| Frank Barchard | b196659 | 2020-05-12 13:47:06 -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 | // Auto-generated file. Do not edit! |
| 7 | // Specification: test/f16-prelu.yaml |
| 8 | // Generator: tools/generate-prelu-test.py |
| 9 | |
| 10 | |
| 11 | #include <gtest/gtest.h> |
| 12 | |
| 13 | #include <xnnpack/common.h> |
| 14 | #include <xnnpack/isa-checks.h> |
| 15 | |
| 16 | #include <xnnpack/prelu.h> |
| 17 | #include "prelu-microkernel-tester.h" |
| 18 | |
| 19 | |
| 20 | #if XNN_ARCH_ARM64 |
| 21 | TEST(F16_PRELU__NEONFP16ARITH_2X8, channels_eq_8) { |
| 22 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 23 | PReLUMicrokernelTester() |
| 24 | .rows(2) |
| 25 | .channels(8) |
| 26 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 27 | } |
| 28 | |
| 29 | TEST(F16_PRELU__NEONFP16ARITH_2X8, channels_div_8) { |
| 30 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 31 | for (size_t channels = 16; channels < 80; channels += 8) { |
| 32 | PReLUMicrokernelTester() |
| 33 | .rows(2) |
| 34 | .channels(channels) |
| 35 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 36 | } |
| 37 | } |
| 38 | |
| 39 | TEST(F16_PRELU__NEONFP16ARITH_2X8, channels_lt_8) { |
| 40 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 41 | for (size_t channels = 1; channels < 8; channels++) { |
| 42 | PReLUMicrokernelTester() |
| 43 | .rows(2) |
| 44 | .channels(channels) |
| 45 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 46 | } |
| 47 | } |
| 48 | |
| 49 | TEST(F16_PRELU__NEONFP16ARITH_2X8, channels_gt_8) { |
| 50 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 51 | for (size_t channels = 9; channels < 16; channels++) { |
| 52 | PReLUMicrokernelTester() |
| 53 | .rows(2) |
| 54 | .channels(channels) |
| 55 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 56 | } |
| 57 | } |
| 58 | |
| 59 | TEST(F16_PRELU__NEONFP16ARITH_2X8, rows_lt_2) { |
| 60 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 61 | for (size_t rows = 1; rows < 2; rows++) { |
| 62 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 63 | PReLUMicrokernelTester() |
| 64 | .rows(rows) |
| 65 | .channels(channels) |
| 66 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 67 | } |
| 68 | } |
| 69 | } |
| 70 | |
| 71 | TEST(F16_PRELU__NEONFP16ARITH_2X8, rows_div_2) { |
| 72 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 73 | for (size_t rows = 4; rows <= 8; rows += 2) { |
| 74 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 75 | PReLUMicrokernelTester() |
| 76 | .rows(rows) |
| 77 | .channels(channels) |
| 78 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 79 | } |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | TEST(F16_PRELU__NEONFP16ARITH_2X8, rows_gt_2) { |
| 84 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 85 | for (size_t rows = 3; rows < 4; rows++) { |
| 86 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 87 | PReLUMicrokernelTester() |
| 88 | .rows(rows) |
| 89 | .channels(channels) |
| 90 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 91 | } |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | TEST(F16_PRELU__NEONFP16ARITH_2X8, input_stride) { |
| 96 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 97 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 98 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 99 | PReLUMicrokernelTester() |
| 100 | .rows(rows) |
| 101 | .channels(channels) |
| 102 | .input_stride(43) |
| 103 | .iterations(1) |
| 104 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 105 | } |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | TEST(F16_PRELU__NEONFP16ARITH_2X8, output_stride) { |
| 110 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 111 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 112 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 113 | PReLUMicrokernelTester() |
| 114 | .rows(rows) |
| 115 | .channels(channels) |
| 116 | .output_stride(43) |
| 117 | .iterations(1) |
| 118 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 119 | } |
| 120 | } |
| 121 | } |
| 122 | |
| 123 | TEST(F16_PRELU__NEONFP16ARITH_2X8, inplace) { |
| 124 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 125 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 126 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 127 | PReLUMicrokernelTester() |
| 128 | .rows(rows) |
| 129 | .channels(channels) |
| 130 | .inplace(true) |
| 131 | .iterations(1) |
| 132 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x8); |
| 133 | } |
| 134 | } |
| 135 | } |
| 136 | #endif // XNN_ARCH_ARM64 |
| 137 | |
| 138 | |
| 139 | #if XNN_ARCH_ARM64 |
| 140 | TEST(F16_PRELU__NEONFP16ARITH_2X16, channels_eq_16) { |
| 141 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 142 | PReLUMicrokernelTester() |
| 143 | .rows(2) |
| 144 | .channels(16) |
| 145 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 146 | } |
| 147 | |
| 148 | TEST(F16_PRELU__NEONFP16ARITH_2X16, channels_div_16) { |
| 149 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 150 | for (size_t channels = 32; channels < 160; channels += 16) { |
| 151 | PReLUMicrokernelTester() |
| 152 | .rows(2) |
| 153 | .channels(channels) |
| 154 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | TEST(F16_PRELU__NEONFP16ARITH_2X16, channels_lt_16) { |
| 159 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 160 | for (size_t channels = 1; channels < 16; channels++) { |
| 161 | PReLUMicrokernelTester() |
| 162 | .rows(2) |
| 163 | .channels(channels) |
| 164 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | TEST(F16_PRELU__NEONFP16ARITH_2X16, channels_gt_16) { |
| 169 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 170 | for (size_t channels = 17; channels < 32; channels++) { |
| 171 | PReLUMicrokernelTester() |
| 172 | .rows(2) |
| 173 | .channels(channels) |
| 174 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | TEST(F16_PRELU__NEONFP16ARITH_2X16, rows_lt_2) { |
| 179 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 180 | for (size_t rows = 1; rows < 2; rows++) { |
| 181 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 182 | PReLUMicrokernelTester() |
| 183 | .rows(rows) |
| 184 | .channels(channels) |
| 185 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 186 | } |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | TEST(F16_PRELU__NEONFP16ARITH_2X16, rows_div_2) { |
| 191 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 192 | for (size_t rows = 4; rows <= 8; rows += 2) { |
| 193 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 194 | PReLUMicrokernelTester() |
| 195 | .rows(rows) |
| 196 | .channels(channels) |
| 197 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 198 | } |
| 199 | } |
| 200 | } |
| 201 | |
| 202 | TEST(F16_PRELU__NEONFP16ARITH_2X16, rows_gt_2) { |
| 203 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 204 | for (size_t rows = 3; rows < 4; rows++) { |
| 205 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 206 | PReLUMicrokernelTester() |
| 207 | .rows(rows) |
| 208 | .channels(channels) |
| 209 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 210 | } |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | TEST(F16_PRELU__NEONFP16ARITH_2X16, input_stride) { |
| 215 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 216 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 217 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 218 | PReLUMicrokernelTester() |
| 219 | .rows(rows) |
| 220 | .channels(channels) |
| 221 | .input_stride(83) |
| 222 | .iterations(1) |
| 223 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 224 | } |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | TEST(F16_PRELU__NEONFP16ARITH_2X16, output_stride) { |
| 229 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 230 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 231 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 232 | PReLUMicrokernelTester() |
| 233 | .rows(rows) |
| 234 | .channels(channels) |
| 235 | .output_stride(83) |
| 236 | .iterations(1) |
| 237 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 238 | } |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | TEST(F16_PRELU__NEONFP16ARITH_2X16, inplace) { |
| 243 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 244 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 245 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 246 | PReLUMicrokernelTester() |
| 247 | .rows(rows) |
| 248 | .channels(channels) |
| 249 | .inplace(true) |
| 250 | .iterations(1) |
| 251 | .Test(xnn_f16_prelu_ukernel__neonfp16arith_2x16); |
| 252 | } |
| 253 | } |
| 254 | } |
| 255 | #endif // XNN_ARCH_ARM64 |
| Marat Dukhan | bd7f9a4 | 2022-01-10 20:04:36 -0800 | [diff] [blame] | 256 | |
| 257 | |
| 258 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 259 | TEST(F16_PRELU__F16C_2X8, channels_eq_8) { |
| 260 | TEST_REQUIRES_X86_F16C; |
| 261 | PReLUMicrokernelTester() |
| 262 | .rows(2) |
| 263 | .channels(8) |
| 264 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 265 | } |
| 266 | |
| 267 | TEST(F16_PRELU__F16C_2X8, channels_div_8) { |
| 268 | TEST_REQUIRES_X86_F16C; |
| 269 | for (size_t channels = 16; channels < 80; channels += 8) { |
| 270 | PReLUMicrokernelTester() |
| 271 | .rows(2) |
| 272 | .channels(channels) |
| 273 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 274 | } |
| 275 | } |
| 276 | |
| 277 | TEST(F16_PRELU__F16C_2X8, channels_lt_8) { |
| 278 | TEST_REQUIRES_X86_F16C; |
| 279 | for (size_t channels = 1; channels < 8; channels++) { |
| 280 | PReLUMicrokernelTester() |
| 281 | .rows(2) |
| 282 | .channels(channels) |
| 283 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | TEST(F16_PRELU__F16C_2X8, channels_gt_8) { |
| 288 | TEST_REQUIRES_X86_F16C; |
| 289 | for (size_t channels = 9; channels < 16; channels++) { |
| 290 | PReLUMicrokernelTester() |
| 291 | .rows(2) |
| 292 | .channels(channels) |
| 293 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | TEST(F16_PRELU__F16C_2X8, rows_lt_2) { |
| 298 | TEST_REQUIRES_X86_F16C; |
| 299 | for (size_t rows = 1; rows < 2; rows++) { |
| 300 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 301 | PReLUMicrokernelTester() |
| 302 | .rows(rows) |
| 303 | .channels(channels) |
| 304 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 305 | } |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | TEST(F16_PRELU__F16C_2X8, rows_div_2) { |
| 310 | TEST_REQUIRES_X86_F16C; |
| 311 | for (size_t rows = 4; rows <= 8; rows += 2) { |
| 312 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 313 | PReLUMicrokernelTester() |
| 314 | .rows(rows) |
| 315 | .channels(channels) |
| 316 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 317 | } |
| 318 | } |
| 319 | } |
| 320 | |
| 321 | TEST(F16_PRELU__F16C_2X8, rows_gt_2) { |
| 322 | TEST_REQUIRES_X86_F16C; |
| 323 | for (size_t rows = 3; rows < 4; rows++) { |
| 324 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 325 | PReLUMicrokernelTester() |
| 326 | .rows(rows) |
| 327 | .channels(channels) |
| 328 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 329 | } |
| 330 | } |
| 331 | } |
| 332 | |
| 333 | TEST(F16_PRELU__F16C_2X8, input_stride) { |
| 334 | TEST_REQUIRES_X86_F16C; |
| 335 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 336 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 337 | PReLUMicrokernelTester() |
| 338 | .rows(rows) |
| 339 | .channels(channels) |
| 340 | .input_stride(43) |
| 341 | .iterations(1) |
| 342 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 343 | } |
| 344 | } |
| 345 | } |
| 346 | |
| 347 | TEST(F16_PRELU__F16C_2X8, output_stride) { |
| 348 | TEST_REQUIRES_X86_F16C; |
| 349 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 350 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 351 | PReLUMicrokernelTester() |
| 352 | .rows(rows) |
| 353 | .channels(channels) |
| 354 | .output_stride(43) |
| 355 | .iterations(1) |
| 356 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 357 | } |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | TEST(F16_PRELU__F16C_2X8, inplace) { |
| 362 | TEST_REQUIRES_X86_F16C; |
| 363 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 364 | for (size_t channels = 1; channels <= 40; channels += 7) { |
| 365 | PReLUMicrokernelTester() |
| 366 | .rows(rows) |
| 367 | .channels(channels) |
| 368 | .inplace(true) |
| 369 | .iterations(1) |
| 370 | .Test(xnn_f16_prelu_ukernel__f16c_2x8); |
| 371 | } |
| 372 | } |
| 373 | } |
| 374 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 375 | |
| 376 | |
| 377 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 378 | TEST(F16_PRELU__F16C_2X16, channels_eq_16) { |
| 379 | TEST_REQUIRES_X86_F16C; |
| 380 | PReLUMicrokernelTester() |
| 381 | .rows(2) |
| 382 | .channels(16) |
| 383 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 384 | } |
| 385 | |
| 386 | TEST(F16_PRELU__F16C_2X16, channels_div_16) { |
| 387 | TEST_REQUIRES_X86_F16C; |
| 388 | for (size_t channels = 32; channels < 160; channels += 16) { |
| 389 | PReLUMicrokernelTester() |
| 390 | .rows(2) |
| 391 | .channels(channels) |
| 392 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 393 | } |
| 394 | } |
| 395 | |
| 396 | TEST(F16_PRELU__F16C_2X16, channels_lt_16) { |
| 397 | TEST_REQUIRES_X86_F16C; |
| 398 | for (size_t channels = 1; channels < 16; channels++) { |
| 399 | PReLUMicrokernelTester() |
| 400 | .rows(2) |
| 401 | .channels(channels) |
| 402 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 403 | } |
| 404 | } |
| 405 | |
| 406 | TEST(F16_PRELU__F16C_2X16, channels_gt_16) { |
| 407 | TEST_REQUIRES_X86_F16C; |
| 408 | for (size_t channels = 17; channels < 32; channels++) { |
| 409 | PReLUMicrokernelTester() |
| 410 | .rows(2) |
| 411 | .channels(channels) |
| 412 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 413 | } |
| 414 | } |
| 415 | |
| 416 | TEST(F16_PRELU__F16C_2X16, rows_lt_2) { |
| 417 | TEST_REQUIRES_X86_F16C; |
| 418 | for (size_t rows = 1; rows < 2; rows++) { |
| 419 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 420 | PReLUMicrokernelTester() |
| 421 | .rows(rows) |
| 422 | .channels(channels) |
| 423 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 424 | } |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | TEST(F16_PRELU__F16C_2X16, rows_div_2) { |
| 429 | TEST_REQUIRES_X86_F16C; |
| 430 | for (size_t rows = 4; rows <= 8; rows += 2) { |
| 431 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 432 | PReLUMicrokernelTester() |
| 433 | .rows(rows) |
| 434 | .channels(channels) |
| 435 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 436 | } |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | TEST(F16_PRELU__F16C_2X16, rows_gt_2) { |
| 441 | TEST_REQUIRES_X86_F16C; |
| 442 | for (size_t rows = 3; rows < 4; rows++) { |
| 443 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 444 | PReLUMicrokernelTester() |
| 445 | .rows(rows) |
| 446 | .channels(channels) |
| 447 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 448 | } |
| 449 | } |
| 450 | } |
| 451 | |
| 452 | TEST(F16_PRELU__F16C_2X16, input_stride) { |
| 453 | TEST_REQUIRES_X86_F16C; |
| 454 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 455 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 456 | PReLUMicrokernelTester() |
| 457 | .rows(rows) |
| 458 | .channels(channels) |
| 459 | .input_stride(83) |
| 460 | .iterations(1) |
| 461 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 462 | } |
| 463 | } |
| 464 | } |
| 465 | |
| 466 | TEST(F16_PRELU__F16C_2X16, output_stride) { |
| 467 | TEST_REQUIRES_X86_F16C; |
| 468 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 469 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 470 | PReLUMicrokernelTester() |
| 471 | .rows(rows) |
| 472 | .channels(channels) |
| 473 | .output_stride(83) |
| 474 | .iterations(1) |
| 475 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 476 | } |
| 477 | } |
| 478 | } |
| 479 | |
| 480 | TEST(F16_PRELU__F16C_2X16, inplace) { |
| 481 | TEST_REQUIRES_X86_F16C; |
| 482 | for (size_t rows = 1; rows <= 6; rows += 1) { |
| 483 | for (size_t channels = 1; channels <= 80; channels += 15) { |
| 484 | PReLUMicrokernelTester() |
| 485 | .rows(rows) |
| 486 | .channels(channels) |
| 487 | .inplace(true) |
| 488 | .iterations(1) |
| 489 | .Test(xnn_f16_prelu_ukernel__f16c_2x16); |
| 490 | } |
| 491 | } |
| 492 | } |
| 493 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |