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 | // Auto-generated file. Do not edit! |
| 7 | // Specification: test/x32-packx.yaml |
| 8 | // Generator: tools/generate-pack-test.py |
| 9 | |
| 10 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 11 | #include <gtest/gtest.h> |
| 12 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 13 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | #include <xnnpack/isa-checks.h> |
| 15 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 16 | #include <xnnpack/packx.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 17 | #include "pack-microkernel-tester.h" |
| 18 | |
| 19 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 20 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 21 | TEST(X32_PACKX_4X__NEON_ST4, k_eq_4) { |
| 22 | TEST_REQUIRES_ARM_NEON; |
| 23 | PackMicrokernelTester() |
| 24 | .mr(4) |
| 25 | .m(4) |
| 26 | .k(4) |
| 27 | .Test(xnn_x32_packx_ukernel_4x__neon_st4); |
| 28 | } |
| 29 | |
| 30 | TEST(X32_PACKX_4X__NEON_ST4, k_eq_4_subtile) { |
| 31 | TEST_REQUIRES_ARM_NEON; |
| 32 | for (size_t m = 1; m <= 4; m++) { |
| 33 | PackMicrokernelTester() |
| 34 | .mr(4) |
| 35 | .m(m) |
| 36 | .k(4) |
| 37 | .Test(xnn_x32_packx_ukernel_4x__neon_st4); |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | TEST(X32_PACKX_4X__NEON_ST4, k_lt_4) { |
| 42 | TEST_REQUIRES_ARM_NEON; |
| 43 | for (size_t k = 1; k < 4; k++) { |
| 44 | PackMicrokernelTester() |
| 45 | .mr(4) |
| 46 | .m(4) |
| 47 | .k(k) |
| 48 | .Test(xnn_x32_packx_ukernel_4x__neon_st4); |
| 49 | } |
| 50 | } |
| 51 | |
| 52 | TEST(X32_PACKX_4X__NEON_ST4, k_lt_4_subtile) { |
| 53 | TEST_REQUIRES_ARM_NEON; |
| 54 | for (size_t k = 1; k < 4; k++) { |
| 55 | for (size_t m = 1; m <= 4; m++) { |
| 56 | PackMicrokernelTester() |
| 57 | .mr(4) |
| 58 | .m(m) |
| 59 | .k(k) |
| 60 | .Test(xnn_x32_packx_ukernel_4x__neon_st4); |
| 61 | } |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | TEST(X32_PACKX_4X__NEON_ST4, k_gt_4) { |
| 66 | TEST_REQUIRES_ARM_NEON; |
| 67 | for (size_t k = 5; k < 8; k++) { |
| 68 | PackMicrokernelTester() |
| 69 | .mr(4) |
| 70 | .m(4) |
| 71 | .k(k) |
| 72 | .Test(xnn_x32_packx_ukernel_4x__neon_st4); |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | TEST(X32_PACKX_4X__NEON_ST4, k_gt_4_subtile) { |
| 77 | TEST_REQUIRES_ARM_NEON; |
| 78 | for (size_t k = 5; k < 8; k++) { |
| 79 | for (size_t m = 1; m <= 4; m++) { |
| 80 | PackMicrokernelTester() |
| 81 | .mr(4) |
| 82 | .m(m) |
| 83 | .k(k) |
| 84 | .Test(xnn_x32_packx_ukernel_4x__neon_st4); |
| 85 | } |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | TEST(X32_PACKX_4X__NEON_ST4, k_div_4) { |
| 90 | TEST_REQUIRES_ARM_NEON; |
| 91 | for (size_t k = 8; k < 40; k += 4) { |
| 92 | PackMicrokernelTester() |
| 93 | .mr(4) |
| 94 | .m(4) |
| 95 | .k(k) |
| 96 | .Test(xnn_x32_packx_ukernel_4x__neon_st4); |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | TEST(X32_PACKX_4X__NEON_ST4, k_div_4_subtile) { |
| 101 | TEST_REQUIRES_ARM_NEON; |
| 102 | for (size_t k = 8; k < 40; k += 4) { |
| 103 | for (size_t m = 1; m <= 4; m++) { |
| 104 | PackMicrokernelTester() |
| 105 | .mr(4) |
| 106 | .m(m) |
| 107 | .k(k) |
| 108 | .Test(xnn_x32_packx_ukernel_4x__neon_st4); |
| 109 | } |
| 110 | } |
| 111 | } |
| 112 | |
| 113 | TEST(X32_PACKX_4X__NEON_ST4, strided_x) { |
| 114 | TEST_REQUIRES_ARM_NEON; |
| 115 | for (size_t k = 1; k <= 20; k += 5) { |
| 116 | PackMicrokernelTester() |
| 117 | .mr(4) |
| 118 | .m(4) |
| 119 | .k(k) |
| 120 | .x_stride(23) |
| 121 | .Test(xnn_x32_packx_ukernel_4x__neon_st4); |
| 122 | } |
| 123 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 124 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 125 | |
| 126 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 127 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 128 | TEST(X32_PACKX_4X__SSE, k_eq_4) { |
| 129 | TEST_REQUIRES_X86_SSE; |
| 130 | PackMicrokernelTester() |
| 131 | .mr(4) |
| 132 | .m(4) |
| 133 | .k(4) |
| 134 | .Test(xnn_x32_packx_ukernel_4x__sse); |
| 135 | } |
| 136 | |
| 137 | TEST(X32_PACKX_4X__SSE, k_eq_4_subtile) { |
| 138 | TEST_REQUIRES_X86_SSE; |
| 139 | for (size_t m = 1; m <= 4; m++) { |
| 140 | PackMicrokernelTester() |
| 141 | .mr(4) |
| 142 | .m(m) |
| 143 | .k(4) |
| 144 | .Test(xnn_x32_packx_ukernel_4x__sse); |
| 145 | } |
| 146 | } |
| 147 | |
| 148 | TEST(X32_PACKX_4X__SSE, k_lt_4) { |
| 149 | TEST_REQUIRES_X86_SSE; |
| 150 | for (size_t k = 1; k < 4; k++) { |
| 151 | PackMicrokernelTester() |
| 152 | .mr(4) |
| 153 | .m(4) |
| 154 | .k(k) |
| 155 | .Test(xnn_x32_packx_ukernel_4x__sse); |
| 156 | } |
| 157 | } |
| 158 | |
| 159 | TEST(X32_PACKX_4X__SSE, k_lt_4_subtile) { |
| 160 | TEST_REQUIRES_X86_SSE; |
| 161 | for (size_t k = 1; k < 4; k++) { |
| 162 | for (size_t m = 1; m <= 4; m++) { |
| 163 | PackMicrokernelTester() |
| 164 | .mr(4) |
| 165 | .m(m) |
| 166 | .k(k) |
| 167 | .Test(xnn_x32_packx_ukernel_4x__sse); |
| 168 | } |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | TEST(X32_PACKX_4X__SSE, k_gt_4) { |
| 173 | TEST_REQUIRES_X86_SSE; |
| 174 | for (size_t k = 5; k < 8; k++) { |
| 175 | PackMicrokernelTester() |
| 176 | .mr(4) |
| 177 | .m(4) |
| 178 | .k(k) |
| 179 | .Test(xnn_x32_packx_ukernel_4x__sse); |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | TEST(X32_PACKX_4X__SSE, k_gt_4_subtile) { |
| 184 | TEST_REQUIRES_X86_SSE; |
| 185 | for (size_t k = 5; k < 8; k++) { |
| 186 | for (size_t m = 1; m <= 4; m++) { |
| 187 | PackMicrokernelTester() |
| 188 | .mr(4) |
| 189 | .m(m) |
| 190 | .k(k) |
| 191 | .Test(xnn_x32_packx_ukernel_4x__sse); |
| 192 | } |
| 193 | } |
| 194 | } |
| 195 | |
| 196 | TEST(X32_PACKX_4X__SSE, k_div_4) { |
| 197 | TEST_REQUIRES_X86_SSE; |
| 198 | for (size_t k = 8; k < 40; k += 4) { |
| 199 | PackMicrokernelTester() |
| 200 | .mr(4) |
| 201 | .m(4) |
| 202 | .k(k) |
| 203 | .Test(xnn_x32_packx_ukernel_4x__sse); |
| 204 | } |
| 205 | } |
| 206 | |
| 207 | TEST(X32_PACKX_4X__SSE, k_div_4_subtile) { |
| 208 | TEST_REQUIRES_X86_SSE; |
| 209 | for (size_t k = 8; k < 40; k += 4) { |
| 210 | for (size_t m = 1; m <= 4; m++) { |
| 211 | PackMicrokernelTester() |
| 212 | .mr(4) |
| 213 | .m(m) |
| 214 | .k(k) |
| 215 | .Test(xnn_x32_packx_ukernel_4x__sse); |
| 216 | } |
| 217 | } |
| 218 | } |
| 219 | |
| 220 | TEST(X32_PACKX_4X__SSE, strided_x) { |
| 221 | TEST_REQUIRES_X86_SSE; |
| 222 | for (size_t k = 1; k <= 20; k += 5) { |
| 223 | PackMicrokernelTester() |
| 224 | .mr(4) |
| 225 | .m(4) |
| 226 | .k(k) |
| 227 | .x_stride(23) |
| 228 | .Test(xnn_x32_packx_ukernel_4x__sse); |
| 229 | } |
| 230 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 231 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 232 | |
| 233 | |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 234 | #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 66d99e9 | 2020-07-16 12:56:21 -0700 | [diff] [blame] | 235 | TEST(X32_PACKX_4X__WASMSIMD, k_eq_4) { |
| 236 | PackMicrokernelTester() |
| 237 | .mr(4) |
| 238 | .m(4) |
| 239 | .k(4) |
| 240 | .Test(xnn_x32_packx_ukernel_4x__wasmsimd); |
| 241 | } |
| 242 | |
| 243 | TEST(X32_PACKX_4X__WASMSIMD, k_eq_4_subtile) { |
| 244 | for (size_t m = 1; m <= 4; m++) { |
| 245 | PackMicrokernelTester() |
| 246 | .mr(4) |
| 247 | .m(m) |
| 248 | .k(4) |
| 249 | .Test(xnn_x32_packx_ukernel_4x__wasmsimd); |
| 250 | } |
| 251 | } |
| 252 | |
| 253 | TEST(X32_PACKX_4X__WASMSIMD, k_lt_4) { |
| 254 | for (size_t k = 1; k < 4; k++) { |
| 255 | PackMicrokernelTester() |
| 256 | .mr(4) |
| 257 | .m(4) |
| 258 | .k(k) |
| 259 | .Test(xnn_x32_packx_ukernel_4x__wasmsimd); |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | TEST(X32_PACKX_4X__WASMSIMD, k_lt_4_subtile) { |
| 264 | for (size_t k = 1; k < 4; k++) { |
| 265 | for (size_t m = 1; m <= 4; m++) { |
| 266 | PackMicrokernelTester() |
| 267 | .mr(4) |
| 268 | .m(m) |
| 269 | .k(k) |
| 270 | .Test(xnn_x32_packx_ukernel_4x__wasmsimd); |
| 271 | } |
| 272 | } |
| 273 | } |
| 274 | |
| 275 | TEST(X32_PACKX_4X__WASMSIMD, k_gt_4) { |
| 276 | for (size_t k = 5; k < 8; k++) { |
| 277 | PackMicrokernelTester() |
| 278 | .mr(4) |
| 279 | .m(4) |
| 280 | .k(k) |
| 281 | .Test(xnn_x32_packx_ukernel_4x__wasmsimd); |
| 282 | } |
| 283 | } |
| 284 | |
| 285 | TEST(X32_PACKX_4X__WASMSIMD, k_gt_4_subtile) { |
| 286 | for (size_t k = 5; k < 8; k++) { |
| 287 | for (size_t m = 1; m <= 4; m++) { |
| 288 | PackMicrokernelTester() |
| 289 | .mr(4) |
| 290 | .m(m) |
| 291 | .k(k) |
| 292 | .Test(xnn_x32_packx_ukernel_4x__wasmsimd); |
| 293 | } |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | TEST(X32_PACKX_4X__WASMSIMD, k_div_4) { |
| 298 | for (size_t k = 8; k < 40; k += 4) { |
| 299 | PackMicrokernelTester() |
| 300 | .mr(4) |
| 301 | .m(4) |
| 302 | .k(k) |
| 303 | .Test(xnn_x32_packx_ukernel_4x__wasmsimd); |
| 304 | } |
| 305 | } |
| 306 | |
| 307 | TEST(X32_PACKX_4X__WASMSIMD, k_div_4_subtile) { |
| 308 | for (size_t k = 8; k < 40; k += 4) { |
| 309 | for (size_t m = 1; m <= 4; m++) { |
| 310 | PackMicrokernelTester() |
| 311 | .mr(4) |
| 312 | .m(m) |
| 313 | .k(k) |
| 314 | .Test(xnn_x32_packx_ukernel_4x__wasmsimd); |
| 315 | } |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | TEST(X32_PACKX_4X__WASMSIMD, strided_x) { |
| 320 | for (size_t k = 1; k <= 20; k += 5) { |
| 321 | PackMicrokernelTester() |
| 322 | .mr(4) |
| 323 | .m(4) |
| 324 | .k(k) |
| 325 | .x_stride(23) |
| 326 | .Test(xnn_x32_packx_ukernel_4x__wasmsimd); |
| 327 | } |
| 328 | } |
Marat Dukhan | 4c61779 | 2021-12-21 15:47:58 -0800 | [diff] [blame] | 329 | #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
Marat Dukhan | 66d99e9 | 2020-07-16 12:56:21 -0700 | [diff] [blame] | 330 | |
| 331 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 332 | TEST(X32_PACKX_2X__SCALAR, k_eq_1) { |
| 333 | PackMicrokernelTester() |
| 334 | .mr(2) |
| 335 | .m(2) |
| 336 | .k(1) |
| 337 | .Test(xnn_x32_packx_ukernel_2x__scalar); |
| 338 | } |
| 339 | |
| 340 | TEST(X32_PACKX_2X__SCALAR, k_eq_1_subtile) { |
| 341 | for (size_t m = 1; m <= 2; m++) { |
| 342 | PackMicrokernelTester() |
| 343 | .mr(2) |
| 344 | .m(m) |
| 345 | .k(1) |
| 346 | .Test(xnn_x32_packx_ukernel_2x__scalar); |
| 347 | } |
| 348 | } |
| 349 | |
| 350 | TEST(X32_PACKX_2X__SCALAR, k_gt_1) { |
| 351 | for (size_t k = 2; k < 10; k++) { |
| 352 | PackMicrokernelTester() |
| 353 | .mr(2) |
| 354 | .m(2) |
| 355 | .k(k) |
| 356 | .Test(xnn_x32_packx_ukernel_2x__scalar); |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | TEST(X32_PACKX_2X__SCALAR, k_gt_1_subtile) { |
| 361 | for (size_t k = 2; k < 10; k++) { |
| 362 | for (size_t m = 1; m <= 2; m++) { |
| 363 | PackMicrokernelTester() |
| 364 | .mr(2) |
| 365 | .m(m) |
| 366 | .k(k) |
| 367 | .Test(xnn_x32_packx_ukernel_2x__scalar); |
| 368 | } |
| 369 | } |
| 370 | } |
| 371 | |
| 372 | TEST(X32_PACKX_2X__SCALAR, strided_x) { |
| 373 | for (size_t k = 1; k <= 5; k += 2) { |
| 374 | PackMicrokernelTester() |
| 375 | .mr(2) |
| 376 | .m(2) |
| 377 | .k(k) |
| 378 | .x_stride(7) |
| 379 | .Test(xnn_x32_packx_ukernel_2x__scalar); |
| 380 | } |
| 381 | } |
| 382 | |
| 383 | TEST(X32_PACKX_3X__SCALAR, k_eq_1) { |
| 384 | PackMicrokernelTester() |
| 385 | .mr(3) |
| 386 | .m(3) |
| 387 | .k(1) |
| 388 | .Test(xnn_x32_packx_ukernel_3x__scalar); |
| 389 | } |
| 390 | |
| 391 | TEST(X32_PACKX_3X__SCALAR, k_eq_1_subtile) { |
| 392 | for (size_t m = 1; m <= 3; m++) { |
| 393 | PackMicrokernelTester() |
| 394 | .mr(3) |
| 395 | .m(m) |
| 396 | .k(1) |
| 397 | .Test(xnn_x32_packx_ukernel_3x__scalar); |
| 398 | } |
| 399 | } |
| 400 | |
| 401 | TEST(X32_PACKX_3X__SCALAR, k_gt_1) { |
| 402 | for (size_t k = 2; k < 10; k++) { |
| 403 | PackMicrokernelTester() |
| 404 | .mr(3) |
| 405 | .m(3) |
| 406 | .k(k) |
| 407 | .Test(xnn_x32_packx_ukernel_3x__scalar); |
| 408 | } |
| 409 | } |
| 410 | |
| 411 | TEST(X32_PACKX_3X__SCALAR, k_gt_1_subtile) { |
| 412 | for (size_t k = 2; k < 10; k++) { |
| 413 | for (size_t m = 1; m <= 3; m++) { |
| 414 | PackMicrokernelTester() |
| 415 | .mr(3) |
| 416 | .m(m) |
| 417 | .k(k) |
| 418 | .Test(xnn_x32_packx_ukernel_3x__scalar); |
| 419 | } |
| 420 | } |
| 421 | } |
| 422 | |
| 423 | TEST(X32_PACKX_3X__SCALAR, strided_x) { |
| 424 | for (size_t k = 1; k <= 5; k += 2) { |
| 425 | PackMicrokernelTester() |
| 426 | .mr(3) |
| 427 | .m(3) |
| 428 | .k(k) |
| 429 | .x_stride(7) |
| 430 | .Test(xnn_x32_packx_ukernel_3x__scalar); |
| 431 | } |
| 432 | } |
| 433 | |
| 434 | TEST(X32_PACKX_4X__SCALAR, k_eq_1) { |
| 435 | PackMicrokernelTester() |
| 436 | .mr(4) |
| 437 | .m(4) |
| 438 | .k(1) |
| 439 | .Test(xnn_x32_packx_ukernel_4x__scalar); |
| 440 | } |
| 441 | |
| 442 | TEST(X32_PACKX_4X__SCALAR, k_eq_1_subtile) { |
| 443 | for (size_t m = 1; m <= 4; m++) { |
| 444 | PackMicrokernelTester() |
| 445 | .mr(4) |
| 446 | .m(m) |
| 447 | .k(1) |
| 448 | .Test(xnn_x32_packx_ukernel_4x__scalar); |
| 449 | } |
| 450 | } |
| 451 | |
| 452 | TEST(X32_PACKX_4X__SCALAR, k_gt_1) { |
| 453 | for (size_t k = 2; k < 10; k++) { |
| 454 | PackMicrokernelTester() |
| 455 | .mr(4) |
| 456 | .m(4) |
| 457 | .k(k) |
| 458 | .Test(xnn_x32_packx_ukernel_4x__scalar); |
| 459 | } |
| 460 | } |
| 461 | |
| 462 | TEST(X32_PACKX_4X__SCALAR, k_gt_1_subtile) { |
| 463 | for (size_t k = 2; k < 10; k++) { |
| 464 | for (size_t m = 1; m <= 4; m++) { |
| 465 | PackMicrokernelTester() |
| 466 | .mr(4) |
| 467 | .m(m) |
| 468 | .k(k) |
| 469 | .Test(xnn_x32_packx_ukernel_4x__scalar); |
| 470 | } |
| 471 | } |
| 472 | } |
| 473 | |
| 474 | TEST(X32_PACKX_4X__SCALAR, strided_x) { |
| 475 | for (size_t k = 1; k <= 5; k += 2) { |
| 476 | PackMicrokernelTester() |
| 477 | .mr(4) |
| 478 | .m(4) |
| 479 | .k(k) |
| 480 | .x_stride(7) |
| 481 | .Test(xnn_x32_packx_ukernel_4x__scalar); |
| 482 | } |
| 483 | } |