Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 1 | /* |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2019 ARM Limited. |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 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 | |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 47 | namespace |
| 48 | { |
Georgios Pinitas | 631c41a | 2017-12-06 11:53:03 +0000 | [diff] [blame] | 49 | Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, |
| 50 | int32_t a_offset, int32_t b_offset) |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 51 | { |
| 52 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32); |
| 53 | |
| 54 | // If a_offset == 0, vector_sum_col can be a nullptr |
| 55 | if(a_offset != 0) |
| 56 | { |
| 57 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32); |
| 58 | ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != mm_result->dimension(0)); |
| 59 | } |
| 60 | |
| 61 | // If b_offset == 0, vector_sum_row can be a nullptr |
| 62 | if(b_offset != 0) |
| 63 | { |
| 64 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32); |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 65 | |
| 66 | // Check if input is a 3D reinterpretation |
| 67 | const bool reinterpret_as_3d = mm_result->num_dimensions() > 1 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x(); |
| 68 | |
| 69 | // Validate input |
| 70 | ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (mm_result->dimension(1) * mm_result->dimension(2))); |
| 71 | ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != mm_result->dimension(1)); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 72 | |
Georgios Pinitas | 4062680 | 2017-12-08 19:02:45 +0000 | [diff] [blame] | 73 | TensorShape output_shape = mm_result->tensor_shape(); |
| 74 | if(output_shape.num_dimensions() > 1) |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 75 | { |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 76 | const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2; |
| 77 | |
Georgios Pinitas | 4062680 | 2017-12-08 19:02:45 +0000 | [diff] [blame] | 78 | TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape(); |
| 79 | vector_sum_row_shape.collapse_from(1); |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 80 | output_shape.collapse_from(output_batch_idx); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 81 | |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 82 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[output_batch_idx], |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 83 | "mm_result tensor must have the same number of batches of output tensor"); |
Georgios Pinitas | 4062680 | 2017-12-08 19:02:45 +0000 | [diff] [blame] | 84 | |
| 85 | if(a_offset != 0) |
| 86 | { |
| 87 | TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape(); |
| 88 | vector_sum_col_shape.collapse_from(1); |
| 89 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 90 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && 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"); |
Georgios Pinitas | 4062680 | 2017-12-08 19:02:45 +0000 | [diff] [blame] | 92 | } |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 93 | } |
| 94 | } |
| 95 | |
Georgios Pinitas | 631c41a | 2017-12-06 11:53:03 +0000 | [diff] [blame] | 96 | return Status{}; |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 97 | } |
| 98 | |
Georgios Pinitas | 631c41a | 2017-12-06 11:53:03 +0000 | [diff] [blame] | 99 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *mm_result, ITensorInfo *vector_sum_col, ITensorInfo *vector_sum_row, |
| 100 | int32_t a_offset, int32_t b_offset) |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 101 | { |
| 102 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 103 | bool window_changed = false; |
| 104 | |
| 105 | // Configure kernel window |
| 106 | Window win = calculate_max_window(*mm_result, Steps(num_elems_processed_per_iteration)); |
| 107 | |
| 108 | AccessWindowHorizontal mm_result_access(mm_result, 0, num_elems_processed_per_iteration); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 109 | window_changed = window_changed || update_window_and_padding(win, mm_result_access); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 110 | |
| 111 | if(a_offset != 0) |
| 112 | { |
| 113 | AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 114 | window_changed = window_changed || update_window_and_padding(win, vector_sum_col_access); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 115 | } |
| 116 | if(b_offset != 0) |
| 117 | { |
| 118 | AccessWindowStatic vector_sum_row_access(vector_sum_row, 0, 0, vector_sum_row->dimension(0), 0); // NOLINT |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 119 | window_changed = window_changed || update_window_and_padding(win, vector_sum_row_access); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 120 | } |
| 121 | |
Georgios Pinitas | 631c41a | 2017-12-06 11:53:03 +0000 | [diff] [blame] | 122 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 123 | return std::make_pair(err, win); |
| 124 | } |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 125 | |
| 126 | template <bool is_gemm3d> |
| 127 | void run_offset_contribution(const Window &window, |
| 128 | ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, |
| 129 | int32_t a_offset, int32_t b_offset, int32_t k_offset, bool slide_vector_sum_col) |
| 130 | { |
| 131 | Window collapsed_window = window.collapse_if_possible(window, Window::DimZ); |
| 132 | |
| 133 | const int height_input = is_gemm3d ? mm_result->info()->dimension(1) : 0; |
| 134 | const int depth_input = is_gemm3d ? mm_result->info()->dimension(2) : 1; |
| 135 | |
| 136 | if((a_offset != 0) && (b_offset != 0) && (vector_sum_col != nullptr) && (vector_sum_row != nullptr)) // true, true |
| 137 | { |
| 138 | // Set window for vector_sum_col |
| 139 | Window win_vector_sum_col(collapsed_window); |
| 140 | win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 141 | win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 142 | |
| 143 | // Set window for vector_sum_row |
| 144 | Window win_vector_sum_row(collapsed_window); |
| 145 | win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 146 | win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 147 | win_vector_sum_row.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 148 | |
| 149 | Iterator vector_sum_col_it(vector_sum_col, win_vector_sum_col); |
| 150 | Iterator vector_sum_row_it(vector_sum_row, win_vector_sum_row); |
| 151 | Iterator mm_result_it(mm_result, window); |
| 152 | |
| 153 | const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y(); |
| 154 | |
| 155 | // Offset in case vector_sum_col is batched |
| 156 | const int vector_sum_col_batch_offset = slide_vector_sum_col ? vector_sum_col->info()->strides_in_bytes().z() : 0; |
| 157 | |
| 158 | execute_window_loop(collapsed_window, [&](const Coordinates & id) |
| 159 | { |
| 160 | const int batch_id = id.z() / depth_input; |
| 161 | const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_batch_offset); |
| 162 | |
| 163 | // Compute the leftover term due to a_offset. |
| 164 | int32x4x4_t a_offset_term_s32 = |
| 165 | { |
| 166 | { |
| 167 | vld1q_s32(vector_sum_col_ptr + 0), |
| 168 | vld1q_s32(vector_sum_col_ptr + 4), |
| 169 | vld1q_s32(vector_sum_col_ptr + 8), |
| 170 | vld1q_s32(vector_sum_col_ptr + 12) |
| 171 | } |
| 172 | }; |
| 173 | |
| 174 | a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], a_offset); |
| 175 | a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], a_offset); |
| 176 | a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], a_offset); |
| 177 | a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], a_offset); |
| 178 | |
| 179 | // Compute the leftover term due to b_offset. |
| 180 | int32x4_t b_offset_term_s32 = vld1q_dup_s32(reinterpret_cast<const int32_t *>(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + id.y() |
| 181 | + (id.z() % depth_input) * height_input); |
| 182 | b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, b_offset); |
| 183 | |
| 184 | // Add a_offset_term_s32 and b_offset_term_s32 |
| 185 | int32x4x4_t offset_term_s32 = |
| 186 | { |
| 187 | { |
| 188 | vdupq_n_s32(k_offset), |
| 189 | vdupq_n_s32(k_offset), |
| 190 | vdupq_n_s32(k_offset), |
| 191 | vdupq_n_s32(k_offset) |
| 192 | } |
| 193 | }; |
| 194 | |
| 195 | offset_term_s32.val[0] = vaddq_s32(offset_term_s32.val[0], vaddq_s32(a_offset_term_s32.val[0], b_offset_term_s32)); |
| 196 | offset_term_s32.val[1] = vaddq_s32(offset_term_s32.val[1], vaddq_s32(a_offset_term_s32.val[1], b_offset_term_s32)); |
| 197 | offset_term_s32.val[2] = vaddq_s32(offset_term_s32.val[2], vaddq_s32(a_offset_term_s32.val[2], b_offset_term_s32)); |
| 198 | offset_term_s32.val[3] = vaddq_s32(offset_term_s32.val[3], vaddq_s32(a_offset_term_s32.val[3], b_offset_term_s32)); |
| 199 | |
| 200 | int32x4x4_t in_s32 = |
| 201 | { |
| 202 | { |
| 203 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 0), |
| 204 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 4), |
| 205 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 8), |
| 206 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 12) |
| 207 | } |
| 208 | }; |
| 209 | |
| 210 | // Add the offset terms to GEMM's result |
| 211 | in_s32.val[0] = vaddq_s32(in_s32.val[0], offset_term_s32.val[0]); |
| 212 | in_s32.val[1] = vaddq_s32(in_s32.val[1], offset_term_s32.val[1]); |
| 213 | in_s32.val[2] = vaddq_s32(in_s32.val[2], offset_term_s32.val[2]); |
| 214 | in_s32.val[3] = vaddq_s32(in_s32.val[3], offset_term_s32.val[3]); |
| 215 | |
| 216 | // Store the result with the offset contribution |
| 217 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 0, in_s32.val[0]); |
| 218 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 4, in_s32.val[1]); |
| 219 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 8, in_s32.val[2]); |
| 220 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 12, in_s32.val[3]); |
| 221 | }, |
| 222 | vector_sum_col_it, vector_sum_row_it, mm_result_it); |
| 223 | } |
| 224 | else if((a_offset == 0) && (b_offset != 0) && (vector_sum_row != nullptr)) // false, true |
| 225 | { |
| 226 | ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_row); |
| 227 | |
| 228 | // Set window for vector_sum_row |
| 229 | Window win_vector_sum_row(collapsed_window); |
| 230 | win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 231 | win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 232 | win_vector_sum_row.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 233 | |
| 234 | Iterator vector_sum_row_it(vector_sum_row, win_vector_sum_row); |
| 235 | Iterator mm_result_it(mm_result, window); |
| 236 | |
| 237 | const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y(); |
| 238 | |
| 239 | execute_window_loop(window, [&](const Coordinates & id) |
| 240 | { |
| 241 | const int batch_id = id.z() / depth_input; |
| 242 | |
| 243 | // Compute the leftover term due to b_offset. |
| 244 | int32x4_t b_offset_term_s32 = vld1q_dup_s32(reinterpret_cast<const int32_t *>(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + id.y() |
| 245 | + (id.z() % depth_input) * height_input); |
| 246 | b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, b_offset); |
| 247 | |
| 248 | int32x4x4_t in_s32 = |
| 249 | { |
| 250 | { |
| 251 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 0), |
| 252 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 4), |
| 253 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 8), |
| 254 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 12) |
| 255 | } |
| 256 | }; |
| 257 | |
| 258 | // Add the offset terms to GEMM's result |
| 259 | in_s32.val[0] = vaddq_s32(in_s32.val[0], b_offset_term_s32); |
| 260 | in_s32.val[1] = vaddq_s32(in_s32.val[1], b_offset_term_s32); |
| 261 | in_s32.val[2] = vaddq_s32(in_s32.val[2], b_offset_term_s32); |
| 262 | in_s32.val[3] = vaddq_s32(in_s32.val[3], b_offset_term_s32); |
| 263 | |
| 264 | // Store the result with the offset contribution |
| 265 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 0, in_s32.val[0]); |
| 266 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 4, in_s32.val[1]); |
| 267 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 8, in_s32.val[2]); |
| 268 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 12, in_s32.val[3]); |
| 269 | }, |
| 270 | vector_sum_row_it, mm_result_it); |
| 271 | } |
| 272 | else if((a_offset != 0) && (b_offset == 0) && (vector_sum_col != nullptr)) // true, false |
| 273 | { |
| 274 | // Set window for vector_sum_col |
| 275 | Window win_vector_sum_col(collapsed_window); |
| 276 | win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 277 | win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 278 | |
| 279 | Iterator vector_sum_col_it(vector_sum_col, win_vector_sum_col); |
| 280 | Iterator mm_result_it(mm_result, window); |
| 281 | |
| 282 | // Offset in case vector_sum_col is batched |
| 283 | const int vector_sum_col_batch_offset = slide_vector_sum_col ? vector_sum_col->info()->strides_in_bytes().z() : 0; |
| 284 | |
| 285 | execute_window_loop(window, [&](const Coordinates & id) |
| 286 | { |
| 287 | const int batch_id = id.z() / depth_input; |
| 288 | const auto vector_sum_col_ptr = reinterpret_cast<const int32_t *>(vector_sum_col_it.ptr() + batch_id * vector_sum_col_batch_offset); |
| 289 | |
| 290 | // Compute the leftover term due to a_offset. |
| 291 | int32x4x4_t a_offset_term_s32 = |
| 292 | { |
| 293 | { |
| 294 | vld1q_s32(vector_sum_col_ptr + 0), |
| 295 | vld1q_s32(vector_sum_col_ptr + 4), |
| 296 | vld1q_s32(vector_sum_col_ptr + 8), |
| 297 | vld1q_s32(vector_sum_col_ptr + 12) |
| 298 | } |
| 299 | }; |
| 300 | |
| 301 | a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], a_offset); |
| 302 | a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], a_offset); |
| 303 | a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], a_offset); |
| 304 | a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], a_offset); |
| 305 | |
| 306 | int32x4x4_t in_s32 = |
| 307 | { |
| 308 | { |
| 309 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 0), |
| 310 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 4), |
| 311 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 8), |
| 312 | vld1q_s32(reinterpret_cast<const int32_t *>(mm_result_it.ptr()) + 12) |
| 313 | } |
| 314 | }; |
| 315 | |
| 316 | // Add the offset terms to GEMM's result |
| 317 | in_s32.val[0] = vaddq_s32(in_s32.val[0], a_offset_term_s32.val[0]); |
| 318 | in_s32.val[1] = vaddq_s32(in_s32.val[1], a_offset_term_s32.val[1]); |
| 319 | in_s32.val[2] = vaddq_s32(in_s32.val[2], a_offset_term_s32.val[2]); |
| 320 | in_s32.val[3] = vaddq_s32(in_s32.val[3], a_offset_term_s32.val[3]); |
| 321 | |
| 322 | // Store the result with the offset contribution |
| 323 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 0, in_s32.val[0]); |
| 324 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 4, in_s32.val[1]); |
| 325 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 8, in_s32.val[2]); |
| 326 | vst1q_s32(reinterpret_cast<int32_t *>(mm_result_it.ptr()) + 12, in_s32.val[3]); |
| 327 | }, |
| 328 | vector_sum_col_it, mm_result_it); |
| 329 | } |
| 330 | else // false, false |
| 331 | { |
| 332 | // No offset contribution from matrix A and matrix B |
| 333 | return; |
| 334 | } |
| 335 | } |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 336 | } // namespace |
| 337 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 338 | NEGEMMLowpOffsetContributionKernel::NEGEMMLowpOffsetContributionKernel() |
| 339 | : _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) |
| 340 | { |
| 341 | } |
| 342 | |
| 343 | 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) |
| 344 | { |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 345 | // Perform validate step |
| 346 | ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result); |
| 347 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(mm_result->info(), |
| 348 | vector_sum_col != nullptr ? vector_sum_col->info() : nullptr, // NOLINT |
| 349 | vector_sum_row != nullptr ? vector_sum_row->info() : nullptr, // NOLINT |
| 350 | a_offset, b_offset)); // NOLINT |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 351 | |
| 352 | _vector_sum_col = vector_sum_col; |
| 353 | _vector_sum_row = vector_sum_row; |
| 354 | _mm_result = mm_result; |
| 355 | _a_offset = a_offset; |
| 356 | _b_offset = b_offset; |
| 357 | _k_offset = a_offset * b_offset * k; |
| 358 | |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 359 | // If a_offset == 0, vector_sum_col can be a nullptr |
| 360 | if(a_offset != 0) |
| 361 | { |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 362 | // Check if vector_sum_col_shape should be slidden or not |
| 363 | // 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 |
| 364 | // This scenario can happen when the the matrix multiplication is used to perform a convolution operation |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 365 | _slide_vector_sum_col = vector_sum_col->info()->tensor_shape().num_dimensions() > 1; |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 366 | } |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 367 | |
| 368 | // Configure kernel window |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 369 | auto win_config = validate_and_configure_window(mm_result->info(), |
| 370 | vector_sum_col != nullptr ? vector_sum_col->info() : nullptr, // NOLINT |
| 371 | vector_sum_row != nullptr ? vector_sum_row->info() : nullptr, // NOLINT |
| 372 | a_offset, b_offset); |
| 373 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 374 | INEKernel::configure(win_config.second); |
| 375 | } |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 376 | |
Georgios Pinitas | 631c41a | 2017-12-06 11:53:03 +0000 | [diff] [blame] | 377 | Status NEGEMMLowpOffsetContributionKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, |
| 378 | int32_t a_offset, int32_t b_offset) |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 379 | { |
| 380 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, a_offset, b_offset)); |
| 381 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(mm_result->clone().get(), |
| 382 | vector_sum_col != nullptr ? vector_sum_col->clone().get() : nullptr, |
| 383 | vector_sum_row != nullptr ? vector_sum_row->clone().get() : nullptr, |
| 384 | a_offset, b_offset) |
| 385 | .first); // NOLINT |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 386 | |
Georgios Pinitas | 631c41a | 2017-12-06 11:53:03 +0000 | [diff] [blame] | 387 | return Status{}; |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 388 | } |
| 389 | |
| 390 | void NEGEMMLowpOffsetContributionKernel::run(const Window &window, const ThreadInfo &info) |
| 391 | { |
| 392 | ARM_COMPUTE_UNUSED(info); |
| 393 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 394 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 395 | |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 396 | // Check if input is a 3D reinterpretation |
| 397 | const bool reinterpret_as_3d = _vector_sum_row != nullptr |
| 398 | && _mm_result->info()->num_dimensions() > 1 |
| 399 | && _mm_result->info()->tensor_shape().y() != _vector_sum_row->info()->tensor_shape().x(); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 400 | |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 401 | if(reinterpret_as_3d) |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 402 | { |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 403 | run_offset_contribution<true>(window, _mm_result, _vector_sum_col, _vector_sum_row, _a_offset, _b_offset, _k_offset, _slide_vector_sum_col); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 404 | } |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 405 | else |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 406 | { |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 407 | run_offset_contribution<false>(window, _mm_result, _vector_sum_col, _vector_sum_row, _a_offset, _b_offset, _k_offset, _slide_vector_sum_col); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 408 | } |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 409 | } |