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_SSSE3)
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_NEON)
101 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
102 const int8x8_t kZero = {0};
105 const Color perspectives[2] = {pos.side_to_move(), ~pos.side_to_move()};
106 for (IndexType p = 0; p < 2; ++p) {
107 const IndexType offset = kHalfDimensions * p;
109 #if defined(USE_AVX2)
110 auto out = reinterpret_cast<__m256i*>(&output[offset]);
111 for (IndexType j = 0; j < kNumChunks; ++j) {
114 #if defined(__MINGW32__) || defined(__MINGW64__)
115 // HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary
116 // compiled with g++ in MSYS2 crashes here because the output memory is not aligned
117 // even though alignas is specified.
123 (&reinterpret_cast<const __m256i*>(
124 accumulation[perspectives[p]][0])[j * 2 + 0]);
127 #if defined(__MINGW32__) || defined(__MINGW64__)
133 (&reinterpret_cast<const __m256i*>(
134 accumulation[perspectives[p]][0])[j * 2 + 1]);
136 #if defined(__MINGW32__) || defined(__MINGW64__)
142 (&out[j], _mm256_permute4x64_epi64(_mm256_max_epi8(
143 _mm256_packs_epi16(sum0, sum1), kZero), kControl));
146 #elif defined(USE_SSSE3)
147 auto out = reinterpret_cast<__m128i*>(&output[offset]);
148 for (IndexType j = 0; j < kNumChunks; ++j) {
149 __m128i sum0 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
150 accumulation[perspectives[p]][0])[j * 2 + 0]);
151 __m128i sum1 = _mm_load_si128(&reinterpret_cast<const __m128i*>(
152 accumulation[perspectives[p]][0])[j * 2 + 1]);
153 const __m128i packedbytes = _mm_packs_epi16(sum0, sum1);
155 _mm_store_si128(&out[j],
158 _mm_max_epi8(packedbytes, kZero)
160 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
166 #elif defined(USE_NEON)
167 const auto out = reinterpret_cast<int8x8_t*>(&output[offset]);
168 for (IndexType j = 0; j < kNumChunks; ++j) {
169 int16x8_t sum = reinterpret_cast<const int16x8_t*>(
170 accumulation[perspectives[p]][0])[j];
171 out[j] = vmax_s8(vqmovn_s16(sum), kZero);
175 for (IndexType j = 0; j < kHalfDimensions; ++j) {
176 BiasType sum = accumulation[static_cast<int>(perspectives[p])][0][j];
177 output[offset + j] = static_cast<OutputType>(
178 std::max<int>(0, std::min<int>(127, sum)));
186 // Calculate cumulative value without using difference calculation
187 void RefreshAccumulator(const Position& pos) const {
188 auto& accumulator = pos.state()->accumulator;
190 Features::IndexList active_indices[2];
191 RawFeatures::AppendActiveIndices(pos, kRefreshTriggers[i],
193 for (Color perspective : { WHITE, BLACK }) {
194 std::memcpy(accumulator.accumulation[perspective][i], biases_,
195 kHalfDimensions * sizeof(BiasType));
196 for (const auto index : active_indices[perspective]) {
197 const IndexType offset = kHalfDimensions * index;
199 #if defined(USE_AVX2)
200 auto accumulation = reinterpret_cast<__m256i*>(
201 &accumulator.accumulation[perspective][i][0]);
202 auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
203 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
204 for (IndexType j = 0; j < kNumChunks; ++j) {
205 #if defined(__MINGW32__) || defined(__MINGW64__)
206 _mm256_storeu_si256(&accumulation[j], _mm256_add_epi16(_mm256_loadu_si256(&accumulation[j]), column[j]));
208 accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
212 #elif defined(USE_SSE2)
213 auto accumulation = reinterpret_cast<__m128i*>(
214 &accumulator.accumulation[perspective][i][0]);
215 auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
216 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
217 for (IndexType j = 0; j < kNumChunks; ++j) {
218 accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
221 #elif defined(USE_NEON)
222 auto accumulation = reinterpret_cast<int16x8_t*>(
223 &accumulator.accumulation[perspective][i][0]);
224 auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
225 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
226 for (IndexType j = 0; j < kNumChunks; ++j) {
227 accumulation[j] = vaddq_s16(accumulation[j], column[j]);
231 for (IndexType j = 0; j < kHalfDimensions; ++j) {
232 accumulator.accumulation[perspective][i][j] += weights_[offset + j];
239 accumulator.computed_accumulation = true;
240 accumulator.computed_score = false;
243 // Calculate cumulative value using difference calculation
244 void UpdateAccumulator(const Position& pos) const {
245 const auto prev_accumulator = pos.state()->previous->accumulator;
246 auto& accumulator = pos.state()->accumulator;
248 Features::IndexList removed_indices[2], added_indices[2];
250 RawFeatures::AppendChangedIndices(pos, kRefreshTriggers[i],
251 removed_indices, added_indices, reset);
252 for (Color perspective : { WHITE, BLACK }) {
254 #if defined(USE_AVX2)
255 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
256 auto accumulation = reinterpret_cast<__m256i*>(
257 &accumulator.accumulation[perspective][i][0]);
259 #elif defined(USE_SSE2)
260 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
261 auto accumulation = reinterpret_cast<__m128i*>(
262 &accumulator.accumulation[perspective][i][0]);
264 #elif defined(USE_NEON)
265 constexpr IndexType kNumChunks = kHalfDimensions / (kSimdWidth / 2);
266 auto accumulation = reinterpret_cast<int16x8_t*>(
267 &accumulator.accumulation[perspective][i][0]);
270 if (reset[perspective]) {
271 std::memcpy(accumulator.accumulation[perspective][i], biases_,
272 kHalfDimensions * sizeof(BiasType));
274 std::memcpy(accumulator.accumulation[perspective][i],
275 prev_accumulator.accumulation[perspective][i],
276 kHalfDimensions * sizeof(BiasType));
277 // Difference calculation for the deactivated features
278 for (const auto index : removed_indices[perspective]) {
279 const IndexType offset = kHalfDimensions * index;
281 #if defined(USE_AVX2)
282 auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
283 for (IndexType j = 0; j < kNumChunks; ++j) {
284 accumulation[j] = _mm256_sub_epi16(accumulation[j], column[j]);
287 #elif defined(USE_SSE2)
288 auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
289 for (IndexType j = 0; j < kNumChunks; ++j) {
290 accumulation[j] = _mm_sub_epi16(accumulation[j], column[j]);
293 #elif defined(USE_NEON)
294 auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
295 for (IndexType j = 0; j < kNumChunks; ++j) {
296 accumulation[j] = vsubq_s16(accumulation[j], column[j]);
300 for (IndexType j = 0; j < kHalfDimensions; ++j) {
301 accumulator.accumulation[perspective][i][j] -=
302 weights_[offset + j];
308 { // Difference calculation for the activated features
309 for (const auto index : added_indices[perspective]) {
310 const IndexType offset = kHalfDimensions * index;
312 #if defined(USE_AVX2)
313 auto column = reinterpret_cast<const __m256i*>(&weights_[offset]);
314 for (IndexType j = 0; j < kNumChunks; ++j) {
315 accumulation[j] = _mm256_add_epi16(accumulation[j], column[j]);
318 #elif defined(USE_SSE2)
319 auto column = reinterpret_cast<const __m128i*>(&weights_[offset]);
320 for (IndexType j = 0; j < kNumChunks; ++j) {
321 accumulation[j] = _mm_add_epi16(accumulation[j], column[j]);
324 #elif defined(USE_NEON)
325 auto column = reinterpret_cast<const int16x8_t*>(&weights_[offset]);
326 for (IndexType j = 0; j < kNumChunks; ++j) {
327 accumulation[j] = vaddq_s16(accumulation[j], column[j]);
331 for (IndexType j = 0; j < kHalfDimensions; ++j) {
332 accumulator.accumulation[perspective][i][j] +=
333 weights_[offset + j];
341 accumulator.computed_accumulation = true;
342 accumulator.computed_score = false;
345 using BiasType = std::int16_t;
346 using WeightType = std::int16_t;
348 alignas(kCacheLineSize) BiasType biases_[kHalfDimensions];
349 alignas(kCacheLineSize)
350 WeightType weights_[kHalfDimensions * kInputDimensions];
353 } // namespace Eval::NNUE
355 #endif // #ifndef NNUE_FEATURE_TRANSFORMER_H_INCLUDED