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> |
| 19 | #include <random> |
| 20 | #include <vector> |
| 21 | |
| 22 | #include <xnnpack.h> |
| 23 | |
| 24 | |
| 25 | class ConvolutionOperatorTester { |
| 26 | public: |
| 27 | inline ConvolutionOperatorTester& padding(uint32_t padding) { |
| 28 | this->padding_top_ = padding; |
| 29 | this->padding_right_ = padding; |
| 30 | this->padding_bottom_ = padding; |
| 31 | this->padding_left_ = padding; |
| 32 | return *this; |
| 33 | } |
| 34 | |
| 35 | inline ConvolutionOperatorTester& padding(uint32_t padding_height, uint32_t padding_width) { |
| 36 | this->padding_top_ = padding_height; |
| 37 | this->padding_right_ = padding_width; |
| 38 | this->padding_bottom_ = padding_height; |
| 39 | this->padding_left_ = padding_width; |
| 40 | return *this; |
| 41 | } |
| 42 | |
| 43 | inline ConvolutionOperatorTester& padding_height(uint32_t padding_height) { |
| 44 | this->padding_top_ = padding_height; |
| 45 | this->padding_bottom_ = padding_height; |
| 46 | return *this; |
| 47 | } |
| 48 | |
| 49 | inline ConvolutionOperatorTester& padding_width(uint32_t padding_width) { |
| 50 | this->padding_right_ = padding_width; |
| 51 | this->padding_left_ = padding_width; |
| 52 | return *this; |
| 53 | } |
| 54 | |
| 55 | inline ConvolutionOperatorTester& padding_top(uint32_t padding_top) { |
| 56 | this->padding_top_ = padding_top; |
| 57 | return *this; |
| 58 | } |
| 59 | |
| 60 | inline uint32_t padding_top() const { |
| 61 | return this->padding_top_; |
| 62 | } |
| 63 | |
| 64 | inline ConvolutionOperatorTester& padding_right(uint32_t padding_right) { |
| 65 | this->padding_right_ = padding_right; |
| 66 | return *this; |
| 67 | } |
| 68 | |
| 69 | inline uint32_t padding_right() const { |
| 70 | return this->padding_right_; |
| 71 | } |
| 72 | |
| 73 | inline ConvolutionOperatorTester& padding_bottom(uint32_t padding_bottom) { |
| 74 | this->padding_bottom_ = padding_bottom; |
| 75 | return *this; |
| 76 | } |
| 77 | |
| 78 | inline uint32_t padding_bottom() const { |
| 79 | return this->padding_bottom_; |
| 80 | } |
| 81 | |
| 82 | inline ConvolutionOperatorTester& padding_left(uint32_t padding_left) { |
| 83 | this->padding_left_ = padding_left; |
| 84 | return *this; |
| 85 | } |
| 86 | |
| 87 | inline uint32_t padding_left() const { |
| 88 | return this->padding_left_; |
| 89 | } |
| 90 | |
| 91 | inline ConvolutionOperatorTester& input_size(uint32_t input_height, uint32_t input_width) { |
| 92 | assert(input_height >= 1); |
| 93 | assert(input_width >= 1); |
| 94 | this->input_height_ = input_height; |
| 95 | this->input_width_ = input_width; |
| 96 | return *this; |
| 97 | } |
| 98 | |
| 99 | inline ConvolutionOperatorTester& input_height(uint32_t input_height) { |
| 100 | assert(input_height >= 1); |
| 101 | this->input_height_ = input_height; |
| 102 | return *this; |
| 103 | } |
| 104 | |
| 105 | inline uint32_t input_height() const { |
| 106 | return this->input_height_; |
| 107 | } |
| 108 | |
| 109 | inline ConvolutionOperatorTester& input_width(uint32_t input_width) { |
| 110 | assert(input_width >= 1); |
| 111 | this->input_width_ = input_width; |
| 112 | return *this; |
| 113 | } |
| 114 | |
| 115 | inline uint32_t input_width() const { |
| 116 | return this->input_width_; |
| 117 | } |
| 118 | |
| 119 | inline ConvolutionOperatorTester& groups(uint32_t groups) { |
| 120 | assert(groups >= 1); |
| 121 | this->groups_ = groups; |
| 122 | return *this; |
| 123 | } |
| 124 | |
| 125 | inline uint32_t groups() const { |
| 126 | return this->groups_; |
| 127 | } |
| 128 | |
| 129 | inline ConvolutionOperatorTester& group_input_channels(size_t group_input_channels) { |
| 130 | assert(group_input_channels >= 1); |
| 131 | this->group_input_channels_ = group_input_channels; |
| 132 | return *this; |
| 133 | } |
| 134 | |
| 135 | inline size_t group_input_channels() const { |
| 136 | return this->group_input_channels_; |
| 137 | } |
| 138 | |
| 139 | inline ConvolutionOperatorTester& group_output_channels(size_t group_output_channels) { |
| 140 | assert(group_output_channels >= 1); |
| 141 | this->group_output_channels_ = group_output_channels; |
| 142 | return *this; |
| 143 | } |
| 144 | |
| 145 | inline size_t group_output_channels() const { |
| 146 | return this->group_output_channels_; |
| 147 | } |
| 148 | |
| 149 | inline ConvolutionOperatorTester& batch_size(size_t batch_size) { |
| 150 | assert(batch_size >= 1); |
| 151 | this->batch_size_ = batch_size; |
| 152 | return *this; |
| 153 | } |
| 154 | |
| 155 | inline size_t batch_size() const { |
| 156 | return this->batch_size_; |
| 157 | } |
| 158 | |
| 159 | inline ConvolutionOperatorTester& kernel_size(uint32_t kernel_size) { |
| 160 | assert(kernel_size >= 1); |
| 161 | this->kernel_height_ = kernel_size; |
| 162 | this->kernel_width_ = kernel_size; |
| 163 | return *this; |
| 164 | } |
| 165 | |
| 166 | inline ConvolutionOperatorTester& kernel_size(uint32_t kernel_height, uint32_t kernel_width) { |
| 167 | assert(kernel_height >= 1); |
| 168 | assert(kernel_width >= 1); |
| 169 | this->kernel_height_ = kernel_height; |
| 170 | this->kernel_width_ = kernel_width; |
| 171 | return *this; |
| 172 | } |
| 173 | |
| 174 | inline ConvolutionOperatorTester& kernel_height(uint32_t kernel_height) { |
| 175 | assert(kernel_height >= 1); |
| 176 | this->kernel_height_ = kernel_height; |
| 177 | return *this; |
| 178 | } |
| 179 | |
| 180 | inline uint32_t kernel_height() const { |
| 181 | return this->kernel_height_; |
| 182 | } |
| 183 | |
| 184 | inline ConvolutionOperatorTester& kernel_width(uint32_t kernel_width) { |
| 185 | assert(kernel_width >= 1); |
| 186 | this->kernel_width_ = kernel_width; |
| 187 | return *this; |
| 188 | } |
| 189 | |
| 190 | inline uint32_t kernel_width() const { |
| 191 | return this->kernel_width_; |
| 192 | } |
| 193 | |
| 194 | inline ConvolutionOperatorTester& dilation(uint32_t dilation) { |
| 195 | assert(dilation >= 1); |
| 196 | this->dilation_height_ = dilation; |
| 197 | this->dilation_width_ = dilation; |
| 198 | return *this; |
| 199 | } |
| 200 | |
| 201 | inline ConvolutionOperatorTester& dilation(uint32_t dilation_height, uint32_t dilation_width) { |
| 202 | assert(dilation_height >= 1); |
| 203 | assert(dilation_width >= 1); |
| 204 | this->dilation_height_ = dilation_height; |
| 205 | this->dilation_width_ = dilation_width; |
| 206 | return *this; |
| 207 | } |
| 208 | |
| 209 | inline ConvolutionOperatorTester& dilation_height(uint32_t dilation_height) { |
| 210 | assert(dilation_height >= 1); |
| 211 | this->dilation_height_ = dilation_height; |
| 212 | return *this; |
| 213 | } |
| 214 | |
| 215 | inline uint32_t dilation_height() const { |
| 216 | return this->dilation_height_; |
| 217 | } |
| 218 | |
| 219 | inline ConvolutionOperatorTester& dilation_width(uint32_t dilation_width) { |
| 220 | assert(dilation_width >= 1); |
| 221 | this->dilation_width_ = dilation_width; |
| 222 | return *this; |
| 223 | } |
| 224 | |
| 225 | inline uint32_t dilation_width() const { |
| 226 | return this->dilation_width_; |
| 227 | } |
| 228 | |
| 229 | inline ConvolutionOperatorTester& subsampling(uint32_t subsampling) { |
| 230 | assert(subsampling >= 1); |
| 231 | this->subsampling_height_ = subsampling; |
| 232 | this->subsampling_width_ = subsampling; |
| 233 | return *this; |
| 234 | } |
| 235 | |
| 236 | inline ConvolutionOperatorTester& subsampling(uint32_t subsampling_height, uint32_t subsampling_width) { |
| 237 | assert(subsampling_height >= 1); |
| 238 | assert(subsampling_width >= 1); |
| 239 | this->subsampling_height_ = subsampling_height; |
| 240 | this->subsampling_width_ = subsampling_width; |
| 241 | return *this; |
| 242 | } |
| 243 | |
| 244 | inline ConvolutionOperatorTester& subsampling_height(uint32_t subsampling_height) { |
| 245 | assert(subsampling_height >= 1); |
| 246 | this->subsampling_height_ = subsampling_height; |
| 247 | return *this; |
| 248 | } |
| 249 | |
| 250 | inline uint32_t subsampling_height() const { |
| 251 | return this->subsampling_height_; |
| 252 | } |
| 253 | |
| 254 | inline ConvolutionOperatorTester& subsampling_width(uint32_t subsampling_width) { |
| 255 | assert(subsampling_width >= 1); |
| 256 | this->subsampling_width_ = subsampling_width; |
| 257 | return *this; |
| 258 | } |
| 259 | |
| 260 | inline uint32_t subsampling_width() const { |
| 261 | return this->subsampling_width_; |
| 262 | } |
| 263 | |
| 264 | inline ConvolutionOperatorTester& input_pixel_stride(size_t input_pixel_stride) { |
| 265 | assert(input_pixel_stride >= 1); |
| 266 | this->input_pixel_stride_ = input_pixel_stride; |
| 267 | return *this; |
| 268 | } |
| 269 | |
| 270 | inline size_t input_pixel_stride() const { |
| 271 | if (this->input_pixel_stride_ == 0) { |
| 272 | return group_input_channels() * groups(); |
| 273 | } else { |
| 274 | assert(this->input_pixel_stride_ >= group_input_channels() * groups()); |
| 275 | return this->input_pixel_stride_; |
| 276 | } |
| 277 | } |
| 278 | |
| 279 | inline ConvolutionOperatorTester& output_pixel_stride(size_t output_pixel_stride) { |
| 280 | assert(output_pixel_stride >= 1); |
| 281 | this->output_pixel_stride_ = output_pixel_stride; |
| 282 | return *this; |
| 283 | } |
| 284 | |
| 285 | inline size_t output_pixel_stride() const { |
| 286 | if (this->output_pixel_stride_ == 0) { |
| 287 | return group_output_channels() * groups(); |
| 288 | } else { |
| 289 | assert(this->output_pixel_stride_ >= group_output_channels() * groups()); |
| 290 | return this->output_pixel_stride_; |
| 291 | } |
| 292 | } |
| 293 | |
| 294 | inline uint32_t dilated_kernel_height() const { |
| 295 | return (kernel_height() - 1) * dilation_height() + 1; |
| 296 | } |
| 297 | |
| 298 | inline uint32_t dilated_kernel_width() const { |
| 299 | return (kernel_width() - 1) * dilation_width() + 1; |
| 300 | } |
| 301 | |
| 302 | inline size_t output_height() const { |
| 303 | const size_t padded_input_height = padding_top() + input_height() + padding_bottom(); |
| 304 | if (padded_input_height <= dilated_kernel_height()) { |
| 305 | return 1; |
| 306 | } else { |
| 307 | return (padded_input_height - dilated_kernel_height()) / subsampling_height() + 1; |
| 308 | } |
| 309 | } |
| 310 | |
| 311 | inline size_t output_width() const { |
| 312 | const size_t padded_input_width = padding_left() + input_width() + padding_right(); |
| 313 | if (padded_input_width <= dilated_kernel_width()) { |
| 314 | return 1; |
| 315 | } else { |
| 316 | return (padded_input_width - dilated_kernel_width()) / subsampling_width() + 1; |
| 317 | } |
| 318 | } |
| 319 | |
| 320 | inline ConvolutionOperatorTester& next_input_size(uint32_t next_input_height, uint32_t next_input_width) { |
| 321 | assert(next_input_height >= 1); |
| 322 | assert(next_input_width >= 1); |
| 323 | this->next_input_height_ = next_input_height; |
| 324 | this->next_input_width_ = next_input_width; |
| 325 | return *this; |
| 326 | } |
| 327 | |
| 328 | inline ConvolutionOperatorTester& next_input_height(uint32_t next_input_height) { |
| 329 | assert(next_input_height >= 1); |
| 330 | this->next_input_height_ = next_input_height; |
| 331 | return *this; |
| 332 | } |
| 333 | |
| 334 | inline uint32_t next_input_height() const { |
| 335 | if (this->next_input_height_ == 0) { |
| 336 | return input_height(); |
| 337 | } else { |
| 338 | return this->next_input_height_; |
| 339 | } |
| 340 | } |
| 341 | |
| 342 | inline ConvolutionOperatorTester& next_input_width(uint32_t next_input_width) { |
| 343 | assert(next_input_width >= 1); |
| 344 | this->next_input_width_ = next_input_width; |
| 345 | return *this; |
| 346 | } |
| 347 | |
| 348 | inline uint32_t next_input_width() const { |
| 349 | if (this->next_input_width_ == 0) { |
| 350 | return input_width(); |
| 351 | } else { |
| 352 | return this->next_input_width_; |
| 353 | } |
| 354 | } |
| 355 | |
| 356 | inline size_t next_output_height() const { |
| 357 | const size_t padded_input_height = padding_top() + next_input_height() + padding_bottom(); |
| 358 | if (padded_input_height <= dilated_kernel_height()) { |
| 359 | return 1; |
| 360 | } else { |
| 361 | return (padded_input_height - dilated_kernel_height()) / subsampling_height() + 1; |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | inline size_t next_output_width() const { |
| 366 | const size_t padded_input_width = padding_left() + next_input_width() + padding_right(); |
| 367 | if (padded_input_width <= dilated_kernel_width()) { |
| 368 | return 1; |
| 369 | } else { |
| 370 | return (padded_input_width - dilated_kernel_width()) / subsampling_width() + 1; |
| 371 | } |
| 372 | } |
| 373 | |
| 374 | inline ConvolutionOperatorTester& next_batch_size(size_t next_batch_size) { |
| 375 | assert(next_batch_size >= 1); |
| 376 | this->next_batch_size_ = next_batch_size; |
| 377 | return *this; |
| 378 | } |
| 379 | |
| 380 | inline size_t next_batch_size() const { |
| 381 | if (this->next_batch_size_ == 0) { |
| 382 | return batch_size(); |
| 383 | } else { |
| 384 | return this->next_batch_size_; |
| 385 | } |
| 386 | } |
| 387 | |
| 388 | inline ConvolutionOperatorTester& qmin(uint8_t qmin) { |
| 389 | this->qmin_ = qmin; |
| 390 | return *this; |
| 391 | } |
| 392 | |
| 393 | inline uint8_t qmin() const { |
| 394 | return this->qmin_; |
| 395 | } |
| 396 | |
| 397 | inline ConvolutionOperatorTester& qmax(uint8_t qmax) { |
| 398 | this->qmax_ = qmax; |
| 399 | return *this; |
| 400 | } |
| 401 | |
| 402 | inline uint8_t qmax() const { |
| 403 | return this->qmax_; |
| 404 | } |
| 405 | |
| 406 | inline ConvolutionOperatorTester& depthwise_layout(bool depthwise_layout) { |
| 407 | this->depthwise_layout_ = depthwise_layout; |
| 408 | return *this; |
| 409 | } |
| 410 | |
| 411 | inline bool depthwise_layout() const { |
| 412 | return this->depthwise_layout_; |
| 413 | } |
| 414 | |
| 415 | inline ConvolutionOperatorTester& iterations(size_t iterations) { |
| 416 | this->iterations_ = iterations; |
| 417 | return *this; |
| 418 | } |
| 419 | |
| 420 | inline size_t iterations() const { |
| 421 | return this->iterations_; |
| 422 | } |
| 423 | |
| 424 | void TestQ8() const { |
| 425 | std::random_device random_device; |
| 426 | auto rng = std::mt19937(random_device()); |
| 427 | auto s32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng); |
| 428 | auto u8rng = std::bind(std::uniform_int_distribution<uint8_t>(), rng); |
| 429 | |
| 430 | std::vector<uint8_t> input(XNN_EXTRA_BYTES / sizeof(uint8_t) + |
| 431 | batch_size() * ((input_height() * input_width() - 1) * input_pixel_stride() + groups() * group_input_channels()) + 8); |
| 432 | std::vector<uint8_t> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 433 | std::vector<int32_t> bias(groups() * group_output_channels()); |
| 434 | std::vector<uint8_t> output(batch_size() * ((output_height() * output_width() - 1) * output_pixel_stride() + groups() * group_output_channels())); |
| 435 | std::vector<int32_t> accumulators(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 436 | std::vector<double> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 437 | |
| 438 | const uint8_t input_zero_point = 127; |
| 439 | const uint8_t kernel_zero_point = 127; |
| 440 | |
| 441 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 442 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 443 | std::generate(kernel.begin(), kernel.end(), std::ref(u8rng)); |
| 444 | std::generate(bias.begin(), bias.end(), std::ref(s32rng)); |
| 445 | std::fill(output.begin(), output.end(), 0xA5); |
| 446 | |
| 447 | // Compute reference results, without renormalization. |
| 448 | for (size_t i = 0; i < batch_size(); i++) { |
| 449 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 450 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 451 | for (size_t g = 0; g < groups(); g++) { |
| 452 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 453 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 454 | bias[g * group_output_channels() + oc]; |
| 455 | } |
| 456 | } |
| 457 | } |
| 458 | } |
| 459 | } |
| 460 | if (depthwise_layout()) { |
| 461 | ASSERT_EQ(group_input_channels(), 1); |
| 462 | |
| 463 | for (size_t i = 0; i < batch_size(); i++) { |
| 464 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 465 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 466 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 467 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 468 | if (iy < input_height()) { |
| 469 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 470 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 471 | if (ix < input_width()) { |
| 472 | for (size_t g = 0; g < groups(); g++) { |
| 473 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 474 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 475 | (int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + g]) - int32_t(input_zero_point)) * |
| 476 | (int32_t(kernel[((ky * kernel_width() + kx) * groups() + g) * group_output_channels() + oc]) - int32_t(kernel_zero_point)); |
| 477 | } |
| 478 | } |
| 479 | } |
| 480 | } |
| 481 | } |
| 482 | } |
| 483 | } |
| 484 | } |
| 485 | } |
| 486 | } else { |
| 487 | for (size_t i = 0; i < batch_size(); i++) { |
| 488 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 489 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 490 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 491 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 492 | if (iy < input_height()) { |
| 493 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 494 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 495 | if (ix < input_width()) { |
| 496 | for (size_t g = 0; g < groups(); g++) { |
| 497 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 498 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 499 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 500 | (int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + g * group_input_channels() + ic]) - int32_t(input_zero_point)) * |
| 501 | (int32_t(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]) - int32_t(kernel_zero_point)); |
| 502 | } |
| 503 | } |
| 504 | } |
| 505 | } |
| 506 | } |
| 507 | } |
| 508 | } |
| 509 | } |
| 510 | } |
| 511 | } |
| 512 | } |
| 513 | |
| 514 | // Compute renormalization parameters. |
| 515 | const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend()); |
| 516 | const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend()); |
| 517 | |
| 518 | const double output_scale = double(uint32_t(accumulated_max - accumulated_min)) / 255.0; |
| 519 | const uint8_t output_zero_point = uint8_t(std::max(std::min( |
| 520 | lrint(127.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale), |
| 521 | long(std::numeric_limits<uint8_t>::max())), long(std::numeric_limits<uint8_t>::min()))); |
| 522 | |
| 523 | // Renormalize reference results. |
| 524 | std::transform(accumulators.cbegin(), accumulators.cend(), output_ref.begin(), |
| 525 | [this, output_scale, output_zero_point](int32_t x) -> double { |
| 526 | return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax()) - output_zero_point), double(qmin()) - output_zero_point); |
| 527 | }); |
| 528 | |
| 529 | // Create, setup, run, and destroy Convolution operator. |
| 530 | ASSERT_EQ(xnn_status_success, xnn_initialize()); |
| 531 | xnn_operator_t convolution_op = nullptr; |
| 532 | |
| 533 | ASSERT_EQ(xnn_status_success, |
| 534 | xnn_create_convolution2d_nhwc_q8( |
| 535 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 536 | kernel_height(), kernel_width(), |
| 537 | subsampling_height(), subsampling_width(), |
| 538 | dilation_height(), dilation_width(), |
| 539 | groups(), group_input_channels(), group_output_channels(), |
| 540 | input_pixel_stride(), output_pixel_stride(), |
| 541 | input_zero_point, 1.0f /* input scale */, |
| 542 | kernel_zero_point, 1.0f /* kernel scale */, |
| 543 | kernel.data(), bias.data(), |
| 544 | output_zero_point, output_scale, qmin(), qmax(), |
| 545 | depthwise_layout() ? XNN_CONVOLUTION_FLAG_DEPTHWISE : 0, |
| 546 | &convolution_op)); |
| 547 | |
| 548 | // Smart pointer to automatically delete convolution_op. |
| 549 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 550 | |
| 551 | ASSERT_EQ(xnn_status_success, |
| 552 | xnn_setup_convolution2d_nhwc_q8( |
| 553 | convolution_op, |
| 554 | batch_size(), input_height(), input_width(), |
| 555 | input.data(), output.data(), |
| 556 | nullptr /* thread pool */)); |
| 557 | |
| 558 | ASSERT_EQ(xnn_status_success, |
| 559 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 560 | |
| 561 | // Verify results. |
| 562 | for (size_t i = 0; i < batch_size(); i++) { |
| 563 | for (size_t y = 0; y < output_height(); y++) { |
| 564 | for (size_t x = 0; x < output_width(); x++) { |
| 565 | for (size_t g = 0; g < groups(); g++) { |
| 566 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 567 | ASSERT_LE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c]), int32_t(qmax())) |
| 568 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 569 | ASSERT_GE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c]), int32_t(qmin())) |
| 570 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 571 | ASSERT_NEAR( |
| 572 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
| 573 | double(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c]) - double(output_zero_point), |
| 574 | 0.9) |
| 575 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 576 | } |
| 577 | } |
| 578 | } |
| 579 | } |
| 580 | } |
| 581 | } |
| 582 | } |
| 583 | |
| 584 | void TestF32() const { |
| 585 | std::random_device random_device; |
| 586 | auto rng = std::mt19937(random_device()); |
| 587 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.1f, 1.0f), rng); |
| 588 | |
| 589 | std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + |
| 590 | batch_size() * ((input_height() * input_width() - 1) * input_pixel_stride() + groups() * group_input_channels())); |
| 591 | std::vector<float> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 592 | std::vector<float> bias(groups() * group_output_channels()); |
| 593 | std::vector<float> output(batch_size() * ((output_height() * output_width() - 1) * output_pixel_stride() + groups() * group_output_channels())); |
| 594 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 595 | |
| 596 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 597 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 598 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 599 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 600 | std::fill(output.begin(), output.end(), nanf("")); |
| 601 | |
| 602 | // Compute reference results, without clamping. |
| 603 | for (size_t i = 0; i < batch_size(); i++) { |
| 604 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 605 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 606 | for (size_t g = 0; g < groups(); g++) { |
| 607 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 608 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 609 | bias[g * group_output_channels() + oc]; |
| 610 | } |
| 611 | } |
| 612 | } |
| 613 | } |
| 614 | } |
| 615 | if (depthwise_layout()) { |
| 616 | ASSERT_EQ(group_input_channels(), 1); |
| 617 | |
| 618 | for (size_t i = 0; i < batch_size(); i++) { |
| 619 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 620 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 621 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 622 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 623 | if (iy < input_height()) { |
| 624 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 625 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 626 | if (ix < input_width()) { |
| 627 | for (size_t g = 0; g < groups(); g++) { |
| 628 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 629 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 630 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + g] * |
| 631 | kernel[((ky * kernel_width() + kx) * groups() + g) * group_output_channels() + oc]; |
| 632 | } |
| 633 | } |
| 634 | } |
| 635 | } |
| 636 | } |
| 637 | } |
| 638 | } |
| 639 | } |
| 640 | } |
| 641 | } else { |
| 642 | for (size_t i = 0; i < batch_size(); i++) { |
| 643 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 644 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 645 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 646 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 647 | if (iy < input_height()) { |
| 648 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 649 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 650 | if (ix < input_width()) { |
| 651 | for (size_t g = 0; g < groups(); g++) { |
| 652 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 653 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 654 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 655 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + g * group_input_channels() + ic] * |
| 656 | kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]; |
| 657 | } |
| 658 | } |
| 659 | } |
| 660 | } |
| 661 | } |
| 662 | } |
| 663 | } |
| 664 | } |
| 665 | } |
| 666 | } |
| 667 | } |
| 668 | |
| 669 | // Compute clamping parameters. |
| 670 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 671 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 672 | |
| 673 | const float output_min = accumulated_min + (accumulated_max - accumulated_min) / 255.0f * float(qmin()); |
| 674 | const float output_max = accumulated_max - (accumulated_max - accumulated_min) / 255.0f * float(255 - qmax()); |
| 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 Convolution operator. |
| 682 | ASSERT_EQ(xnn_status_success, xnn_initialize()); |
| 683 | xnn_operator_t convolution_op = nullptr; |
| 684 | |
| 685 | ASSERT_EQ(xnn_status_success, |
| 686 | xnn_create_convolution2d_nhwc_f32( |
| 687 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 688 | kernel_height(), kernel_width(), |
| 689 | subsampling_height(), subsampling_width(), |
| 690 | dilation_height(), dilation_width(), |
| 691 | groups(), group_input_channels(), group_output_channels(), |
| 692 | input_pixel_stride(), output_pixel_stride(), |
| 693 | kernel.data(), bias.data(), |
| 694 | output_min, output_max, |
| 695 | depthwise_layout() ? XNN_CONVOLUTION_FLAG_DEPTHWISE : 0, |
| 696 | &convolution_op)); |
| 697 | |
| 698 | // Smart pointer to automatically delete convolution_op. |
| 699 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 700 | |
| 701 | ASSERT_EQ(xnn_status_success, |
| 702 | xnn_setup_convolution2d_nhwc_f32( |
| 703 | convolution_op, |
| 704 | batch_size(), input_height(), input_width(), |
| 705 | input.data(), output.data(), |
| 706 | nullptr /* thread pool */)); |
| 707 | |
| 708 | ASSERT_EQ(xnn_status_success, |
| 709 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 710 | |
| 711 | // Verify results. |
| 712 | for (size_t i = 0; i < batch_size(); i++) { |
| 713 | for (size_t y = 0; y < output_height(); y++) { |
| 714 | for (size_t x = 0; x < output_width(); x++) { |
| 715 | for (size_t g = 0; g < groups(); g++) { |
| 716 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 717 | ASSERT_GE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c], output_min) |
| 718 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 719 | ASSERT_LE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c], output_max) |
| 720 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 721 | ASSERT_NEAR( |
| 722 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
| 723 | output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c], |
| 724 | 1.0e-4 * std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c])) |
| 725 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 726 | } |
| 727 | } |
| 728 | } |
| 729 | } |
| 730 | } |
| 731 | } |
| 732 | } |
| 733 | |
| 734 | void TestSetupQ8() const { |
| 735 | ASSERT_FALSE(depthwise_layout()); |
| 736 | |
| 737 | std::random_device random_device; |
| 738 | auto rng = std::mt19937(random_device()); |
| 739 | auto s32rng = std::bind(std::uniform_int_distribution<int32_t>(-10000, 10000), rng); |
| 740 | auto u8rng = std::bind(std::uniform_int_distribution<uint8_t>(), rng); |
| 741 | |
| 742 | std::vector<uint8_t> input(XNN_EXTRA_BYTES / sizeof(uint8_t) + std::max( |
| 743 | batch_size() * ((input_height() * input_width() - 1) * input_pixel_stride() + groups() * group_input_channels()), |
| 744 | next_batch_size() * ((next_input_height() * next_input_width() - 1) * input_pixel_stride() + groups() * group_input_channels())) + 8); |
| 745 | std::vector<uint8_t> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 746 | std::vector<int32_t> bias(groups() * group_output_channels()); |
| 747 | std::vector<uint8_t> output(std::max( |
| 748 | batch_size() * ((output_height() * output_width() - 1) * output_pixel_stride() + groups() * group_output_channels()), |
| 749 | next_batch_size() * ((next_output_height() * next_output_width() - 1) * output_pixel_stride() + groups() * group_output_channels()))); |
| 750 | std::vector<int32_t> accumulators(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 751 | std::vector<double> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 752 | std::vector<int32_t> next_accumulators(next_batch_size() * next_output_height() * next_output_width() * groups() * group_output_channels()); |
| 753 | std::vector<double> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * groups() * group_output_channels()); |
| 754 | |
| 755 | const uint8_t input_zero_point = 127; |
| 756 | const uint8_t kernel_zero_point = 127; |
| 757 | |
| 758 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 759 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 760 | std::generate(kernel.begin(), kernel.end(), std::ref(u8rng)); |
| 761 | std::generate(bias.begin(), bias.end(), std::ref(s32rng)); |
| 762 | std::fill(output.begin(), output.end(), 0xA5); |
| 763 | |
| 764 | // Compute reference results, without renormalization. |
| 765 | for (size_t i = 0; i < batch_size(); i++) { |
| 766 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 767 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 768 | for (size_t g = 0; g < groups(); g++) { |
| 769 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 770 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 771 | bias[g * group_output_channels() + oc]; |
| 772 | } |
| 773 | } |
| 774 | } |
| 775 | } |
| 776 | } |
| 777 | for (size_t i = 0; i < batch_size(); i++) { |
| 778 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 779 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 780 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 781 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 782 | if (iy < input_height()) { |
| 783 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 784 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 785 | if (ix < input_width()) { |
| 786 | for (size_t g = 0; g < groups(); g++) { |
| 787 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 788 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 789 | accumulators[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 790 | (int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + g * group_input_channels() + ic]) - int32_t(input_zero_point)) * |
| 791 | (int32_t(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]) - int32_t(kernel_zero_point)); |
| 792 | } |
| 793 | } |
| 794 | } |
| 795 | } |
| 796 | } |
| 797 | } |
| 798 | } |
| 799 | } |
| 800 | } |
| 801 | } |
| 802 | |
| 803 | // Compute renormalization parameters. |
| 804 | const int32_t accumulated_min = *std::min_element(accumulators.cbegin(), accumulators.cend()); |
| 805 | const int32_t accumulated_max = *std::max_element(accumulators.cbegin(), accumulators.cend()); |
| 806 | |
| 807 | const double output_scale = double(uint32_t(accumulated_max - accumulated_min)) / 255.0; |
| 808 | const uint8_t output_zero_point = uint8_t(std::max(std::min( |
| 809 | lrint(127.5 - 0.5 * double(accumulated_min + accumulated_max) / output_scale), |
| 810 | long(std::numeric_limits<uint8_t>::max())), long(std::numeric_limits<uint8_t>::min()))); |
| 811 | |
| 812 | // Renormalize reference results. |
| 813 | std::transform(accumulators.cbegin(), accumulators.cend(), output_ref.begin(), |
| 814 | [this, output_scale, output_zero_point](int32_t x) -> double { |
| 815 | return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax()) - output_zero_point), double(qmin()) - output_zero_point); |
| 816 | }); |
| 817 | |
| 818 | // Create, setup, and run Convolution operator once. |
| 819 | ASSERT_EQ(xnn_status_success, xnn_initialize()); |
| 820 | xnn_operator_t convolution_op = nullptr; |
| 821 | |
| 822 | ASSERT_EQ(xnn_status_success, |
| 823 | xnn_create_convolution2d_nhwc_q8( |
| 824 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 825 | kernel_height(), kernel_width(), |
| 826 | subsampling_height(), subsampling_width(), |
| 827 | dilation_height(), dilation_width(), |
| 828 | groups(), group_input_channels(), group_output_channels(), |
| 829 | input_pixel_stride(), output_pixel_stride(), |
| 830 | input_zero_point, 1.0f /* input scale */, |
| 831 | kernel_zero_point, 1.0f /* kernel scale */, |
| 832 | kernel.data(), bias.data(), |
| 833 | output_zero_point, output_scale, qmin(), qmax(), |
| 834 | 0, &convolution_op)); |
| 835 | |
| 836 | // Smart pointer to automatically delete convolution_op. |
| 837 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 838 | |
| 839 | ASSERT_EQ(xnn_status_success, |
| 840 | xnn_setup_convolution2d_nhwc_q8( |
| 841 | convolution_op, |
| 842 | batch_size(), input_height(), input_width(), |
| 843 | input.data(), output.data(), |
| 844 | nullptr /* thread pool */)); |
| 845 | |
| 846 | ASSERT_EQ(xnn_status_success, |
| 847 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 848 | |
| 849 | // Verify results of the first run. |
| 850 | for (size_t i = 0; i < batch_size(); i++) { |
| 851 | for (size_t y = 0; y < output_height(); y++) { |
| 852 | for (size_t x = 0; x < output_width(); x++) { |
| 853 | for (size_t g = 0; g < groups(); g++) { |
| 854 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 855 | ASSERT_LE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c]), int32_t(qmax())) |
| 856 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 857 | ASSERT_GE(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c]), int32_t(qmin())) |
| 858 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 859 | ASSERT_NEAR( |
| 860 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
| 861 | double(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c]) - double(output_zero_point), |
| 862 | 0.9) |
| 863 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 864 | } |
| 865 | } |
| 866 | } |
| 867 | } |
| 868 | } |
| 869 | |
| 870 | // Re-generate data for the second run. |
| 871 | std::generate(input.begin(), input.end(), std::ref(u8rng)); |
| 872 | std::fill(output.begin(), output.end(), 0xA5); |
| 873 | |
| 874 | // Compute reference results for the second run, including renormalization. |
| 875 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 876 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 877 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 878 | for (size_t g = 0; g < groups(); g++) { |
| 879 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 880 | next_accumulators[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 881 | bias[g * group_output_channels() + oc]; |
| 882 | } |
| 883 | } |
| 884 | } |
| 885 | } |
| 886 | } |
| 887 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 888 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 889 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 890 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 891 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 892 | if (iy < next_input_height()) { |
| 893 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 894 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 895 | if (ix < next_input_width()) { |
| 896 | for (size_t g = 0; g < groups(); g++) { |
| 897 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 898 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 899 | next_accumulators[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 900 | (int32_t(input[((i * next_input_height() + iy) * next_input_width() + ix) * input_pixel_stride() + g * group_input_channels() + ic]) - int32_t(input_zero_point)) * |
| 901 | (int32_t(kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]) - int32_t(kernel_zero_point)); |
| 902 | } |
| 903 | } |
| 904 | } |
| 905 | } |
| 906 | } |
| 907 | } |
| 908 | } |
| 909 | } |
| 910 | } |
| 911 | } |
| 912 | std::transform(next_accumulators.cbegin(), next_accumulators.cend(), next_output_ref.begin(), |
| 913 | [this, output_scale, output_zero_point](int32_t x) -> double { |
| 914 | return std::max<double>(std::min<double>(double(x) / output_scale, double(qmax()) - output_zero_point), double(qmin()) - output_zero_point); |
| 915 | }); |
| 916 | |
| 917 | // Setup and run Convolution operator the second time, and destroy the operator. |
| 918 | ASSERT_EQ(xnn_status_success, |
| 919 | xnn_setup_convolution2d_nhwc_q8( |
| 920 | convolution_op, |
| 921 | next_batch_size(), next_input_height(), next_input_width(), |
| 922 | input.data(), output.data(), |
| 923 | nullptr /* thread pool */)); |
| 924 | |
| 925 | ASSERT_EQ(xnn_status_success, |
| 926 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 927 | |
| 928 | // Verify results of the second run. |
| 929 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 930 | for (size_t y = 0; y < next_output_height(); y++) { |
| 931 | for (size_t x = 0; x < next_output_width(); x++) { |
| 932 | for (size_t g = 0; g < groups(); g++) { |
| 933 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 934 | ASSERT_LE(int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + g * group_output_channels() + c]), int32_t(qmax())) |
| 935 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 936 | ASSERT_GE(int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + g * group_output_channels() + c]), int32_t(qmin())) |
| 937 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 938 | ASSERT_NEAR( |
| 939 | next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c], |
| 940 | double(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + g * group_output_channels() + c]) - double(output_zero_point), |
| 941 | 0.9) |
| 942 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 943 | } |
| 944 | } |
| 945 | } |
| 946 | } |
| 947 | } |
| 948 | } |
| 949 | } |
| 950 | |
| 951 | void TestSetupF32() const { |
| 952 | ASSERT_FALSE(depthwise_layout()); |
| 953 | |
| 954 | std::random_device random_device; |
| 955 | auto rng = std::mt19937(random_device()); |
| 956 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.1f, 1.0f), rng); |
| 957 | |
| 958 | std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + std::max( |
| 959 | batch_size() * ((input_height() * input_width() - 1) * input_pixel_stride() + groups() * group_input_channels()), |
| 960 | next_batch_size() * ((next_input_height() * next_input_width() - 1) * input_pixel_stride() + groups() * group_input_channels()))); |
| 961 | std::vector<float> kernel(groups() * group_output_channels() * kernel_height() * kernel_width() * group_input_channels()); |
| 962 | std::vector<float> bias(groups() * group_output_channels()); |
| 963 | std::vector<float> output(std::max( |
| 964 | batch_size() * ((output_height() * output_width() - 1) * output_pixel_stride() + groups() * group_output_channels()), |
| 965 | next_batch_size() * ((next_output_height() * next_output_width() - 1) * output_pixel_stride() + groups() * group_output_channels()))); |
| 966 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * groups() * group_output_channels()); |
| 967 | std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * groups() * group_output_channels()); |
| 968 | |
| 969 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 970 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 971 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 972 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 973 | std::fill(output.begin(), output.end(), nanf("")); |
| 974 | |
| 975 | // Compute reference results, without clamping. |
| 976 | for (size_t i = 0; i < batch_size(); i++) { |
| 977 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 978 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 979 | for (size_t g = 0; g < groups(); g++) { |
| 980 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 981 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 982 | bias[g * group_output_channels() + oc]; |
| 983 | } |
| 984 | } |
| 985 | } |
| 986 | } |
| 987 | } |
| 988 | for (size_t i = 0; i < batch_size(); i++) { |
| 989 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 990 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 991 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 992 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 993 | if (iy < input_height()) { |
| 994 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 995 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 996 | if (ix < input_width()) { |
| 997 | for (size_t g = 0; g < groups(); g++) { |
| 998 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 999 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1000 | output_ref[(((i * output_height() + oy) * output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 1001 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + g * group_input_channels() + ic] * |
| 1002 | kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]; |
| 1003 | } |
| 1004 | } |
| 1005 | } |
| 1006 | } |
| 1007 | } |
| 1008 | } |
| 1009 | } |
| 1010 | } |
| 1011 | } |
| 1012 | } |
| 1013 | |
| 1014 | // Compute clamping parameters. |
| 1015 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 1016 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 1017 | |
| 1018 | const float output_min = accumulated_min + (accumulated_max - accumulated_min) / 255.0f * float(qmin()); |
| 1019 | const float output_max = accumulated_max - (accumulated_max - accumulated_min) / 255.0f * float(255 - qmax()); |
| 1020 | |
| 1021 | // Clamp reference results. |
| 1022 | for (float& value : output_ref) { |
| 1023 | value = std::max(std::min(value, output_max), output_min); |
| 1024 | } |
| 1025 | |
| 1026 | // Create, setup, and run Convolution operator once. |
| 1027 | ASSERT_EQ(xnn_status_success, xnn_initialize()); |
| 1028 | xnn_operator_t convolution_op = nullptr; |
| 1029 | |
| 1030 | ASSERT_EQ(xnn_status_success, |
| 1031 | xnn_create_convolution2d_nhwc_f32( |
| 1032 | padding_top(), padding_right(), padding_bottom(), padding_left(), |
| 1033 | kernel_height(), kernel_width(), |
| 1034 | subsampling_height(), subsampling_width(), |
| 1035 | dilation_height(), dilation_width(), |
| 1036 | groups(), group_input_channels(), group_output_channels(), |
| 1037 | input_pixel_stride(), output_pixel_stride(), |
| 1038 | kernel.data(), bias.data(), |
| 1039 | output_min, output_max, |
| 1040 | 0, &convolution_op)); |
| 1041 | |
| 1042 | // Smart pointer to automatically delete convolution_op. |
| 1043 | std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_convolution_op(convolution_op, xnn_delete_operator); |
| 1044 | |
| 1045 | ASSERT_EQ(xnn_status_success, |
| 1046 | xnn_setup_convolution2d_nhwc_f32( |
| 1047 | convolution_op, |
| 1048 | batch_size(), input_height(), input_width(), |
| 1049 | input.data(), output.data(), |
| 1050 | nullptr /* thread pool */)); |
| 1051 | |
| 1052 | ASSERT_EQ(xnn_status_success, |
| 1053 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 1054 | |
| 1055 | // Verify results of the first run. |
| 1056 | for (size_t i = 0; i < batch_size(); i++) { |
| 1057 | for (size_t y = 0; y < output_height(); y++) { |
| 1058 | for (size_t x = 0; x < output_width(); x++) { |
| 1059 | for (size_t g = 0; g < groups(); g++) { |
| 1060 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 1061 | ASSERT_GE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c], output_min) |
| 1062 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1063 | ASSERT_LE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c], output_max) |
| 1064 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1065 | ASSERT_NEAR( |
| 1066 | output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c], |
| 1067 | output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + g * group_output_channels() + c], |
| 1068 | 1.0e-4 * std::abs(output_ref[(((i * output_height() + y) * output_width() + x) * groups() + g) * group_output_channels() + c])) |
| 1069 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1070 | } |
| 1071 | } |
| 1072 | } |
| 1073 | } |
| 1074 | } |
| 1075 | |
| 1076 | // Re-generate data for the second run. |
| 1077 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 1078 | std::fill(output.begin(), output.end(), nanf("")); |
| 1079 | |
| 1080 | // Compute reference results for the second run, including clamping. |
| 1081 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1082 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1083 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1084 | for (size_t g = 0; g < groups(); g++) { |
| 1085 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1086 | next_output_ref[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] = |
| 1087 | bias[g * group_output_channels() + oc]; |
| 1088 | } |
| 1089 | } |
| 1090 | } |
| 1091 | } |
| 1092 | } |
| 1093 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1094 | for (size_t oy = 0; oy < next_output_height(); oy++) { |
| 1095 | for (size_t ox = 0; ox < next_output_width(); ox++) { |
| 1096 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 1097 | const size_t iy = oy * subsampling_height() + ky * dilation_height() - padding_top(); |
| 1098 | if (iy < next_input_height()) { |
| 1099 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 1100 | const size_t ix = ox * subsampling_width() + kx * dilation_width() - padding_left(); |
| 1101 | if (ix < next_input_width()) { |
| 1102 | for (size_t g = 0; g < groups(); g++) { |
| 1103 | for (size_t oc = 0; oc < group_output_channels(); oc++) { |
| 1104 | for (size_t ic = 0; ic < group_input_channels(); ic++) { |
| 1105 | next_output_ref[(((i * next_output_height() + oy) * next_output_width() + ox) * groups() + g) * group_output_channels() + oc] += |
| 1106 | input[((i * next_input_height() + iy) * next_input_width() + ix) * input_pixel_stride() + g * group_input_channels() + ic] * |
| 1107 | kernel[(((g * group_output_channels() + oc) * kernel_height() + ky) * kernel_width() + kx) * group_input_channels() + ic]; |
| 1108 | } |
| 1109 | } |
| 1110 | } |
| 1111 | } |
| 1112 | } |
| 1113 | } |
| 1114 | } |
| 1115 | } |
| 1116 | } |
| 1117 | } |
| 1118 | for (float& value : next_output_ref) { |
| 1119 | value = std::max(std::min(value, output_max), output_min); |
| 1120 | } |
| 1121 | |
| 1122 | // Setup and run Convolution operator the second time, and destroy the operator. |
| 1123 | ASSERT_EQ(xnn_status_success, |
| 1124 | xnn_setup_convolution2d_nhwc_f32( |
| 1125 | convolution_op, |
| 1126 | next_batch_size(), next_input_height(), next_input_width(), |
| 1127 | input.data(), output.data(), |
| 1128 | nullptr /* thread pool */)); |
| 1129 | |
| 1130 | ASSERT_EQ(xnn_status_success, |
| 1131 | xnn_run_operator(convolution_op, nullptr /* thread pool */)); |
| 1132 | |
| 1133 | // Verify results of the second run. |
| 1134 | for (size_t i = 0; i < next_batch_size(); i++) { |
| 1135 | for (size_t y = 0; y < next_output_height(); y++) { |
| 1136 | for (size_t x = 0; x < next_output_width(); x++) { |
| 1137 | for (size_t g = 0; g < groups(); g++) { |
| 1138 | for (size_t c = 0; c < group_output_channels(); c++) { |
| 1139 | ASSERT_GE(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + g * group_output_channels() + c], output_min) |
| 1140 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1141 | ASSERT_LE(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + g * group_output_channels() + c], output_max) |
| 1142 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1143 | ASSERT_NEAR( |
| 1144 | next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c], |
| 1145 | output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + g * group_output_channels() + c], |
| 1146 | 1.0e-4 * std::abs(next_output_ref[(((i * next_output_height() + y) * next_output_width() + x) * groups() + g) * group_output_channels() + c])) |
| 1147 | << "(x, y) = (" << x << ", " << y << "), group = " << g << ", channel = " << c; |
| 1148 | } |
| 1149 | } |
| 1150 | } |
| 1151 | } |
| 1152 | } |
| 1153 | } |
| 1154 | } |
| 1155 | |
| 1156 | private: |
| 1157 | uint32_t padding_top_{0}; |
| 1158 | uint32_t padding_right_{0}; |
| 1159 | uint32_t padding_bottom_{0}; |
| 1160 | uint32_t padding_left_{0}; |
| 1161 | size_t input_height_{1}; |
| 1162 | size_t input_width_{1}; |
| 1163 | uint32_t groups_{1}; |
| 1164 | size_t group_input_channels_{1}; |
| 1165 | size_t input_pixel_stride_{0}; |
| 1166 | size_t group_output_channels_{1}; |
| 1167 | size_t output_pixel_stride_{0}; |
| 1168 | size_t batch_size_{1}; |
| 1169 | uint32_t kernel_height_{1}; |
| 1170 | uint32_t kernel_width_{1}; |
| 1171 | uint32_t dilation_height_{1}; |
| 1172 | uint32_t dilation_width_{1}; |
| 1173 | uint32_t subsampling_height_{1}; |
| 1174 | uint32_t subsampling_width_{1}; |
| 1175 | size_t next_input_height_{0}; |
| 1176 | size_t next_input_width_{0}; |
| 1177 | size_t next_batch_size_{0}; |
| 1178 | uint8_t qmin_{0}; |
| 1179 | uint8_t qmax_{255}; |
| 1180 | bool depthwise_layout_{false}; |
| 1181 | size_t iterations_{1}; |
| 1182 | }; |