XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include <gtest/gtest.h> |
| 9 | |
| 10 | #include <algorithm> |
| 11 | #include <cassert> |
| 12 | #include <cmath> |
| 13 | #include <cstddef> |
| 14 | #include <cstdlib> |
| 15 | #include <functional> |
| 16 | #include <limits> |
| 17 | #include <random> |
| 18 | #include <vector> |
| 19 | |
| 20 | #include <xnnpack/AlignedAllocator.h> |
| 21 | #include <xnnpack/pack.h> |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 22 | #include <xnnpack/params-init.h> |
Frank Barchard | e0601b5 | 2019-10-25 17:43:34 -0700 | [diff] [blame] | 23 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | #include <xnnpack.h> |
| 25 | |
| 26 | |
| 27 | class ConvHWCMicrokernelTester { |
| 28 | public: |
| 29 | enum class Variant { |
| 30 | Native, |
| 31 | Scalar, |
| 32 | }; |
| 33 | |
| 34 | inline ConvHWCMicrokernelTester& output_channels_tile(uint32_t output_channels_tile) { |
| 35 | this->output_channels_tile_ = output_channels_tile; |
| 36 | return *this; |
| 37 | } |
| 38 | |
| 39 | inline uint32_t output_channels_tile() const { |
| 40 | return this->output_channels_tile_; |
| 41 | } |
| 42 | |
| 43 | inline ConvHWCMicrokernelTester& padding(uint32_t padding) { |
| 44 | this->padding_top_ = padding; |
| 45 | this->padding_right_ = padding; |
| 46 | this->padding_bottom_ = padding; |
| 47 | this->padding_left_ = padding; |
| 48 | return *this; |
| 49 | } |
| 50 | |
| 51 | inline ConvHWCMicrokernelTester& padding_height(uint32_t padding_height) { |
| 52 | this->padding_top_ = padding_height; |
| 53 | this->padding_bottom_ = padding_height; |
| 54 | return *this; |
| 55 | } |
| 56 | |
| 57 | inline ConvHWCMicrokernelTester& padding_width(uint32_t padding_width) { |
| 58 | this->padding_right_ = padding_width; |
| 59 | this->padding_left_ = padding_width; |
| 60 | return *this; |
| 61 | } |
| 62 | |
| 63 | inline ConvHWCMicrokernelTester& padding_top(uint32_t padding_top) { |
| 64 | this->padding_top_ = padding_top; |
| 65 | return *this; |
| 66 | } |
| 67 | |
| 68 | inline uint32_t padding_top() const { |
| 69 | return this->padding_top_; |
| 70 | } |
| 71 | |
| 72 | inline ConvHWCMicrokernelTester& padding_right(uint32_t padding_right) { |
| 73 | this->padding_right_ = padding_right; |
| 74 | return *this; |
| 75 | } |
| 76 | |
| 77 | inline uint32_t padding_right() const { |
| 78 | return this->padding_right_; |
| 79 | } |
| 80 | |
| 81 | inline ConvHWCMicrokernelTester& padding_bottom(uint32_t padding_bottom) { |
| 82 | this->padding_bottom_ = padding_bottom; |
| 83 | return *this; |
| 84 | } |
| 85 | |
| 86 | inline uint32_t padding_bottom() const { |
| 87 | return this->padding_bottom_; |
| 88 | } |
| 89 | |
| 90 | inline ConvHWCMicrokernelTester& padding_left(uint32_t padding_left) { |
| 91 | this->padding_left_ = padding_left; |
| 92 | return *this; |
| 93 | } |
| 94 | |
| 95 | inline uint32_t padding_left() const { |
| 96 | return this->padding_left_; |
| 97 | } |
| 98 | |
| 99 | inline ConvHWCMicrokernelTester& input_size(uint32_t input_height, uint32_t input_width) { |
| 100 | assert(input_height >= 1); |
| 101 | assert(input_width >= 1); |
| 102 | this->input_height_ = input_height; |
| 103 | this->input_width_ = input_width; |
| 104 | return *this; |
| 105 | } |
| 106 | |
| 107 | inline ConvHWCMicrokernelTester& input_height(uint32_t input_height) { |
| 108 | assert(input_height >= 1); |
| 109 | this->input_height_ = input_height; |
| 110 | return *this; |
| 111 | } |
| 112 | |
| 113 | inline uint32_t input_height() const { |
| 114 | return this->input_height_; |
| 115 | } |
| 116 | |
| 117 | inline ConvHWCMicrokernelTester& input_width(uint32_t input_width) { |
| 118 | assert(input_width >= 1); |
| 119 | this->input_width_ = input_width; |
| 120 | return *this; |
| 121 | } |
| 122 | |
| 123 | inline uint32_t input_width() const { |
| 124 | return this->input_width_; |
| 125 | } |
| 126 | |
| 127 | inline ConvHWCMicrokernelTester& input_channels(size_t input_channels) { |
| 128 | assert(input_channels >= 1); |
| 129 | this->input_channels_ = input_channels; |
| 130 | return *this; |
| 131 | } |
| 132 | |
| 133 | inline size_t input_channels() const { |
| 134 | return this->input_channels_; |
| 135 | } |
| 136 | |
| 137 | inline ConvHWCMicrokernelTester& output_channels(size_t output_channels) { |
| 138 | assert(output_channels >= 1); |
| 139 | this->output_channels_ = output_channels; |
| 140 | return *this; |
| 141 | } |
| 142 | |
| 143 | inline size_t output_channels() const { |
| 144 | return this->output_channels_; |
| 145 | } |
| 146 | |
| 147 | inline size_t packed_output_channels() const { |
| 148 | return output_channels() % output_channels_tile() == 0 ? output_channels() : output_channels() / output_channels_tile() * output_channels_tile() + output_channels_tile(); |
| 149 | } |
| 150 | |
| 151 | inline ConvHWCMicrokernelTester& batch_size(size_t batch_size) { |
| 152 | assert(batch_size >= 1); |
| 153 | this->batch_size_ = batch_size; |
| 154 | return *this; |
| 155 | } |
| 156 | |
| 157 | inline size_t batch_size() const { |
| 158 | return this->batch_size_; |
| 159 | } |
| 160 | |
| 161 | inline ConvHWCMicrokernelTester& kernel_size(uint32_t kernel_size) { |
| 162 | assert(kernel_size >= 1); |
| 163 | this->kernel_height_ = kernel_size; |
| 164 | this->kernel_width_ = kernel_size; |
| 165 | return *this; |
| 166 | } |
| 167 | |
| 168 | inline ConvHWCMicrokernelTester& kernel_height(uint32_t kernel_height) { |
| 169 | assert(kernel_height >= 1); |
| 170 | this->kernel_height_ = kernel_height; |
| 171 | return *this; |
| 172 | } |
| 173 | |
| 174 | inline uint32_t kernel_height() const { |
| 175 | return this->kernel_height_; |
| 176 | } |
| 177 | |
| 178 | inline ConvHWCMicrokernelTester& kernel_width(uint32_t kernel_width) { |
| 179 | assert(kernel_width >= 1); |
| 180 | this->kernel_width_ = kernel_width; |
| 181 | return *this; |
| 182 | } |
| 183 | |
| 184 | inline uint32_t kernel_width() const { |
| 185 | return this->kernel_width_; |
| 186 | } |
| 187 | |
| 188 | inline ConvHWCMicrokernelTester& subsampling(uint32_t subsampling) { |
| 189 | assert(subsampling >= 1); |
| 190 | this->subsampling_height_ = subsampling; |
| 191 | this->subsampling_width_ = subsampling; |
| 192 | return *this; |
| 193 | } |
| 194 | |
| 195 | inline ConvHWCMicrokernelTester& subsampling_height(uint32_t subsampling_height) { |
| 196 | assert(subsampling_height >= 1); |
| 197 | this->subsampling_height_ = subsampling_height; |
| 198 | return *this; |
| 199 | } |
| 200 | |
| 201 | inline uint32_t subsampling_height() const { |
| 202 | return this->subsampling_height_; |
| 203 | } |
| 204 | |
| 205 | inline ConvHWCMicrokernelTester& subsampling_width(uint32_t subsampling_width) { |
| 206 | assert(subsampling_width >= 1); |
| 207 | this->subsampling_width_ = subsampling_width; |
| 208 | return *this; |
| 209 | } |
| 210 | |
| 211 | inline uint32_t subsampling_width() const { |
| 212 | return this->subsampling_width_; |
| 213 | } |
| 214 | |
| 215 | inline ConvHWCMicrokernelTester& output_y_start(uint32_t output_y_start) { |
| 216 | this->output_y_start_ = output_y_start; |
| 217 | return *this; |
| 218 | } |
| 219 | |
| 220 | inline uint32_t output_y_start() const { |
| 221 | return this->output_y_start_; |
| 222 | } |
| 223 | |
| 224 | inline ConvHWCMicrokernelTester& output_y_end(uint32_t output_y_end) { |
| 225 | this->output_y_end_ = output_y_end; |
| 226 | return *this; |
| 227 | } |
| 228 | |
| 229 | inline uint32_t output_y_end() const { |
| 230 | if (this->output_y_end_ == std::numeric_limits<uint32_t>::max()) { |
| 231 | return output_height(); |
| 232 | } else { |
| 233 | return this->output_y_end_; |
| 234 | } |
| 235 | } |
| 236 | |
| 237 | inline size_t input_pixel_stride() const { |
| 238 | return input_channels(); |
| 239 | } |
| 240 | |
| 241 | inline size_t output_pixel_stride() const { |
| 242 | return output_channels(); |
| 243 | } |
| 244 | |
| 245 | inline size_t output_height() const { |
| 246 | const size_t padded_input_height = padding_top() + input_height() + padding_bottom(); |
Marat Dukhan | 441e221 | 2019-12-04 18:30:49 -0800 | [diff] [blame] | 247 | if (padded_input_height < kernel_height()) { |
| 248 | return 0; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 249 | } else { |
| 250 | return (padded_input_height - kernel_height()) / subsampling_height() + 1; |
| 251 | } |
| 252 | } |
| 253 | |
| 254 | inline size_t output_width() const { |
| 255 | const size_t padded_input_width = padding_left() + input_width() + padding_right(); |
Marat Dukhan | 441e221 | 2019-12-04 18:30:49 -0800 | [diff] [blame] | 256 | if (padded_input_width < kernel_width()) { |
| 257 | return 0; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 258 | } else { |
| 259 | return (padded_input_width - kernel_width()) / subsampling_width() + 1; |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | inline ConvHWCMicrokernelTester& qmin(uint8_t qmin) { |
| 264 | this->qmin_ = qmin; |
| 265 | return *this; |
| 266 | } |
| 267 | |
| 268 | inline uint8_t qmin() const { |
| 269 | return this->qmin_; |
| 270 | } |
| 271 | |
| 272 | inline ConvHWCMicrokernelTester& qmax(uint8_t qmax) { |
| 273 | this->qmax_ = qmax; |
| 274 | return *this; |
| 275 | } |
| 276 | |
| 277 | inline uint8_t qmax() const { |
| 278 | return this->qmax_; |
| 279 | } |
| 280 | |
| 281 | inline ConvHWCMicrokernelTester& iterations(size_t iterations) { |
| 282 | this->iterations_ = iterations; |
| 283 | return *this; |
| 284 | } |
| 285 | |
| 286 | inline size_t iterations() const { |
| 287 | return this->iterations_; |
| 288 | } |
| 289 | |
| 290 | void Test(xnn_f32_conv_hwc_ukernel_function conv, Variant variant = Variant::Native) const { |
| 291 | ASSERT_LT(output_y_start(), output_height()); |
| 292 | ASSERT_LE(output_y_end(), output_height()); |
| 293 | ASSERT_GT(output_y_end(), output_y_start()); |
Marat Dukhan | 441e221 | 2019-12-04 18:30:49 -0800 | [diff] [blame] | 294 | ASSERT_GE(output_width(), 1); |
| 295 | ASSERT_GE(output_height(), 1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 296 | |
| 297 | std::random_device random_device; |
| 298 | auto rng = std::mt19937(random_device()); |
| 299 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.1f, 1.0f), rng); |
| 300 | |
| 301 | std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + |
| 302 | batch_size() * ((input_height() * input_width() - 1) * input_pixel_stride() + input_channels())); |
| 303 | std::vector<float> zero(XNN_EXTRA_BYTES / sizeof(float) + input_width() * input_channels()); |
| 304 | std::vector<float> kernel(output_channels() * kernel_height() * kernel_width() * input_channels()); |
| 305 | std::vector<float> bias(output_channels()); |
| 306 | std::vector<float> output(batch_size() * ((output_height() * output_width() - 1) * output_pixel_stride() + output_channels())); |
| 307 | std::vector<float> output_ref(batch_size() * output_height() * output_width() * output_channels()); |
Marat Dukhan | 9594db0 | 2019-12-05 14:32:37 -0800 | [diff] [blame] | 308 | std::vector<float, AlignedAllocator<float, 64>> packed_weights((input_channels() * kernel_height() * kernel_width() + 1) * packed_output_channels()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 309 | |
| 310 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 311 | std::generate(input.begin(), input.end(), std::ref(f32rng)); |
| 312 | std::generate(kernel.begin(), kernel.end(), std::ref(f32rng)); |
| 313 | std::generate(bias.begin(), bias.end(), std::ref(f32rng)); |
| 314 | std::fill(output.begin(), output.end(), nanf("")); |
| 315 | std::fill(packed_weights.begin(), packed_weights.end(), 0.0f); |
| 316 | |
| 317 | xnn_pack_f32_dconv_oki_w( |
| 318 | output_channels(), |
| 319 | input_channels(), |
| 320 | output_channels_tile(), |
| 321 | kernel_height(), kernel_width(), |
| 322 | kernel.data(), bias.data(), packed_weights.data()); |
| 323 | |
| 324 | // Compute reference results, without clamping. |
| 325 | for (size_t i = 0; i < batch_size(); i++) { |
| 326 | for (size_t oy = 0; oy < output_height(); oy++) { |
| 327 | for (size_t ox = 0; ox < output_width(); ox++) { |
| 328 | for (size_t oc = 0; oc < output_channels(); oc++) { |
| 329 | float acc = bias[oc]; |
| 330 | for (size_t ky = 0; ky < kernel_height(); ky++) { |
| 331 | const size_t iy = oy * subsampling_height() + ky - padding_top(); |
| 332 | if (iy < input_height()) { |
| 333 | for (size_t kx = 0; kx < kernel_width(); kx++) { |
| 334 | const size_t ix = ox * subsampling_width() + kx - padding_left(); |
| 335 | if (ix < input_width()) { |
| 336 | for (size_t ic = 0; ic < input_channels(); ic++) { |
| 337 | acc += |
| 338 | input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + ic] * |
| 339 | kernel[((oc * kernel_height() + ky) * kernel_width() + kx) * input_channels() + ic]; |
| 340 | } |
| 341 | } |
| 342 | } |
| 343 | } |
| 344 | } |
| 345 | output_ref[((i * output_height() + oy) * output_width() + ox) * output_channels() + oc] = acc; |
| 346 | } |
| 347 | } |
| 348 | } |
| 349 | } |
| 350 | |
| 351 | // Compute clamping parameters. |
| 352 | const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend()); |
| 353 | const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend()); |
| 354 | |
| 355 | const float output_min = accumulated_min + (accumulated_max - accumulated_min) / 255.0f * float(qmin()); |
| 356 | const float output_max = accumulated_max - (accumulated_max - accumulated_min) / 255.0f * float(255 - qmax()); |
| 357 | |
| 358 | // Clamp reference results. |
| 359 | for (float& value : output_ref) { |
| 360 | value = std::max(std::min(value, output_max), output_min); |
| 361 | } |
| 362 | |
| 363 | // Prepare output parameters. |
| 364 | xnn_f32_output_params output_params = { }; |
| 365 | switch (variant) { |
| 366 | case Variant::Native: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 367 | output_params = xnn_init_f32_output_params(output_min, output_max); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 368 | break; |
| 369 | case Variant::Scalar: |
Marat Dukhan | eeaa7bd | 2019-10-25 17:31:25 -0700 | [diff] [blame] | 370 | output_params = xnn_init_scalar_f32_output_params(output_min, output_max); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 371 | break; |
| 372 | } |
| 373 | |
| 374 | // Call optimized micro-kernel. |
| 375 | conv( |
| 376 | input_height(), input_width(), |
| 377 | output_y_start(), output_y_end(), |
| 378 | input.data(), zero.data(), packed_weights.data(), output.data(), |
| 379 | padding_top(), output_channels(), |
| 380 | output_pixel_stride() * output_width() * sizeof(float), |
| 381 | output_pixel_stride() * sizeof(float), |
| 382 | &output_params); |
| 383 | |
| 384 | // Verify results. |
| 385 | for (size_t i = 0; i < batch_size(); i++) { |
| 386 | for (size_t y = output_y_start(); y < output_y_end(); y++) { |
| 387 | for (size_t x = 0; x < output_width(); x++) { |
| 388 | for (size_t c = 0; c < output_channels(); c++) { |
| 389 | ASSERT_GE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_min) |
| 390 | << "(x, y) = (" << x << ", " << y << "), channel = " << c; |
| 391 | ASSERT_LE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_max) |
| 392 | << "(x, y) = (" << x << ", " << y << "), channel = " << c; |
| 393 | ASSERT_NEAR( |
| 394 | output_ref[((i * output_height() + y) * output_width() + x) * output_channels() + c], |
| 395 | output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], |
| 396 | 1.0e-4 * std::abs(output_ref[((i * output_height() + y) * output_width() + x) * output_channels() + c])) |
| 397 | << "(x, y) = (" << x << ", " << y << "), channel = " << c; |
| 398 | } |
| 399 | } |
| 400 | } |
| 401 | } |
| 402 | } |
| 403 | } |
| 404 | |
| 405 | private: |
| 406 | uint32_t padding_top_{0}; |
| 407 | uint32_t padding_right_{0}; |
| 408 | uint32_t padding_bottom_{0}; |
| 409 | uint32_t padding_left_{0}; |
| 410 | size_t input_height_{1}; |
| 411 | size_t input_width_{1}; |
| 412 | size_t input_channels_{1}; |
| 413 | size_t output_channels_{1}; |
| 414 | uint32_t output_channels_tile_{1}; |
| 415 | size_t batch_size_{1}; |
| 416 | uint32_t kernel_height_{1}; |
| 417 | uint32_t kernel_width_{1}; |
| 418 | uint32_t subsampling_height_{1}; |
| 419 | uint32_t subsampling_width_{1}; |
| 420 | uint32_t output_y_start_{0}; |
| 421 | uint32_t output_y_end_{std::numeric_limits<uint32_t>::max()}; |
| 422 | uint8_t qmin_{0}; |
| 423 | uint8_t qmax_{255}; |
| 424 | size_t iterations_{1}; |
| 425 | }; |