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_loadA_si512(a)
40 #define vec_store(a,b) _mm512_storeA_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_loadA_si256(a)
48 #define vec_store(a,b) _mm256_storeA_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 // Proceed with the difference calculation if possible
123 bool UpdateAccumulatorIfPossible(const Position& pos) const {
125 const auto now = pos.state();
126 if (now->accumulator.computed_accumulation)
129 const auto prev = now->previous;
131 if (prev->accumulator.computed_accumulation) {
132 UpdateAccumulator(pos);
134 } else if (prev->previous && prev->previous->accumulator.computed_accumulation) {
135 UpdateAccumulator(pos);
143 // Convert input features
144 void Transform(const Position& pos, OutputType* output) const {
146 if (!UpdateAccumulatorIfPossible(pos))
147 RefreshAccumulator(pos);
149 const auto& accumulation = pos.state()->accumulator.accumulation;
151 #if defined(USE_AVX2)
152 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
153 constexpr int kControl = 0b11011000;
154 const __m256i kZero = _mm256_setzero_si256();
156 #elif defined(USE_SSE2)
157 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
160 const __m128i kZero = _mm_setzero_si128();
162 const __m128i k0x80s = _mm_set1_epi8(-128);
165 #elif defined(USE_MMX)
166 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
167 const __m64 k0x80s = _mm_set1_pi8(-128);
169 #elif defined(USE_NEON)
170 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
171 const int8x8_t kZero = {0};
174 const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
175 for (IndexType p = 0; p < 2; ++p) {
176 const IndexType offset = kHalfDimensions * p;
178 #if defined(USE_AVX2)
179 auto out = reinterpret_cast<__m256i*>(&output[offset]);
180 for (IndexType j = 0; j < kNumChunks; ++j) {
181 __m256i sum0 = _mm256_loadA_si256(
182 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 0]);
183 __m256i sum1 = _mm256_loadA_si256(
184 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 1]);
185 _mm256_storeA_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
186 _mm256_packs_epi16(sum0, sum1), kZero), kControl));
189 #elif defined(USE_SSE2)
190 auto out = reinterpret_cast<__m128i*>(&output[offset]);
191 for (IndexType j = 0; j < kNumChunks; ++j) {
192 __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
193 accumulation[perspectives[p]][0])[j * 2 + 0]);
194 __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
195 accumulation[perspectives[p]][0])[j * 2 + 1]);
196 const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
198 _mm_store_si128(&out[j],
201 _mm_max_epi8(packedbytes, kZero)
203 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
209 #elif defined(USE_MMX)
210 auto out = reinterpret_cast<__m64*>(&output[offset]);
211 for (IndexType j = 0; j < kNumChunks; ++j) {
212 __m64 sum0 = *(&reinterpret_cast<const __m64*>(
213 accumulation[perspectives[p]][0])[j * 2 + 0]);
214 __m64 sum1 = *(&reinterpret_cast<const __m64*>(
215 accumulation[perspectives[p]][0])[j * 2 + 1]);
216 const __m64 packedbytes = _mm_packs_pi16(sum0, sum1);
217 out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
220 #elif defined(USE_NEON)
221 const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
222 for (IndexType j = 0; j < kNumChunks; ++j) {
223 int16x8_t sum = reinterpret_cast<const int16x8_t*>(
224 accumulation[perspectives[p]][0])[j];
225 out[j] = vmax_s8(vqmovn_s16(sum), kZero);
229 for (IndexType j = 0; j < kHalfDimensions; ++j) {
230 BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
231 output[offset + j] = static_cast<OutputType>(
232 std::max<int>(0, std::min<int>(127, sum)));
243 // Calculate cumulative value without using difference calculation
244 void RefreshAccumulator(const Position& pos) const {
246 auto& accumulator = pos.state()->accumulator;
248 Features::IndexList active_indices[2];
249 RawFeatures::AppendActiveIndices(pos, kRefreshTriggers[i],
251 for (Color perspective : { WHITE, BLACK }) {
253 for (unsigned j = 0; j < kHalfDimensions / kTileHeight; ++j) {
254 auto biasesTile = reinterpret_cast<const vec_t*>(
255 &biases_[j * kTileHeight]);
256 auto accTile = reinterpret_cast<vec_t*>(
257 &accumulator.accumulation[perspective][i][j * kTileHeight]);
260 for (unsigned k = 0; k < kNumRegs; ++k)
261 acc[k] = biasesTile[k];
263 for (const auto index : active_indices[perspective]) {
264 const IndexType offset = kHalfDimensions * index + j * kTileHeight;
265 auto column = reinterpret_cast<const vec_t*>(&weights_[offset]);
267 for (unsigned k = 0; k < kNumRegs; ++k)
268 acc[k] = vec_add_16(acc[k], column[k]);
271 for (unsigned k = 0; k < kNumRegs; k++)
272 vec_store(&accTile[k], acc[k]);
275 std::memcpy(accumulator.accumulation[perspective][i], biases_,
276 kHalfDimensions * sizeof(BiasType));
278 for (const auto index : active_indices[perspective]) {
279 const IndexType offset = kHalfDimensions * index;
281 for (IndexType j = 0; j < kHalfDimensions; ++j)
282 accumulator.accumulation[perspective][i][j] += weights_[offset + j];
291 accumulator.computed_accumulation = true;
294 // Calculate cumulative value using difference calculation
295 void UpdateAccumulator(const Position& pos) const {
297 Accumulator* prev_accumulator;
298 assert(pos.state()->previous);
299 if (pos.state()->previous->accumulator.computed_accumulation) {
300 prev_accumulator = &pos.state()->previous->accumulator;
303 assert(pos.state()->previous->previous);
304 assert(pos.state()->previous->previous->accumulator.computed_accumulation);
305 prev_accumulator = &pos.state()->previous->previous->accumulator;
308 auto& accumulator = pos.state()->accumulator;
310 Features::IndexList removed_indices[2], added_indices[2];
311 bool reset[2] = { false, false };
312 RawFeatures::AppendChangedIndices(pos, kRefreshTriggers[i],
313 removed_indices, added_indices, reset);
316 for (IndexType j = 0; j < kHalfDimensions / kTileHeight; ++j) {
317 for (Color perspective : { WHITE, BLACK }) {
318 auto accTile = reinterpret_cast<vec_t*>(
319 &accumulator.accumulation[perspective][i][j * kTileHeight]);
322 if (reset[perspective]) {
323 auto biasesTile = reinterpret_cast<const vec_t*>(
324 &biases_[j * kTileHeight]);
325 for (unsigned k = 0; k < kNumRegs; ++k)
326 acc[k] = biasesTile[k];
328 auto prevAccTile = reinterpret_cast<const vec_t*>(
329 &prev_accumulator->accumulation[perspective][i][j * kTileHeight]);
330 for (IndexType k = 0; k < kNumRegs; ++k)
331 acc[k] = vec_load(&prevAccTile[k]);
333 // Difference calculation for the deactivated features
334 for (const auto index : removed_indices[perspective]) {
335 const IndexType offset = kHalfDimensions * index + j * kTileHeight;
336 auto column = reinterpret_cast<const vec_t*>(&weights_[offset]);
338 for (IndexType k = 0; k < kNumRegs; ++k)
339 acc[k] = vec_sub_16(acc[k], column[k]);
342 { // Difference calculation for the activated features
343 for (const auto index : added_indices[perspective]) {
344 const IndexType offset = kHalfDimensions * index + j * kTileHeight;
345 auto column = reinterpret_cast<const vec_t*>(&weights_[offset]);
347 for (IndexType k = 0; k < kNumRegs; ++k)
348 acc[k] = vec_add_16(acc[k], column[k]);
352 for (IndexType k = 0; k < kNumRegs; ++k)
353 vec_store(&accTile[k], acc[k]);
361 for (Color perspective : { WHITE, BLACK }) {
363 if (reset[perspective]) {
364 std::memcpy(accumulator.accumulation[perspective][i], biases_,
365 kHalfDimensions * sizeof(BiasType));
367 std::memcpy(accumulator.accumulation[perspective][i],
368 prev_accumulator->accumulation[perspective][i],
369 kHalfDimensions * sizeof(BiasType));
370 // Difference calculation for the deactivated features
371 for (const auto index : removed_indices[perspective]) {
372 const IndexType offset = kHalfDimensions * index;
374 for (IndexType j = 0; j < kHalfDimensions; ++j)
375 accumulator.accumulation[perspective][i][j] -= weights_[offset + j];
378 { // Difference calculation for the activated features
379 for (const auto index : added_indices[perspective]) {
380 const IndexType offset = kHalfDimensions * index;
382 for (IndexType j = 0; j < kHalfDimensions; ++j)
383 accumulator.accumulation[perspective][i][j] += weights_[offset + j];
389 accumulator.computed_accumulation = true;
392 using BiasType = std::int16_t;
393 using WeightType = std::int16_t;
395 alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
396 alignas(kCacheLineSize)
397 WeightType weights_[kHalfDimensions * kInputDimensions];
400 } // namespace Eval::NNUE
402 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED