2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2021 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
24 #include "nnue_common.h"
25 #include "nnue_architecture.h"
29 #include <cstring> // std::memset()
31 namespace Stockfish::Eval::NNUE {
33 // If vector instructions are enabled, we update and refresh the
34 // accumulator tile by tile such that each tile fits in the CPU's
38 static_assert(PSQTBuckets == 8, "Assumed by the current choice of constants.");
41 typedef __m512i vec_t;
42 typedef __m256i psqt_vec_t;
43 #define vec_load(a) _mm512_load_si512(a)
44 #define vec_store(a,b) _mm512_store_si512(a,b)
45 #define vec_add_16(a,b) _mm512_add_epi16(a,b)
46 #define vec_sub_16(a,b) _mm512_sub_epi16(a,b)
47 #define vec_load_psqt(a) _mm256_load_si256(a)
48 #define vec_store_psqt(a,b) _mm256_store_si256(a,b)
49 #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
50 #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
51 #define vec_zero_psqt() _mm256_setzero_si256()
52 static constexpr IndexType NumRegs = 8; // only 8 are needed
53 static constexpr IndexType NumPsqtRegs = 1;
56 typedef __m256i vec_t;
57 typedef __m256i psqt_vec_t;
58 #define vec_load(a) _mm256_load_si256(a)
59 #define vec_store(a,b) _mm256_store_si256(a,b)
60 #define vec_add_16(a,b) _mm256_add_epi16(a,b)
61 #define vec_sub_16(a,b) _mm256_sub_epi16(a,b)
62 #define vec_load_psqt(a) _mm256_load_si256(a)
63 #define vec_store_psqt(a,b) _mm256_store_si256(a,b)
64 #define vec_add_psqt_32(a,b) _mm256_add_epi32(a,b)
65 #define vec_sub_psqt_32(a,b) _mm256_sub_epi32(a,b)
66 #define vec_zero_psqt() _mm256_setzero_si256()
67 static constexpr IndexType NumRegs = 16;
68 static constexpr IndexType NumPsqtRegs = 1;
71 typedef __m128i vec_t;
72 typedef __m128i psqt_vec_t;
73 #define vec_load(a) (*(a))
74 #define vec_store(a,b) *(a)=(b)
75 #define vec_add_16(a,b) _mm_add_epi16(a,b)
76 #define vec_sub_16(a,b) _mm_sub_epi16(a,b)
77 #define vec_load_psqt(a) (*(a))
78 #define vec_store_psqt(a,b) *(a)=(b)
79 #define vec_add_psqt_32(a,b) _mm_add_epi32(a,b)
80 #define vec_sub_psqt_32(a,b) _mm_sub_epi32(a,b)
81 #define vec_zero_psqt() _mm_setzero_si128()
82 static constexpr IndexType NumRegs = Is64Bit ? 16 : 8;
83 static constexpr IndexType NumPsqtRegs = 2;
87 typedef __m64 psqt_vec_t;
88 #define vec_load(a) (*(a))
89 #define vec_store(a,b) *(a)=(b)
90 #define vec_add_16(a,b) _mm_add_pi16(a,b)
91 #define vec_sub_16(a,b) _mm_sub_pi16(a,b)
92 #define vec_load_psqt(a) (*(a))
93 #define vec_store_psqt(a,b) *(a)=(b)
94 #define vec_add_psqt_32(a,b) _mm_add_pi32(a,b)
95 #define vec_sub_psqt_32(a,b) _mm_sub_pi32(a,b)
96 #define vec_zero_psqt() _mm_setzero_si64()
97 static constexpr IndexType NumRegs = 8;
98 static constexpr IndexType NumPsqtRegs = 4;
101 typedef int16x8_t vec_t;
102 typedef int32x4_t psqt_vec_t;
103 #define vec_load(a) (*(a))
104 #define vec_store(a,b) *(a)=(b)
105 #define vec_add_16(a,b) vaddq_s16(a,b)
106 #define vec_sub_16(a,b) vsubq_s16(a,b)
107 #define vec_load_psqt(a) (*(a))
108 #define vec_store_psqt(a,b) *(a)=(b)
109 #define vec_add_psqt_32(a,b) vaddq_s32(a,b)
110 #define vec_sub_psqt_32(a,b) vsubq_s32(a,b)
111 #define vec_zero_psqt() psqt_vec_t{0}
112 static constexpr IndexType NumRegs = 16;
113 static constexpr IndexType NumPsqtRegs = 2;
120 // Input feature converter
121 class FeatureTransformer {
124 // Number of output dimensions for one side
125 static constexpr IndexType HalfDimensions = TransformedFeatureDimensions;
128 static constexpr IndexType TileHeight = NumRegs * sizeof(vec_t) / 2;
129 static constexpr IndexType PsqtTileHeight = NumPsqtRegs * sizeof(psqt_vec_t) / 4;
130 static_assert(HalfDimensions % TileHeight == 0, "TileHeight must divide HalfDimensions");
131 static_assert(PSQTBuckets % PsqtTileHeight == 0, "PsqtTileHeight must divide PSQTBuckets");
136 using OutputType = TransformedFeatureType;
138 // Number of input/output dimensions
139 static constexpr IndexType InputDimensions = FeatureSet::Dimensions;
140 static constexpr IndexType OutputDimensions = HalfDimensions * 2;
142 // Size of forward propagation buffer
143 static constexpr std::size_t BufferSize =
144 OutputDimensions * sizeof(OutputType);
146 // Hash value embedded in the evaluation file
147 static constexpr std::uint32_t get_hash_value() {
148 return FeatureSet::HashValue ^ OutputDimensions;
151 // Read network parameters
152 bool read_parameters(std::istream& stream) {
153 for (std::size_t i = 0; i < HalfDimensions; ++i)
154 biases[i] = read_little_endian<BiasType>(stream);
155 for (std::size_t i = 0; i < HalfDimensions * InputDimensions; ++i)
156 weights[i] = read_little_endian<WeightType>(stream);
157 for (std::size_t i = 0; i < PSQTBuckets * InputDimensions; ++i)
158 psqtWeights[i] = read_little_endian<PSQTWeightType>(stream);
159 return !stream.fail();
162 // Write network parameters
163 bool write_parameters(std::ostream& stream) const {
164 for (std::size_t i = 0; i < HalfDimensions; ++i)
165 write_little_endian<BiasType>(stream, biases[i]);
166 for (std::size_t i = 0; i < HalfDimensions * InputDimensions; ++i)
167 write_little_endian<WeightType>(stream, weights[i]);
168 return !stream.fail();
171 // Convert input features
172 std::pair<std::int32_t, bool> transform(const Position& pos, OutputType* output, int bucket, Value lazyThreshold) const {
173 update_accumulator(pos, WHITE);
174 update_accumulator(pos, BLACK);
176 const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
177 const auto& accumulation = pos.state()->accumulator.accumulation;
178 const auto& psqtAccumulation = pos.state()->accumulator.psqtAccumulation;
181 psqtAccumulation[static_cast<int>(perspectives[0])][bucket]
182 - psqtAccumulation[static_cast<int>(perspectives[1])][bucket]
185 if (abs(psqt) > (int)lazyThreshold * OutputScale)
186 return { psqt, true };
188 #if defined(USE_AVX512)
189 constexpr IndexType NumChunks = HalfDimensions / (SimdWidth * 2);
190 static_assert(HalfDimensions % (SimdWidth * 2) == 0);
191 const __m512i Control = _mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7);
192 const __m512i Zero = _mm512_setzero_si512();
194 #elif defined(USE_AVX2)
195 constexpr IndexType NumChunks = HalfDimensions / SimdWidth;
196 constexpr int Control = 0b11011000;
197 const __m256i Zero = _mm256_setzero_si256();
199 #elif defined(USE_SSE2)
200 constexpr IndexType NumChunks = HalfDimensions / SimdWidth;
203 const __m128i Zero = _mm_setzero_si128();
205 const __m128i k0x80s = _mm_set1_epi8(-128);
208 #elif defined(USE_MMX)
209 constexpr IndexType NumChunks = HalfDimensions / SimdWidth;
210 const __m64 k0x80s = _mm_set1_pi8(-128);
212 #elif defined(USE_NEON)
213 constexpr IndexType NumChunks = HalfDimensions / (SimdWidth / 2);
214 const int8x8_t Zero = {0};
217 for (IndexType p = 0; p < 2; ++p) {
218 const IndexType offset = HalfDimensions * p;
220 #if defined(USE_AVX512)
221 auto out = reinterpret_cast<__m512i*>(&output[offset]);
222 for (IndexType j = 0; j < NumChunks; ++j) {
223 __m512i sum0 = _mm512_load_si512(
224 &reinterpret_cast<const __m512i*>(accumulation[perspectives[p]])[j * 2 + 0]);
225 __m512i sum1 = _mm512_load_si512(
226 &reinterpret_cast<const __m512i*>(accumulation[perspectives[p]])[j * 2 + 1]);
227 _mm512_store_si512(&out[j], _mm512_permutexvar_epi64(Control,
228 _mm512_max_epi8(_mm512_packs_epi16(sum0, sum1), Zero)));
231 #elif defined(USE_AVX2)
232 auto out = reinterpret_cast<__m256i*>(&output[offset]);
233 for (IndexType j = 0; j < NumChunks; ++j) {
234 __m256i sum0 = _mm256_load_si256(
235 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]])[j * 2 + 0]);
236 __m256i sum1 = _mm256_load_si256(
237 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]])[j * 2 + 1]);
238 _mm256_store_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
239 _mm256_packs_epi16(sum0, sum1), Zero), Control));
242 #elif defined(USE_SSE2)
243 auto out = reinterpret_cast<__m128i*>(&output[offset]);
244 for (IndexType j = 0; j < NumChunks; ++j) {
245 __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
246 accumulation[perspectives[p]])[j * 2 + 0]);
247 __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
248 accumulation[perspectives[p]])[j * 2 + 1]);
249 const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
251 _mm_store_si128(&out[j],
254 _mm_max_epi8(packedbytes, Zero)
256 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
262 #elif defined(USE_MMX)
263 auto out = reinterpret_cast<__m64*>(&output[offset]);
264 for (IndexType j = 0; j < NumChunks; ++j) {
265 __m64 sum0 = *(&reinterpret_cast<const __m64*>(
266 accumulation[perspectives[p]])[j * 2 + 0]);
267 __m64 sum1 = *(&reinterpret_cast<const __m64*>(
268 accumulation[perspectives[p]])[j * 2 + 1]);
269 const __m64 packedbytes = _mm_packs_pi16(sum0, sum1);
270 out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
273 #elif defined(USE_NEON)
274 const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
275 for (IndexType j = 0; j < NumChunks; ++j) {
276 int16x8_t sum = reinterpret_cast<const int16x8_t*>(
277 accumulation[perspectives[p]])[j];
278 out[j] = vmax_s8(vqmovn_s16(sum), Zero);
282 for (IndexType j = 0; j < HalfDimensions; ++j) {
283 BiasType sum = accumulation[static_cast<int>(perspectives[p])][j];
284 output[offset + j] = static_cast<OutputType>(
285 std::max<int>(0, std::min<int>(127, sum)));
294 return { psqt, false };
298 void update_accumulator(const Position& pos, const Color perspective) const {
300 // The size must be enough to contain the largest possible update.
301 // That might depend on the feature set and generally relies on the
302 // feature set's update cost calculation to be correct and never
303 // allow updates with more added/removed features than MaxActiveDimensions.
304 using IndexList = ValueList<IndexType, FeatureSet::MaxActiveDimensions>;
307 // Gcc-10.2 unnecessarily spills AVX2 registers if this array
308 // is defined in the VECTOR code below, once in each branch
310 psqt_vec_t psqt[NumPsqtRegs];
313 // Look for a usable accumulator of an earlier position. We keep track
314 // of the estimated gain in terms of features to be added/subtracted.
315 StateInfo *st = pos.state(), *next = nullptr;
316 int gain = FeatureSet::refresh_cost(pos);
317 while (st->accumulator.state[perspective] == EMPTY)
319 // This governs when a full feature refresh is needed and how many
320 // updates are better than just one full refresh.
321 if ( FeatureSet::requires_refresh(st, perspective)
322 || (gain -= FeatureSet::update_cost(st) + 1) < 0)
328 if (st->accumulator.state[perspective] == COMPUTED)
333 // Update incrementally in two steps. First, we update the "next"
334 // accumulator. Then, we update the current accumulator (pos.state()).
336 // Gather all features to be updated.
337 const Square ksq = pos.square<KING>(perspective);
338 IndexList removed[2], added[2];
339 FeatureSet::append_changed_indices(
340 ksq, next, perspective, removed[0], added[0]);
341 for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous)
342 FeatureSet::append_changed_indices(
343 ksq, st2, perspective, removed[1], added[1]);
345 // Mark the accumulators as computed.
346 next->accumulator.state[perspective] = COMPUTED;
347 pos.state()->accumulator.state[perspective] = COMPUTED;
349 // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
350 StateInfo *states_to_update[3] =
351 { next, next == pos.state() ? nullptr : pos.state(), nullptr };
353 for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
356 auto accTile = reinterpret_cast<vec_t*>(
357 &st->accumulator.accumulation[perspective][j * TileHeight]);
358 for (IndexType k = 0; k < NumRegs; ++k)
359 acc[k] = vec_load(&accTile[k]);
361 for (IndexType i = 0; states_to_update[i]; ++i)
363 // Difference calculation for the deactivated features
364 for (const auto index : removed[i])
366 const IndexType offset = HalfDimensions * index + j * TileHeight;
367 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
368 for (IndexType k = 0; k < NumRegs; ++k)
369 acc[k] = vec_sub_16(acc[k], column[k]);
372 // Difference calculation for the activated features
373 for (const auto index : added[i])
375 const IndexType offset = HalfDimensions * index + j * TileHeight;
376 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
377 for (IndexType k = 0; k < NumRegs; ++k)
378 acc[k] = vec_add_16(acc[k], column[k]);
382 accTile = reinterpret_cast<vec_t*>(
383 &states_to_update[i]->accumulator.accumulation[perspective][j * TileHeight]);
384 for (IndexType k = 0; k < NumRegs; ++k)
385 vec_store(&accTile[k], acc[k]);
389 for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
392 auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
393 &st->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
394 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
395 psqt[k] = vec_load_psqt(&accTilePsqt[k]);
397 for (IndexType i = 0; states_to_update[i]; ++i)
399 // Difference calculation for the deactivated features
400 for (const auto index : removed[i])
402 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
403 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
404 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
405 psqt[k] = vec_sub_psqt_32(psqt[k], columnPsqt[k]);
408 // Difference calculation for the activated features
409 for (const auto index : added[i])
411 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
412 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
413 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
414 psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
418 accTilePsqt = reinterpret_cast<psqt_vec_t*>(
419 &states_to_update[i]->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
420 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
421 vec_store_psqt(&accTilePsqt[k], psqt[k]);
426 for (IndexType i = 0; states_to_update[i]; ++i)
428 std::memcpy(states_to_update[i]->accumulator.accumulation[perspective],
429 st->accumulator.accumulation[perspective],
430 HalfDimensions * sizeof(BiasType));
432 for (std::size_t k = 0; k < PSQTBuckets; ++k)
433 states_to_update[i]->accumulator.psqtAccumulation[perspective][k] = st->accumulator.psqtAccumulation[perspective][k];
435 st = states_to_update[i];
437 // Difference calculation for the deactivated features
438 for (const auto index : removed[i])
440 const IndexType offset = HalfDimensions * index;
442 for (IndexType j = 0; j < HalfDimensions; ++j)
443 st->accumulator.accumulation[perspective][j] -= weights[offset + j];
445 for (std::size_t k = 0; k < PSQTBuckets; ++k)
446 st->accumulator.psqtAccumulation[perspective][k] -= psqtWeights[index * PSQTBuckets + k];
449 // Difference calculation for the activated features
450 for (const auto index : added[i])
452 const IndexType offset = HalfDimensions * index;
454 for (IndexType j = 0; j < HalfDimensions; ++j)
455 st->accumulator.accumulation[perspective][j] += weights[offset + j];
457 for (std::size_t k = 0; k < PSQTBuckets; ++k)
458 st->accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
465 // Refresh the accumulator
466 auto& accumulator = pos.state()->accumulator;
467 accumulator.state[perspective] = COMPUTED;
469 FeatureSet::append_active_indices(pos, perspective, active);
472 for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
474 auto biasesTile = reinterpret_cast<const vec_t*>(
475 &biases[j * TileHeight]);
476 for (IndexType k = 0; k < NumRegs; ++k)
477 acc[k] = biasesTile[k];
479 for (const auto index : active)
481 const IndexType offset = HalfDimensions * index + j * TileHeight;
482 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
484 for (unsigned k = 0; k < NumRegs; ++k)
485 acc[k] = vec_add_16(acc[k], column[k]);
488 auto accTile = reinterpret_cast<vec_t*>(
489 &accumulator.accumulation[perspective][j * TileHeight]);
490 for (unsigned k = 0; k < NumRegs; k++)
491 vec_store(&accTile[k], acc[k]);
494 for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
496 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
497 psqt[k] = vec_zero_psqt();
499 for (const auto index : active)
501 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
502 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
504 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
505 psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
508 auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
509 &accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
510 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
511 vec_store_psqt(&accTilePsqt[k], psqt[k]);
515 std::memcpy(accumulator.accumulation[perspective], biases,
516 HalfDimensions * sizeof(BiasType));
518 for (std::size_t k = 0; k < PSQTBuckets; ++k)
519 accumulator.psqtAccumulation[perspective][k] = 0;
521 for (const auto index : active)
523 const IndexType offset = HalfDimensions * index;
525 for (IndexType j = 0; j < HalfDimensions; ++j)
526 accumulator.accumulation[perspective][j] += weights[offset + j];
528 for (std::size_t k = 0; k < PSQTBuckets; ++k)
529 accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
539 using BiasType = std::int16_t;
540 using WeightType = std::int16_t;
541 using PSQTWeightType = std::int32_t;
543 alignas(CacheLineSize) BiasType biases[HalfDimensions];
544 alignas(CacheLineSize) WeightType weights[HalfDimensions * InputDimensions];
545 alignas(CacheLineSize) PSQTWeightType psqtWeights[InputDimensions * PSQTBuckets];
548 } // namespace Stockfish::Eval::NNUE
550 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED