2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
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.
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.
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/>.
19 // A class that converts the input features of the NNUE evaluation function
21 #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED
22 #define NNUE_FEATURE_TRANSFORMER_H_INCLUDED
31 #include "../position.h"
33 #include "nnue_accumulator.h"
34 #include "nnue_architecture.h"
35 #include "nnue_common.h"
37 namespace Stockfish::Eval::NNUE {
39 using BiasType = std::int16_t;
40 using WeightType = std::int16_t;
41 using PSQTWeightType = std::int32_t;
43 // If vector instructions are enabled, we update and refresh the
44 // accumulator tile by tile such that each tile fits in the CPU's
48 static_assert(PSQTBuckets % 8 == 0,
49 "Per feature PSQT values cannot be processed at granularity lower than 8 at a time.");
52 using vec_t = __m512i;
53 using psqt_vec_t = __m256i;
54 #define vec_load(a) _mm512_load_si512(a)
55 #define vec_store(a,b) _mm512_store_si512(a,b)
56 #define vec_add_16(a,b) _mm512_add_epi16(a,b)
57 #define vec_sub_16(a,b) _mm512_sub_epi16(a,b)
58 #define vec_mul_16(a,b) _mm512_mullo_epi16(a,b)
59 #define vec_zero() _mm512_setzero_epi32()
60 #define vec_set_16(a) _mm512_set1_epi16(a)
61 #define vec_max_16(a,b) _mm512_max_epi16(a,b)
62 #define vec_min_16(a,b) _mm512_min_epi16(a,b)
63 inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
64 vec_t compacted = _mm512_packs_epi16(_mm512_srli_epi16(a,7),_mm512_srli_epi16(b,7));
65 return _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), compacted);
67 #define vec_load_psqt(a) _mm256_load_si256(a)
68 #define vec_store_psqt(a,b) _mm256_store_si256(a,b)
69 #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
70 #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
71 #define vec_zero_psqt() _mm256_setzero_si256()
72 #define NumRegistersSIMD 16
73 #define MaxChunkSize 64
76 using vec_t = __m256i;
77 using psqt_vec_t = __m256i;
78 #define vec_load(a) _mm256_load_si256(a)
79 #define vec_store(a,b) _mm256_store_si256(a,b)
80 #define vec_add_16(a,b) _mm256_add_epi16(a,b)
81 #define vec_sub_16(a,b) _mm256_sub_epi16(a,b)
82 #define vec_mul_16(a,b) _mm256_mullo_epi16(a,b)
83 #define vec_zero() _mm256_setzero_si256()
84 #define vec_set_16(a) _mm256_set1_epi16(a)
85 #define vec_max_16(a,b) _mm256_max_epi16(a,b)
86 #define vec_min_16(a,b) _mm256_min_epi16(a,b)
87 inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
88 vec_t compacted = _mm256_packs_epi16(_mm256_srli_epi16(a,7), _mm256_srli_epi16(b,7));
89 return _mm256_permute4x64_epi64(compacted, 0b11011000);
91 #define vec_load_psqt(a) _mm256_load_si256(a)
92 #define vec_store_psqt(a,b) _mm256_store_si256(a,b)
93 #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
94 #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
95 #define vec_zero_psqt() _mm256_setzero_si256()
96 #define NumRegistersSIMD 16
97 #define MaxChunkSize 32
100 using vec_t = __m128i;
101 using psqt_vec_t = __m128i;
102 #define vec_load(a) (*(a))
103 #define vec_store(a,b) *(a)=(b)
104 #define vec_add_16(a,b) _mm_add_epi16(a,b)
105 #define vec_sub_16(a,b) _mm_sub_epi16(a,b)
106 #define vec_mul_16(a,b) _mm_mullo_epi16(a,b)
107 #define vec_zero() _mm_setzero_si128()
108 #define vec_set_16(a) _mm_set1_epi16(a)
109 #define vec_max_16(a,b) _mm_max_epi16(a,b)
110 #define vec_min_16(a,b) _mm_min_epi16(a,b)
111 #define vec_msb_pack_16(a,b) _mm_packs_epi16(_mm_srli_epi16(a,7),_mm_srli_epi16(b,7))
112 #define vec_load_psqt(a) (*(a))
113 #define vec_store_psqt(a,b) *(a)=(b)
114 #define vec_add_psqt_32(a,b) _mm_add_epi32(a,b)
115 #define vec_sub_psqt_32(a,b) _mm_sub_epi32(a,b)
116 #define vec_zero_psqt() _mm_setzero_si128()
117 #define NumRegistersSIMD (Is64Bit ? 16 : 8)
118 #define MaxChunkSize 16
122 using psqt_vec_t = __m64;
123 #define vec_load(a) (*(a))
124 #define vec_store(a,b) *(a)=(b)
125 #define vec_add_16(a,b) _mm_add_pi16(a,b)
126 #define vec_sub_16(a,b) _mm_sub_pi16(a,b)
127 #define vec_mul_16(a,b) _mm_mullo_pi16(a,b)
128 #define vec_zero() _mm_setzero_si64()
129 #define vec_set_16(a) _mm_set1_pi16(a)
130 inline vec_t vec_max_16(vec_t a,vec_t b){
131 vec_t comparison = _mm_cmpgt_pi16(a,b);
132 return _mm_or_si64(_mm_and_si64(comparison, a), _mm_andnot_si64(comparison, b));
134 inline vec_t vec_min_16(vec_t a,vec_t b){
135 vec_t comparison = _mm_cmpgt_pi16(a,b);
136 return _mm_or_si64(_mm_and_si64(comparison, b), _mm_andnot_si64(comparison, a));
138 #define vec_msb_pack_16(a,b) _mm_packs_pi16(_mm_srli_pi16(a,7),_mm_srli_pi16(b,7))
139 #define vec_load_psqt(a) (*(a))
140 #define vec_store_psqt(a,b) *(a)=(b)
141 #define vec_add_psqt_32(a,b) _mm_add_pi32(a,b)
142 #define vec_sub_psqt_32(a,b) _mm_sub_pi32(a,b)
143 #define vec_zero_psqt() _mm_setzero_si64()
144 #define vec_cleanup() _mm_empty()
145 #define NumRegistersSIMD 8
146 #define MaxChunkSize 8
149 using vec_t = int16x8_t;
150 using psqt_vec_t = int32x4_t;
151 #define vec_load(a) (*(a))
152 #define vec_store(a,b) *(a)=(b)
153 #define vec_add_16(a,b) vaddq_s16(a,b)
154 #define vec_sub_16(a,b) vsubq_s16(a,b)
155 #define vec_mul_16(a,b) vmulq_s16(a,b)
156 #define vec_zero() vec_t{0}
157 #define vec_set_16(a) vdupq_n_s16(a)
158 #define vec_max_16(a,b) vmaxq_s16(a,b)
159 #define vec_min_16(a,b) vminq_s16(a,b)
160 inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
161 const int8x8_t shifta = vshrn_n_s16(a, 7);
162 const int8x8_t shiftb = vshrn_n_s16(b, 7);
163 const int8x16_t compacted = vcombine_s8(shifta,shiftb);
164 return *reinterpret_cast<const vec_t*> (&compacted);
166 #define vec_load_psqt(a) (*(a))
167 #define vec_store_psqt(a,b) *(a)=(b)
168 #define vec_add_psqt_32(a,b) vaddq_s32(a,b)
169 #define vec_sub_psqt_32(a,b) vsubq_s32(a,b)
170 #define vec_zero_psqt() psqt_vec_t{0}
171 #define NumRegistersSIMD 16
172 #define MaxChunkSize 16
182 // Compute optimal SIMD register count for feature transformer accumulation.
184 // We use __m* types as template arguments, which causes GCC to emit warnings
185 // about losing some attribute information. This is irrelevant to us as we
186 // only take their size, so the following pragma are harmless.
187 #if defined(__GNUC__)
188 #pragma GCC diagnostic push
189 #pragma GCC diagnostic ignored "-Wignored-attributes"
192 template <typename SIMDRegisterType,
196 static constexpr int BestRegisterCount()
198 #define RegisterSize sizeof(SIMDRegisterType)
199 #define LaneSize sizeof(LaneType)
201 static_assert(RegisterSize >= LaneSize);
202 static_assert(MaxRegisters <= NumRegistersSIMD);
203 static_assert(MaxRegisters > 0);
204 static_assert(NumRegistersSIMD > 0);
205 static_assert(RegisterSize % LaneSize == 0);
206 static_assert((NumLanes * LaneSize) % RegisterSize == 0);
208 const int ideal = (NumLanes * LaneSize) / RegisterSize;
209 if (ideal <= MaxRegisters)
212 // Look for the largest divisor of the ideal register count that is smaller than MaxRegisters
213 for (int divisor = MaxRegisters; divisor > 1; --divisor)
214 if (ideal % divisor == 0)
220 static constexpr int NumRegs = BestRegisterCount<vec_t, WeightType, TransformedFeatureDimensions, NumRegistersSIMD>();
221 static constexpr int NumPsqtRegs = BestRegisterCount<psqt_vec_t, PSQTWeightType, PSQTBuckets, NumRegistersSIMD>();
222 #if defined(__GNUC__)
223 #pragma GCC diagnostic pop
229 // Input feature converter
230 class FeatureTransformer {
233 // Number of output dimensions for one side
234 static constexpr IndexType HalfDimensions = TransformedFeatureDimensions;
237 static constexpr IndexType TileHeight = NumRegs * sizeof(vec_t) / 2;
238 static constexpr IndexType PsqtTileHeight = NumPsqtRegs * sizeof(psqt_vec_t) / 4;
239 static_assert(HalfDimensions % TileHeight == 0, "TileHeight must divide HalfDimensions");
240 static_assert(PSQTBuckets % PsqtTileHeight == 0, "PsqtTileHeight must divide PSQTBuckets");
245 using OutputType = TransformedFeatureType;
247 // Number of input/output dimensions
248 static constexpr IndexType InputDimensions = FeatureSet::Dimensions;
249 static constexpr IndexType OutputDimensions = HalfDimensions;
251 // Size of forward propagation buffer
252 static constexpr std::size_t BufferSize =
253 OutputDimensions * sizeof(OutputType);
255 // Hash value embedded in the evaluation file
256 static constexpr std::uint32_t get_hash_value() {
257 return FeatureSet::HashValue ^ (OutputDimensions * 2);
260 // Read network parameters
261 bool read_parameters(std::istream& stream) {
263 read_leb_128<BiasType >(stream, biases , HalfDimensions );
264 read_leb_128<WeightType >(stream, weights , HalfDimensions * InputDimensions);
265 read_leb_128<PSQTWeightType>(stream, psqtWeights, PSQTBuckets * InputDimensions);
267 return !stream.fail();
270 // Write network parameters
271 bool write_parameters(std::ostream& stream) const {
273 write_leb_128<BiasType >(stream, biases , HalfDimensions );
274 write_leb_128<WeightType >(stream, weights , HalfDimensions * InputDimensions);
275 write_leb_128<PSQTWeightType>(stream, psqtWeights, PSQTBuckets * InputDimensions);
277 return !stream.fail();
280 // Convert input features
281 std::int32_t transform(const Position& pos, OutputType* output, int bucket) const {
282 update_accumulator<WHITE>(pos);
283 update_accumulator<BLACK>(pos);
285 const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
286 const auto& accumulation = pos.state()->accumulator.accumulation;
287 const auto& psqtAccumulation = pos.state()->accumulator.psqtAccumulation;
290 psqtAccumulation[perspectives[0]][bucket]
291 - psqtAccumulation[perspectives[1]][bucket]
295 for (IndexType p = 0; p < 2; ++p)
297 const IndexType offset = (HalfDimensions / 2) * p;
301 constexpr IndexType OutputChunkSize = MaxChunkSize;
302 static_assert((HalfDimensions / 2) % OutputChunkSize == 0);
303 constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize;
305 vec_t Zero = vec_zero();
306 vec_t One = vec_set_16(127);
308 const vec_t* in0 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][0]));
309 const vec_t* in1 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][HalfDimensions / 2]));
310 vec_t* out = reinterpret_cast< vec_t*>(output + offset);
312 for (IndexType j = 0; j < NumOutputChunks; j += 1)
314 const vec_t sum0a = vec_max_16(vec_min_16(in0[j * 2 + 0], One), Zero);
315 const vec_t sum0b = vec_max_16(vec_min_16(in0[j * 2 + 1], One), Zero);
316 const vec_t sum1a = vec_max_16(vec_min_16(in1[j * 2 + 0], One), Zero);
317 const vec_t sum1b = vec_max_16(vec_min_16(in1[j * 2 + 1], One), Zero);
319 const vec_t pa = vec_mul_16(sum0a, sum1a);
320 const vec_t pb = vec_mul_16(sum0b, sum1b);
322 out[j] = vec_msb_pack_16(pa, pb);
327 for (IndexType j = 0; j < HalfDimensions / 2; ++j) {
328 BiasType sum0 = accumulation[static_cast<int>(perspectives[p])][j + 0];
329 BiasType sum1 = accumulation[static_cast<int>(perspectives[p])][j + HalfDimensions / 2];
330 sum0 = std::clamp<BiasType>(sum0, 0, 127);
331 sum1 = std::clamp<BiasType>(sum1, 0, 127);
332 output[offset + j] = static_cast<OutputType>(unsigned(sum0 * sum1) / 128);
338 #if defined(vec_cleanup)
343 } // end of function transform()
345 void hint_common_access(const Position& pos) const {
346 hint_common_access_for_perspective<WHITE>(pos);
347 hint_common_access_for_perspective<BLACK>(pos);
351 template<Color Perspective>
352 [[nodiscard]] std::pair<StateInfo*, StateInfo*> try_find_computed_accumulator(const Position& pos) const {
353 // Look for a usable accumulator of an earlier position. We keep track
354 // of the estimated gain in terms of features to be added/subtracted.
355 StateInfo *st = pos.state(), *next = nullptr;
356 int gain = FeatureSet::refresh_cost(pos);
357 while (st->previous && !st->accumulator.computed[Perspective])
359 // This governs when a full feature refresh is needed and how many
360 // updates are better than just one full refresh.
361 if ( FeatureSet::requires_refresh(st, Perspective)
362 || (gain -= FeatureSet::update_cost(st) + 1) < 0)
370 // NOTE: The parameter states_to_update is an array of position states, ending with nullptr.
371 // All states must be sequential, that is states_to_update[i] must either be reachable
372 // by repeatedly applying ->previous from states_to_update[i+1] or states_to_update[i] == nullptr.
373 // computed_st must be reachable by repeatedly applying ->previous on states_to_update[0], if not nullptr.
374 template<Color Perspective, size_t N>
375 void update_accumulator_incremental(const Position& pos, StateInfo* computed_st, StateInfo* states_to_update[N]) const {
376 static_assert(N > 0);
377 assert(states_to_update[N-1] == nullptr);
380 // Gcc-10.2 unnecessarily spills AVX2 registers if this array
381 // is defined in the VECTOR code below, once in each branch
383 psqt_vec_t psqt[NumPsqtRegs];
386 if (states_to_update[0] == nullptr)
389 // Update incrementally going back through states_to_update.
391 // Gather all features to be updated.
392 const Square ksq = pos.square<KING>(Perspective);
394 // The size must be enough to contain the largest possible update.
395 // That might depend on the feature set and generally relies on the
396 // feature set's update cost calculation to be correct and never
397 // allow updates with more added/removed features than MaxActiveDimensions.
398 FeatureSet::IndexList removed[N-1], added[N-1];
401 int i = N-2; // last potential state to update. Skip last element because it must be nullptr.
402 while (states_to_update[i] == nullptr)
405 StateInfo *st2 = states_to_update[i];
409 states_to_update[i]->accumulator.computed[Perspective] = true;
411 StateInfo* end_state = i == 0 ? computed_st : states_to_update[i - 1];
413 for (; st2 != end_state; st2 = st2->previous)
414 FeatureSet::append_changed_indices<Perspective>(
415 ksq, st2->dirtyPiece, removed[i], added[i]);
419 StateInfo* st = computed_st;
421 // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
423 for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
426 auto accTile = reinterpret_cast<vec_t*>(
427 &st->accumulator.accumulation[Perspective][j * TileHeight]);
428 for (IndexType k = 0; k < NumRegs; ++k)
429 acc[k] = vec_load(&accTile[k]);
431 for (IndexType i = 0; states_to_update[i]; ++i)
433 // Difference calculation for the deactivated features
434 for (const auto index : removed[i])
436 const IndexType offset = HalfDimensions * index + j * TileHeight;
437 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
438 for (IndexType k = 0; k < NumRegs; ++k)
439 acc[k] = vec_sub_16(acc[k], column[k]);
442 // Difference calculation for the activated features
443 for (const auto index : added[i])
445 const IndexType offset = HalfDimensions * index + j * TileHeight;
446 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
447 for (IndexType k = 0; k < NumRegs; ++k)
448 acc[k] = vec_add_16(acc[k], column[k]);
452 accTile = reinterpret_cast<vec_t*>(
453 &states_to_update[i]->accumulator.accumulation[Perspective][j * TileHeight]);
454 for (IndexType k = 0; k < NumRegs; ++k)
455 vec_store(&accTile[k], acc[k]);
459 for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
462 auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
463 &st->accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
464 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
465 psqt[k] = vec_load_psqt(&accTilePsqt[k]);
467 for (IndexType i = 0; states_to_update[i]; ++i)
469 // Difference calculation for the deactivated features
470 for (const auto index : removed[i])
472 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
473 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
474 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
475 psqt[k] = vec_sub_psqt_32(psqt[k], columnPsqt[k]);
478 // Difference calculation for the activated features
479 for (const auto index : added[i])
481 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
482 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
483 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
484 psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
488 accTilePsqt = reinterpret_cast<psqt_vec_t*>(
489 &states_to_update[i]->accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
490 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
491 vec_store_psqt(&accTilePsqt[k], psqt[k]);
496 for (IndexType i = 0; states_to_update[i]; ++i)
498 std::memcpy(states_to_update[i]->accumulator.accumulation[Perspective],
499 st->accumulator.accumulation[Perspective],
500 HalfDimensions * sizeof(BiasType));
502 for (std::size_t k = 0; k < PSQTBuckets; ++k)
503 states_to_update[i]->accumulator.psqtAccumulation[Perspective][k] = st->accumulator.psqtAccumulation[Perspective][k];
505 st = states_to_update[i];
507 // Difference calculation for the deactivated features
508 for (const auto index : removed[i])
510 const IndexType offset = HalfDimensions * index;
512 for (IndexType j = 0; j < HalfDimensions; ++j)
513 st->accumulator.accumulation[Perspective][j] -= weights[offset + j];
515 for (std::size_t k = 0; k < PSQTBuckets; ++k)
516 st->accumulator.psqtAccumulation[Perspective][k] -= psqtWeights[index * PSQTBuckets + k];
519 // Difference calculation for the activated features
520 for (const auto index : added[i])
522 const IndexType offset = HalfDimensions * index;
524 for (IndexType j = 0; j < HalfDimensions; ++j)
525 st->accumulator.accumulation[Perspective][j] += weights[offset + j];
527 for (std::size_t k = 0; k < PSQTBuckets; ++k)
528 st->accumulator.psqtAccumulation[Perspective][k] += psqtWeights[index * PSQTBuckets + k];
538 template<Color Perspective>
539 void update_accumulator_refresh(const Position& pos) const {
541 // Gcc-10.2 unnecessarily spills AVX2 registers if this array
542 // is defined in the VECTOR code below, once in each branch
544 psqt_vec_t psqt[NumPsqtRegs];
547 // Refresh the accumulator
548 // Could be extracted to a separate function because it's done in 2 places,
549 // but it's unclear if compilers would correctly handle register allocation.
550 auto& accumulator = pos.state()->accumulator;
551 accumulator.computed[Perspective] = true;
552 FeatureSet::IndexList active;
553 FeatureSet::append_active_indices<Perspective>(pos, active);
556 for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
558 auto biasesTile = reinterpret_cast<const vec_t*>(
559 &biases[j * TileHeight]);
560 for (IndexType k = 0; k < NumRegs; ++k)
561 acc[k] = biasesTile[k];
563 for (const auto index : active)
565 const IndexType offset = HalfDimensions * index + j * TileHeight;
566 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
568 for (unsigned k = 0; k < NumRegs; ++k)
569 acc[k] = vec_add_16(acc[k], column[k]);
572 auto accTile = reinterpret_cast<vec_t*>(
573 &accumulator.accumulation[Perspective][j * TileHeight]);
574 for (unsigned k = 0; k < NumRegs; k++)
575 vec_store(&accTile[k], acc[k]);
578 for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
580 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
581 psqt[k] = vec_zero_psqt();
583 for (const auto index : active)
585 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
586 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
588 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
589 psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
592 auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
593 &accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
594 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
595 vec_store_psqt(&accTilePsqt[k], psqt[k]);
599 std::memcpy(accumulator.accumulation[Perspective], biases,
600 HalfDimensions * sizeof(BiasType));
602 for (std::size_t k = 0; k < PSQTBuckets; ++k)
603 accumulator.psqtAccumulation[Perspective][k] = 0;
605 for (const auto index : active)
607 const IndexType offset = HalfDimensions * index;
609 for (IndexType j = 0; j < HalfDimensions; ++j)
610 accumulator.accumulation[Perspective][j] += weights[offset + j];
612 for (std::size_t k = 0; k < PSQTBuckets; ++k)
613 accumulator.psqtAccumulation[Perspective][k] += psqtWeights[index * PSQTBuckets + k];
622 template<Color Perspective>
623 void hint_common_access_for_perspective(const Position& pos) const {
625 // Works like update_accumulator, but performs less work.
626 // Updates ONLY the accumulator for pos.
628 // Look for a usable accumulator of an earlier position. We keep track
629 // of the estimated gain in terms of features to be added/subtracted.
631 if (pos.state()->accumulator.computed[Perspective])
634 auto [oldest_st, _] = try_find_computed_accumulator<Perspective>(pos);
636 if (oldest_st->accumulator.computed[Perspective])
638 // Only update current position accumulator to minimize work.
639 StateInfo* states_to_update[2] = { pos.state(), nullptr };
640 update_accumulator_incremental<Perspective, 2>(pos, oldest_st, states_to_update);
644 update_accumulator_refresh<Perspective>(pos);
648 template<Color Perspective>
649 void update_accumulator(const Position& pos) const {
651 auto [oldest_st, next] = try_find_computed_accumulator<Perspective>(pos);
653 if (oldest_st->accumulator.computed[Perspective])
658 // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
659 // Currently we update 2 accumulators.
660 // 1. for the current position
661 // 2. the next accumulator after the computed one
662 // The heuristic may change in the future.
663 StateInfo *states_to_update[3] =
664 { next, next == pos.state() ? nullptr : pos.state(), nullptr };
666 update_accumulator_incremental<Perspective, 3>(pos, oldest_st, states_to_update);
670 update_accumulator_refresh<Perspective>(pos);
674 alignas(CacheLineSize) BiasType biases[HalfDimensions];
675 alignas(CacheLineSize) WeightType weights[HalfDimensions * InputDimensions];
676 alignas(CacheLineSize) PSQTWeightType psqtWeights[InputDimensions * PSQTBuckets];
679 } // namespace Stockfish::Eval::NNUE
681 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED