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&) { return true; }
58 // Write network parameters
59 bool write_parameters(std::ostream&) const { return true; }
61 // Forward propagation
62 void propagate(const InputType* input, OutputType* output) const {
65 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)
74 const __m256i words0 =
75 _mm256_srai_epi16(_mm256_packs_epi32(_mm256_load_si256(&in[i * 4 + 0]),
76 _mm256_load_si256(&in[i * 4 + 1])),
78 const __m256i words1 =
79 _mm256_srai_epi16(_mm256_packs_epi32(_mm256_load_si256(&in[i * 4 + 2]),
80 _mm256_load_si256(&in[i * 4 + 3])),
83 &out[i], _mm256_permutevar8x32_epi32(
84 _mm256_max_epi8(_mm256_packs_epi16(words0, words1), Zero), Offsets));
89 constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
90 const __m128i Zero = _mm_setzero_si128();
91 const auto in = reinterpret_cast<const __m128i*>(input);
92 const auto out = reinterpret_cast<__m128i*>(output);
93 for (IndexType i = 0; i < NumChunks; ++i)
95 const __m128i words0 = _mm_srai_epi16(
96 _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])),
98 const __m128i words1 = _mm_srai_epi16(
99 _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])),
101 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
102 _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero));
105 constexpr IndexType Start = InputDimensions % SimdWidth == 0
106 ? InputDimensions / SimdWidth * SimdWidth
107 : InputDimensions / (SimdWidth / 2) * (SimdWidth / 2);
109 #elif defined(USE_SSE2)
110 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
113 const __m128i Zero = _mm_setzero_si128();
115 const __m128i k0x80s = _mm_set1_epi8(-128);
118 const auto in = reinterpret_cast<const __m128i*>(input);
119 const auto out = reinterpret_cast<__m128i*>(output);
120 for (IndexType i = 0; i < NumChunks; ++i)
122 const __m128i words0 = _mm_srai_epi16(
123 _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])),
125 const __m128i words1 = _mm_srai_epi16(
126 _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])),
128 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
129 _mm_store_si128(&out[i],
132 _mm_max_epi8(packedbytes, Zero)
134 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
139 constexpr IndexType Start = NumChunks * SimdWidth;
141 #elif defined(USE_NEON)
142 constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
143 const int8x8_t Zero = {0};
144 const auto in = reinterpret_cast<const int32x4_t*>(input);
145 const auto out = reinterpret_cast<int8x8_t*>(output);
146 for (IndexType i = 0; i < NumChunks; ++i)
149 const auto pack = reinterpret_cast<int16x4_t*>(&shifted);
150 pack[0] = vqshrn_n_s32(in[i * 2 + 0], WeightScaleBits);
151 pack[1] = vqshrn_n_s32(in[i * 2 + 1], WeightScaleBits);
152 out[i] = vmax_s8(vqmovn_s16(shifted), Zero);
154 constexpr IndexType Start = NumChunks * (SimdWidth / 2);
156 constexpr IndexType Start = 0;
159 for (IndexType i = Start; i < InputDimensions; ++i)
161 output[i] = static_cast<OutputType>(std::clamp(input[i] >> WeightScaleBits, 0, 127));
166 } // namespace Stockfish::Eval::NNUE::Layers
168 #endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED