Marat Dukhan | 0461f2d | 2021-08-08 12:36:29 -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 | #include <gtest/gtest.h> |
| 7 | |
| 8 | #include <xnnpack/common.h> |
| 9 | #include <xnnpack/isa-checks.h> |
| 10 | |
| 11 | #include <xnnpack/pad.h> |
| 12 | #include "pad-microkernel-tester.h" |
| 13 | |
| 14 | |
| 15 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 16 | TEST(XX_PAD__NEON, fulltile_copy_channels_eq_16) { |
| 17 | TEST_REQUIRES_ARM_NEON; |
| 18 | PadMicrokernelTester() |
| 19 | .rows(1) |
| 20 | .input_channels(16) |
| 21 | .Test(xnn_xx_pad_ukernel__neon); |
| 22 | } |
| 23 | |
| 24 | TEST(XX_PAD__NEON, fulltile_copy_channels_div_16) { |
| 25 | TEST_REQUIRES_ARM_NEON; |
| 26 | for (size_t channels = 32; channels <= 48; channels += 16) { |
| 27 | PadMicrokernelTester() |
| 28 | .rows(1) |
| 29 | .input_channels(channels) |
| 30 | .Test(xnn_xx_pad_ukernel__neon); |
| 31 | } |
| 32 | } |
| 33 | |
| 34 | TEST(XX_PAD__NEON, fulltile_copy_channels_lt_16) { |
| 35 | TEST_REQUIRES_ARM_NEON; |
| 36 | for (size_t channels = 1; channels < 16; channels++) { |
| 37 | PadMicrokernelTester() |
| 38 | .rows(1) |
| 39 | .input_channels(channels) |
| 40 | .Test(xnn_xx_pad_ukernel__neon); |
| 41 | } |
| 42 | } |
| 43 | |
| 44 | TEST(XX_PAD__NEON, fulltile_copy_channels_gt_16) { |
| 45 | TEST_REQUIRES_ARM_NEON; |
| 46 | for (size_t channels = 17; channels < 32; channels++) { |
| 47 | PadMicrokernelTester() |
| 48 | .rows(1) |
| 49 | .input_channels(channels) |
| 50 | .Test(xnn_xx_pad_ukernel__neon); |
| 51 | } |
| 52 | } |
| 53 | |
| 54 | TEST(XX_PAD__NEON, fulltile_pre_padding_eq_1) { |
| 55 | TEST_REQUIRES_ARM_NEON; |
| 56 | PadMicrokernelTester() |
| 57 | .rows(1) |
| 58 | .input_channels(1) |
| 59 | .pre_padding(1) |
| 60 | .Test(xnn_xx_pad_ukernel__neon); |
| 61 | } |
| 62 | |
| 63 | TEST(XX_PAD__NEON, fulltile_pre_padding_eq_2) { |
| 64 | TEST_REQUIRES_ARM_NEON; |
| 65 | PadMicrokernelTester() |
| 66 | .rows(1) |
| 67 | .input_channels(1) |
| 68 | .pre_padding(2) |
| 69 | .Test(xnn_xx_pad_ukernel__neon); |
| 70 | } |
| 71 | |
| 72 | TEST(XX_PAD__NEON, fulltile_pre_padding_eq_4) { |
| 73 | TEST_REQUIRES_ARM_NEON; |
| 74 | PadMicrokernelTester() |
| 75 | .rows(1) |
| 76 | .input_channels(1) |
| 77 | .pre_padding(4) |
| 78 | .Test(xnn_xx_pad_ukernel__neon); |
| 79 | } |
| 80 | |
| 81 | TEST(XX_PAD__NEON, fulltile_pre_padding_eq_16) { |
| 82 | TEST_REQUIRES_ARM_NEON; |
| 83 | PadMicrokernelTester() |
| 84 | .rows(1) |
| 85 | .input_channels(1) |
| 86 | .pre_padding(16) |
| 87 | .Test(xnn_xx_pad_ukernel__neon); |
| 88 | } |
| 89 | |
| 90 | TEST(XX_PAD__NEON, fulltile_pre_padding_div_16) { |
| 91 | TEST_REQUIRES_ARM_NEON; |
| 92 | for (size_t pre_padding = 32; pre_padding <= 48; pre_padding += 16) { |
| 93 | PadMicrokernelTester() |
| 94 | .rows(1) |
| 95 | .input_channels(1) |
| 96 | .pre_padding(pre_padding) |
| 97 | .Test(xnn_xx_pad_ukernel__neon); |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | TEST(XX_PAD__NEON, fulltile_pre_padding_lt_16) { |
| 102 | TEST_REQUIRES_ARM_NEON; |
| 103 | for (size_t pre_padding = 1; pre_padding < 16; pre_padding++) { |
| 104 | PadMicrokernelTester() |
| 105 | .rows(1) |
| 106 | .input_channels(1) |
| 107 | .pre_padding(pre_padding) |
| 108 | .Test(xnn_xx_pad_ukernel__neon); |
| 109 | } |
| 110 | } |
| 111 | |
| 112 | TEST(XX_PAD__NEON, fulltile_pre_padding_gt_16) { |
| 113 | TEST_REQUIRES_ARM_NEON; |
| 114 | for (size_t pre_padding = 17; pre_padding < 32; pre_padding++) { |
| 115 | PadMicrokernelTester() |
| 116 | .rows(1) |
| 117 | .input_channels(1) |
| 118 | .pre_padding(pre_padding) |
| 119 | .Test(xnn_xx_pad_ukernel__neon); |
| 120 | } |
| 121 | } |
| 122 | |
| 123 | TEST(XX_PAD__NEON, fulltile_post_padding_eq_1) { |
| 124 | TEST_REQUIRES_ARM_NEON; |
| 125 | PadMicrokernelTester() |
| 126 | .rows(1) |
| 127 | .input_channels(1) |
| 128 | .post_padding(1) |
| 129 | .Test(xnn_xx_pad_ukernel__neon); |
| 130 | } |
| 131 | |
| 132 | TEST(XX_PAD__NEON, fulltile_post_padding_eq_2) { |
| 133 | TEST_REQUIRES_ARM_NEON; |
| 134 | PadMicrokernelTester() |
| 135 | .rows(1) |
| 136 | .input_channels(1) |
| 137 | .post_padding(2) |
| 138 | .Test(xnn_xx_pad_ukernel__neon); |
| 139 | } |
| 140 | |
| 141 | TEST(XX_PAD__NEON, fulltile_post_padding_eq_4) { |
| 142 | TEST_REQUIRES_ARM_NEON; |
| 143 | PadMicrokernelTester() |
| 144 | .rows(1) |
| 145 | .input_channels(1) |
| 146 | .post_padding(4) |
| 147 | .Test(xnn_xx_pad_ukernel__neon); |
| 148 | } |
| 149 | |
| 150 | TEST(XX_PAD__NEON, fulltile_post_padding_eq_16) { |
| 151 | TEST_REQUIRES_ARM_NEON; |
| 152 | PadMicrokernelTester() |
| 153 | .rows(1) |
| 154 | .input_channels(1) |
| 155 | .post_padding(16) |
| 156 | .Test(xnn_xx_pad_ukernel__neon); |
| 157 | } |
| 158 | |
| 159 | TEST(XX_PAD__NEON, fulltile_post_padding_div_16) { |
| 160 | TEST_REQUIRES_ARM_NEON; |
| 161 | for (size_t post_padding = 32; post_padding <= 48; post_padding += 16) { |
| 162 | PadMicrokernelTester() |
| 163 | .rows(1) |
| 164 | .input_channels(1) |
| 165 | .post_padding(post_padding) |
| 166 | .Test(xnn_xx_pad_ukernel__neon); |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | TEST(XX_PAD__NEON, fulltile_post_padding_lt_16) { |
| 171 | TEST_REQUIRES_ARM_NEON; |
| 172 | for (size_t post_padding = 1; post_padding < 16; post_padding++) { |
| 173 | PadMicrokernelTester() |
| 174 | .rows(1) |
| 175 | .input_channels(1) |
| 176 | .post_padding(post_padding) |
| 177 | .Test(xnn_xx_pad_ukernel__neon); |
| 178 | } |
| 179 | } |
| 180 | |
| 181 | TEST(XX_PAD__NEON, fulltile_post_padding_gt_16) { |
| 182 | TEST_REQUIRES_ARM_NEON; |
| 183 | for (size_t post_padding = 17; post_padding < 32; post_padding++) { |
| 184 | PadMicrokernelTester() |
| 185 | .rows(1) |
| 186 | .input_channels(1) |
| 187 | .pre_padding(post_padding) |
| 188 | .Test(xnn_xx_pad_ukernel__neon); |
| 189 | } |
| 190 | } |
| 191 | |
| 192 | TEST(XX_PAD__NEON, multitile) { |
| 193 | TEST_REQUIRES_ARM_NEON; |
| 194 | for (size_t rows = 2; rows <= 5; rows++) { |
| 195 | for (size_t channels = 1; channels < 48; channels += 3) { |
| 196 | PadMicrokernelTester() |
| 197 | .rows(rows) |
| 198 | .input_channels(channels) |
| 199 | .pre_padding(channels) |
| 200 | .post_padding(channels) |
| 201 | .Test(xnn_xx_pad_ukernel__neon); |
| 202 | } |
| 203 | } |
| 204 | } |
| 205 | |
| 206 | TEST(XX_PAD__NEON, multitile_with_input_stride) { |
| 207 | TEST_REQUIRES_ARM_NEON; |
| 208 | for (size_t rows = 2; rows <= 5; rows++) { |
| 209 | for (size_t channels = 1; channels < 48; channels += 3) { |
| 210 | PadMicrokernelTester() |
| 211 | .rows(rows) |
| 212 | .input_channels(channels) |
| 213 | .pre_padding(channels) |
| 214 | .post_padding(channels) |
| 215 | .input_stride(51) |
| 216 | .Test(xnn_xx_pad_ukernel__neon); |
| 217 | } |
| 218 | } |
| 219 | } |
| 220 | |
| 221 | TEST(XX_PAD__NEON, multitile_with_output_stride) { |
| 222 | TEST_REQUIRES_ARM_NEON; |
| 223 | for (size_t rows = 2; rows <= 5; rows++) { |
| 224 | for (size_t channels = 1; channels < 48; channels += 3) { |
| 225 | PadMicrokernelTester() |
| 226 | .rows(rows) |
| 227 | .input_channels(2 * channels) |
| 228 | .pre_padding(channels) |
| 229 | .post_padding(channels) |
| 230 | .output_stride(193) |
| 231 | .Test(xnn_xx_pad_ukernel__neon); |
| 232 | } |
| 233 | } |
| 234 | } |
| 235 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 236 | |
| 237 | |
| 238 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 239 | TEST(XX_PAD__SSE2, fulltile_copy_channels_eq_16) { |
| 240 | TEST_REQUIRES_X86_SSE2; |
| 241 | PadMicrokernelTester() |
| 242 | .rows(1) |
| 243 | .input_channels(16) |
| 244 | .Test(xnn_xx_pad_ukernel__sse2); |
| 245 | } |
| 246 | |
| 247 | TEST(XX_PAD__SSE2, fulltile_copy_channels_div_16) { |
| 248 | TEST_REQUIRES_X86_SSE2; |
| 249 | for (size_t channels = 32; channels <= 48; channels += 16) { |
| 250 | PadMicrokernelTester() |
| 251 | .rows(1) |
| 252 | .input_channels(channels) |
| 253 | .Test(xnn_xx_pad_ukernel__sse2); |
| 254 | } |
| 255 | } |
| 256 | |
| 257 | TEST(XX_PAD__SSE2, fulltile_copy_channels_lt_16) { |
| 258 | TEST_REQUIRES_X86_SSE2; |
| 259 | for (size_t channels = 1; channels < 16; channels++) { |
| 260 | PadMicrokernelTester() |
| 261 | .rows(1) |
| 262 | .input_channels(channels) |
| 263 | .Test(xnn_xx_pad_ukernel__sse2); |
| 264 | } |
| 265 | } |
| 266 | |
| 267 | TEST(XX_PAD__SSE2, fulltile_copy_channels_gt_16) { |
| 268 | TEST_REQUIRES_X86_SSE2; |
| 269 | for (size_t channels = 17; channels < 32; channels++) { |
| 270 | PadMicrokernelTester() |
| 271 | .rows(1) |
| 272 | .input_channels(channels) |
| 273 | .Test(xnn_xx_pad_ukernel__sse2); |
| 274 | } |
| 275 | } |
| 276 | |
| 277 | TEST(XX_PAD__SSE2, fulltile_pre_padding_eq_1) { |
| 278 | TEST_REQUIRES_X86_SSE2; |
| 279 | PadMicrokernelTester() |
| 280 | .rows(1) |
| 281 | .input_channels(1) |
| 282 | .pre_padding(1) |
| 283 | .Test(xnn_xx_pad_ukernel__sse2); |
| 284 | } |
| 285 | |
| 286 | TEST(XX_PAD__SSE2, fulltile_pre_padding_eq_2) { |
| 287 | TEST_REQUIRES_X86_SSE2; |
| 288 | PadMicrokernelTester() |
| 289 | .rows(1) |
| 290 | .input_channels(1) |
| 291 | .pre_padding(2) |
| 292 | .Test(xnn_xx_pad_ukernel__sse2); |
| 293 | } |
| 294 | |
| 295 | TEST(XX_PAD__SSE2, fulltile_pre_padding_eq_4) { |
| 296 | TEST_REQUIRES_X86_SSE2; |
| 297 | PadMicrokernelTester() |
| 298 | .rows(1) |
| 299 | .input_channels(1) |
| 300 | .pre_padding(4) |
| 301 | .Test(xnn_xx_pad_ukernel__sse2); |
| 302 | } |
| 303 | |
| 304 | TEST(XX_PAD__SSE2, fulltile_pre_padding_eq_16) { |
| 305 | TEST_REQUIRES_X86_SSE2; |
| 306 | PadMicrokernelTester() |
| 307 | .rows(1) |
| 308 | .input_channels(1) |
| 309 | .pre_padding(16) |
| 310 | .Test(xnn_xx_pad_ukernel__sse2); |
| 311 | } |
| 312 | |
| 313 | TEST(XX_PAD__SSE2, fulltile_pre_padding_div_16) { |
| 314 | TEST_REQUIRES_X86_SSE2; |
| 315 | for (size_t pre_padding = 32; pre_padding <= 48; pre_padding += 16) { |
| 316 | PadMicrokernelTester() |
| 317 | .rows(1) |
| 318 | .input_channels(1) |
| 319 | .pre_padding(pre_padding) |
| 320 | .Test(xnn_xx_pad_ukernel__sse2); |
| 321 | } |
| 322 | } |
| 323 | |
| 324 | TEST(XX_PAD__SSE2, fulltile_pre_padding_lt_16) { |
| 325 | TEST_REQUIRES_X86_SSE2; |
| 326 | for (size_t pre_padding = 1; pre_padding < 16; pre_padding++) { |
| 327 | PadMicrokernelTester() |
| 328 | .rows(1) |
| 329 | .input_channels(1) |
| 330 | .pre_padding(pre_padding) |
| 331 | .Test(xnn_xx_pad_ukernel__sse2); |
| 332 | } |
| 333 | } |
| 334 | |
| 335 | TEST(XX_PAD__SSE2, fulltile_pre_padding_gt_16) { |
| 336 | TEST_REQUIRES_X86_SSE2; |
| 337 | for (size_t pre_padding = 17; pre_padding < 32; pre_padding++) { |
| 338 | PadMicrokernelTester() |
| 339 | .rows(1) |
| 340 | .input_channels(1) |
| 341 | .pre_padding(pre_padding) |
| 342 | .Test(xnn_xx_pad_ukernel__sse2); |
| 343 | } |
| 344 | } |
| 345 | |
| 346 | TEST(XX_PAD__SSE2, fulltile_post_padding_eq_1) { |
| 347 | TEST_REQUIRES_X86_SSE2; |
| 348 | PadMicrokernelTester() |
| 349 | .rows(1) |
| 350 | .input_channels(1) |
| 351 | .post_padding(1) |
| 352 | .Test(xnn_xx_pad_ukernel__sse2); |
| 353 | } |
| 354 | |
| 355 | TEST(XX_PAD__SSE2, fulltile_post_padding_eq_2) { |
| 356 | TEST_REQUIRES_X86_SSE2; |
| 357 | PadMicrokernelTester() |
| 358 | .rows(1) |
| 359 | .input_channels(1) |
| 360 | .post_padding(2) |
| 361 | .Test(xnn_xx_pad_ukernel__sse2); |
| 362 | } |
| 363 | |
| 364 | TEST(XX_PAD__SSE2, fulltile_post_padding_eq_4) { |
| 365 | TEST_REQUIRES_X86_SSE2; |
| 366 | PadMicrokernelTester() |
| 367 | .rows(1) |
| 368 | .input_channels(1) |
| 369 | .post_padding(4) |
| 370 | .Test(xnn_xx_pad_ukernel__sse2); |
| 371 | } |
| 372 | |
| 373 | TEST(XX_PAD__SSE2, fulltile_post_padding_eq_16) { |
| 374 | TEST_REQUIRES_X86_SSE2; |
| 375 | PadMicrokernelTester() |
| 376 | .rows(1) |
| 377 | .input_channels(1) |
| 378 | .post_padding(16) |
| 379 | .Test(xnn_xx_pad_ukernel__sse2); |
| 380 | } |
| 381 | |
| 382 | TEST(XX_PAD__SSE2, fulltile_post_padding_div_16) { |
| 383 | TEST_REQUIRES_X86_SSE2; |
| 384 | for (size_t post_padding = 32; post_padding <= 48; post_padding += 16) { |
| 385 | PadMicrokernelTester() |
| 386 | .rows(1) |
| 387 | .input_channels(1) |
| 388 | .post_padding(post_padding) |
| 389 | .Test(xnn_xx_pad_ukernel__sse2); |
| 390 | } |
| 391 | } |
| 392 | |
| 393 | TEST(XX_PAD__SSE2, fulltile_post_padding_lt_16) { |
| 394 | TEST_REQUIRES_X86_SSE2; |
| 395 | for (size_t post_padding = 1; post_padding < 16; post_padding++) { |
| 396 | PadMicrokernelTester() |
| 397 | .rows(1) |
| 398 | .input_channels(1) |
| 399 | .post_padding(post_padding) |
| 400 | .Test(xnn_xx_pad_ukernel__sse2); |
| 401 | } |
| 402 | } |
| 403 | |
| 404 | TEST(XX_PAD__SSE2, fulltile_post_padding_gt_16) { |
| 405 | TEST_REQUIRES_X86_SSE2; |
| 406 | for (size_t post_padding = 17; post_padding < 32; post_padding++) { |
| 407 | PadMicrokernelTester() |
| 408 | .rows(1) |
| 409 | .input_channels(1) |
| 410 | .pre_padding(post_padding) |
| 411 | .Test(xnn_xx_pad_ukernel__sse2); |
| 412 | } |
| 413 | } |
| 414 | |
| 415 | TEST(XX_PAD__SSE2, multitile) { |
| 416 | TEST_REQUIRES_X86_SSE2; |
| 417 | for (size_t rows = 2; rows <= 5; rows++) { |
| 418 | for (size_t channels = 1; channels < 48; channels += 3) { |
| 419 | PadMicrokernelTester() |
| 420 | .rows(rows) |
| 421 | .input_channels(channels) |
| 422 | .pre_padding(channels) |
| 423 | .post_padding(channels) |
| 424 | .Test(xnn_xx_pad_ukernel__sse2); |
| 425 | } |
| 426 | } |
| 427 | } |
| 428 | |
| 429 | TEST(XX_PAD__SSE2, multitile_with_input_stride) { |
| 430 | TEST_REQUIRES_X86_SSE2; |
| 431 | for (size_t rows = 2; rows <= 5; rows++) { |
| 432 | for (size_t channels = 1; channels < 48; channels += 3) { |
| 433 | PadMicrokernelTester() |
| 434 | .rows(rows) |
| 435 | .input_channels(channels) |
| 436 | .pre_padding(channels) |
| 437 | .post_padding(channels) |
| 438 | .input_stride(51) |
| 439 | .Test(xnn_xx_pad_ukernel__sse2); |
| 440 | } |
| 441 | } |
| 442 | } |
| 443 | |
| 444 | TEST(XX_PAD__SSE2, multitile_with_output_stride) { |
| 445 | TEST_REQUIRES_X86_SSE2; |
| 446 | for (size_t rows = 2; rows <= 5; rows++) { |
| 447 | for (size_t channels = 1; channels < 48; channels += 3) { |
| 448 | PadMicrokernelTester() |
| 449 | .rows(rows) |
| 450 | .input_channels(2 * channels) |
| 451 | .pre_padding(channels) |
| 452 | .post_padding(channels) |
| 453 | .output_stride(193) |
| 454 | .Test(xnn_xx_pad_ukernel__sse2); |
| 455 | } |
| 456 | } |
| 457 | } |
| 458 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 459 | |
| 460 | |
| 461 | #if XNN_ARCH_WASMSIMD |
| 462 | TEST(XX_PAD__WASMSIMD, fulltile_copy_channels_eq_16) { |
| 463 | PadMicrokernelTester() |
| 464 | .rows(1) |
| 465 | .input_channels(16) |
| 466 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 467 | } |
| 468 | |
| 469 | TEST(XX_PAD__WASMSIMD, fulltile_copy_channels_div_16) { |
| 470 | for (size_t channels = 32; channels <= 48; channels += 16) { |
| 471 | PadMicrokernelTester() |
| 472 | .rows(1) |
| 473 | .input_channels(channels) |
| 474 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 475 | } |
| 476 | } |
| 477 | |
| 478 | TEST(XX_PAD__WASMSIMD, fulltile_copy_channels_lt_16) { |
| 479 | for (size_t channels = 1; channels < 16; channels++) { |
| 480 | PadMicrokernelTester() |
| 481 | .rows(1) |
| 482 | .input_channels(channels) |
| 483 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 484 | } |
| 485 | } |
| 486 | |
| 487 | TEST(XX_PAD__WASMSIMD, fulltile_copy_channels_gt_16) { |
| 488 | for (size_t channels = 17; channels < 32; channels++) { |
| 489 | PadMicrokernelTester() |
| 490 | .rows(1) |
| 491 | .input_channels(channels) |
| 492 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 493 | } |
| 494 | } |
| 495 | |
| 496 | TEST(XX_PAD__WASMSIMD, fulltile_pre_padding_eq_1) { |
| 497 | PadMicrokernelTester() |
| 498 | .rows(1) |
| 499 | .input_channels(1) |
| 500 | .pre_padding(1) |
| 501 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 502 | } |
| 503 | |
| 504 | TEST(XX_PAD__WASMSIMD, fulltile_pre_padding_eq_2) { |
| 505 | PadMicrokernelTester() |
| 506 | .rows(1) |
| 507 | .input_channels(1) |
| 508 | .pre_padding(2) |
| 509 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 510 | } |
| 511 | |
| 512 | TEST(XX_PAD__WASMSIMD, fulltile_pre_padding_eq_4) { |
| 513 | PadMicrokernelTester() |
| 514 | .rows(1) |
| 515 | .input_channels(1) |
| 516 | .pre_padding(4) |
| 517 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 518 | } |
| 519 | |
| 520 | TEST(XX_PAD__WASMSIMD, fulltile_pre_padding_eq_16) { |
| 521 | PadMicrokernelTester() |
| 522 | .rows(1) |
| 523 | .input_channels(1) |
| 524 | .pre_padding(16) |
| 525 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 526 | } |
| 527 | |
| 528 | TEST(XX_PAD__WASMSIMD, fulltile_pre_padding_div_16) { |
| 529 | for (size_t pre_padding = 32; pre_padding <= 48; pre_padding += 16) { |
| 530 | PadMicrokernelTester() |
| 531 | .rows(1) |
| 532 | .input_channels(1) |
| 533 | .pre_padding(pre_padding) |
| 534 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 535 | } |
| 536 | } |
| 537 | |
| 538 | TEST(XX_PAD__WASMSIMD, fulltile_pre_padding_lt_16) { |
| 539 | for (size_t pre_padding = 1; pre_padding < 16; pre_padding++) { |
| 540 | PadMicrokernelTester() |
| 541 | .rows(1) |
| 542 | .input_channels(1) |
| 543 | .pre_padding(pre_padding) |
| 544 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 545 | } |
| 546 | } |
| 547 | |
| 548 | TEST(XX_PAD__WASMSIMD, fulltile_pre_padding_gt_16) { |
| 549 | for (size_t pre_padding = 17; pre_padding < 32; pre_padding++) { |
| 550 | PadMicrokernelTester() |
| 551 | .rows(1) |
| 552 | .input_channels(1) |
| 553 | .pre_padding(pre_padding) |
| 554 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 555 | } |
| 556 | } |
| 557 | |
| 558 | TEST(XX_PAD__WASMSIMD, fulltile_post_padding_eq_1) { |
| 559 | PadMicrokernelTester() |
| 560 | .rows(1) |
| 561 | .input_channels(1) |
| 562 | .post_padding(1) |
| 563 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 564 | } |
| 565 | |
| 566 | TEST(XX_PAD__WASMSIMD, fulltile_post_padding_eq_2) { |
| 567 | PadMicrokernelTester() |
| 568 | .rows(1) |
| 569 | .input_channels(1) |
| 570 | .post_padding(2) |
| 571 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 572 | } |
| 573 | |
| 574 | TEST(XX_PAD__WASMSIMD, fulltile_post_padding_eq_4) { |
| 575 | PadMicrokernelTester() |
| 576 | .rows(1) |
| 577 | .input_channels(1) |
| 578 | .post_padding(4) |
| 579 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 580 | } |
| 581 | |
| 582 | TEST(XX_PAD__WASMSIMD, fulltile_post_padding_eq_16) { |
| 583 | PadMicrokernelTester() |
| 584 | .rows(1) |
| 585 | .input_channels(1) |
| 586 | .post_padding(16) |
| 587 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 588 | } |
| 589 | |
| 590 | TEST(XX_PAD__WASMSIMD, fulltile_post_padding_div_16) { |
| 591 | for (size_t post_padding = 32; post_padding <= 48; post_padding += 16) { |
| 592 | PadMicrokernelTester() |
| 593 | .rows(1) |
| 594 | .input_channels(1) |
| 595 | .post_padding(post_padding) |
| 596 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 597 | } |
| 598 | } |
| 599 | |
| 600 | TEST(XX_PAD__WASMSIMD, fulltile_post_padding_lt_16) { |
| 601 | for (size_t post_padding = 1; post_padding < 16; post_padding++) { |
| 602 | PadMicrokernelTester() |
| 603 | .rows(1) |
| 604 | .input_channels(1) |
| 605 | .post_padding(post_padding) |
| 606 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 607 | } |
| 608 | } |
| 609 | |
| 610 | TEST(XX_PAD__WASMSIMD, fulltile_post_padding_gt_16) { |
| 611 | for (size_t post_padding = 17; post_padding < 32; post_padding++) { |
| 612 | PadMicrokernelTester() |
| 613 | .rows(1) |
| 614 | .input_channels(1) |
| 615 | .pre_padding(post_padding) |
| 616 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 617 | } |
| 618 | } |
| 619 | |
| 620 | TEST(XX_PAD__WASMSIMD, multitile) { |
| 621 | for (size_t rows = 2; rows <= 5; rows++) { |
| 622 | for (size_t channels = 1; channels < 48; channels += 3) { |
| 623 | PadMicrokernelTester() |
| 624 | .rows(rows) |
| 625 | .input_channels(channels) |
| 626 | .pre_padding(channels) |
| 627 | .post_padding(channels) |
| 628 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 629 | } |
| 630 | } |
| 631 | } |
| 632 | |
| 633 | TEST(XX_PAD__WASMSIMD, multitile_with_input_stride) { |
| 634 | for (size_t rows = 2; rows <= 5; rows++) { |
| 635 | for (size_t channels = 1; channels < 48; channels += 3) { |
| 636 | PadMicrokernelTester() |
| 637 | .rows(rows) |
| 638 | .input_channels(channels) |
| 639 | .pre_padding(channels) |
| 640 | .post_padding(channels) |
| 641 | .input_stride(51) |
| 642 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 643 | } |
| 644 | } |
| 645 | } |
| 646 | |
| 647 | TEST(XX_PAD__WASMSIMD, multitile_with_output_stride) { |
| 648 | for (size_t rows = 2; rows <= 5; rows++) { |
| 649 | for (size_t channels = 1; channels < 48; channels += 3) { |
| 650 | PadMicrokernelTester() |
| 651 | .rows(rows) |
| 652 | .input_channels(2 * channels) |
| 653 | .pre_padding(channels) |
| 654 | .post_padding(channels) |
| 655 | .output_stride(193) |
| 656 | .Test(xnn_xx_pad_ukernel__wasmsimd); |
| 657 | } |
| 658 | } |
| 659 | } |
| 660 | #endif // XNN_ARCH_WASMSIMD |
| 661 | |
| 662 | |
| 663 | TEST(XX_PAD__SCALAR, fulltile_copy_channels_eq_16) { |
| 664 | PadMicrokernelTester() |
| 665 | .rows(1) |
| 666 | .input_channels(16) |
| 667 | .Test(xnn_xx_pad_ukernel__scalar); |
| 668 | } |
| 669 | |
| 670 | TEST(XX_PAD__SCALAR, fulltile_copy_channels_div_16) { |
| 671 | for (size_t channels = 32; channels <= 48; channels += 16) { |
| 672 | PadMicrokernelTester() |
| 673 | .rows(1) |
| 674 | .input_channels(channels) |
| 675 | .Test(xnn_xx_pad_ukernel__scalar); |
| 676 | } |
| 677 | } |
| 678 | |
| 679 | TEST(XX_PAD__SCALAR, fulltile_copy_channels_lt_16) { |
| 680 | for (size_t channels = 1; channels < 16; channels++) { |
| 681 | PadMicrokernelTester() |
| 682 | .rows(1) |
| 683 | .input_channels(channels) |
| 684 | .Test(xnn_xx_pad_ukernel__scalar); |
| 685 | } |
| 686 | } |
| 687 | |
| 688 | TEST(XX_PAD__SCALAR, fulltile_copy_channels_gt_16) { |
| 689 | for (size_t channels = 17; channels < 32; channels++) { |
| 690 | PadMicrokernelTester() |
| 691 | .rows(1) |
| 692 | .input_channels(channels) |
| 693 | .Test(xnn_xx_pad_ukernel__scalar); |
| 694 | } |
| 695 | } |
| 696 | |
| 697 | TEST(XX_PAD__SCALAR, fulltile_pre_padding_eq_1) { |
| 698 | PadMicrokernelTester() |
| 699 | .rows(1) |
| 700 | .input_channels(1) |
| 701 | .pre_padding(1) |
| 702 | .Test(xnn_xx_pad_ukernel__scalar); |
| 703 | } |
| 704 | |
| 705 | TEST(XX_PAD__SCALAR, fulltile_pre_padding_eq_2) { |
| 706 | PadMicrokernelTester() |
| 707 | .rows(1) |
| 708 | .input_channels(1) |
| 709 | .pre_padding(2) |
| 710 | .Test(xnn_xx_pad_ukernel__scalar); |
| 711 | } |
| 712 | |
| 713 | TEST(XX_PAD__SCALAR, fulltile_pre_padding_eq_4) { |
| 714 | PadMicrokernelTester() |
| 715 | .rows(1) |
| 716 | .input_channels(1) |
| 717 | .pre_padding(4) |
| 718 | .Test(xnn_xx_pad_ukernel__scalar); |
| 719 | } |
| 720 | |
| 721 | TEST(XX_PAD__SCALAR, fulltile_pre_padding_div_4) { |
| 722 | for (size_t pre_padding = 8; pre_padding <= 12; pre_padding += 4) { |
| 723 | PadMicrokernelTester() |
| 724 | .rows(1) |
| 725 | .input_channels(1) |
| 726 | .pre_padding(pre_padding) |
| 727 | .Test(xnn_xx_pad_ukernel__scalar); |
| 728 | } |
| 729 | } |
| 730 | |
| 731 | TEST(XX_PAD__SCALAR, fulltile_pre_padding_lt_4) { |
| 732 | for (size_t pre_padding = 1; pre_padding < 4; pre_padding++) { |
| 733 | PadMicrokernelTester() |
| 734 | .rows(1) |
| 735 | .input_channels(1) |
| 736 | .pre_padding(pre_padding) |
| 737 | .Test(xnn_xx_pad_ukernel__scalar); |
| 738 | } |
| 739 | } |
| 740 | |
| 741 | TEST(XX_PAD__SCALAR, fulltile_pre_padding_gt_4) { |
| 742 | for (size_t pre_padding = 5; pre_padding < 8; pre_padding++) { |
| 743 | PadMicrokernelTester() |
| 744 | .rows(1) |
| 745 | .input_channels(1) |
| 746 | .pre_padding(pre_padding) |
| 747 | .Test(xnn_xx_pad_ukernel__scalar); |
| 748 | } |
| 749 | } |
| 750 | |
| 751 | TEST(XX_PAD__SCALAR, fulltile_post_padding_eq_1) { |
| 752 | PadMicrokernelTester() |
| 753 | .rows(1) |
| 754 | .input_channels(1) |
| 755 | .post_padding(1) |
| 756 | .Test(xnn_xx_pad_ukernel__scalar); |
| 757 | } |
| 758 | |
| 759 | TEST(XX_PAD__SCALAR, fulltile_post_padding_eq_2) { |
| 760 | PadMicrokernelTester() |
| 761 | .rows(1) |
| 762 | .input_channels(1) |
| 763 | .post_padding(2) |
| 764 | .Test(xnn_xx_pad_ukernel__scalar); |
| 765 | } |
| 766 | |
| 767 | TEST(XX_PAD__SCALAR, fulltile_post_padding_eq_4) { |
| 768 | PadMicrokernelTester() |
| 769 | .rows(1) |
| 770 | .input_channels(1) |
| 771 | .post_padding(4) |
| 772 | .Test(xnn_xx_pad_ukernel__scalar); |
| 773 | } |
| 774 | |
| 775 | TEST(XX_PAD__SCALAR, fulltile_post_padding_div_4) { |
| 776 | for (size_t post_padding = 8; post_padding <= 12; post_padding += 4) { |
| 777 | PadMicrokernelTester() |
| 778 | .rows(1) |
| 779 | .input_channels(1) |
| 780 | .post_padding(post_padding) |
| 781 | .Test(xnn_xx_pad_ukernel__scalar); |
| 782 | } |
| 783 | } |
| 784 | |
| 785 | TEST(XX_PAD__SCALAR, fulltile_post_padding_lt_4) { |
| 786 | for (size_t post_padding = 1; post_padding < 4; post_padding++) { |
| 787 | PadMicrokernelTester() |
| 788 | .rows(1) |
| 789 | .input_channels(1) |
| 790 | .post_padding(post_padding) |
| 791 | .Test(xnn_xx_pad_ukernel__scalar); |
| 792 | } |
| 793 | } |
| 794 | |
| 795 | TEST(XX_PAD__SCALAR, fulltile_post_padding_gt_4) { |
| 796 | for (size_t post_padding = 5; post_padding < 8; post_padding++) { |
| 797 | PadMicrokernelTester() |
| 798 | .rows(1) |
| 799 | .input_channels(1) |
| 800 | .pre_padding(post_padding) |
| 801 | .Test(xnn_xx_pad_ukernel__scalar); |
| 802 | } |
| 803 | } |
| 804 | |
| 805 | TEST(XX_PAD__SCALAR, multitile) { |
| 806 | for (size_t rows = 2; rows <= 5; rows++) { |
| 807 | for (size_t channels = 1; channels < 12; channels += 3) { |
| 808 | PadMicrokernelTester() |
| 809 | .rows(rows) |
| 810 | .input_channels(channels) |
| 811 | .pre_padding(channels) |
| 812 | .post_padding(channels) |
| 813 | .Test(xnn_xx_pad_ukernel__scalar); |
| 814 | } |
| 815 | } |
| 816 | } |
| 817 | |
| 818 | TEST(XX_PAD__SCALAR, multitile_with_input_stride) { |
| 819 | for (size_t rows = 2; rows <= 5; rows++) { |
| 820 | for (size_t channels = 1; channels < 12; channels += 3) { |
| 821 | PadMicrokernelTester() |
| 822 | .rows(rows) |
| 823 | .input_channels(channels) |
| 824 | .pre_padding(channels) |
| 825 | .post_padding(channels) |
| 826 | .input_stride(51) |
| 827 | .Test(xnn_xx_pad_ukernel__scalar); |
| 828 | } |
| 829 | } |
| 830 | } |
| 831 | |
| 832 | TEST(XX_PAD__SCALAR, multitile_with_output_stride) { |
| 833 | for (size_t rows = 2; rows <= 5; rows++) { |
| 834 | for (size_t channels = 1; channels < 12; channels += 3) { |
| 835 | PadMicrokernelTester() |
| 836 | .rows(rows) |
| 837 | .input_channels(2 * channels) |
| 838 | .pre_padding(channels) |
| 839 | .post_padding(channels) |
| 840 | .output_stride(193) |
| 841 | .Test(xnn_xx_pad_ukernel__scalar); |
| 842 | } |
| 843 | } |
| 844 | } |