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 for (std::size_t i = 0; i < PSQTBuckets * InputDimensions; ++i)
169 write_little_endian<PSQTWeightType>(stream, psqtWeights[i]);
170 return !stream.fail();
173 // Convert input features
174 std::int32_t transform(const Position& pos, OutputType* output, int bucket) const {
175 update_accumulator(pos, WHITE);
176 update_accumulator(pos, BLACK);
178 const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
179 const auto& accumulation = pos.state()->accumulator.accumulation;
180 const auto& psqtAccumulation = pos.state()->accumulator.psqtAccumulation;
183 psqtAccumulation[static_cast<int>(perspectives[0])][bucket]
184 - psqtAccumulation[static_cast<int>(perspectives[1])][bucket]
187 #if defined(USE_AVX512)
188 constexpr IndexType NumChunks = HalfDimensions / (SimdWidth * 2);
189 static_assert(HalfDimensions % (SimdWidth * 2) == 0);
190 const __m512i Control = _mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7);
191 const __m512i Zero = _mm512_setzero_si512();
193 #elif defined(USE_AVX2)
194 constexpr IndexType NumChunks = HalfDimensions / SimdWidth;
195 constexpr int Control = 0b11011000;
196 const __m256i Zero = _mm256_setzero_si256();
198 #elif defined(USE_SSE2)
199 constexpr IndexType NumChunks = HalfDimensions / SimdWidth;
202 const __m128i Zero = _mm_setzero_si128();
204 const __m128i k0x80s = _mm_set1_epi8(-128);
207 #elif defined(USE_MMX)
208 constexpr IndexType NumChunks = HalfDimensions / SimdWidth;
209 const __m64 k0x80s = _mm_set1_pi8(-128);
211 #elif defined(USE_NEON)
212 constexpr IndexType NumChunks = HalfDimensions / (SimdWidth / 2);
213 const int8x8_t Zero = {0};
216 for (IndexType p = 0; p < 2; ++p) {
217 const IndexType offset = HalfDimensions * p;
219 #if defined(USE_AVX512)
220 auto out = reinterpret_cast<__m512i*>(&output[offset]);
221 for (IndexType j = 0; j < NumChunks; ++j) {
222 __m512i sum0 = _mm512_load_si512(
223 &reinterpret_cast<const __m512i*>(accumulation[perspectives[p]])[j * 2 + 0]);
224 __m512i sum1 = _mm512_load_si512(
225 &reinterpret_cast<const __m512i*>(accumulation[perspectives[p]])[j * 2 + 1]);
226 _mm512_store_si512(&out[j], _mm512_permutexvar_epi64(Control,
227 _mm512_max_epi8(_mm512_packs_epi16(sum0, sum1), Zero)));
230 #elif defined(USE_AVX2)
231 auto out = reinterpret_cast<__m256i*>(&output[offset]);
232 for (IndexType j = 0; j < NumChunks; ++j) {
233 __m256i sum0 = _mm256_load_si256(
234 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]])[j * 2 + 0]);
235 __m256i sum1 = _mm256_load_si256(
236 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]])[j * 2 + 1]);
237 _mm256_store_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
238 _mm256_packs_epi16(sum0, sum1), Zero), Control));
241 #elif defined(USE_SSE2)
242 auto out = reinterpret_cast<__m128i*>(&output[offset]);
243 for (IndexType j = 0; j < NumChunks; ++j) {
244 __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
245 accumulation[perspectives[p]])[j * 2 + 0]);
246 __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
247 accumulation[perspectives[p]])[j * 2 + 1]);
248 const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
250 _mm_store_si128(&out[j],
253 _mm_max_epi8(packedbytes, Zero)
255 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
261 #elif defined(USE_MMX)
262 auto out = reinterpret_cast<__m64*>(&output[offset]);
263 for (IndexType j = 0; j < NumChunks; ++j) {
264 __m64 sum0 = *(&reinterpret_cast<const __m64*>(
265 accumulation[perspectives[p]])[j * 2 + 0]);
266 __m64 sum1 = *(&reinterpret_cast<const __m64*>(
267 accumulation[perspectives[p]])[j * 2 + 1]);
268 const __m64 packedbytes = _mm_packs_pi16(sum0, sum1);
269 out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
272 #elif defined(USE_NEON)
273 const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
274 for (IndexType j = 0; j < NumChunks; ++j) {
275 int16x8_t sum = reinterpret_cast<const int16x8_t*>(
276 accumulation[perspectives[p]])[j];
277 out[j] = vmax_s8(vqmovn_s16(sum), Zero);
281 for (IndexType j = 0; j < HalfDimensions; ++j) {
282 BiasType sum = accumulation[static_cast<int>(perspectives[p])][j];
283 output[offset + j] = static_cast<OutputType>(
284 std::max<int>(0, std::min<int>(127, sum)));
297 void update_accumulator(const Position& pos, const Color perspective) const {
299 // The size must be enough to contain the largest possible update.
300 // That might depend on the feature set and generally relies on the
301 // feature set's update cost calculation to be correct and never
302 // allow updates with more added/removed features than MaxActiveDimensions.
303 using IndexList = ValueList<IndexType, FeatureSet::MaxActiveDimensions>;
306 // Gcc-10.2 unnecessarily spills AVX2 registers if this array
307 // is defined in the VECTOR code below, once in each branch
309 psqt_vec_t psqt[NumPsqtRegs];
312 // Look for a usable accumulator of an earlier position. We keep track
313 // of the estimated gain in terms of features to be added/subtracted.
314 StateInfo *st = pos.state(), *next = nullptr;
315 int gain = FeatureSet::refresh_cost(pos);
316 while (st->accumulator.state[perspective] == EMPTY)
318 // This governs when a full feature refresh is needed and how many
319 // updates are better than just one full refresh.
320 if ( FeatureSet::requires_refresh(st, perspective)
321 || (gain -= FeatureSet::update_cost(st) + 1) < 0)
327 if (st->accumulator.state[perspective] == COMPUTED)
332 // Update incrementally in two steps. First, we update the "next"
333 // accumulator. Then, we update the current accumulator (pos.state()).
335 // Gather all features to be updated.
336 const Square ksq = pos.square<KING>(perspective);
337 IndexList removed[2], added[2];
338 FeatureSet::append_changed_indices(
339 ksq, next, perspective, removed[0], added[0]);
340 for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous)
341 FeatureSet::append_changed_indices(
342 ksq, st2, perspective, removed[1], added[1]);
344 // Mark the accumulators as computed.
345 next->accumulator.state[perspective] = COMPUTED;
346 pos.state()->accumulator.state[perspective] = COMPUTED;
348 // Now update the accumulators listed in states_to_update[], where the last element is a sentinel.
349 StateInfo *states_to_update[3] =
350 { next, next == pos.state() ? nullptr : pos.state(), nullptr };
352 for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
355 auto accTile = reinterpret_cast<vec_t*>(
356 &st->accumulator.accumulation[perspective][j * TileHeight]);
357 for (IndexType k = 0; k < NumRegs; ++k)
358 acc[k] = vec_load(&accTile[k]);
360 for (IndexType i = 0; states_to_update[i]; ++i)
362 // Difference calculation for the deactivated features
363 for (const auto index : removed[i])
365 const IndexType offset = HalfDimensions * index + j * TileHeight;
366 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
367 for (IndexType k = 0; k < NumRegs; ++k)
368 acc[k] = vec_sub_16(acc[k], column[k]);
371 // Difference calculation for the activated features
372 for (const auto index : added[i])
374 const IndexType offset = HalfDimensions * index + j * TileHeight;
375 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
376 for (IndexType k = 0; k < NumRegs; ++k)
377 acc[k] = vec_add_16(acc[k], column[k]);
381 accTile = reinterpret_cast<vec_t*>(
382 &states_to_update[i]->accumulator.accumulation[perspective][j * TileHeight]);
383 for (IndexType k = 0; k < NumRegs; ++k)
384 vec_store(&accTile[k], acc[k]);
388 for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
391 auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
392 &st->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
393 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
394 psqt[k] = vec_load_psqt(&accTilePsqt[k]);
396 for (IndexType i = 0; states_to_update[i]; ++i)
398 // Difference calculation for the deactivated features
399 for (const auto index : removed[i])
401 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
402 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
403 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
404 psqt[k] = vec_sub_psqt_32(psqt[k], columnPsqt[k]);
407 // Difference calculation for the activated features
408 for (const auto index : added[i])
410 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
411 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
412 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
413 psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
417 accTilePsqt = reinterpret_cast<psqt_vec_t*>(
418 &states_to_update[i]->accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
419 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
420 vec_store_psqt(&accTilePsqt[k], psqt[k]);
425 for (IndexType i = 0; states_to_update[i]; ++i)
427 std::memcpy(states_to_update[i]->accumulator.accumulation[perspective],
428 st->accumulator.accumulation[perspective],
429 HalfDimensions * sizeof(BiasType));
431 for (std::size_t k = 0; k < PSQTBuckets; ++k)
432 states_to_update[i]->accumulator.psqtAccumulation[perspective][k] = st->accumulator.psqtAccumulation[perspective][k];
434 st = states_to_update[i];
436 // Difference calculation for the deactivated features
437 for (const auto index : removed[i])
439 const IndexType offset = HalfDimensions * index;
441 for (IndexType j = 0; j < HalfDimensions; ++j)
442 st->accumulator.accumulation[perspective][j] -= weights[offset + j];
444 for (std::size_t k = 0; k < PSQTBuckets; ++k)
445 st->accumulator.psqtAccumulation[perspective][k] -= psqtWeights[index * PSQTBuckets + k];
448 // Difference calculation for the activated features
449 for (const auto index : added[i])
451 const IndexType offset = HalfDimensions * index;
453 for (IndexType j = 0; j < HalfDimensions; ++j)
454 st->accumulator.accumulation[perspective][j] += weights[offset + j];
456 for (std::size_t k = 0; k < PSQTBuckets; ++k)
457 st->accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
464 // Refresh the accumulator
465 auto& accumulator = pos.state()->accumulator;
466 accumulator.state[perspective] = COMPUTED;
468 FeatureSet::append_active_indices(pos, perspective, active);
471 for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j)
473 auto biasesTile = reinterpret_cast<const vec_t*>(
474 &biases[j * TileHeight]);
475 for (IndexType k = 0; k < NumRegs; ++k)
476 acc[k] = biasesTile[k];
478 for (const auto index : active)
480 const IndexType offset = HalfDimensions * index + j * TileHeight;
481 auto column = reinterpret_cast<const vec_t*>(&weights[offset]);
483 for (unsigned k = 0; k < NumRegs; ++k)
484 acc[k] = vec_add_16(acc[k], column[k]);
487 auto accTile = reinterpret_cast<vec_t*>(
488 &accumulator.accumulation[perspective][j * TileHeight]);
489 for (unsigned k = 0; k < NumRegs; k++)
490 vec_store(&accTile[k], acc[k]);
493 for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j)
495 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
496 psqt[k] = vec_zero_psqt();
498 for (const auto index : active)
500 const IndexType offset = PSQTBuckets * index + j * PsqtTileHeight;
501 auto columnPsqt = reinterpret_cast<const psqt_vec_t*>(&psqtWeights[offset]);
503 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
504 psqt[k] = vec_add_psqt_32(psqt[k], columnPsqt[k]);
507 auto accTilePsqt = reinterpret_cast<psqt_vec_t*>(
508 &accumulator.psqtAccumulation[perspective][j * PsqtTileHeight]);
509 for (std::size_t k = 0; k < NumPsqtRegs; ++k)
510 vec_store_psqt(&accTilePsqt[k], psqt[k]);
514 std::memcpy(accumulator.accumulation[perspective], biases,
515 HalfDimensions * sizeof(BiasType));
517 for (std::size_t k = 0; k < PSQTBuckets; ++k)
518 accumulator.psqtAccumulation[perspective][k] = 0;
520 for (const auto index : active)
522 const IndexType offset = HalfDimensions * index;
524 for (IndexType j = 0; j < HalfDimensions; ++j)
525 accumulator.accumulation[perspective][j] += weights[offset + j];
527 for (std::size_t k = 0; k < PSQTBuckets; ++k)
528 accumulator.psqtAccumulation[perspective][k] += psqtWeights[index * PSQTBuckets + k];
538 using BiasType = std::int16_t;
539 using WeightType = std::int16_t;
540 using PSQTWeightType = std::int32_t;
542 alignas(CacheLineSize) BiasType biases[HalfDimensions];
543 alignas(CacheLineSize) WeightType weights[HalfDimensions * InputDimensions];
544 alignas(CacheLineSize) PSQTWeightType psqtWeights[InputDimensions * PSQTBuckets];
547 } // namespace Stockfish::Eval::NNUE
549 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED