2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2020 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 Eval::NNUE::Layers {
29 template <typename PreviousLayer>
33 using InputType = typename PreviousLayer::OutputType;
34 using OutputType = std::uint8_t;
35 static_assert(std::is_same<InputType, std::int32_t>::value, "");
37 // Number of input/output dimensions
38 static constexpr IndexType kInputDimensions =
39 PreviousLayer::kOutputDimensions;
40 static constexpr IndexType kOutputDimensions = kInputDimensions;
42 // Size of forward propagation buffer used in this layer
43 static constexpr std::size_t kSelfBufferSize =
44 CeilToMultiple(kOutputDimensions * sizeof(OutputType), kCacheLineSize);
46 // Size of the forward propagation buffer used from the input layer to this layer
47 static constexpr std::size_t kBufferSize =
48 PreviousLayer::kBufferSize + kSelfBufferSize;
50 // Hash value embedded in the evaluation file
51 static constexpr std::uint32_t GetHashValue() {
52 std::uint32_t hash_value = 0x538D24C7u;
53 hash_value += PreviousLayer::GetHashValue();
57 // Read network parameters
58 bool ReadParameters(std::istream& stream) {
59 return previous_layer_.ReadParameters(stream);
62 // Forward propagation
63 const OutputType* Propagate(
64 const TransformedFeatureType* transformed_features, char* buffer) const {
65 const auto input = previous_layer_.Propagate(
66 transformed_features, buffer + kSelfBufferSize);
67 const auto output = reinterpret_cast<OutputType*>(buffer);
70 constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth;
71 const __m256i kZero = _mm256_setzero_si256();
72 const __m256i kOffsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
73 const auto in = reinterpret_cast<const __m256i*>(input);
74 const auto out = reinterpret_cast<__m256i*>(output);
75 for (IndexType i = 0; i < kNumChunks; ++i) {
76 const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
78 #if defined(__MINGW32__) || defined(__MINGW64__)
79 // HACK: Use _mm256_loadu_si256() instead of _mm256_load_si256. Because the binary
80 // compiled with g++ in MSYS2 crashes here because the output memory is not aligned
81 // even though alignas is specified.
89 #if defined(__MINGW32__) || defined(__MINGW64__)
95 (&in[i * 4 + 1])), kWeightScaleBits);
96 const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
98 #if defined(__MINGW32__) || defined(__MINGW64__)
106 #if defined(__MINGW32__) || defined(__MINGW64__)
112 (&in[i * 4 + 3])), kWeightScaleBits);
114 #if defined(__MINGW32__) || defined(__MINGW64__)
120 (&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
121 _mm256_packs_epi16(words0, words1), kZero), kOffsets));
123 constexpr IndexType kStart = kNumChunks * kSimdWidth;
125 #elif defined(USE_SSSE3)
126 constexpr IndexType kNumChunks = kInputDimensions / kSimdWidth;
129 const __m128i kZero = _mm_setzero_si128();
131 const __m128i k0x80s = _mm_set1_epi8(-128);
134 const auto in = reinterpret_cast<const __m128i*>(input);
135 const auto out = reinterpret_cast<__m128i*>(output);
136 for (IndexType i = 0; i < kNumChunks; ++i) {
137 const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
138 _mm_load_si128(&in[i * 4 + 0]),
139 _mm_load_si128(&in[i * 4 + 1])), kWeightScaleBits);
140 const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
141 _mm_load_si128(&in[i * 4 + 2]),
142 _mm_load_si128(&in[i * 4 + 3])), kWeightScaleBits);
143 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
144 _mm_store_si128(&out[i],
147 _mm_max_epi8(packedbytes, kZero)
149 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
154 constexpr IndexType kStart = kNumChunks * kSimdWidth;
156 #elif defined(USE_NEON)
157 constexpr IndexType kNumChunks = kInputDimensions / (kSimdWidth / 2);
158 const int8x8_t kZero = {0};
159 const auto in = reinterpret_cast<const int32x4_t*>(input);
160 const auto out = reinterpret_cast<int8x8_t*>(output);
161 for (IndexType i = 0; i < kNumChunks; ++i) {
163 const auto pack = reinterpret_cast<int16x4_t*>(&shifted);
164 pack[0] = vqshrn_n_s32(in[i * 2 + 0], kWeightScaleBits);
165 pack[1] = vqshrn_n_s32(in[i * 2 + 1], kWeightScaleBits);
166 out[i] = vmax_s8(vqmovn_s16(shifted), kZero);
168 constexpr IndexType kStart = kNumChunks * (kSimdWidth / 2);
170 constexpr IndexType kStart = 0;
173 for (IndexType i = kStart; i < kInputDimensions; ++i) {
174 output[i] = static_cast<OutputType>(
175 std::max(0, std::min(127, input[i] >> kWeightScaleBits)));
181 PreviousLayer previous_layer_;
184 } // namespace Eval::NNUE::Layers
186 #endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED