfDimensions; ++j)
 233             accumulator.accumulation[perspective][i][j] += weights_[offset + j];
 234   #endif
 235
 236         }
 237       }
 238   #if defined(USE_MMX)
 239       _mm_empty();
 240   #endif
 241
 242       accumulator.computed_accumulation = true;
 243       accumulator.computed_score = false;
 244     }
 245
 246     // Calculate cumulative value using difference calculation
 247     void UpdateAccumulator(const Position& pos) const {
 248       const auto prev_accumulator = pos.state()->previous->accumulator;
 249       auto& accumulator = pos.state()->accumulator;
 250       IndexType i = 0;
 251       Features::IndexList removed_indices[2], added_indices[2];
 252       bool reset[2];
 253       RawFeatures::AppendChangedIndices(pos, kRefreshTriggers[i],
 254                                         removed_indices, added_indices, reset);
 255       for (Color perspective : { WHITE, BLACK }) {
 256
 257   #if defined(USE_AVX2)
 258         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 259         auto accumulation = reinterpret_cast<__m256i*>(
 260             &accumulator.accumulation[perspective][i][0]);
 261
 262   #elif defined(USE_SSE2)
 263         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 264         auto accumulation = reinterpret_cast<__m128i*>(
 265             &accumulator.accumulation[perspective][i][0]);
 266
 267   #elif defined(USE_MMX)
 268         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 269         auto accumulation = reinterpret_cast<__m64*>(
 270             &accumulator.accumulation[perspective][i][0]);
 271
 272   #elif defined(USE_NEON)
 273         constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
 274         auto accumulation = reinterpret_cast<int16x8_t*>(
 275             &accumulator.accumulation[perspective][i][0]);
 276   #endif
 277
 278         if (reset[perspective]) {
 279           std::memcpy(accumulator.accumulation[perspective][i], biases_,
 280                       kHalfDimensions * sizeof(BiasType));
 281         } else {
 282           std::memcpy(accumulator.accumulation[perspective][i],
 283                       prev_accumulator.accumulation[perspective][i],
 284                       kHalfDimensions * sizeof(BiasType));
 285           // Difference calculation for the deactivated features
 286           for (const auto index : removed_indices[perspective]) {
 287             const IndexType offset = kHalfDimensions * index;
 288
 289   #if defined(USE_AVX2)
 290             auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 291             for (IndexType j = 0; j < kNumChunks; ++j) {
 292               accumulation[j] = _mm256_sub_epi16(accumulation[j], column[j]);
 293             }
 294
 295   #elif defined(USE_SSE2)
 296             auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 297             for (IndexType j = 0; j < kNumChunks; ++j) {
 298               accumulation[j] = _mm_sub_epi16(accumulation[j], column[j]);
 299             }
 300
 301   #elif defined(USE_MMX)
 302             auto column = reinterpret_cast<const __m64*>(&weights_[offset]);
 303             for (IndexType j = 0; j < kNumChunks; ++j) {
 304               accumulation[j] = _mm_sub_pi16(accumulation[j], column[j]);
 305             }
 306
 307   #elif defined(USE_NEON)
 308             auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 309             for (IndexType j = 0; j < kNumChunks; ++j) {
 310               accumulation[j] = vsubq_s16(accumulation[j], column[j]);
 311             }
 312
 313   #else
 314             for (IndexType j = 0; j < kHalfDimensions; ++j) {
 315               accumulator.accumulation[perspective][i][j] -=
 316                   weights_[offset + j];
 317             }
 318   #endif
 319
 320           }
 321         }
 322         { // Difference calculation for the activated features
 323           for (const auto index : added_indices[perspective]) {
 324             const IndexType offset = kHalfDimensions * index;
 325
 326   #if defined(USE_AVX2)
 327             auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
 328             for (IndexType j = 0; j < kNumChunks; ++j) {
 329               accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
 330             }
 331
 332   #elif defined(USE_SSE2)
 333             auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
 334             for (IndexType j = 0; j < kNumChunks; ++j) {
 335               accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
 336             }
 337
 338   #elif defined(USE_MMX)
 339             auto column = reinterpret_cast<const __m64*>(&weights_[offset]);
 340             for (IndexType j = 0; j < kNumChunks; ++j) {
 341               accumulation[j] = _mm_add_pi16(accumulation[j], column[j]);
 342             }
 343
 344   #elif defined(USE_NEON)
 345             auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
 346             for (IndexType j = 0; j < kNumChunks; ++j) {
 347               accumulation[j] = vaddq_s16(accumulation[j], column[j]);
 348             }
 349
 350   #else
 351             for (IndexType j = 0; j < kHalfDimensions; ++j) {
 352               accumulator.accumulation[perspective][i][j] +=
 353                   weights_[offset + j];
 354             }
 355   #endif
 356
 357           }
 358         }
 359       }
 360   #if defined(USE_MMX)
 361       _mm_empty();
 362   #endif
 363
 364       accumulator.computed_accumulation = true;
 365       accumulator.computed_score = false;
 366     }
 367
 368     using BiasType = std::int16_t;
 369     using WeightType = std::int16_t;
 370
 371     alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
 372     alignas(kCacheLineSize)
 373         WeightType weights_[kHalfDimensions * kInputDimensions];
 374   };
 375
 376 }  // namespace Eval::NNUE
 377
 378 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED