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 (USE_SSSE3 | (USE_NEON >= 8))
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_SSSE3)
54 #if defined (USE_AVX512)
55 using vec_t = __m512i;
56 #define vec_nnz(a) _mm512_cmpgt_epi32_mask(a, _mm512_setzero_si512())
57 #elif defined (USE_AVX2)
58 using vec_t = __m256i;
59 #if defined(USE_VNNI) && !defined(USE_AVXVNNI)
60 #define vec_nnz(a) _mm256_cmpgt_epi32_mask(a, _mm256_setzero_si256())
62 #define vec_nnz(a) _mm256_movemask_ps(_mm256_castsi256_ps(_mm256_cmpgt_epi32(a, _mm256_setzero_si256())))
64 #elif defined (USE_SSSE3)
65 using vec_t = __m128i;
66 #define vec_nnz(a) _mm_movemask_ps(_mm_castsi128_ps(_mm_cmpgt_epi32(a, _mm_setzero_si128())))
68 using vec128_t = __m128i;
69 #define vec128_zero _mm_setzero_si128()
70 #define vec128_set_16(a) _mm_set1_epi16(a)
71 #define vec128_load(a) _mm_load_si128(a)
72 #define vec128_storeu(a, b) _mm_storeu_si128(a, b)
73 #define vec128_add(a, b) _mm_add_epi16(a, b)
74 #elif defined (USE_NEON)
75 using vec_t = uint32x4_t;
76 static const std::uint32_t Mask[4] = {1, 2, 4, 8};
77 #define vec_nnz(a) vaddvq_u32(vandq_u32(vtstq_u32(a, a), vld1q_u32(Mask)))
78 using vec128_t = uint16x8_t;
79 #define vec128_zero vdupq_n_u16(0)
80 #define vec128_set_16(a) vdupq_n_u16(a)
81 #define vec128_load(a) vld1q_u16(reinterpret_cast<const std::uint16_t*>(a))
82 #define vec128_storeu(a, b) vst1q_u16(reinterpret_cast<std::uint16_t*>(a), b)
83 #define vec128_add(a, b) vaddq_u16(a, b)
85 constexpr IndexType InputSimdWidth = sizeof(vec_t) / sizeof(std::int32_t);
86 // Inputs are processed InputSimdWidth at a time and outputs are processed 8 at a time so we process in chunks of max(InputSimdWidth, 8)
87 constexpr IndexType ChunkSize = std::max<IndexType>(InputSimdWidth, 8);
88 constexpr IndexType NumChunks = InputDimensions / ChunkSize;
89 constexpr IndexType InputsPerChunk = ChunkSize / InputSimdWidth;
90 constexpr IndexType OutputsPerChunk = ChunkSize / 8;
92 const auto inputVector = reinterpret_cast<const vec_t*>(input);
94 vec128_t base = vec128_zero;
95 const vec128_t increment = vec128_set_16(8);
96 for (IndexType i = 0; i < NumChunks; ++i)
98 // bitmask of nonzero values in this chunk
100 for (IndexType j = 0; j < InputsPerChunk; ++j)
102 const vec_t inputChunk = inputVector[i * InputsPerChunk + j];
103 nnz |= (unsigned)vec_nnz(inputChunk) << (j * InputSimdWidth);
105 for (IndexType j = 0; j < OutputsPerChunk; ++j)
107 const auto lookup = (nnz >> (j * 8)) & 0xFF;
108 const auto offsets = vec128_load(reinterpret_cast<const vec128_t*>(&lookup_indices[lookup]));
109 vec128_storeu(reinterpret_cast<vec128_t*>(out + count), vec128_add(base, offsets));
110 count += popcount(lookup);
111 base = vec128_add(base, increment);
118 # undef vec128_set_16
120 # undef vec128_storeu
124 // Sparse input implementation
125 template <IndexType InDims, IndexType OutDims>
126 class AffineTransformSparseInput {
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_assert(OutputDimensions % 16 == 0, "Only implemented for OutputDimensions divisible by 16.");
138 static constexpr IndexType PaddedInputDimensions =
139 ceil_to_multiple<IndexType>(InputDimensions, MaxSimdWidth);
140 static constexpr IndexType PaddedOutputDimensions =
141 ceil_to_multiple<IndexType>(OutputDimensions, MaxSimdWidth);
143 #if (USE_SSSE3 | (USE_NEON >= 8))
144 static constexpr IndexType ChunkSize = 4;
146 static constexpr IndexType ChunkSize = 1;
149 using OutputBuffer = OutputType[PaddedOutputDimensions];
151 // Hash value embedded in the evaluation file
152 static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
153 std::uint32_t hashValue = 0xCC03DAE4u;
154 hashValue += OutputDimensions;
155 hashValue ^= prevHash >> 1;
156 hashValue ^= prevHash << 31;
160 static constexpr IndexType get_weight_index_scrambled(IndexType i)
163 (i / ChunkSize) % (PaddedInputDimensions / ChunkSize) * OutputDimensions * ChunkSize +
164 i / PaddedInputDimensions * ChunkSize +
168 static constexpr IndexType get_weight_index(IndexType i)
170 #if (USE_SSSE3 | (USE_NEON >= 8))
171 return get_weight_index_scrambled(i);
177 // Read network parameters
178 bool read_parameters(std::istream& stream) {
179 read_little_endian<BiasType>(stream, biases, OutputDimensions);
180 for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
181 weights[get_weight_index(i)] = read_little_endian<WeightType>(stream);
183 return !stream.fail();
186 // Write network parameters
187 bool write_parameters(std::ostream& stream) const {
188 write_little_endian<BiasType>(stream, biases, OutputDimensions);
190 for (IndexType i = 0; i < OutputDimensions * PaddedInputDimensions; ++i)
191 write_little_endian<WeightType>(stream, weights[get_weight_index(i)]);
193 return !stream.fail();
195 // Forward propagation
197 const InputType* input, OutputType* output) const {
199 #if (USE_SSSE3 | (USE_NEON >= 8))
200 #if defined (USE_AVX512)
201 using invec_t = __m512i;
202 using outvec_t = __m512i;
203 #define vec_set_32 _mm512_set1_epi32
204 #define vec_add_dpbusd_32 Simd::m512_add_dpbusd_epi32
205 #elif defined (USE_AVX2)
206 using invec_t = __m256i;
207 using outvec_t = __m256i;
208 #define vec_set_32 _mm256_set1_epi32
209 #define vec_add_dpbusd_32 Simd::m256_add_dpbusd_epi32
210 #elif defined (USE_SSSE3)
211 using invec_t = __m128i;
212 using outvec_t = __m128i;
213 #define vec_set_32 _mm_set1_epi32
214 #define vec_add_dpbusd_32 Simd::m128_add_dpbusd_epi32
215 #elif defined (USE_NEON_DOTPROD)
216 using invec_t = int8x16_t;
217 using outvec_t = int32x4_t;
218 #define vec_set_32(a) vreinterpretq_s8_u32(vdupq_n_u32(a))
219 #define vec_add_dpbusd_32 Simd::dotprod_m128_add_dpbusd_epi32
220 #elif defined (USE_NEON)
221 using invec_t = int8x16_t;
222 using outvec_t = int32x4_t;
223 #define vec_set_32(a) vreinterpretq_s8_u32(vdupq_n_u32(a))
224 #define vec_add_dpbusd_32 Simd::neon_m128_add_dpbusd_epi32
226 static constexpr IndexType OutputSimdWidth = sizeof(outvec_t) / sizeof(OutputType);
228 constexpr IndexType NumChunks = ceil_to_multiple<IndexType>(InputDimensions, 8) / ChunkSize;
229 constexpr IndexType NumRegs = OutputDimensions / OutputSimdWidth;
230 std::uint16_t nnz[NumChunks];
233 const auto input32 = reinterpret_cast<const std::int32_t*>(input);
235 // Find indices of nonzero 32bit blocks
236 find_nnz<NumChunks>(input32, nnz, count);
238 const outvec_t* biasvec = reinterpret_cast<const outvec_t*>(biases);
239 outvec_t acc[NumRegs];
240 for (IndexType k = 0; k < NumRegs; ++k)
243 for (IndexType j = 0; j < count; ++j)
245 const auto i = nnz[j];
246 const invec_t in = vec_set_32(input32[i]);
247 const auto col = reinterpret_cast<const invec_t*>(&weights[i * OutputDimensions * ChunkSize]);
248 for (IndexType k = 0; k < NumRegs; ++k)
249 vec_add_dpbusd_32(acc[k], in, col[k]);
252 outvec_t* outptr = reinterpret_cast<outvec_t*>(output);
253 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);
267 using BiasType = OutputType;
268 using WeightType = std::int8_t;
270 alignas(CacheLineSize) BiasType biases[OutputDimensions];
271 alignas(CacheLineSize) WeightType weights[OutputDimensions * PaddedInputDimensions];
274 } // namespace Stockfish::Eval::NNUE::Layers
276 #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_SPARSE_INPUT_H_INCLUDED