git.sesse.net Git - stockfish/blob - src/nnue/layers/clipped_relu.h

   1 /*
   2   Stockfish, a UCI chess playing engine derived from Glaurung 2.1
   3   Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
   4
   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.
   9
  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.
  14
  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/>.
  17 */
  18
  19 // Definition of layer ClippedReLU of NNUE evaluation function
  20
  21 #ifndef NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED
  22 #define NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED
  23
  24 #include <algorithm>
  25 #include <cstdint>
  26 #include <iosfwd>
  27
  28 #include "../nnue_common.h"
  29
  30 namespace Stockfish::Eval::NNUE::Layers {
  31
  32 // Clipped ReLU
  33 template<IndexType InDims>
  34 class ClippedReLU {
  35    public:
  36     // Input/output type
  37     using InputType  = std::int32_t;
  38     using OutputType = std::uint8_t;
  39
  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);
  45
  46     using OutputBuffer = OutputType[PaddedOutputDimensions];
  47
  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;
  52         return hashValue;
  53     }
  54
  55     // Read network parameters
  56     bool read_parameters(std::istream&) { return true; }
  57
  58     // Write network parameters
  59     bool write_parameters(std::ostream&) const { return true; }
  60
  61     // Forward propagation
  62     void propagate(const InputType* input, OutputType* output) const {
  63
  64 #if defined(USE_AVX2)
  65         if constexpr (InputDimensions % SimdWidth == 0)
  66         {
  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             {
  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])),
  77                                     WeightScaleBits);
  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])),
  81                                     WeightScaleBits);
  82                 _mm256_store_si256(
  83                   &out[i], _mm256_permutevar8x32_epi32(
  84                              _mm256_max_epi8(_mm256_packs_epi16(words0, words1), Zero), Offsets));
  85             }
  86         }
  87         else
  88         {
  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)
  94             {
  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])),
  97                   WeightScaleBits);
  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])),
 100                   WeightScaleBits);
 101                 const __m128i packedbytes = _mm_packs_epi16(words0, words1);
 102                 _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero));
 103             }
 104         }
 105         constexpr IndexType Start = InputDimensions % SimdWidth == 0
 106                                     ? InputDimensions / SimdWidth * SimdWidth
 107                                     : InputDimensions / (SimdWidth / 2) * (SimdWidth / 2);
 108
 109 #elif defined(USE_SSE2)
 110         constexpr IndexType NumChunks = InputDimensions / SimdWidth;
 111
 112     #ifdef USE_SSE41
 113         const __m128i Zero = _mm_setzero_si128();
 114     #else
 115         const __m128i k0x80s = _mm_set1_epi8(-128);
 116     #endif
 117
 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)
 121         {
 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])),
 124               WeightScaleBits);
 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])),
 127               WeightScaleBits);
 128             const __m128i packedbytes = _mm_packs_epi16(words0, words1);
 129             _mm_store_si128(&out[i],
 130
 131     #ifdef USE_SSE41
 132                             _mm_max_epi8(packedbytes, Zero)
 133     #else
 134                             _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)
 135     #endif
 136
 137             );
 138         }
 139         constexpr IndexType Start = NumChunks * SimdWidth;
 140
 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)
 147         {
 148             int16x8_t  shifted;
 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);
 153         }
 154         constexpr IndexType Start = NumChunks * (SimdWidth / 2);
 155 #else
 156         constexpr IndexType Start = 0;
 157 #endif
 158
 159         for (IndexType i = Start; i < InputDimensions; ++i)
 160         {
 161             output[i] = static_cast<OutputType>(std::clamp(input[i] >> WeightScaleBits, 0, 127));
 162         }
 163     }
 164 };
 165
 166 }  // namespace Stockfish::Eval::NNUE::Layers
 167
 168 #endif  // NNUE_LAYERS_CLIPPED_RELU_H_INCLUDED