2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2020 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"
26 #include "features/index_list.h"
28 #include <cstring> // std::memset()
30 namespace Eval::NNUE {
32 // If vector instructions are enabled, we update and refresh the
33 // accumulator tile by tile such that each tile fits in the CPU's
38 typedef __m512i vec_t;
39 #define vec_load(a) _mm512_load_si512(a)
40 #define vec_store(a,b) _mm512_store_si512(a,b)
41 #define vec_add_16(a,b) _mm512_add_epi16(a,b)
42 #define vec_sub_16(a,b) _mm512_sub_epi16(a,b)
43 static constexpr IndexType kNumRegs = 8; // only 8 are needed
46 typedef __m256i vec_t;
47 #define vec_load(a) _mm256_load_si256(a)
48 #define vec_store(a,b) _mm256_store_si256(a,b)
49 #define vec_add_16(a,b) _mm256_add_epi16(a,b)
50 #define vec_sub_16(a,b) _mm256_sub_epi16(a,b)
51 static constexpr IndexType kNumRegs = 16;
54 typedef __m128i vec_t;
55 #define vec_load(a) (*(a))
56 #define vec_store(a,b) *(a)=(b)
57 #define vec_add_16(a,b) _mm_add_epi16(a,b)
58 #define vec_sub_16(a,b) _mm_sub_epi16(a,b)
59 static constexpr IndexType kNumRegs = Is64Bit ? 16 : 8;
63 #define vec_load(a) (*(a))
64 #define vec_store(a,b) *(a)=(b)
65 #define vec_add_16(a,b) _mm_add_pi16(a,b)
66 #define vec_sub_16(a,b) _mm_sub_pi16(a,b)
67 static constexpr IndexType kNumRegs = 8;
70 typedef int16x8_t vec_t;
71 #define vec_load(a) (*(a))
72 #define vec_store(a,b) *(a)=(b)
73 #define vec_add_16(a,b) vaddq_s16(a,b)
74 #define vec_sub_16(a,b) vsubq_s16(a,b)
75 static constexpr IndexType kNumRegs = 16;
82 // Input feature converter
83 class FeatureTransformer {
86 // Number of output dimensions for one side
87 static constexpr IndexType kHalfDimensions = kTransformedFeatureDimensions;
90 static constexpr IndexType kTileHeight = kNumRegs * sizeof(vec_t) / 2;
91 static_assert(kHalfDimensions % kTileHeight == 0, "kTileHeight must divide kHalfDimensions");
96 using OutputType = TransformedFeatureType;
98 // Number of input/output dimensions
99 static constexpr IndexType kInputDimensions = RawFeatures::kDimensions;
100 static constexpr IndexType kOutputDimensions = kHalfDimensions * 2;
102 // Size of forward propagation buffer
103 static constexpr std::size_t kBufferSize =
104 kOutputDimensions * sizeof(OutputType);
106 // Hash value embedded in the evaluation file
107 static constexpr std::uint32_t GetHashValue() {
109 return RawFeatures::kHashValue ^ kOutputDimensions;
112 // Read network parameters
113 bool ReadParameters(std::istream& stream) {
115 for (std::size_t i = 0; i < kHalfDimensions; ++i)
116 biases_[i] = read_little_endian<BiasType>(stream);
117 for (std::size_t i = 0; i < kHalfDimensions * kInputDimensions; ++i)
118 weights_[i] = read_little_endian<WeightType>(stream);
119 return !stream.fail();
122 // Convert input features
123 void Transform(const Position& pos, OutputType* output) const {
125 UpdateAccumulator(pos, WHITE);
126 UpdateAccumulator(pos, BLACK);
128 const auto& accumulation = pos.state()->accumulator.accumulation;
130 #if defined(USE_AVX512)
131 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth * 2);
132 static_assert(kHalfDimensions % (kSimdWidth * 2) == 0);
133 const __m512i kControl = _mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7);
134 const __m512i kZero = _mm512_setzero_si512();
136 #elif defined(USE_AVX2)
137 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
138 constexpr int kControl = 0b11011000;
139 const __m256i kZero = _mm256_setzero_si256();
141 #elif defined(USE_SSE2)
142 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
145 const __m128i kZero = _mm_setzero_si128();
147 const __m128i k0x80s = _mm_set1_epi8(-128);
150 #elif defined(USE_MMX)
151 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
152 const __m64 k0x80s = _mm_set1_pi8(-128);
154 #elif defined(USE_NEON)
155 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
156 const int8x8_t kZero = {0};
159 const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
160 for (IndexType p = 0; p < 2; ++p) {
161 const IndexType offset = kHalfDimensions * p;
163 #if defined(USE_AVX512)
164 auto out = reinterpret_cast<__m512i*>(&output[offset]);
165 for (IndexType j = 0; j < kNumChunks; ++j) {
166 __m512i sum0 = _mm512_load_si512(
167 &reinterpret_cast<const __m512i*>(accumulation[perspectives[p]][0])[j * 2 + 0]);
168 __m512i sum1 = _mm512_load_si512(
169 &reinterpret_cast<const __m512i*>(accumulation[perspectives[p]][0])[j * 2 + 1]);
170 _mm512_store_si512(&out[j], _mm512_permutexvar_epi64(kControl,
171 _mm512_max_epi8(_mm512_packs_epi16(sum0, sum1), kZero)));
174 #elif defined(USE_AVX2)
175 auto out = reinterpret_cast<__m256i*>(&output[offset]);
176 for (IndexType j = 0; j < kNumChunks; ++j) {
177 __m256i sum0 = _mm256_load_si256(
178 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 0]);
179 __m256i sum1 = _mm256_load_si256(
180 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 1]);
181 _mm256_store_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
182 _mm256_packs_epi16(sum0, sum1), kZero), kControl));
185 #elif defined(USE_SSE2)
186 auto out = reinterpret_cast<__m128i*>(&output[offset]);
187 for (IndexType j = 0; j < kNumChunks; ++j) {
188 __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
189 accumulation[perspectives[p]][0])[j * 2 + 0]);
190 __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
191 accumulation[perspectives[p]][0])[j * 2 + 1]);
192 const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
194 _mm_store_si128(&out[j],
197 _mm_max_epi8(packedbytes, kZero)
199 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
205 #elif defined(USE_MMX)
206 auto out = reinterpret_cast<__m64*>(&output[offset]);
207 for (IndexType j = 0; j < kNumChunks; ++j) {
208 __m64 sum0 = *(&reinterpret_cast<const __m64*>(
209 accumulation[perspectives[p]][0])[j * 2 + 0]);
210 __m64 sum1 = *(&reinterpret_cast<const __m64*>(
211 accumulation[perspectives[p]][0])[j * 2 + 1]);
212 const __m64 packedbytes = _mm_packs_pi16(sum0, sum1);
213 out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
216 #elif defined(USE_NEON)
217 const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
218 for (IndexType j = 0; j < kNumChunks; ++j) {
219 int16x8_t sum = reinterpret_cast<const int16x8_t*>(
220 accumulation[perspectives[p]][0])[j];
221 out[j] = vmax_s8(vqmovn_s16(sum), kZero);
225 for (IndexType j = 0; j < kHalfDimensions; ++j) {
226 BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
227 output[offset + j] = static_cast<OutputType>(
228 std::max<int>(0, std::min<int>(127, sum)));
239 void UpdateAccumulator(const Position& pos, const Color c) const {
242 // Gcc-10.2 unnecessarily spills AVX2 registers if this array
243 // is defined in the VECTOR code below, once in each branch
247 // Look for a usable accumulator of an earlier position. We keep track
248 // of the estimated gain in terms of features to be added/subtracted.
249 StateInfo *st = pos.state(), *next = nullptr;
250 int gain = pos.count<ALL_PIECES>() - 2;
251 while (st->accumulator.state[c] == EMPTY)
253 auto& dp = st->dirtyPiece;
254 // The first condition tests whether an incremental update is
255 // possible at all: if this side's king has moved, it is not possible.
256 static_assert(std::is_same_v<RawFeatures::SortedTriggerSet,
257 Features::CompileTimeList<Features::TriggerEvent, Features::TriggerEvent::kFriendKingMoved>>,
258 "Current code assumes that only kFriendlyKingMoved refresh trigger is being used.");
259 if ( dp.piece[0] == make_piece(c, KING)
260 || (gain -= dp.dirty_num + 1) < 0)
266 if (st->accumulator.state[c] == COMPUTED)
271 // Update incrementally in two steps. First, we update the "next"
272 // accumulator. Then, we update the current accumulator (pos.state()).
274 // Gather all features to be updated. This code assumes HalfKP features
275 // only and doesn't support refresh triggers.
276 static_assert(std::is_same_v<Features::FeatureSet<Features::HalfKP<Features::Side::kFriend>>,
278 Features::IndexList removed[2], added[2];
279 Features::HalfKP<Features::Side::kFriend>::AppendChangedIndices(pos,
280 next->dirtyPiece, c, &removed[0], &added[0]);
281 for (StateInfo *st2 = pos.state(); st2 != next; st2 = st2->previous)
282 Features::HalfKP<Features::Side::kFriend>::AppendChangedIndices(pos,
283 st2->dirtyPiece, c, &removed[1], &added[1]);
285 // Mark the accumulators as computed.
286 next->accumulator.state[c] = COMPUTED;
287 pos.state()->accumulator.state[c] = COMPUTED;
289 // Now update the accumulators listed in info[], where the last element is a sentinel.
291 { next, next == pos.state() ? nullptr : pos.state(), nullptr };
293 for (IndexType j = 0; j < kHalfDimensions / kTileHeight; ++j)
296 auto accTile = reinterpret_cast<vec_t*>(
297 &st->accumulator.accumulation[c][0][j * kTileHeight]);
298 for (IndexType k = 0; k < kNumRegs; ++k)
299 acc[k] = vec_load(&accTile[k]);
301 for (IndexType i = 0; info[i]; ++i)
303 // Difference calculation for the deactivated features
304 for (const auto index : removed[i])
306 const IndexType offset = kHalfDimensions * index + j * kTileHeight;
307 auto column = reinterpret_cast<const vec_t*>(&weights_[offset]);
308 for (IndexType k = 0; k < kNumRegs; ++k)
309 acc[k] = vec_sub_16(acc[k], column[k]);
312 // Difference calculation for the activated features
313 for (const auto index : added[i])
315 const IndexType offset = kHalfDimensions * index + j * kTileHeight;
316 auto column = reinterpret_cast<const vec_t*>(&weights_[offset]);
317 for (IndexType k = 0; k < kNumRegs; ++k)
318 acc[k] = vec_add_16(acc[k], column[k]);
322 accTile = reinterpret_cast<vec_t*>(
323 &info[i]->accumulator.accumulation[c][0][j * kTileHeight]);
324 for (IndexType k = 0; k < kNumRegs; ++k)
325 vec_store(&accTile[k], acc[k]);
330 for (IndexType i = 0; info[i]; ++i)
332 std::memcpy(info[i]->accumulator.accumulation[c][0],
333 st->accumulator.accumulation[c][0],
334 kHalfDimensions * sizeof(BiasType));
337 // Difference calculation for the deactivated features
338 for (const auto index : removed[i])
340 const IndexType offset = kHalfDimensions * index;
342 for (IndexType j = 0; j < kHalfDimensions; ++j)
343 st->accumulator.accumulation[c][0][j] -= weights_[offset + j];
346 // Difference calculation for the activated features
347 for (const auto index : added[i])
349 const IndexType offset = kHalfDimensions * index;
351 for (IndexType j = 0; j < kHalfDimensions; ++j)
352 st->accumulator.accumulation[c][0][j] += weights_[offset + j];
359 // Refresh the accumulator
360 auto& accumulator = pos.state()->accumulator;
361 accumulator.state[c] = COMPUTED;
362 Features::IndexList active;
363 Features::HalfKP<Features::Side::kFriend>::AppendActiveIndices(pos, c, &active);
366 for (IndexType j = 0; j < kHalfDimensions / kTileHeight; ++j)
368 auto biasesTile = reinterpret_cast<const vec_t*>(
369 &biases_[j * kTileHeight]);
370 for (IndexType k = 0; k < kNumRegs; ++k)
371 acc[k] = biasesTile[k];
373 for (const auto index : active)
375 const IndexType offset = kHalfDimensions * index + j * kTileHeight;
376 auto column = reinterpret_cast<const vec_t*>(&weights_[offset]);
378 for (unsigned k = 0; k < kNumRegs; ++k)
379 acc[k] = vec_add_16(acc[k], column[k]);
382 auto accTile = reinterpret_cast<vec_t*>(
383 &accumulator.accumulation[c][0][j * kTileHeight]);
384 for (unsigned k = 0; k < kNumRegs; k++)
385 vec_store(&accTile[k], acc[k]);
389 std::memcpy(accumulator.accumulation[c][0], biases_,
390 kHalfDimensions * sizeof(BiasType));
392 for (const auto index : active)
394 const IndexType offset = kHalfDimensions * index;
396 for (IndexType j = 0; j < kHalfDimensions; ++j)
397 accumulator.accumulation[c][0][j] += weights_[offset + j];
407 using BiasType = std::int16_t;
408 using WeightType = std::int16_t;
410 alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
411 alignas(kCacheLineSize)
412 WeightType weights_[kHalfDimensions * kInputDimensions];
415 } // namespace Eval::NNUE
417 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED