Support VNNI on 256bit vectors
[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 auto input_vector = reinterpret_cast<const __m256i*>(input);
89   #if !defined(USE_VNNI)
90       const __m256i kOnes = _mm256_set1_epi16(1);
91   #endif
92
93   #elif defined(USE_SSE2)
94       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
95   #ifndef USE_SSSE3
96       const __m128i kZeros = _mm_setzero_si128();
97   #else
98       const __m128i kOnes = _mm_set1_epi16(1);
99   #endif
100       const auto input_vector = reinterpret_cast<const __m128i*>(input);
101
102   #elif defined(USE_MMX)
103       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
104       const __m64 kZeros = _mm_setzero_si64();
105       const auto input_vector = reinterpret_cast<const __m64*>(input);
106
107   #elif defined(USE_NEON)
108       constexpr IndexType kNumChunks = kPaddedInputDimensions / kSimdWidth;
109       const auto input_vector = reinterpret_cast<const int8x8_t*>(input);
110   #endif
111
112       for (IndexType i = 0; i < kOutputDimensions; ++i) {
113         const IndexType offset = i * kPaddedInputDimensions;
114
115   #if defined(USE_AVX512)
116         __m512i sum = _mm512_setzero_si512();
117         const auto row = reinterpret_cast<const __m512i*>(&weights_[offset]);
118         for (IndexType j = 0; j < kNumChunks; ++j) {
119   #if defined(USE_VNNI)
120             sum = _mm512_dpbusd_epi32(sum, _mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
121   #else
122             __m512i product = _mm512_maddubs_epi16(_mm512_loadA_si512(&input_vector[j]), _mm512_load_si512(&row[j]));
123             product = _mm512_madd_epi16(product, kOnes);
124             sum = _mm512_add_epi32(sum, product);
125   #endif
126         }
127
128         // Note: Changing kMaxSimdWidth from 32 to 64 breaks loading existing networks.
129         // As a result kPaddedInputDimensions may not be an even multiple of 64(512bit)
130         // and we have to do one more 256bit chunk.
131         if (kPaddedInputDimensions != kNumChunks * kSimdWidth * 2)
132         {
133             const auto iv256  = reinterpret_cast<const __m256i*>(&input_vector[kNumChunks]);
134             const auto row256 = reinterpret_cast<const __m256i*>(&row[kNumChunks]);
135   #if defined(USE_VNNI)
136             __m256i product256 = _mm256_dpbusd_epi32(
137                 _mm512_castsi512_si256(sum), _mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
138             sum = _mm512_inserti32x8(sum, product256, 0);
139   #else
140             __m256i product256 = _mm256_maddubs_epi16(_mm256_loadA_si256(&iv256[0]), _mm256_load_si256(&row256[0]));
141             sum = _mm512_add_epi32(sum, _mm512_cvtepi16_epi32(product256));
142   #endif
143         }
144         output[i] = _mm512_reduce_add_epi32(sum) + biases_[i];
145
146   #elif defined(USE_AVX2)
147         __m256i sum = _mm256_setzero_si256();
148         const auto row = reinterpret_cast<const __m256i*>(&weights_[offset]);
149         for (IndexType j = 0; j < kNumChunks; ++j) {
150   #if defined(USE_VNNI)
151           sum = _mm256_dpbusd_epi32(sum, _mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j]));
152   #else
153           __m256i product = _mm256_maddubs_epi16(_mm256_loadA_si256(&input_vector[j]), _mm256_load_si256(&row[j]));
154           product = _mm256_madd_epi16(product, kOnes);
155           sum = _mm256_add_epi32(sum, product);
156   #endif
157         }
158         __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
159         sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));
160         sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB));
161         output[i] = _mm_cvtsi128_si32(sum128) + biases_[i];
162
163   #elif defined(USE_SSSE3)
164         __m128i sum = _mm_setzero_si128();
165         const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
166         for (int j = 0; j < (int)kNumChunks - 1; j += 2) {
167           __m128i product0 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j]), _mm_load_si128(&row[j]));
168           product0 = _mm_madd_epi16(product0, kOnes);
169           sum = _mm_add_epi32(sum, product0);
170           __m128i product1 = _mm_maddubs_epi16(_mm_load_si128(&input_vector[j+1]), _mm_load_si128(&row[j+1]));
171           product1 = _mm_madd_epi16(product1, kOnes);
172           sum = _mm_add_epi32(sum, product1);
173         }
174         if (kNumChunks & 0x1) {
175           __m128i product = _mm_maddubs_epi16(_mm_load_si128(&input_vector[kNumChunks-1]), _mm_load_si128(&row[kNumChunks-1]));
176           product = _mm_madd_epi16(product, kOnes);
177           sum = _mm_add_epi32(sum, product);
178         }
179         sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC
180         sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
181         output[i] = _mm_cvtsi128_si32(sum) + biases_[i];
182
183   #elif defined(USE_SSE2)
184         __m128i sum_lo = _mm_cvtsi32_si128(biases_[i]);
185         __m128i sum_hi = kZeros;
186         const auto row = reinterpret_cast<const __m128i*>(&weights_[offset]);
187         for (IndexType j = 0; j < kNumChunks; ++j) {
188           __m128i row_j = _mm_load_si128(&row[j]);
189           __m128i input_j = _mm_load_si128(&input_vector[j]);
190           __m128i row_signs = _mm_cmpgt_epi8(kZeros, row_j);
191           __m128i extended_row_lo = _mm_unpacklo_epi8(row_j, row_signs);
192           __m128i extended_row_hi = _mm_unpackhi_epi8(row_j, row_signs);
193           __m128i extended_input_lo = _mm_unpacklo_epi8(input_j, kZeros);
194           __m128i extended_input_hi = _mm_unpackhi_epi8(input_j, kZeros);
195           __m128i product_lo = _mm_madd_epi16(extended_row_lo, extended_input_lo);
196           __m128i product_hi = _mm_madd_epi16(extended_row_hi, extended_input_hi);
197           sum_lo = _mm_add_epi32(sum_lo, product_lo);
198           sum_hi = _mm_add_epi32(sum_hi, product_hi);
199         }
200         __m128i sum = _mm_add_epi32(sum_lo, sum_hi);
201         __m128i sum_high_64 = _mm_shuffle_epi32(sum, _MM_SHUFFLE(1, 0, 3, 2));
202         sum = _mm_add_epi32(sum, sum_high_64);
203         __m128i sum_second_32 = _mm_shufflelo_epi16(sum, _MM_SHUFFLE(1, 0, 3, 2));
204         sum = _mm_add_epi32(sum, sum_second_32);
205         output[i] = _mm_cvtsi128_si32(sum);
206
207   #elif defined(USE_MMX)
208         __m64 sum_lo = _mm_cvtsi32_si64(biases_[i]);
209         __m64 sum_hi = kZeros;
210         const auto row = reinterpret_cast<const __m64*>(&weights_[offset]);
211         for (IndexType j = 0; j < kNumChunks; ++j) {
212           __m64 row_j = row[j];
213           __m64 input_j = input_vector[j];
214           __m64 row_signs = _mm_cmpgt_pi8(kZeros, row_j);
215           __m64 extended_row_lo = _mm_unpacklo_pi8(row_j, row_signs);
216           __m64 extended_row_hi = _mm_unpackhi_pi8(row_j, row_signs);
217           __m64 extended_input_lo = _mm_unpacklo_pi8(input_j, kZeros);
218           __m64 extended_input_hi = _mm_unpackhi_pi8(input_j, kZeros);
219           __m64 product_lo = _mm_madd_pi16(extended_row_lo, extended_input_lo);
220           __m64 product_hi = _mm_madd_pi16(extended_row_hi, extended_input_hi);
221           sum_lo = _mm_add_pi32(sum_lo, product_lo);
222           sum_hi = _mm_add_pi32(sum_hi, product_hi);
223         }
224         __m64 sum = _mm_add_pi32(sum_lo, sum_hi);
225         sum = _mm_add_pi32(sum, _mm_unpackhi_pi32(sum, sum));
226         output[i] = _mm_cvtsi64_si32(sum);
227
228   #elif defined(USE_NEON)
229         int32x4_t sum = {biases_[i]};
230         const auto row = reinterpret_cast<const int8x8_t*>(&weights_[offset]);
231         for (IndexType j = 0; j < kNumChunks; ++j) {
232           int16x8_t product = vmull_s8(input_vector[j * 2], row[j * 2]);
233           product = vmlal_s8(product, input_vector[j * 2 + 1], row[j * 2 + 1]);
234           sum = vpadalq_s16(sum, product);
235         }
236         output[i] = sum[0] + sum[1] + sum[2] + sum[3];
237
238   #else
239         OutputType sum = biases_[i];
240         for (IndexType j = 0; j < kInputDimensions; ++j) {
241           sum += weights_[offset + j] * input[j];
242         }
243         output[i] = sum;
244   #endif
245
246       }
247   #if defined(USE_MMX)
248       _mm_empty();
249   #endif
250       return output;
251     }
252
253    private:
254     using BiasType = OutputType;
255     using WeightType = std::int8_t;
256
257     PreviousLayer previous_layer_;
258
259     alignas(kCacheLineSize) BiasType biases_[kOutputDimensions];
260     alignas(kCacheLineSize)
261         WeightType weights_[kOutputDimensions * kPaddedInputDimensions];
262   };
263
264 }  // namespace Eval::NNUE::Layers
265
266 #endif // #ifndef NNUE_LAYERS_AFFINE_TRANSFORM_H_INCLUDED