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
27 #include <type_traits>
28 #include "../nnue_common.h"
29 #include "affine_transform.h"
33 This file contains the definition for a fully connected layer (aka affine transform) with block sparse input.
36 namespace Stockfish::Eval::NNUE::Layers {
38 #if defined(USE_SSSE3)
39 alignas(CacheLineSize) static inline const std::array<std::array<std::uint16_t, 8>, 256> lookup_indices = [](){
40 std::array<std::array<std::uint16_t, 8>, 256> v{};
41 for (unsigned i = 0; i < 256; ++i)
43 std::uint64_t j = i, k = 0;
45 v[i][k++] = pop_lsb(j);
50 // Find indices of nonzero numbers in an int32_t array
51 template<const IndexType InputDimensions>
52 void find_nnz(const std::int32_t* input, std::uint16_t* out, IndexType& count_out) {
53 #if defined (USE_AVX512)
54 using vec_t = __m512i;
55 #define vec_nnz(a) _mm512_cmpgt_epi32_mask(a, _mm512_setzero_si512())
56 #elif defined (USE_AVX2)
57 using vec_t = __m256i;
58 #if defined(USE_VNNI) && !defined(USE_AVXVNNI)
59 #define vec_nnz(a) _mm256_cmpgt_epi32_mask(a, _mm256_setzero_si256())
61 #define vec_nnz(a) _mm256_movemask_ps(_mm256_castsi256_ps(_mm256_cmpgt_epi32(a, _mm256_setzero_si256())))
63 #elif defined (USE_SSSE3)
64 using vec_t = __m128i;
65 #define vec_nnz(a) _mm_movemask_ps(_mm_castsi128_ps(_mm_cmpgt_epi32(a, _mm_setzero_si128())))
67 constexpr IndexType InputSimdWidth = sizeof(vec_t) / sizeof(std::int32_t);
68 // Inputs are processed InputSimdWidth at a time and outputs are processed 8 at a time so we process in chunks of max(InputSimdWidth, 8)
69 constexpr IndexType ChunkSize = std::max<IndexType>(InputSimdWidth, 8);
70 constexpr IndexType NumChunks = InputDimensions / ChunkSize;
71 constexpr IndexType InputsPerChunk = ChunkSize / InputSimdWidth;
72 constexpr IndexType OutputsPerChunk = ChunkSize / 8;
74 const auto inputVector = reinterpret_cast<const vec_t*>(input);
76 __m128i base = _mm_setzero_si128();
77 const __m128i increment = _mm_set1_epi16(8);
78 for (IndexType i = 0; i < NumChunks; ++i)
80 // bitmask of nonzero values in this chunk
82 for (IndexType j = 0; j < InputsPerChunk; ++j)
84 const vec_t inputChunk = inputVector[i * InputsPerChunk + j];
85 nnz |= (unsigned)vec_nnz(inputChunk) << (j * InputSimdWidth);
87 for (IndexType j = 0; j < OutputsPerChunk; ++j)
89 const auto lookup = (nnz >> (j * 8)) & 0xFF;
90 const auto offsets = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&lookup_indices[lookup]));
91 _mm_storeu_si128(reinterpret_cast<__m128i*>(out + count), _mm_add_epi16(base, offsets));
92 count += popcount(lookup);
93 base = _mm_add_epi16(base, increment);
101 // Sparse input implementation
102 template <IndexType InDims, IndexType OutDims>
103 class AffineTransformSparseInput {
106 using InputType = std::uint8_t;
107 using OutputType = std::int32_t;
109 // Number of input/output dimensions
110 static constexpr IndexType InputDimensions = InDims;
111 static constexpr IndexType OutputDimensions = OutDims;
113 static_assert(OutputDimensions % 16 == 0, "Only implemented for OutputDimensions divisible by 16.");
115 static constexpr IndexType PaddedInputDimensions =
116 ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
117 static constexpr IndexType PaddedOutputDimensions =
118 ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
120 #if defined (USE_SSSE3)
121 static constexpr IndexType ChunkSize = 4;
123 static constexpr IndexType ChunkSize = 1;
126 using OutputBuffer = OutputType[PaddedOutputDimensions];
128 // Hash value embedded in the evaluation file
129 static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
130 std::uint32_t hashValue = 0xCC03DAE4u;
131 hashValue += OutputDimensions;
132 hashValue ^= prevHash >> 1;
133 hashValue ^= prevHash << 31;
137 static constexpr IndexType get_weight_index_scrambled(IndexType i)
140 (i / ChunkSize) % (PaddedInputDimensions / ChunkSize) * OutputDimensions * ChunkSize +
141 i / PaddedInputDimensions * ChunkSize +
145 static constexpr IndexType get_weight_index(IndexType i)
147 #if defined (USE_SSSE3)
148 return get_weight_index_scrambled(i);
154 // Read network parameters
155 bool read_parameters(std::istream& stream) {
156 read_little_endian<BiasType>(stream, biases, OutputDimensions);
157 for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
158 weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
160 return !stream.fail();
163 // Write network parameters
164 bool write_parameters(std::ostream& stream) const {
165 write_little_endian<BiasType>(stream, biases, OutputDimensions);
167 for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
168 write_little_endian<WeightType>(stream, weights[get_weight_index(i)]);
170 return !stream.fail();
172 // Forward propagation
174 const InputType* input, OutputType* output) const {
176 #if defined (USE_SSSE3)
177 #if defined (USE_AVX512)
178 using vec_t = __m512i;
179 #define vec_setzero _mm512_setzero_si512
180 #define vec_set_32 _mm512_set1_epi32
181 #define vec_add_dpbusd_32 Simd::m512_add_dpbusd_epi32
182 #elif defined (USE_AVX2)
183 using vec_t = __m256i;
184 #define vec_setzero _mm256_setzero_si256
185 #define vec_set_32 _mm256_set1_epi32
186 #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
187 #elif defined (USE_SSSE3)
188 using vec_t = __m128i;
189 #define vec_setzero _mm_setzero_si128
190 #define vec_set_32 _mm_set1_epi32
191 #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
193 static constexpr IndexType OutputSimdWidth = sizeof(vec_t) / sizeof(OutputType);
195 constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / ChunkSize;
196 constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth;
197 std::uint16_t nnz[NumChunks];
200 const auto input32 = reinterpret_cast<const std::int32_t*>(input);
202 // Find indices of nonzero 32bit blocks
203 find_nnz<NumChunks>(input32, nnz, count);
205 const vec_t* biasvec = reinterpret_cast<const vec_t*>(biases);
207 for (IndexType k = 0; k < NumRegs; ++k)
210 for (IndexType j = 0; j < count; ++j)
212 const auto i = nnz[j];
213 const vec_t in = vec_set_32(input32[i]);
214 const auto col = reinterpret_cast<const vec_t*>(&weights[i * OutputDimensions * ChunkSize]);
215 for (IndexType k = 0; k < NumRegs; ++k)
216 vec_add_dpbusd_32(acc[k], in, col[k]);
219 vec_t* outptr = reinterpret_cast<vec_t*>(output);
220 for (IndexType k = 0; k < NumRegs; ++k)
224 # undef vec_add_dpbusd_32
226 // Use dense implementation for the other architectures.
227 affine_transform_non_ssse3<
229 PaddedInputDimensions,
230 OutputDimensions>(output, weights, biases, input);
235 using BiasType = OutputType;
236 using WeightType = std::int8_t;
238 alignas(CacheLineSize) BiasType biases[OutputDimensions];
239 alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
242 } // namespace Stockfish::Eval::NNUE::Layers
244 #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_SPARSE_INPUT_H_INCLUDED