2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2022 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 ClippedReLU of NNUE evaluation function
21 #ifndef NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED
22 #define NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED
24 #include "../nnue_common.h"
26 namespace Stockfish::Eval::NNUE::Layers {
29 template <IndexType InDims>
33 using InputType = std::int32_t;
34 using OutputType = std::uint8_t;
36 // Number of input/output dimensions
37 static constexpr IndexType InputDimensions = InDims;
38 static constexpr IndexType OutputDimensions = InputDimensions;
39 static constexpr IndexType PaddedOutputDimensions =
40 ceil_to_multiple<IndexType>(OutputDimensions, 32);
42 using OutputBuffer = OutputType[PaddedOutputDimensions];
44 // Hash value embedded in the evaluation file
45 static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
46 std::uint32_t hashValue = 0x538D24C7u;
47 hashValue += prevHash;
51 // Read network parameters
52 bool read_parameters(std::istream&) {
56 // Write network parameters
57 bool write_parameters(std::ostream&) const {
61 // Forward propagation
62 const OutputType* propagate(
63 const InputType* input, OutputType* output) const {
66 if constexpr (InputDimensions % SimdWidth == 0) {
67 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
68 const __m256i Zero = _mm256_setzero_si256();
69 const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
70 const auto in = reinterpret_cast<const __m256i*>(input);
71 const auto out = reinterpret_cast<__m256i*>(output);
72 for (IndexType i = 0; i < NumChunks; ++i) {
73 const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
74 _mm256_load_si256(&in[i * 4 + 0]),
75 _mm256_load_si256(&in[i * 4 + 1])), WeightScaleBits);
76 const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
77 _mm256_load_si256(&in[i * 4 + 2]),
78 _mm256_load_si256(&in[i * 4 + 3])), WeightScaleBits);
79 _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
80 _mm256_packs_epi16(words0, words1), Zero), Offsets));
83 constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
84 const __m128i Zero = _mm_setzero_si128();
85 const auto in = reinterpret_cast<const __m128i*>(input);
86 const auto out = reinterpret_cast<__m128i*>(output);
87 for (IndexType i = 0; i < NumChunks; ++i) {
88 const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
89 _mm_load_si128(&in[i * 4 + 0]),
90 _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
91 const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
92 _mm_load_si128(&in[i * 4 + 2]),
93 _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
94 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
95 _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero));
98 constexpr IndexType Start =
99 InputDimensions % SimdWidth == 0
100 ? InputDimensions / SimdWidth * SimdWidth
101 : InputDimensions / (SimdWidth / 2) * (SimdWidth / 2);
103 #elif defined(USE_SSE2)
104 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
107 const __m128i Zero = _mm_setzero_si128();
109 const __m128i k0x80s = _mm_set1_epi8(-128);
112 const auto in = reinterpret_cast<const __m128i*>(input);
113 const auto out = reinterpret_cast<__m128i*>(output);
114 for (IndexType i = 0; i < NumChunks; ++i) {
115 const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
116 _mm_load_si128(&in[i * 4 + 0]),
117 _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
118 const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
119 _mm_load_si128(&in[i * 4 + 2]),
120 _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
121 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
122 _mm_store_si128(&out[i],
125 _mm_max_epi8(packedbytes, Zero)
127 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
132 constexpr IndexType Start = NumChunks * SimdWidth;
134 #elif defined(USE_MMX)
135 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
136 const __m64 k0x80s = _mm_set1_pi8(-128);
137 const auto in = reinterpret_cast<const __m64*>(input);
138 const auto out = reinterpret_cast<__m64*>(output);
139 for (IndexType i = 0; i < NumChunks; ++i) {
140 const __m64 words0 = _mm_srai_pi16(
141 _mm_packs_pi32(in[i * 4 + 0], in[i * 4 + 1]),
143 const __m64 words1 = _mm_srai_pi16(
144 _mm_packs_pi32(in[i * 4 + 2], in[i * 4 + 3]),
146 const __m64 packedbytes = _mm_packs_pi16(words0, words1);
147 out[i] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
150 constexpr IndexType Start = NumChunks * SimdWidth;
152 #elif defined(USE_NEON)
153 constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
154 const int8x8_t Zero = {0};
155 const auto in = reinterpret_cast<const int32x4_t*>(input);
156 const auto out = reinterpret_cast<int8x8_t*>(output);
157 for (IndexType i = 0; i < NumChunks; ++i) {
159 const auto pack = reinterpret_cast<int16x4_t*>(&shifted);
160 pack[0] = vqshrn_n_s32(in[i * 2 + 0], WeightScaleBits);
161 pack[1] = vqshrn_n_s32(in[i * 2 + 1], WeightScaleBits);
162 out[i] = vmax_s8(vqmovn_s16(shifted), Zero);
164 constexpr IndexType Start = NumChunks * (SimdWidth / 2);
166 constexpr IndexType Start = 0;
169 for (IndexType i = Start; i < InputDimensions; ++i) {
170 output[i] = static_cast<OutputType>(
171 std::max(0, std::min(127, input[i] >> WeightScaleBits)));
174 // Affine transform layers expect that there is at least
175 // ceil_to_multiple(OutputDimensions, 32) initialized values.
176 // We cannot do this in the affine transform because it requires
177 // preallocating space here.
178 for (IndexType i = OutputDimensions; i < PaddedOutputDimensions; ++i) {
186 } // namespace Stockfish::Eval::NNUE::Layers
188 #endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED