Reformat code in little-endian patch
[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       for (std::size_t i = 0; i < kOutputDimensions; ++i)
66         biases_[i] = read_little_endian<BiasType>(stream);
67       for (std::size_t i = 0; i < kOutputDimensions * kPaddedInputDimensions; ++i)
68         weights_[i] = read_little_endian<WeightType>(stream);
69       return !stream.fail();
70     }
71
72     // Forward propagation
73     const OutputType* Propagate(
74         const TransformedFeatureType* transformed_features, char* buffer) const {
75       const auto input = previous_layer_.Propagate(
76           transformed_features, buffer + kSelfBufferSize);
77       const auto output = reinterpret_cast<OutputType*>(buffer);
78
79   #if defined(USE_AVX512)
80       constexpr IndexType kNumChunks = kPaddedInputDimensions / (kSimdWidth * 2);
81       const auto input_vector = reinterpret_cast<const __m512i*>(input);
82   #if !defined(USE_VNNI)
83       const __m512i kOnes = _mm512_set1_epi16(1);
84   #endif
85
86   #elif defined(USE_AVX2)
87       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
88       const __m256i kOnes = _mm256_set1_epi16(1);
89       const auto input_vector = reinterpret_cast<const __m256i*>(input);
90
91   #elif defined(USE_SSE2)
92       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
93   #ifndef USE_SSSE3
94       const __m128i kZeros = _mm_setzero_si128();
95   #else
96       const __m128i kOnes = _mm_set1_epi16(1);
97   #endif
98       const auto input_vector = reinterpret_cast<const __m128i*>(input);
99
100   #elif defined(USE_MMX)
101       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
102       const __m64 kZeros = _mm_setzero_si64();
103       const auto input_vector = reinterpret_cast<const __m64*>(input);
104
105   #elif defined(USE_NEON)
106       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
107       const auto input_vector = reinterpret_cast<const int8x8_t*>(input);
108   #endif
109
110       for (IndexType i = 0; i < kOutputDimensions; ++i) {
111         const IndexType offset = i * kPaddedInputDimensions;
112
113   #if defined(USE_AVX512)
114         __m512i sum = _mm512_setzero_si512();
115         const auto row = reinterpret_cast<const __m512i*>(&weights_[offset]);
116         for (IndexType j = 0; j < kNumChunks; ++j) {
117   #if defined(USE_VNNI)
118             sum = _mm512_dpbusd_epi32(sum, _mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
119   #else
120             __m512i product = _mm512_maddubs_epi16(_mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
121             product = _mm512_madd_epi16(product, kOnes);
122             sum = _mm512_add_epi32(sum, product);
123   #endif
124         }
125
126         // Note: Changing kMaxSimdWidth from 32 to 64 breaks loading existing networks.
127         // As a result kPaddedInputDimensions may not be an even multiple of 64(512bit)
128         // and we have to do one more 256bit chunk.
129         if (kPaddedInputDimensions != kNumChunks * kSimdWidth * 2)
130         {
131             const auto iv256  = reinterpret_cast<const __m256i*>(&input_vector[kNumChunks]);
132             const auto row256 = reinterpret_cast<const __m256i*>(&row[kNumChunks]);
133   #if defined(USE_VNNI)
134             __m256i product256 = _mm256_dpbusd_epi32(
135                 _mm512_castsi512_si256(sum), _mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
136             sum = _mm512_inserti32x8(sum, product256, 0);
137   #else
138             __m256i product256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
139             sum = _mm512_add_epi32(sum, _mm512_cvtepi16_epi32(product256));
140   #endif
141         }
142         output[i] = _mm512_reduce_add_epi32(sum) + biases_[i];
143
144   #elif defined(USE_AVX2)
145         __m256i sum = _mm256_setzero_si256();
146         const auto row = reinterpret_cast<const __m256i*>(&weights_[offset]);
147         for (IndexType j = 0; j < kNumChunks; ++j) {
148           __m256i product = _mm256_maddubs_epi16(_mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j]));
149           product = _mm256_madd_epi16(product, kOnes);
150           sum = _mm256_add_epi32(sum, product);
151         }
152         __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
153         sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));
154         sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB));
155         output[i] = _mm_cvtsi128_si32(sum128) + biases_[i];
156
157   #elif defined(USE_SSSE3)
158         __m128i sum = _mm_setzero_si128();
159         const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
160         for (int j = 0; j < (int)kNumChunks - 1; j += 2) {
161           __m128i product0 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j]), _mm_load_si128(&row[j]));
162           product0 = _mm_madd_epi16(product0, kOnes);
163           sum = _mm_add_epi32(sum, product0);
164           __m128i product1 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j+1]), _mm_load_si128(&row[j+1]));
165           product1 = _mm_madd_epi16(product1, kOnes);
166           sum = _mm_add_epi32(sum, product1);
167         }
168         if (kNumChunks & 0x1) {
169           __m128i product = _mm_maddubs_epi16(_mm_load_si128(&input_vector[kNumChunks-1]), _mm_load_si128(&row[kNumChunks-1]));
170           product = _mm_madd_epi16(product, kOnes);
171           sum = _mm_add_epi32(sum, product);
172         }
173         sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC
174         sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
175         output[i] = _mm_cvtsi128_si32(sum) + biases_[i];
176
177   #elif defined(USE_SSE2)
178         __m128i sum_lo = _mm_cvtsi32_si128(biases_[i]);
179         __m128i sum_hi = kZeros;
180         const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
181         for (IndexType j = 0; j < kNumChunks; ++j) {
182           __m128i row_j = _mm_load_si128(&row[j]);
183           __m128i input_j = _mm_load_si128(&input_vector[j]);
184           __m128i row_signs = _mm_cmpgt_epi8(kZeros, row_j);
185           __m128i extended_row_lo = _mm_unpacklo_epi8(row_j, row_signs);
186           __m128i extended_row_hi = _mm_unpackhi_epi8(row_j, row_signs);
187           __m128i extended_input_lo = _mm_unpacklo_epi8(input_j, kZeros);
188           __m128i extended_input_hi = _mm_unpackhi_epi8(input_j, kZeros);
189           __m128i product_lo = _mm_madd_epi16(extended_row_lo, extended_input_lo);
190           __m128i product_hi = _mm_madd_epi16(extended_row_hi, extended_input_hi);
191           sum_lo = _mm_add_epi32(sum_lo, product_lo);
192           sum_hi = _mm_add_epi32(sum_hi, product_hi);
193         }
194         __m128i sum = _mm_add_epi32(sum_lo, sum_hi);
195         __m128i sum_high_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2));
196         sum = _mm_add_epi32(sum, sum_high_64);
197         __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2));
198         sum = _mm_add_epi32(sum, sum_second_32);
199         output[i] = _mm_cvtsi128_si32(sum);
200
201   #elif defined(USE_MMX)
202         __m64 sum_lo = _mm_cvtsi32_si64(biases_[i]);
203         __m64 sum_hi = kZeros;
204         const auto row = reinterpret_cast<const __m64*>(&weights_[offset]);
205         for (IndexType j = 0; j < kNumChunks; ++j) {
206           __m64 row_j = row[j];
207           __m64 input_j = input_vector[j];
208           __m64 row_signs = _mm_cmpgt_pi8(kZeros, row_j);
209           __m64 extended_row_lo = _mm_unpacklo_pi8(row_j, row_signs);
210           __m64 extended_row_hi = _mm_unpackhi_pi8(row_j, row_signs);
211           __m64 extended_input_lo = _mm_unpacklo_pi8(input_j, kZeros);
212           __m64 extended_input_hi = _mm_unpackhi_pi8(input_j, kZeros);
213           __m64 product_lo = _mm_madd_pi16(extended_row_lo, extended_input_lo);
214           __m64 product_hi = _mm_madd_pi16(extended_row_hi, extended_input_hi);
215           sum_lo = _mm_add_pi32(sum_lo, product_lo);
216           sum_hi = _mm_add_pi32(sum_hi, product_hi);
217         }
218         __m64 sum = _mm_add_pi32(sum_lo, sum_hi);
219         sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum));
220         output[i] = _mm_cvtsi64_si32(sum);
221
222   #elif defined(USE_NEON)
223         int32x4_t sum = {biases_[i]};
224         const auto row = reinterpret_cast<const int8x8_t*>(&weights_[offset]);
225         for (IndexType j = 0; j < kNumChunks; ++j) {
226           int16x8_t product = vmull_s8(input_vector[j * 2], row[j * 2]);
227           product = vmlal_s8(product, input_vector[j * 2 + 1], row[j * 2 + 1]);
228           sum = vpadalq_s16(sum, product);
229         }
230         output[i] = sum[0] + sum[1] + sum[2] + sum[3];
231
232   #else
233         OutputType sum = biases_[i];
234         for (IndexType j = 0; j < kInputDimensions; ++j) {
235           sum += weights_[offset + j] * input[j];
236         }
237         output[i] = sum;
238   #endif
239
240       }
241   #if defined(USE_MMX)
242       _mm_empty();
243   #endif
244       return output;
245     }
246
247    private:
248     using BiasType = OutputType;
249     using WeightType = std::int8_t;
250
251     PreviousLayer previous_layer_;
252
253     alignas(kCacheLineSize) BiasType biases_[kOutputDimensions];
254     alignas(kCacheLineSize)
255         WeightType weights_[kOutputDimensions * kPaddedInputDimensions];
256   };
257
258 }  // namespace Eval::NNUE::Layers
259
260 #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED