Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "arm_compute/core/Helpers.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
| 30 | #include "arm_compute/core/TensorInfo.h" |
| 31 | #include "arm_compute/core/Types.h" |
| 32 | #include "arm_compute/core/Utils.h" |
| 33 | #include "arm_compute/core/Validate.h" |
| 34 | #include "arm_compute/core/Window.h" |
| 35 | |
| 36 | #include <arm_neon.h> |
| 37 | #include <cstddef> |
| 38 | #include <cstdint> |
| 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
| 42 | namespace arm_compute |
| 43 | { |
| 44 | class Coordinates; |
| 45 | } // namespace arm_compute |
| 46 | |
| 47 | NEGEMMLowpOffsetContributionKernel::NEGEMMLowpOffsetContributionKernel() |
| 48 | : _vector_sum_col(nullptr), _vector_sum_row(nullptr), _mm_result(nullptr), _a_offset(0), _b_offset(0), _k_offset(0), _slide_vector_sum_col(true) |
| 49 | { |
| 50 | } |
| 51 | |
| 52 | void NEGEMMLowpOffsetContributionKernel::configure(ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, int32_t k, int32_t a_offset, int32_t b_offset) |
| 53 | { |
| 54 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32); |
| 55 | |
| 56 | // If a_offset == 0, vector_sum_col can be a nullptr |
| 57 | if(a_offset != 0) |
| 58 | { |
| 59 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32); |
| 60 | ARM_COMPUTE_ERROR_ON(vector_sum_col->info()->dimension(0) != mm_result->info()->dimension(0)); |
| 61 | |
| 62 | TensorShape vector_sum_col_shape = vector_sum_col->info()->tensor_shape(); |
| 63 | vector_sum_col_shape.collapse(1); |
| 64 | |
| 65 | // Check if vector_sum_col_shape should be slidden or not |
| 66 | // Don't slide vector_sum_col_shape along the y dimension if vector_sum_col_shape has just 1 dimension and vector_sum_row_shape more than 1 |
| 67 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
| 68 | _slide_vector_sum_col = vector_sum_col_shape[1] != 1; |
| 69 | } |
| 70 | |
| 71 | // If b_offset == 0, vector_sum_row can be a nullptr |
| 72 | if(b_offset != 0) |
| 73 | { |
| 74 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32); |
| 75 | ARM_COMPUTE_ERROR_ON(vector_sum_row->info()->dimension(0) != mm_result->info()->dimension(1)); |
| 76 | |
| 77 | TensorShape output_shape = mm_result->info()->tensor_shape(); |
| 78 | TensorShape vector_sum_row_shape = vector_sum_row->info()->tensor_shape(); |
| 79 | vector_sum_row_shape.collapse(1); |
| 80 | output_shape.collapse(2); |
| 81 | |
| 82 | ARM_COMPUTE_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[2], "mm_result tensor must have the same number of batches of output tensor"); |
| 83 | |
| 84 | if(a_offset != 0) |
| 85 | { |
| 86 | TensorShape vector_sum_col_shape = vector_sum_col->info()->tensor_shape(); |
| 87 | vector_sum_col_shape.collapse(1); |
| 88 | |
| 89 | ARM_COMPUTE_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 |
| 90 | && vector_sum_col_shape[1] != vector_sum_row_shape[1], |
| 91 | "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1"); |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | _vector_sum_col = vector_sum_col; |
| 96 | _vector_sum_row = vector_sum_row; |
| 97 | _mm_result = mm_result; |
| 98 | _a_offset = a_offset; |
| 99 | _b_offset = b_offset; |
| 100 | _k_offset = a_offset * b_offset * k; |
| 101 | |
| 102 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 103 | |
| 104 | // Configure kernel window |
| 105 | Window win = calculate_max_window(*mm_result->info(), Steps(num_elems_processed_per_iteration)); |
| 106 | |
| 107 | AccessWindowHorizontal mm_result_access(mm_result->info(), 0, num_elems_processed_per_iteration); |
| 108 | |
| 109 | // Accordingly with a_offset and b_offset, we can have 4 cases: |
| 110 | // a_offset != 0 && b_offset != 0 |
| 111 | // a_offset = 0 && b_offset != 0 |
| 112 | // a_offset != 0 && b_offset = 0 |
| 113 | // a_offset = 0 && b_offset = 0 |
| 114 | if(a_offset != 0 && b_offset != 0) |
| 115 | { |
| 116 | AccessWindowStatic vector_sum_row_access(vector_sum_row->info(), 0, 0, vector_sum_row->info()->dimension(0), 0); |
| 117 | AccessWindowHorizontal vector_sum_col_access(vector_sum_col->info(), 0, num_elems_processed_per_iteration); |
| 118 | |
| 119 | update_window_and_padding(win, |
| 120 | vector_sum_col_access, |
| 121 | vector_sum_row_access, |
| 122 | mm_result_access); |
| 123 | } |
| 124 | else if(a_offset == 0 && b_offset != 0) |
| 125 | { |
| 126 | AccessWindowStatic vector_sum_row_access(vector_sum_row->info(), 0, 0, vector_sum_row->info()->dimension(0), 0); |
| 127 | |
| 128 | update_window_and_padding(win, |
| 129 | vector_sum_row_access, |
| 130 | mm_result_access); |
| 131 | } |
| 132 | else if(a_offset != 0 && b_offset == 0) |
| 133 | { |
| 134 | AccessWindowHorizontal vector_sum_col_access(vector_sum_col->info(), 0, num_elems_processed_per_iteration); |
| 135 | |
| 136 | update_window_and_padding(win, |
| 137 | vector_sum_col_access, |
| 138 | mm_result_access); |
| 139 | } |
| 140 | else |
| 141 | { |
| 142 | update_window_and_padding(win, |
| 143 | mm_result_access); |
| 144 | } |
| 145 | |
| 146 | INEKernel::configure(win); |
| 147 | } |
| 148 | |
| 149 | void NEGEMMLowpOffsetContributionKernel::run(const Window &window, const ThreadInfo &info) |
| 150 | { |
| 151 | ARM_COMPUTE_UNUSED(info); |
| 152 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 153 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 154 | |
| 155 | Window collapsed_window = window.collapse_if_possible(IKernel::window(), Window::DimZ); |
| 156 | |
| 157 | if(_a_offset != 0 && _b_offset != 0) // true, true |
| 158 | { |
| 159 | // Set window for vector_sum_col |
| 160 | Window win_vector_sum_col(collapsed_window); |
| 161 | win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 162 | if(!_slide_vector_sum_col) |
| 163 | { |
| 164 | win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 165 | } |
| 166 | |
| 167 | // Set window for vector_sum_row |
| 168 | Window win_vector_sum_row(collapsed_window); |
| 169 | win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 170 | win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 171 | |
| 172 | Iterator vector_sum_col(_vector_sum_col, win_vector_sum_col); |
| 173 | Iterator vector_sum_row(_vector_sum_row, win_vector_sum_row); |
| 174 | Iterator mm_result(_mm_result, window); |
| 175 | |
| 176 | execute_window_loop(window, [&](const Coordinates & id) |
| 177 | { |
| 178 | // Compute the leftover term due to a_offset. |
| 179 | int32x4x4_t a_offset_term_s32 = |
| 180 | { |
| 181 | { |
| 182 | vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 0), |
| 183 | vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 4), |
| 184 | vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 8), |
| 185 | vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 12) |
| 186 | } |
| 187 | }; |
| 188 | |
| 189 | a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], _a_offset); |
| 190 | a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], _a_offset); |
| 191 | a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], _a_offset); |
| 192 | a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], _a_offset); |
| 193 | |
| 194 | // Compute the leftover term due to b_offset. |
| 195 | int32x4_t b_offset_term_s32 = vld1q_dup_s32(reinterpret_cast<const int32_t *>(vector_sum_row.ptr()) + id.y()); |
| 196 | b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, _b_offset); |
| 197 | |
| 198 | // Add a_offset_term_s32 and b_offset_term_s32 |
| 199 | int32x4x4_t offset_term_s32 = |
| 200 | { |
| 201 | { |
| 202 | vdupq_n_s32(_k_offset), |
| 203 | vdupq_n_s32(_k_offset), |
| 204 | vdupq_n_s32(_k_offset), |
| 205 | vdupq_n_s32(_k_offset) |
| 206 | } |
| 207 | }; |
| 208 | |
| 209 | offset_term_s32.val[0] = vaddq_s32(offset_term_s32.val[0], vaddq_s32(a_offset_term_s32.val[0], b_offset_term_s32)); |
| 210 | offset_term_s32.val[1] = vaddq_s32(offset_term_s32.val[1], vaddq_s32(a_offset_term_s32.val[1], b_offset_term_s32)); |
| 211 | offset_term_s32.val[2] = vaddq_s32(offset_term_s32.val[2], vaddq_s32(a_offset_term_s32.val[2], b_offset_term_s32)); |
| 212 | offset_term_s32.val[3] = vaddq_s32(offset_term_s32.val[3], vaddq_s32(a_offset_term_s32.val[3], b_offset_term_s32)); |
| 213 | |
| 214 | int32x4x4_t in_s32 = |
| 215 | { |
| 216 | { |
| 217 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 0), |
| 218 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 4), |
| 219 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 8), |
| 220 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 12) |
| 221 | } |
| 222 | }; |
| 223 | |
| 224 | // Add the offset terms to GEMM's result |
| 225 | in_s32.val[0] = vaddq_s32(in_s32.val[0], offset_term_s32.val[0]); |
| 226 | in_s32.val[1] = vaddq_s32(in_s32.val[1], offset_term_s32.val[1]); |
| 227 | in_s32.val[2] = vaddq_s32(in_s32.val[2], offset_term_s32.val[2]); |
| 228 | in_s32.val[3] = vaddq_s32(in_s32.val[3], offset_term_s32.val[3]); |
| 229 | |
| 230 | // Store the result with the offset contribution |
| 231 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 0, in_s32.val[0]); |
| 232 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 4, in_s32.val[1]); |
| 233 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 8, in_s32.val[2]); |
| 234 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 12, in_s32.val[3]); |
| 235 | }, |
| 236 | vector_sum_col, vector_sum_row, mm_result); |
| 237 | } |
| 238 | else if((_a_offset == 0) && (_b_offset != 0)) // false, true |
| 239 | { |
| 240 | // Set window for vector_sum_row |
| 241 | Window win_vector_sum_row(collapsed_window); |
| 242 | win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 243 | win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 244 | |
| 245 | Iterator vector_sum_row(_vector_sum_row, win_vector_sum_row); |
| 246 | Iterator mm_result(_mm_result, window); |
| 247 | |
| 248 | execute_window_loop(window, [&](const Coordinates & id) |
| 249 | { |
| 250 | // Compute the leftover term due to b_offset. |
| 251 | int32x4_t b_offset_term_s32 = vld1q_dup_s32(reinterpret_cast<const int32_t *>(vector_sum_row.ptr()) + id.y()); |
| 252 | b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, _b_offset); |
| 253 | |
| 254 | int32x4x4_t in_s32 = |
| 255 | { |
| 256 | { |
| 257 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 0), |
| 258 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 4), |
| 259 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 8), |
| 260 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 12) |
| 261 | } |
| 262 | }; |
| 263 | |
| 264 | // Add the offset terms to GEMM's result |
| 265 | in_s32.val[0] = vaddq_s32(in_s32.val[0], b_offset_term_s32); |
| 266 | in_s32.val[1] = vaddq_s32(in_s32.val[1], b_offset_term_s32); |
| 267 | in_s32.val[2] = vaddq_s32(in_s32.val[2], b_offset_term_s32); |
| 268 | in_s32.val[3] = vaddq_s32(in_s32.val[3], b_offset_term_s32); |
| 269 | |
| 270 | // Store the result with the offset contribution |
| 271 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 0, in_s32.val[0]); |
| 272 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 4, in_s32.val[1]); |
| 273 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 8, in_s32.val[2]); |
| 274 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 12, in_s32.val[3]); |
| 275 | }, |
| 276 | vector_sum_row, mm_result); |
| 277 | } |
| 278 | else if((_a_offset != 0) && (_b_offset == 0)) // true, false |
| 279 | { |
| 280 | // Set window for vector_sum_col |
| 281 | Window win_vector_sum_col(collapsed_window); |
| 282 | win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 283 | if(!_slide_vector_sum_col) |
| 284 | { |
| 285 | win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 286 | } |
| 287 | |
| 288 | Iterator vector_sum_col(_vector_sum_col, win_vector_sum_col); |
| 289 | Iterator mm_result(_mm_result, window); |
| 290 | |
| 291 | execute_window_loop(window, [&](const Coordinates & id) |
| 292 | { |
| 293 | // Compute the leftover term due to a_offset. |
| 294 | int32x4x4_t a_offset_term_s32 = |
| 295 | { |
| 296 | { |
| 297 | vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 0), |
| 298 | vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 4), |
| 299 | vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 8), |
| 300 | vld1q_s32(reinterpret_cast<const int32_t *>(vector_sum_col.ptr()) + 12) |
| 301 | } |
| 302 | }; |
| 303 | |
| 304 | a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], _a_offset); |
| 305 | a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], _a_offset); |
| 306 | a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], _a_offset); |
| 307 | a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], _a_offset); |
| 308 | |
| 309 | int32x4x4_t in_s32 = |
| 310 | { |
| 311 | { |
| 312 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 0), |
| 313 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 4), |
| 314 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 8), |
| 315 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result.ptr()) + 12) |
| 316 | } |
| 317 | }; |
| 318 | |
| 319 | // Add the offset terms to GEMM's result |
| 320 | in_s32.val[0] = vaddq_s32(in_s32.val[0], a_offset_term_s32.val[0]); |
| 321 | in_s32.val[1] = vaddq_s32(in_s32.val[1], a_offset_term_s32.val[1]); |
| 322 | in_s32.val[2] = vaddq_s32(in_s32.val[2], a_offset_term_s32.val[2]); |
| 323 | in_s32.val[3] = vaddq_s32(in_s32.val[3], a_offset_term_s32.val[3]); |
| 324 | |
| 325 | // Store the result with the offset contribution |
| 326 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 0, in_s32.val[0]); |
| 327 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 4, in_s32.val[1]); |
| 328 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 8, in_s32.val[2]); |
| 329 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result.ptr()) + 12, in_s32.val[3]); |
| 330 | }, |
| 331 | vector_sum_col, mm_result); |
| 332 | } |
| 333 | else // false, false |
| 334 | { |
| 335 | // No offset contribution from matrix A and matrix B |
| 336 | return; |
| 337 | } |
| 338 | } |