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 // Input feature converter
33 class FeatureTransformer {
36 // Number of output dimensions for one side
37 static constexpr IndexType kHalfDimensions = kTransformedFeatureDimensions;
41 using OutputType = TransformedFeatureType;
43 // Number of input/output dimensions
44 static constexpr IndexType kInputDimensions = RawFeatures::kDimensions;
45 static constexpr IndexType kOutputDimensions = kHalfDimensions * 2;
47 // Size of forward propagation buffer
48 static constexpr std::size_t kBufferSize =
49 kOutputDimensions * sizeof(OutputType);
51 // Hash value embedded in the evaluation file
52 static constexpr std::uint32_t GetHashValue() {
53 return RawFeatures::kHashValue ^ kOutputDimensions;
56 // Read network parameters
57 bool ReadParameters(std::istream& stream) {
58 stream.read(reinterpret_cast<char*>(biases_),
59 kHalfDimensions * sizeof(BiasType));
60 stream.read(reinterpret_cast<char*>(weights_),
61 kHalfDimensions * kInputDimensions * sizeof(WeightType));
62 return !stream.fail();
65 // Proceed with the difference calculation if possible
66 bool UpdateAccumulatorIfPossible(const Position& pos) const {
67 const auto now = pos.state();
68 if (now->accumulator.computed_accumulation) {
71 const auto prev = now->previous;
72 if (prev && prev->accumulator.computed_accumulation) {
73 UpdateAccumulator(pos);
79 // Convert input features
80 void Transform(const Position& pos, OutputType* output, bool refresh) const {
81 if (refresh || !UpdateAccumulatorIfPossible(pos)) {
82 RefreshAccumulator(pos);
84 const auto& accumulation = pos.state()->accumulator.accumulation;
87 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
88 constexpr int kControl = 0b11011000;
89 const __m256i kZero = _mm256_setzero_si256();
91 #elif defined(USE_SSE2)
92 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
95 const __m128i kZero = _mm_setzero_si128();
97 const __m128i k0x80s = _mm_set1_epi8(-128);
100 #elif defined(USE_MMX)
101 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
102 const __m64 k0x80s = _mm_set1_pi8(-128);
104 #elif defined(USE_NEON)
105 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
106 const int8x8_t kZero = {0};
109 const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
110 for (IndexType p = 0; p < 2; ++p) {
111 const IndexType offset = kHalfDimensions * p;
113 #if defined(USE_AVX2)
114 auto out = reinterpret_cast<__m256i*>(&output[offset]);
115 for (IndexType j = 0; j < kNumChunks; ++j) {
116 __m256i sum0 = _mm256_loadA_si256(
117 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 0]);
118 __m256i sum1 = _mm256_loadA_si256(
119 &reinterpret_cast<const __m256i*>(accumulation[perspectives[p]][0])[j * 2 + 1]);
120 _mm256_storeA_si256(&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
121 _mm256_packs_epi16(sum0, sum1), kZero), kControl));
124 #elif defined(USE_SSE2)
125 auto out = reinterpret_cast<__m128i*>(&output[offset]);
126 for (IndexType j = 0; j < kNumChunks; ++j) {
127 __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
128 accumulation[perspectives[p]][0])[j * 2 + 0]);
129 __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
130 accumulation[perspectives[p]][0])[j * 2 + 1]);
131 const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
133 _mm_store_si128(&out[j],
136 _mm_max_epi8(packedbytes, kZero)
138 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
144 #elif defined(USE_MMX)
145 auto out = reinterpret_cast<__m64*>(&output[offset]);
146 for (IndexType j = 0; j < kNumChunks; ++j) {
147 __m64 sum0 = *(&reinterpret_cast<const __m64*>(
148 accumulation[perspectives[p]][0])[j * 2 + 0]);
149 __m64 sum1 = *(&reinterpret_cast<const __m64*>(
150 accumulation[perspectives[p]][0])[j * 2 + 1]);
151 const __m64 packedbytes = _mm_packs_pi16(sum0, sum1);
152 out[j] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
155 #elif defined(USE_NEON)
156 const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
157 for (IndexType j = 0; j < kNumChunks; ++j) {
158 int16x8_t sum = reinterpret_cast<const int16x8_t*>(
159 accumulation[perspectives[p]][0])[j];
160 out[j] = vmax_s8(vqmovn_s16(sum), kZero);
164 for (IndexType j = 0; j < kHalfDimensions; ++j) {
165 BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
166 output[offset + j] = static_cast<OutputType>(
167 std::max<int>(0, std::min<int>(127, sum)));
178 // Calculate cumulative value without using difference calculation
179 void RefreshAccumulator(const Position& pos) const {
180 auto& accumulator = pos.state()->accumulator;
182 Features::IndexList active_indices[2];
183 RawFeatures::AppendActiveIndices(pos, kRefreshTriggers[i],
185 for (Color perspective : { WHITE, BLACK }) {
186 std::memcpy(accumulator.accumulation[perspective][i], biases_,
187 kHalfDimensions * sizeof(BiasType));
188 for (const auto index : active_indices[perspective]) {
189 const IndexType offset = kHalfDimensions * index;
190 #if defined(USE_AVX512)
191 auto accumulation = reinterpret_cast<__m512i*>(
192 &accumulator.accumulation[perspective][i][0]);
193 auto column = reinterpret_cast<const __m512i*>(&weights_[offset]);
194 constexpr IndexType kNumChunks = kHalfDimensions / kSimdWidth;
195 for (IndexType j = 0; j < kNumChunks; ++j)
196 _mm512_storeA_si512(&accumulation[j], _mm512_add_epi16(_mm512_loadA_si512(&accumulation[j]), column[j]));
198 #elif defined(USE_AVX2)
199 auto accumulation = reinterpret_cast<__m256i*>(
200 &accumulator.accumulation[perspective][i][0]);
201 auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
202 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
203 for (IndexType j = 0; j < kNumChunks; ++j)
204 _mm256_storeA_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadA_si256(&accumulation[j]), column[j]));
206 #elif defined(USE_SSE2)
207 auto accumulation = reinterpret_cast<__m128i*>(
208 &accumulator.accumulation[perspective][i][0]);
209 auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
210 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
211 for (IndexType j = 0; j < kNumChunks; ++j)
212 accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
214 #elif defined(USE_MMX)
215 auto accumulation = reinterpret_cast<__m64*>(
216 &accumulator.accumulation[perspective][i][0]);
217 auto column = reinterpret_cast<const __m64*>(&weights_[offset]);
218 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
219 for (IndexType j = 0; j < kNumChunks; ++j) {
220 accumulation[j] = _mm_add_pi16(accumulation[j], column[j]);
223 #elif defined(USE_NEON)
224 auto accumulation = reinterpret_cast<int16x8_t*>(
225 &accumulator.accumulation[perspective][i][0]);
226 auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
227 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
228 for (IndexType j = 0; j < kNumChunks; ++j)
229 accumulation[j] = vaddq_s16(accumulation[j], column[j]);
232 for (IndexType j = 0; j < kHalfDimensions; ++j)
233 accumulator.accumulation[perspective][i][j] += weights_[offset + j];
242 accumulator.computed_accumulation = true;
243 accumulator.computed_score = false;
246 // Calculate cumulative value using difference calculation
247 void UpdateAccumulator(const Position& pos) const {
248 const auto prev_accumulator = pos.state()->previous->accumulator;
249 auto& accumulator = pos.state()->accumulator;
251 Features::IndexList removed_indices[2], added_indices[2];
253 RawFeatures::AppendChangedIndices(pos, kRefreshTriggers[i],
254 removed_indices, added_indices, reset);
255 for (Color perspective : { WHITE, BLACK }) {
257 #if defined(USE_AVX2)
258 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
259 auto accumulation = reinterpret_cast<__m256i*>(
260 &accumulator.accumulation[perspective][i][0]);
262 #elif defined(USE_SSE2)
263 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
264 auto accumulation = reinterpret_cast<__m128i*>(
265 &accumulator.accumulation[perspective][i][0]);
267 #elif defined(USE_MMX)
268 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
269 auto accumulation = reinterpret_cast<__m64*>(
270 &accumulator.accumulation[perspective][i][0]);
272 #elif defined(USE_NEON)
273 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
274 auto accumulation = reinterpret_cast<int16x8_t*>(
275 &accumulator.accumulation[perspective][i][0]);
278 if (reset[perspective]) {
279 std::memcpy(accumulator.accumulation[perspective][i], biases_,
280 kHalfDimensions * sizeof(BiasType));
282 std::memcpy(accumulator.accumulation[perspective][i],
283 prev_accumulator.accumulation[perspective][i],
284 kHalfDimensions * sizeof(BiasType));
285 // Difference calculation for the deactivated features
286 for (const auto index : removed_indices[perspective]) {
287 const IndexType offset = kHalfDimensions * index;
289 #if defined(USE_AVX2)
290 auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
291 for (IndexType j = 0; j < kNumChunks; ++j) {
292 accumulation[j] = _mm256_sub_epi16(accumulation[j], column[j]);
295 #elif defined(USE_SSE2)
296 auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
297 for (IndexType j = 0; j < kNumChunks; ++j) {
298 accumulation[j] = _mm_sub_epi16(accumulation[j], column[j]);
301 #elif defined(USE_MMX)
302 auto column = reinterpret_cast<const __m64*>(&weights_[offset]);
303 for (IndexType j = 0; j < kNumChunks; ++j) {
304 accumulation[j] = _mm_sub_pi16(accumulation[j], column[j]);
307 #elif defined(USE_NEON)
308 auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
309 for (IndexType j = 0; j < kNumChunks; ++j) {
310 accumulation[j] = vsubq_s16(accumulation[j], column[j]);
314 for (IndexType j = 0; j < kHalfDimensions; ++j) {
315 accumulator.accumulation[perspective][i][j] -=
316 weights_[offset + j];
322 { // Difference calculation for the activated features
323 for (const auto index : added_indices[perspective]) {
324 const IndexType offset = kHalfDimensions * index;
326 #if defined(USE_AVX2)
327 auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
328 for (IndexType j = 0; j < kNumChunks; ++j) {
329 accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
332 #elif defined(USE_SSE2)
333 auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
334 for (IndexType j = 0; j < kNumChunks; ++j) {
335 accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
338 #elif defined(USE_MMX)
339 auto column = reinterpret_cast<const __m64*>(&weights_[offset]);
340 for (IndexType j = 0; j < kNumChunks; ++j) {
341 accumulation[j] = _mm_add_pi16(accumulation[j], column[j]);
344 #elif defined(USE_NEON)
345 auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
346 for (IndexType j = 0; j < kNumChunks; ++j) {
347 accumulation[j] = vaddq_s16(accumulation[j], column[j]);
351 for (IndexType j = 0; j < kHalfDimensions; ++j) {
352 accumulator.accumulation[perspective][i][j] +=
353 weights_[offset + j];
364 accumulator.computed_accumulation = true;
365 accumulator.computed_score = false;
368 using BiasType = std::int16_t;
369 using WeightType = std::int16_t;
371 alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
372 alignas(kCacheLineSize)
373 WeightType weights_[kHalfDimensions * kInputDimensions];
376 } // namespace Eval::NNUE
378 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED