XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #pragma once |
| 10 | |
| 11 | #include <gtest/gtest.h> |
| 12 | |
| 13 | #include <algorithm> |
| 14 | #include <cassert> |
| 15 | #include <cmath> |
| 16 | #include <cstddef> |
| 17 | #include <cstdlib> |
| 18 | #include <functional> |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 19 | #include <limits> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | #include <random> |
| 21 | #include <vector> |
| 22 | |
Frank Barchard | 49b4dcc | 2020-06-26 14:07:19 -0700 | [diff] [blame] | 23 | #include <fp16.h> |
| 24 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 25 | #include <xnnpack.h> |
| 26 | |
| 27 | |
| 28 | class ConvolutionOperatorTester { |
| 29 | public: |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 30 | inline ConvolutionOperatorTester& padding_tf_same(bool padding_same) { |
| 31 | if (padding_same) { |
| 32 | assert(padding_top() == 0); |
| 33 | assert(padding_left() == 0); |
| 34 | assert(padding_bottom() == 0); |
| 35 | assert(padding_right() == 0); |
| 36 | } |
| 37 | this->padding_tf_same_ = padding_same; |
| 38 | return *this; |
| 39 | } |
| 40 | |
| 41 | inline bool padding_tf_same() const { |
| 42 | return this->padding_tf_same_; |
| 43 | } |
| 44 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 45 | inline ConvolutionOperatorTester& padding(uint32_t padding) { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 46 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | this->padding_top_ = padding; |
| 48 | this->padding_right_ = padding; |
| 49 | this->padding_bottom_ = padding; |
| 50 | this->padding_left_ = padding; |
| 51 | return *this; |
| 52 | } |
| 53 | |
| 54 | inline ConvolutionOperatorTester& padding(uint32_t padding_height, uint32_t padding_width) { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 55 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 56 | this->padding_top_ = padding_height; |
| 57 | this->padding_right_ = padding_width; |
| 58 | this->padding_bottom_ = padding_height; |
| 59 | this->padding_left_ = padding_width; |
| 60 | return *this; |
| 61 | } |
| 62 | |
| 63 | inline ConvolutionOperatorTester& padding_height(uint32_t padding_height) { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 64 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 65 | this->padding_top_ = padding_height; |
| 66 | this->padding_bottom_ = padding_height; |
| 67 | return *this; |
| 68 | } |
| 69 | |
| 70 | inline ConvolutionOperatorTester& padding_width(uint32_t padding_width) { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 71 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 72 | this->padding_right_ = padding_width; |
| 73 | this->padding_left_ = padding_width; |
| 74 | return *this; |
| 75 | } |
| 76 | |
| 77 | inline ConvolutionOperatorTester& padding_top(uint32_t padding_top) { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 78 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 79 | this->padding_top_ = padding_top; |
| 80 | return *this; |
| 81 | } |
| 82 | |
| 83 | inline uint32_t padding_top() const { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 84 | if (padding_tf_same()) { |
| 85 | const uint32_t total_padding_height = |
| 86 | (output_height() - 1) * subsampling_height() + dilated_kernel_height() - input_height(); |
| 87 | return total_padding_height / 2; |
| 88 | } else { |
| 89 | return this->padding_top_; |
| 90 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 91 | } |
| 92 | |
| 93 | inline ConvolutionOperatorTester& padding_left(uint32_t padding_left) { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 94 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 95 | this->padding_left_ = padding_left; |
| 96 | return *this; |
| 97 | } |
| 98 | |
| 99 | inline uint32_t padding_left() const { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 100 | if (padding_tf_same()) { |
| 101 | const uint32_t total_padding_width = |
| 102 | (output_width() - 1) * subsampling_width() + dilated_kernel_width() - input_width(); |
| 103 | return total_padding_width / 2; |
| 104 | } else { |
| 105 | return this->padding_left_; |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | inline ConvolutionOperatorTester& padding_bottom(uint32_t padding_bottom) { |
| 110 | assert(!padding_tf_same()); |
| 111 | this->padding_bottom_ = padding_bottom; |
| 112 | return *this; |
| 113 | } |
| 114 | |
| 115 | inline uint32_t padding_bottom() const { |
| 116 | if (padding_tf_same()) { |
| 117 | const uint32_t total_padding_height = |
| 118 | (output_height() - 1) * subsampling_height() + dilated_kernel_height() - input_height(); |
| 119 | return total_padding_height - total_padding_height / 2; |
| 120 | } else { |
| 121 | return this->padding_bottom_; |
| 122 | } |
| 123 | } |
| 124 | |
| 125 | inline ConvolutionOperatorTester& padding_right(uint32_t padding_right) { |
| 126 | assert(!padding_tf_same()); |
| 127 | this->padding_right_ = padding_right; |
| 128 | return *this; |
| 129 | } |
| 130 | |
| 131 | inline uint32_t padding_right() const { |
| 132 | if (padding_tf_same()) { |
| 133 | const uint32_t total_padding_width = |
| 134 | (output_width() - 1) * subsampling_width() + dilated_kernel_width() - input_width(); |
| 135 | return total_padding_width - total_padding_width / 2; |
| 136 | } else { |
| 137 | return this->padding_right_; |
| 138 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 139 | } |
| 140 | |
| 141 | inline ConvolutionOperatorTester& input_size(uint32_t input_height, uint32_t input_width) { |
| 142 | assert(input_height >= 1); |
| 143 | assert(input_width >= 1); |
| 144 | this->input_height_ = input_height; |
| 145 | this->input_width_ = input_width; |
| 146 | return *this; |
| 147 | } |
| 148 | |
| 149 | inline ConvolutionOperatorTester& input_height(uint32_t input_height) { |
| 150 | assert(input_height >= 1); |
| 151 | this->input_height_ = input_height; |
| 152 | return *this; |
| 153 | } |
| 154 | |
| 155 | inline uint32_t input_height() const { |
| 156 | return this->input_height_; |
| 157 | } |
| 158 | |
| 159 | inline ConvolutionOperatorTester& input_width(uint32_t input_width) { |
| 160 | assert(input_width >= 1); |
| 161 | this->input_width_ = input_width; |
| 162 | return *this; |
| 163 | } |
| 164 | |
| 165 | inline uint32_t input_width() const { |
| 166 | return this->input_width_; |
| 167 | } |
| 168 | |
| 169 | inline ConvolutionOperatorTester& groups(uint32_t groups) { |
| 170 | assert(groups >= 1); |
| 171 | this->groups_ = groups; |
| 172 | return *this; |
| 173 | } |
| 174 | |
| 175 | inline uint32_t groups() const { |
| 176 | return this->groups_; |
| 177 | } |
| 178 | |
| 179 | inline ConvolutionOperatorTester& group_input_channels(size_t group_input_channels) { |
| 180 | assert(group_input_channels >= 1); |
| 181 | this->group_input_channels_ = group_input_channels; |
| 182 | return *this; |
| 183 | } |
| 184 | |
| 185 | inline size_t group_input_channels() const { |
| 186 | return this->group_input_channels_; |
| 187 | } |
| 188 | |
| 189 | inline ConvolutionOperatorTester& group_output_channels(size_t group_output_channels) { |
| 190 | assert(group_output_channels >= 1); |
| 191 | this->group_output_channels_ = group_output_channels; |
| 192 | return *this; |
| 193 | } |
| 194 | |
| 195 | inline size_t group_output_channels() const { |
| 196 | return this->group_output_channels_; |
| 197 | } |
| 198 | |
| 199 | inline ConvolutionOperatorTester& batch_size(size_t batch_size) { |
| 200 | assert(batch_size >= 1); |
| 201 | this->batch_size_ = batch_size; |
| 202 | return *this; |
| 203 | } |
| 204 | |
| 205 | inline size_t batch_size() const { |
| 206 | return this->batch_size_; |
| 207 | } |
| 208 | |
| 209 | inline ConvolutionOperatorTester& kernel_size(uint32_t kernel_size) { |
| 210 | assert(kernel_size >= 1); |
| 211 | this->kernel_height_ = kernel_size; |
| 212 | this->kernel_width_ = kernel_size; |
| 213 | return *this; |
| 214 | } |
| 215 | |
| 216 | inline ConvolutionOperatorTester& kernel_size(uint32_t kernel_height, uint32_t kernel_width) { |
| 217 | assert(kernel_height >= 1); |
| 218 | assert(kernel_width >= 1); |
| 219 | this->kernel_height_ = kernel_height; |
| 220 | this->kernel_width_ = kernel_width; |
| 221 | return *this; |
| 222 | } |
| 223 | |
| 224 | inline ConvolutionOperatorTester& kernel_height(uint32_t kernel_height) { |
| 225 | assert(kernel_height >= 1); |
| 226 | this->kernel_height_ = kernel_height; |
| 227 | return *this; |
| 228 | } |
| 229 | |
| 230 | inline uint32_t kernel_height() const { |
| 231 | return this->kernel_height_; |
| 232 | } |
| 233 | |
| 234 | inline ConvolutionOperatorTester& kernel_width(uint32_t kernel_width) { |
| 235 | assert(kernel_width >= 1); |
| 236 | this->kernel_width_ = kernel_width; |
| 237 | return *this; |
| 238 | } |
| 239 | |
| 240 | inline uint32_t kernel_width() const { |
| 241 | return this->kernel_width_; |
| 242 | } |
| 243 | |
| 244 | inline ConvolutionOperatorTester& dilation(uint32_t dilation) { |
| 245 | assert(dilation >= 1); |
| 246 | this->dilation_height_ = dilation; |
| 247 | this->dilation_width_ = dilation; |
| 248 | return *this; |
| 249 | } |
| 250 | |
| 251 | inline ConvolutionOperatorTester& dilation(uint32_t dilation_height, uint32_t dilation_width) { |
| 252 | assert(dilation_height >= 1); |
| 253 | assert(dilation_width >= 1); |
| 254 | this->dilation_height_ = dilation_height; |
| 255 | this->dilation_width_ = dilation_width; |
| 256 | return *this; |
| 257 | } |
| 258 | |
| 259 | inline ConvolutionOperatorTester& dilation_height(uint32_t dilation_height) { |
| 260 | assert(dilation_height >= 1); |
| 261 | this->dilation_height_ = dilation_height; |
| 262 | return *this; |
| 263 | } |
| 264 | |
| 265 | inline uint32_t dilation_height() const { |
| 266 | return this->dilation_height_; |
| 267 | } |
| 268 | |
| 269 | inline ConvolutionOperatorTester& dilation_width(uint32_t dilation_width) { |
| 270 | assert(dilation_width >= 1); |
| 271 | this->dilation_width_ = dilation_width; |
| 272 | return *this; |
| 273 | } |
| 274 | |
| 275 | inline uint32_t dilation_width() const { |
| 276 | return this->dilation_width_; |
| 277 | } |
| 278 | |
| 279 | inline ConvolutionOperatorTester& subsampling(uint32_t subsampling) { |
| 280 | assert(subsampling >= 1); |
| 281 | this->subsampling_height_ = subsampling; |
| 282 | this->subsampling_width_ = subsampling; |
| 283 | return *this; |
| 284 | } |
| 285 | |
| 286 | inline ConvolutionOperatorTester& subsampling(uint32_t subsampling_height, uint32_t subsampling_width) { |
| 287 | assert(subsampling_height >= 1); |
| 288 | assert(subsampling_width >= 1); |
| 289 | this->subsampling_height_ = subsampling_height; |
| 290 | this->subsampling_width_ = subsampling_width; |
| 291 | return *this; |
| 292 | } |
| 293 | |
| 294 | inline ConvolutionOperatorTester& subsampling_height(uint32_t subsampling_height) { |
| 295 | assert(subsampling_height >= 1); |
| 296 | this->subsampling_height_ = subsampling_height; |
| 297 | return *this; |
| 298 | } |
| 299 | |
| 300 | inline uint32_t subsampling_height() const { |
| 301 | return this->subsampling_height_; |
| 302 | } |
| 303 | |
| 304 | inline ConvolutionOperatorTester& subsampling_width(uint32_t subsampling_width) { |
| 305 | assert(subsampling_width >= 1); |
| 306 | this->subsampling_width_ = subsampling_width; |
| 307 | return *this; |
| 308 | } |
| 309 | |
| 310 | inline uint32_t subsampling_width() const { |
| 311 | return this->subsampling_width_; |
| 312 | } |
| 313 | |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 314 | inline ConvolutionOperatorTester& input_channel_stride(size_t input_channel_stride) { |
| 315 | assert(input_channel_stride >= 1); |
| 316 | this->input_channel_stride_ = input_channel_stride; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 317 | return *this; |
| 318 | } |
| 319 | |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 320 | inline size_t input_channel_stride() const { |
| 321 | if (this->input_channel_stride_ == 0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 322 | return group_input_channels() * groups(); |
| 323 | } else { |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 324 | assert(this->input_channel_stride_ >= group_input_channels() * groups()); |
| 325 | return this->input_channel_stride_; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 326 | } |
| 327 | } |
| 328 | |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 329 | inline ConvolutionOperatorTester& output_channel_stride(size_t output_channel_stride) { |
| 330 | assert(output_channel_stride >= 1); |
| 331 | this->output_channel_stride_ = output_channel_stride; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 332 | return *this; |
| 333 | } |
| 334 | |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 335 | inline size_t output_channel_stride() const { |
| 336 | if (this->output_channel_stride_ == 0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 337 | return group_output_channels() * groups(); |
| 338 | } else { |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 339 | assert(this->output_channel_stride_ >= group_output_channels() * groups()); |
| 340 | return this->output_channel_stride_; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 341 | } |
| 342 | } |
| 343 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 344 | inline uint32_t dilated_kernel_height() const { |
| 345 | return (kernel_height() - 1) * dilation_height() + 1; |
| 346 | } |
| 347 | |
| 348 | inline uint32_t dilated_kernel_width() const { |
| 349 | return (kernel_width() - 1) * dilation_width() + 1; |
| 350 | } |
| 351 | |
| 352 | inline size_t output_height() const { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 353 | if (padding_tf_same()) { |
| 354 | return (input_height() + subsampling_height() - 1) / subsampling_height(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 355 | } else { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 356 | const size_t padded_input_height = padding_top() + input_height() + padding_bottom(); |
| 357 | if (padded_input_height <= dilated_kernel_height()) { |
| 358 | return 1; |
| 359 | } else { |
| 360 | return (padded_input_height - dilated_kernel_height()) / subsampling_height() + 1; |
| 361 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 362 | } |
| 363 | } |
| 364 | |
| 365 | inline size_t output_width() const { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 366 | if (padding_tf_same()) { |
| 367 | return (input_width() + subsampling_width() - 1) / subsampling_width(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 368 | } else { |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 369 | const size_t padded_input_width = padding_left() + input_width() + padding_right(); |
| 370 | if (padded_input_width <= dilated_kernel_width()) { |
| 371 | return 1; |
| 372 | } else { |
| 373 | return (padded_input_width - dilated_kernel_width()) / subsampling_width() + 1; |
| 374 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 375 | } |
| 376 | } |
| 377 | |
| 378 | inline ConvolutionOperatorTester& next_input_size(uint32_t next_input_height, uint32_t next_input_width) { |
| 379 | assert(next_input_height >= 1); |
| 380 | assert(next_input_width >= 1); |
| 381 | this->next_input_height_ = next_input_height; |
| 382 | this->next_input_width_ = next_input_width; |
| 383 | return *this; |
| 384 | } |
| 385 | |
| 386 | inline ConvolutionOperatorTester& next_input_height(uint32_t next_input_height) { |
| 387 | assert(next_input_height >= 1); |
| 388 | this->next_input_height_ = next_input_height; |
| 389 | return *this; |
| 390 | } |
| 391 | |
| 392 | inline uint32_t next_input_height() const { |
| 393 | if (this->next_input_height_ == 0) { |
| 394 | return input_height(); |
| 395 | } else { |
| 396 | return this->next_input_height_; |
| 397 | } |
| 398 | } |
| 399 | |
| 400 | inline ConvolutionOperatorTester& next_input_width(uint32_t next_input_width) { |
| 401 | assert(next_input_width >= 1); |
| 402 | this->next_input_width_ = next_input_width; |
| 403 | return *this; |
| 404 | } |
| 405 | |
| 406 | inline uint32_t next_input_width() const { |
| 407 | if (this->next_input_width_ == 0) { |
| 408 | return input_width(); |
| 409 | } else { |
| 410 | return this->next_input_width_; |
| 411 | } |
| 412 | } |
| 413 | |
| 414 | inline size_t next_output_height() const { |
| 415 | const size_t padded_input_height = padding_top() + next_input_height() + padding_bottom(); |
| 416 | if (padded_input_height <= dilated_kernel_height()) { |
| 417 | return 1; |
| 418 | } else { |
| 419 | return (padded_input_height - dilated_kernel_height()) / subsampling_height() + 1; |
| 420 | } |
| 421 | } |
| 422 | |
| 423 | inline size_t next_output_width() const { |
| 424 | const size_t padded_input_width = padding_left() + next_input_width() + padding_right(); |
| 425 | if (padded_input_width <= dilated_kernel_width()) { |
| 426 | return 1; |
| 427 | } else { |
| 428 | return (padded_input_width - dilated_kernel_width()) / subsampling_width() + 1; |
| 429 | } |
| 430 | } |
| 431 | |
| 432 | inline ConvolutionOperatorTester& next_batch_size(size_t next_batch_size) { |
| 433 | assert(next_batch_size >= 1); |
| 434 | this->next_batch_size_ = next_batch_size; |
| 435 | return *this; |
| 436 | } |
| 437 | |
| 438 | inline size_t next_batch_size() const { |
| 439 | if (this->next_batch_size_ == 0) { |
| 440 | return batch_size(); |
| 441 | } else { |
| 442 | return this->next_batch_size_; |
| 443 | } |
| 444 | } |
| 445 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 446 | inline ConvolutionOperatorTester& sparsity(float sparsity) { |
| 447 | this->sparsity_ = sparsity; |
| 448 | return *this; |
| 449 | } |
| 450 | |
| 451 | inline float sparsity() const { |
| 452 | return this->sparsity_; |
| 453 | } |
| 454 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 455 | inline ConvolutionOperatorTester& qmin(uint8_t qmin) { |
| 456 | this->qmin_ = qmin; |
| 457 | return *this; |
| 458 | } |
| 459 | |
| 460 | inline uint8_t qmin() const { |
| 461 | return this->qmin_; |
| 462 | } |
| 463 | |
| 464 | inline ConvolutionOperatorTester& qmax(uint8_t qmax) { |
| 465 | this->qmax_ = qmax; |
| 466 | return *this; |
| 467 | } |
| 468 | |
| 469 | inline uint8_t qmax() const { |
| 470 | return this->qmax_; |
| 471 | } |
| 472 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 473 | inline ConvolutionOperatorTester& force_nhwc_input(bool force_nhwc_input) { |
| 474 | this->force_nhwc_input_ = force_nhwc_input; |
| 475 | return *this; |
| 476 | } |
| 477 | |
| 478 | inline bool force_nhwc_input() const { |
| 479 | return this->force_nhwc_input_; |
| 480 | } |
| 481 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 482 | inline ConvolutionOperatorTester& depthwise_layout(bool depthwise_layout) { |
| 483 | this->depthwise_layout_ = depthwise_layout; |
| 484 | return *this; |
| 485 | } |
| 486 | |
| 487 | inline bool depthwise_layout() const { |
| 488 | return this->depthwise_layout_; |
| 489 | } |
| 490 | |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 491 | inline ConvolutionOperatorTester& has_bias(bool has_bias) { |
| 492 | this->has_bias_ = has_bias; |
| 493 | return *this; |
| 494 | } |
| 495 | |
| 496 | inline bool has_bias() const { |
| 497 | return this->has_bias_; |
| 498 | } |
| 499 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 500 | inline ConvolutionOperatorTester& iterations(size_t iterations) { |
| 501 | this->iterations_ = iterations; |
| 502 | return *this; |
| 503 | } |
| 504 | |
| 505 | inline size_t iterations() const { |
| 506 | return this->iterations_; |
| 507 | } |
| 508 | |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 509 | void TestNHWCxQS8() const { |
| 510 | std::random_device random_device; |
| 511 | auto rng = std::mt19937(random_device()); |
| 512 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng); |
| 513 | auto i8rng = std::bind( |
| 514 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), rng); |
| 515 | |
| 516 | std::vector<int8_t> input(XNN_EXTRA_BYTES / sizeof(int8_t) + |
| 517 | batch_size() * ((input_height() * input_width() - 1) * input_channel_stride() + groups() * group_input_channels()) + 8); |
| 518 | std::vector<int8_t> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 519 | std::vector<int32_t> bias(groups() * group_output_channels()); |
| 520 | std::vector<int8_t> output(batch_size() * ((output_height() * output_width() - 1) * output_channel_stride() + groups() * group_output_channels())); |
| 521 | std::vector<int32_t> accumulators(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 522 | std::vector<double> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 523 | |
| 524 | const int8_t input_zero_point = -1; |
| 525 | |
| 526 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 527 | std::generate(input.begin(), input.end(), std::ref(i8rng)); |
| 528 | std::generate(kernel.begin(), kernel.end(), std::ref(i8rng)); |
| 529 | std::generate(bias.begin(), bias.end(), std::ref(i32rng)); |
| 530 | std::fill(output.begin(), output.end(), 0xA5); |
| 531 | |
| 532 | // Compute reference results, without renormalization. |
| 533 | if (has_bias()) { |
| 534 | for (size_t i = 0; i < batch_size(); i++) { |
| 535 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 536 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 537 | for (size_t g = 0; g < groups(); g++) { |
| 538 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 539 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 540 | bias[g * group_output_channels() + oc]; |
| 541 | } |
| 542 | } |
| 543 | } |
| 544 | } |
| 545 | } |
| 546 | } else { |
| 547 | std::fill(accumulators.begin(), accumulators.end(), 0); |
| 548 | } |
| 549 | if (depthwise_layout()) { |
| 550 | ASSERT_EQ(group_input_channels(), 1); |
| 551 | |
| 552 | for (size_t i = 0; i < batch_size(); i++) { |
| 553 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 554 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 555 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 556 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 557 | if (iy < input_height()) { |
| 558 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 559 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 560 | if (ix < input_width()) { |
| 561 | for (size_t g = 0; g < groups(); g++) { |
| 562 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 563 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 564 | (int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g]) - int32_t(input_zero_point)) * |
| 565 | int32_t(kernel[((ky * kernel_width() + kx) * groups() + g) * group_output_channels() + oc]); |
| 566 | } |
| 567 | } |
| 568 | } |
| 569 | } |
| 570 | } |
| 571 | } |
| 572 | } |
| 573 | } |
| 574 | } |
| 575 | } else { |
| 576 | for (size_t i = 0; i < batch_size(); i++) { |
| 577 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 578 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 579 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 580 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 581 | if (iy < input_height()) { |
| 582 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 583 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 584 | if (ix < input_width()) { |
| 585 | for (size_t g = 0; g < groups(); g++) { |
| 586 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 587 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 588 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 589 | (int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic]) - int32_t(input_zero_point)) * |
| 590 | int32_t(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]); |
| 591 | } |
| 592 | } |
| 593 | } |
| 594 | } |
| 595 | } |
| 596 | } |
| 597 | } |
| 598 | } |
| 599 | } |
| 600 | } |
| 601 | } |
| 602 | |
| 603 | // Compute renormalization parameters. |
| 604 | const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend()); |
| 605 | const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend()); |
| 606 | |
| 607 | const double output_scale = double(uint32_t(accumulated_max - accumulated_min)) / 255.0; |
| 608 | const int8_t output_zero_point = int8_t(std::max(std::min( |
| 609 | lrint(-0.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale), |
| 610 | long(std::numeric_limits<int8_t>::max())), long(std::numeric_limits<int8_t>::min()))); |
| 611 | |
| 612 | // Renormalize reference results. |
| 613 | std::transform(accumulators.cbegin(), accumulators.cend(), output_ref.begin(), |
| 614 | [this, output_scale, output_zero_point](int32_t x) -> double { |
| 615 | return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax() - 0x80) - output_zero_point), double(qmin() - 0x80) - output_zero_point); |
| 616 | }); |
| 617 | |
| 618 | // Create, setup, run, and destroy Convolution operator. |
| 619 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 620 | xnn_operator_t convolution_op = nullptr; |
| 621 | |
| 622 | xnn_status status = xnn_create_convolution2d_nhwc_qs8( |
| 623 | padding_tf_same() ? 0 : padding_top(), padding_tf_same() ? 0 : padding_right(), |
| 624 | padding_tf_same() ? 0 : padding_bottom(), padding_tf_same() ? 0 : padding_left(), |
| 625 | kernel_height(), kernel_width(), |
| 626 | subsampling_height(), subsampling_width(), |
| 627 | dilation_height(), dilation_width(), |
| 628 | groups(), group_input_channels(), group_output_channels(), |
| 629 | input_channel_stride(), output_channel_stride(), |
| 630 | input_zero_point, 1.0f /* input scale */, 1.0f /* kernel scale */, |
| 631 | kernel.data(), has_bias() ? bias.data() : nullptr, |
| 632 | output_zero_point, output_scale, int8_t(qmin() - 0x80), int8_t(qmax() - 0x80), |
| 633 | (depthwise_layout() ? XNN_FLAG_DEPTHWISE_CONVOLUTION : 0) | (padding_tf_same() ? XNN_FLAG_TENSORFLOW_SAME_PADDING : 0), |
| 634 | &convolution_op); |
| 635 | if (status == xnn_status_unsupported_hardware) { |
| 636 | GTEST_SKIP(); |
| 637 | } |
| 638 | ASSERT_EQ(xnn_status_success, status); |
| 639 | ASSERT_NE(nullptr, convolution_op); |
| 640 | |
| 641 | // Smart pointer to automatically delete convolution_op. |
| 642 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 643 | |
| 644 | ASSERT_EQ(xnn_status_success, |
| 645 | xnn_setup_convolution2d_nhwc_qs8( |
| 646 | convolution_op, |
| 647 | batch_size(), input_height(), input_width(), |
| 648 | input.data(), output.data(), |
| 649 | nullptr /* thread pool */)); |
| 650 | |
| 651 | ASSERT_EQ(xnn_status_success, |
| 652 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 653 | |
| 654 | // Verify results. |
| 655 | for (size_t i = 0; i < batch_size(); i++) { |
| 656 | for (size_t y = 0; y < output_height(); y++) { |
| 657 | for (size_t x = 0; x < output_width(); x++) { |
| 658 | for (size_t g = 0; g < groups(); g++) { |
| 659 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 660 | ASSERT_LE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmax() - 0x80)) |
| 661 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 662 | ASSERT_GE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmin() - 0x80)) |
| 663 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 664 | ASSERT_NEAR( |
| 665 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
| 666 | double(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]) - double(output_zero_point), |
| 667 | 0.9) |
| 668 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 669 | } |
| 670 | } |
| 671 | } |
| 672 | } |
| 673 | } |
| 674 | } |
| 675 | } |
| 676 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 677 | void TestNHWCxQU8() const { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 678 | std::random_device random_device; |
| 679 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 680 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng); |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 681 | auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), rng); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 682 | |
| 683 | std::vector<uint8_t> input(XNN_EXTRA_BYTES / sizeof(uint8_t) + |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 684 | batch_size() * ((input_height() * input_width() - 1) * input_channel_stride() + groups() * group_input_channels()) + 8); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 685 | std::vector<uint8_t> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 686 | std::vector<int32_t> bias(groups() * group_output_channels()); |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 687 | std::vector<uint8_t> output(batch_size() * ((output_height() * output_width() - 1) * output_channel_stride() + groups() * group_output_channels())); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 688 | std::vector<int32_t> accumulators(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 689 | std::vector<double> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 690 | |
| 691 | const uint8_t input_zero_point = 127; |
| 692 | const uint8_t kernel_zero_point = 127; |
| 693 | |
| 694 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 695 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 696 | std::generate(kernel.begin(), kernel.end(), std::ref(u8rng)); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 697 | std::generate(bias.begin(), bias.end(), std::ref(i32rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 698 | std::fill(output.begin(), output.end(), 0xA5); |
| 699 | |
| 700 | // Compute reference results, without renormalization. |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 701 | if (has_bias()) { |
| 702 | for (size_t i = 0; i < batch_size(); i++) { |
| 703 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 704 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 705 | for (size_t g = 0; g < groups(); g++) { |
| 706 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 707 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 708 | bias[g * group_output_channels() + oc]; |
| 709 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 710 | } |
| 711 | } |
| 712 | } |
| 713 | } |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 714 | } else { |
| 715 | std::fill(accumulators.begin(), accumulators.end(), 0); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 716 | } |
| 717 | if (depthwise_layout()) { |
| 718 | ASSERT_EQ(group_input_channels(), 1); |
| 719 | |
| 720 | for (size_t i = 0; i < batch_size(); i++) { |
| 721 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 722 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 723 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 724 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 725 | if (iy < input_height()) { |
| 726 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 727 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 728 | if (ix < input_width()) { |
| 729 | for (size_t g = 0; g < groups(); g++) { |
| 730 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 731 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 732 | (int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g]) - int32_t(input_zero_point)) * |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 733 | (int32_t(kernel[((ky * kernel_width() + kx) * groups() + g) * group_output_channels() + oc]) - int32_t(kernel_zero_point)); |
| 734 | } |
| 735 | } |
| 736 | } |
| 737 | } |
| 738 | } |
| 739 | } |
| 740 | } |
| 741 | } |
| 742 | } |
| 743 | } else { |
| 744 | for (size_t i = 0; i < batch_size(); i++) { |
| 745 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 746 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 747 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 748 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 749 | if (iy < input_height()) { |
| 750 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 751 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 752 | if (ix < input_width()) { |
| 753 | for (size_t g = 0; g < groups(); g++) { |
| 754 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 755 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 756 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 757 | (int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic]) - int32_t(input_zero_point)) * |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 758 | (int32_t(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]) - int32_t(kernel_zero_point)); |
| 759 | } |
| 760 | } |
| 761 | } |
| 762 | } |
| 763 | } |
| 764 | } |
| 765 | } |
| 766 | } |
| 767 | } |
| 768 | } |
| 769 | } |
| 770 | |
| 771 | // Compute renormalization parameters. |
| 772 | const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend()); |
| 773 | const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend()); |
| 774 | |
| 775 | const double output_scale = double(uint32_t(accumulated_max - accumulated_min)) / 255.0; |
| 776 | const uint8_t output_zero_point = uint8_t(std::max(std::min( |
| 777 | lrint(127.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale), |
| 778 | long(std::numeric_limits<uint8_t>::max())), long(std::numeric_limits<uint8_t>::min()))); |
| 779 | |
| 780 | // Renormalize reference results. |
| 781 | std::transform(accumulators.cbegin(), accumulators.cend(), output_ref.begin(), |
| 782 | [this, output_scale, output_zero_point](int32_t x) -> double { |
| 783 | return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax()) - output_zero_point), double(qmin()) - output_zero_point); |
| 784 | }); |
| 785 | |
| 786 | // Create, setup, run, and destroy Convolution operator. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 787 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 788 | xnn_operator_t convolution_op = nullptr; |
| 789 | |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 790 | xnn_status status = xnn_create_convolution2d_nhwc_qu8( |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 791 | padding_tf_same() ? 0 : padding_top(), padding_tf_same() ? 0 : padding_right(), |
| 792 | padding_tf_same() ? 0 : padding_bottom(), padding_tf_same() ? 0 : padding_left(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 793 | kernel_height(), kernel_width(), |
| 794 | subsampling_height(), subsampling_width(), |
| 795 | dilation_height(), dilation_width(), |
| 796 | groups(), group_input_channels(), group_output_channels(), |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 797 | input_channel_stride(), output_channel_stride(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 798 | input_zero_point, 1.0f /* input scale */, |
| 799 | kernel_zero_point, 1.0f /* kernel scale */, |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 800 | kernel.data(), has_bias() ? bias.data() : nullptr, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 801 | output_zero_point, output_scale, qmin(), qmax(), |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 802 | (depthwise_layout() ? XNN_FLAG_DEPTHWISE_CONVOLUTION : 0) | (padding_tf_same() ? XNN_FLAG_TENSORFLOW_SAME_PADDING : 0), |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 803 | &convolution_op); |
| 804 | if (status == xnn_status_unsupported_hardware) { |
| 805 | GTEST_SKIP(); |
| 806 | } |
| 807 | ASSERT_EQ(xnn_status_success, status); |
| 808 | ASSERT_NE(nullptr, convolution_op); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 809 | |
| 810 | // Smart pointer to automatically delete convolution_op. |
| 811 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 812 | |
| 813 | ASSERT_EQ(xnn_status_success, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 814 | xnn_setup_convolution2d_nhwc_qu8( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 815 | convolution_op, |
| 816 | batch_size(), input_height(), input_width(), |
| 817 | input.data(), output.data(), |
| 818 | nullptr /* thread pool */)); |
| 819 | |
| 820 | ASSERT_EQ(xnn_status_success, |
| 821 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 822 | |
| 823 | // Verify results. |
| 824 | for (size_t i = 0; i < batch_size(); i++) { |
| 825 | for (size_t y = 0; y < output_height(); y++) { |
| 826 | for (size_t x = 0; x < output_width(); x++) { |
| 827 | for (size_t g = 0; g < groups(); g++) { |
| 828 | for (size_t c = 0; c < group_output_channels(); c++) { |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 829 | ASSERT_LE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmax())) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 830 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 831 | ASSERT_GE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmin())) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 832 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 833 | ASSERT_NEAR( |
| 834 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 835 | double(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]) - double(output_zero_point), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 836 | 0.9) |
| 837 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 838 | } |
| 839 | } |
| 840 | } |
| 841 | } |
| 842 | } |
| 843 | } |
| 844 | } |
| 845 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 846 | void TestNHWCxF32() const { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 847 | std::random_device random_device; |
| 848 | auto rng = std::mt19937(random_device()); |
| 849 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.1f, 1.0f), rng); |
| 850 | |
| 851 | std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 852 | batch_size() * ((input_height() * input_width() - 1) * input_channel_stride() + groups() * group_input_channels())); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 853 | std::vector<float> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 854 | std::vector<float> bias(groups() * group_output_channels()); |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 855 | std::vector<float> output(batch_size() * ((output_height() * output_width() - 1) * output_channel_stride() + groups() * group_output_channels())); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 856 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 857 | |
| 858 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 859 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 860 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 861 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 862 | std::fill(output.begin(), output.end(), nanf("")); |
| 863 | |
| 864 | // Compute reference results, without clamping. |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 865 | if (has_bias()) { |
| 866 | for (size_t i = 0; i < batch_size(); i++) { |
| 867 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 868 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 869 | for (size_t g = 0; g < groups(); g++) { |
| 870 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 871 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 872 | bias[g * group_output_channels() + oc]; |
| 873 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 874 | } |
| 875 | } |
| 876 | } |
| 877 | } |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 878 | } else { |
| 879 | std::fill(output_ref.begin(), output_ref.end(), 0.0f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 880 | } |
| 881 | if (depthwise_layout()) { |
| 882 | ASSERT_EQ(group_input_channels(), 1); |
| 883 | |
| 884 | for (size_t i = 0; i < batch_size(); i++) { |
| 885 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 886 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 887 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 888 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 889 | if (iy < input_height()) { |
| 890 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 891 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 892 | if (ix < input_width()) { |
| 893 | for (size_t g = 0; g < groups(); g++) { |
| 894 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 895 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 896 | input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g] * |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 897 | kernel[((ky * kernel_width() + kx) * groups() + g) * group_output_channels() + oc]; |
| 898 | } |
| 899 | } |
| 900 | } |
| 901 | } |
| 902 | } |
| 903 | } |
| 904 | } |
| 905 | } |
| 906 | } |
| 907 | } else { |
| 908 | for (size_t i = 0; i < batch_size(); i++) { |
| 909 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 910 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 911 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 912 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 913 | if (iy < input_height()) { |
| 914 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 915 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 916 | if (ix < input_width()) { |
| 917 | for (size_t g = 0; g < groups(); g++) { |
| 918 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 919 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 920 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 921 | input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic] * |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 922 | kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]; |
| 923 | } |
| 924 | } |
| 925 | } |
| 926 | } |
| 927 | } |
| 928 | } |
| 929 | } |
| 930 | } |
| 931 | } |
| 932 | } |
| 933 | } |
| 934 | |
| 935 | // Compute clamping parameters. |
| 936 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 937 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 938 | |
| 939 | const float output_min = accumulated_min + (accumulated_max - accumulated_min) / 255.0f * float(qmin()); |
| 940 | const float output_max = accumulated_max - (accumulated_max - accumulated_min) / 255.0f * float(255 - qmax()); |
| 941 | |
| 942 | // Clamp reference results. |
| 943 | for (float& value : output_ref) { |
| 944 | value = std::max(std::min(value, output_max), output_min); |
| 945 | } |
| 946 | |
| 947 | // Create, setup, run, and destroy Convolution operator. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 948 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 949 | xnn_operator_t convolution_op = nullptr; |
| 950 | |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 951 | xnn_status status = xnn_create_convolution2d_nhwc_f32( |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 952 | padding_tf_same() ? 0 : padding_top(), padding_tf_same() ? 0 : padding_right(), |
| 953 | padding_tf_same() ? 0 : padding_bottom(), padding_tf_same() ? 0 : padding_left(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 954 | kernel_height(), kernel_width(), |
| 955 | subsampling_height(), subsampling_width(), |
| 956 | dilation_height(), dilation_width(), |
| 957 | groups(), group_input_channels(), group_output_channels(), |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 958 | input_channel_stride(), output_channel_stride(), |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 959 | kernel.data(), has_bias() ? bias.data() : nullptr, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 960 | output_min, output_max, |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 961 | (depthwise_layout() ? XNN_FLAG_DEPTHWISE_CONVOLUTION : 0) | (padding_tf_same() ? XNN_FLAG_TENSORFLOW_SAME_PADDING : 0), |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 962 | &convolution_op); |
| 963 | if (status == xnn_status_unsupported_hardware) { |
| 964 | GTEST_SKIP(); |
| 965 | } |
| 966 | ASSERT_EQ(xnn_status_success, status); |
| 967 | ASSERT_NE(nullptr, convolution_op); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 968 | |
| 969 | // Smart pointer to automatically delete convolution_op. |
| 970 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 971 | |
| 972 | ASSERT_EQ(xnn_status_success, |
| 973 | xnn_setup_convolution2d_nhwc_f32( |
| 974 | convolution_op, |
| 975 | batch_size(), input_height(), input_width(), |
| 976 | input.data(), output.data(), |
| 977 | nullptr /* thread pool */)); |
| 978 | |
| 979 | ASSERT_EQ(xnn_status_success, |
| 980 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 981 | |
| 982 | // Verify results. |
| 983 | for (size_t i = 0; i < batch_size(); i++) { |
| 984 | for (size_t y = 0; y < output_height(); y++) { |
| 985 | for (size_t x = 0; x < output_width(); x++) { |
| 986 | for (size_t g = 0; g < groups(); g++) { |
| 987 | for (size_t c = 0; c < group_output_channels(); c++) { |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 988 | ASSERT_GE(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c], output_min) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 989 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 990 | ASSERT_LE(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c], output_max) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 991 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 992 | ASSERT_NEAR( |
| 993 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 994 | output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c], |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 995 | 1.0e-4 * std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c])) |
| 996 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 997 | } |
| 998 | } |
| 999 | } |
| 1000 | } |
| 1001 | } |
| 1002 | } |
| 1003 | } |
| 1004 | |
Frank Barchard | 49b4dcc | 2020-06-26 14:07:19 -0700 | [diff] [blame] | 1005 | void TestNHWCxF16() const { |
| 1006 | std::random_device random_device; |
| 1007 | auto rng = std::mt19937(random_device()); |
| 1008 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.1f, 1.0f), rng); |
| 1009 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 1010 | |
| 1011 | std::vector<uint16_t> input(XNN_EXTRA_BYTES / sizeof(uint16_t) + |
| 1012 | batch_size() * ((input_height() * input_width() - 1) * input_channel_stride() + groups() * group_input_channels())); |
| 1013 | std::vector<uint16_t> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 1014 | std::vector<uint16_t> bias(groups() * group_output_channels()); |
| 1015 | std::vector<uint16_t> output(batch_size() * ((output_height() * output_width() - 1) * output_channel_stride() + groups() * group_output_channels())); |
| 1016 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 1017 | |
| 1018 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 1019 | std::generate(input.begin(), input.end(), std::ref(f16rng)); |
| 1020 | std::generate(kernel.begin(), kernel.end(), std::ref(f16rng)); |
| 1021 | std::generate(bias.begin(), bias.end(), std::ref(f16rng)); |
| 1022 | std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */); |
| 1023 | |
| 1024 | // Compute reference results, without clamping. |
| 1025 | if (has_bias()) { |
| 1026 | for (size_t i = 0; i < batch_size(); i++) { |
| 1027 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1028 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1029 | for (size_t g = 0; g < groups(); g++) { |
| 1030 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1031 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 1032 | fp16_ieee_to_fp32_value(bias[g * group_output_channels() + oc]); |
| 1033 | } |
| 1034 | } |
| 1035 | } |
| 1036 | } |
| 1037 | } |
| 1038 | } else { |
| 1039 | std::fill(output_ref.begin(), output_ref.end(), 0.0f); |
| 1040 | } |
| 1041 | if (depthwise_layout()) { |
| 1042 | ASSERT_EQ(group_input_channels(), 1); |
| 1043 | |
| 1044 | for (size_t i = 0; i < batch_size(); i++) { |
| 1045 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1046 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1047 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1048 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1049 | if (iy < input_height()) { |
| 1050 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1051 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1052 | if (ix < input_width()) { |
| 1053 | for (size_t g = 0; g < groups(); g++) { |
| 1054 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1055 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 1056 | fp16_ieee_to_fp32_value(input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g]) * |
| 1057 | fp16_ieee_to_fp32_value(kernel[((ky * kernel_width() + kx) * groups() + g) * group_output_channels() + oc]); |
| 1058 | } |
| 1059 | } |
| 1060 | } |
| 1061 | } |
| 1062 | } |
| 1063 | } |
| 1064 | } |
| 1065 | } |
| 1066 | } |
| 1067 | } else { |
| 1068 | for (size_t i = 0; i < batch_size(); i++) { |
| 1069 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1070 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1071 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1072 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1073 | if (iy < input_height()) { |
| 1074 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1075 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1076 | if (ix < input_width()) { |
| 1077 | for (size_t g = 0; g < groups(); g++) { |
| 1078 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1079 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1080 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 1081 | fp16_ieee_to_fp32_value(input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic]) * |
| 1082 | fp16_ieee_to_fp32_value(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]); |
| 1083 | } |
| 1084 | } |
| 1085 | } |
| 1086 | } |
| 1087 | } |
| 1088 | } |
| 1089 | } |
| 1090 | } |
| 1091 | } |
| 1092 | } |
| 1093 | } |
| 1094 | |
| 1095 | // Compute clamping parameters. |
| 1096 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 1097 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 1098 | const float accumulated_range = accumulated_max - accumulated_min; |
| 1099 | const float scaled_min = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_min + accumulated_range / 255.0f * float(qmin()))); |
| 1100 | const float scaled_max = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_max - accumulated_range / 255.0f * float(255 - qmax()))); |
| 1101 | const float output_min = scaled_min == scaled_max ? -std::numeric_limits<float>::infinity() : scaled_min; |
| 1102 | const float output_max = scaled_min == scaled_max ? +std::numeric_limits<float>::infinity() : scaled_max; |
| 1103 | |
| 1104 | // Clamp reference results. |
| 1105 | for (float& value : output_ref) { |
| 1106 | value = std::max(std::min(value, output_max), output_min); |
| 1107 | } |
| 1108 | |
| 1109 | // Create, setup, run, and destroy Convolution operator. |
| 1110 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 1111 | xnn_operator_t convolution_op = nullptr; |
| 1112 | |
| 1113 | xnn_status status = xnn_create_convolution2d_nhwc_f16( |
| 1114 | padding_tf_same() ? 0 : padding_top(), padding_tf_same() ? 0 : padding_right(), |
| 1115 | padding_tf_same() ? 0 : padding_bottom(), padding_tf_same() ? 0 : padding_left(), |
| 1116 | kernel_height(), kernel_width(), |
| 1117 | subsampling_height(), subsampling_width(), |
| 1118 | dilation_height(), dilation_width(), |
| 1119 | groups(), group_input_channels(), group_output_channels(), |
| 1120 | input_channel_stride(), output_channel_stride(), |
| 1121 | kernel.data(), has_bias() ? bias.data() : nullptr, |
| 1122 | output_min, output_max, |
| 1123 | (depthwise_layout() ? XNN_FLAG_DEPTHWISE_CONVOLUTION : 0) | (padding_tf_same() ? XNN_FLAG_TENSORFLOW_SAME_PADDING : 0), |
| 1124 | &convolution_op); |
Frank Barchard | 49b4dcc | 2020-06-26 14:07:19 -0700 | [diff] [blame] | 1125 | if (status == xnn_status_unsupported_hardware) { |
| 1126 | GTEST_SKIP(); |
| 1127 | } |
| 1128 | ASSERT_EQ(xnn_status_success, status); |
| 1129 | ASSERT_NE(nullptr, convolution_op); |
| 1130 | |
| 1131 | // Smart pointer to automatically delete convolution_op. |
| 1132 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 1133 | |
| 1134 | ASSERT_EQ(xnn_status_success, |
| 1135 | xnn_setup_convolution2d_nhwc_f16( |
| 1136 | convolution_op, |
| 1137 | batch_size(), input_height(), input_width(), |
| 1138 | input.data(), output.data(), |
| 1139 | nullptr /* thread pool */)); |
| 1140 | |
| 1141 | ASSERT_EQ(xnn_status_success, |
| 1142 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 1143 | |
| 1144 | // Verify results. |
| 1145 | for (size_t i = 0; i < batch_size(); i++) { |
| 1146 | for (size_t y = 0; y < output_height(); y++) { |
| 1147 | for (size_t x = 0; x < output_width(); x++) { |
| 1148 | for (size_t g = 0; g < groups(); g++) { |
| 1149 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 1150 | // ASSERT_GE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), output_min) |
| 1151 | // << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1152 | // ASSERT_LE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), output_max) |
| 1153 | // << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1154 | ASSERT_NEAR( |
| 1155 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
| 1156 | fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), |
| 1157 | 1.0e-2 * std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c])) |
| 1158 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1159 | } |
| 1160 | } |
| 1161 | } |
| 1162 | } |
| 1163 | } |
| 1164 | } |
| 1165 | } |
| 1166 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1167 | void TestNCHWxF32() const { |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1168 | std::random_device random_device; |
| 1169 | auto rng = std::mt19937(random_device()); |
| 1170 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.1f, 1.0f), rng); |
| 1171 | auto prng = std::bind(std::uniform_real_distribution<float>(), rng); |
| 1172 | |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1173 | std::vector<float> input(2 * XNN_EXTRA_BYTES / sizeof(float) + |
| 1174 | ((batch_size() - 1) * input_channel_stride() + groups() * group_input_channels()) * input_height() * input_width()); |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1175 | std::vector<float> kernel( |
| 1176 | groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 1177 | std::vector<float> bias(groups() * group_output_channels()); |
| 1178 | std::vector<float> output( |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1179 | ((batch_size() - 1) * output_channel_stride() + groups() * group_output_channels()) * output_height() * output_width()); |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1180 | std::vector<float> output_ref(batch_size() * groups() * group_output_channels() * output_height() * output_width()); |
| 1181 | |
| 1182 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 1183 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 1184 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 1185 | for (float& k : kernel) { |
| 1186 | if (prng() <= sparsity()) { |
| 1187 | k = 0.0f; |
| 1188 | } |
| 1189 | } |
| 1190 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 1191 | std::fill(output.begin(), output.end(), nanf("")); |
| 1192 | |
| 1193 | // Compute reference results, without clamping. |
| 1194 | if (has_bias()) { |
| 1195 | for (size_t i = 0; i < batch_size(); i++) { |
| 1196 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1197 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1198 | for (size_t g = 0; g < groups(); g++) { |
| 1199 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1200 | output_ref[(((i * groups() + g) * group_output_channels() + oc) * output_height() + oy) * output_width() + ox] = |
| 1201 | bias[g * group_output_channels() + oc]; |
| 1202 | } |
| 1203 | } |
| 1204 | } |
| 1205 | } |
| 1206 | } |
| 1207 | } else { |
| 1208 | std::fill(output_ref.begin(), output_ref.end(), 0.0f); |
| 1209 | } |
| 1210 | if (force_nhwc_input()) { |
| 1211 | for (size_t i = 0; i < batch_size(); i++) { |
| 1212 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1213 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1214 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1215 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1216 | if (iy < input_height()) { |
| 1217 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1218 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1219 | if (ix < input_width()) { |
| 1220 | for (size_t g = 0; g < groups(); g++) { |
| 1221 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1222 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1223 | output_ref[(((i * groups() + g) * group_output_channels() + oc) * output_height() + oy) * output_width() + ox] += |
| 1224 | input[((((i * input_height() + iy) * input_width() + ix) * groups() + g) * group_input_channels() + ic)] * |
| 1225 | kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]; |
| 1226 | } |
| 1227 | } |
| 1228 | } |
| 1229 | } |
| 1230 | } |
| 1231 | } |
| 1232 | } |
| 1233 | } |
| 1234 | } |
| 1235 | } |
Marat Dukhan | 3303271 | 2020-06-18 11:06:04 -0700 | [diff] [blame] | 1236 | } else if (depthwise_layout()) { |
| 1237 | ASSERT_EQ(group_input_channels(), 1); |
| 1238 | |
| 1239 | for (size_t i = 0; i < batch_size(); i++) { |
| 1240 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1241 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1242 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1243 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1244 | if (iy < input_height()) { |
| 1245 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1246 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1247 | if (ix < input_width()) { |
| 1248 | for (size_t g = 0; g < groups(); g++) { |
| 1249 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1250 | output_ref[(((i * groups() + g) * group_output_channels() + oc) * output_height() + oy) * output_width() + ox] += |
| 1251 | input[((i * input_channel_stride() + g) * input_height() + iy) * input_width() + ix] * |
| 1252 | kernel[((ky * kernel_width() + kx) * groups() + g) * group_output_channels() + oc]; |
| 1253 | } |
| 1254 | } |
| 1255 | } |
| 1256 | } |
| 1257 | } |
| 1258 | } |
| 1259 | } |
| 1260 | } |
| 1261 | } |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1262 | } else { |
| 1263 | for (size_t i = 0; i < batch_size(); i++) { |
| 1264 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1265 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1266 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1267 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1268 | if (iy < input_height()) { |
| 1269 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1270 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1271 | if (ix < input_width()) { |
| 1272 | for (size_t g = 0; g < groups(); g++) { |
| 1273 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1274 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1275 | output_ref[(((i * groups() + g) * group_output_channels() + oc) * output_height() + oy) * output_width() + ox] += |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1276 | input[((i * input_channel_stride() + g * group_input_channels() + ic) * input_height() + iy) * input_width() + ix] * |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1277 | kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]; |
| 1278 | } |
| 1279 | } |
| 1280 | } |
| 1281 | } |
| 1282 | } |
| 1283 | } |
| 1284 | } |
| 1285 | } |
| 1286 | } |
| 1287 | } |
| 1288 | } |
| 1289 | |
| 1290 | // Compute clamping parameters. |
| 1291 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 1292 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 1293 | |
Marat Dukhan | 869c62d | 2020-04-09 17:17:55 -0700 | [diff] [blame] | 1294 | const float output_min = qmin() == 0 ? -std::numeric_limits<float>::infinity() : |
| 1295 | accumulated_min + (accumulated_max - accumulated_min) / 255.0f * float(qmin()); |
| 1296 | const float output_max = qmax() == 255 ? std::numeric_limits<float>::infinity() : |
| 1297 | accumulated_max - (accumulated_max - accumulated_min) / 255.0f * float(255 - qmax()); |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1298 | |
| 1299 | // Clamp reference results. |
| 1300 | for (float& value : output_ref) { |
| 1301 | value = std::max(std::min(value, output_max), output_min); |
| 1302 | } |
| 1303 | |
| 1304 | // Create, setup, run, and destroy Convolution operator. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1305 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1306 | xnn_operator_t convolution_op = nullptr; |
| 1307 | |
| 1308 | xnn_status status = xnn_create_convolution2d_nchw_f32( |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 1309 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 1310 | kernel_height(), kernel_width(), |
| 1311 | subsampling_height(), subsampling_width(), |
| 1312 | dilation_height(), dilation_width(), |
| 1313 | groups(), group_input_channels(), group_output_channels(), |
| 1314 | input_channel_stride(), output_channel_stride(), |
| 1315 | kernel.data(), has_bias() ? bias.data() : nullptr, |
| 1316 | output_min, output_max, |
| 1317 | (depthwise_layout() ? XNN_FLAG_DEPTHWISE_CONVOLUTION : 0) | (force_nhwc_input() ? XNN_FLAG_INPUT_NHWC : 0), |
| 1318 | &convolution_op); |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1319 | if (status == xnn_status_unsupported_parameter) { |
| 1320 | GTEST_SKIP(); |
| 1321 | } |
| 1322 | ASSERT_EQ(xnn_status_success, status); |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 1323 | ASSERT_NE(nullptr, convolution_op); |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1324 | |
| 1325 | // Smart pointer to automatically delete convolution_op. |
| 1326 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 1327 | |
| 1328 | ASSERT_EQ(xnn_status_success, |
| 1329 | xnn_setup_convolution2d_nchw_f32( |
| 1330 | convolution_op, |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1331 | batch_size(), input_height(), input_width(), |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1332 | input.data(), output.data(), |
| 1333 | nullptr /* thread pool */)); |
| 1334 | |
| 1335 | ASSERT_EQ(xnn_status_success, |
| 1336 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 1337 | |
| 1338 | // Verify results. |
| 1339 | for (size_t i = 0; i < batch_size(); i++) { |
| 1340 | for (size_t y = 0; y < output_height(); y++) { |
| 1341 | for (size_t x = 0; x < output_width(); x++) { |
| 1342 | for (size_t g = 0; g < groups(); g++) { |
| 1343 | for (size_t c = 0; c < group_output_channels(); c++) { |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1344 | ASSERT_GE(output[((i * output_channel_stride() + g * group_output_channels() + c) * output_height() + y) * output_width() + x], output_min) |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1345 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c << ", image = " << i; |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1346 | ASSERT_LE(output[((i * output_channel_stride() + g * group_output_channels() + c) * output_height() + y) * output_width() + x], output_max) |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1347 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c << ", image = " << i; |
| 1348 | ASSERT_NEAR( |
| 1349 | output_ref[(((i * groups() + g) * group_output_channels() + c) * output_height() + y) * output_width() + x], |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1350 | output[((i * output_channel_stride() + g * group_output_channels() + c) * output_height() + y) * output_width() + x], |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1351 | 1.0e-4 * std::abs(output_ref[(((i * groups() + g) * group_output_channels() + c) * output_height() + y) * output_width() + x])) |
| 1352 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c << ", image = " << i; |
| 1353 | } |
| 1354 | } |
| 1355 | } |
| 1356 | } |
| 1357 | } |
| 1358 | } |
| 1359 | } |
| 1360 | |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 1361 | void TestSetupNHWCxQS8() const { |
| 1362 | ASSERT_FALSE(depthwise_layout()); |
| 1363 | |
| 1364 | std::random_device random_device; |
| 1365 | auto rng = std::mt19937(random_device()); |
| 1366 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng); |
| 1367 | auto i8rng = std::bind( |
| 1368 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), rng); |
| 1369 | |
| 1370 | std::vector<int8_t> input(XNN_EXTRA_BYTES / sizeof(int8_t) + std::max( |
| 1371 | batch_size() * ((input_height() * input_width() - 1) * input_channel_stride() + groups() * group_input_channels()), |
| 1372 | next_batch_size() * ((next_input_height() * next_input_width() - 1) * input_channel_stride() + groups() * group_input_channels())) + 8); |
| 1373 | std::vector<int8_t> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 1374 | std::vector<int32_t> bias(groups() * group_output_channels()); |
| 1375 | std::vector<int8_t> output(std::max( |
| 1376 | batch_size() * ((output_height() * output_width() - 1) * output_channel_stride() + groups() * group_output_channels()), |
| 1377 | next_batch_size() * ((next_output_height() * next_output_width() - 1) * output_channel_stride() + groups() * group_output_channels()))); |
| 1378 | std::vector<int32_t> accumulators(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 1379 | std::vector<double> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 1380 | std::vector<int32_t> next_accumulators(next_batch_size() * next_output_height() * next_output_width() * groups() * group_output_channels()); |
| 1381 | std::vector<double> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * groups() * group_output_channels()); |
| 1382 | |
| 1383 | const int8_t input_zero_point = -1; |
| 1384 | |
| 1385 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 1386 | std::generate(input.begin(), input.end(), std::ref(i8rng)); |
| 1387 | std::generate(kernel.begin(), kernel.end(), std::ref(i8rng)); |
| 1388 | std::generate(bias.begin(), bias.end(), std::ref(i32rng)); |
| 1389 | std::fill(output.begin(), output.end(), 0xA5); |
| 1390 | |
| 1391 | // Compute reference results, without renormalization. |
| 1392 | if (has_bias()) { |
| 1393 | for (size_t i = 0; i < batch_size(); i++) { |
| 1394 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1395 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1396 | for (size_t g = 0; g < groups(); g++) { |
| 1397 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1398 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 1399 | bias[g * group_output_channels() + oc]; |
| 1400 | } |
| 1401 | } |
| 1402 | } |
| 1403 | } |
| 1404 | } |
| 1405 | } else { |
| 1406 | std::fill(accumulators.begin(), accumulators.end(), 0); |
| 1407 | } |
| 1408 | for (size_t i = 0; i < batch_size(); i++) { |
| 1409 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1410 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1411 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1412 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1413 | if (iy < input_height()) { |
| 1414 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1415 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1416 | if (ix < input_width()) { |
| 1417 | for (size_t g = 0; g < groups(); g++) { |
| 1418 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1419 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1420 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 1421 | (int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic]) - int32_t(input_zero_point)) * |
| 1422 | int32_t(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]); |
| 1423 | } |
| 1424 | } |
| 1425 | } |
| 1426 | } |
| 1427 | } |
| 1428 | } |
| 1429 | } |
| 1430 | } |
| 1431 | } |
| 1432 | } |
| 1433 | |
| 1434 | // Compute renormalization parameters. |
| 1435 | const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend()); |
| 1436 | const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend()); |
| 1437 | |
| 1438 | const double output_scale = double(uint32_t(accumulated_max - accumulated_min)) / 255.0; |
| 1439 | const int8_t output_zero_point = int8_t(std::max(std::min( |
| 1440 | lrint(-0.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale), |
| 1441 | long(std::numeric_limits<int8_t>::max())), long(std::numeric_limits<int8_t>::min()))); |
| 1442 | |
| 1443 | // Renormalize reference results. |
| 1444 | std::transform(accumulators.cbegin(), accumulators.cend(), output_ref.begin(), |
| 1445 | [this, output_scale, output_zero_point](int32_t x) -> double { |
| 1446 | return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax() - 0x80) - output_zero_point), double(qmin() - 0x80) - output_zero_point); |
| 1447 | }); |
| 1448 | |
| 1449 | // Create, setup, and run Convolution operator once. |
| 1450 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 1451 | xnn_operator_t convolution_op = nullptr; |
| 1452 | |
| 1453 | xnn_status status = xnn_create_convolution2d_nhwc_qs8( |
| 1454 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 1455 | kernel_height(), kernel_width(), |
| 1456 | subsampling_height(), subsampling_width(), |
| 1457 | dilation_height(), dilation_width(), |
| 1458 | groups(), group_input_channels(), group_output_channels(), |
| 1459 | input_channel_stride(), output_channel_stride(), |
| 1460 | input_zero_point, 1.0f /* input scale */, 1.0f /* kernel scale */, |
| 1461 | kernel.data(), has_bias() ? bias.data() : nullptr, |
| 1462 | output_zero_point, output_scale, int8_t(qmin() - 0x80), int8_t(qmax() - 0x80), |
| 1463 | 0, &convolution_op); |
| 1464 | if (status == xnn_status_unsupported_hardware) { |
| 1465 | GTEST_SKIP(); |
| 1466 | } |
| 1467 | ASSERT_EQ(xnn_status_success, status); |
| 1468 | ASSERT_NE(nullptr, convolution_op); |
| 1469 | |
| 1470 | // Smart pointer to automatically delete convolution_op. |
| 1471 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 1472 | |
| 1473 | ASSERT_EQ(xnn_status_success, |
| 1474 | xnn_setup_convolution2d_nhwc_qs8( |
| 1475 | convolution_op, |
| 1476 | batch_size(), input_height(), input_width(), |
| 1477 | input.data(), output.data(), |
| 1478 | nullptr /* thread pool */)); |
| 1479 | |
| 1480 | ASSERT_EQ(xnn_status_success, |
| 1481 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 1482 | |
| 1483 | // Verify results of the first run. |
| 1484 | for (size_t i = 0; i < batch_size(); i++) { |
| 1485 | for (size_t y = 0; y < output_height(); y++) { |
| 1486 | for (size_t x = 0; x < output_width(); x++) { |
| 1487 | for (size_t g = 0; g < groups(); g++) { |
| 1488 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 1489 | ASSERT_LE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmax() - 0x80)) |
| 1490 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1491 | ASSERT_GE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmin() - 0x80)) |
| 1492 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1493 | ASSERT_NEAR( |
| 1494 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
| 1495 | double(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]) - double(output_zero_point), |
| 1496 | 0.9) |
| 1497 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1498 | } |
| 1499 | } |
| 1500 | } |
| 1501 | } |
| 1502 | } |
| 1503 | |
| 1504 | // Re-generate data for the second run. |
| 1505 | std::generate(input.begin(), input.end(), std::ref(i8rng)); |
| 1506 | std::fill(output.begin(), output.end(), 0xA5); |
| 1507 | |
| 1508 | // Compute reference results for the second run, including renormalization. |
| 1509 | if (has_bias()) { |
| 1510 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1511 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1512 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1513 | for (size_t g = 0; g < groups(); g++) { |
| 1514 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1515 | next_accumulators[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 1516 | bias[g * group_output_channels() + oc]; |
| 1517 | } |
| 1518 | } |
| 1519 | } |
| 1520 | } |
| 1521 | } |
| 1522 | } else { |
| 1523 | std::fill(next_accumulators.begin(), next_accumulators.end(), 0); |
| 1524 | } |
| 1525 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1526 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1527 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1528 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1529 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1530 | if (iy < next_input_height()) { |
| 1531 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1532 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1533 | if (ix < next_input_width()) { |
| 1534 | for (size_t g = 0; g < groups(); g++) { |
| 1535 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1536 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1537 | next_accumulators[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 1538 | (int32_t(input[((i * next_input_height() + iy) * next_input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic]) - int32_t(input_zero_point)) * |
| 1539 | int32_t(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]); |
| 1540 | } |
| 1541 | } |
| 1542 | } |
| 1543 | } |
| 1544 | } |
| 1545 | } |
| 1546 | } |
| 1547 | } |
| 1548 | } |
| 1549 | } |
| 1550 | std::transform(next_accumulators.cbegin(), next_accumulators.cend(), next_output_ref.begin(), |
| 1551 | [this, output_scale, output_zero_point](int32_t x) -> double { |
| 1552 | return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax() - 0x80) - output_zero_point), double(qmin() - 0x80) - output_zero_point); |
| 1553 | }); |
| 1554 | |
| 1555 | // Setup and run Convolution operator the second time, and destroy the operator. |
| 1556 | ASSERT_EQ(xnn_status_success, |
| 1557 | xnn_setup_convolution2d_nhwc_qs8( |
| 1558 | convolution_op, |
| 1559 | next_batch_size(), next_input_height(), next_input_width(), |
| 1560 | input.data(), output.data(), |
| 1561 | nullptr /* thread pool */)); |
| 1562 | |
| 1563 | ASSERT_EQ(xnn_status_success, |
| 1564 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 1565 | |
| 1566 | // Verify results of the second run. |
| 1567 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1568 | for (size_t y = 0; y < next_output_height(); y++) { |
| 1569 | for (size_t x = 0; x < next_output_width(); x++) { |
| 1570 | for (size_t g = 0; g < groups(); g++) { |
| 1571 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 1572 | ASSERT_LE(int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmax() - 0x80)) |
| 1573 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1574 | ASSERT_GE(int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmin() - 0x80)) |
| 1575 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1576 | ASSERT_NEAR( |
| 1577 | next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c], |
| 1578 | double(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]) - double(output_zero_point), |
| 1579 | 0.9) |
| 1580 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1581 | } |
| 1582 | } |
| 1583 | } |
| 1584 | } |
| 1585 | } |
| 1586 | } |
| 1587 | } |
| 1588 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1589 | void TestSetupNHWCxQU8() const { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1590 | ASSERT_FALSE(depthwise_layout()); |
| 1591 | |
| 1592 | std::random_device random_device; |
| 1593 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 1594 | auto i32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng); |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 1595 | auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), rng); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1596 | |
| 1597 | std::vector<uint8_t> input(XNN_EXTRA_BYTES / sizeof(uint8_t) + std::max( |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1598 | batch_size() * ((input_height() * input_width() - 1) * input_channel_stride() + groups() * group_input_channels()), |
| 1599 | next_batch_size() * ((next_input_height() * next_input_width() - 1) * input_channel_stride() + groups() * group_input_channels())) + 8); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1600 | std::vector<uint8_t> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 1601 | std::vector<int32_t> bias(groups() * group_output_channels()); |
| 1602 | std::vector<uint8_t> output(std::max( |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1603 | batch_size() * ((output_height() * output_width() - 1) * output_channel_stride() + groups() * group_output_channels()), |
| 1604 | next_batch_size() * ((next_output_height() * next_output_width() - 1) * output_channel_stride() + groups() * group_output_channels()))); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1605 | std::vector<int32_t> accumulators(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 1606 | std::vector<double> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 1607 | std::vector<int32_t> next_accumulators(next_batch_size() * next_output_height() * next_output_width() * groups() * group_output_channels()); |
| 1608 | std::vector<double> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * groups() * group_output_channels()); |
| 1609 | |
| 1610 | const uint8_t input_zero_point = 127; |
| 1611 | const uint8_t kernel_zero_point = 127; |
| 1612 | |
| 1613 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 1614 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 1615 | std::generate(kernel.begin(), kernel.end(), std::ref(u8rng)); |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 1616 | std::generate(bias.begin(), bias.end(), std::ref(i32rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1617 | std::fill(output.begin(), output.end(), 0xA5); |
| 1618 | |
| 1619 | // Compute reference results, without renormalization. |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 1620 | if (has_bias()) { |
| 1621 | for (size_t i = 0; i < batch_size(); i++) { |
| 1622 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1623 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1624 | for (size_t g = 0; g < groups(); g++) { |
| 1625 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1626 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 1627 | bias[g * group_output_channels() + oc]; |
| 1628 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1629 | } |
| 1630 | } |
| 1631 | } |
| 1632 | } |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 1633 | } else { |
| 1634 | std::fill(accumulators.begin(), accumulators.end(), 0); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1635 | } |
| 1636 | for (size_t i = 0; i < batch_size(); i++) { |
| 1637 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1638 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1639 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1640 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1641 | if (iy < input_height()) { |
| 1642 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1643 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1644 | if (ix < input_width()) { |
| 1645 | for (size_t g = 0; g < groups(); g++) { |
| 1646 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1647 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1648 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1649 | (int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic]) - int32_t(input_zero_point)) * |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1650 | (int32_t(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]) - int32_t(kernel_zero_point)); |
| 1651 | } |
| 1652 | } |
| 1653 | } |
| 1654 | } |
| 1655 | } |
| 1656 | } |
| 1657 | } |
| 1658 | } |
| 1659 | } |
| 1660 | } |
| 1661 | |
| 1662 | // Compute renormalization parameters. |
| 1663 | const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend()); |
| 1664 | const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend()); |
| 1665 | |
| 1666 | const double output_scale = double(uint32_t(accumulated_max - accumulated_min)) / 255.0; |
| 1667 | const uint8_t output_zero_point = uint8_t(std::max(std::min( |
| 1668 | lrint(127.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale), |
| 1669 | long(std::numeric_limits<uint8_t>::max())), long(std::numeric_limits<uint8_t>::min()))); |
| 1670 | |
| 1671 | // Renormalize reference results. |
| 1672 | std::transform(accumulators.cbegin(), accumulators.cend(), output_ref.begin(), |
| 1673 | [this, output_scale, output_zero_point](int32_t x) -> double { |
| 1674 | return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax()) - output_zero_point), double(qmin()) - output_zero_point); |
| 1675 | }); |
| 1676 | |
| 1677 | // Create, setup, and run Convolution operator once. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1678 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1679 | xnn_operator_t convolution_op = nullptr; |
| 1680 | |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 1681 | xnn_status status = xnn_create_convolution2d_nhwc_qu8( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1682 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 1683 | kernel_height(), kernel_width(), |
| 1684 | subsampling_height(), subsampling_width(), |
| 1685 | dilation_height(), dilation_width(), |
| 1686 | groups(), group_input_channels(), group_output_channels(), |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1687 | input_channel_stride(), output_channel_stride(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1688 | input_zero_point, 1.0f /* input scale */, |
| 1689 | kernel_zero_point, 1.0f /* kernel scale */, |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 1690 | kernel.data(), has_bias() ? bias.data() : nullptr, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1691 | output_zero_point, output_scale, qmin(), qmax(), |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 1692 | 0, &convolution_op); |
| 1693 | if (status == xnn_status_unsupported_hardware) { |
| 1694 | GTEST_SKIP(); |
| 1695 | } |
| 1696 | ASSERT_EQ(xnn_status_success, status); |
| 1697 | ASSERT_NE(nullptr, convolution_op); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1698 | |
| 1699 | // Smart pointer to automatically delete convolution_op. |
| 1700 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 1701 | |
| 1702 | ASSERT_EQ(xnn_status_success, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1703 | xnn_setup_convolution2d_nhwc_qu8( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1704 | convolution_op, |
| 1705 | batch_size(), input_height(), input_width(), |
| 1706 | input.data(), output.data(), |
| 1707 | nullptr /* thread pool */)); |
| 1708 | |
| 1709 | ASSERT_EQ(xnn_status_success, |
| 1710 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 1711 | |
| 1712 | // Verify results of the first run. |
| 1713 | for (size_t i = 0; i < batch_size(); i++) { |
| 1714 | for (size_t y = 0; y < output_height(); y++) { |
| 1715 | for (size_t x = 0; x < output_width(); x++) { |
| 1716 | for (size_t g = 0; g < groups(); g++) { |
| 1717 | for (size_t c = 0; c < group_output_channels(); c++) { |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1718 | ASSERT_LE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmax())) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1719 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1720 | ASSERT_GE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmin())) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1721 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1722 | ASSERT_NEAR( |
| 1723 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1724 | double(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]) - double(output_zero_point), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1725 | 0.9) |
| 1726 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1727 | } |
| 1728 | } |
| 1729 | } |
| 1730 | } |
| 1731 | } |
| 1732 | |
| 1733 | // Re-generate data for the second run. |
| 1734 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 1735 | std::fill(output.begin(), output.end(), 0xA5); |
| 1736 | |
| 1737 | // Compute reference results for the second run, including renormalization. |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 1738 | if (has_bias()) { |
| 1739 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1740 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1741 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1742 | for (size_t g = 0; g < groups(); g++) { |
| 1743 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1744 | next_accumulators[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 1745 | bias[g * group_output_channels() + oc]; |
| 1746 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1747 | } |
| 1748 | } |
| 1749 | } |
| 1750 | } |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 1751 | } else { |
| 1752 | std::fill(next_accumulators.begin(), next_accumulators.end(), 0); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1753 | } |
| 1754 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1755 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1756 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1757 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1758 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1759 | if (iy < next_input_height()) { |
| 1760 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1761 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1762 | if (ix < next_input_width()) { |
| 1763 | for (size_t g = 0; g < groups(); g++) { |
| 1764 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1765 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1766 | next_accumulators[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1767 | (int32_t(input[((i * next_input_height() + iy) * next_input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic]) - int32_t(input_zero_point)) * |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1768 | (int32_t(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]) - int32_t(kernel_zero_point)); |
| 1769 | } |
| 1770 | } |
| 1771 | } |
| 1772 | } |
| 1773 | } |
| 1774 | } |
| 1775 | } |
| 1776 | } |
| 1777 | } |
| 1778 | } |
| 1779 | std::transform(next_accumulators.cbegin(), next_accumulators.cend(), next_output_ref.begin(), |
| 1780 | [this, output_scale, output_zero_point](int32_t x) -> double { |
| 1781 | return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax()) - output_zero_point), double(qmin()) - output_zero_point); |
| 1782 | }); |
| 1783 | |
| 1784 | // Setup and run Convolution operator the second time, and destroy the operator. |
| 1785 | ASSERT_EQ(xnn_status_success, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1786 | xnn_setup_convolution2d_nhwc_qu8( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1787 | convolution_op, |
| 1788 | next_batch_size(), next_input_height(), next_input_width(), |
| 1789 | input.data(), output.data(), |
| 1790 | nullptr /* thread pool */)); |
| 1791 | |
| 1792 | ASSERT_EQ(xnn_status_success, |
| 1793 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 1794 | |
| 1795 | // Verify results of the second run. |
| 1796 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1797 | for (size_t y = 0; y < next_output_height(); y++) { |
| 1798 | for (size_t x = 0; x < next_output_width(); x++) { |
| 1799 | for (size_t g = 0; g < groups(); g++) { |
| 1800 | for (size_t c = 0; c < group_output_channels(); c++) { |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1801 | ASSERT_LE(int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmax())) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1802 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1803 | ASSERT_GE(int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), int32_t(qmin())) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1804 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1805 | ASSERT_NEAR( |
| 1806 | next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c], |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 1807 | double(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]) - double(output_zero_point), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1808 | 0.9) |
| 1809 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1810 | } |
| 1811 | } |
| 1812 | } |
| 1813 | } |
| 1814 | } |
| 1815 | } |
| 1816 | } |
| 1817 | |
Frank Barchard | 49b4dcc | 2020-06-26 14:07:19 -0700 | [diff] [blame] | 1818 | void TestSetupNHWCxF16() const { |
| 1819 | ASSERT_FALSE(depthwise_layout()); |
| 1820 | |
| 1821 | std::random_device random_device; |
| 1822 | auto rng = std::mt19937(random_device()); |
| 1823 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.1f, 1.0f), rng); |
| 1824 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 1825 | |
| 1826 | std::vector<uint16_t> input(XNN_EXTRA_BYTES / sizeof(uint16_t) + std::max( |
| 1827 | batch_size() * ((input_height() * input_width() - 1) * input_channel_stride() + groups() * group_input_channels()), |
| 1828 | next_batch_size() * ((next_input_height() * next_input_width() - 1) * input_channel_stride() + groups() * group_input_channels()))); |
| 1829 | std::vector<uint16_t> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 1830 | std::vector<uint16_t> bias(groups() * group_output_channels()); |
| 1831 | std::vector<uint16_t> output(std::max( |
| 1832 | batch_size() * ((output_height() * output_width() - 1) * output_channel_stride() + groups() * group_output_channels()), |
| 1833 | next_batch_size() * ((next_output_height() * next_output_width() - 1) * output_channel_stride() + groups() * group_output_channels()))); |
| 1834 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 1835 | std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * groups() * group_output_channels()); |
| 1836 | |
| 1837 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 1838 | std::generate(input.begin(), input.end(), std::ref(f16rng)); |
| 1839 | std::generate(kernel.begin(), kernel.end(), std::ref(f16rng)); |
| 1840 | std::generate(bias.begin(), bias.end(), std::ref(f16rng)); |
| 1841 | std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */); |
| 1842 | |
| 1843 | // Compute reference results, without clamping. |
| 1844 | if (has_bias()) { |
| 1845 | for (size_t i = 0; i < batch_size(); i++) { |
| 1846 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1847 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1848 | for (size_t g = 0; g < groups(); g++) { |
| 1849 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1850 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 1851 | fp16_ieee_to_fp32_value(bias[g * group_output_channels() + oc]); |
| 1852 | } |
| 1853 | } |
| 1854 | } |
| 1855 | } |
| 1856 | } |
| 1857 | } else { |
| 1858 | std::fill(output_ref.begin(), output_ref.end(), 0.0f); |
| 1859 | } |
| 1860 | for (size_t i = 0; i < batch_size(); i++) { |
| 1861 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1862 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1863 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1864 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1865 | if (iy < input_height()) { |
| 1866 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1867 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1868 | if (ix < input_width()) { |
| 1869 | for (size_t g = 0; g < groups(); g++) { |
| 1870 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1871 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1872 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 1873 | fp16_ieee_to_fp32_value(input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic]) * |
| 1874 | fp16_ieee_to_fp32_value(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]); |
| 1875 | } |
| 1876 | } |
| 1877 | } |
| 1878 | } |
| 1879 | } |
| 1880 | } |
| 1881 | } |
| 1882 | } |
| 1883 | } |
| 1884 | } |
| 1885 | |
| 1886 | // Compute clamping parameters. |
| 1887 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 1888 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 1889 | const float accumulated_range = accumulated_max - accumulated_min; |
| 1890 | const float scaled_min = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_min + accumulated_range / 255.0f * float(qmin()))); |
| 1891 | const float scaled_max = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_max - accumulated_range / 255.0f * float(255 - qmax()))); |
| 1892 | const float output_min = scaled_min == scaled_max ? -std::numeric_limits<float>::infinity() : scaled_min; |
| 1893 | const float output_max = scaled_min == scaled_max ? +std::numeric_limits<float>::infinity() : scaled_max; |
| 1894 | |
| 1895 | for (float& output_value : output_ref) { |
| 1896 | output_value = std::min(std::max(output_value, output_min), output_max); |
| 1897 | } |
| 1898 | |
| 1899 | // Create, setup, and run Convolution operator once. |
| 1900 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 1901 | xnn_operator_t convolution_op = nullptr; |
| 1902 | |
| 1903 | xnn_status status = xnn_create_convolution2d_nhwc_f16( |
| 1904 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 1905 | kernel_height(), kernel_width(), |
| 1906 | subsampling_height(), subsampling_width(), |
| 1907 | dilation_height(), dilation_width(), |
| 1908 | groups(), group_input_channels(), group_output_channels(), |
| 1909 | input_channel_stride(), output_channel_stride(), |
| 1910 | kernel.data(), has_bias() ? bias.data() : nullptr, |
| 1911 | output_min, output_max, |
| 1912 | 0, &convolution_op); |
Frank Barchard | 49b4dcc | 2020-06-26 14:07:19 -0700 | [diff] [blame] | 1913 | if (status == xnn_status_unsupported_hardware) { |
| 1914 | GTEST_SKIP(); |
| 1915 | } |
| 1916 | ASSERT_EQ(xnn_status_success, status); |
| 1917 | ASSERT_NE(nullptr, convolution_op); |
| 1918 | |
| 1919 | // Smart pointer to automatically delete convolution_op. |
| 1920 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 1921 | |
| 1922 | ASSERT_EQ(xnn_status_success, |
| 1923 | xnn_setup_convolution2d_nhwc_f16( |
| 1924 | convolution_op, |
| 1925 | batch_size(), input_height(), input_width(), |
| 1926 | input.data(), output.data(), |
| 1927 | nullptr /* thread pool */)); |
| 1928 | |
| 1929 | ASSERT_EQ(xnn_status_success, |
| 1930 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 1931 | |
| 1932 | // Verify results of the first run. |
| 1933 | for (size_t i = 0; i < batch_size(); i++) { |
| 1934 | for (size_t y = 0; y < output_height(); y++) { |
| 1935 | for (size_t x = 0; x < output_width(); x++) { |
| 1936 | for (size_t g = 0; g < groups(); g++) { |
| 1937 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 1938 | ASSERT_GE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), output_min) |
| 1939 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1940 | ASSERT_LE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), output_max) |
| 1941 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1942 | ASSERT_NEAR( |
| 1943 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
| 1944 | fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), |
| 1945 | 1.0e-2 * std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c])) |
| 1946 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1947 | } |
| 1948 | } |
| 1949 | } |
| 1950 | } |
| 1951 | } |
| 1952 | |
| 1953 | // Re-generate data for the second run. |
| 1954 | std::generate(input.begin(), input.end(), std::ref(f16rng)); |
| 1955 | std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */); |
| 1956 | |
| 1957 | // Compute reference results for the second run, including clamping. |
| 1958 | if (has_bias()) { |
| 1959 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1960 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1961 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1962 | for (size_t g = 0; g < groups(); g++) { |
| 1963 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1964 | next_output_ref[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 1965 | fp16_ieee_to_fp32_value(bias[g * group_output_channels() + oc]); |
| 1966 | } |
| 1967 | } |
| 1968 | } |
| 1969 | } |
| 1970 | } |
| 1971 | } else { |
| 1972 | std::fill(next_output_ref.begin(), next_output_ref.end(), 0.0f); |
| 1973 | } |
| 1974 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1975 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1976 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1977 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1978 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1979 | if (iy < next_input_height()) { |
| 1980 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1981 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1982 | if (ix < next_input_width()) { |
| 1983 | for (size_t g = 0; g < groups(); g++) { |
| 1984 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1985 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1986 | next_output_ref[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 1987 | fp16_ieee_to_fp32_value(input[((i * next_input_height() + iy) * next_input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic]) * |
| 1988 | fp16_ieee_to_fp32_value(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]); |
| 1989 | } |
| 1990 | } |
| 1991 | } |
| 1992 | } |
| 1993 | } |
| 1994 | } |
| 1995 | } |
| 1996 | } |
| 1997 | } |
| 1998 | } |
| 1999 | for (float& value : next_output_ref) { |
| 2000 | value = std::max(std::min(value, output_max), output_min); |
| 2001 | } |
| 2002 | |
| 2003 | // Setup and run Convolution operator the second time, and destroy the operator. |
| 2004 | ASSERT_EQ(xnn_status_success, |
| 2005 | xnn_setup_convolution2d_nhwc_f16( |
| 2006 | convolution_op, |
| 2007 | next_batch_size(), next_input_height(), next_input_width(), |
| 2008 | input.data(), output.data(), |
| 2009 | nullptr /* thread pool */)); |
| 2010 | |
| 2011 | ASSERT_EQ(xnn_status_success, |
| 2012 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 2013 | |
| 2014 | // Verify results of the second run. |
| 2015 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 2016 | for (size_t y = 0; y < next_output_height(); y++) { |
| 2017 | for (size_t x = 0; x < next_output_width(); x++) { |
| 2018 | for (size_t g = 0; g < groups(); g++) { |
| 2019 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 2020 | ASSERT_GE(fp16_ieee_to_fp32_value(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), output_min) |
| 2021 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 2022 | ASSERT_LE(fp16_ieee_to_fp32_value(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), output_max) |
| 2023 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 2024 | ASSERT_NEAR( |
| 2025 | next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c], |
| 2026 | fp16_ieee_to_fp32_value(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c]), |
| 2027 | 1.0e-2 * std::abs(next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c])) |
| 2028 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 2029 | } |
| 2030 | } |
| 2031 | } |
| 2032 | } |
| 2033 | } |
| 2034 | } |
| 2035 | } |
| 2036 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2037 | void TestSetupNHWCxF32() const { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2038 | ASSERT_FALSE(depthwise_layout()); |
| 2039 | |
| 2040 | std::random_device random_device; |
| 2041 | auto rng = std::mt19937(random_device()); |
| 2042 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.1f, 1.0f), rng); |
| 2043 | |
| 2044 | std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + std::max( |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2045 | batch_size() * ((input_height() * input_width() - 1) * input_channel_stride() + groups() * group_input_channels()), |
| 2046 | next_batch_size() * ((next_input_height() * next_input_width() - 1) * input_channel_stride() + groups() * group_input_channels()))); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2047 | std::vector<float> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 2048 | std::vector<float> bias(groups() * group_output_channels()); |
| 2049 | std::vector<float> output(std::max( |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2050 | batch_size() * ((output_height() * output_width() - 1) * output_channel_stride() + groups() * group_output_channels()), |
| 2051 | next_batch_size() * ((next_output_height() * next_output_width() - 1) * output_channel_stride() + groups() * group_output_channels()))); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2052 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 2053 | std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * groups() * group_output_channels()); |
| 2054 | |
| 2055 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 2056 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 2057 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 2058 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 2059 | std::fill(output.begin(), output.end(), nanf("")); |
| 2060 | |
| 2061 | // Compute reference results, without clamping. |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 2062 | if (has_bias()) { |
| 2063 | for (size_t i = 0; i < batch_size(); i++) { |
| 2064 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 2065 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 2066 | for (size_t g = 0; g < groups(); g++) { |
| 2067 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 2068 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 2069 | bias[g * group_output_channels() + oc]; |
| 2070 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2071 | } |
| 2072 | } |
| 2073 | } |
| 2074 | } |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 2075 | } else { |
| 2076 | std::fill(output_ref.begin(), output_ref.end(), 0.0f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2077 | } |
| 2078 | for (size_t i = 0; i < batch_size(); i++) { |
| 2079 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 2080 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 2081 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 2082 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 2083 | if (iy < input_height()) { |
| 2084 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 2085 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 2086 | if (ix < input_width()) { |
| 2087 | for (size_t g = 0; g < groups(); g++) { |
| 2088 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 2089 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 2090 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2091 | input[((i * input_height() + iy) * input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic] * |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2092 | kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]; |
| 2093 | } |
| 2094 | } |
| 2095 | } |
| 2096 | } |
| 2097 | } |
| 2098 | } |
| 2099 | } |
| 2100 | } |
| 2101 | } |
| 2102 | } |
| 2103 | |
| 2104 | // Compute clamping parameters. |
| 2105 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 2106 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 2107 | |
| 2108 | const float output_min = accumulated_min + (accumulated_max - accumulated_min) / 255.0f * float(qmin()); |
| 2109 | const float output_max = accumulated_max - (accumulated_max - accumulated_min) / 255.0f * float(255 - qmax()); |
| 2110 | |
| 2111 | // Clamp reference results. |
| 2112 | for (float& value : output_ref) { |
| 2113 | value = std::max(std::min(value, output_max), output_min); |
| 2114 | } |
| 2115 | |
| 2116 | // Create, setup, and run Convolution operator once. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2117 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2118 | xnn_operator_t convolution_op = nullptr; |
| 2119 | |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 2120 | xnn_status status = xnn_create_convolution2d_nhwc_f32( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2121 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 2122 | kernel_height(), kernel_width(), |
| 2123 | subsampling_height(), subsampling_width(), |
| 2124 | dilation_height(), dilation_width(), |
| 2125 | groups(), group_input_channels(), group_output_channels(), |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2126 | input_channel_stride(), output_channel_stride(), |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 2127 | kernel.data(), has_bias() ? bias.data() : nullptr, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2128 | output_min, output_max, |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame^] | 2129 | 0, &convolution_op); |
| 2130 | if (status == xnn_status_unsupported_hardware) { |
| 2131 | GTEST_SKIP(); |
| 2132 | } |
| 2133 | ASSERT_EQ(xnn_status_success, status); |
| 2134 | ASSERT_NE(nullptr, convolution_op); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2135 | |
| 2136 | // Smart pointer to automatically delete convolution_op. |
| 2137 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 2138 | |
| 2139 | ASSERT_EQ(xnn_status_success, |
| 2140 | xnn_setup_convolution2d_nhwc_f32( |
| 2141 | convolution_op, |
| 2142 | batch_size(), input_height(), input_width(), |
| 2143 | input.data(), output.data(), |
| 2144 | nullptr /* thread pool */)); |
| 2145 | |
| 2146 | ASSERT_EQ(xnn_status_success, |
| 2147 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 2148 | |
| 2149 | // Verify results of the first run. |
| 2150 | for (size_t i = 0; i < batch_size(); i++) { |
| 2151 | for (size_t y = 0; y < output_height(); y++) { |
| 2152 | for (size_t x = 0; x < output_width(); x++) { |
| 2153 | for (size_t g = 0; g < groups(); g++) { |
| 2154 | for (size_t c = 0; c < group_output_channels(); c++) { |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2155 | ASSERT_GE(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c], output_min) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2156 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2157 | ASSERT_LE(output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c], output_max) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2158 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 2159 | ASSERT_NEAR( |
| 2160 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2161 | output[((i * output_height() + y) * output_width() + x) * output_channel_stride() + g * group_output_channels() + c], |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2162 | 1.0e-4 * std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c])) |
| 2163 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 2164 | } |
| 2165 | } |
| 2166 | } |
| 2167 | } |
| 2168 | } |
| 2169 | |
| 2170 | // Re-generate data for the second run. |
| 2171 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 2172 | std::fill(output.begin(), output.end(), nanf("")); |
| 2173 | |
| 2174 | // Compute reference results for the second run, including clamping. |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 2175 | if (has_bias()) { |
| 2176 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 2177 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 2178 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 2179 | for (size_t g = 0; g < groups(); g++) { |
| 2180 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 2181 | next_output_ref[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 2182 | bias[g * group_output_channels() + oc]; |
| 2183 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2184 | } |
| 2185 | } |
| 2186 | } |
| 2187 | } |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 2188 | } else { |
| 2189 | std::fill(next_output_ref.begin(), next_output_ref.end(), 0.0f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2190 | } |
| 2191 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 2192 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 2193 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 2194 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 2195 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 2196 | if (iy < next_input_height()) { |
| 2197 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 2198 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 2199 | if (ix < next_input_width()) { |
| 2200 | for (size_t g = 0; g < groups(); g++) { |
| 2201 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 2202 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 2203 | next_output_ref[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2204 | input[((i * next_input_height() + iy) * next_input_width() + ix) * input_channel_stride() + g * group_input_channels() + ic] * |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2205 | kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]; |
| 2206 | } |
| 2207 | } |
| 2208 | } |
| 2209 | } |
| 2210 | } |
| 2211 | } |
| 2212 | } |
| 2213 | } |
| 2214 | } |
| 2215 | } |
| 2216 | for (float& value : next_output_ref) { |
| 2217 | value = std::max(std::min(value, output_max), output_min); |
| 2218 | } |
| 2219 | |
| 2220 | // Setup and run Convolution operator the second time, and destroy the operator. |
| 2221 | ASSERT_EQ(xnn_status_success, |
| 2222 | xnn_setup_convolution2d_nhwc_f32( |
| 2223 | convolution_op, |
| 2224 | next_batch_size(), next_input_height(), next_input_width(), |
| 2225 | input.data(), output.data(), |
| 2226 | nullptr /* thread pool */)); |
| 2227 | |
| 2228 | ASSERT_EQ(xnn_status_success, |
| 2229 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 2230 | |
| 2231 | // Verify results of the second run. |
| 2232 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 2233 | for (size_t y = 0; y < next_output_height(); y++) { |
| 2234 | for (size_t x = 0; x < next_output_width(); x++) { |
| 2235 | for (size_t g = 0; g < groups(); g++) { |
| 2236 | for (size_t c = 0; c < group_output_channels(); c++) { |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2237 | ASSERT_GE(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c], output_min) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2238 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2239 | ASSERT_LE(output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c], output_max) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2240 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 2241 | ASSERT_NEAR( |
| 2242 | next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c], |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2243 | output[((i * next_output_height() + y) * next_output_width() + x) * output_channel_stride() + g * group_output_channels() + c], |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2244 | 1.0e-4 * std::abs(next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c])) |
| 2245 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 2246 | } |
| 2247 | } |
| 2248 | } |
| 2249 | } |
| 2250 | } |
| 2251 | } |
| 2252 | } |
| 2253 | |
| 2254 | private: |
| 2255 | uint32_t padding_top_{0}; |
| 2256 | uint32_t padding_right_{0}; |
| 2257 | uint32_t padding_bottom_{0}; |
| 2258 | uint32_t padding_left_{0}; |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 2259 | bool padding_tf_same_{false}; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2260 | size_t input_height_{1}; |
| 2261 | size_t input_width_{1}; |
| 2262 | uint32_t groups_{1}; |
| 2263 | size_t group_input_channels_{1}; |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2264 | size_t input_channel_stride_{0}; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2265 | size_t group_output_channels_{1}; |
Marat Dukhan | c3d52cf | 2020-06-18 07:56:25 -0700 | [diff] [blame] | 2266 | size_t output_channel_stride_{0}; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2267 | size_t batch_size_{1}; |
| 2268 | uint32_t kernel_height_{1}; |
| 2269 | uint32_t kernel_width_{1}; |
| 2270 | uint32_t dilation_height_{1}; |
| 2271 | uint32_t dilation_width_{1}; |
| 2272 | uint32_t subsampling_height_{1}; |
| 2273 | uint32_t subsampling_width_{1}; |
| 2274 | size_t next_input_height_{0}; |
| 2275 | size_t next_input_width_{0}; |
| 2276 | size_t next_batch_size_{0}; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2277 | float sparsity_{0.0f}; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2278 | uint8_t qmin_{0}; |
| 2279 | uint8_t qmax_{255}; |
| 2280 | bool depthwise_layout_{false}; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2281 | bool force_nhwc_input_{false}; |
Marat Dukhan | f568f08 | 2019-10-30 09:47:07 -0700 | [diff] [blame] | 2282 | bool has_bias_{true}; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2283 | size_t iterations_{1}; |
| 2284 | }; |