Michele Di Giorgio | d02d5ed | 2021-01-22 09:47:04 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 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 | |
| 25 | #pragma once |
| 26 | |
| 27 | #include "src/core/NEON/kernels/arm_gemm/utils.hpp" |
| 28 | |
| 29 | #ifdef CYCLE_PROFILING |
| 30 | #include "profiler.hpp" |
| 31 | #endif |
| 32 | |
Michele Di Giorgio | 0f033df | 2021-07-16 15:00:08 +0100 | [diff] [blame] | 33 | #include <limits> |
| 34 | |
Michele Di Giorgio | d02d5ed | 2021-01-22 09:47:04 +0000 | [diff] [blame] | 35 | namespace arm_conv { |
| 36 | namespace depthwise { |
| 37 | |
| 38 | namespace common |
| 39 | { |
| 40 | template <typename strategy, typename F> |
| 41 | void depthwise_multiplier_execute( |
| 42 | const F execute_tile, |
| 43 | typename strategy::input_type pad_value, |
| 44 | const DepthwiseArgs &args, |
| 45 | const unsigned int batches, |
| 46 | const unsigned int input_height, |
| 47 | const unsigned int input_width, |
| 48 | const unsigned int input_channels, |
| 49 | const PaddingValues &padding, |
| 50 | const void *const _input, |
| 51 | const size_t ld_input_col, |
| 52 | const size_t ld_input_row, |
| 53 | const size_t ld_input_batch, |
| 54 | const void *const parameters, |
| 55 | const size_t param_stride, |
| 56 | const unsigned int output_height, |
| 57 | const unsigned int output_width, |
| 58 | void *const _output, |
| 59 | const size_t ld_output_col, |
| 60 | const size_t ld_output_row, |
| 61 | const size_t ld_output_batch, |
| 62 | void *const _working_space, |
| 63 | const unsigned int thread_id, |
| 64 | const unsigned int n_threads |
| 65 | ) |
| 66 | { |
| 67 | using TInput = typename strategy::input_type; |
| 68 | using TOutput = typename strategy::return_type; |
| 69 | |
| 70 | // Determine what portion of the work to do. |
| 71 | const unsigned int n_rows_per_thread = arm_gemm::iceildiv(output_height, n_threads); |
| 72 | const int start_out_height = std::min(thread_id * n_rows_per_thread, output_height); |
| 73 | const int end_out_height = std::min(start_out_height + n_rows_per_thread, output_height); |
| 74 | |
| 75 | // Cast input and output pointers into the right types |
| 76 | const TInput *const inptr = static_cast<const TInput *>(_input); |
| 77 | TOutput *const outptr = static_cast<TOutput *>(_output); |
| 78 | |
| 79 | // To simplify the kernel, we process padded or non-NCHW-ordered input into |
| 80 | // a form which can be consumed by the kernel. This data is stored here and |
| 81 | // passed into the kernel as an array of N pointers (one per row of the |
| 82 | // input). |
| 83 | TInput rearranged_input[strategy::input_rows][strategy::input_col_quads*(16 / sizeof(TInput))]; |
| 84 | const TInput *inptrs[strategy::input_rows]; |
| 85 | |
| 86 | // Create an array for the output pointers |
| 87 | TOutput * _outptr_array[strategy::output_rows * strategy::output_cols]; |
| 88 | TOutput **const outptr_array = _outptr_array; |
| 89 | |
| 90 | // Allocate portions of the working space |
| 91 | uint8_t *const working_space = static_cast<uint8_t *>(_working_space); |
| 92 | TOutput *const output_buffer = reinterpret_cast<TOutput *>(working_space); |
| 93 | |
| 94 | // For each output tile, construct the requisite set of pointers and call |
| 95 | // into the kernel. |
| 96 | for (unsigned int batch = 0; batch < batches; batch++) |
| 97 | { |
| 98 | // Get batch pointers |
| 99 | const auto inptr_batch = inptr + batch * ld_input_batch; |
| 100 | const auto outptr_batch = outptr + batch * ld_output_batch; |
| 101 | |
| 102 | for (int start_out_i = start_out_height; |
| 103 | start_out_i < end_out_height; |
| 104 | start_out_i += static_cast<int>(strategy::output_rows)) |
| 105 | { |
| 106 | const int end_out_i = start_out_i + strategy::output_rows; |
| 107 | const int start_in_i = start_out_i * strategy::stride_rows - padding.top; |
| 108 | const int end_in_i = start_in_i + strategy::input_rows; |
| 109 | |
| 110 | // Compute top/bottom padding |
| 111 | const auto pad_top = static_cast<unsigned int>(-std::min(start_in_i, 0)); |
| 112 | const auto pad_bottom = static_cast<unsigned int>(-std::min(static_cast<int>(input_height) - end_in_i, 0)); |
| 113 | const unsigned int valid_output_rows = std::min( |
| 114 | end_out_i - start_out_i, |
| 115 | static_cast<int>(output_height) - start_out_i |
| 116 | ); |
| 117 | |
| 118 | for (int start_out_j = 0; start_out_j < static_cast<int>(output_width);) |
| 119 | { |
| 120 | const int start_in_j = start_out_j * strategy::stride_cols - args.padding.left; |
| 121 | const int pad_left = -std::min(0, start_in_j); |
| 122 | |
| 123 | const int end_out_j = start_out_j + strategy::output_cols; |
| 124 | const int end_in_j = start_in_j + strategy::input_cols; |
| 125 | |
| 126 | const auto pad_right = static_cast<unsigned int>(-std::min(static_cast<int>(input_width) - end_in_j, 0)); |
| 127 | const unsigned int valid_output_cols = std::min( |
| 128 | end_out_j - start_out_j, |
| 129 | static_cast<int>(output_width) - start_out_j |
| 130 | ); |
| 131 | |
| 132 | // Construct the output pointer array. |
| 133 | TOutput **outptr_pos = outptr_array; |
| 134 | for (auto i = 0u; i < valid_output_rows; i++) |
| 135 | { |
| 136 | unsigned int j = 0u; |
| 137 | TOutput *colptr = outptr_batch + (start_out_i + i) * ld_output_row + start_out_j * ld_output_col; |
| 138 | for (; j < valid_output_cols; j++) |
| 139 | { |
| 140 | *(outptr_pos++) = colptr; |
| 141 | colptr += ld_output_col; |
| 142 | } |
| 143 | for (; j < strategy::output_cols; j++) |
| 144 | { |
| 145 | *(outptr_pos++) = output_buffer; |
| 146 | } |
| 147 | } |
| 148 | for (auto i = valid_output_rows; i < strategy::output_rows; i++) |
| 149 | { |
| 150 | for (auto j = 0u; j < strategy::output_cols; j++) |
| 151 | { |
| 152 | *(outptr_pos++) = output_buffer; |
| 153 | } |
| 154 | } |
| 155 | |
| 156 | start_out_j += strategy::output_cols; |
| 157 | |
| 158 | const uint8_t *params = static_cast<const uint8_t *>(parameters); |
| 159 | |
| 160 | // Loop over the input channels |
| 161 | for (unsigned int in_c = 0; in_c < input_channels; in_c++) |
| 162 | { |
| 163 | // Construct the input array - first fill with padding values and |
| 164 | // then fill in correct values. |
| 165 | for (unsigned int i = 0; i < strategy::input_rows; i++) |
| 166 | { |
| 167 | for (unsigned int j = 0; |
| 168 | j < (16 / sizeof(TInput)) * strategy::input_col_quads; j++) |
| 169 | { |
| 170 | rearranged_input[i][j] = pad_value; |
| 171 | } |
| 172 | inptrs[i] = rearranged_input[i]; |
| 173 | } |
| 174 | |
| 175 | auto inptr_row = inptr_batch + in_c + |
| 176 | (start_in_i + pad_top) * ld_input_row + |
| 177 | (start_in_j + pad_left) * ld_input_col; |
| 178 | if (ld_input_col == 1 && !pad_left && |
| 179 | start_in_j + (16 / sizeof(TInput)) * strategy::input_col_quads < input_width) |
| 180 | { |
| 181 | // The input tensor is already in NCHW format, and we're reading |
| 182 | // an unpadded section of it - allow the kernel to read it |
| 183 | // directly. |
| 184 | for (unsigned int i = pad_top; i < strategy::input_rows - pad_bottom; i++) |
| 185 | { |
| 186 | inptrs[i] = inptr_row; |
| 187 | inptr_row += ld_input_row; |
| 188 | } |
| 189 | } |
| 190 | else |
| 191 | { |
| 192 | // Either the input tensor isn't in NCHW format, or we're reading |
| 193 | // a padded section. Copy the relevant portion of the input here |
| 194 | // and allow the kernel to read this. |
| 195 | for (unsigned int i = pad_top; i < strategy::input_rows - pad_bottom; i++) |
| 196 | { |
| 197 | auto inptr_col = inptr_row; |
| 198 | for (unsigned int j = pad_left; j < strategy::input_cols - pad_right; j++) |
| 199 | { |
| 200 | rearranged_input[i][j] = *inptr_col; |
| 201 | inptr_col += ld_input_col; |
| 202 | } |
| 203 | inptr_row += ld_input_row; |
| 204 | } |
| 205 | } |
| 206 | |
| 207 | execute_tile(inptrs, outptr_array, params); |
| 208 | |
| 209 | // Progress the output pointers |
| 210 | TOutput **outptr_pos = outptr_array; |
| 211 | for (auto i = 0u; i < strategy::output_rows * strategy::output_cols; i++) |
| 212 | { |
| 213 | outptr_pos[i] += args.channel_multiplier; |
| 214 | } |
| 215 | |
| 216 | // Progress the pointer into the parameters |
| 217 | params += param_stride; |
| 218 | } |
| 219 | } |
| 220 | } |
| 221 | } |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | template <class strategy> |
| 226 | class DepthwiseDepthfirstWithMultiplier : |
| 227 | public DepthwiseCommon<typename strategy::input_type, |
| 228 | typename strategy::weight_type, |
| 229 | typename strategy::return_type> |
| 230 | { |
| 231 | using TInput = typename strategy::input_type; |
| 232 | using TWeight = typename strategy::weight_type; |
| 233 | using TOutput = typename strategy::return_type; |
| 234 | using TAccum = typename strategy::bias_type; |
| 235 | |
| 236 | size_t sizeof_output_buffer(unsigned int n_channels) const |
| 237 | { |
| 238 | const unsigned int vl = arm_gemm::utils::get_vector_length<TOutput>(strategy::vl_type); |
| 239 | const auto rounded_channels = arm_gemm::roundup(n_channels, vl); |
| 240 | return sizeof(TOutput) * rounded_channels; |
| 241 | } |
| 242 | |
| 243 | public: |
| 244 | DepthwiseDepthfirstWithMultiplier(const DepthwiseArgs &args) : DepthwiseCommon<TInput, TWeight, TOutput>(args) |
| 245 | { |
| 246 | } |
| 247 | |
| 248 | DepthwiseDepthfirstWithMultiplier(DepthwiseDepthfirstWithMultiplier &) = delete; |
| 249 | DepthwiseDepthfirstWithMultiplier &operator=(DepthwiseDepthfirstWithMultiplier &) = delete; |
| 250 | |
| 251 | size_t get_storage_size(void) const override |
| 252 | { |
| 253 | // TODO What if we insert extra padding? Biases are a different size to the inputs, ... |
| 254 | const unsigned int vl = arm_gemm::utils::get_vector_length<TInput>(strategy::vl_type); |
| 255 | const auto rounded_channels = this->m_args.input_channels * arm_gemm::roundup(this->m_args.channel_multiplier, vl); |
| 256 | return (1 + this->m_args.kernel_rows * this->m_args.kernel_cols) * rounded_channels * sizeof(TWeight); |
| 257 | } |
| 258 | |
| 259 | void pack_parameters(void *_buffer, const void *_biases, const void *_weights, size_t ld_weight_col, size_t ld_weight_row) override |
| 260 | { |
| 261 | // TODO What if the kernel needs a different packing function? |
| 262 | |
| 263 | // Cast the pointers |
| 264 | float *buffer = static_cast<float *>(_buffer); |
| 265 | const float *biases = static_cast<const float *>(_biases); |
| 266 | const float *const weights = static_cast<const float *>(_weights); |
| 267 | |
| 268 | const unsigned int vl = arm_gemm::utils::get_vector_length<TInput>(strategy::vl_type); |
| 269 | ld_weight_col = (ld_weight_col == 0) ? this->m_args.channel_multiplier * this->m_args.input_channels : ld_weight_col; |
| 270 | ld_weight_row = (ld_weight_row == 0) ? this->m_args.kernel_cols * ld_weight_col : ld_weight_row; |
| 271 | |
| 272 | for (unsigned int in_c = 0; in_c < this->m_args.input_channels; in_c++) |
| 273 | { |
| 274 | for (unsigned int n = 0; n < this->m_args.channel_multiplier; n += vl) |
| 275 | { |
| 276 | const unsigned int out_c = in_c * this->m_args.channel_multiplier + n; |
| 277 | const unsigned int todo = std::min(vl, this->m_args.channel_multiplier - n); |
| 278 | |
| 279 | // Copy across the correct amount of bias (or 0) |
| 280 | for (unsigned int i = 0; i < todo; i++) |
| 281 | { |
| 282 | buffer[i] = (biases == nullptr) ? 0 : biases[out_c + i]; |
| 283 | } |
| 284 | buffer += vl; |
| 285 | |
| 286 | // Copy each of the weights in turn |
| 287 | auto weights_row = weights + out_c; |
| 288 | for (unsigned int i = 0; i < this->m_args.kernel_rows; i++) |
| 289 | { |
| 290 | auto weights_col = weights_row; |
| 291 | |
| 292 | for (unsigned int j = 0; j < this->m_args.kernel_cols; j++) |
| 293 | { |
| 294 | for (unsigned int m = 0; m < todo; m++) |
| 295 | { |
| 296 | buffer[m] = weights_col[m]; |
| 297 | } |
| 298 | buffer += vl; |
| 299 | |
| 300 | weights_col += ld_weight_col; |
| 301 | } |
| 302 | |
| 303 | weights_row += ld_weight_row; |
| 304 | } |
| 305 | } |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | size_t get_working_size(const unsigned int n_threads, const unsigned int n_channels) const override |
| 310 | { |
| 311 | const unsigned int n_output_channels = n_channels * this->m_args.channel_multiplier; |
| 312 | return n_threads * sizeof_output_buffer(n_output_channels); |
| 313 | } |
| 314 | |
| 315 | using DepthwiseCommon<typename strategy::input_type, typename strategy::weight_type, typename strategy::return_type>::execute; |
| 316 | void execute( |
| 317 | const unsigned int batches, |
| 318 | const unsigned int input_height, |
| 319 | const unsigned int input_width, |
| 320 | const unsigned int input_channels, |
| 321 | const PaddingValues &padding, |
| 322 | const void *const _input, |
| 323 | const size_t ld_input_col, |
| 324 | const size_t ld_input_row, |
| 325 | const size_t ld_input_batch, |
| 326 | const void *const parameters, |
| 327 | const unsigned int output_height, |
| 328 | const unsigned int output_width, |
| 329 | void *const _output, |
| 330 | const size_t ld_output_col, |
| 331 | const size_t ld_output_row, |
| 332 | const size_t ld_output_batch, |
| 333 | void *const _working_space, |
| 334 | const unsigned int thread_id, |
| 335 | const unsigned int n_threads |
| 336 | ) const override |
| 337 | { |
| 338 | strategy strat(this->m_args.cpu_info); |
| 339 | #ifdef CYCLE_PROFILING |
| 340 | arm_gemm::profiler prof; |
| 341 | #endif |
| 342 | |
| 343 | // Compute activation values |
| 344 | TAccum activation_min = std::numeric_limits<TAccum>::has_infinity ? -std::numeric_limits<TAccum>::infinity() : std::numeric_limits<TAccum>::min(); |
| 345 | TAccum activation_max = std::numeric_limits<TAccum>::has_infinity ? std::numeric_limits<TAccum>::infinity() : std::numeric_limits<TAccum>::max(); |
| 346 | |
| 347 | switch (this->m_args.activation.type) |
| 348 | { |
| 349 | case arm_gemm::Activation::Type::BoundedReLU: |
| 350 | activation_max = static_cast<TAccum>(this->m_args.activation.param1); |
| 351 | // Fall through |
| 352 | case arm_gemm::Activation::Type::ReLU: |
| 353 | activation_min = static_cast<TAccum>(0); |
| 354 | break; |
| 355 | default: |
| 356 | break; |
| 357 | } |
| 358 | |
| 359 | // Determine what portion of the work to do. |
| 360 | const unsigned int n_rows_per_thread = arm_gemm::iceildiv(output_height, n_threads); |
| 361 | const int start_out_height = std::min(thread_id * n_rows_per_thread, output_height); |
| 362 | const int end_out_height = std::min(start_out_height + n_rows_per_thread, output_height); |
| 363 | |
| 364 | // Need a stride over blocks of parameters |
| 365 | const unsigned int vl = arm_gemm::utils::get_vector_length<TOutput>(strategy::vl_type); |
| 366 | const unsigned int param_stride = |
| 367 | arm_gemm::roundup(this->m_args.channel_multiplier, vl) * |
| 368 | (sizeof(TAccum) + sizeof(TWeight) * strategy::kernel_rows * strategy::kernel_cols); |
| 369 | |
| 370 | // Cast input and output pointers into the right types |
| 371 | const TInput *const inptr = static_cast<const TInput *>(_input); |
| 372 | TOutput *const outptr = static_cast<TOutput *>(_output); |
| 373 | |
| 374 | // To simplify the kernel, we process padded or non-NCHW-ordered input into |
| 375 | // a form which can be consumed by the kernel. This data is stored here and |
| 376 | // passed into the kernel as an array of N pointers (one per row of the |
| 377 | // input). |
| 378 | TInput rearranged_input[strategy::input_rows][strategy::input_col_quads*4]; |
| 379 | const TInput *inptrs[strategy::input_rows]; |
| 380 | |
| 381 | // Create an array for the output pointers |
| 382 | TOutput * _outptr_array[strategy::output_rows * strategy::output_cols]; |
| 383 | TOutput **const outptr_array = _outptr_array; |
| 384 | |
| 385 | // Allocate portions of the working space |
| 386 | uint8_t *const working_space = static_cast<uint8_t *>(_working_space) + get_working_size(thread_id, input_channels); |
| 387 | TOutput *const output_buffer = reinterpret_cast<TOutput *>(working_space); |
| 388 | |
| 389 | // For each output tile, construct the requisite set of pointers and call |
| 390 | // into the kernel. |
| 391 | for (unsigned int batch = 0; batch < batches; batch++) |
| 392 | { |
| 393 | // Get batch pointers |
| 394 | const auto inptr_batch = inptr + batch * ld_input_batch; |
| 395 | const auto outptr_batch = outptr + batch * ld_output_batch; |
| 396 | |
| 397 | for (int start_out_i = start_out_height; |
| 398 | start_out_i < end_out_height; |
| 399 | start_out_i += static_cast<int>(strategy::output_rows)) |
| 400 | { |
| 401 | const int end_out_i = start_out_i + strategy::output_rows; |
| 402 | const int start_in_i = start_out_i * strategy::stride_rows - padding.top; |
| 403 | const int end_in_i = start_in_i + strategy::input_rows; |
| 404 | |
| 405 | // Compute top/bottom padding |
| 406 | const auto pad_top = static_cast<unsigned int>(-std::min(start_in_i, 0)); |
| 407 | const auto pad_bottom = static_cast<unsigned int>(-std::min(static_cast<int>(input_height) - end_in_i, 0)); |
| 408 | const unsigned int valid_output_rows = std::min( |
| 409 | end_out_i - start_out_i, |
| 410 | static_cast<int>(output_height) - start_out_i |
| 411 | ); |
| 412 | |
| 413 | for (int start_out_j = 0; start_out_j < static_cast<int>(output_width);) |
| 414 | { |
| 415 | const int start_in_j = start_out_j * strategy::stride_cols - this->m_args.padding.left; |
| 416 | const int pad_left = -std::min(0, start_in_j); |
| 417 | |
| 418 | const int end_out_j = start_out_j + strategy::output_cols; |
| 419 | const int end_in_j = start_in_j + strategy::input_cols; |
| 420 | |
| 421 | const auto pad_right = static_cast<unsigned int>(-std::min(static_cast<int>(input_width) - end_in_j, 0)); |
| 422 | const unsigned int valid_output_cols = std::min( |
| 423 | end_out_j - start_out_j, |
| 424 | static_cast<int>(output_width) - start_out_j |
| 425 | ); |
| 426 | |
| 427 | // Construct the output pointer array. |
| 428 | TOutput **outptr_pos = outptr_array; |
| 429 | for (auto i = 0u; i < valid_output_rows; i++) |
| 430 | { |
| 431 | unsigned int j = 0u; |
| 432 | TOutput *colptr = outptr_batch + (start_out_i + i) * ld_output_row + start_out_j * ld_output_col; |
| 433 | for (; j < valid_output_cols; j++) |
| 434 | { |
| 435 | *(outptr_pos++) = colptr; |
| 436 | colptr += ld_output_col; |
| 437 | } |
| 438 | for (; j < strategy::output_cols; j++) |
| 439 | { |
| 440 | *(outptr_pos++) = output_buffer; |
| 441 | } |
| 442 | } |
| 443 | for (auto i = valid_output_rows; i < strategy::output_rows; i++) |
| 444 | { |
| 445 | for (auto j = 0u; j < strategy::output_cols; j++) |
| 446 | { |
| 447 | *(outptr_pos++) = output_buffer; |
| 448 | } |
| 449 | } |
| 450 | |
| 451 | start_out_j += strategy::output_cols; |
| 452 | |
| 453 | const uint8_t *params = static_cast<const uint8_t *>(parameters); |
| 454 | |
| 455 | // Loop over the input channels |
| 456 | for (unsigned int in_c = 0; in_c < input_channels; in_c++) |
| 457 | { |
| 458 | // Construct the input array - first fill with padding values and |
| 459 | // then fill in correct values. |
| 460 | for (unsigned int i = 0; i < strategy::input_rows; i++) |
| 461 | { |
| 462 | for (unsigned int j = 0; j < 4 * strategy::input_col_quads; j++) |
| 463 | { |
| 464 | rearranged_input[i][j] = static_cast<TInput>(0); |
| 465 | } |
| 466 | inptrs[i] = rearranged_input[i]; |
| 467 | } |
| 468 | |
| 469 | auto inptr_row = inptr_batch + in_c + |
| 470 | (start_in_i + pad_top) * ld_input_row + |
| 471 | (start_in_j + pad_left) * ld_input_col; |
| 472 | if (ld_input_col == 1 && !pad_left && |
| 473 | start_in_j + 4 * strategy::input_col_quads < input_width) |
| 474 | { |
| 475 | // The input tensor is already in NCHW format, and we're reading |
| 476 | // an unpadded section of it - allow the kernel to read it |
| 477 | // directly. |
| 478 | for (unsigned int i = pad_top; i < strategy::input_rows - pad_bottom; i++) |
| 479 | { |
| 480 | inptrs[i] = inptr_row; |
| 481 | inptr_row += ld_input_row; |
| 482 | } |
| 483 | } |
| 484 | else |
| 485 | { |
| 486 | // Either the input tensor isn't in NCHW format, or we're reading |
| 487 | // a padded section. Copy the relevant portion of the input here |
| 488 | // and allow the kernel to read this. |
| 489 | for (unsigned int i = pad_top; i < strategy::input_rows - pad_bottom; i++) |
| 490 | { |
| 491 | auto inptr_col = inptr_row; |
| 492 | for (unsigned int j = pad_left; j < strategy::input_cols - pad_right; j++) |
| 493 | { |
| 494 | rearranged_input[i][j] = *inptr_col; |
| 495 | inptr_col += ld_input_col; |
| 496 | } |
| 497 | inptr_row += ld_input_row; |
| 498 | } |
| 499 | } |
| 500 | |
| 501 | { |
| 502 | #ifdef CYCLE_PROFILING |
| 503 | auto p = prof.ScopedProfiler(PROFILE_KERNEL, (unsigned long)(strategy::output_rows * strategy::output_cols * this->m_args.channel_multiplier * strategy::kernel_rows * strategy::kernel_cols)); |
| 504 | #endif |
| 505 | strat.kernel( |
| 506 | inptrs, outptr_array, params, |
| 507 | this->m_args.channel_multiplier, |
| 508 | activation_min, activation_max |
| 509 | ); |
| 510 | } |
| 511 | |
| 512 | // Progress the output pointers |
| 513 | TOutput **outptr_pos = outptr_array; |
| 514 | for (auto i = 0u; i < strategy::output_rows * strategy::output_cols; i++) |
| 515 | { |
| 516 | outptr_pos[i] += this->m_args.channel_multiplier; |
| 517 | } |
| 518 | |
| 519 | // Progress the pointer into the parameters |
| 520 | params += param_stride; |
| 521 | } |
| 522 | } |
| 523 | } |
| 524 | } |
| 525 | } |
| 526 | }; |
| 527 | |
| 528 | } // namespace depthwise |
| 529 | } // namespace arm_conv |