ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 1 | /* |
Michael Tyler | 74921ee | 2023-04-12 17:43:17 +0100 | [diff] [blame^] | 2 | * Copyright (c) 2022-2023 Arm Limited. |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [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 | |
| 25 | #include "addressing.hpp" |
Michael Tyler | 74921ee | 2023-04-12 17:43:17 +0100 | [diff] [blame^] | 26 | #include "utils.hpp" |
| 27 | #include <algorithm> |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 28 | #include <cstring> |
| 29 | |
| 30 | namespace arm_conv { |
| 31 | namespace addressing { |
| 32 | |
| 33 | void fill_pointer_array( |
| 34 | size_t element_size, |
| 35 | void **dest_raw, const unsigned int array_rows, const unsigned int array_cols, |
| 36 | void *base_ptr_raw, size_t ld_row, size_t ld_col, |
| 37 | void *pad_buffer_raw, |
| 38 | const unsigned int pad_top, const unsigned int valid_rows, |
| 39 | const unsigned int pad_left, const unsigned int valid_cols |
| 40 | ) |
| 41 | { |
| 42 | auto dest = reinterpret_cast<char **>(dest_raw); |
| 43 | auto base_ptr = reinterpret_cast<char *>(base_ptr_raw); |
| 44 | auto pad_buffer = reinterpret_cast<char *>(pad_buffer_raw); |
| 45 | ld_row *= element_size; |
| 46 | ld_col *= element_size; |
| 47 | |
| 48 | const auto last_valid_row = std::min(pad_top + valid_rows, array_rows); |
| 49 | const auto last_valid_col = std::min(pad_left + valid_cols, array_cols); |
| 50 | |
| 51 | unsigned int i = 0; |
| 52 | for (; i < pad_top; i++) |
| 53 | { |
| 54 | for (unsigned int j = 0; j < array_cols; j++) |
| 55 | { |
| 56 | *(dest++) = pad_buffer; |
| 57 | } |
| 58 | } |
| 59 | for (; i < last_valid_row; i++) |
| 60 | { |
| 61 | unsigned int j = 0; |
| 62 | auto colptr = base_ptr; |
| 63 | base_ptr += ld_row; |
| 64 | |
| 65 | for (; j < pad_left; j++) |
| 66 | { |
| 67 | *(dest++) = pad_buffer; |
| 68 | } |
| 69 | for (; j < last_valid_col; j++) |
| 70 | { |
| 71 | *(dest++) = colptr; |
| 72 | colptr += ld_col; |
| 73 | } |
| 74 | for (; j < array_cols; j++) |
| 75 | { |
| 76 | *(dest++) = pad_buffer; |
| 77 | } |
| 78 | } |
| 79 | for (; i < array_rows; i++) |
| 80 | { |
| 81 | for (unsigned int j = 0; j < array_cols; j++) |
| 82 | { |
| 83 | *(dest++) = pad_buffer; |
| 84 | } |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | |
| 89 | void fill_pointer_array_generic_kernel( |
| 90 | const size_t element_size, |
| 91 | void **dest_raw, |
| 92 | const unsigned int output_rows, const unsigned int output_cols, |
| 93 | const unsigned int kernel_rows, const unsigned int kernel_cols, |
| 94 | const unsigned int stride_rows, const unsigned int stride_cols, |
| 95 | void *base_ptr_raw, size_t ld_row, size_t ld_col, |
| 96 | void *pad_buffer_raw, |
| 97 | const unsigned int pad_top, const unsigned int valid_rows, |
| 98 | const unsigned int pad_left, const unsigned int valid_cols |
| 99 | ) |
| 100 | { |
| 101 | auto dest = reinterpret_cast<char **>(dest_raw); |
| 102 | auto base_ptr = reinterpret_cast<char *>(base_ptr_raw); |
| 103 | auto pad_buffer = reinterpret_cast<char *>(pad_buffer_raw); |
| 104 | ld_row *= element_size; |
| 105 | ld_col *= element_size; |
| 106 | |
| 107 | const auto last_valid_row = pad_top + valid_rows; |
| 108 | const auto last_valid_col = pad_left + valid_cols; |
| 109 | const auto point_stride = output_rows * output_cols; |
| 110 | |
| 111 | // Iterate over the output points, after every point increment the pointer |
| 112 | // into the address array. |
| 113 | for (unsigned int oi = 0; oi < output_rows; oi++) |
| 114 | { |
| 115 | for (unsigned int oj = 0; oj < output_cols; oj++) |
| 116 | { |
| 117 | auto point_dest = dest; |
| 118 | dest++; |
| 119 | |
| 120 | // Iterate over kernel points and fill in the pointer array. |
| 121 | unsigned int ki = 0, ii = oi*stride_rows; |
| 122 | for (; ii < pad_top && ki < kernel_rows; ii++, ki++) |
| 123 | { |
| 124 | // Fill with padding |
| 125 | for (unsigned int j = 0; j < kernel_cols; j++) |
| 126 | { |
| 127 | *point_dest = pad_buffer; |
| 128 | point_dest += point_stride; |
| 129 | } |
| 130 | } |
| 131 | for (; ii < last_valid_row && ki < kernel_rows; ii++, ki++) |
| 132 | { |
| 133 | unsigned int kj = 0, ij = oj*stride_cols; |
| 134 | for (; ij < pad_left && kj < kernel_cols; ij++, kj++) |
| 135 | { |
| 136 | // Padding |
| 137 | *point_dest = pad_buffer; |
| 138 | point_dest += point_stride; |
| 139 | } |
| 140 | for (; ij < last_valid_col && kj < kernel_cols; ij++, kj++) |
| 141 | { |
| 142 | *point_dest = base_ptr + (ii - pad_top)*ld_row + (ij - pad_left)*ld_col; |
| 143 | point_dest += point_stride; |
| 144 | } |
| 145 | for (; kj < kernel_cols; kj++) |
| 146 | { |
| 147 | // Padding |
| 148 | *point_dest = pad_buffer; |
| 149 | point_dest += point_stride; |
| 150 | } |
| 151 | } |
| 152 | for (; ki < kernel_rows; ki++) |
| 153 | { |
| 154 | // Fill with padding |
| 155 | for (unsigned int j = 0; j < kernel_cols; j++) |
| 156 | { |
| 157 | *point_dest = pad_buffer; |
| 158 | point_dest += point_stride; |
| 159 | } |
| 160 | } |
| 161 | } |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | /* Patch array constructor |
| 166 | * |
| 167 | * Some depthwise kernels require an NCHW-ordered patch of input. Here we |
| 168 | * construct such a patch, and fill in an array of pointers to the rows of the |
| 169 | * patch. |
| 170 | */ |
| 171 | void fill_nchw_patch_array( |
| 172 | size_t element_size, |
| 173 | const void **dest_row_pointers_raw, // Array of pointers to each row of the patch |
| 174 | void *dest_patch_raw, // Pointer to space which can be used to construct the patch |
| 175 | const unsigned int patch_rows, unsigned int patch_cols, // Patch size |
| 176 | const void *src_ptr_raw, size_t ld_row, size_t ld_col, // Source tensor |
| 177 | const void *pad_row, // Pointer to a row of padding values |
| 178 | const unsigned int pad_top, const unsigned int valid_rows, |
| 179 | const unsigned int pad_left, const unsigned int valid_cols |
| 180 | ) |
| 181 | { |
| 182 | // Convert into more useful types |
| 183 | auto row_pointers = reinterpret_cast<const char **>(dest_row_pointers_raw); |
| 184 | auto dest_patch = reinterpret_cast<char *>(dest_patch_raw); |
| 185 | auto src = reinterpret_cast<const char *>(src_ptr_raw); |
| 186 | ld_row *= element_size; |
| 187 | ld_col *= element_size; |
| 188 | |
| 189 | // Round up the patch columns to be a full quad |
| 190 | patch_cols = arm_gemm::roundup<unsigned int>(patch_cols, 16 / element_size); |
| 191 | |
| 192 | const auto last_valid_row = std::min(pad_top + valid_rows, patch_rows); |
| 193 | const auto last_valid_col = std::min(pad_left + valid_cols, patch_cols); |
| 194 | |
| 195 | // Construct the patch and row pointer array together |
| 196 | unsigned int i = 0; |
| 197 | for (; i < pad_top; i++) |
| 198 | { |
| 199 | // Insert pointers into the padding row |
| 200 | *(row_pointers++) = reinterpret_cast<const char *>(pad_row); |
| 201 | } |
| 202 | for (; i < last_valid_row; i++) |
| 203 | { |
| 204 | // Get a copy of the pointer for this row |
| 205 | auto colptr = src; |
| 206 | src += ld_row; |
| 207 | |
| 208 | // If the input is already in NCHW format (ld_col == element_size) AND |
| 209 | // there is no padding, then we just use a pointer to the source tensor; |
| 210 | // otherwise we need to construct a patch and provide a pointer to it. |
| 211 | if (ld_col == element_size && pad_left == 0 && last_valid_col == patch_cols) |
| 212 | { |
| 213 | *(row_pointers++) = colptr; |
| 214 | } |
| 215 | else |
| 216 | { |
| 217 | auto patch_col = dest_patch; |
| 218 | *(row_pointers++) = dest_patch; |
| 219 | dest_patch += element_size * patch_cols; // Move the patch pointer on |
| 220 | |
| 221 | // Construct the patch; fill the entirety with padding and then copy in |
| 222 | // the valid elements. |
| 223 | memcpy(patch_col, pad_row, element_size * patch_cols); |
| 224 | patch_col += pad_left * element_size; // Move over the left padding |
| 225 | |
| 226 | if (ld_col == element_size) |
| 227 | { |
| 228 | // If the input is NCHW then copy across as many columns as we can. |
| 229 | memcpy(patch_col, colptr, (last_valid_col - pad_left) * element_size); |
| 230 | } |
| 231 | else |
| 232 | { |
| 233 | // If the input is NHWC then copy columns across in turn. |
| 234 | for (auto j = pad_left; j < last_valid_col; j++) |
| 235 | { |
| 236 | memcpy(patch_col, colptr, element_size); // Copy the valid element |
| 237 | patch_col += element_size; // Progress the patch destination |
| 238 | colptr += ld_col; // Progress the patch source |
| 239 | } |
| 240 | } |
| 241 | } |
| 242 | } |
| 243 | for (; i < patch_rows; i++) |
| 244 | { |
| 245 | // Insert pointers into the padding row |
| 246 | *(row_pointers++) = reinterpret_cast<const char *>(pad_row); |
| 247 | } |
| 248 | } |
| 249 | |
| 250 | |
| 251 | /* Patch array constructor (generic kernels) |
| 252 | * |
| 253 | * Construct an array of pointers; one pointer for each output row for each |
| 254 | * kernel point. Pointers should point at a whole number of QUADS containing an |
| 255 | * input point for each output point. If the kernel column stride is 1 and the |
| 256 | * data is NCHW then the input tensor might be addressed directly, otherwise a |
| 257 | * new patch sample might need to be constructed. |
| 258 | */ |
| 259 | void fill_patch_array_generic_kernel( |
| 260 | size_t element_size, |
| 261 | const void **dest_pointers_raw, // Pointers: one per output row per kernel point |
| 262 | void *patch_raw, // Pointer to space which can be used to construct the patch |
| 263 | const unsigned int output_rows, const unsigned int output_cols, |
| 264 | const unsigned int kernel_rows, const unsigned int kernel_cols, |
| 265 | const unsigned int stride_rows, const unsigned int stride_cols, |
| 266 | const void *src_ptr_raw, size_t ld_row, size_t ld_col, // Source tensor |
| 267 | const void *pad_row, // Pointer to a row of padding values |
| 268 | const unsigned int pad_top, const unsigned int valid_rows, |
| 269 | const unsigned int pad_left, const unsigned int valid_cols |
| 270 | ) |
| 271 | { |
| 272 | auto dest = reinterpret_cast<const char **>(dest_pointers_raw); |
| 273 | auto patch = reinterpret_cast<char *>(patch_raw); |
| 274 | auto src_ptr = reinterpret_cast<const char *>(src_ptr_raw); |
| 275 | ld_row *= element_size; |
| 276 | ld_col *= element_size; |
| 277 | |
| 278 | // Round up the patch columns to a multiple of quad-length |
| 279 | const auto patch_cols = arm_gemm::roundup<unsigned int>(output_cols, 16 / element_size); |
| 280 | |
| 281 | const auto input_rows = kernel_rows + (output_rows - 1) * stride_rows; |
| 282 | const auto last_valid_row = std::min(pad_top + valid_rows, input_rows); |
| 283 | |
| 284 | const auto input_cols = kernel_cols + (output_cols - 1) * stride_cols; |
| 285 | const auto last_valid_col = std::min(pad_left + valid_cols, input_cols); |
| 286 | |
| 287 | for (auto ki = 0u; ki < kernel_rows; ki++) |
| 288 | { |
| 289 | for (auto kj = 0u; kj < kernel_cols; kj++) |
| 290 | { |
| 291 | auto oi = 0u, ii = ki; |
| 292 | for (; oi < output_rows && ii < pad_top; oi++, ii += stride_rows) |
| 293 | { |
| 294 | // Insert a pointer to the padding row |
| 295 | *(dest++) = reinterpret_cast<const char *>(pad_row); |
| 296 | } |
| 297 | for (; oi < output_rows && ii < last_valid_row; oi++, ii += stride_rows) |
| 298 | { |
| 299 | auto rowptr = src_ptr + (ii - pad_top) * ld_row; |
| 300 | |
| 301 | // Construct a sample of the input here |
| 302 | auto patch_pos = patch; |
| 303 | *(dest++) = patch; |
| 304 | patch += patch_cols * element_size; |
| 305 | |
| 306 | // Fill with padding |
| 307 | memcpy(patch_pos, pad_row, patch_cols * element_size); |
| 308 | |
| 309 | // Fill in the valid elements |
| 310 | auto oj = 0u, ij = kj; |
| 311 | for (; oj < patch_cols && ij < pad_left; oj++, ij += stride_cols) |
| 312 | { |
| 313 | // Do nothing for padding |
| 314 | patch_pos += element_size; |
| 315 | } |
| 316 | for (; oj < patch_cols && ij < last_valid_col; oj++, ij += stride_cols) |
| 317 | { |
| 318 | // Copy from the source tensor |
| 319 | memcpy(patch_pos, rowptr + (ij - pad_left)*ld_col, element_size); |
| 320 | patch_pos += element_size; |
| 321 | } |
| 322 | // No action required for right-hand padding |
| 323 | } |
| 324 | for (; oi < output_rows; oi++) |
| 325 | { |
| 326 | *(dest++) = reinterpret_cast<const char *>(pad_row); |
| 327 | } |
| 328 | } |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | } // namespace addressing |
| 333 | } // namespace arm_conv |