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 // Definition of layer AffineTransform of NNUE evaluation function
21 #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
22 #define NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
25 #include "../nnue_common.h"
27 namespace Stockfish::Eval::NNUE::Layers {
29 // Affine transformation layer
30 template <typename PreviousLayer, IndexType OutDims>
31 class AffineTransform {
34 using InputType = typename PreviousLayer::OutputType;
35 using OutputType = std::int32_t;
36 static_assert(std::is_same<InputType, std::uint8_t>::value, "");
38 // Number of input/output dimensions
39 static constexpr IndexType InputDimensions =
40 PreviousLayer::OutputDimensions;
41 static constexpr IndexType OutputDimensions = OutDims;
42 static constexpr IndexType PaddedInputDimensions =
43 ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
44 #if defined (USE_AVX512)
45 static constexpr const IndexType OutputSimdWidth = SimdWidth / 2;
46 #elif defined (USE_SSSE3)
47 static constexpr const IndexType OutputSimdWidth = SimdWidth / 4;
50 // Size of forward propagation buffer used in this layer
51 static constexpr std::size_t SelfBufferSize =
52 ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);
54 // Size of the forward propagation buffer used from the input layer to this layer
55 static constexpr std::size_t BufferSize =
56 PreviousLayer::BufferSize + SelfBufferSize;
58 // Hash value embedded in the evaluation file
59 static constexpr std::uint32_t get_hash_value() {
60 std::uint32_t hashValue = 0xCC03DAE4u;
61 hashValue += OutputDimensions;
62 hashValue ^= PreviousLayer::get_hash_value() >> 1;
63 hashValue ^= PreviousLayer::get_hash_value() << 31;
67 // Read network parameters
68 bool read_parameters(std::istream& stream) {
69 if (!previousLayer.read_parameters(stream)) return false;
70 for (std::size_t i = 0; i < OutputDimensions; ++i)
71 biases[i] = read_little_endian<BiasType>(stream);
72 for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
73 #if !defined (USE_SSSE3)
74 weights[i] = read_little_endian<WeightType>(stream);
77 (i / 4) % (PaddedInputDimensions / 4) * OutputDimensions * 4 +
78 i / PaddedInputDimensions * 4 +
80 ] = read_little_endian<WeightType>(stream);
83 return !stream.fail();
86 // Write network parameters
87 bool write_parameters(std::ostream& stream) const {
88 if (!previousLayer.write_parameters(stream)) return false;
89 for (std::size_t i = 0; i < OutputDimensions; ++i)
90 write_little_endian<BiasType>(stream, biases[i]);
91 #if !defined (USE_SSSE3)
92 for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
93 write_little_endian<WeightType>(stream, weights[i]);
95 std::unique_ptr<WeightType[]> unscrambledWeights = std::make_unique<WeightType[]>(OutputDimensions * PaddedInputDimensions);
96 for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i) {
97 unscrambledWeights[i] =
99 (i / 4) % (PaddedInputDimensions / 4) * OutputDimensions * 4 +
100 i / PaddedInputDimensions * 4 +
105 for (std::size_t i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
106 write_little_endian<WeightType>(stream, unscrambledWeights[i]);
109 return !stream.fail();
112 // Forward propagation
113 const OutputType* propagate(
114 const TransformedFeatureType* transformedFeatures, char* buffer) const {
115 const auto input = previousLayer.propagate(
116 transformedFeatures, buffer + SelfBufferSize);
118 #if defined (USE_AVX512)
120 [[maybe_unused]] const __m512i Ones512 = _mm512_set1_epi16(1);
122 [[maybe_unused]] auto m512_hadd = [](__m512i sum, int bias) -> int {
123 return _mm512_reduce_add_epi32(sum) + bias;
126 [[maybe_unused]] auto m512_add_dpbusd_epi32 = [=](__m512i& acc, __m512i a, __m512i b) {
127 #if defined (USE_VNNI)
128 acc = _mm512_dpbusd_epi32(acc, a, b);
130 __m512i product0 = _mm512_maddubs_epi16(a, b);
131 product0 = _mm512_madd_epi16(product0, Ones512);
132 acc = _mm512_add_epi32(acc, product0);
136 [[maybe_unused]] auto m512_add_dpbusd_epi32x4 = [=](__m512i& acc, __m512i a0, __m512i b0, __m512i a1, __m512i b1,
137 __m512i a2, __m512i b2, __m512i a3, __m512i b3) {
138 #if defined (USE_VNNI)
139 acc = _mm512_dpbusd_epi32(acc, a0, b0);
140 acc = _mm512_dpbusd_epi32(acc, a1, b1);
141 acc = _mm512_dpbusd_epi32(acc, a2, b2);
142 acc = _mm512_dpbusd_epi32(acc, a3, b3);
144 __m512i product0 = _mm512_maddubs_epi16(a0, b0);
145 __m512i product1 = _mm512_maddubs_epi16(a1, b1);
146 __m512i product2 = _mm512_maddubs_epi16(a2, b2);
147 __m512i product3 = _mm512_maddubs_epi16(a3, b3);
148 product0 = _mm512_adds_epi16(product0, product1);
149 product0 = _mm512_madd_epi16(product0, Ones512);
150 product2 = _mm512_adds_epi16(product2, product3);
151 product2 = _mm512_madd_epi16(product2, Ones512);
152 acc = _mm512_add_epi32(acc, _mm512_add_epi32(product0, product2));
157 #if defined (USE_AVX2)
159 [[maybe_unused]] const __m256i Ones256 = _mm256_set1_epi16(1);
161 [[maybe_unused]] auto m256_hadd = [](__m256i sum, int bias) -> int {
162 __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
163 sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));
164 sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB));
165 return _mm_cvtsi128_si32(sum128) + bias;
168 [[maybe_unused]] auto m256_add_dpbusd_epi32 = [=](__m256i& acc, __m256i a, __m256i b) {
169 #if defined (USE_VNNI)
170 acc = _mm256_dpbusd_epi32(acc, a, b);
172 __m256i product0 = _mm256_maddubs_epi16(a, b);
173 product0 = _mm256_madd_epi16(product0, Ones256);
174 acc = _mm256_add_epi32(acc, product0);
178 [[maybe_unused]] auto m256_add_dpbusd_epi32x4 = [=](__m256i& acc, __m256i a0, __m256i b0, __m256i a1, __m256i b1,
179 __m256i a2, __m256i b2, __m256i a3, __m256i b3) {
180 #if defined (USE_VNNI)
181 acc = _mm256_dpbusd_epi32(acc, a0, b0);
182 acc = _mm256_dpbusd_epi32(acc, a1, b1);
183 acc = _mm256_dpbusd_epi32(acc, a2, b2);
184 acc = _mm256_dpbusd_epi32(acc, a3, b3);
186 __m256i product0 = _mm256_maddubs_epi16(a0, b0);
187 __m256i product1 = _mm256_maddubs_epi16(a1, b1);
188 __m256i product2 = _mm256_maddubs_epi16(a2, b2);
189 __m256i product3 = _mm256_maddubs_epi16(a3, b3);
190 product0 = _mm256_adds_epi16(product0, product1);
191 product0 = _mm256_madd_epi16(product0, Ones256);
192 product2 = _mm256_adds_epi16(product2, product3);
193 product2 = _mm256_madd_epi16(product2, Ones256);
194 acc = _mm256_add_epi32(acc, _mm256_add_epi32(product0, product2));
199 #if defined (USE_SSSE3)
201 [[maybe_unused]] const __m128i Ones128 = _mm_set1_epi16(1);
203 [[maybe_unused]] auto m128_hadd = [](__m128i sum, int bias) -> int {
204 sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC
205 sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
206 return _mm_cvtsi128_si32(sum) + bias;
209 [[maybe_unused]] auto m128_add_dpbusd_epi32 = [=](__m128i& acc, __m128i a, __m128i b) {
210 __m128i product0 = _mm_maddubs_epi16(a, b);
211 product0 = _mm_madd_epi16(product0, Ones128);
212 acc = _mm_add_epi32(acc, product0);
215 [[maybe_unused]] auto m128_add_dpbusd_epi32x4 = [=](__m128i& acc, __m128i a0, __m128i b0, __m128i a1, __m128i b1,
216 __m128i a2, __m128i b2, __m128i a3, __m128i b3) {
217 __m128i product0 = _mm_maddubs_epi16(a0, b0);
218 __m128i product1 = _mm_maddubs_epi16(a1, b1);
219 __m128i product2 = _mm_maddubs_epi16(a2, b2);
220 __m128i product3 = _mm_maddubs_epi16(a3, b3);
221 product0 = _mm_adds_epi16(product0, product1);
222 product0 = _mm_madd_epi16(product0, Ones128);
223 product2 = _mm_adds_epi16(product2, product3);
224 product2 = _mm_madd_epi16(product2, Ones128);
225 acc = _mm_add_epi32(acc, _mm_add_epi32(product0, product2));
230 #if defined (USE_AVX512)
231 using vec_t = __m512i;
232 #define vec_setzero _mm512_setzero_si512
233 #define vec_set_32 _mm512_set1_epi32
234 auto& vec_add_dpbusd_32 = m512_add_dpbusd_epi32;
235 auto& vec_add_dpbusd_32x4 = m512_add_dpbusd_epi32x4;
236 auto& vec_hadd = m512_hadd;
237 #elif defined (USE_AVX2)
238 using vec_t = __m256i;
239 #define vec_setzero _mm256_setzero_si256
240 #define vec_set_32 _mm256_set1_epi32
241 auto& vec_add_dpbusd_32 = m256_add_dpbusd_epi32;
242 auto& vec_add_dpbusd_32x4 = m256_add_dpbusd_epi32x4;
243 auto& vec_hadd = m256_hadd;
244 #elif defined (USE_SSSE3)
245 using vec_t = __m128i;
246 #define vec_setzero _mm_setzero_si128
247 #define vec_set_32 _mm_set1_epi32
248 auto& vec_add_dpbusd_32 = m128_add_dpbusd_epi32;
249 auto& vec_add_dpbusd_32x4 = m128_add_dpbusd_epi32x4;
250 auto& vec_hadd = m128_hadd;
253 #if defined (USE_SSSE3)
254 const auto output = reinterpret_cast<OutputType*>(buffer);
255 const auto inputVector = reinterpret_cast<const vec_t*>(input);
257 static_assert(OutputDimensions % OutputSimdWidth == 0 || OutputDimensions == 1);
259 // OutputDimensions is either 1 or a multiple of SimdWidth
260 // because then it is also an input dimension.
261 if constexpr (OutputDimensions % OutputSimdWidth == 0)
263 static_assert(InputDimensions % 16 == 0);
265 constexpr IndexType NumChunks = InputDimensions / 4;
266 constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth;
268 const auto input32 = reinterpret_cast<const std::int32_t*>(input);
269 const vec_t* biasvec = reinterpret_cast<const vec_t*>(biases);
271 for (IndexType k = 0; k < NumRegs; ++k)
272 outs[k] = biasvec[k];
274 for (IndexType i = 0; i < NumChunks; i += 4)
276 const vec_t in0 = vec_set_32(input32[i + 0]);
277 const vec_t in1 = vec_set_32(input32[i + 1]);
278 const vec_t in2 = vec_set_32(input32[i + 2]);
279 const vec_t in3 = vec_set_32(input32[i + 3]);
280 const auto col0 = reinterpret_cast<const vec_t*>(&weights[(i + 0) * OutputDimensions * 4]);
281 const auto col1 = reinterpret_cast<const vec_t*>(&weights[(i + 1) * OutputDimensions * 4]);
282 const auto col2 = reinterpret_cast<const vec_t*>(&weights[(i + 2) * OutputDimensions * 4]);
283 const auto col3 = reinterpret_cast<const vec_t*>(&weights[(i + 3) * OutputDimensions * 4]);
284 for (IndexType k = 0; k < NumRegs; ++k)
285 vec_add_dpbusd_32x4(outs[k], in0, col0[k], in1, col1[k], in2, col2[k], in3, col3[k]);
288 vec_t* outptr = reinterpret_cast<vec_t*>(output);
289 for (IndexType k = 0; k < NumRegs; ++k)
292 else if constexpr (OutputDimensions == 1)
294 static_assert(InputDimensions % 4 == 0);
296 #if defined (USE_AVX512)
297 if constexpr (PaddedInputDimensions % (SimdWidth * 2) != 0)
299 constexpr IndexType NumChunks = PaddedInputDimensions / SimdWidth;
300 const auto inputVector256 = reinterpret_cast<const __m256i*>(input);
302 __m256i sum0 = _mm256_setzero_si256();
303 const auto row0 = reinterpret_cast<const __m256i*>(&weights[0]);
305 for (int j = 0; j < (int)NumChunks; ++j)
307 const __m256i in = inputVector256[j];
308 m256_add_dpbusd_epi32(sum0, in, row0[j]);
310 output[0] = m256_hadd(sum0, biases[0]);
315 #if defined (USE_AVX512)
316 constexpr IndexType NumChunks = PaddedInputDimensions / (SimdWidth * 2);
318 constexpr IndexType NumChunks = PaddedInputDimensions / SimdWidth;
320 vec_t sum0 = vec_setzero();
321 const auto row0 = reinterpret_cast<const vec_t*>(&weights[0]);
323 for (int j = 0; j < (int)NumChunks; ++j)
325 const vec_t in = inputVector[j];
326 vec_add_dpbusd_32(sum0, in, row0[j]);
328 output[0] = vec_hadd(sum0, biases[0]);
334 // Use old implementation for the other architectures.
336 auto output = reinterpret_cast<OutputType*>(buffer);
338 #if defined(USE_SSE2)
339 // At least a multiple of 16, with SSE2.
340 static_assert(InputDimensions % SimdWidth == 0);
341 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
342 const __m128i Zeros = _mm_setzero_si128();
343 const auto inputVector = reinterpret_cast<const __m128i*>(input);
345 #elif defined(USE_MMX)
346 static_assert(InputDimensions % SimdWidth == 0);
347 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
348 const __m64 Zeros = _mm_setzero_si64();
349 const auto inputVector = reinterpret_cast<const __m64*>(input);
351 #elif defined(USE_NEON)
352 static_assert(InputDimensions % SimdWidth == 0);
353 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
354 const auto inputVector = reinterpret_cast<const int8x8_t*>(input);
357 for (IndexType i = 0; i < OutputDimensions; ++i) {
358 const IndexType offset = i * PaddedInputDimensions;
360 #if defined(USE_SSE2)
361 __m128i sumLo = _mm_cvtsi32_si128(biases[i]);
362 __m128i sumHi = Zeros;
363 const auto row = reinterpret_cast<const __m128i*>(&weights[offset]);
364 for (IndexType j = 0; j < NumChunks; ++j) {
365 __m128i row_j = _mm_load_si128(&row[j]);
366 __m128i input_j = _mm_load_si128(&inputVector[j]);
367 __m128i extendedRowLo = _mm_srai_epi16(_mm_unpacklo_epi8(row_j, row_j), 8);
368 __m128i extendedRowHi = _mm_srai_epi16(_mm_unpackhi_epi8(row_j, row_j), 8);
369 __m128i extendedInputLo = _mm_unpacklo_epi8(input_j, Zeros);
370 __m128i extendedInputHi = _mm_unpackhi_epi8(input_j, Zeros);
371 __m128i productLo = _mm_madd_epi16(extendedRowLo, extendedInputLo);
372 __m128i productHi = _mm_madd_epi16(extendedRowHi, extendedInputHi);
373 sumLo = _mm_add_epi32(sumLo, productLo);
374 sumHi = _mm_add_epi32(sumHi, productHi);
376 __m128i sum = _mm_add_epi32(sumLo, sumHi);
377 __m128i sumHigh_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2));
378 sum = _mm_add_epi32(sum, sumHigh_64);
379 __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2));
380 sum = _mm_add_epi32(sum, sum_second_32);
381 output[i] = _mm_cvtsi128_si32(sum);
383 #elif defined(USE_MMX)
384 __m64 sumLo = _mm_cvtsi32_si64(biases[i]);
386 const auto row = reinterpret_cast<const __m64*>(&weights[offset]);
387 for (IndexType j = 0; j < NumChunks; ++j) {
388 __m64 row_j = row[j];
389 __m64 input_j = inputVector[j];
390 __m64 extendedRowLo = _mm_srai_pi16(_mm_unpacklo_pi8(row_j, row_j), 8);
391 __m64 extendedRowHi = _mm_srai_pi16(_mm_unpackhi_pi8(row_j, row_j), 8);
392 __m64 extendedInputLo = _mm_unpacklo_pi8(input_j, Zeros);
393 __m64 extendedInputHi = _mm_unpackhi_pi8(input_j, Zeros);
394 __m64 productLo = _mm_madd_pi16(extendedRowLo, extendedInputLo);
395 __m64 productHi = _mm_madd_pi16(extendedRowHi, extendedInputHi);
396 sumLo = _mm_add_pi32(sumLo, productLo);
397 sumHi = _mm_add_pi32(sumHi, productHi);
399 __m64 sum = _mm_add_pi32(sumLo, sumHi);
400 sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum));
401 output[i] = _mm_cvtsi64_si32(sum);
403 #elif defined(USE_NEON)
404 int32x4_t sum = {biases[i]};
405 const auto row = reinterpret_cast<const int8x8_t*>(&weights[offset]);
406 for (IndexType j = 0; j < NumChunks; ++j) {
407 int16x8_t product = vmull_s8(inputVector[j * 2], row[j * 2]);
408 product = vmlal_s8(product, inputVector[j * 2 + 1], row[j * 2 + 1]);
409 sum = vpadalq_s16(sum, product);
411 output[i] = sum[0] + sum[1] + sum[2] + sum[3];
414 OutputType sum = biases[i];
415 for (IndexType j = 0; j < InputDimensions; ++j) {
416 sum += weights[offset + j] * input[j];
432 using BiasType = OutputType;
433 using WeightType = std::int8_t;
435 PreviousLayer previousLayer;
437 alignas(CacheLineSize) BiasType biases[OutputDimensions];
438 alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
441 } // namespace Stockfish::Eval::NNUE::Layers
443 #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED