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 AffineTransformSparseInput of NNUE evaluation function
21 #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_SPARSE_INPUT_H_INCLUDED
22 #define NNUE_LAYERS_AFFINE_TRANSFORM_SPARSE_INPUT_H_INCLUDED
28 #include <type_traits>
29 #include "../nnue_common.h"
30 #include "affine_transform.h"
34 This file contains the definition for a fully connected layer (aka affine transform) with block sparse input.
37 namespace Stockfish::Eval::NNUE::Layers {
38 #if defined(__GNUC__) // GCC, Clang, ICC
40 static inline IndexType lsb_(std::uint32_t b) {
42 return IndexType(__builtin_ctzl(b));
45 #elif defined(_MSC_VER) // MSVC
47 static inline IndexType lsb_(std::uint32_t b) {
50 _BitScanForward(&idx, b);
51 return (IndexType) idx;
54 #else // Compiler is neither GCC nor MSVC compatible
56 #error "Compiler not supported."
61 #if defined(USE_SSSE3)
62 alignas(CacheLineSize) static inline const std::array<std::array<std::uint16_t, 8>, 256> lookup_indices = [](){
63 std::array<std::array<std::uint16_t, 8>, 256> v{};
64 for (int i = 0; i < 256; ++i)
70 const IndexType lsbIndex = lsb_(std::uint32_t(j));
78 alignas(CacheLineSize) static inline const std::array<unsigned, 256> lookup_count = [](){
79 std::array<unsigned, 256> v;
80 for (int i = 0; i < 256; ++i)
81 v[i] = unsigned(std::bitset<8>(i).count());
85 // Find indices of nonzero numbers in an int32_t array
86 template<const IndexType InputDimensions>
87 void find_nnz(const std::int32_t* input, std::uint16_t* out, IndexType& count_out) {
88 #if defined (USE_AVX512)
89 using vec_t = __m512i;
90 #define vec_nnz(a) _mm512_cmpgt_epi32_mask(a, _mm512_setzero_si512())
91 #elif defined (USE_AVX2)
92 using vec_t = __m256i;
93 #define vec_nnz(a) _mm256_movemask_ps(_mm256_castsi256_ps(_mm256_cmpgt_epi32(a, _mm256_setzero_si256())))
94 #elif defined (USE_SSSE3)
95 using vec_t = __m128i;
96 #define vec_nnz(a) _mm_movemask_ps(_mm_castsi128_ps(_mm_cmpgt_epi32(a, _mm_setzero_si128())))
98 constexpr IndexType InputSimdWidth = sizeof(vec_t) / sizeof(std::int32_t);
99 // Inputs are processed InputSimdWidth at a time and outputs are processed 8 at a time so we process in chunks of max(InputSimdWidth, 8)
100 constexpr IndexType ChunkSize = std::max<IndexType>(InputSimdWidth, 8);
101 constexpr IndexType NumChunks = InputDimensions / ChunkSize;
102 constexpr IndexType InputsPerChunk = ChunkSize / InputSimdWidth;
103 constexpr IndexType OutputsPerChunk = ChunkSize / 8;
105 const auto inputVector = reinterpret_cast<const vec_t*>(input);
107 __m128i base = _mm_set1_epi16(0);
108 __m128i increment = _mm_set1_epi16(8);
109 for (IndexType i = 0; i < NumChunks; ++i)
111 // bitmask of nonzero values in this chunk
113 for (IndexType j = 0; j < InputsPerChunk; ++j)
115 const vec_t inputChunk = inputVector[i * InputsPerChunk + j];
116 nnz |= (unsigned)vec_nnz(inputChunk) << (j * InputSimdWidth);
118 for (IndexType j = 0; j < OutputsPerChunk; ++j)
120 const auto lookup = (nnz >> (j * 8)) & 0xFF;
121 const auto offsets = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&lookup_indices[lookup]));
122 _mm_storeu_si128(reinterpret_cast<__m128i*>(out + count), _mm_add_epi16(base, offsets));
123 count += lookup_count[lookup];
124 base = _mm_add_epi16(base, increment);
132 // Sparse input implementation
133 template <IndexType InDims, IndexType OutDims>
134 class AffineTransformSparseInput {
138 using InputType = std::uint8_t;
139 using OutputType = std::int32_t;
141 // Number of input/output dimensions
142 static constexpr IndexType InputDimensions = InDims;
143 static constexpr IndexType OutputDimensions = OutDims;
145 static_assert(OutputDimensions % 16 == 0, "Only implemented for OutputDimensions divisible by 16.");
147 static constexpr IndexType PaddedInputDimensions =
148 ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
149 static constexpr IndexType PaddedOutputDimensions =
150 ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
152 #if defined (USE_SSSE3)
153 static constexpr IndexType ChunkSize = 4;
155 static constexpr IndexType ChunkSize = 1;
158 using OutputBuffer = OutputType[PaddedOutputDimensions];
160 // Hash value embedded in the evaluation file
161 static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
162 std::uint32_t hashValue = 0xCC03DAE4u;
163 hashValue += OutputDimensions;
164 hashValue ^= prevHash >> 1;
165 hashValue ^= prevHash << 31;
169 static IndexType get_weight_index_scrambled(IndexType i)
172 (i / ChunkSize) % (PaddedInputDimensions / ChunkSize) * OutputDimensions * ChunkSize +
173 i / PaddedInputDimensions * ChunkSize +
177 static IndexType get_weight_index(IndexType i)
179 #if defined (USE_SSSE3)
180 return get_weight_index_scrambled(i);
186 // Read network parameters
187 bool read_parameters(std::istream& stream) {
188 read_little_endian<BiasType>(stream, biases, OutputDimensions);
189 for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
190 weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
192 return !stream.fail();
195 // Write network parameters
196 bool write_parameters(std::ostream& stream) const {
197 write_little_endian<BiasType>(stream, biases, OutputDimensions);
199 for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
200 write_little_endian<WeightType>(stream, weights[get_weight_index(i)]);
202 return !stream.fail();
204 // Forward propagation
205 const OutputType* propagate(
206 const InputType* input, OutputType* output) const {
208 #if defined (USE_SSSE3)
209 #if defined (USE_AVX512)
210 using vec_t = __m512i;
211 #define vec_setzero _mm512_setzero_si512
212 #define vec_set_32 _mm512_set1_epi32
213 #define vec_add_dpbusd_32 Simd::m512_add_dpbusd_epi32
214 #elif defined (USE_AVX2)
215 using vec_t = __m256i;
216 #define vec_setzero _mm256_setzero_si256
217 #define vec_set_32 _mm256_set1_epi32
218 #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
219 #elif defined (USE_SSSE3)
220 using vec_t = __m128i;
221 #define vec_setzero _mm_setzero_si128
222 #define vec_set_32 _mm_set1_epi32
223 #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
225 static constexpr IndexType OutputSimdWidth = sizeof(vec_t) / sizeof(OutputType);
227 constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / ChunkSize;
228 constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth;
229 std::uint16_t nnz[NumChunks];
232 const auto input32 = reinterpret_cast<const std::int32_t*>(input);
234 // Find indices of nonzero 32bit blocks
235 find_nnz<NumChunks>(input32, nnz, count);
237 const vec_t* biasvec = reinterpret_cast<const vec_t*>(biases);
239 for (IndexType k = 0; k < NumRegs; ++k)
242 for (IndexType j = 0; j < count; ++j)
244 const auto i = nnz[j];
245 const vec_t in = vec_set_32(input32[i]);
246 const auto col = reinterpret_cast<const vec_t*>(&weights[i * OutputDimensions * ChunkSize]);
247 for (IndexType k = 0; k < NumRegs; ++k)
248 vec_add_dpbusd_32(acc[k], in, col[k]);
251 vec_t* outptr = reinterpret_cast<vec_t*>(output);
252 for (IndexType k = 0; k < NumRegs; ++k)
256 # undef vec_add_dpbusd_32
258 // Use dense implementation for the other architectures.
259 affine_transform_non_ssse3<
261 PaddedInputDimensions,
262 OutputDimensions>(output, weights, biases, input);
269 using BiasType = OutputType;
270 using WeightType = std::int8_t;
272 alignas(CacheLineSize) BiasType biases[OutputDimensions];
273 alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
276 } // namespace Stockfish::Eval::NNUE::Layers
278 #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_SPARSE_INPUT_H_INCLUDED