2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2023 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
27 #include "../nnue_common.h"
31 This file contains the definition for a fully connected layer (aka affine transform).
33 - expected use-case is for when PaddedInputDimensions == 32 and InputDimensions <= 32.
34 - that's why AVX512 is hard to implement
35 - expected use-case is small layers
36 - inputs are processed in chunks of 4, weights are respectively transposed
37 - accumulation happens directly to int32s
40 namespace Stockfish::Eval::NNUE::Layers {
42 // Fallback implementation for older/other architectures.
43 // Requires the input to be padded to at least 16 values.
44 #if !defined(USE_SSSE3)
45 template <IndexType InputDimensions, IndexType PaddedInputDimensions, IndexType OutputDimensions>
46 static void affine_transform_non_ssse3(std::int32_t* output, const std::int8_t* weights, const std::int32_t* biases, const std::uint8_t* input)
48 # if defined(USE_SSE2) || defined(USE_NEON_DOTPROD) || defined(USE_NEON)
49 # if defined(USE_SSE2)
50 // At least a multiple of 16, with SSE2.
51 constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
52 const __m128i Zeros = _mm_setzero_si128();
53 const auto inputVector = reinterpret_cast<const __m128i*>(input);
55 # elif defined(USE_NEON_DOTPROD)
56 constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
57 const auto inputVector = reinterpret_cast<const int8x16_t*>(input);
59 # elif defined(USE_NEON)
60 constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 16) / 16;
61 const auto inputVector = reinterpret_cast<const int8x8_t*>(input);
64 for (IndexType i = 0; i < OutputDimensions; ++i) {
65 const IndexType offset = i * PaddedInputDimensions;
67 # if defined(USE_SSE2)
68 __m128i sumLo = _mm_cvtsi32_si128(biases[i]);
69 __m128i sumHi = Zeros;
70 const auto row = reinterpret_cast<const __m128i*>(&weights[offset]);
71 for (IndexType j = 0; j < NumChunks; ++j) {
72 __m128i row_j = _mm_load_si128(&row[j]);
73 __m128i input_j = _mm_load_si128(&inputVector[j]);
74 __m128i extendedRowLo = _mm_srai_epi16(_mm_unpacklo_epi8(row_j, row_j), 8);
75 __m128i extendedRowHi = _mm_srai_epi16(_mm_unpackhi_epi8(row_j, row_j), 8);
76 __m128i extendedInputLo = _mm_unpacklo_epi8(input_j, Zeros);
77 __m128i extendedInputHi = _mm_unpackhi_epi8(input_j, Zeros);
78 __m128i productLo = _mm_madd_epi16(extendedRowLo, extendedInputLo);
79 __m128i productHi = _mm_madd_epi16(extendedRowHi, extendedInputHi);
80 sumLo = _mm_add_epi32(sumLo, productLo);
81 sumHi = _mm_add_epi32(sumHi, productHi);
83 __m128i sum = _mm_add_epi32(sumLo, sumHi);
84 __m128i sumHigh_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2));
85 sum = _mm_add_epi32(sum, sumHigh_64);
86 __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2));
87 sum = _mm_add_epi32(sum, sum_second_32);
88 output[i] = _mm_cvtsi128_si32(sum);
90 # elif defined(USE_NEON_DOTPROD)
91 int32x4_t sum = {biases[i]};
92 const auto row = reinterpret_cast<const int8x16_t*>(&weights[offset]);
93 for (IndexType j = 0; j < NumChunks; ++j) {
94 sum = vdotq_s32(sum, inputVector[j], row[j]);
96 output[i] = vaddvq_s32(sum);
98 # elif defined(USE_NEON)
99 int32x4_t sum = {biases[i]};
100 const auto row = reinterpret_cast<const int8x8_t*>(&weights[offset]);
101 for (IndexType j = 0; j < NumChunks; ++j) {
102 int16x8_t product = vmull_s8(inputVector[j * 2], row[j * 2]);
103 product = vmlal_s8(product, inputVector[j * 2 + 1], row[j * 2 + 1]);
104 sum = vpadalq_s16(sum, product);
106 output[i] = sum[0] + sum[1] + sum[2] + sum[3];
111 std::memcpy(output, biases, sizeof(std::int32_t) * OutputDimensions);
113 // Traverse weights in transpose order to take advantage of input sparsity
114 for (IndexType i = 0; i < InputDimensions; ++i)
116 const std::int8_t* w = &weights[i];
117 const int in = input[i];
118 for (IndexType j = 0; j < OutputDimensions; ++j)
119 output[j] += w[j * PaddedInputDimensions] * in;
125 template <IndexType InDims, IndexType OutDims>
126 class AffineTransform {
129 using InputType = std::uint8_t;
130 using OutputType = std::int32_t;
132 // Number of input/output dimensions
133 static constexpr IndexType InputDimensions = InDims;
134 static constexpr IndexType OutputDimensions = OutDims;
136 static constexpr IndexType PaddedInputDimensions =
137 ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
138 static constexpr IndexType PaddedOutputDimensions =
139 ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
141 using OutputBuffer = OutputType[PaddedOutputDimensions];
143 // Hash value embedded in the evaluation file
144 static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
145 std::uint32_t hashValue = 0xCC03DAE4u;
146 hashValue += OutputDimensions;
147 hashValue ^= prevHash >> 1;
148 hashValue ^= prevHash << 31;
152 static constexpr IndexType get_weight_index_scrambled(IndexType i)
155 (i / 4) % (PaddedInputDimensions / 4) * OutputDimensions * 4 +
156 i / PaddedInputDimensions * 4 +
160 static constexpr IndexType get_weight_index(IndexType i)
162 #if defined (USE_SSSE3)
163 return get_weight_index_scrambled(i);
169 // Read network parameters
170 bool read_parameters(std::istream& stream) {
171 read_little_endian<BiasType>(stream, biases, OutputDimensions);
172 for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
173 weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
175 return !stream.fail();
178 // Write network parameters
179 bool write_parameters(std::ostream& stream) const {
180 write_little_endian<BiasType>(stream, biases, OutputDimensions);
182 for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
183 write_little_endian<WeightType>(stream, weights[get_weight_index(i)]);
185 return !stream.fail();
187 // Forward propagation
189 const InputType* input, OutputType* output) const {
191 #if defined (USE_SSSE3)
193 if constexpr (OutputDimensions > 1)
196 #if defined (USE_AVX512)
197 using vec_t = __m512i;
198 #define vec_setzero _mm512_setzero_si512
199 #define vec_set_32 _mm512_set1_epi32
200 #define vec_add_dpbusd_32 Simd::m512_add_dpbusd_epi32
201 #define vec_add_dpbusd_32x2 Simd::m512_add_dpbusd_epi32x2
202 #define vec_hadd Simd::m512_hadd
203 #elif defined (USE_AVX2)
204 using vec_t = __m256i;
205 #define vec_setzero _mm256_setzero_si256
206 #define vec_set_32 _mm256_set1_epi32
207 #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
208 #define vec_add_dpbusd_32x2 Simd::m256_add_dpbusd_epi32x2
209 #define vec_hadd Simd::m256_hadd
210 #elif defined (USE_SSSE3)
211 using vec_t = __m128i;
212 #define vec_setzero _mm_setzero_si128
213 #define vec_set_32 _mm_set1_epi32
214 #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
215 #define vec_add_dpbusd_32x2 Simd::m128_add_dpbusd_epi32x2
216 #define vec_hadd Simd::m128_hadd
219 static constexpr IndexType OutputSimdWidth = sizeof(vec_t) / sizeof(OutputType);
221 static_assert(OutputDimensions % OutputSimdWidth == 0);
223 constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / 4;
224 constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth;
226 const auto input32 = reinterpret_cast<const std::int32_t*>(input);
227 const vec_t* biasvec = reinterpret_cast<const vec_t*>(biases);
229 for (IndexType k = 0; k < NumRegs; ++k)
232 for (IndexType i = 0; i < NumChunks; i += 2)
234 const vec_t in0 = vec_set_32(input32[i + 0]);
235 const vec_t in1 = vec_set_32(input32[i + 1]);
236 const auto col0 = reinterpret_cast<const vec_t*>(&weights[(i + 0) * OutputDimensions * 4]);
237 const auto col1 = reinterpret_cast<const vec_t*>(&weights[(i + 1) * OutputDimensions * 4]);
238 for (IndexType k = 0; k < NumRegs; ++k)
239 vec_add_dpbusd_32x2(acc[k], in0, col0[k], in1, col1[k]);
242 vec_t* outptr = reinterpret_cast<vec_t*>(output);
243 for (IndexType k = 0; k < NumRegs; ++k)
248 # undef vec_add_dpbusd_32
249 # undef vec_add_dpbusd_32x2
253 else if constexpr (OutputDimensions == 1)
256 // We cannot use AVX512 for the last layer because there's only 32 inputs and the buffer is not padded to 64 elements.
257 #if defined (USE_AVX2)
258 using vec_t = __m256i;
259 #define vec_setzero _mm256_setzero_si256
260 #define vec_set_32 _mm256_set1_epi32
261 #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
262 #define vec_add_dpbusd_32x2 Simd::m256_add_dpbusd_epi32x2
263 #define vec_hadd Simd::m256_hadd
264 #elif defined (USE_SSSE3)
265 using vec_t = __m128i;
266 #define vec_setzero _mm_setzero_si128
267 #define vec_set_32 _mm_set1_epi32
268 #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
269 #define vec_add_dpbusd_32x2 Simd::m128_add_dpbusd_epi32x2
270 #define vec_hadd Simd::m128_hadd
273 const auto inputVector = reinterpret_cast<const vec_t*>(input);
275 static constexpr IndexType InputSimdWidth = sizeof(vec_t) / sizeof(InputType);
277 static_assert(PaddedInputDimensions % InputSimdWidth == 0);
279 constexpr IndexType NumChunks = PaddedInputDimensions / InputSimdWidth;
280 vec_t sum0 = vec_setzero();
281 const auto row0 = reinterpret_cast<const vec_t*>(&weights[0]);
283 for (int j = 0; j < int(NumChunks); ++j)
285 const vec_t in = inputVector[j];
286 vec_add_dpbusd_32(sum0, in, row0[j]);
288 output[0] = vec_hadd(sum0, biases[0]);
292 # undef vec_add_dpbusd_32
293 # undef vec_add_dpbusd_32x2
298 // Use old implementation for the other architectures.
299 affine_transform_non_ssse3<
301 PaddedInputDimensions,
302 OutputDimensions>(output, weights, biases, input);
307 using BiasType = OutputType;
308 using WeightType = std::int8_t;
310 alignas(CacheLineSize) BiasType biases[OutputDimensions];
311 alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
314 } // namespace Stockfish::Eval::NNUE::Layers
316 #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED