Provide vectorized NNUE code for SSE2 and MMX targets
[stockfish] / src / nnue / layers / affine_transform.h
1 /*
2   Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3   Copyright (C) 2004-2020 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 AffineTransform of NNUE evaluation function
20
21 #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
22 #define NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED
23
24 #include <iostream>
25 #include "../nnue_common.h"
26
27 namespace Eval::NNUE::Layers {
28
29   // Affine transformation layer
30   template <typename PreviousLayer, IndexType OutputDimensions>
31   class AffineTransform {
32    public:
33     // Input/output type
34     using InputType = typename PreviousLayer::OutputType;
35     using OutputType = std::int32_t;
36     static_assert(std::is_same<InputType, std::uint8_t>::value, "");
37
38     // Number of input/output dimensions
39     static constexpr IndexType kInputDimensions =
40         PreviousLayer::kOutputDimensions;
41     static constexpr IndexType kOutputDimensions = OutputDimensions;
42     static constexpr IndexType kPaddedInputDimensions =
43         CeilToMultiple<IndexType>(kInputDimensions, kMaxSimdWidth);
44
45     // Size of forward propagation buffer used in this layer
46     static constexpr std::size_t kSelfBufferSize =
47         CeilToMultiple(kOutputDimensions * sizeof(OutputType), kCacheLineSize);
48
49     // Size of the forward propagation buffer used from the input layer to this layer
50     static constexpr std::size_t kBufferSize =
51         PreviousLayer::kBufferSize + kSelfBufferSize;
52
53     // Hash value embedded in the evaluation file
54     static constexpr std::uint32_t GetHashValue() {
55       std::uint32_t hash_value = 0xCC03DAE4u;
56       hash_value += kOutputDimensions;
57       hash_value ^= PreviousLayer::GetHashValue() >> 1;
58       hash_value ^= PreviousLayer::GetHashValue() << 31;
59       return hash_value;
60     }
61
62    // Read network parameters
63     bool ReadParameters(std::istream& stream) {
64       if (!previous_layer_.ReadParameters(stream)) return false;
65       stream.read(reinterpret_cast<char*>(biases_),
66                   kOutputDimensions * sizeof(BiasType));
67       stream.read(reinterpret_cast<char*>(weights_),
68                   kOutputDimensions * kPaddedInputDimensions *
69                   sizeof(WeightType));
70       return !stream.fail();
71     }
72
73     // Forward propagation
74     const OutputType* Propagate(
75         const TransformedFeatureType* transformed_features, char* buffer) const {
76       const auto input = previous_layer_.Propagate(
77           transformed_features, buffer + kSelfBufferSize);
78       const auto output = reinterpret_cast<OutputType*>(buffer);
79
80   #if defined(USE_AVX512)
81       constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2);
82       const __m512i kOnes = _mm512_set1_epi16(1);
83       const auto input_vector = reinterpret_cast<const __m512i*>(input);
84
85   #elif defined(USE_AVX2)
86       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
87       const __m256i kOnes = _mm256_set1_epi16(1);
88       const auto input_vector = reinterpret_cast<const __m256i*>(input);
89
90   #elif defined(USE_SSE2)
91       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
92   #ifndef USE_SSSE3
93       const __m128i kZeros = _mm_setzero_si128();
94   #else
95       const __m128i kOnes = _mm_set1_epi16(1);
96   #endif
97       const auto input_vector = reinterpret_cast<const __m128i*>(input);
98
99   #elif defined(USE_MMX)
100       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
101       const __m64 kZeros = _mm_setzero_si64();
102       const auto input_vector = reinterpret_cast<const __m64*>(input);
103
104   #elif defined(USE_NEON)
105       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
106       const auto input_vector = reinterpret_cast<const int8x8_t*>(input);
107   #endif
108
109       for (IndexType i = 0; i < kOutputDimensions; ++i) {
110         const IndexType offset = i * kPaddedInputDimensions;
111
112   #if defined(USE_AVX512)
113         __m512i sum = _mm512_setzero_si512();
114         const auto row = reinterpret_cast<const __m512i*>(&weights_[offset]);
115         for (IndexType j = 0; j < kNumChunks; ++j) {
116             __m512i product = _mm512_maddubs_epi16(_mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
117             product = _mm512_madd_epi16(product, kOnes);
118             sum = _mm512_add_epi32(sum, product);
119         }
120
121         // Note: Changing kMaxSimdWidth from 32 to 64 breaks loading existing networks.
122         // As a result kPaddedInputDimensions may not be an even multiple of 64(512bit)
123         // and we have to do one more 256bit chunk.
124         if (kPaddedInputDimensions != kNumChunks * kSimdWidth * 2)
125         {
126             const auto iv256  = reinterpret_cast<const __m256i*>(&input_vector[kNumChunks]);
127             const auto row256 = reinterpret_cast<const __m256i*>(&row[kNumChunks]);
128             __m256i product256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
129             product256 = _mm256_madd_epi16(product256, _mm256_set1_epi16(1));
130             sum = _mm512_add_epi32(sum, _mm512_zextsi256_si512(product256));
131         }
132         output[i] = _mm512_reduce_add_epi32(sum) + biases_[i];
133
134   #elif defined(USE_AVX2)
135         __m256i sum = _mm256_setzero_si256();
136         const auto row = reinterpret_cast<const __m256i*>(&weights_[offset]);
137         for (IndexType j = 0; j < kNumChunks; ++j) {
138           __m256i product = _mm256_maddubs_epi16(_mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j]));
139           product = _mm256_madd_epi16(product, kOnes);
140           sum = _mm256_add_epi32(sum, product);
141         }
142         __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
143         sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));
144         sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB));
145         output[i] = _mm_cvtsi128_si32(sum128) + biases_[i];
146
147   #elif defined(USE_SSSE3)
148         __m128i sum = _mm_setzero_si128();
149         const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
150         for (int j = 0; j < (int)kNumChunks - 1; j += 2) {
151           __m128i product0 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j]), _mm_load_si128(&row[j]));
152           product0 = _mm_madd_epi16(product0, kOnes);
153           sum = _mm_add_epi32(sum, product0);
154           __m128i product1 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j+1]), _mm_load_si128(&row[j+1]));
155           product1 = _mm_madd_epi16(product1, kOnes);
156           sum = _mm_add_epi32(sum, product1);
157         }
158         if (kNumChunks & 0x1) {
159           __m128i product = _mm_maddubs_epi16(_mm_load_si128(&input_vector[kNumChunks-1]), _mm_load_si128(&row[kNumChunks-1]));
160           product = _mm_madd_epi16(product, kOnes);
161           sum = _mm_add_epi32(sum, product);
162         }
163         sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC
164         sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
165         output[i] = _mm_cvtsi128_si32(sum) + biases_[i];
166
167   #elif defined(USE_SSE2)
168         __m128i sum_lo = _mm_cvtsi32_si128(biases_[i]);
169         __m128i sum_hi = kZeros;
170         const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
171         for (IndexType j = 0; j < kNumChunks; ++j) {
172           __m128i row_j = _mm_load_si128(&row[j]);
173           __m128i input_j = _mm_load_si128(&input_vector[j]);
174           __m128i row_signs = _mm_cmpgt_epi8(kZeros, row_j);
175           __m128i extended_row_lo = _mm_unpacklo_epi8(row_j, row_signs);
176           __m128i extended_row_hi = _mm_unpackhi_epi8(row_j, row_signs);
177           __m128i extended_input_lo = _mm_unpacklo_epi8(input_j, kZeros);
178           __m128i extended_input_hi = _mm_unpackhi_epi8(input_j, kZeros);
179           __m128i product_lo = _mm_madd_epi16(extended_row_lo, extended_input_lo);
180           __m128i product_hi = _mm_madd_epi16(extended_row_hi, extended_input_hi);
181           sum_lo = _mm_add_epi32(sum_lo, product_lo);
182           sum_hi = _mm_add_epi32(sum_hi, product_hi);
183         }
184         __m128i sum = _mm_add_epi32(sum_lo, sum_hi);
185         __m128i sum_high_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2));
186         sum = _mm_add_epi32(sum, sum_high_64);
187         __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2));
188         sum = _mm_add_epi32(sum, sum_second_32);
189         output[i] = _mm_cvtsi128_si32(sum);
190
191   #elif defined(USE_MMX)
192         __m64 sum_lo = _mm_cvtsi32_si64(biases_[i]);
193         __m64 sum_hi = kZeros;
194         const auto row = reinterpret_cast<const __m64*>(&weights_[offset]);
195         for (IndexType j = 0; j < kNumChunks; ++j) {
196           __m64 row_j = row[j];
197           __m64 input_j = input_vector[j];
198           __m64 row_signs = _mm_cmpgt_pi8(kZeros, row_j);
199           __m64 extended_row_lo = _mm_unpacklo_pi8(row_j, row_signs);
200           __m64 extended_row_hi = _mm_unpackhi_pi8(row_j, row_signs);
201           __m64 extended_input_lo = _mm_unpacklo_pi8(input_j, kZeros);
202           __m64 extended_input_hi = _mm_unpackhi_pi8(input_j, kZeros);
203           __m64 product_lo = _mm_madd_pi16(extended_row_lo, extended_input_lo);
204           __m64 product_hi = _mm_madd_pi16(extended_row_hi, extended_input_hi);
205           sum_lo = _mm_add_pi32(sum_lo, product_lo);
206           sum_hi = _mm_add_pi32(sum_hi, product_hi);
207         }
208         __m64 sum = _mm_add_pi32(sum_lo, sum_hi);
209         sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum));
210         output[i] = _mm_cvtsi64_si32(sum);
211
212   #elif defined(USE_NEON)
213         int32x4_t sum = {biases_[i]};
214         const auto row = reinterpret_cast<const int8x8_t*>(&weights_[offset]);
215         for (IndexType j = 0; j < kNumChunks; ++j) {
216           int16x8_t product = vmull_s8(input_vector[j * 2], row[j * 2]);
217           product = vmlal_s8(product, input_vector[j * 2 + 1], row[j * 2 + 1]);
218           sum = vpadalq_s16(sum, product);
219         }
220         output[i] = sum[0] + sum[1] + sum[2] + sum[3];
221
222   #else
223         OutputType sum = biases_[i];
224         for (IndexType j = 0; j < kInputDimensions; ++j) {
225           sum += weights_[offset + j] * input[j];
226         }
227         output[i] = sum;
228   #endif
229
230       }
231   #if defined(USE_MMX)
232       _mm_empty();
233   #endif
234       return output;
235     }
236
237    private:
238     using BiasType = OutputType;
239     using WeightType = std::int8_t;
240
241     PreviousLayer previous_layer_;
242
243     alignas(kCacheLineSize) BiasType biases_[kOutputDimensions];
244     alignas(kCacheLineSize)
245         WeightType weights_[kOutputDimensions * kPaddedInputDimensions];
246   };
247
248 }  // namespace Eval::NNUE::Layers
249
250 #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED