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
121 using vec_t = int16x8_t;
122 using psqt_vec_t = int32x4_t;
123 #define vec_load(a) (*(a))
124 #define vec_store(a,b) *(a)=(b)
125 #define vec_add_16(a,b) vaddq_s16(a,b)
126 #define vec_sub_16(a,b) vsubq_s16(a,b)
127 #define vec_mul_16(a,b) vmulq_s16(a,b)
128 #define vec_zero() vec_t{0}
129 #define vec_set_16(a) vdupq_n_s16(a)
130 #define vec_max_16(a,b) vmaxq_s16(a,b)
131 #define vec_min_16(a,b) vminq_s16(a,b)
132 inline vec_t vec_msb_pack_16(vec_t a, vec_t b){
133 const int8x8_t shifta = vshrn_n_s16(a, 7);
134 const int8x8_t shiftb = vshrn_n_s16(b, 7);
135 const int8x16_t compacted = vcombine_s8(shifta,shiftb);
136 return *reinterpret_cast<const vec_t*> (&compacted);
138 #define vec_load_psqt(a) (*(a))
139 #define vec_store_psqt(a,b) *(a)=(b)
140 #define vec_add_psqt_32(a,b) vaddq_s32(a,b)
141 #define vec_sub_psqt_32(a,b) vsubq_s32(a,b)
142 #define vec_zero_psqt() psqt_vec_t{0}
143 #define NumRegistersSIMD 16
144 #define MaxChunkSize 16
154 // Compute optimal SIMD register count for feature transformer accumulation.
156 // We use __m* types as template arguments, which causes GCC to emit warnings
157 // about losing some attribute information. This is irrelevant to us as we
158 // only take their size, so the following pragma are harmless.
159 #if defined(__GNUC__)
160 #pragma GCC diagnostic push
161 #pragma GCC diagnostic ignored "-Wignored-attributes"
164 template <typename SIMDRegisterType,
168 static constexpr int BestRegisterCount()
170 #define RegisterSize sizeof(SIMDRegisterType)
171 #define LaneSize sizeof(LaneType)
173 static_assert(RegisterSize >= LaneSize);
174 static_assert(MaxRegisters <= NumRegistersSIMD);
175 static_assert(MaxRegisters > 0);
176 static_assert(NumRegistersSIMD > 0);
177 static_assert(RegisterSize % LaneSize == 0);
178 static_assert((NumLanes * LaneSize) % RegisterSize == 0);
180 const int ideal = (NumLanes * LaneSize) / RegisterSize;
181 if (ideal <= MaxRegisters)
184 // Look for the largest divisor of the ideal register count that is smaller than MaxRegisters
185 for (int divisor = MaxRegisters; divisor > 1; --divisor)
186 if (ideal % divisor == 0)
192 static constexpr int NumRegs = BestRegisterCount<vec_t, WeightType, TransformedFeatureDimensions, NumRegistersSIMD>();
193 static constexpr int NumPsqtRegs = BestRegisterCount<psqt_vec_t, PSQTWeightType, PSQTBuckets, NumRegistersSIMD>();
194 #if defined(__GNUC__)
195 #pragma GCC diagnostic pop
201 // Input feature converter
202 class FeatureTransformer {
205 // Number of output dimensions for one side
206 static constexpr IndexType HalfDimensions = TransformedFeatureDimensions;
209 static constexpr IndexType TileHeight = NumRegs * sizeof(vec_t) / 2;
210 static constexpr IndexType PsqtTileHeight = NumPsqtRegs * sizeof(psqt_vec_t) / 4;
211 static_assert(HalfDimensions % TileHeight == 0, "TileHeight must divide HalfDimensions");
212 static_assert(PSQTBuckets % PsqtTileHeight == 0, "PsqtTileHeight must divide PSQTBuckets");
217 using OutputType = TransformedFeatureType;
219 // Number of input/output dimensions
220 static constexpr IndexType InputDimensions = FeatureSet::Dimensions;
221 static constexpr IndexType OutputDimensions = HalfDimensions;
223 // Size of forward propagation buffer
224 static constexpr std::size_t BufferSize =
225 OutputDimensions * sizeof(OutputType);
227 // Hash value embedded in the evaluation file
228 static constexpr std::uint32_t get_hash_value() {
229 return FeatureSet::HashValue ^ (OutputDimensions * 2);
232 // Read network parameters
233 bool read_parameters(std::istream& stream) {
235 read_leb_128<BiasType >(stream, biases , HalfDimensions );
236 read_leb_128<WeightType >(stream, weights , HalfDimensions * InputDimensions);
237 read_leb_128<PSQTWeightType>(stream, psqtWeights, PSQTBuckets * InputDimensions);
239 return !stream.fail();
242 // Write network parameters
243 bool write_parameters(std::ostream& stream) const {
245 write_leb_128<BiasType >(stream, biases , HalfDimensions );
246 write_leb_128<WeightType >(stream, weights , HalfDimensions * InputDimensions);
247 write_leb_128<PSQTWeightType>(stream, psqtWeights, PSQTBuckets * InputDimensions);
249 return !stream.fail();
252 // Convert input features
253 std::int32_t transform(const Position& pos, OutputType* output, int bucket) const {
254 update_accumulator<WHITE>(pos);
255 update_accumulator<BLACK>(pos);
257 const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
258 const auto& accumulation = pos.state()->accumulator.accumulation;
259 const auto& psqtAccumulation = pos.state()->accumulator.psqtAccumulation;
262 psqtAccumulation[perspectives[0]][bucket]
263 - psqtAccumulation[perspectives[1]][bucket]
267 for (IndexType p = 0; p < 2; ++p)
269 const IndexType offset = (HalfDimensions / 2) * p;
273 constexpr IndexType OutputChunkSize = MaxChunkSize;
274 static_assert((HalfDimensions / 2) % OutputChunkSize == 0);
275 constexpr IndexType NumOutputChunks = HalfDimensions / 2 / OutputChunkSize;
277 vec_t Zero = vec_zero();
278 vec_t One = vec_set_16(127);
280 const vec_t* in0 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][0]));
281 const vec_t* in1 = reinterpret_cast<const vec_t*>(&(accumulation[perspectives[p]][HalfDimensions / 2]));
282 vec_t* out = reinterpret_cast< vec_t*>(output + offset);
284 for (IndexType j = 0; j < NumOutputChunks; j += 1)
286 const vec_t sum0a = vec_max_16(vec_min_16(in0[j * 2 + 0], One), Zero);
287 const vec_t sum0b = vec_max_16(vec_min_16(in0[j * 2 + 1], One), Zero);
288 const vec_t sum1a = vec_max_16(vec_min_16(in1[j * 2 + 0], One), Zero);
289 const vec_t sum1b = vec_max_16(vec_min_16(in1[j * 2 + 1], One), Zero);
291 const vec_t pa = vec_mul_16(sum0a, sum1a);
292 const vec_t pb = vec_mul_16(sum0b, sum1b);
294 out[j] = vec_msb_pack_16(pa, pb);
299 for (IndexType j = 0; j < HalfDimensions / 2; ++j) {
300 BiasType sum0 = accumulation[static_cast<int>(perspectives[p])][j + 0];
301 BiasType sum1 = accumulation[static_cast<int>(perspectives[p])][j + HalfDimensions / 2];
302 sum0 = std::clamp<BiasType>(sum0, 0, 127);
303 sum1 = std::clamp<BiasType>(sum1, 0, 127);
304 output[offset + j] = static_cast<OutputType>(unsigned(sum0 * sum1) / 128);
311 } // end of function transform()
313 void hint_common_access(const Position& pos) const {
314 hint_common_access_for_perspective<WHITE>(pos);
315 hint_common_access_for_perspective<BLACK>(pos);
319 template<Color Perspective>
320 [[nodiscard]] std::pair<StateInfo*, StateInfo*> try_find_computed_accumulator(const Position& pos) const {
321 // Look for a usable accumulator of an earlier position. We keep track
322 // of the estimated gain in terms of features to be added/subtracted.
323 StateInfo *st = pos.state(), *next = nullptr;
324 int gain = FeatureSet::refresh_cost(pos);
325 while (st->previous && !st->accumulator.computed[Perspective])
327 // This governs when a full feature refresh is needed and how many
328 // updates are better than just one full refresh.
329 if ( FeatureSet::requires_refresh(st, Perspective)
330 || (gain -= FeatureSet::update_cost(st) + 1) < 0)
338 // NOTE: The parameter states_to_update is an array of position states, ending with nullptr.
339 // All states must be sequential, that is states_to_update[i] must either be reachable
340 // by repeatedly applying ->previous from states_to_update[i+1] or states_to_update[i] == nullptr.
341 // computed_st must be reachable by repeatedly applying ->previous on states_to_update[0], if not nullptr.
342 template<Color Perspective, size_t N>
343 void update_accumulator_incremental(const Position& pos, StateInfo* computed_st, StateInfo* states_to_update[N]) const {
344 static_assert(N > 0);
345 assert(states_to_update[N-1] == nullptr);
348 // Gcc-10.2 unnecessarily spills AVX2 registers if this array
349 // is defined in the VECTOR code below, once in each branch
351 psqt_vec_t psqt[NumPsqtRegs];
354 if (states_to_update[0] == nullptr)
357 // Update incrementally going back through states_to_update.
359 // Gather all features to be updated.
360 const Square ksq = pos.square<KING>(Perspective);
362 // The size must be enough to contain the largest possible update.
363 // That might depend on the feature set and generally relies on the
364 // feature set's update cost calculation to be correct and never
365 // allow updates with more added/removed features than MaxActiveDimensions.
366 FeatureSet::IndexList removed[N-1], added[N-1];
369 int i = N-2; // last potential state to update. Skip last element because it must be nullptr.
370 while (states_to_update[i] == nullptr)
373 StateInfo* st2 = states_to_update[i];
377 states_to_update[i]->accumulator.computed[Perspective] = true;
379 const StateInfo* end_state = i == 0 ? computed_st : states_to_update[i - 1];
381 for (; st2 != end_state; st2 = st2->previous)
382 FeatureSet::append_changed_indices<Perspective>(
383 ksq, st2->dirtyPiece, removed[i], added[i]);
387 StateInfo* st = computed_st;
389 // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
392 if ( states_to_update[1] == nullptr
393 && (removed[0].size() == 1 || removed[0].size() == 2)
394 && added[0].size() == 1)
396 assert(states_to_update[0]);
398 auto accTileIn = reinterpret_cast<const vec_t*>(
399 &st->accumulator.accumulation[Perspective][0]);
400 auto accTileOut = reinterpret_cast<vec_t*>(
401 &states_to_update[0]->accumulator.accumulation[Perspective][0]);
403 const IndexType offsetR0 = HalfDimensions * removed[0][0];
404 auto columnR0 = reinterpret_cast<const vec_t*>(&weights[offsetR0]);
405 const IndexType offsetA = HalfDimensions * added[0][0];
406 auto columnA = reinterpret_cast<const vec_t*>(&weights[offsetA]);
408 if (removed[0].size() == 1)
410 for (IndexType k = 0; k < HalfDimensions * sizeof(std::int16_t) / sizeof(vec_t); ++k)
411 accTileOut[k] = vec_add_16(vec_sub_16(accTileIn[k], columnR0[k]), columnA[k]);
415 const IndexType offsetR1 = HalfDimensions * removed[0][1];
416 auto columnR1 = reinterpret_cast<const vec_t*>(&weights[offsetR1]);
418 for (IndexType k = 0; k < HalfDimensions * sizeof(std::int16_t) / sizeof(vec_t); ++k)
419 accTileOut[k] = vec_sub_16(
420 vec_add_16(accTileIn[k], columnA[k]),
421 vec_add_16(columnR0[k], columnR1[k]));
424 auto accTilePsqtIn = reinterpret_cast<const psqt_vec_t*>(
425 &st->accumulator.psqtAccumulation[Perspective][0]);
426 auto accTilePsqtOut = reinterpret_cast<psqt_vec_t*>(
427 &states_to_update[0]->accumulator.psqtAccumulation[Perspective][0]);
429 const IndexType offsetPsqtR0 = PSQTBuckets * removed[0][0];
430 auto columnPsqtR0 = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offsetPsqtR0]);
431 const IndexType offsetPsqtA = PSQTBuckets * added[0][0];
432 auto columnPsqtA = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offsetPsqtA]);
434 if (removed[0].size() == 1)
436 for (std::size_t k = 0; k < PSQTBuckets * sizeof(std::int32_t) / sizeof(psqt_vec_t); ++k)
437 accTilePsqtOut[k] = vec_add_psqt_32(vec_sub_psqt_32(
438 accTilePsqtIn[k], columnPsqtR0[k]), columnPsqtA[k]);
442 const IndexType offsetPsqtR1 = PSQTBuckets * removed[0][1];
443 auto columnPsqtR1 = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offsetPsqtR1]);
445 for (std::size_t k = 0; k < PSQTBuckets * sizeof(std::int32_t) / sizeof(psqt_vec_t); ++k)
446 accTilePsqtOut[k] = vec_sub_psqt_32(
447 vec_add_psqt_32(accTilePsqtIn[k], columnPsqtA[k]),
448 vec_add_psqt_32(columnPsqtR0[k], columnPsqtR1[k]));
453 for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
456 auto accTileIn = reinterpret_cast<const vec_t*>(
457 &st->accumulator.accumulation[Perspective][j * TileHeight]);
458 for (IndexType k = 0; k < NumRegs; ++k)
459 acc[k] = vec_load(&accTileIn[k]);
461 for (IndexType i = 0; states_to_update[i]; ++i)
463 // Difference calculation for the deactivated features
464 for (const auto index : removed[i])
466 const IndexType offset = HalfDimensions * index + j * TileHeight;
467 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
468 for (IndexType k = 0; k < NumRegs; ++k)
469 acc[k] = vec_sub_16(acc[k], column[k]);
472 // Difference calculation for the activated features
473 for (const auto index : added[i])
475 const IndexType offset = HalfDimensions * index + j * TileHeight;
476 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
477 for (IndexType k = 0; k < NumRegs; ++k)
478 acc[k] = vec_add_16(acc[k], column[k]);
482 auto accTileOut = reinterpret_cast<vec_t*>(
483 &states_to_update[i]->accumulator.accumulation[Perspective][j * TileHeight]);
484 for (IndexType k = 0; k < NumRegs; ++k)
485 vec_store(&accTileOut[k], acc[k]);
489 for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
492 auto accTilePsqtIn = reinterpret_cast<const psqt_vec_t*>(
493 &st->accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
494 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
495 psqt[k] = vec_load_psqt(&accTilePsqtIn[k]);
497 for (IndexType i = 0; states_to_update[i]; ++i)
499 // Difference calculation for the deactivated features
500 for (const auto index : removed[i])
502 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
503 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
504 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
505 psqt[k] = vec_sub_psqt_32(psqt[k], columnPsqt[k]);
508 // Difference calculation for the activated features
509 for (const auto index : added[i])
511 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
512 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
513 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
514 psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
518 auto accTilePsqtOut = reinterpret_cast<psqt_vec_t*>(
519 &states_to_update[i]->accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
520 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
521 vec_store_psqt(&accTilePsqtOut[k], psqt[k]);
526 for (IndexType i = 0; states_to_update[i]; ++i)
528 std::memcpy(states_to_update[i]->accumulator.accumulation[Perspective],
529 st->accumulator.accumulation[Perspective],
530 HalfDimensions * sizeof(BiasType));
532 for (std::size_t k = 0; k < PSQTBuckets; ++k)
533 states_to_update[i]->accumulator.psqtAccumulation[Perspective][k] = st->accumulator.psqtAccumulation[Perspective][k];
535 st = states_to_update[i];
537 // Difference calculation for the deactivated features
538 for (const auto index : removed[i])
540 const IndexType offset = HalfDimensions * index;
542 for (IndexType j = 0; j < HalfDimensions; ++j)
543 st->accumulator.accumulation[Perspective][j] -= weights[offset + j];
545 for (std::size_t k = 0; k < PSQTBuckets; ++k)
546 st->accumulator.psqtAccumulation[Perspective][k] -= psqtWeights[index * PSQTBuckets + k];
549 // Difference calculation for the activated features
550 for (const auto index : added[i])
552 const IndexType offset = HalfDimensions * index;
554 for (IndexType j = 0; j < HalfDimensions; ++j)
555 st->accumulator.accumulation[Perspective][j] += weights[offset + j];
557 for (std::size_t k = 0; k < PSQTBuckets; ++k)
558 st->accumulator.psqtAccumulation[Perspective][k] += psqtWeights[index * PSQTBuckets + k];
564 template<Color Perspective>
565 void update_accumulator_refresh(const Position& pos) const {
567 // Gcc-10.2 unnecessarily spills AVX2 registers if this array
568 // is defined in the VECTOR code below, once in each branch
570 psqt_vec_t psqt[NumPsqtRegs];
573 // Refresh the accumulator
574 // Could be extracted to a separate function because it's done in 2 places,
575 // but it's unclear if compilers would correctly handle register allocation.
576 auto& accumulator = pos.state()->accumulator;
577 accumulator.computed[Perspective] = true;
578 FeatureSet::IndexList active;
579 FeatureSet::append_active_indices<Perspective>(pos, active);
582 for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
584 auto biasesTile = reinterpret_cast<const vec_t*>(
585 &biases[j * TileHeight]);
586 for (IndexType k = 0; k < NumRegs; ++k)
587 acc[k] = biasesTile[k];
589 for (const auto index : active)
591 const IndexType offset = HalfDimensions * index + j * TileHeight;
592 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
594 for (unsigned k = 0; k < NumRegs; ++k)
595 acc[k] = vec_add_16(acc[k], column[k]);
598 auto accTile = reinterpret_cast<vec_t*>(
599 &accumulator.accumulation[Perspective][j * TileHeight]);
600 for (unsigned k = 0; k < NumRegs; k++)
601 vec_store(&accTile[k], acc[k]);
604 for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
606 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
607 psqt[k] = vec_zero_psqt();
609 for (const auto index : active)
611 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
612 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
614 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
615 psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
618 auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
619 &accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]);
620 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
621 vec_store_psqt(&accTilePsqt[k], psqt[k]);
625 std::memcpy(accumulator.accumulation[Perspective], biases,
626 HalfDimensions * sizeof(BiasType));
628 for (std::size_t k = 0; k < PSQTBuckets; ++k)
629 accumulator.psqtAccumulation[Perspective][k] = 0;
631 for (const auto index : active)
633 const IndexType offset = HalfDimensions * index;
635 for (IndexType j = 0; j < HalfDimensions; ++j)
636 accumulator.accumulation[Perspective][j] += weights[offset + j];
638 for (std::size_t k = 0; k < PSQTBuckets; ++k)
639 accumulator.psqtAccumulation[Perspective][k] += psqtWeights[index * PSQTBuckets + k];
644 template<Color Perspective>
645 void hint_common_access_for_perspective(const Position& pos) const {
647 // Works like update_accumulator, but performs less work.
648 // Updates ONLY the accumulator for pos.
650 // Look for a usable accumulator of an earlier position. We keep track
651 // of the estimated gain in terms of features to be added/subtracted.
653 if (pos.state()->accumulator.computed[Perspective])
656 auto [oldest_st, _] = try_find_computed_accumulator<Perspective>(pos);
658 if (oldest_st->accumulator.computed[Perspective])
660 // Only update current position accumulator to minimize work.
661 StateInfo* states_to_update[2] = { pos.state(), nullptr };
662 update_accumulator_incremental<Perspective, 2>(pos, oldest_st, states_to_update);
666 update_accumulator_refresh<Perspective>(pos);
670 template<Color Perspective>
671 void update_accumulator(const Position& pos) const {
673 auto [oldest_st, next] = try_find_computed_accumulator<Perspective>(pos);
675 if (oldest_st->accumulator.computed[Perspective])
680 // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
681 // Currently we update 2 accumulators.
682 // 1. for the current position
683 // 2. the next accumulator after the computed one
684 // The heuristic may change in the future.
685 StateInfo *states_to_update[3] =
686 { next, next == pos.state() ? nullptr : pos.state(), nullptr };
688 update_accumulator_incremental<Perspective, 3>(pos, oldest_st, states_to_update);
692 update_accumulator_refresh<Perspective>(pos);
696 alignas(CacheLineSize) BiasType biases[HalfDimensions];
697 alignas(CacheLineSize) WeightType weights[HalfDimensions * InputDimensions];
698 alignas(CacheLineSize) PSQTWeightType psqtWeights[InputDimensions * PSQTBuckets];
701 } // namespace Stockfish::Eval::NNUE
703 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED