2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2021 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 <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 InputDimensions = PreviousLayer::OutputDimensions;
39 static constexpr IndexType OutputDimensions = InputDimensions;
40 static constexpr IndexType PaddedOutputDimensions =
41 ceil_to_multiple<IndexType>(OutputDimensions, 32);
43 // Size of forward propagation buffer used in this layer
44 static constexpr std::size_t SelfBufferSize =
45 ceil_to_multiple(OutputDimensions * sizeof(OutputType), CacheLineSize);
47 // Size of the forward propagation buffer used from the input layer to this layer
48 static constexpr std::size_t BufferSize =
49 PreviousLayer::BufferSize + SelfBufferSize;
51 // Hash value embedded in the evaluation file
52 static constexpr std::uint32_t get_hash_value() {
53 std::uint32_t hashValue = 0x538D24C7u;
54 hashValue += PreviousLayer::get_hash_value();
58 // Read network parameters
59 bool read_parameters(std::istream& stream) {
60 return previousLayer.read_parameters(stream);
63 // Write network parameters
64 bool write_parameters(std::ostream& stream) const {
65 return previousLayer.write_parameters(stream);
68 // Forward propagation
69 const OutputType* propagate(
70 const TransformedFeatureType* transformedFeatures, char* buffer) const {
71 const auto input = previousLayer.propagate(
72 transformedFeatures, buffer + SelfBufferSize);
73 const auto output = reinterpret_cast<OutputType*>(buffer);
76 if constexpr (InputDimensions % SimdWidth == 0) {
77 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
78 const __m256i Zero = _mm256_setzero_si256();
79 const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
80 const auto in = reinterpret_cast<const __m256i*>(input);
81 const auto out = reinterpret_cast<__m256i*>(output);
82 for (IndexType i = 0; i < NumChunks; ++i) {
83 const __m256i words0 = _mm256_srai_epi16(_mm256_packs_epi32(
84 _mm256_load_si256(&in[i * 4 + 0]),
85 _mm256_load_si256(&in[i * 4 + 1])), WeightScaleBits);
86 const __m256i words1 = _mm256_srai_epi16(_mm256_packs_epi32(
87 _mm256_load_si256(&in[i * 4 + 2]),
88 _mm256_load_si256(&in[i * 4 + 3])), WeightScaleBits);
89 _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32(_mm256_max_epi8(
90 _mm256_packs_epi16(words0, words1), Zero), Offsets));
93 constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
94 const __m128i Zero = _mm_setzero_si128();
95 const auto in = reinterpret_cast<const __m128i*>(input);
96 const auto out = reinterpret_cast<__m128i*>(output);
97 for (IndexType i = 0; i < NumChunks; ++i) {
98 const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
99 _mm_load_si128(&in[i * 4 + 0]),
100 _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
101 const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
102 _mm_load_si128(&in[i * 4 + 2]),
103 _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
104 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
105 _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero));
108 constexpr IndexType Start =
109 InputDimensions % SimdWidth == 0
110 ? InputDimensions / SimdWidth * SimdWidth
111 : InputDimensions / (SimdWidth / 2) * (SimdWidth / 2);
113 #elif defined(USE_SSE2)
114 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
117 const __m128i Zero = _mm_setzero_si128();
119 const __m128i k0x80s = _mm_set1_epi8(-128);
122 const auto in = reinterpret_cast<const __m128i*>(input);
123 const auto out = reinterpret_cast<__m128i*>(output);
124 for (IndexType i = 0; i < NumChunks; ++i) {
125 const __m128i words0 = _mm_srai_epi16(_mm_packs_epi32(
126 _mm_load_si128(&in[i * 4 + 0]),
127 _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits);
128 const __m128i words1 = _mm_srai_epi16(_mm_packs_epi32(
129 _mm_load_si128(&in[i * 4 + 2]),
130 _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits);
131 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
132 _mm_store_si128(&out[i],
135 _mm_max_epi8(packedbytes, Zero)
137 _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
142 constexpr IndexType Start = NumChunks * SimdWidth;
144 #elif defined(USE_MMX)
145 constexpr IndexType NumChunks = InputDimensions / SimdWidth;
146 const __m64 k0x80s = _mm_set1_pi8(-128);
147 const auto in = reinterpret_cast<const __m64*>(input);
148 const auto out = reinterpret_cast<__m64*>(output);
149 for (IndexType i = 0; i < NumChunks; ++i) {
150 const __m64 words0 = _mm_srai_pi16(
151 _mm_packs_pi32(in[i * 4 + 0], in[i * 4 + 1]),
153 const __m64 words1 = _mm_srai_pi16(
154 _mm_packs_pi32(in[i * 4 + 2], in[i * 4 + 3]),
156 const __m64 packedbytes = _mm_packs_pi16(words0, words1);
157 out[i] = _mm_subs_pi8(_mm_adds_pi8(packedbytes, k0x80s), k0x80s);
160 constexpr IndexType Start = NumChunks * SimdWidth;
162 #elif defined(USE_NEON)
163 constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2);
164 const int8x8_t Zero = {0};
165 const auto in = reinterpret_cast<const int32x4_t*>(input);
166 const auto out = reinterpret_cast<int8x8_t*>(output);
167 for (IndexType i = 0; i < NumChunks; ++i) {
169 const auto pack = reinterpret_cast<int16x4_t*>(&shifted);
170 pack[0] = vqshrn_n_s32(in[i * 2 + 0], WeightScaleBits);
171 pack[1] = vqshrn_n_s32(in[i * 2 + 1], WeightScaleBits);
172 out[i] = vmax_s8(vqmovn_s16(shifted), Zero);
174 constexpr IndexType Start = NumChunks * (SimdWidth / 2);
176 constexpr IndexType Start = 0;
179 for (IndexType i = Start; i < InputDimensions; ++i) {
180 output[i] = static_cast<OutputType>(
181 std::max(0, std::min(127, input[i] >> WeightScaleBits)));
184 // Affine transform layers expect that there is at least
185 // ceil_to_multiple(OutputDimensions, 32) initialized values.
186 // We cannot do this in the affine transform because it requires
187 // preallocating space here.
188 for (IndexType i = OutputDimensions; i < PaddedOutputDimensions; ++i) {
196 PreviousLayer previousLayer;
199 } // namespace Stockfish::Eval::NNUE::Layers
201 #endif // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED