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 | |
Marat Dukhan | 5756a92 | 2022-02-04 01:55:53 -0800 | [diff] [blame] | 13 | #include <fp16.h> |
| 14 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 15 | #include <algorithm> |
| 16 | #include <cassert> |
| 17 | #include <cstddef> |
| 18 | #include <cstdlib> |
| 19 | #include <functional> |
| 20 | #include <limits> |
| 21 | #include <random> |
| 22 | #include <vector> |
| 23 | |
| 24 | #include <xnnpack.h> |
| 25 | |
| 26 | |
| 27 | class MaxPoolingOperatorTester { |
| 28 | public: |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 29 | inline MaxPoolingOperatorTester& padding_tf_same(bool padding_same) { |
| 30 | if (padding_same) { |
| 31 | assert(padding_top() == 0); |
| 32 | assert(padding_left() == 0); |
| 33 | assert(padding_bottom() == 0); |
| 34 | assert(padding_right() == 0); |
| 35 | } |
| 36 | this->padding_tf_same_ = padding_same; |
| 37 | return *this; |
| 38 | } |
| 39 | |
| 40 | inline bool padding_tf_same() const { |
| 41 | return this->padding_tf_same_; |
| 42 | } |
| 43 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 44 | inline MaxPoolingOperatorTester& padding(uint32_t padding) { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 45 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 46 | this->padding_top_ = padding; |
| 47 | this->padding_right_ = padding; |
| 48 | this->padding_bottom_ = padding; |
| 49 | this->padding_left_ = padding; |
| 50 | return *this; |
| 51 | } |
| 52 | |
| 53 | inline MaxPoolingOperatorTester& padding(uint32_t padding_height, uint32_t padding_width) { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 54 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 55 | this->padding_top_ = padding_height; |
| 56 | this->padding_right_ = padding_width; |
| 57 | this->padding_bottom_ = padding_height; |
| 58 | this->padding_left_ = padding_width; |
| 59 | return *this; |
| 60 | } |
| 61 | |
| 62 | inline MaxPoolingOperatorTester& padding_height(uint32_t padding_height) { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 63 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 64 | this->padding_top_ = padding_height; |
| 65 | this->padding_bottom_ = padding_height; |
| 66 | return *this; |
| 67 | } |
| 68 | |
| 69 | inline MaxPoolingOperatorTester& padding_width(uint32_t padding_width) { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 70 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 71 | this->padding_right_ = padding_width; |
| 72 | this->padding_left_ = padding_width; |
| 73 | return *this; |
| 74 | } |
| 75 | |
| 76 | inline MaxPoolingOperatorTester& padding_top(uint32_t padding_top) { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 77 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 78 | this->padding_top_ = padding_top; |
| 79 | return *this; |
| 80 | } |
| 81 | |
| 82 | inline uint32_t padding_top() const { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 83 | if (padding_tf_same()) { |
| 84 | const uint32_t total_padding_height = |
| 85 | (output_height() - 1) * stride_height() + dilated_pooling_height() - input_height(); |
| 86 | return total_padding_height / 2; |
| 87 | } else { |
| 88 | return this->padding_top_; |
| 89 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 90 | } |
| 91 | |
| 92 | inline MaxPoolingOperatorTester& padding_left(uint32_t padding_left) { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 93 | assert(!padding_tf_same()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 94 | this->padding_left_ = padding_left; |
| 95 | return *this; |
| 96 | } |
| 97 | |
| 98 | inline uint32_t padding_left() const { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 99 | if (padding_tf_same()) { |
| 100 | const uint32_t total_padding_width = |
| 101 | (output_width() - 1) * stride_width() + dilated_pooling_width() - input_width(); |
| 102 | return total_padding_width / 2; |
| 103 | } else { |
| 104 | return this->padding_left_; |
| 105 | } |
| 106 | } |
| 107 | |
| 108 | inline MaxPoolingOperatorTester& padding_bottom(uint32_t padding_bottom) { |
| 109 | assert(!padding_tf_same()); |
| 110 | this->padding_bottom_ = padding_bottom; |
| 111 | return *this; |
| 112 | } |
| 113 | |
| 114 | inline uint32_t padding_bottom() const { |
| 115 | if (padding_tf_same()) { |
| 116 | const uint32_t total_padding_height = |
| 117 | (output_height() - 1) * stride_height() + dilated_pooling_height() - input_height(); |
| 118 | return total_padding_height - total_padding_height / 2; |
| 119 | } else { |
| 120 | return this->padding_bottom_; |
| 121 | } |
| 122 | } |
| 123 | |
| 124 | inline MaxPoolingOperatorTester& padding_right(uint32_t padding_right) { |
| 125 | assert(!padding_tf_same()); |
| 126 | this->padding_right_ = padding_right; |
| 127 | return *this; |
| 128 | } |
| 129 | |
| 130 | inline uint32_t padding_right() const { |
| 131 | if (padding_tf_same()) { |
| 132 | const uint32_t total_padding_width = |
| 133 | (output_width() - 1) * stride_width() + dilated_pooling_width() - input_width(); |
| 134 | return total_padding_width - total_padding_width / 2; |
| 135 | } else { |
| 136 | return this->padding_right_; |
| 137 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 138 | } |
| 139 | |
| 140 | inline MaxPoolingOperatorTester& input_size(size_t input_height, size_t input_width) { |
| 141 | assert(input_height >= 1); |
| 142 | assert(input_width >= 1); |
| 143 | this->input_height_ = input_height; |
| 144 | this->input_width_ = input_width; |
| 145 | return *this; |
| 146 | } |
| 147 | |
| 148 | inline MaxPoolingOperatorTester& input_height(size_t input_height) { |
| 149 | assert(input_height >= 1); |
| 150 | this->input_height_ = input_height; |
| 151 | return *this; |
| 152 | } |
| 153 | |
| 154 | inline size_t input_height() const { |
| 155 | return this->input_height_; |
| 156 | } |
| 157 | |
| 158 | inline MaxPoolingOperatorTester& input_width(size_t input_width) { |
| 159 | assert(input_width >= 1); |
| 160 | this->input_width_ = input_width; |
| 161 | return *this; |
| 162 | } |
| 163 | |
| 164 | inline size_t input_width() const { |
| 165 | return this->input_width_; |
| 166 | } |
| 167 | |
| 168 | inline MaxPoolingOperatorTester& channels(size_t channels) { |
| 169 | assert(channels != 0); |
| 170 | this->channels_ = channels; |
| 171 | return *this; |
| 172 | } |
| 173 | |
| 174 | inline size_t channels() const { |
| 175 | return this->channels_; |
| 176 | } |
| 177 | |
| 178 | inline MaxPoolingOperatorTester& batch_size(size_t batch_size) { |
| 179 | assert(batch_size != 0); |
| 180 | this->batch_size_ = batch_size; |
| 181 | return *this; |
| 182 | } |
| 183 | |
| 184 | inline size_t batch_size() const { |
| 185 | return this->batch_size_; |
| 186 | } |
| 187 | |
| 188 | inline MaxPoolingOperatorTester& pooling_size(uint32_t pooling_size) { |
| 189 | assert(pooling_size >= 1); |
| 190 | this->pooling_height_ = pooling_size; |
| 191 | this->pooling_width_ = pooling_size; |
| 192 | return *this; |
| 193 | } |
| 194 | |
| 195 | inline MaxPoolingOperatorTester& pooling_size(uint32_t pooling_height, uint32_t pooling_width) { |
| 196 | assert(pooling_height >= 1); |
| 197 | assert(pooling_width >= 1); |
| 198 | this->pooling_height_ = pooling_height; |
| 199 | this->pooling_width_ = pooling_width; |
| 200 | return *this; |
| 201 | } |
| 202 | |
| 203 | inline MaxPoolingOperatorTester& pooling_height(uint32_t pooling_height) { |
| 204 | assert(pooling_height >= 1); |
| 205 | this->pooling_height_ = pooling_height; |
| 206 | return *this; |
| 207 | } |
| 208 | |
| 209 | inline uint32_t pooling_height() const { |
| 210 | return this->pooling_height_; |
| 211 | } |
| 212 | |
| 213 | inline MaxPoolingOperatorTester& pooling_width(uint32_t pooling_width) { |
| 214 | assert(pooling_width >= 1); |
| 215 | this->pooling_width_ = pooling_width; |
| 216 | return *this; |
| 217 | } |
| 218 | |
| 219 | inline uint32_t pooling_width() const { |
| 220 | return this->pooling_width_; |
| 221 | } |
| 222 | |
| 223 | inline MaxPoolingOperatorTester& stride(uint32_t stride) { |
| 224 | assert(stride >= 1); |
| 225 | this->stride_height_ = stride; |
| 226 | this->stride_width_ = stride; |
| 227 | return *this; |
| 228 | } |
| 229 | |
| 230 | inline MaxPoolingOperatorTester& stride(uint32_t stride_height, uint32_t stride_width) { |
| 231 | assert(stride_height >= 1); |
| 232 | assert(stride_width >= 1); |
| 233 | this->stride_height_ = stride_height; |
| 234 | this->stride_width_ = stride_width; |
| 235 | return *this; |
| 236 | } |
| 237 | |
| 238 | inline MaxPoolingOperatorTester& stride_height(uint32_t stride_height) { |
| 239 | assert(stride_height >= 1); |
| 240 | this->stride_height_ = stride_height; |
| 241 | return *this; |
| 242 | } |
| 243 | |
| 244 | inline uint32_t stride_height() const { |
| 245 | return this->stride_height_; |
| 246 | } |
| 247 | |
| 248 | inline MaxPoolingOperatorTester& stride_width(uint32_t stride_width) { |
| 249 | assert(stride_width >= 1); |
| 250 | this->stride_width_ = stride_width; |
| 251 | return *this; |
| 252 | } |
| 253 | |
| 254 | inline uint32_t stride_width() const { |
| 255 | return this->stride_width_; |
| 256 | } |
| 257 | |
| 258 | inline MaxPoolingOperatorTester& dilation(uint32_t dilation) { |
| 259 | assert(dilation >= 1); |
| 260 | this->dilation_height_ = dilation; |
| 261 | this->dilation_width_ = dilation; |
| 262 | return *this; |
| 263 | } |
| 264 | |
| 265 | inline MaxPoolingOperatorTester& dilation(uint32_t dilation_height, uint32_t dilation_width) { |
| 266 | assert(dilation_height >= 1); |
| 267 | assert(dilation_width >= 1); |
| 268 | this->dilation_height_ = dilation_height; |
| 269 | this->dilation_width_ = dilation_width; |
| 270 | return *this; |
| 271 | } |
| 272 | |
| 273 | inline MaxPoolingOperatorTester& dilation_height(uint32_t dilation_height) { |
| 274 | assert(dilation_height >= 1); |
| 275 | this->dilation_height_ = dilation_height; |
| 276 | return *this; |
| 277 | } |
| 278 | |
| 279 | inline uint32_t dilation_height() const { |
| 280 | return this->dilation_height_; |
| 281 | } |
| 282 | |
| 283 | inline MaxPoolingOperatorTester& dilation_width(uint32_t dilation_width) { |
| 284 | assert(dilation_width >= 1); |
| 285 | this->dilation_width_ = dilation_width; |
| 286 | return *this; |
| 287 | } |
| 288 | |
| 289 | inline uint32_t dilation_width() const { |
| 290 | return this->dilation_width_; |
| 291 | } |
| 292 | |
| 293 | inline uint32_t dilated_pooling_height() const { |
| 294 | return (pooling_height() - 1) * dilation_height() + 1; |
| 295 | } |
| 296 | |
| 297 | inline uint32_t dilated_pooling_width() const { |
| 298 | return (pooling_width() - 1) * dilation_width() + 1; |
| 299 | } |
| 300 | |
| 301 | inline size_t output_height() const { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 302 | if (padding_tf_same()) { |
| 303 | return (input_height() + stride_height() - 1) / stride_height(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 304 | } else { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 305 | const size_t padded_input_height = padding_top() + input_height() + padding_bottom(); |
| 306 | if (padded_input_height <= dilated_pooling_height()) { |
| 307 | return 1; |
| 308 | } else { |
| 309 | return (padded_input_height - dilated_pooling_height()) / stride_height() + 1; |
| 310 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 311 | } |
| 312 | } |
| 313 | |
| 314 | inline size_t output_width() const { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 315 | if (padding_tf_same()) { |
| 316 | return (input_width() + stride_width() - 1) / stride_width(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 317 | } else { |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 318 | const size_t padded_input_width = padding_left() + input_width() + padding_right(); |
| 319 | if (padded_input_width <= dilated_pooling_width()) { |
| 320 | return 1; |
| 321 | } else { |
| 322 | return (padded_input_width - dilated_pooling_width()) / stride_width() + 1; |
| 323 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 324 | } |
| 325 | } |
| 326 | |
| 327 | inline MaxPoolingOperatorTester& input_pixel_stride(size_t input_pixel_stride) { |
| 328 | assert(input_pixel_stride != 0); |
| 329 | this->input_pixel_stride_ = input_pixel_stride; |
| 330 | return *this; |
| 331 | } |
| 332 | |
| 333 | inline size_t input_pixel_stride() const { |
| 334 | if (this->input_pixel_stride_ == 0) { |
| 335 | return channels(); |
| 336 | } else { |
| 337 | assert(this->input_pixel_stride_ >= channels()); |
| 338 | return this->input_pixel_stride_; |
| 339 | } |
| 340 | } |
| 341 | |
| 342 | inline MaxPoolingOperatorTester& output_pixel_stride(size_t output_pixel_stride) { |
| 343 | assert(output_pixel_stride != 0); |
| 344 | this->output_pixel_stride_ = output_pixel_stride; |
| 345 | return *this; |
| 346 | } |
| 347 | |
| 348 | inline size_t output_pixel_stride() const { |
| 349 | if (this->output_pixel_stride_ == 0) { |
| 350 | return channels(); |
| 351 | } else { |
| 352 | assert(this->output_pixel_stride_ >= channels()); |
| 353 | return this->output_pixel_stride_; |
| 354 | } |
| 355 | } |
| 356 | |
| 357 | inline MaxPoolingOperatorTester& next_input_size(uint32_t next_input_height, uint32_t next_input_width) { |
| 358 | assert(next_input_height >= 1); |
| 359 | assert(next_input_width >= 1); |
| 360 | this->next_input_height_ = next_input_height; |
| 361 | this->next_input_width_ = next_input_width; |
| 362 | return *this; |
| 363 | } |
| 364 | |
| 365 | inline MaxPoolingOperatorTester& next_input_height(uint32_t next_input_height) { |
| 366 | assert(next_input_height >= 1); |
| 367 | this->next_input_height_ = next_input_height; |
| 368 | return *this; |
| 369 | } |
| 370 | |
| 371 | inline uint32_t next_input_height() const { |
| 372 | if (this->next_input_height_ == 0) { |
| 373 | return input_height(); |
| 374 | } else { |
| 375 | return this->next_input_height_; |
| 376 | } |
| 377 | } |
| 378 | |
| 379 | inline MaxPoolingOperatorTester& next_input_width(uint32_t next_input_width) { |
| 380 | assert(next_input_width >= 1); |
| 381 | this->next_input_width_ = next_input_width; |
| 382 | return *this; |
| 383 | } |
| 384 | |
| 385 | inline uint32_t next_input_width() const { |
| 386 | if (this->next_input_width_ == 0) { |
| 387 | return input_width(); |
| 388 | } else { |
| 389 | return this->next_input_width_; |
| 390 | } |
| 391 | } |
| 392 | |
| 393 | inline size_t next_output_height() const { |
| 394 | const size_t padded_next_input_height = padding_top() + next_input_height() + padding_bottom(); |
| 395 | if (padded_next_input_height <= dilated_pooling_height()) { |
| 396 | return 1; |
| 397 | } else { |
| 398 | return (padded_next_input_height - dilated_pooling_height()) / stride_height() + 1; |
| 399 | } |
| 400 | } |
| 401 | |
| 402 | inline size_t next_output_width() const { |
| 403 | const size_t padded_next_input_width = padding_left() + next_input_width() + padding_right(); |
| 404 | if (padded_next_input_width <= dilated_pooling_width()) { |
| 405 | return 1; |
| 406 | } else { |
| 407 | return (padded_next_input_width - dilated_pooling_width()) / stride_width() + 1; |
| 408 | } |
| 409 | } |
| 410 | |
| 411 | inline MaxPoolingOperatorTester& next_batch_size(size_t next_batch_size) { |
| 412 | assert(next_batch_size >= 1); |
| 413 | this->next_batch_size_ = next_batch_size; |
| 414 | return *this; |
| 415 | } |
| 416 | |
| 417 | inline size_t next_batch_size() const { |
| 418 | if (this->next_batch_size_ == 0) { |
| 419 | return batch_size(); |
| 420 | } else { |
| 421 | return this->next_batch_size_; |
| 422 | } |
| 423 | } |
| 424 | |
| 425 | inline MaxPoolingOperatorTester& qmin(uint8_t qmin) { |
| 426 | this->qmin_ = qmin; |
| 427 | return *this; |
| 428 | } |
| 429 | |
| 430 | inline uint8_t qmin() const { |
| 431 | return this->qmin_; |
| 432 | } |
| 433 | |
| 434 | inline MaxPoolingOperatorTester& qmax(uint8_t qmax) { |
| 435 | this->qmax_ = qmax; |
| 436 | return *this; |
| 437 | } |
| 438 | |
| 439 | inline uint8_t qmax() const { |
| 440 | return this->qmax_; |
| 441 | } |
| 442 | |
| 443 | inline MaxPoolingOperatorTester& iterations(size_t iterations) { |
| 444 | this->iterations_ = iterations; |
| 445 | return *this; |
| 446 | } |
| 447 | |
| 448 | inline size_t iterations() const { |
| 449 | return this->iterations_; |
| 450 | } |
| 451 | |
Marat Dukhan | dc5c148 | 2021-08-16 09:03:15 -0700 | [diff] [blame] | 452 | void TestS8() const { |
| 453 | std::random_device random_device; |
| 454 | auto rng = std::mt19937(random_device()); |
| 455 | auto i8rng = std::bind( |
| 456 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), |
| 457 | std::ref(rng)); |
| 458 | |
| 459 | std::vector<int8_t> input((batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(int8_t)); |
| 460 | std::vector<int8_t> output((batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(int8_t)); |
| 461 | std::vector<int8_t> output_ref(batch_size() * output_height() * output_width() * channels()); |
| 462 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 463 | std::generate(input.begin(), input.end(), std::ref(i8rng)); |
| 464 | std::fill(output.begin(), output.end(), 0xA5); |
| 465 | |
| 466 | // Compute reference results. |
| 467 | for (size_t i = 0; i < batch_size(); i++) { |
| 468 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 469 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 470 | for (size_t c = 0; c < channels(); c++) { |
| 471 | int8_t max_value = std::numeric_limits<int8_t>::min(); |
| 472 | for (size_t py = 0; py < pooling_height(); py++) { |
| 473 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 474 | for (size_t px = 0; px < pooling_width(); px++) { |
| 475 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 476 | if (ix < input_width() && iy < input_height()) { |
| 477 | max_value = std::max(max_value, |
| 478 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c]); |
| 479 | } |
| 480 | } |
| 481 | } |
| 482 | max_value = std::min(max_value, int8_t(qmax() - 0x80)); |
| 483 | max_value = std::max(max_value, int8_t(qmin() - 0x80)); |
| 484 | output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = max_value; |
| 485 | } |
| 486 | } |
| 487 | } |
| 488 | } |
| 489 | |
| 490 | // Create, setup, run, and destroy Max Pooling operator. |
| 491 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 492 | xnn_operator_t max_pooling_op = nullptr; |
| 493 | |
| 494 | ASSERT_EQ(xnn_status_success, |
| 495 | xnn_create_max_pooling2d_nhwc_s8( |
| 496 | padding_tf_same() ? 0 : padding_top(), padding_tf_same() ? 0 : padding_right(), |
| 497 | padding_tf_same() ? 0 : padding_bottom(), padding_tf_same() ? 0 : padding_left(), |
| 498 | pooling_height(), pooling_width(), |
| 499 | stride_height(), stride_width(), |
| 500 | dilation_height(), dilation_width(), |
| 501 | channels(), input_pixel_stride(), output_pixel_stride(), |
| 502 | int8_t(qmin() - 0x80), int8_t(qmax() - 0x80), |
| 503 | padding_tf_same() ? XNN_FLAG_TENSORFLOW_SAME_PADDING : 0, |
| 504 | &max_pooling_op)); |
| 505 | ASSERT_NE(nullptr, max_pooling_op); |
| 506 | |
| 507 | // Smart pointer to automatically delete max_pooling_op. |
| 508 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_max_pooling_op(max_pooling_op, xnn_delete_operator); |
| 509 | |
| 510 | ASSERT_EQ(xnn_status_success, |
| 511 | xnn_setup_max_pooling2d_nhwc_s8( |
| 512 | max_pooling_op, |
| 513 | batch_size(), input_height(), input_width(), |
| 514 | input.data(), output.data(), |
| 515 | nullptr /* thread pool */)); |
| 516 | |
| 517 | ASSERT_EQ(xnn_status_success, |
| 518 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 519 | |
| 520 | // Verify results. |
| 521 | for (size_t i = 0; i < batch_size(); i++) { |
| 522 | for (size_t y = 0; y < output_height(); y++) { |
| 523 | for (size_t x = 0; x < output_width(); x++) { |
| 524 | for (size_t c = 0; c < channels(); c++) { |
| 525 | ASSERT_LE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), int32_t(qmax() - 0x80)); |
| 526 | ASSERT_GE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), int32_t(qmin() - 0x80)); |
| 527 | ASSERT_EQ(int32_t(output_ref[((i * output_height() + y) * output_width() + x) * channels() + c]), |
| 528 | int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c])) << |
| 529 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c; |
| 530 | } |
| 531 | } |
| 532 | } |
| 533 | } |
| 534 | } |
| 535 | } |
| 536 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 537 | void TestU8() const { |
| 538 | std::random_device random_device; |
| 539 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 540 | 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] | 541 | |
| 542 | std::vector<uint8_t> input((batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
| 543 | std::vector<uint8_t> output((batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
| 544 | std::vector<uint8_t> output_ref(batch_size() * output_height() * output_width() * channels()); |
| 545 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 546 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 547 | std::fill(output.begin(), output.end(), 0xA5); |
| 548 | |
| 549 | // Compute reference results. |
| 550 | for (size_t i = 0; i < batch_size(); i++) { |
| 551 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 552 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 553 | for (size_t c = 0; c < channels(); c++) { |
| 554 | uint8_t max_value = 0; |
| 555 | for (size_t py = 0; py < pooling_height(); py++) { |
| 556 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 557 | for (size_t px = 0; px < pooling_width(); px++) { |
| 558 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 559 | if (ix < input_width() && iy < input_height()) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 560 | max_value = std::max(max_value, |
| 561 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c]); |
| 562 | } |
| 563 | } |
| 564 | } |
| 565 | max_value = std::min(max_value, qmax()); |
| 566 | max_value = std::max(max_value, qmin()); |
| 567 | output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = max_value; |
| 568 | } |
| 569 | } |
| 570 | } |
| 571 | } |
| 572 | |
| 573 | // Create, setup, run, and destroy Max Pooling operator. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 574 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 575 | xnn_operator_t max_pooling_op = nullptr; |
| 576 | |
| 577 | ASSERT_EQ(xnn_status_success, |
| 578 | xnn_create_max_pooling2d_nhwc_u8( |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 579 | padding_tf_same() ? 0 : padding_top(), padding_tf_same() ? 0 : padding_right(), |
| 580 | padding_tf_same() ? 0 : padding_bottom(), padding_tf_same() ? 0 : padding_left(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 581 | pooling_height(), pooling_width(), |
| 582 | stride_height(), stride_width(), |
| 583 | dilation_height(), dilation_width(), |
| 584 | channels(), input_pixel_stride(), output_pixel_stride(), |
| 585 | qmin(), qmax(), |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 586 | padding_tf_same() ? XNN_FLAG_TENSORFLOW_SAME_PADDING : 0, |
| 587 | &max_pooling_op)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 588 | ASSERT_NE(nullptr, max_pooling_op); |
| 589 | |
| 590 | // Smart pointer to automatically delete max_pooling_op. |
| 591 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_max_pooling_op(max_pooling_op, xnn_delete_operator); |
| 592 | |
| 593 | ASSERT_EQ(xnn_status_success, |
| 594 | xnn_setup_max_pooling2d_nhwc_u8( |
| 595 | max_pooling_op, |
| 596 | batch_size(), input_height(), input_width(), |
| 597 | input.data(), output.data(), |
| 598 | nullptr /* thread pool */)); |
| 599 | |
| 600 | ASSERT_EQ(xnn_status_success, |
| 601 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 602 | |
| 603 | // Verify results. |
| 604 | for (size_t i = 0; i < batch_size(); i++) { |
| 605 | for (size_t y = 0; y < output_height(); y++) { |
| 606 | for (size_t x = 0; x < output_width(); x++) { |
| 607 | for (size_t c = 0; c < channels(); c++) { |
| 608 | ASSERT_LE(uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), uint32_t(qmax())); |
| 609 | ASSERT_GE(uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), uint32_t(qmin())); |
| 610 | ASSERT_EQ(uint32_t(output_ref[((i * output_height() + y) * output_width() + x) * channels() + c]), |
| 611 | uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c])) << |
| 612 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c; |
| 613 | } |
| 614 | } |
| 615 | } |
| 616 | } |
| 617 | } |
| 618 | } |
| 619 | |
Marat Dukhan | 5756a92 | 2022-02-04 01:55:53 -0800 | [diff] [blame] | 620 | void TestF16() const { |
| 621 | std::random_device random_device; |
| 622 | auto rng = std::mt19937(random_device()); |
| 623 | // Note: we need to avoid FP16 denormals in the generated tensor because they might be processed differently in |
| 624 | // native vs emulated arithmetics, and we use exact comparison to verify the results against reference. |
| 625 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.001f, 1.0f), rng); |
| 626 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 627 | |
| 628 | std::vector<uint16_t> input((batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 629 | std::vector<uint16_t> output((batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 630 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels()); |
| 631 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 632 | std::generate(input.begin(), input.end(), std::ref(f16rng)); |
| 633 | std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */); |
| 634 | |
| 635 | // Compute reference results, without clamping. |
| 636 | for (size_t i = 0; i < batch_size(); i++) { |
| 637 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 638 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 639 | for (size_t c = 0; c < channels(); c++) { |
| 640 | float max_value = -std::numeric_limits<float>::infinity(); |
| 641 | for (size_t py = 0; py < pooling_height(); py++) { |
| 642 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 643 | for (size_t px = 0; px < pooling_width(); px++) { |
| 644 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 645 | if (ix < input_width() && iy < input_height()) { |
| 646 | max_value = std::max(max_value, |
| 647 | fp16_ieee_to_fp32_value(input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c])); |
| 648 | } |
| 649 | } |
| 650 | } |
| 651 | output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = max_value; |
| 652 | } |
| 653 | } |
| 654 | } |
| 655 | } |
| 656 | |
| 657 | // Compute clamping parameters. |
| 658 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 659 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 660 | const float accumulated_range = accumulated_max - accumulated_min; |
| 661 | float output_min = accumulated_min + accumulated_range / 255.0f * float(qmin()); |
| 662 | float output_max = accumulated_max - accumulated_range / 255.0f * float(255 - qmax()); |
| 663 | output_min = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(output_min)); |
| 664 | output_max = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(output_max)); |
| 665 | if (accumulated_range == 0.0f) { |
| 666 | output_min = -std::numeric_limits<float>::infinity(); |
| 667 | output_max = +std::numeric_limits<float>::infinity(); |
| 668 | } |
| 669 | if (qmin() == std::numeric_limits<uint8_t>::min()) { |
| 670 | output_min = -std::numeric_limits<float>::infinity(); |
| 671 | } |
| 672 | if (qmax() == std::numeric_limits<uint8_t>::max()) { |
| 673 | output_max = +std::numeric_limits<float>::infinity(); |
| 674 | } |
| 675 | |
| 676 | // Clamp reference results. |
| 677 | for (float& value : output_ref) { |
| 678 | value = std::max(std::min(value, output_max), output_min); |
| 679 | } |
| 680 | |
| 681 | // Create, setup, run, and destroy Max Pooling operator. |
| 682 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 683 | xnn_operator_t max_pooling_op = nullptr; |
| 684 | |
| 685 | const xnn_status status = xnn_create_max_pooling2d_nhwc_f16( |
| 686 | padding_tf_same() ? 0 : padding_top(), padding_tf_same() ? 0 : padding_right(), |
| 687 | padding_tf_same() ? 0 : padding_bottom(), padding_tf_same() ? 0 : padding_left(), |
| 688 | pooling_height(), pooling_width(), |
| 689 | stride_height(), stride_width(), |
| 690 | dilation_height(), dilation_width(), |
| 691 | channels(), input_pixel_stride(), output_pixel_stride(), |
| 692 | output_min, output_max, |
| 693 | padding_tf_same() ? XNN_FLAG_TENSORFLOW_SAME_PADDING : 0, |
| 694 | &max_pooling_op); |
| 695 | if (status == xnn_status_unsupported_hardware) { |
| 696 | GTEST_SKIP(); |
| 697 | } |
| 698 | ASSERT_EQ(xnn_status_success, status); |
| 699 | ASSERT_NE(nullptr, max_pooling_op); |
| 700 | |
| 701 | // Smart pointer to automatically delete max_pooling_op. |
| 702 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_max_pooling_op(max_pooling_op, xnn_delete_operator); |
| 703 | |
| 704 | ASSERT_EQ(xnn_status_success, |
| 705 | xnn_setup_max_pooling2d_nhwc_f16( |
| 706 | max_pooling_op, |
| 707 | batch_size(), input_height(), input_width(), |
| 708 | input.data(), output.data(), |
| 709 | nullptr /* thread pool */)); |
| 710 | |
| 711 | ASSERT_EQ(xnn_status_success, |
| 712 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 713 | |
| 714 | // Verify results. |
| 715 | for (size_t i = 0; i < batch_size(); i++) { |
| 716 | for (size_t y = 0; y < output_height(); y++) { |
| 717 | for (size_t x = 0; x < output_width(); x++) { |
| 718 | for (size_t c = 0; c < channels(); c++) { |
| 719 | ASSERT_LE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), output_max); |
| 720 | ASSERT_GE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), output_min); |
| 721 | ASSERT_EQ( |
| 722 | fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), |
| 723 | output_ref[((i * output_height() + y) * output_width() + x) * channels() + c]) << |
| 724 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c |
| 725 | << ", min = " << output_min << ", max = " << output_max; |
| 726 | } |
| 727 | } |
| 728 | } |
| 729 | } |
| 730 | } |
| 731 | } |
| 732 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 733 | void TestF32() const { |
| 734 | std::random_device random_device; |
| 735 | auto rng = std::mt19937(random_device()); |
| 736 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 737 | |
| 738 | std::vector<float> input((batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(float)); |
| 739 | std::vector<float> output((batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(float)); |
| 740 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels()); |
| 741 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 742 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 743 | std::fill(output.begin(), output.end(), nanf("")); |
| 744 | |
| 745 | // Compute reference results, without clamping. |
| 746 | for (size_t i = 0; i < batch_size(); i++) { |
| 747 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 748 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 749 | for (size_t c = 0; c < channels(); c++) { |
| 750 | float max_value = -std::numeric_limits<float>::infinity(); |
| 751 | for (size_t py = 0; py < pooling_height(); py++) { |
| 752 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 753 | for (size_t px = 0; px < pooling_width(); px++) { |
| 754 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 755 | if (ix < input_width() && iy < input_height()) { |
| 756 | max_value = std::max(max_value, |
| 757 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c]); |
| 758 | } |
| 759 | } |
| 760 | } |
| 761 | output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = max_value; |
| 762 | } |
| 763 | } |
| 764 | } |
| 765 | } |
| 766 | |
| 767 | // Compute clamping parameters. |
| 768 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 769 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 770 | const float accumulated_range = accumulated_max - accumulated_min; |
| 771 | const float output_min = accumulated_range == 0.0f ? |
| 772 | -std::numeric_limits<float>::infinity() : |
| 773 | accumulated_min + accumulated_range / 255.0f * float(qmin()); |
| 774 | const float output_max = accumulated_range == 0.0f ? |
| 775 | +std::numeric_limits<float>::infinity() : |
| 776 | accumulated_max - accumulated_range / 255.0f * float(255 - qmax()); |
| 777 | |
| 778 | // Clamp reference results. |
| 779 | for (float& value : output_ref) { |
| 780 | value = std::max(std::min(value, output_max), output_min); |
| 781 | } |
| 782 | |
| 783 | // Create, setup, run, and destroy Max Pooling operator. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 784 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 785 | xnn_operator_t max_pooling_op = nullptr; |
| 786 | |
| 787 | ASSERT_EQ(xnn_status_success, |
| 788 | xnn_create_max_pooling2d_nhwc_f32( |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 789 | padding_tf_same() ? 0 : padding_top(), padding_tf_same() ? 0 : padding_right(), |
| 790 | padding_tf_same() ? 0 : padding_bottom(), padding_tf_same() ? 0 : padding_left(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 791 | pooling_height(), pooling_width(), |
| 792 | stride_height(), stride_width(), |
| 793 | dilation_height(), dilation_width(), |
| 794 | channels(), input_pixel_stride(), output_pixel_stride(), |
| 795 | output_min, output_max, |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 796 | padding_tf_same() ? XNN_FLAG_TENSORFLOW_SAME_PADDING : 0, |
| 797 | &max_pooling_op)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 798 | ASSERT_NE(nullptr, max_pooling_op); |
| 799 | |
| 800 | // Smart pointer to automatically delete max_pooling_op. |
| 801 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_max_pooling_op(max_pooling_op, xnn_delete_operator); |
| 802 | |
| 803 | ASSERT_EQ(xnn_status_success, |
| 804 | xnn_setup_max_pooling2d_nhwc_f32( |
| 805 | max_pooling_op, |
| 806 | batch_size(), input_height(), input_width(), |
| 807 | input.data(), output.data(), |
| 808 | nullptr /* thread pool */)); |
| 809 | |
| 810 | ASSERT_EQ(xnn_status_success, |
| 811 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 812 | |
| 813 | // Verify results. |
| 814 | for (size_t i = 0; i < batch_size(); i++) { |
| 815 | for (size_t y = 0; y < output_height(); y++) { |
| 816 | for (size_t x = 0; x < output_width(); x++) { |
| 817 | for (size_t c = 0; c < channels(); c++) { |
| 818 | ASSERT_LE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_max); |
| 819 | ASSERT_GE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_min); |
| 820 | ASSERT_EQ(output_ref[((i * output_height() + y) * output_width() + x) * channels() + c], |
| 821 | output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]) << |
| 822 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c |
| 823 | << ", min = " << output_min << ", max = " << output_max; |
| 824 | } |
| 825 | } |
| 826 | } |
| 827 | } |
| 828 | } |
| 829 | } |
| 830 | |
Marat Dukhan | dc5c148 | 2021-08-16 09:03:15 -0700 | [diff] [blame] | 831 | void TestSetupS8() const { |
| 832 | std::random_device random_device; |
| 833 | auto rng = std::mt19937(random_device()); |
| 834 | auto i8rng = std::bind( |
| 835 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), |
| 836 | std::ref(rng)); |
| 837 | |
| 838 | std::vector<int8_t> input(XNN_EXTRA_BYTES / sizeof(int8_t) + std::max( |
| 839 | (batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels(), |
| 840 | (next_batch_size() * next_input_height() * next_input_width() - 1) * input_pixel_stride() + channels())); |
| 841 | std::vector<int8_t> output(XNN_EXTRA_BYTES / sizeof(int8_t) + std::max( |
| 842 | (batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels(), |
| 843 | (next_batch_size() * next_output_height() * next_output_width() - 1) * output_pixel_stride() + channels())); |
| 844 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels()); |
| 845 | std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * channels()); |
| 846 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 847 | std::generate(input.begin(), input.end(), std::ref(i8rng)); |
| 848 | std::fill(output.begin(), output.end(), 0xA5); |
| 849 | |
| 850 | // Compute reference results. |
| 851 | for (size_t i = 0; i < batch_size(); i++) { |
| 852 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 853 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 854 | for (size_t c = 0; c < channels(); c++) { |
| 855 | int8_t max_value = std::numeric_limits<int8_t>::min(); |
| 856 | for (size_t py = 0; py < pooling_height(); py++) { |
| 857 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 858 | for (size_t px = 0; px < pooling_width(); px++) { |
| 859 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 860 | if (ix < input_width() && iy < input_height()) { |
| 861 | max_value = std::max(max_value, |
| 862 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c]); |
| 863 | } |
| 864 | } |
| 865 | } |
| 866 | max_value = std::min(max_value, int8_t(qmax() - 0x80)); |
| 867 | max_value = std::max(max_value, int8_t(qmin() - 0x80)); |
| 868 | output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = max_value; |
| 869 | } |
| 870 | } |
| 871 | } |
| 872 | } |
| 873 | |
| 874 | // Create, setup, and run Max Pooling operator once. |
| 875 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 876 | xnn_operator_t max_pooling_op = nullptr; |
| 877 | |
| 878 | ASSERT_EQ(xnn_status_success, |
| 879 | xnn_create_max_pooling2d_nhwc_s8( |
| 880 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 881 | pooling_height(), pooling_width(), |
| 882 | stride_height(), stride_width(), |
| 883 | dilation_height(), dilation_width(), |
| 884 | channels(), input_pixel_stride(), output_pixel_stride(), |
| 885 | int8_t(qmin() - 0x80), int8_t(qmax() - 0x80), |
| 886 | 0, &max_pooling_op)); |
| 887 | ASSERT_NE(nullptr, max_pooling_op); |
| 888 | |
| 889 | // Smart pointer to automatically delete max_pooling_op. |
| 890 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_max_pooling_op(max_pooling_op, xnn_delete_operator); |
| 891 | |
| 892 | ASSERT_EQ(xnn_status_success, |
| 893 | xnn_setup_max_pooling2d_nhwc_s8( |
| 894 | max_pooling_op, |
| 895 | batch_size(), input_height(), input_width(), |
| 896 | input.data(), output.data(), |
| 897 | nullptr /* thread pool */)); |
| 898 | |
| 899 | ASSERT_EQ(xnn_status_success, |
| 900 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 901 | |
| 902 | // Verify results of the first run. |
| 903 | for (size_t i = 0; i < batch_size(); i++) { |
| 904 | for (size_t y = 0; y < output_height(); y++) { |
| 905 | for (size_t x = 0; x < output_width(); x++) { |
| 906 | for (size_t c = 0; c < channels(); c++) { |
| 907 | ASSERT_LE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), int32_t(qmax() - 0x80)); |
| 908 | ASSERT_GE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), int32_t(qmin() - 0x80)); |
| 909 | ASSERT_EQ(int32_t(output_ref[((i * output_height() + y) * output_width() + x) * channels() + c]), |
| 910 | int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c])) << |
| 911 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c; |
| 912 | } |
| 913 | } |
| 914 | } |
| 915 | } |
| 916 | |
| 917 | // Re-generate data for the second run. |
| 918 | std::generate(input.begin(), input.end(), std::ref(i8rng)); |
| 919 | std::fill(output.begin(), output.end(), 0xA5); |
| 920 | |
| 921 | // Compute reference results for the second run. |
| 922 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 923 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 924 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 925 | for (size_t c = 0; c < channels(); c++) { |
| 926 | int8_t max_value = std::numeric_limits<int8_t>::min(); |
| 927 | for (size_t py = 0; py < pooling_height(); py++) { |
| 928 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 929 | for (size_t px = 0; px < pooling_width(); px++) { |
| 930 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 931 | if (ix < next_input_width() && iy < next_input_height()) { |
| 932 | max_value = std::max(max_value, |
| 933 | input[((i * next_input_height() + iy) * next_input_width() + ix) * input_pixel_stride() + c]); |
| 934 | } |
| 935 | } |
| 936 | } |
| 937 | max_value = std::min(max_value, int8_t(qmax() - 0x80)); |
| 938 | max_value = std::max(max_value, int8_t(qmin() - 0x80)); |
| 939 | next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c] = max_value; |
| 940 | } |
| 941 | } |
| 942 | } |
| 943 | } |
| 944 | |
| 945 | // Setup and run Max Pooling operator the second time, and destroy the operator. |
| 946 | ASSERT_EQ(xnn_status_success, |
| 947 | xnn_setup_max_pooling2d_nhwc_s8( |
| 948 | max_pooling_op, |
| 949 | next_batch_size(), next_input_height(), next_input_width(), |
| 950 | input.data(), output.data(), |
| 951 | nullptr /* thread pool */)); |
| 952 | |
| 953 | ASSERT_EQ(xnn_status_success, |
| 954 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 955 | |
| 956 | // Verify results of the second run. |
| 957 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 958 | for (size_t y = 0; y < next_output_height(); y++) { |
| 959 | for (size_t x = 0; x < next_output_width(); x++) { |
| 960 | for (size_t c = 0; c < channels(); c++) { |
| 961 | ASSERT_LE(int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]), int32_t(qmax() - 0x80)); |
| 962 | ASSERT_GE(int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]), int32_t(qmin() - 0x80)); |
| 963 | ASSERT_EQ(int32_t(next_output_ref[((i * next_output_height() + y) * next_output_width() + x) * channels() + c]), |
| 964 | int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c])) << |
| 965 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c; |
| 966 | } |
| 967 | } |
| 968 | } |
| 969 | } |
| 970 | } |
| 971 | } |
| 972 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 973 | void TestSetupU8() const { |
| 974 | std::random_device random_device; |
| 975 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 976 | 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] | 977 | |
| 978 | std::vector<uint8_t> input(XNN_EXTRA_BYTES / sizeof(uint8_t) + std::max( |
| 979 | (batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels(), |
| 980 | (next_batch_size() * next_input_height() * next_input_width() - 1) * input_pixel_stride() + channels())); |
| 981 | std::vector<uint8_t> output(XNN_EXTRA_BYTES / sizeof(uint8_t) + std::max( |
| 982 | (batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels(), |
| 983 | (next_batch_size() * next_output_height() * next_output_width() - 1) * output_pixel_stride() + channels())); |
| 984 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels()); |
| 985 | std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * channels()); |
| 986 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 987 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 988 | std::fill(output.begin(), output.end(), 0xA5); |
| 989 | |
| 990 | // Compute reference results. |
| 991 | for (size_t i = 0; i < batch_size(); i++) { |
| 992 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 993 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 994 | for (size_t c = 0; c < channels(); c++) { |
| 995 | uint8_t max_value = 0; |
| 996 | for (size_t py = 0; py < pooling_height(); py++) { |
| 997 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 998 | for (size_t px = 0; px < pooling_width(); px++) { |
| 999 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 1000 | if (ix < input_width() && iy < input_height()) { |
| 1001 | max_value = std::max(max_value, |
| 1002 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c]); |
| 1003 | } |
| 1004 | } |
| 1005 | } |
| 1006 | max_value = std::min(max_value, qmax()); |
| 1007 | max_value = std::max(max_value, qmin()); |
| 1008 | output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = max_value; |
| 1009 | } |
| 1010 | } |
| 1011 | } |
| 1012 | } |
| 1013 | |
| 1014 | // Create, setup, and run Max Pooling operator once. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1015 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1016 | xnn_operator_t max_pooling_op = nullptr; |
| 1017 | |
| 1018 | ASSERT_EQ(xnn_status_success, |
| 1019 | xnn_create_max_pooling2d_nhwc_u8( |
| 1020 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 1021 | pooling_height(), pooling_width(), |
| 1022 | stride_height(), stride_width(), |
| 1023 | dilation_height(), dilation_width(), |
| 1024 | channels(), input_pixel_stride(), output_pixel_stride(), |
| 1025 | qmin(), qmax(), |
| 1026 | 0, &max_pooling_op)); |
| 1027 | ASSERT_NE(nullptr, max_pooling_op); |
| 1028 | |
| 1029 | // Smart pointer to automatically delete max_pooling_op. |
| 1030 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_max_pooling_op(max_pooling_op, xnn_delete_operator); |
| 1031 | |
| 1032 | ASSERT_EQ(xnn_status_success, |
| 1033 | xnn_setup_max_pooling2d_nhwc_u8( |
| 1034 | max_pooling_op, |
| 1035 | batch_size(), input_height(), input_width(), |
| 1036 | input.data(), output.data(), |
| 1037 | nullptr /* thread pool */)); |
| 1038 | |
| 1039 | ASSERT_EQ(xnn_status_success, |
| 1040 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 1041 | |
| 1042 | // Verify results of the first run. |
| 1043 | for (size_t i = 0; i < batch_size(); i++) { |
| 1044 | for (size_t y = 0; y < output_height(); y++) { |
| 1045 | for (size_t x = 0; x < output_width(); x++) { |
| 1046 | for (size_t c = 0; c < channels(); c++) { |
| 1047 | ASSERT_LE(uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), uint32_t(qmax())); |
| 1048 | ASSERT_GE(uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), uint32_t(qmin())); |
| 1049 | ASSERT_EQ(uint32_t(output_ref[((i * output_height() + y) * output_width() + x) * channels() + c]), |
| 1050 | uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c])) << |
| 1051 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c; |
| 1052 | } |
| 1053 | } |
| 1054 | } |
| 1055 | } |
| 1056 | |
| 1057 | // Re-generate data for the second run. |
| 1058 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 1059 | std::fill(output.begin(), output.end(), 0xA5); |
| 1060 | |
| 1061 | // Compute reference results for the second run. |
| 1062 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1063 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1064 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1065 | for (size_t c = 0; c < channels(); c++) { |
| 1066 | uint8_t max_value = 0; |
| 1067 | for (size_t py = 0; py < pooling_height(); py++) { |
| 1068 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 1069 | for (size_t px = 0; px < pooling_width(); px++) { |
| 1070 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 1071 | if (ix < next_input_width() && iy < next_input_height()) { |
| 1072 | max_value = std::max(max_value, |
| 1073 | input[((i * next_input_height() + iy) * next_input_width() + ix) * input_pixel_stride() + c]); |
| 1074 | } |
| 1075 | } |
| 1076 | } |
| 1077 | max_value = std::min(max_value, qmax()); |
| 1078 | max_value = std::max(max_value, qmin()); |
| 1079 | next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c] = max_value; |
| 1080 | } |
| 1081 | } |
| 1082 | } |
| 1083 | } |
| 1084 | |
| 1085 | // Setup and run Max Pooling operator the second time, and destroy the operator. |
| 1086 | ASSERT_EQ(xnn_status_success, |
| 1087 | xnn_setup_max_pooling2d_nhwc_u8( |
| 1088 | max_pooling_op, |
| 1089 | next_batch_size(), next_input_height(), next_input_width(), |
| 1090 | input.data(), output.data(), |
| 1091 | nullptr /* thread pool */)); |
| 1092 | |
| 1093 | ASSERT_EQ(xnn_status_success, |
| 1094 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 1095 | |
| 1096 | // Verify results of the second run. |
| 1097 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1098 | for (size_t y = 0; y < next_output_height(); y++) { |
| 1099 | for (size_t x = 0; x < next_output_width(); x++) { |
| 1100 | for (size_t c = 0; c < channels(); c++) { |
| 1101 | ASSERT_LE(uint32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]), uint32_t(qmax())); |
| 1102 | ASSERT_GE(uint32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]), uint32_t(qmin())); |
| 1103 | ASSERT_EQ(uint32_t(next_output_ref[((i * next_output_height() + y) * next_output_width() + x) * channels() + c]), |
| 1104 | uint32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c])) << |
| 1105 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c; |
| 1106 | } |
| 1107 | } |
| 1108 | } |
| 1109 | } |
| 1110 | } |
| 1111 | } |
| 1112 | |
Marat Dukhan | 5756a92 | 2022-02-04 01:55:53 -0800 | [diff] [blame] | 1113 | void TestSetupF16() const { |
| 1114 | std::random_device random_device; |
| 1115 | auto rng = std::mt19937(random_device()); |
| 1116 | // Note: we need to avoid FP16 denormals in the generated tensor because they might be processed differently in |
| 1117 | // native vs emulated arithmetics, and we use exact comparison to verify the results against reference. |
| 1118 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.001f, 1.0f), rng); |
| 1119 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 1120 | |
| 1121 | std::vector<uint16_t> input(XNN_EXTRA_BYTES / sizeof(uint16_t) + std::max( |
| 1122 | (batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels(), |
| 1123 | (next_batch_size() * next_input_height() * next_input_width() - 1) * input_pixel_stride() + channels())); |
| 1124 | std::vector<uint16_t> output(XNN_EXTRA_BYTES / sizeof(uint16_t) + std::max( |
| 1125 | (batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels(), |
| 1126 | (next_batch_size() * next_output_height() * next_output_width() - 1) * output_pixel_stride() + channels())); |
| 1127 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels()); |
| 1128 | std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * channels()); |
| 1129 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 1130 | std::generate(input.begin(), input.end(), std::ref(f16rng)); |
| 1131 | std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */); |
| 1132 | |
| 1133 | // Compute reference results, without clamping. |
| 1134 | for (size_t i = 0; i < batch_size(); i++) { |
| 1135 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1136 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1137 | for (size_t c = 0; c < channels(); c++) { |
| 1138 | float max_value = -std::numeric_limits<float>::infinity(); |
| 1139 | for (size_t py = 0; py < pooling_height(); py++) { |
| 1140 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 1141 | for (size_t px = 0; px < pooling_width(); px++) { |
| 1142 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 1143 | if (ix < input_width() && iy < input_height()) { |
| 1144 | max_value = std::max(max_value, |
| 1145 | fp16_ieee_to_fp32_value(input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c])); |
| 1146 | } |
| 1147 | } |
| 1148 | } |
| 1149 | output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = max_value; |
| 1150 | } |
| 1151 | } |
| 1152 | } |
| 1153 | } |
| 1154 | |
| 1155 | // Compute clamping parameters. |
| 1156 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 1157 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 1158 | const float accumulated_range = accumulated_max - accumulated_min; |
| 1159 | float output_min = accumulated_min + accumulated_range / 255.0f * float(qmin()); |
| 1160 | float output_max = accumulated_max - accumulated_range / 255.0f * float(255 - qmax()); |
| 1161 | output_min = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(output_min)); |
| 1162 | output_max = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(output_max)); |
| 1163 | if (accumulated_range == 0.0f) { |
| 1164 | output_min = -std::numeric_limits<float>::infinity(); |
| 1165 | output_max = +std::numeric_limits<float>::infinity(); |
| 1166 | } |
| 1167 | if (qmin() == std::numeric_limits<uint8_t>::min()) { |
| 1168 | output_min = -std::numeric_limits<float>::infinity(); |
| 1169 | } |
| 1170 | if (qmax() == std::numeric_limits<uint8_t>::max()) { |
| 1171 | output_max = +std::numeric_limits<float>::infinity(); |
| 1172 | } |
| 1173 | |
| 1174 | // Clamp reference results. |
| 1175 | for (float& value : output_ref) { |
| 1176 | value = std::max(std::min(value, output_max), output_min); |
| 1177 | } |
| 1178 | |
| 1179 | // Create, setup, and run Max Pooling operator once. |
| 1180 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
| 1181 | xnn_operator_t max_pooling_op = nullptr; |
| 1182 | |
| 1183 | const xnn_status status = xnn_create_max_pooling2d_nhwc_f16( |
| 1184 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 1185 | pooling_height(), pooling_width(), |
| 1186 | stride_height(), stride_width(), |
| 1187 | dilation_height(), dilation_width(), |
| 1188 | channels(), input_pixel_stride(), output_pixel_stride(), |
| 1189 | output_min, output_max, |
| 1190 | 0, &max_pooling_op); |
| 1191 | if (status == xnn_status_unsupported_hardware) { |
| 1192 | GTEST_SKIP(); |
| 1193 | } |
| 1194 | ASSERT_EQ(xnn_status_success, status); |
| 1195 | ASSERT_NE(nullptr, max_pooling_op); |
| 1196 | |
| 1197 | // Smart pointer to automatically delete max_pooling_op. |
| 1198 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_max_pooling_op(max_pooling_op, xnn_delete_operator); |
| 1199 | |
| 1200 | ASSERT_EQ(xnn_status_success, |
| 1201 | xnn_setup_max_pooling2d_nhwc_f16( |
| 1202 | max_pooling_op, |
| 1203 | batch_size(), input_height(), input_width(), |
| 1204 | input.data(), output.data(), |
| 1205 | nullptr /* thread pool */)); |
| 1206 | |
| 1207 | ASSERT_EQ(xnn_status_success, |
| 1208 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 1209 | |
| 1210 | // Verify results of the first run. |
| 1211 | for (size_t i = 0; i < batch_size(); i++) { |
| 1212 | for (size_t y = 0; y < output_height(); y++) { |
| 1213 | for (size_t x = 0; x < output_width(); x++) { |
| 1214 | for (size_t c = 0; c < channels(); c++) { |
| 1215 | ASSERT_LE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), output_max); |
| 1216 | ASSERT_GE(fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), output_min); |
| 1217 | ASSERT_EQ( |
| 1218 | fp16_ieee_to_fp32_value(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), |
| 1219 | output_ref[((i * output_height() + y) * output_width() + x) * channels() + c]) << |
| 1220 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c |
| 1221 | << ", min = " << output_min << ", max = " << output_max; |
| 1222 | } |
| 1223 | } |
| 1224 | } |
| 1225 | } |
| 1226 | |
| 1227 | // Re-generate data for the second run. |
| 1228 | std::generate(input.begin(), input.end(), std::ref(f16rng)); |
| 1229 | std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */); |
| 1230 | |
| 1231 | // Compute reference results for the second run, including clamping. |
| 1232 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1233 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1234 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1235 | for (size_t c = 0; c < channels(); c++) { |
| 1236 | float max_value = -std::numeric_limits<float>::infinity(); |
| 1237 | for (size_t py = 0; py < pooling_height(); py++) { |
| 1238 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 1239 | for (size_t px = 0; px < pooling_width(); px++) { |
| 1240 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 1241 | if (ix < next_input_width() && iy < next_input_height()) { |
| 1242 | max_value = std::max(max_value, |
| 1243 | fp16_ieee_to_fp32_value(input[((i * next_input_height() + iy) * next_input_width() + ix) * input_pixel_stride() + c])); |
| 1244 | } |
| 1245 | } |
| 1246 | } |
| 1247 | max_value = std::min(max_value, output_max); |
| 1248 | max_value = std::max(max_value, output_min); |
| 1249 | next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c] = max_value; |
| 1250 | } |
| 1251 | } |
| 1252 | } |
| 1253 | } |
| 1254 | |
| 1255 | // Setup and run Max Pooling operator the second time, and destroy the operator. |
| 1256 | ASSERT_EQ(xnn_status_success, |
| 1257 | xnn_setup_max_pooling2d_nhwc_f16( |
| 1258 | max_pooling_op, |
| 1259 | next_batch_size(), next_input_height(), next_input_width(), |
| 1260 | input.data(), output.data(), |
| 1261 | nullptr /* thread pool */)); |
| 1262 | |
| 1263 | ASSERT_EQ(xnn_status_success, |
| 1264 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 1265 | |
| 1266 | // Verify results of the second run. |
| 1267 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1268 | for (size_t y = 0; y < next_output_height(); y++) { |
| 1269 | for (size_t x = 0; x < next_output_width(); x++) { |
| 1270 | for (size_t c = 0; c < channels(); c++) { |
| 1271 | ASSERT_LE(fp16_ieee_to_fp32_value(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]), output_max); |
| 1272 | ASSERT_GE(fp16_ieee_to_fp32_value(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]), output_min); |
| 1273 | ASSERT_EQ( |
| 1274 | fp16_ieee_to_fp32_value(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]), |
| 1275 | next_output_ref[((i * next_output_height() + y) * next_output_width() + x) * channels() + c]) << |
| 1276 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c |
| 1277 | << ", min = " << output_min << ", max = " << output_max; |
| 1278 | } |
| 1279 | } |
| 1280 | } |
| 1281 | } |
| 1282 | } |
| 1283 | } |
| 1284 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1285 | void TestSetupF32() const { |
| 1286 | std::random_device random_device; |
| 1287 | auto rng = std::mt19937(random_device()); |
| 1288 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 1289 | |
| 1290 | std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + std::max( |
| 1291 | (batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels(), |
| 1292 | (next_batch_size() * next_input_height() * next_input_width() - 1) * input_pixel_stride() + channels())); |
| 1293 | std::vector<float> output(XNN_EXTRA_BYTES / sizeof(float) + std::max( |
| 1294 | (batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels(), |
| 1295 | (next_batch_size() * next_output_height() * next_output_width() - 1) * output_pixel_stride() + channels())); |
| 1296 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels()); |
| 1297 | std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * channels()); |
| 1298 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 1299 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 1300 | std::fill(output.begin(), output.end(), nanf("")); |
| 1301 | |
| 1302 | // Compute reference results, without clamping. |
| 1303 | for (size_t i = 0; i < batch_size(); i++) { |
| 1304 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 1305 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 1306 | for (size_t c = 0; c < channels(); c++) { |
| 1307 | float max_value = -std::numeric_limits<float>::infinity(); |
| 1308 | for (size_t py = 0; py < pooling_height(); py++) { |
| 1309 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 1310 | for (size_t px = 0; px < pooling_width(); px++) { |
| 1311 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 1312 | if (ix < input_width() && iy < input_height()) { |
| 1313 | max_value = std::max(max_value, |
| 1314 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c]); |
| 1315 | } |
| 1316 | } |
| 1317 | } |
| 1318 | output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = max_value; |
| 1319 | } |
| 1320 | } |
| 1321 | } |
| 1322 | } |
| 1323 | |
| 1324 | // Compute clamping parameters. |
| 1325 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 1326 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 1327 | const float accumulated_range = accumulated_max - accumulated_min; |
| 1328 | const float output_min = accumulated_range == 0.0f ? |
| 1329 | -std::numeric_limits<float>::infinity() : |
| 1330 | accumulated_min + accumulated_range / 255.0f * float(qmin()); |
| 1331 | const float output_max = accumulated_range == 0.0f ? |
| 1332 | +std::numeric_limits<float>::infinity() : |
| 1333 | accumulated_max - accumulated_range / 255.0f * float(255 - qmax()); |
| 1334 | |
| 1335 | // Clamp reference results. |
| 1336 | for (float& value : output_ref) { |
| 1337 | value = std::max(std::min(value, output_max), output_min); |
| 1338 | } |
| 1339 | |
| 1340 | // Create, setup, and run Max Pooling operator once. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1341 | ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1342 | xnn_operator_t max_pooling_op = nullptr; |
| 1343 | |
| 1344 | ASSERT_EQ(xnn_status_success, |
| 1345 | xnn_create_max_pooling2d_nhwc_f32( |
| 1346 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 1347 | pooling_height(), pooling_width(), |
| 1348 | stride_height(), stride_width(), |
| 1349 | dilation_height(), dilation_width(), |
| 1350 | channels(), input_pixel_stride(), output_pixel_stride(), |
| 1351 | output_min, output_max, |
| 1352 | 0, &max_pooling_op)); |
| 1353 | ASSERT_NE(nullptr, max_pooling_op); |
| 1354 | |
| 1355 | // Smart pointer to automatically delete max_pooling_op. |
| 1356 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_max_pooling_op(max_pooling_op, xnn_delete_operator); |
| 1357 | |
| 1358 | ASSERT_EQ(xnn_status_success, |
| 1359 | xnn_setup_max_pooling2d_nhwc_f32( |
| 1360 | max_pooling_op, |
| 1361 | batch_size(), input_height(), input_width(), |
| 1362 | input.data(), output.data(), |
| 1363 | nullptr /* thread pool */)); |
| 1364 | |
| 1365 | ASSERT_EQ(xnn_status_success, |
| 1366 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 1367 | |
| 1368 | // Verify results of the first run. |
| 1369 | for (size_t i = 0; i < batch_size(); i++) { |
| 1370 | for (size_t y = 0; y < output_height(); y++) { |
| 1371 | for (size_t x = 0; x < output_width(); x++) { |
| 1372 | for (size_t c = 0; c < channels(); c++) { |
| 1373 | ASSERT_LE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_max); |
| 1374 | ASSERT_GE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_min); |
| 1375 | ASSERT_EQ(output_ref[((i * output_height() + y) * output_width() + x) * channels() + c], |
| 1376 | output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]) << |
| 1377 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c; |
| 1378 | } |
| 1379 | } |
| 1380 | } |
| 1381 | } |
| 1382 | |
| 1383 | // Re-generate data for the second run. |
| 1384 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 1385 | std::fill(output.begin(), output.end(), 0xA5); |
| 1386 | |
| 1387 | // Compute reference results for the second run, including clamping. |
| 1388 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1389 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1390 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1391 | for (size_t c = 0; c < channels(); c++) { |
| 1392 | float max_value = -std::numeric_limits<float>::infinity(); |
| 1393 | for (size_t py = 0; py < pooling_height(); py++) { |
| 1394 | const size_t iy = oy * stride_height() + py * dilation_height() - padding_top(); |
| 1395 | for (size_t px = 0; px < pooling_width(); px++) { |
| 1396 | const size_t ix = ox * stride_width() + px * dilation_width() - padding_left(); |
| 1397 | if (ix < next_input_width() && iy < next_input_height()) { |
| 1398 | max_value = std::max(max_value, |
| 1399 | input[((i * next_input_height() + iy) * next_input_width() + ix) * input_pixel_stride() + c]); |
| 1400 | } |
| 1401 | } |
| 1402 | } |
| 1403 | max_value = std::min(max_value, output_max); |
| 1404 | max_value = std::max(max_value, output_min); |
| 1405 | next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c] = max_value; |
| 1406 | } |
| 1407 | } |
| 1408 | } |
| 1409 | } |
| 1410 | |
| 1411 | // Setup and run Max Pooling operator the second time, and destroy the operator. |
| 1412 | ASSERT_EQ(xnn_status_success, |
| 1413 | xnn_setup_max_pooling2d_nhwc_f32( |
| 1414 | max_pooling_op, |
| 1415 | next_batch_size(), next_input_height(), next_input_width(), |
| 1416 | input.data(), output.data(), |
| 1417 | nullptr /* thread pool */)); |
| 1418 | |
| 1419 | ASSERT_EQ(xnn_status_success, |
| 1420 | xnn_run_operator(max_pooling_op, nullptr /* thread pool */)); |
| 1421 | |
| 1422 | // Verify results of the second run. |
| 1423 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1424 | for (size_t y = 0; y < next_output_height(); y++) { |
| 1425 | for (size_t x = 0; x < next_output_width(); x++) { |
| 1426 | for (size_t c = 0; c < channels(); c++) { |
| 1427 | ASSERT_LE(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c], output_max); |
| 1428 | ASSERT_GE(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c], output_min); |
| 1429 | ASSERT_EQ(next_output_ref[((i * next_output_height() + y) * next_output_width() + x) * channels() + c], |
| 1430 | output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]) << |
| 1431 | "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c; |
| 1432 | } |
| 1433 | } |
| 1434 | } |
| 1435 | } |
| 1436 | } |
| 1437 | } |
| 1438 | |
| 1439 | private: |
| 1440 | uint32_t padding_top_{0}; |
| 1441 | uint32_t padding_right_{0}; |
| 1442 | uint32_t padding_bottom_{0}; |
| 1443 | uint32_t padding_left_{0}; |
Marat Dukhan | bee7825 | 2020-02-27 23:52:08 -0800 | [diff] [blame] | 1444 | bool padding_tf_same_{false}; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1445 | size_t input_height_{1}; |
| 1446 | size_t input_width_{1}; |
| 1447 | size_t channels_{1}; |
| 1448 | size_t batch_size_{1}; |
| 1449 | size_t input_pixel_stride_{0}; |
| 1450 | size_t output_pixel_stride_{0}; |
| 1451 | uint32_t pooling_height_{1}; |
| 1452 | uint32_t pooling_width_{1}; |
| 1453 | uint32_t stride_height_{1}; |
| 1454 | uint32_t stride_width_{1}; |
| 1455 | uint32_t dilation_height_{1}; |
| 1456 | uint32_t dilation_width_{1}; |
| 1457 | size_t next_input_height_{0}; |
| 1458 | size_t next_input_width_{0}; |
| 1459 | size_t next_batch_size_{0}; |
| 1460 | uint8_t qmin_{0}; |
| 1461 | uint8_t qmax_{255}; |
| 1462 | size_t iterations_{1}; |
| 1463 | }; |