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 ClippedReLU of NNUE evaluation function
21 #ifndef NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED
22 #define NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED
28 #include "../nnue_common.h"
30 namespace Stockfish::Eval::NNUE::Layers {
33 template <IndexType InDims>
37 using InputType = std::int32_t;
38 using OutputType = std::uint8_t;
40 // Number of input/output dimensions
41 static constexpr IndexType InputDimensions = InDims;
42 static constexpr IndexType OutputDimensions = InputDimensions;
43 static constexpr IndexType PaddedOutputDimensions =
44 ceil_to_multiple<IndexType>(OutputDimensions, 32);
46 using OutputBuffer = OutputType[PaddedOutputDimensions];
48 // Hash value embedded in the evaluation file
49 static constexpr std::uint32_t get_hash_value(std::uint32_t prevHash) {
50 std::uint32_t hashValue = 0x538D24C7u;
51 hashValue += prevHash;
55 // Read network parameters
56 bool read_parameters(std::istream&) {
60 // Write network parameters
61 bool write_parameters(std::ostream&) const {
65 // Forward propagation
67 const InputType* input, OutputType* output) const {
70 if constexpr (InputDimensions % SimdWidth == 0) {
71 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
72 const __m256i Zero = _mm256_setzero_si256();
73 const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
74 const auto in = reinterpret_cast<const __m256i*>(input);
75 const auto out = reinterpret_cast<__m256i*>(output);
76 for (IndexType i = 0; i < NumChunks; ++i) {
77 const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
78 _mm256_load_si256(&in[i * 4 + 0]),
79 _mm256_load_si256(&in[i * 4 + 1])), WeightScaleBits);
80 const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
81 _mm256_load_si256(&in[i * 4 + 2]),
82 _mm256_load_si256(&in[i * 4 + 3])), WeightScaleBits);
83 _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
84 _mm256_packs_epi16(words0, words1), Zero), Offsets));
87 constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
88 const __m128i Zero = _mm_setzero_si128();
89 const auto in = reinterpret_cast<const __m128i*>(input);
90 const auto out = reinterpret_cast<__m128i*>(output);
91 for (IndexType i = 0; i < NumChunks; ++i) {
92 const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
93 _mm_load_si128(&in[i * 4 + 0]),
94 _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
95 const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
96 _mm_load_si128(&in[i * 4 + 2]),
97 _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
98 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
99 _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero));
102 constexpr IndexType Start =
103 InputDimensions % SimdWidth == 0
104 ? InputDimensions / SimdWidth * SimdWidth
105 : InputDimensions / (SimdWidth / 2) * (SimdWidth / 2);
107 #elif defined(USE_SSE2)
108 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
111 const __m128i Zero = _mm_setzero_si128();
113 const __m128i k0x80s = _mm_set1_epi8(-128);
116 const auto in = reinterpret_cast<const __m128i*>(input);
117 const auto out = reinterpret_cast<__m128i*>(output);
118 for (IndexType i = 0; i < NumChunks; ++i) {
119 const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
120 _mm_load_si128(&in[i * 4 + 0]),
121 _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
122 const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
123 _mm_load_si128(&in[i * 4 + 2]),
124 _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
125 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
126 _mm_store_si128(&out[i],
129 _mm_max_epi8(packedbytes, Zero)
131 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
136 constexpr IndexType Start = NumChunks * SimdWidth;
138 #elif defined(USE_NEON)
139 constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
140 const int8x8_t Zero = {0};
141 const auto in = reinterpret_cast<const int32x4_t*>(input);
142 const auto out = reinterpret_cast<int8x8_t*>(output);
143 for (IndexType i = 0; i < NumChunks; ++i) {
145 const auto pack = reinterpret_cast<int16x4_t*>(&shifted);
146 pack[0] = vqshrn_n_s32(in[i * 2 + 0], WeightScaleBits);
147 pack[1] = vqshrn_n_s32(in[i * 2 + 1], WeightScaleBits);
148 out[i] = vmax_s8(vqmovn_s16(shifted), Zero);
150 constexpr IndexType Start = NumChunks * (SimdWidth / 2);
152 constexpr IndexType Start = 0;
155 for (IndexType i = Start; i < InputDimensions; ++i) {
156 output[i] = static_cast<OutputType>(
157 std::clamp(input[i] >> WeightScaleBits, 0, 127));
162 } // namespace Stockfish::Eval::NNUE::Layers
164 #endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED