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