| 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 | // Auto-generated file. Do not edit! |
| 10 | // Specification: test/f32-ppmm.yaml |
| 11 | // Generator: tools/generate-gemm-test.py |
| 12 | |
| 13 | |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | #include <gtest/gtest.h> |
| 15 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 16 | #include <xnnpack/common.h> |
| 17 | #include <xnnpack/isa-checks.h> |
| 18 | |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 19 | #include <xnnpack/gemm.h> |
| 20 | #include <xnnpack/igemm.h> |
| 21 | #include <xnnpack/ppmm.h> |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 22 | #include "gemm-microkernel-tester.h" |
| 23 | |
| 24 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 25 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | TEST(F32_PPMM_4X8__NEON, k_eq_1) { |
| 27 | TEST_REQUIRES_ARM_NEON; |
| 28 | GemmMicrokernelTester() |
| 29 | .mr(4) |
| 30 | .nr(8) |
| 31 | .kr(1) |
| 32 | .sr(1) |
| 33 | .m(4) |
| 34 | .n(8) |
| 35 | .k(1) |
| 36 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 37 | } |
| 38 | |
| 39 | TEST(F32_PPMM_4X8__NEON, strided_cn) { |
| 40 | TEST_REQUIRES_ARM_NEON; |
| 41 | GemmMicrokernelTester() |
| 42 | .mr(4) |
| 43 | .nr(8) |
| 44 | .kr(1) |
| 45 | .sr(1) |
| 46 | .m(4) |
| 47 | .n(8) |
| 48 | .k(1) |
| 49 | .cn_stride(11) |
| 50 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 51 | } |
| 52 | |
| 53 | TEST(F32_PPMM_4X8__NEON, k_eq_1_strided_a) { |
| 54 | TEST_REQUIRES_ARM_NEON; |
| 55 | GemmMicrokernelTester() |
| 56 | .mr(4) |
| 57 | .nr(8) |
| 58 | .kr(1) |
| 59 | .sr(1) |
| 60 | .m(4) |
| 61 | .n(8) |
| 62 | .k(1) |
| 63 | .a_stride(3) |
| 64 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 65 | } |
| 66 | |
| 67 | TEST(F32_PPMM_4X8__NEON, k_eq_1_subtile) { |
| 68 | TEST_REQUIRES_ARM_NEON; |
| 69 | for (uint32_t m = 1; m <= 4; m++) { |
| 70 | for (uint32_t n = 1; n <= 8; n++) { |
| 71 | GemmMicrokernelTester() |
| 72 | .mr(4) |
| 73 | .nr(8) |
| 74 | .kr(1) |
| 75 | .sr(1) |
| 76 | .m(m) |
| 77 | .n(n) |
| 78 | .k(1) |
| 79 | .iterations(1) |
| 80 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 81 | } |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | TEST(F32_PPMM_4X8__NEON, k_eq_1_subtile_m) { |
| 86 | TEST_REQUIRES_ARM_NEON; |
| 87 | for (uint32_t m = 1; m <= 4; m++) { |
| 88 | GemmMicrokernelTester() |
| 89 | .mr(4) |
| 90 | .nr(8) |
| 91 | .kr(1) |
| 92 | .sr(1) |
| 93 | .m(m) |
| 94 | .n(8) |
| 95 | .k(1) |
| 96 | .iterations(1) |
| 97 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | TEST(F32_PPMM_4X8__NEON, k_eq_1_subtile_n) { |
| 102 | TEST_REQUIRES_ARM_NEON; |
| 103 | for (uint32_t n = 1; n <= 8; n++) { |
| 104 | GemmMicrokernelTester() |
| 105 | .mr(4) |
| 106 | .nr(8) |
| 107 | .kr(1) |
| 108 | .sr(1) |
| 109 | .m(4) |
| 110 | .n(n) |
| 111 | .k(1) |
| 112 | .iterations(1) |
| 113 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | TEST(F32_PPMM_4X8__NEON, k_gt_1) { |
| 118 | TEST_REQUIRES_ARM_NEON; |
| 119 | for (size_t k = 2; k < 10; k++) { |
| 120 | GemmMicrokernelTester() |
| 121 | .mr(4) |
| 122 | .nr(8) |
| 123 | .kr(1) |
| 124 | .sr(1) |
| 125 | .m(4) |
| 126 | .n(8) |
| 127 | .k(k) |
| 128 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | TEST(F32_PPMM_4X8__NEON, k_gt_1_subtile) { |
| 133 | TEST_REQUIRES_ARM_NEON; |
| 134 | for (size_t k = 2; k < 10; k++) { |
| 135 | for (uint32_t m = 1; m <= 4; m++) { |
| 136 | for (uint32_t n = 1; n <= 8; n++) { |
| 137 | GemmMicrokernelTester() |
| 138 | .mr(4) |
| 139 | .nr(8) |
| 140 | .kr(1) |
| 141 | .sr(1) |
| 142 | .m(m) |
| 143 | .n(n) |
| 144 | .k(k) |
| 145 | .iterations(1) |
| 146 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 147 | } |
| 148 | } |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | TEST(F32_PPMM_4X8__NEON, n_gt_8) { |
| 153 | TEST_REQUIRES_ARM_NEON; |
| 154 | for (uint32_t n = 9; n < 16; n++) { |
| 155 | for (size_t k = 1; k <= 5; k += 2) { |
| 156 | GemmMicrokernelTester() |
| 157 | .mr(4) |
| 158 | .nr(8) |
| 159 | .kr(1) |
| 160 | .sr(1) |
| 161 | .m(4) |
| 162 | .n(8) |
| 163 | .k(k) |
| 164 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 165 | } |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | TEST(F32_PPMM_4X8__NEON, n_gt_8_strided_cn) { |
| 170 | TEST_REQUIRES_ARM_NEON; |
| 171 | for (uint32_t n = 9; n < 16; n++) { |
| 172 | for (size_t k = 1; k <= 5; k += 2) { |
| 173 | GemmMicrokernelTester() |
| 174 | .mr(4) |
| 175 | .nr(8) |
| 176 | .kr(1) |
| 177 | .sr(1) |
| 178 | .m(4) |
| 179 | .n(8) |
| 180 | .k(k) |
| 181 | .cn_stride(11) |
| 182 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 183 | } |
| 184 | } |
| 185 | } |
| 186 | |
| 187 | TEST(F32_PPMM_4X8__NEON, n_gt_8_strided_a) { |
| 188 | TEST_REQUIRES_ARM_NEON; |
| 189 | for (uint32_t n = 9; n < 16; n++) { |
| 190 | for (size_t k = 1; k <= 5; k += 2) { |
| 191 | GemmMicrokernelTester() |
| 192 | .mr(4) |
| 193 | .nr(8) |
| 194 | .kr(1) |
| 195 | .sr(1) |
| 196 | .m(4) |
| 197 | .n(n) |
| 198 | .k(k) |
| 199 | .a_stride(7) |
| 200 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 201 | } |
| 202 | } |
| 203 | } |
| 204 | |
| 205 | TEST(F32_PPMM_4X8__NEON, n_gt_8_subtile) { |
| 206 | TEST_REQUIRES_ARM_NEON; |
| 207 | for (uint32_t n = 9; n < 16; n++) { |
| 208 | for (size_t k = 1; k <= 5; k += 2) { |
| 209 | for (uint32_t m = 1; m <= 4; m++) { |
| 210 | GemmMicrokernelTester() |
| 211 | .mr(4) |
| 212 | .nr(8) |
| 213 | .kr(1) |
| 214 | .sr(1) |
| 215 | .m(m) |
| 216 | .n(n) |
| 217 | .k(k) |
| 218 | .iterations(1) |
| 219 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 220 | } |
| 221 | } |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | TEST(F32_PPMM_4X8__NEON, n_div_8) { |
| 226 | TEST_REQUIRES_ARM_NEON; |
| 227 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 228 | for (size_t k = 1; k <= 5; k += 2) { |
| 229 | GemmMicrokernelTester() |
| 230 | .mr(4) |
| 231 | .nr(8) |
| 232 | .kr(1) |
| 233 | .sr(1) |
| 234 | .m(4) |
| 235 | .n(8) |
| 236 | .k(k) |
| 237 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 238 | } |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | TEST(F32_PPMM_4X8__NEON, n_div_8_strided_cn) { |
| 243 | TEST_REQUIRES_ARM_NEON; |
| 244 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 245 | for (size_t k = 1; k <= 5; k += 2) { |
| 246 | GemmMicrokernelTester() |
| 247 | .mr(4) |
| 248 | .nr(8) |
| 249 | .kr(1) |
| 250 | .sr(1) |
| 251 | .m(4) |
| 252 | .n(n) |
| 253 | .k(k) |
| 254 | .cn_stride(11) |
| 255 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 256 | } |
| 257 | } |
| 258 | } |
| 259 | |
| 260 | TEST(F32_PPMM_4X8__NEON, n_div_8_strided_a) { |
| 261 | TEST_REQUIRES_ARM_NEON; |
| 262 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 263 | for (size_t k = 1; k <= 5; k += 2) { |
| 264 | GemmMicrokernelTester() |
| 265 | .mr(4) |
| 266 | .nr(8) |
| 267 | .kr(1) |
| 268 | .sr(1) |
| 269 | .m(4) |
| 270 | .n(n) |
| 271 | .k(k) |
| 272 | .a_stride(7) |
| 273 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 274 | } |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | TEST(F32_PPMM_4X8__NEON, n_div_8_subtile) { |
| 279 | TEST_REQUIRES_ARM_NEON; |
| 280 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 281 | for (size_t k = 1; k <= 5; k += 2) { |
| 282 | for (uint32_t m = 1; m <= 4; m++) { |
| 283 | GemmMicrokernelTester() |
| 284 | .mr(4) |
| 285 | .nr(8) |
| 286 | .kr(1) |
| 287 | .sr(1) |
| 288 | .m(m) |
| 289 | .n(n) |
| 290 | .k(k) |
| 291 | .iterations(1) |
| 292 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 293 | } |
| 294 | } |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | TEST(F32_PPMM_4X8__NEON, strided_cm_subtile) { |
| 299 | TEST_REQUIRES_ARM_NEON; |
| 300 | for (size_t k = 1; k <= 5; k += 2) { |
| 301 | for (uint32_t m = 1; m <= 4; m++) { |
| 302 | for (uint32_t n = 1; n <= 8; n++) { |
| 303 | GemmMicrokernelTester() |
| 304 | .mr(4) |
| 305 | .nr(8) |
| 306 | .kr(1) |
| 307 | .sr(1) |
| 308 | .m(m) |
| 309 | .n(n) |
| 310 | .k(k) |
| 311 | .cm_stride(11) |
| 312 | .iterations(1) |
| 313 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 314 | } |
| 315 | } |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | TEST(F32_PPMM_4X8__NEON, qmin) { |
| 320 | TEST_REQUIRES_ARM_NEON; |
| 321 | GemmMicrokernelTester() |
| 322 | .mr(4) |
| 323 | .nr(8) |
| 324 | .kr(1) |
| 325 | .sr(1) |
| 326 | .m(4) |
| 327 | .n(8) |
| 328 | .k(1) |
| 329 | .qmin(128) |
| 330 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 331 | } |
| 332 | |
| 333 | TEST(F32_PPMM_4X8__NEON, qmax) { |
| 334 | TEST_REQUIRES_ARM_NEON; |
| 335 | GemmMicrokernelTester() |
| 336 | .mr(4) |
| 337 | .nr(8) |
| 338 | .kr(1) |
| 339 | .sr(1) |
| 340 | .m(4) |
| 341 | .n(8) |
| 342 | .k(1) |
| 343 | .qmax(128) |
| 344 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 345 | } |
| 346 | |
| 347 | TEST(F32_PPMM_4X8__NEON, strided_cm) { |
| 348 | TEST_REQUIRES_ARM_NEON; |
| 349 | GemmMicrokernelTester() |
| 350 | .mr(4) |
| 351 | .nr(8) |
| 352 | .kr(1) |
| 353 | .sr(1) |
| 354 | .m(4) |
| 355 | .n(8) |
| 356 | .k(1) |
| 357 | .cm_stride(11) |
| 358 | .Test(xnn_f32_ppmm_ukernel_4x8__neon); |
| 359 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 360 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 361 | |
| 362 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 363 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 364 | TEST(F32_PPMM_4X8__NEONFMA, k_eq_1) { |
| 365 | TEST_REQUIRES_ARM_NEON_FMA; |
| 366 | GemmMicrokernelTester() |
| 367 | .mr(4) |
| 368 | .nr(8) |
| 369 | .kr(1) |
| 370 | .sr(1) |
| 371 | .m(4) |
| 372 | .n(8) |
| 373 | .k(1) |
| 374 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 375 | } |
| 376 | |
| 377 | TEST(F32_PPMM_4X8__NEONFMA, strided_cn) { |
| 378 | TEST_REQUIRES_ARM_NEON_FMA; |
| 379 | GemmMicrokernelTester() |
| 380 | .mr(4) |
| 381 | .nr(8) |
| 382 | .kr(1) |
| 383 | .sr(1) |
| 384 | .m(4) |
| 385 | .n(8) |
| 386 | .k(1) |
| 387 | .cn_stride(11) |
| 388 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 389 | } |
| 390 | |
| 391 | TEST(F32_PPMM_4X8__NEONFMA, k_eq_1_strided_a) { |
| 392 | TEST_REQUIRES_ARM_NEON_FMA; |
| 393 | GemmMicrokernelTester() |
| 394 | .mr(4) |
| 395 | .nr(8) |
| 396 | .kr(1) |
| 397 | .sr(1) |
| 398 | .m(4) |
| 399 | .n(8) |
| 400 | .k(1) |
| 401 | .a_stride(3) |
| 402 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 403 | } |
| 404 | |
| 405 | TEST(F32_PPMM_4X8__NEONFMA, k_eq_1_subtile) { |
| 406 | TEST_REQUIRES_ARM_NEON_FMA; |
| 407 | for (uint32_t m = 1; m <= 4; m++) { |
| 408 | for (uint32_t n = 1; n <= 8; n++) { |
| 409 | GemmMicrokernelTester() |
| 410 | .mr(4) |
| 411 | .nr(8) |
| 412 | .kr(1) |
| 413 | .sr(1) |
| 414 | .m(m) |
| 415 | .n(n) |
| 416 | .k(1) |
| 417 | .iterations(1) |
| 418 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 419 | } |
| 420 | } |
| 421 | } |
| 422 | |
| 423 | TEST(F32_PPMM_4X8__NEONFMA, k_eq_1_subtile_m) { |
| 424 | TEST_REQUIRES_ARM_NEON_FMA; |
| 425 | for (uint32_t m = 1; m <= 4; m++) { |
| 426 | GemmMicrokernelTester() |
| 427 | .mr(4) |
| 428 | .nr(8) |
| 429 | .kr(1) |
| 430 | .sr(1) |
| 431 | .m(m) |
| 432 | .n(8) |
| 433 | .k(1) |
| 434 | .iterations(1) |
| 435 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 436 | } |
| 437 | } |
| 438 | |
| 439 | TEST(F32_PPMM_4X8__NEONFMA, k_eq_1_subtile_n) { |
| 440 | TEST_REQUIRES_ARM_NEON_FMA; |
| 441 | for (uint32_t n = 1; n <= 8; n++) { |
| 442 | GemmMicrokernelTester() |
| 443 | .mr(4) |
| 444 | .nr(8) |
| 445 | .kr(1) |
| 446 | .sr(1) |
| 447 | .m(4) |
| 448 | .n(n) |
| 449 | .k(1) |
| 450 | .iterations(1) |
| 451 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 452 | } |
| 453 | } |
| 454 | |
| 455 | TEST(F32_PPMM_4X8__NEONFMA, k_gt_1) { |
| 456 | TEST_REQUIRES_ARM_NEON_FMA; |
| 457 | for (size_t k = 2; k < 10; k++) { |
| 458 | GemmMicrokernelTester() |
| 459 | .mr(4) |
| 460 | .nr(8) |
| 461 | .kr(1) |
| 462 | .sr(1) |
| 463 | .m(4) |
| 464 | .n(8) |
| 465 | .k(k) |
| 466 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 467 | } |
| 468 | } |
| 469 | |
| 470 | TEST(F32_PPMM_4X8__NEONFMA, k_gt_1_subtile) { |
| 471 | TEST_REQUIRES_ARM_NEON_FMA; |
| 472 | for (size_t k = 2; k < 10; k++) { |
| 473 | for (uint32_t m = 1; m <= 4; m++) { |
| 474 | for (uint32_t n = 1; n <= 8; n++) { |
| 475 | GemmMicrokernelTester() |
| 476 | .mr(4) |
| 477 | .nr(8) |
| 478 | .kr(1) |
| 479 | .sr(1) |
| 480 | .m(m) |
| 481 | .n(n) |
| 482 | .k(k) |
| 483 | .iterations(1) |
| 484 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 485 | } |
| 486 | } |
| 487 | } |
| 488 | } |
| 489 | |
| 490 | TEST(F32_PPMM_4X8__NEONFMA, n_gt_8) { |
| 491 | TEST_REQUIRES_ARM_NEON_FMA; |
| 492 | for (uint32_t n = 9; n < 16; n++) { |
| 493 | for (size_t k = 1; k <= 5; k += 2) { |
| 494 | GemmMicrokernelTester() |
| 495 | .mr(4) |
| 496 | .nr(8) |
| 497 | .kr(1) |
| 498 | .sr(1) |
| 499 | .m(4) |
| 500 | .n(8) |
| 501 | .k(k) |
| 502 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 503 | } |
| 504 | } |
| 505 | } |
| 506 | |
| 507 | TEST(F32_PPMM_4X8__NEONFMA, n_gt_8_strided_cn) { |
| 508 | TEST_REQUIRES_ARM_NEON_FMA; |
| 509 | for (uint32_t n = 9; n < 16; n++) { |
| 510 | for (size_t k = 1; k <= 5; k += 2) { |
| 511 | GemmMicrokernelTester() |
| 512 | .mr(4) |
| 513 | .nr(8) |
| 514 | .kr(1) |
| 515 | .sr(1) |
| 516 | .m(4) |
| 517 | .n(8) |
| 518 | .k(k) |
| 519 | .cn_stride(11) |
| 520 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 521 | } |
| 522 | } |
| 523 | } |
| 524 | |
| 525 | TEST(F32_PPMM_4X8__NEONFMA, n_gt_8_strided_a) { |
| 526 | TEST_REQUIRES_ARM_NEON_FMA; |
| 527 | for (uint32_t n = 9; n < 16; n++) { |
| 528 | for (size_t k = 1; k <= 5; k += 2) { |
| 529 | GemmMicrokernelTester() |
| 530 | .mr(4) |
| 531 | .nr(8) |
| 532 | .kr(1) |
| 533 | .sr(1) |
| 534 | .m(4) |
| 535 | .n(n) |
| 536 | .k(k) |
| 537 | .a_stride(7) |
| 538 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 539 | } |
| 540 | } |
| 541 | } |
| 542 | |
| 543 | TEST(F32_PPMM_4X8__NEONFMA, n_gt_8_subtile) { |
| 544 | TEST_REQUIRES_ARM_NEON_FMA; |
| 545 | for (uint32_t n = 9; n < 16; n++) { |
| 546 | for (size_t k = 1; k <= 5; k += 2) { |
| 547 | for (uint32_t m = 1; m <= 4; m++) { |
| 548 | GemmMicrokernelTester() |
| 549 | .mr(4) |
| 550 | .nr(8) |
| 551 | .kr(1) |
| 552 | .sr(1) |
| 553 | .m(m) |
| 554 | .n(n) |
| 555 | .k(k) |
| 556 | .iterations(1) |
| 557 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 558 | } |
| 559 | } |
| 560 | } |
| 561 | } |
| 562 | |
| 563 | TEST(F32_PPMM_4X8__NEONFMA, n_div_8) { |
| 564 | TEST_REQUIRES_ARM_NEON_FMA; |
| 565 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 566 | for (size_t k = 1; k <= 5; k += 2) { |
| 567 | GemmMicrokernelTester() |
| 568 | .mr(4) |
| 569 | .nr(8) |
| 570 | .kr(1) |
| 571 | .sr(1) |
| 572 | .m(4) |
| 573 | .n(8) |
| 574 | .k(k) |
| 575 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 576 | } |
| 577 | } |
| 578 | } |
| 579 | |
| 580 | TEST(F32_PPMM_4X8__NEONFMA, n_div_8_strided_cn) { |
| 581 | TEST_REQUIRES_ARM_NEON_FMA; |
| 582 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 583 | for (size_t k = 1; k <= 5; k += 2) { |
| 584 | GemmMicrokernelTester() |
| 585 | .mr(4) |
| 586 | .nr(8) |
| 587 | .kr(1) |
| 588 | .sr(1) |
| 589 | .m(4) |
| 590 | .n(n) |
| 591 | .k(k) |
| 592 | .cn_stride(11) |
| 593 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 594 | } |
| 595 | } |
| 596 | } |
| 597 | |
| 598 | TEST(F32_PPMM_4X8__NEONFMA, n_div_8_strided_a) { |
| 599 | TEST_REQUIRES_ARM_NEON_FMA; |
| 600 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 601 | for (size_t k = 1; k <= 5; k += 2) { |
| 602 | GemmMicrokernelTester() |
| 603 | .mr(4) |
| 604 | .nr(8) |
| 605 | .kr(1) |
| 606 | .sr(1) |
| 607 | .m(4) |
| 608 | .n(n) |
| 609 | .k(k) |
| 610 | .a_stride(7) |
| 611 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 612 | } |
| 613 | } |
| 614 | } |
| 615 | |
| 616 | TEST(F32_PPMM_4X8__NEONFMA, n_div_8_subtile) { |
| 617 | TEST_REQUIRES_ARM_NEON_FMA; |
| 618 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 619 | for (size_t k = 1; k <= 5; k += 2) { |
| 620 | for (uint32_t m = 1; m <= 4; m++) { |
| 621 | GemmMicrokernelTester() |
| 622 | .mr(4) |
| 623 | .nr(8) |
| 624 | .kr(1) |
| 625 | .sr(1) |
| 626 | .m(m) |
| 627 | .n(n) |
| 628 | .k(k) |
| 629 | .iterations(1) |
| 630 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 631 | } |
| 632 | } |
| 633 | } |
| 634 | } |
| 635 | |
| 636 | TEST(F32_PPMM_4X8__NEONFMA, strided_cm_subtile) { |
| 637 | TEST_REQUIRES_ARM_NEON_FMA; |
| 638 | for (size_t k = 1; k <= 5; k += 2) { |
| 639 | for (uint32_t m = 1; m <= 4; m++) { |
| 640 | for (uint32_t n = 1; n <= 8; n++) { |
| 641 | GemmMicrokernelTester() |
| 642 | .mr(4) |
| 643 | .nr(8) |
| 644 | .kr(1) |
| 645 | .sr(1) |
| 646 | .m(m) |
| 647 | .n(n) |
| 648 | .k(k) |
| 649 | .cm_stride(11) |
| 650 | .iterations(1) |
| 651 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 652 | } |
| 653 | } |
| 654 | } |
| 655 | } |
| 656 | |
| 657 | TEST(F32_PPMM_4X8__NEONFMA, qmin) { |
| 658 | TEST_REQUIRES_ARM_NEON_FMA; |
| 659 | GemmMicrokernelTester() |
| 660 | .mr(4) |
| 661 | .nr(8) |
| 662 | .kr(1) |
| 663 | .sr(1) |
| 664 | .m(4) |
| 665 | .n(8) |
| 666 | .k(1) |
| 667 | .qmin(128) |
| 668 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 669 | } |
| 670 | |
| 671 | TEST(F32_PPMM_4X8__NEONFMA, qmax) { |
| 672 | TEST_REQUIRES_ARM_NEON_FMA; |
| 673 | GemmMicrokernelTester() |
| 674 | .mr(4) |
| 675 | .nr(8) |
| 676 | .kr(1) |
| 677 | .sr(1) |
| 678 | .m(4) |
| 679 | .n(8) |
| 680 | .k(1) |
| 681 | .qmax(128) |
| 682 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 683 | } |
| 684 | |
| 685 | TEST(F32_PPMM_4X8__NEONFMA, strided_cm) { |
| 686 | TEST_REQUIRES_ARM_NEON_FMA; |
| 687 | GemmMicrokernelTester() |
| 688 | .mr(4) |
| 689 | .nr(8) |
| 690 | .kr(1) |
| 691 | .sr(1) |
| 692 | .m(4) |
| 693 | .n(8) |
| 694 | .k(1) |
| 695 | .cm_stride(11) |
| 696 | .Test(xnn_f32_ppmm_ukernel_4x8__neonfma); |
| 697 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 698 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 699 | |
| 700 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 701 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 702 | TEST(F32_PPMM_8X8__NEON, k_eq_1) { |
| 703 | TEST_REQUIRES_ARM_NEON; |
| 704 | GemmMicrokernelTester() |
| 705 | .mr(8) |
| 706 | .nr(8) |
| 707 | .kr(1) |
| 708 | .sr(1) |
| 709 | .m(8) |
| 710 | .n(8) |
| 711 | .k(1) |
| 712 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 713 | } |
| 714 | |
| 715 | TEST(F32_PPMM_8X8__NEON, strided_cn) { |
| 716 | TEST_REQUIRES_ARM_NEON; |
| 717 | GemmMicrokernelTester() |
| 718 | .mr(8) |
| 719 | .nr(8) |
| 720 | .kr(1) |
| 721 | .sr(1) |
| 722 | .m(8) |
| 723 | .n(8) |
| 724 | .k(1) |
| 725 | .cn_stride(11) |
| 726 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 727 | } |
| 728 | |
| 729 | TEST(F32_PPMM_8X8__NEON, k_eq_1_strided_a) { |
| 730 | TEST_REQUIRES_ARM_NEON; |
| 731 | GemmMicrokernelTester() |
| 732 | .mr(8) |
| 733 | .nr(8) |
| 734 | .kr(1) |
| 735 | .sr(1) |
| 736 | .m(8) |
| 737 | .n(8) |
| 738 | .k(1) |
| 739 | .a_stride(3) |
| 740 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 741 | } |
| 742 | |
| 743 | TEST(F32_PPMM_8X8__NEON, k_eq_1_subtile) { |
| 744 | TEST_REQUIRES_ARM_NEON; |
| 745 | for (uint32_t m = 1; m <= 8; m++) { |
| 746 | for (uint32_t n = 1; n <= 8; n++) { |
| 747 | GemmMicrokernelTester() |
| 748 | .mr(8) |
| 749 | .nr(8) |
| 750 | .kr(1) |
| 751 | .sr(1) |
| 752 | .m(m) |
| 753 | .n(n) |
| 754 | .k(1) |
| 755 | .iterations(1) |
| 756 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 757 | } |
| 758 | } |
| 759 | } |
| 760 | |
| 761 | TEST(F32_PPMM_8X8__NEON, k_eq_1_subtile_m) { |
| 762 | TEST_REQUIRES_ARM_NEON; |
| 763 | for (uint32_t m = 1; m <= 8; m++) { |
| 764 | GemmMicrokernelTester() |
| 765 | .mr(8) |
| 766 | .nr(8) |
| 767 | .kr(1) |
| 768 | .sr(1) |
| 769 | .m(m) |
| 770 | .n(8) |
| 771 | .k(1) |
| 772 | .iterations(1) |
| 773 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 774 | } |
| 775 | } |
| 776 | |
| 777 | TEST(F32_PPMM_8X8__NEON, k_eq_1_subtile_n) { |
| 778 | TEST_REQUIRES_ARM_NEON; |
| 779 | for (uint32_t n = 1; n <= 8; n++) { |
| 780 | GemmMicrokernelTester() |
| 781 | .mr(8) |
| 782 | .nr(8) |
| 783 | .kr(1) |
| 784 | .sr(1) |
| 785 | .m(8) |
| 786 | .n(n) |
| 787 | .k(1) |
| 788 | .iterations(1) |
| 789 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 790 | } |
| 791 | } |
| 792 | |
| 793 | TEST(F32_PPMM_8X8__NEON, k_gt_1) { |
| 794 | TEST_REQUIRES_ARM_NEON; |
| 795 | for (size_t k = 2; k < 10; k++) { |
| 796 | GemmMicrokernelTester() |
| 797 | .mr(8) |
| 798 | .nr(8) |
| 799 | .kr(1) |
| 800 | .sr(1) |
| 801 | .m(8) |
| 802 | .n(8) |
| 803 | .k(k) |
| 804 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 805 | } |
| 806 | } |
| 807 | |
| 808 | TEST(F32_PPMM_8X8__NEON, k_gt_1_subtile) { |
| 809 | TEST_REQUIRES_ARM_NEON; |
| 810 | for (size_t k = 2; k < 10; k++) { |
| 811 | for (uint32_t m = 1; m <= 8; m++) { |
| 812 | for (uint32_t n = 1; n <= 8; n++) { |
| 813 | GemmMicrokernelTester() |
| 814 | .mr(8) |
| 815 | .nr(8) |
| 816 | .kr(1) |
| 817 | .sr(1) |
| 818 | .m(m) |
| 819 | .n(n) |
| 820 | .k(k) |
| 821 | .iterations(1) |
| 822 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 823 | } |
| 824 | } |
| 825 | } |
| 826 | } |
| 827 | |
| 828 | TEST(F32_PPMM_8X8__NEON, n_gt_8) { |
| 829 | TEST_REQUIRES_ARM_NEON; |
| 830 | for (uint32_t n = 9; n < 16; n++) { |
| 831 | for (size_t k = 1; k <= 5; k += 2) { |
| 832 | GemmMicrokernelTester() |
| 833 | .mr(8) |
| 834 | .nr(8) |
| 835 | .kr(1) |
| 836 | .sr(1) |
| 837 | .m(8) |
| 838 | .n(8) |
| 839 | .k(k) |
| 840 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 841 | } |
| 842 | } |
| 843 | } |
| 844 | |
| 845 | TEST(F32_PPMM_8X8__NEON, n_gt_8_strided_cn) { |
| 846 | TEST_REQUIRES_ARM_NEON; |
| 847 | for (uint32_t n = 9; n < 16; n++) { |
| 848 | for (size_t k = 1; k <= 5; k += 2) { |
| 849 | GemmMicrokernelTester() |
| 850 | .mr(8) |
| 851 | .nr(8) |
| 852 | .kr(1) |
| 853 | .sr(1) |
| 854 | .m(8) |
| 855 | .n(8) |
| 856 | .k(k) |
| 857 | .cn_stride(11) |
| 858 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 859 | } |
| 860 | } |
| 861 | } |
| 862 | |
| 863 | TEST(F32_PPMM_8X8__NEON, n_gt_8_strided_a) { |
| 864 | TEST_REQUIRES_ARM_NEON; |
| 865 | for (uint32_t n = 9; n < 16; n++) { |
| 866 | for (size_t k = 1; k <= 5; k += 2) { |
| 867 | GemmMicrokernelTester() |
| 868 | .mr(8) |
| 869 | .nr(8) |
| 870 | .kr(1) |
| 871 | .sr(1) |
| 872 | .m(8) |
| 873 | .n(n) |
| 874 | .k(k) |
| 875 | .a_stride(7) |
| 876 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 877 | } |
| 878 | } |
| 879 | } |
| 880 | |
| 881 | TEST(F32_PPMM_8X8__NEON, n_gt_8_subtile) { |
| 882 | TEST_REQUIRES_ARM_NEON; |
| 883 | for (uint32_t n = 9; n < 16; n++) { |
| 884 | for (size_t k = 1; k <= 5; k += 2) { |
| 885 | for (uint32_t m = 1; m <= 8; m++) { |
| 886 | GemmMicrokernelTester() |
| 887 | .mr(8) |
| 888 | .nr(8) |
| 889 | .kr(1) |
| 890 | .sr(1) |
| 891 | .m(m) |
| 892 | .n(n) |
| 893 | .k(k) |
| 894 | .iterations(1) |
| 895 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 896 | } |
| 897 | } |
| 898 | } |
| 899 | } |
| 900 | |
| 901 | TEST(F32_PPMM_8X8__NEON, n_div_8) { |
| 902 | TEST_REQUIRES_ARM_NEON; |
| 903 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 904 | for (size_t k = 1; k <= 5; k += 2) { |
| 905 | GemmMicrokernelTester() |
| 906 | .mr(8) |
| 907 | .nr(8) |
| 908 | .kr(1) |
| 909 | .sr(1) |
| 910 | .m(8) |
| 911 | .n(8) |
| 912 | .k(k) |
| 913 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 914 | } |
| 915 | } |
| 916 | } |
| 917 | |
| 918 | TEST(F32_PPMM_8X8__NEON, n_div_8_strided_cn) { |
| 919 | TEST_REQUIRES_ARM_NEON; |
| 920 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 921 | for (size_t k = 1; k <= 5; k += 2) { |
| 922 | GemmMicrokernelTester() |
| 923 | .mr(8) |
| 924 | .nr(8) |
| 925 | .kr(1) |
| 926 | .sr(1) |
| 927 | .m(8) |
| 928 | .n(n) |
| 929 | .k(k) |
| 930 | .cn_stride(11) |
| 931 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 932 | } |
| 933 | } |
| 934 | } |
| 935 | |
| 936 | TEST(F32_PPMM_8X8__NEON, n_div_8_strided_a) { |
| 937 | TEST_REQUIRES_ARM_NEON; |
| 938 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 939 | for (size_t k = 1; k <= 5; k += 2) { |
| 940 | GemmMicrokernelTester() |
| 941 | .mr(8) |
| 942 | .nr(8) |
| 943 | .kr(1) |
| 944 | .sr(1) |
| 945 | .m(8) |
| 946 | .n(n) |
| 947 | .k(k) |
| 948 | .a_stride(7) |
| 949 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 950 | } |
| 951 | } |
| 952 | } |
| 953 | |
| 954 | TEST(F32_PPMM_8X8__NEON, n_div_8_subtile) { |
| 955 | TEST_REQUIRES_ARM_NEON; |
| 956 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 957 | for (size_t k = 1; k <= 5; k += 2) { |
| 958 | for (uint32_t m = 1; m <= 8; m++) { |
| 959 | GemmMicrokernelTester() |
| 960 | .mr(8) |
| 961 | .nr(8) |
| 962 | .kr(1) |
| 963 | .sr(1) |
| 964 | .m(m) |
| 965 | .n(n) |
| 966 | .k(k) |
| 967 | .iterations(1) |
| 968 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 969 | } |
| 970 | } |
| 971 | } |
| 972 | } |
| 973 | |
| 974 | TEST(F32_PPMM_8X8__NEON, strided_cm_subtile) { |
| 975 | TEST_REQUIRES_ARM_NEON; |
| 976 | for (size_t k = 1; k <= 5; k += 2) { |
| 977 | for (uint32_t m = 1; m <= 8; m++) { |
| 978 | for (uint32_t n = 1; n <= 8; n++) { |
| 979 | GemmMicrokernelTester() |
| 980 | .mr(8) |
| 981 | .nr(8) |
| 982 | .kr(1) |
| 983 | .sr(1) |
| 984 | .m(m) |
| 985 | .n(n) |
| 986 | .k(k) |
| 987 | .cm_stride(11) |
| 988 | .iterations(1) |
| 989 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 990 | } |
| 991 | } |
| 992 | } |
| 993 | } |
| 994 | |
| 995 | TEST(F32_PPMM_8X8__NEON, qmin) { |
| 996 | TEST_REQUIRES_ARM_NEON; |
| 997 | GemmMicrokernelTester() |
| 998 | .mr(8) |
| 999 | .nr(8) |
| 1000 | .kr(1) |
| 1001 | .sr(1) |
| 1002 | .m(8) |
| 1003 | .n(8) |
| 1004 | .k(1) |
| 1005 | .qmin(128) |
| 1006 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 1007 | } |
| 1008 | |
| 1009 | TEST(F32_PPMM_8X8__NEON, qmax) { |
| 1010 | TEST_REQUIRES_ARM_NEON; |
| 1011 | GemmMicrokernelTester() |
| 1012 | .mr(8) |
| 1013 | .nr(8) |
| 1014 | .kr(1) |
| 1015 | .sr(1) |
| 1016 | .m(8) |
| 1017 | .n(8) |
| 1018 | .k(1) |
| 1019 | .qmax(128) |
| 1020 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 1021 | } |
| 1022 | |
| 1023 | TEST(F32_PPMM_8X8__NEON, strided_cm) { |
| 1024 | TEST_REQUIRES_ARM_NEON; |
| 1025 | GemmMicrokernelTester() |
| 1026 | .mr(8) |
| 1027 | .nr(8) |
| 1028 | .kr(1) |
| 1029 | .sr(1) |
| 1030 | .m(8) |
| 1031 | .n(8) |
| 1032 | .k(1) |
| 1033 | .cm_stride(11) |
| 1034 | .Test(xnn_f32_ppmm_ukernel_8x8__neon); |
| 1035 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1036 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1037 | |
| 1038 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1039 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1040 | TEST(F32_PPMM_8X8__NEONFMA, k_eq_1) { |
| 1041 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1042 | GemmMicrokernelTester() |
| 1043 | .mr(8) |
| 1044 | .nr(8) |
| 1045 | .kr(1) |
| 1046 | .sr(1) |
| 1047 | .m(8) |
| 1048 | .n(8) |
| 1049 | .k(1) |
| 1050 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1051 | } |
| 1052 | |
| 1053 | TEST(F32_PPMM_8X8__NEONFMA, strided_cn) { |
| 1054 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1055 | GemmMicrokernelTester() |
| 1056 | .mr(8) |
| 1057 | .nr(8) |
| 1058 | .kr(1) |
| 1059 | .sr(1) |
| 1060 | .m(8) |
| 1061 | .n(8) |
| 1062 | .k(1) |
| 1063 | .cn_stride(11) |
| 1064 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1065 | } |
| 1066 | |
| 1067 | TEST(F32_PPMM_8X8__NEONFMA, k_eq_1_strided_a) { |
| 1068 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1069 | GemmMicrokernelTester() |
| 1070 | .mr(8) |
| 1071 | .nr(8) |
| 1072 | .kr(1) |
| 1073 | .sr(1) |
| 1074 | .m(8) |
| 1075 | .n(8) |
| 1076 | .k(1) |
| 1077 | .a_stride(3) |
| 1078 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1079 | } |
| 1080 | |
| 1081 | TEST(F32_PPMM_8X8__NEONFMA, k_eq_1_subtile) { |
| 1082 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1083 | for (uint32_t m = 1; m <= 8; m++) { |
| 1084 | for (uint32_t n = 1; n <= 8; n++) { |
| 1085 | GemmMicrokernelTester() |
| 1086 | .mr(8) |
| 1087 | .nr(8) |
| 1088 | .kr(1) |
| 1089 | .sr(1) |
| 1090 | .m(m) |
| 1091 | .n(n) |
| 1092 | .k(1) |
| 1093 | .iterations(1) |
| 1094 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1095 | } |
| 1096 | } |
| 1097 | } |
| 1098 | |
| 1099 | TEST(F32_PPMM_8X8__NEONFMA, k_eq_1_subtile_m) { |
| 1100 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1101 | for (uint32_t m = 1; m <= 8; m++) { |
| 1102 | GemmMicrokernelTester() |
| 1103 | .mr(8) |
| 1104 | .nr(8) |
| 1105 | .kr(1) |
| 1106 | .sr(1) |
| 1107 | .m(m) |
| 1108 | .n(8) |
| 1109 | .k(1) |
| 1110 | .iterations(1) |
| 1111 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1112 | } |
| 1113 | } |
| 1114 | |
| 1115 | TEST(F32_PPMM_8X8__NEONFMA, k_eq_1_subtile_n) { |
| 1116 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1117 | for (uint32_t n = 1; n <= 8; n++) { |
| 1118 | GemmMicrokernelTester() |
| 1119 | .mr(8) |
| 1120 | .nr(8) |
| 1121 | .kr(1) |
| 1122 | .sr(1) |
| 1123 | .m(8) |
| 1124 | .n(n) |
| 1125 | .k(1) |
| 1126 | .iterations(1) |
| 1127 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1128 | } |
| 1129 | } |
| 1130 | |
| 1131 | TEST(F32_PPMM_8X8__NEONFMA, k_gt_1) { |
| 1132 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1133 | for (size_t k = 2; k < 10; k++) { |
| 1134 | GemmMicrokernelTester() |
| 1135 | .mr(8) |
| 1136 | .nr(8) |
| 1137 | .kr(1) |
| 1138 | .sr(1) |
| 1139 | .m(8) |
| 1140 | .n(8) |
| 1141 | .k(k) |
| 1142 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1143 | } |
| 1144 | } |
| 1145 | |
| 1146 | TEST(F32_PPMM_8X8__NEONFMA, k_gt_1_subtile) { |
| 1147 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1148 | for (size_t k = 2; k < 10; k++) { |
| 1149 | for (uint32_t m = 1; m <= 8; m++) { |
| 1150 | for (uint32_t n = 1; n <= 8; n++) { |
| 1151 | GemmMicrokernelTester() |
| 1152 | .mr(8) |
| 1153 | .nr(8) |
| 1154 | .kr(1) |
| 1155 | .sr(1) |
| 1156 | .m(m) |
| 1157 | .n(n) |
| 1158 | .k(k) |
| 1159 | .iterations(1) |
| 1160 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1161 | } |
| 1162 | } |
| 1163 | } |
| 1164 | } |
| 1165 | |
| 1166 | TEST(F32_PPMM_8X8__NEONFMA, n_gt_8) { |
| 1167 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1168 | for (uint32_t n = 9; n < 16; n++) { |
| 1169 | for (size_t k = 1; k <= 5; k += 2) { |
| 1170 | GemmMicrokernelTester() |
| 1171 | .mr(8) |
| 1172 | .nr(8) |
| 1173 | .kr(1) |
| 1174 | .sr(1) |
| 1175 | .m(8) |
| 1176 | .n(8) |
| 1177 | .k(k) |
| 1178 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1179 | } |
| 1180 | } |
| 1181 | } |
| 1182 | |
| 1183 | TEST(F32_PPMM_8X8__NEONFMA, n_gt_8_strided_cn) { |
| 1184 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1185 | for (uint32_t n = 9; n < 16; n++) { |
| 1186 | for (size_t k = 1; k <= 5; k += 2) { |
| 1187 | GemmMicrokernelTester() |
| 1188 | .mr(8) |
| 1189 | .nr(8) |
| 1190 | .kr(1) |
| 1191 | .sr(1) |
| 1192 | .m(8) |
| 1193 | .n(8) |
| 1194 | .k(k) |
| 1195 | .cn_stride(11) |
| 1196 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1197 | } |
| 1198 | } |
| 1199 | } |
| 1200 | |
| 1201 | TEST(F32_PPMM_8X8__NEONFMA, n_gt_8_strided_a) { |
| 1202 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1203 | for (uint32_t n = 9; n < 16; n++) { |
| 1204 | for (size_t k = 1; k <= 5; k += 2) { |
| 1205 | GemmMicrokernelTester() |
| 1206 | .mr(8) |
| 1207 | .nr(8) |
| 1208 | .kr(1) |
| 1209 | .sr(1) |
| 1210 | .m(8) |
| 1211 | .n(n) |
| 1212 | .k(k) |
| 1213 | .a_stride(7) |
| 1214 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1215 | } |
| 1216 | } |
| 1217 | } |
| 1218 | |
| 1219 | TEST(F32_PPMM_8X8__NEONFMA, n_gt_8_subtile) { |
| 1220 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1221 | for (uint32_t n = 9; n < 16; n++) { |
| 1222 | for (size_t k = 1; k <= 5; k += 2) { |
| 1223 | for (uint32_t m = 1; m <= 8; m++) { |
| 1224 | GemmMicrokernelTester() |
| 1225 | .mr(8) |
| 1226 | .nr(8) |
| 1227 | .kr(1) |
| 1228 | .sr(1) |
| 1229 | .m(m) |
| 1230 | .n(n) |
| 1231 | .k(k) |
| 1232 | .iterations(1) |
| 1233 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1234 | } |
| 1235 | } |
| 1236 | } |
| 1237 | } |
| 1238 | |
| 1239 | TEST(F32_PPMM_8X8__NEONFMA, n_div_8) { |
| 1240 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1241 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1242 | for (size_t k = 1; k <= 5; k += 2) { |
| 1243 | GemmMicrokernelTester() |
| 1244 | .mr(8) |
| 1245 | .nr(8) |
| 1246 | .kr(1) |
| 1247 | .sr(1) |
| 1248 | .m(8) |
| 1249 | .n(8) |
| 1250 | .k(k) |
| 1251 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1252 | } |
| 1253 | } |
| 1254 | } |
| 1255 | |
| 1256 | TEST(F32_PPMM_8X8__NEONFMA, n_div_8_strided_cn) { |
| 1257 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1258 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1259 | for (size_t k = 1; k <= 5; k += 2) { |
| 1260 | GemmMicrokernelTester() |
| 1261 | .mr(8) |
| 1262 | .nr(8) |
| 1263 | .kr(1) |
| 1264 | .sr(1) |
| 1265 | .m(8) |
| 1266 | .n(n) |
| 1267 | .k(k) |
| 1268 | .cn_stride(11) |
| 1269 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1270 | } |
| 1271 | } |
| 1272 | } |
| 1273 | |
| 1274 | TEST(F32_PPMM_8X8__NEONFMA, n_div_8_strided_a) { |
| 1275 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1276 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1277 | for (size_t k = 1; k <= 5; k += 2) { |
| 1278 | GemmMicrokernelTester() |
| 1279 | .mr(8) |
| 1280 | .nr(8) |
| 1281 | .kr(1) |
| 1282 | .sr(1) |
| 1283 | .m(8) |
| 1284 | .n(n) |
| 1285 | .k(k) |
| 1286 | .a_stride(7) |
| 1287 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1288 | } |
| 1289 | } |
| 1290 | } |
| 1291 | |
| 1292 | TEST(F32_PPMM_8X8__NEONFMA, n_div_8_subtile) { |
| 1293 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1294 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1295 | for (size_t k = 1; k <= 5; k += 2) { |
| 1296 | for (uint32_t m = 1; m <= 8; m++) { |
| 1297 | GemmMicrokernelTester() |
| 1298 | .mr(8) |
| 1299 | .nr(8) |
| 1300 | .kr(1) |
| 1301 | .sr(1) |
| 1302 | .m(m) |
| 1303 | .n(n) |
| 1304 | .k(k) |
| 1305 | .iterations(1) |
| 1306 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1307 | } |
| 1308 | } |
| 1309 | } |
| 1310 | } |
| 1311 | |
| 1312 | TEST(F32_PPMM_8X8__NEONFMA, strided_cm_subtile) { |
| 1313 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1314 | for (size_t k = 1; k <= 5; k += 2) { |
| 1315 | for (uint32_t m = 1; m <= 8; m++) { |
| 1316 | for (uint32_t n = 1; n <= 8; n++) { |
| 1317 | GemmMicrokernelTester() |
| 1318 | .mr(8) |
| 1319 | .nr(8) |
| 1320 | .kr(1) |
| 1321 | .sr(1) |
| 1322 | .m(m) |
| 1323 | .n(n) |
| 1324 | .k(k) |
| 1325 | .cm_stride(11) |
| 1326 | .iterations(1) |
| 1327 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1328 | } |
| 1329 | } |
| 1330 | } |
| 1331 | } |
| 1332 | |
| 1333 | TEST(F32_PPMM_8X8__NEONFMA, qmin) { |
| 1334 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1335 | GemmMicrokernelTester() |
| 1336 | .mr(8) |
| 1337 | .nr(8) |
| 1338 | .kr(1) |
| 1339 | .sr(1) |
| 1340 | .m(8) |
| 1341 | .n(8) |
| 1342 | .k(1) |
| 1343 | .qmin(128) |
| 1344 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1345 | } |
| 1346 | |
| 1347 | TEST(F32_PPMM_8X8__NEONFMA, qmax) { |
| 1348 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1349 | GemmMicrokernelTester() |
| 1350 | .mr(8) |
| 1351 | .nr(8) |
| 1352 | .kr(1) |
| 1353 | .sr(1) |
| 1354 | .m(8) |
| 1355 | .n(8) |
| 1356 | .k(1) |
| 1357 | .qmax(128) |
| 1358 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1359 | } |
| 1360 | |
| 1361 | TEST(F32_PPMM_8X8__NEONFMA, strided_cm) { |
| 1362 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1363 | GemmMicrokernelTester() |
| 1364 | .mr(8) |
| 1365 | .nr(8) |
| 1366 | .kr(1) |
| 1367 | .sr(1) |
| 1368 | .m(8) |
| 1369 | .n(8) |
| 1370 | .k(1) |
| 1371 | .cm_stride(11) |
| 1372 | .Test(xnn_f32_ppmm_ukernel_8x8__neonfma); |
| 1373 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1374 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1375 | |
| 1376 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1377 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1378 | TEST(F32_PPMM_4X8__SSE, k_eq_1) { |
| 1379 | TEST_REQUIRES_X86_SSE; |
| 1380 | GemmMicrokernelTester() |
| 1381 | .mr(4) |
| 1382 | .nr(8) |
| 1383 | .kr(1) |
| 1384 | .sr(1) |
| 1385 | .m(4) |
| 1386 | .n(8) |
| 1387 | .k(1) |
| 1388 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1389 | } |
| 1390 | |
| 1391 | TEST(F32_PPMM_4X8__SSE, strided_cn) { |
| 1392 | TEST_REQUIRES_X86_SSE; |
| 1393 | GemmMicrokernelTester() |
| 1394 | .mr(4) |
| 1395 | .nr(8) |
| 1396 | .kr(1) |
| 1397 | .sr(1) |
| 1398 | .m(4) |
| 1399 | .n(8) |
| 1400 | .k(1) |
| 1401 | .cn_stride(11) |
| 1402 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1403 | } |
| 1404 | |
| 1405 | TEST(F32_PPMM_4X8__SSE, k_eq_1_strided_a) { |
| 1406 | TEST_REQUIRES_X86_SSE; |
| 1407 | GemmMicrokernelTester() |
| 1408 | .mr(4) |
| 1409 | .nr(8) |
| 1410 | .kr(1) |
| 1411 | .sr(1) |
| 1412 | .m(4) |
| 1413 | .n(8) |
| 1414 | .k(1) |
| 1415 | .a_stride(3) |
| 1416 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1417 | } |
| 1418 | |
| 1419 | TEST(F32_PPMM_4X8__SSE, k_eq_1_subtile) { |
| 1420 | TEST_REQUIRES_X86_SSE; |
| 1421 | for (uint32_t m = 1; m <= 4; m++) { |
| 1422 | for (uint32_t n = 1; n <= 8; n++) { |
| 1423 | GemmMicrokernelTester() |
| 1424 | .mr(4) |
| 1425 | .nr(8) |
| 1426 | .kr(1) |
| 1427 | .sr(1) |
| 1428 | .m(m) |
| 1429 | .n(n) |
| 1430 | .k(1) |
| 1431 | .iterations(1) |
| 1432 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1433 | } |
| 1434 | } |
| 1435 | } |
| 1436 | |
| 1437 | TEST(F32_PPMM_4X8__SSE, k_eq_1_subtile_m) { |
| 1438 | TEST_REQUIRES_X86_SSE; |
| 1439 | for (uint32_t m = 1; m <= 4; m++) { |
| 1440 | GemmMicrokernelTester() |
| 1441 | .mr(4) |
| 1442 | .nr(8) |
| 1443 | .kr(1) |
| 1444 | .sr(1) |
| 1445 | .m(m) |
| 1446 | .n(8) |
| 1447 | .k(1) |
| 1448 | .iterations(1) |
| 1449 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1450 | } |
| 1451 | } |
| 1452 | |
| 1453 | TEST(F32_PPMM_4X8__SSE, k_eq_1_subtile_n) { |
| 1454 | TEST_REQUIRES_X86_SSE; |
| 1455 | for (uint32_t n = 1; n <= 8; n++) { |
| 1456 | GemmMicrokernelTester() |
| 1457 | .mr(4) |
| 1458 | .nr(8) |
| 1459 | .kr(1) |
| 1460 | .sr(1) |
| 1461 | .m(4) |
| 1462 | .n(n) |
| 1463 | .k(1) |
| 1464 | .iterations(1) |
| 1465 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1466 | } |
| 1467 | } |
| 1468 | |
| 1469 | TEST(F32_PPMM_4X8__SSE, k_gt_1) { |
| 1470 | TEST_REQUIRES_X86_SSE; |
| 1471 | for (size_t k = 2; k < 10; k++) { |
| 1472 | GemmMicrokernelTester() |
| 1473 | .mr(4) |
| 1474 | .nr(8) |
| 1475 | .kr(1) |
| 1476 | .sr(1) |
| 1477 | .m(4) |
| 1478 | .n(8) |
| 1479 | .k(k) |
| 1480 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1481 | } |
| 1482 | } |
| 1483 | |
| 1484 | TEST(F32_PPMM_4X8__SSE, k_gt_1_subtile) { |
| 1485 | TEST_REQUIRES_X86_SSE; |
| 1486 | for (size_t k = 2; k < 10; k++) { |
| 1487 | for (uint32_t m = 1; m <= 4; m++) { |
| 1488 | for (uint32_t n = 1; n <= 8; n++) { |
| 1489 | GemmMicrokernelTester() |
| 1490 | .mr(4) |
| 1491 | .nr(8) |
| 1492 | .kr(1) |
| 1493 | .sr(1) |
| 1494 | .m(m) |
| 1495 | .n(n) |
| 1496 | .k(k) |
| 1497 | .iterations(1) |
| 1498 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1499 | } |
| 1500 | } |
| 1501 | } |
| 1502 | } |
| 1503 | |
| 1504 | TEST(F32_PPMM_4X8__SSE, n_gt_8) { |
| 1505 | TEST_REQUIRES_X86_SSE; |
| 1506 | for (uint32_t n = 9; n < 16; n++) { |
| 1507 | for (size_t k = 1; k <= 5; k += 2) { |
| 1508 | GemmMicrokernelTester() |
| 1509 | .mr(4) |
| 1510 | .nr(8) |
| 1511 | .kr(1) |
| 1512 | .sr(1) |
| 1513 | .m(4) |
| 1514 | .n(8) |
| 1515 | .k(k) |
| 1516 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1517 | } |
| 1518 | } |
| 1519 | } |
| 1520 | |
| 1521 | TEST(F32_PPMM_4X8__SSE, n_gt_8_strided_cn) { |
| 1522 | TEST_REQUIRES_X86_SSE; |
| 1523 | for (uint32_t n = 9; n < 16; n++) { |
| 1524 | for (size_t k = 1; k <= 5; k += 2) { |
| 1525 | GemmMicrokernelTester() |
| 1526 | .mr(4) |
| 1527 | .nr(8) |
| 1528 | .kr(1) |
| 1529 | .sr(1) |
| 1530 | .m(4) |
| 1531 | .n(8) |
| 1532 | .k(k) |
| 1533 | .cn_stride(11) |
| 1534 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1535 | } |
| 1536 | } |
| 1537 | } |
| 1538 | |
| 1539 | TEST(F32_PPMM_4X8__SSE, n_gt_8_strided_a) { |
| 1540 | TEST_REQUIRES_X86_SSE; |
| 1541 | for (uint32_t n = 9; n < 16; n++) { |
| 1542 | for (size_t k = 1; k <= 5; k += 2) { |
| 1543 | GemmMicrokernelTester() |
| 1544 | .mr(4) |
| 1545 | .nr(8) |
| 1546 | .kr(1) |
| 1547 | .sr(1) |
| 1548 | .m(4) |
| 1549 | .n(n) |
| 1550 | .k(k) |
| 1551 | .a_stride(7) |
| 1552 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1553 | } |
| 1554 | } |
| 1555 | } |
| 1556 | |
| 1557 | TEST(F32_PPMM_4X8__SSE, n_gt_8_subtile) { |
| 1558 | TEST_REQUIRES_X86_SSE; |
| 1559 | for (uint32_t n = 9; n < 16; n++) { |
| 1560 | for (size_t k = 1; k <= 5; k += 2) { |
| 1561 | for (uint32_t m = 1; m <= 4; m++) { |
| 1562 | GemmMicrokernelTester() |
| 1563 | .mr(4) |
| 1564 | .nr(8) |
| 1565 | .kr(1) |
| 1566 | .sr(1) |
| 1567 | .m(m) |
| 1568 | .n(n) |
| 1569 | .k(k) |
| 1570 | .iterations(1) |
| 1571 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1572 | } |
| 1573 | } |
| 1574 | } |
| 1575 | } |
| 1576 | |
| 1577 | TEST(F32_PPMM_4X8__SSE, n_div_8) { |
| 1578 | TEST_REQUIRES_X86_SSE; |
| 1579 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1580 | for (size_t k = 1; k <= 5; k += 2) { |
| 1581 | GemmMicrokernelTester() |
| 1582 | .mr(4) |
| 1583 | .nr(8) |
| 1584 | .kr(1) |
| 1585 | .sr(1) |
| 1586 | .m(4) |
| 1587 | .n(8) |
| 1588 | .k(k) |
| 1589 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1590 | } |
| 1591 | } |
| 1592 | } |
| 1593 | |
| 1594 | TEST(F32_PPMM_4X8__SSE, n_div_8_strided_cn) { |
| 1595 | TEST_REQUIRES_X86_SSE; |
| 1596 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1597 | for (size_t k = 1; k <= 5; k += 2) { |
| 1598 | GemmMicrokernelTester() |
| 1599 | .mr(4) |
| 1600 | .nr(8) |
| 1601 | .kr(1) |
| 1602 | .sr(1) |
| 1603 | .m(4) |
| 1604 | .n(n) |
| 1605 | .k(k) |
| 1606 | .cn_stride(11) |
| 1607 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1608 | } |
| 1609 | } |
| 1610 | } |
| 1611 | |
| 1612 | TEST(F32_PPMM_4X8__SSE, n_div_8_strided_a) { |
| 1613 | TEST_REQUIRES_X86_SSE; |
| 1614 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1615 | for (size_t k = 1; k <= 5; k += 2) { |
| 1616 | GemmMicrokernelTester() |
| 1617 | .mr(4) |
| 1618 | .nr(8) |
| 1619 | .kr(1) |
| 1620 | .sr(1) |
| 1621 | .m(4) |
| 1622 | .n(n) |
| 1623 | .k(k) |
| 1624 | .a_stride(7) |
| 1625 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1626 | } |
| 1627 | } |
| 1628 | } |
| 1629 | |
| 1630 | TEST(F32_PPMM_4X8__SSE, n_div_8_subtile) { |
| 1631 | TEST_REQUIRES_X86_SSE; |
| 1632 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1633 | for (size_t k = 1; k <= 5; k += 2) { |
| 1634 | for (uint32_t m = 1; m <= 4; m++) { |
| 1635 | GemmMicrokernelTester() |
| 1636 | .mr(4) |
| 1637 | .nr(8) |
| 1638 | .kr(1) |
| 1639 | .sr(1) |
| 1640 | .m(m) |
| 1641 | .n(n) |
| 1642 | .k(k) |
| 1643 | .iterations(1) |
| 1644 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1645 | } |
| 1646 | } |
| 1647 | } |
| 1648 | } |
| 1649 | |
| 1650 | TEST(F32_PPMM_4X8__SSE, strided_cm_subtile) { |
| 1651 | TEST_REQUIRES_X86_SSE; |
| 1652 | for (size_t k = 1; k <= 5; k += 2) { |
| 1653 | for (uint32_t m = 1; m <= 4; m++) { |
| 1654 | for (uint32_t n = 1; n <= 8; n++) { |
| 1655 | GemmMicrokernelTester() |
| 1656 | .mr(4) |
| 1657 | .nr(8) |
| 1658 | .kr(1) |
| 1659 | .sr(1) |
| 1660 | .m(m) |
| 1661 | .n(n) |
| 1662 | .k(k) |
| 1663 | .cm_stride(11) |
| 1664 | .iterations(1) |
| 1665 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1666 | } |
| 1667 | } |
| 1668 | } |
| 1669 | } |
| 1670 | |
| 1671 | TEST(F32_PPMM_4X8__SSE, qmin) { |
| 1672 | TEST_REQUIRES_X86_SSE; |
| 1673 | GemmMicrokernelTester() |
| 1674 | .mr(4) |
| 1675 | .nr(8) |
| 1676 | .kr(1) |
| 1677 | .sr(1) |
| 1678 | .m(4) |
| 1679 | .n(8) |
| 1680 | .k(1) |
| 1681 | .qmin(128) |
| 1682 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1683 | } |
| 1684 | |
| 1685 | TEST(F32_PPMM_4X8__SSE, qmax) { |
| 1686 | TEST_REQUIRES_X86_SSE; |
| 1687 | GemmMicrokernelTester() |
| 1688 | .mr(4) |
| 1689 | .nr(8) |
| 1690 | .kr(1) |
| 1691 | .sr(1) |
| 1692 | .m(4) |
| 1693 | .n(8) |
| 1694 | .k(1) |
| 1695 | .qmax(128) |
| 1696 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1697 | } |
| 1698 | |
| 1699 | TEST(F32_PPMM_4X8__SSE, strided_cm) { |
| 1700 | TEST_REQUIRES_X86_SSE; |
| 1701 | GemmMicrokernelTester() |
| 1702 | .mr(4) |
| 1703 | .nr(8) |
| 1704 | .kr(1) |
| 1705 | .sr(1) |
| 1706 | .m(4) |
| 1707 | .n(8) |
| 1708 | .k(1) |
| 1709 | .cm_stride(11) |
| 1710 | .Test(xnn_f32_ppmm_ukernel_4x8__sse); |
| 1711 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1712 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1713 | |
| 1714 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1715 | #if !XNN_ARCH_ASMJS && !XNN_ARCH_WASM |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1716 | TEST(F32_PPMM_4X8__PSIMD, k_eq_1) { |
| 1717 | TEST_REQUIRES_PSIMD; |
| 1718 | GemmMicrokernelTester() |
| 1719 | .mr(4) |
| 1720 | .nr(8) |
| 1721 | .kr(1) |
| 1722 | .sr(1) |
| 1723 | .m(4) |
| 1724 | .n(8) |
| 1725 | .k(1) |
| 1726 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1727 | } |
| 1728 | |
| 1729 | TEST(F32_PPMM_4X8__PSIMD, strided_cn) { |
| 1730 | TEST_REQUIRES_PSIMD; |
| 1731 | GemmMicrokernelTester() |
| 1732 | .mr(4) |
| 1733 | .nr(8) |
| 1734 | .kr(1) |
| 1735 | .sr(1) |
| 1736 | .m(4) |
| 1737 | .n(8) |
| 1738 | .k(1) |
| 1739 | .cn_stride(11) |
| 1740 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1741 | } |
| 1742 | |
| 1743 | TEST(F32_PPMM_4X8__PSIMD, k_eq_1_strided_a) { |
| 1744 | TEST_REQUIRES_PSIMD; |
| 1745 | GemmMicrokernelTester() |
| 1746 | .mr(4) |
| 1747 | .nr(8) |
| 1748 | .kr(1) |
| 1749 | .sr(1) |
| 1750 | .m(4) |
| 1751 | .n(8) |
| 1752 | .k(1) |
| 1753 | .a_stride(3) |
| 1754 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1755 | } |
| 1756 | |
| 1757 | TEST(F32_PPMM_4X8__PSIMD, k_eq_1_subtile) { |
| 1758 | TEST_REQUIRES_PSIMD; |
| 1759 | for (uint32_t m = 1; m <= 4; m++) { |
| 1760 | for (uint32_t n = 1; n <= 8; n++) { |
| 1761 | GemmMicrokernelTester() |
| 1762 | .mr(4) |
| 1763 | .nr(8) |
| 1764 | .kr(1) |
| 1765 | .sr(1) |
| 1766 | .m(m) |
| 1767 | .n(n) |
| 1768 | .k(1) |
| 1769 | .iterations(1) |
| 1770 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1771 | } |
| 1772 | } |
| 1773 | } |
| 1774 | |
| 1775 | TEST(F32_PPMM_4X8__PSIMD, k_eq_1_subtile_m) { |
| 1776 | TEST_REQUIRES_PSIMD; |
| 1777 | for (uint32_t m = 1; m <= 4; m++) { |
| 1778 | GemmMicrokernelTester() |
| 1779 | .mr(4) |
| 1780 | .nr(8) |
| 1781 | .kr(1) |
| 1782 | .sr(1) |
| 1783 | .m(m) |
| 1784 | .n(8) |
| 1785 | .k(1) |
| 1786 | .iterations(1) |
| 1787 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1788 | } |
| 1789 | } |
| 1790 | |
| 1791 | TEST(F32_PPMM_4X8__PSIMD, k_eq_1_subtile_n) { |
| 1792 | TEST_REQUIRES_PSIMD; |
| 1793 | for (uint32_t n = 1; n <= 8; n++) { |
| 1794 | GemmMicrokernelTester() |
| 1795 | .mr(4) |
| 1796 | .nr(8) |
| 1797 | .kr(1) |
| 1798 | .sr(1) |
| 1799 | .m(4) |
| 1800 | .n(n) |
| 1801 | .k(1) |
| 1802 | .iterations(1) |
| 1803 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1804 | } |
| 1805 | } |
| 1806 | |
| 1807 | TEST(F32_PPMM_4X8__PSIMD, k_gt_1) { |
| 1808 | TEST_REQUIRES_PSIMD; |
| 1809 | for (size_t k = 2; k < 10; k++) { |
| 1810 | GemmMicrokernelTester() |
| 1811 | .mr(4) |
| 1812 | .nr(8) |
| 1813 | .kr(1) |
| 1814 | .sr(1) |
| 1815 | .m(4) |
| 1816 | .n(8) |
| 1817 | .k(k) |
| 1818 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1819 | } |
| 1820 | } |
| 1821 | |
| 1822 | TEST(F32_PPMM_4X8__PSIMD, k_gt_1_subtile) { |
| 1823 | TEST_REQUIRES_PSIMD; |
| 1824 | for (size_t k = 2; k < 10; k++) { |
| 1825 | for (uint32_t m = 1; m <= 4; m++) { |
| 1826 | for (uint32_t n = 1; n <= 8; n++) { |
| 1827 | GemmMicrokernelTester() |
| 1828 | .mr(4) |
| 1829 | .nr(8) |
| 1830 | .kr(1) |
| 1831 | .sr(1) |
| 1832 | .m(m) |
| 1833 | .n(n) |
| 1834 | .k(k) |
| 1835 | .iterations(1) |
| 1836 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1837 | } |
| 1838 | } |
| 1839 | } |
| 1840 | } |
| 1841 | |
| 1842 | TEST(F32_PPMM_4X8__PSIMD, n_gt_8) { |
| 1843 | TEST_REQUIRES_PSIMD; |
| 1844 | for (uint32_t n = 9; n < 16; n++) { |
| 1845 | for (size_t k = 1; k <= 5; k += 2) { |
| 1846 | GemmMicrokernelTester() |
| 1847 | .mr(4) |
| 1848 | .nr(8) |
| 1849 | .kr(1) |
| 1850 | .sr(1) |
| 1851 | .m(4) |
| 1852 | .n(8) |
| 1853 | .k(k) |
| 1854 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1855 | } |
| 1856 | } |
| 1857 | } |
| 1858 | |
| 1859 | TEST(F32_PPMM_4X8__PSIMD, n_gt_8_strided_cn) { |
| 1860 | TEST_REQUIRES_PSIMD; |
| 1861 | for (uint32_t n = 9; n < 16; n++) { |
| 1862 | for (size_t k = 1; k <= 5; k += 2) { |
| 1863 | GemmMicrokernelTester() |
| 1864 | .mr(4) |
| 1865 | .nr(8) |
| 1866 | .kr(1) |
| 1867 | .sr(1) |
| 1868 | .m(4) |
| 1869 | .n(8) |
| 1870 | .k(k) |
| 1871 | .cn_stride(11) |
| 1872 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1873 | } |
| 1874 | } |
| 1875 | } |
| 1876 | |
| 1877 | TEST(F32_PPMM_4X8__PSIMD, n_gt_8_strided_a) { |
| 1878 | TEST_REQUIRES_PSIMD; |
| 1879 | for (uint32_t n = 9; n < 16; n++) { |
| 1880 | for (size_t k = 1; k <= 5; k += 2) { |
| 1881 | GemmMicrokernelTester() |
| 1882 | .mr(4) |
| 1883 | .nr(8) |
| 1884 | .kr(1) |
| 1885 | .sr(1) |
| 1886 | .m(4) |
| 1887 | .n(n) |
| 1888 | .k(k) |
| 1889 | .a_stride(7) |
| 1890 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1891 | } |
| 1892 | } |
| 1893 | } |
| 1894 | |
| 1895 | TEST(F32_PPMM_4X8__PSIMD, n_gt_8_subtile) { |
| 1896 | TEST_REQUIRES_PSIMD; |
| 1897 | for (uint32_t n = 9; n < 16; n++) { |
| 1898 | for (size_t k = 1; k <= 5; k += 2) { |
| 1899 | for (uint32_t m = 1; m <= 4; m++) { |
| 1900 | GemmMicrokernelTester() |
| 1901 | .mr(4) |
| 1902 | .nr(8) |
| 1903 | .kr(1) |
| 1904 | .sr(1) |
| 1905 | .m(m) |
| 1906 | .n(n) |
| 1907 | .k(k) |
| 1908 | .iterations(1) |
| 1909 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1910 | } |
| 1911 | } |
| 1912 | } |
| 1913 | } |
| 1914 | |
| 1915 | TEST(F32_PPMM_4X8__PSIMD, n_div_8) { |
| 1916 | TEST_REQUIRES_PSIMD; |
| 1917 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1918 | for (size_t k = 1; k <= 5; k += 2) { |
| 1919 | GemmMicrokernelTester() |
| 1920 | .mr(4) |
| 1921 | .nr(8) |
| 1922 | .kr(1) |
| 1923 | .sr(1) |
| 1924 | .m(4) |
| 1925 | .n(8) |
| 1926 | .k(k) |
| 1927 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1928 | } |
| 1929 | } |
| 1930 | } |
| 1931 | |
| 1932 | TEST(F32_PPMM_4X8__PSIMD, n_div_8_strided_cn) { |
| 1933 | TEST_REQUIRES_PSIMD; |
| 1934 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1935 | for (size_t k = 1; k <= 5; k += 2) { |
| 1936 | GemmMicrokernelTester() |
| 1937 | .mr(4) |
| 1938 | .nr(8) |
| 1939 | .kr(1) |
| 1940 | .sr(1) |
| 1941 | .m(4) |
| 1942 | .n(n) |
| 1943 | .k(k) |
| 1944 | .cn_stride(11) |
| 1945 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1946 | } |
| 1947 | } |
| 1948 | } |
| 1949 | |
| 1950 | TEST(F32_PPMM_4X8__PSIMD, n_div_8_strided_a) { |
| 1951 | TEST_REQUIRES_PSIMD; |
| 1952 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1953 | for (size_t k = 1; k <= 5; k += 2) { |
| 1954 | GemmMicrokernelTester() |
| 1955 | .mr(4) |
| 1956 | .nr(8) |
| 1957 | .kr(1) |
| 1958 | .sr(1) |
| 1959 | .m(4) |
| 1960 | .n(n) |
| 1961 | .k(k) |
| 1962 | .a_stride(7) |
| 1963 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1964 | } |
| 1965 | } |
| 1966 | } |
| 1967 | |
| 1968 | TEST(F32_PPMM_4X8__PSIMD, n_div_8_subtile) { |
| 1969 | TEST_REQUIRES_PSIMD; |
| 1970 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1971 | for (size_t k = 1; k <= 5; k += 2) { |
| 1972 | for (uint32_t m = 1; m <= 4; m++) { |
| 1973 | GemmMicrokernelTester() |
| 1974 | .mr(4) |
| 1975 | .nr(8) |
| 1976 | .kr(1) |
| 1977 | .sr(1) |
| 1978 | .m(m) |
| 1979 | .n(n) |
| 1980 | .k(k) |
| 1981 | .iterations(1) |
| 1982 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 1983 | } |
| 1984 | } |
| 1985 | } |
| 1986 | } |
| 1987 | |
| 1988 | TEST(F32_PPMM_4X8__PSIMD, strided_cm_subtile) { |
| 1989 | TEST_REQUIRES_PSIMD; |
| 1990 | for (size_t k = 1; k <= 5; k += 2) { |
| 1991 | for (uint32_t m = 1; m <= 4; m++) { |
| 1992 | for (uint32_t n = 1; n <= 8; n++) { |
| 1993 | GemmMicrokernelTester() |
| 1994 | .mr(4) |
| 1995 | .nr(8) |
| 1996 | .kr(1) |
| 1997 | .sr(1) |
| 1998 | .m(m) |
| 1999 | .n(n) |
| 2000 | .k(k) |
| 2001 | .cm_stride(11) |
| 2002 | .iterations(1) |
| 2003 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 2004 | } |
| 2005 | } |
| 2006 | } |
| 2007 | } |
| 2008 | |
| 2009 | TEST(F32_PPMM_4X8__PSIMD, qmin) { |
| 2010 | TEST_REQUIRES_PSIMD; |
| 2011 | GemmMicrokernelTester() |
| 2012 | .mr(4) |
| 2013 | .nr(8) |
| 2014 | .kr(1) |
| 2015 | .sr(1) |
| 2016 | .m(4) |
| 2017 | .n(8) |
| 2018 | .k(1) |
| 2019 | .qmin(128) |
| 2020 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 2021 | } |
| 2022 | |
| 2023 | TEST(F32_PPMM_4X8__PSIMD, qmax) { |
| 2024 | TEST_REQUIRES_PSIMD; |
| 2025 | GemmMicrokernelTester() |
| 2026 | .mr(4) |
| 2027 | .nr(8) |
| 2028 | .kr(1) |
| 2029 | .sr(1) |
| 2030 | .m(4) |
| 2031 | .n(8) |
| 2032 | .k(1) |
| 2033 | .qmax(128) |
| 2034 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 2035 | } |
| 2036 | |
| 2037 | TEST(F32_PPMM_4X8__PSIMD, strided_cm) { |
| 2038 | TEST_REQUIRES_PSIMD; |
| 2039 | GemmMicrokernelTester() |
| 2040 | .mr(4) |
| 2041 | .nr(8) |
| 2042 | .kr(1) |
| 2043 | .sr(1) |
| 2044 | .m(4) |
| 2045 | .n(8) |
| 2046 | .k(1) |
| 2047 | .cm_stride(11) |
| 2048 | .Test(xnn_f32_ppmm_ukernel_4x8__psimd, GemmMicrokernelTester::Variant::Scalar); |
| 2049 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2050 | #endif // !XNN_ARCH_ASMJS && !XNN_ARCH_WASM |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2051 | |
| 2052 | |
| 2053 | TEST(F32_PPMM_4X2__SCALAR, k_eq_1) { |
| 2054 | GemmMicrokernelTester() |
| 2055 | .mr(4) |
| 2056 | .nr(2) |
| 2057 | .kr(1) |
| 2058 | .sr(1) |
| 2059 | .m(4) |
| 2060 | .n(2) |
| 2061 | .k(1) |
| 2062 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2063 | } |
| 2064 | |
| 2065 | TEST(F32_PPMM_4X2__SCALAR, strided_cn) { |
| 2066 | GemmMicrokernelTester() |
| 2067 | .mr(4) |
| 2068 | .nr(2) |
| 2069 | .kr(1) |
| 2070 | .sr(1) |
| 2071 | .m(4) |
| 2072 | .n(2) |
| 2073 | .k(1) |
| 2074 | .cn_stride(5) |
| 2075 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2076 | } |
| 2077 | |
| 2078 | TEST(F32_PPMM_4X2__SCALAR, k_eq_1_strided_a) { |
| 2079 | GemmMicrokernelTester() |
| 2080 | .mr(4) |
| 2081 | .nr(2) |
| 2082 | .kr(1) |
| 2083 | .sr(1) |
| 2084 | .m(4) |
| 2085 | .n(2) |
| 2086 | .k(1) |
| 2087 | .a_stride(3) |
| 2088 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2089 | } |
| 2090 | |
| 2091 | TEST(F32_PPMM_4X2__SCALAR, k_eq_1_subtile) { |
| 2092 | for (uint32_t m = 1; m <= 4; m++) { |
| 2093 | for (uint32_t n = 1; n <= 2; n++) { |
| 2094 | GemmMicrokernelTester() |
| 2095 | .mr(4) |
| 2096 | .nr(2) |
| 2097 | .kr(1) |
| 2098 | .sr(1) |
| 2099 | .m(m) |
| 2100 | .n(n) |
| 2101 | .k(1) |
| 2102 | .iterations(1) |
| 2103 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2104 | } |
| 2105 | } |
| 2106 | } |
| 2107 | |
| 2108 | TEST(F32_PPMM_4X2__SCALAR, k_eq_1_subtile_m) { |
| 2109 | for (uint32_t m = 1; m <= 4; m++) { |
| 2110 | GemmMicrokernelTester() |
| 2111 | .mr(4) |
| 2112 | .nr(2) |
| 2113 | .kr(1) |
| 2114 | .sr(1) |
| 2115 | .m(m) |
| 2116 | .n(2) |
| 2117 | .k(1) |
| 2118 | .iterations(1) |
| 2119 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2120 | } |
| 2121 | } |
| 2122 | |
| 2123 | TEST(F32_PPMM_4X2__SCALAR, k_eq_1_subtile_n) { |
| 2124 | for (uint32_t n = 1; n <= 2; n++) { |
| 2125 | GemmMicrokernelTester() |
| 2126 | .mr(4) |
| 2127 | .nr(2) |
| 2128 | .kr(1) |
| 2129 | .sr(1) |
| 2130 | .m(4) |
| 2131 | .n(n) |
| 2132 | .k(1) |
| 2133 | .iterations(1) |
| 2134 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2135 | } |
| 2136 | } |
| 2137 | |
| 2138 | TEST(F32_PPMM_4X2__SCALAR, k_gt_1) { |
| 2139 | for (size_t k = 2; k < 10; k++) { |
| 2140 | GemmMicrokernelTester() |
| 2141 | .mr(4) |
| 2142 | .nr(2) |
| 2143 | .kr(1) |
| 2144 | .sr(1) |
| 2145 | .m(4) |
| 2146 | .n(2) |
| 2147 | .k(k) |
| 2148 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2149 | } |
| 2150 | } |
| 2151 | |
| 2152 | TEST(F32_PPMM_4X2__SCALAR, k_gt_1_subtile) { |
| 2153 | for (size_t k = 2; k < 10; k++) { |
| 2154 | for (uint32_t m = 1; m <= 4; m++) { |
| 2155 | for (uint32_t n = 1; n <= 2; n++) { |
| 2156 | GemmMicrokernelTester() |
| 2157 | .mr(4) |
| 2158 | .nr(2) |
| 2159 | .kr(1) |
| 2160 | .sr(1) |
| 2161 | .m(m) |
| 2162 | .n(n) |
| 2163 | .k(k) |
| 2164 | .iterations(1) |
| 2165 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2166 | } |
| 2167 | } |
| 2168 | } |
| 2169 | } |
| 2170 | |
| 2171 | TEST(F32_PPMM_4X2__SCALAR, n_gt_2) { |
| 2172 | for (uint32_t n = 3; n < 4; n++) { |
| 2173 | for (size_t k = 1; k <= 5; k += 2) { |
| 2174 | GemmMicrokernelTester() |
| 2175 | .mr(4) |
| 2176 | .nr(2) |
| 2177 | .kr(1) |
| 2178 | .sr(1) |
| 2179 | .m(4) |
| 2180 | .n(2) |
| 2181 | .k(k) |
| 2182 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2183 | } |
| 2184 | } |
| 2185 | } |
| 2186 | |
| 2187 | TEST(F32_PPMM_4X2__SCALAR, n_gt_2_strided_cn) { |
| 2188 | for (uint32_t n = 3; n < 4; n++) { |
| 2189 | for (size_t k = 1; k <= 5; k += 2) { |
| 2190 | GemmMicrokernelTester() |
| 2191 | .mr(4) |
| 2192 | .nr(2) |
| 2193 | .kr(1) |
| 2194 | .sr(1) |
| 2195 | .m(4) |
| 2196 | .n(2) |
| 2197 | .k(k) |
| 2198 | .cn_stride(5) |
| 2199 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2200 | } |
| 2201 | } |
| 2202 | } |
| 2203 | |
| 2204 | TEST(F32_PPMM_4X2__SCALAR, n_gt_2_strided_a) { |
| 2205 | for (uint32_t n = 3; n < 4; n++) { |
| 2206 | for (size_t k = 1; k <= 5; k += 2) { |
| 2207 | GemmMicrokernelTester() |
| 2208 | .mr(4) |
| 2209 | .nr(2) |
| 2210 | .kr(1) |
| 2211 | .sr(1) |
| 2212 | .m(4) |
| 2213 | .n(n) |
| 2214 | .k(k) |
| 2215 | .a_stride(7) |
| 2216 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2217 | } |
| 2218 | } |
| 2219 | } |
| 2220 | |
| 2221 | TEST(F32_PPMM_4X2__SCALAR, n_gt_2_subtile) { |
| 2222 | for (uint32_t n = 3; n < 4; n++) { |
| 2223 | for (size_t k = 1; k <= 5; k += 2) { |
| 2224 | for (uint32_t m = 1; m <= 4; m++) { |
| 2225 | GemmMicrokernelTester() |
| 2226 | .mr(4) |
| 2227 | .nr(2) |
| 2228 | .kr(1) |
| 2229 | .sr(1) |
| 2230 | .m(m) |
| 2231 | .n(n) |
| 2232 | .k(k) |
| 2233 | .iterations(1) |
| 2234 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2235 | } |
| 2236 | } |
| 2237 | } |
| 2238 | } |
| 2239 | |
| 2240 | TEST(F32_PPMM_4X2__SCALAR, n_div_2) { |
| 2241 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2242 | for (size_t k = 1; k <= 5; k += 2) { |
| 2243 | GemmMicrokernelTester() |
| 2244 | .mr(4) |
| 2245 | .nr(2) |
| 2246 | .kr(1) |
| 2247 | .sr(1) |
| 2248 | .m(4) |
| 2249 | .n(2) |
| 2250 | .k(k) |
| 2251 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2252 | } |
| 2253 | } |
| 2254 | } |
| 2255 | |
| 2256 | TEST(F32_PPMM_4X2__SCALAR, n_div_2_strided_cn) { |
| 2257 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2258 | for (size_t k = 1; k <= 5; k += 2) { |
| 2259 | GemmMicrokernelTester() |
| 2260 | .mr(4) |
| 2261 | .nr(2) |
| 2262 | .kr(1) |
| 2263 | .sr(1) |
| 2264 | .m(4) |
| 2265 | .n(n) |
| 2266 | .k(k) |
| 2267 | .cn_stride(5) |
| 2268 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2269 | } |
| 2270 | } |
| 2271 | } |
| 2272 | |
| 2273 | TEST(F32_PPMM_4X2__SCALAR, n_div_2_strided_a) { |
| 2274 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2275 | for (size_t k = 1; k <= 5; k += 2) { |
| 2276 | GemmMicrokernelTester() |
| 2277 | .mr(4) |
| 2278 | .nr(2) |
| 2279 | .kr(1) |
| 2280 | .sr(1) |
| 2281 | .m(4) |
| 2282 | .n(n) |
| 2283 | .k(k) |
| 2284 | .a_stride(7) |
| 2285 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2286 | } |
| 2287 | } |
| 2288 | } |
| 2289 | |
| 2290 | TEST(F32_PPMM_4X2__SCALAR, n_div_2_subtile) { |
| 2291 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2292 | for (size_t k = 1; k <= 5; k += 2) { |
| 2293 | for (uint32_t m = 1; m <= 4; m++) { |
| 2294 | GemmMicrokernelTester() |
| 2295 | .mr(4) |
| 2296 | .nr(2) |
| 2297 | .kr(1) |
| 2298 | .sr(1) |
| 2299 | .m(m) |
| 2300 | .n(n) |
| 2301 | .k(k) |
| 2302 | .iterations(1) |
| 2303 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2304 | } |
| 2305 | } |
| 2306 | } |
| 2307 | } |
| 2308 | |
| 2309 | TEST(F32_PPMM_4X2__SCALAR, strided_cm_subtile) { |
| 2310 | for (size_t k = 1; k <= 5; k += 2) { |
| 2311 | for (uint32_t m = 1; m <= 4; m++) { |
| 2312 | for (uint32_t n = 1; n <= 2; n++) { |
| 2313 | GemmMicrokernelTester() |
| 2314 | .mr(4) |
| 2315 | .nr(2) |
| 2316 | .kr(1) |
| 2317 | .sr(1) |
| 2318 | .m(m) |
| 2319 | .n(n) |
| 2320 | .k(k) |
| 2321 | .cm_stride(5) |
| 2322 | .iterations(1) |
| 2323 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2324 | } |
| 2325 | } |
| 2326 | } |
| 2327 | } |
| 2328 | |
| 2329 | TEST(F32_PPMM_4X2__SCALAR, qmin) { |
| 2330 | GemmMicrokernelTester() |
| 2331 | .mr(4) |
| 2332 | .nr(2) |
| 2333 | .kr(1) |
| 2334 | .sr(1) |
| 2335 | .m(4) |
| 2336 | .n(2) |
| 2337 | .k(1) |
| 2338 | .qmin(128) |
| 2339 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2340 | } |
| 2341 | |
| 2342 | TEST(F32_PPMM_4X2__SCALAR, qmax) { |
| 2343 | GemmMicrokernelTester() |
| 2344 | .mr(4) |
| 2345 | .nr(2) |
| 2346 | .kr(1) |
| 2347 | .sr(1) |
| 2348 | .m(4) |
| 2349 | .n(2) |
| 2350 | .k(1) |
| 2351 | .qmax(128) |
| 2352 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2353 | } |
| 2354 | |
| 2355 | TEST(F32_PPMM_4X2__SCALAR, strided_cm) { |
| 2356 | GemmMicrokernelTester() |
| 2357 | .mr(4) |
| 2358 | .nr(2) |
| 2359 | .kr(1) |
| 2360 | .sr(1) |
| 2361 | .m(4) |
| 2362 | .n(2) |
| 2363 | .k(1) |
| 2364 | .cm_stride(5) |
| 2365 | .Test(xnn_f32_ppmm_ukernel_4x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2366 | } |
| 2367 | |
| 2368 | |
| 2369 | TEST(F32_PPMM_2X4__SCALAR, k_eq_1) { |
| 2370 | GemmMicrokernelTester() |
| 2371 | .mr(2) |
| 2372 | .nr(4) |
| 2373 | .kr(1) |
| 2374 | .sr(1) |
| 2375 | .m(2) |
| 2376 | .n(4) |
| 2377 | .k(1) |
| 2378 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2379 | } |
| 2380 | |
| 2381 | TEST(F32_PPMM_2X4__SCALAR, strided_cn) { |
| 2382 | GemmMicrokernelTester() |
| 2383 | .mr(2) |
| 2384 | .nr(4) |
| 2385 | .kr(1) |
| 2386 | .sr(1) |
| 2387 | .m(2) |
| 2388 | .n(4) |
| 2389 | .k(1) |
| 2390 | .cn_stride(7) |
| 2391 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2392 | } |
| 2393 | |
| 2394 | TEST(F32_PPMM_2X4__SCALAR, k_eq_1_strided_a) { |
| 2395 | GemmMicrokernelTester() |
| 2396 | .mr(2) |
| 2397 | .nr(4) |
| 2398 | .kr(1) |
| 2399 | .sr(1) |
| 2400 | .m(2) |
| 2401 | .n(4) |
| 2402 | .k(1) |
| 2403 | .a_stride(3) |
| 2404 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2405 | } |
| 2406 | |
| 2407 | TEST(F32_PPMM_2X4__SCALAR, k_eq_1_subtile) { |
| 2408 | for (uint32_t m = 1; m <= 2; m++) { |
| 2409 | for (uint32_t n = 1; n <= 4; n++) { |
| 2410 | GemmMicrokernelTester() |
| 2411 | .mr(2) |
| 2412 | .nr(4) |
| 2413 | .kr(1) |
| 2414 | .sr(1) |
| 2415 | .m(m) |
| 2416 | .n(n) |
| 2417 | .k(1) |
| 2418 | .iterations(1) |
| 2419 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2420 | } |
| 2421 | } |
| 2422 | } |
| 2423 | |
| 2424 | TEST(F32_PPMM_2X4__SCALAR, k_eq_1_subtile_m) { |
| 2425 | for (uint32_t m = 1; m <= 2; m++) { |
| 2426 | GemmMicrokernelTester() |
| 2427 | .mr(2) |
| 2428 | .nr(4) |
| 2429 | .kr(1) |
| 2430 | .sr(1) |
| 2431 | .m(m) |
| 2432 | .n(4) |
| 2433 | .k(1) |
| 2434 | .iterations(1) |
| 2435 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2436 | } |
| 2437 | } |
| 2438 | |
| 2439 | TEST(F32_PPMM_2X4__SCALAR, k_eq_1_subtile_n) { |
| 2440 | for (uint32_t n = 1; n <= 4; n++) { |
| 2441 | GemmMicrokernelTester() |
| 2442 | .mr(2) |
| 2443 | .nr(4) |
| 2444 | .kr(1) |
| 2445 | .sr(1) |
| 2446 | .m(2) |
| 2447 | .n(n) |
| 2448 | .k(1) |
| 2449 | .iterations(1) |
| 2450 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2451 | } |
| 2452 | } |
| 2453 | |
| 2454 | TEST(F32_PPMM_2X4__SCALAR, k_gt_1) { |
| 2455 | for (size_t k = 2; k < 10; k++) { |
| 2456 | GemmMicrokernelTester() |
| 2457 | .mr(2) |
| 2458 | .nr(4) |
| 2459 | .kr(1) |
| 2460 | .sr(1) |
| 2461 | .m(2) |
| 2462 | .n(4) |
| 2463 | .k(k) |
| 2464 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2465 | } |
| 2466 | } |
| 2467 | |
| 2468 | TEST(F32_PPMM_2X4__SCALAR, k_gt_1_subtile) { |
| 2469 | for (size_t k = 2; k < 10; k++) { |
| 2470 | for (uint32_t m = 1; m <= 2; m++) { |
| 2471 | for (uint32_t n = 1; n <= 4; n++) { |
| 2472 | GemmMicrokernelTester() |
| 2473 | .mr(2) |
| 2474 | .nr(4) |
| 2475 | .kr(1) |
| 2476 | .sr(1) |
| 2477 | .m(m) |
| 2478 | .n(n) |
| 2479 | .k(k) |
| 2480 | .iterations(1) |
| 2481 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2482 | } |
| 2483 | } |
| 2484 | } |
| 2485 | } |
| 2486 | |
| 2487 | TEST(F32_PPMM_2X4__SCALAR, n_gt_4) { |
| 2488 | for (uint32_t n = 5; n < 8; n++) { |
| 2489 | for (size_t k = 1; k <= 5; k += 2) { |
| 2490 | GemmMicrokernelTester() |
| 2491 | .mr(2) |
| 2492 | .nr(4) |
| 2493 | .kr(1) |
| 2494 | .sr(1) |
| 2495 | .m(2) |
| 2496 | .n(4) |
| 2497 | .k(k) |
| 2498 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2499 | } |
| 2500 | } |
| 2501 | } |
| 2502 | |
| 2503 | TEST(F32_PPMM_2X4__SCALAR, n_gt_4_strided_cn) { |
| 2504 | for (uint32_t n = 5; n < 8; n++) { |
| 2505 | for (size_t k = 1; k <= 5; k += 2) { |
| 2506 | GemmMicrokernelTester() |
| 2507 | .mr(2) |
| 2508 | .nr(4) |
| 2509 | .kr(1) |
| 2510 | .sr(1) |
| 2511 | .m(2) |
| 2512 | .n(4) |
| 2513 | .k(k) |
| 2514 | .cn_stride(7) |
| 2515 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2516 | } |
| 2517 | } |
| 2518 | } |
| 2519 | |
| 2520 | TEST(F32_PPMM_2X4__SCALAR, n_gt_4_strided_a) { |
| 2521 | for (uint32_t n = 5; n < 8; n++) { |
| 2522 | for (size_t k = 1; k <= 5; k += 2) { |
| 2523 | GemmMicrokernelTester() |
| 2524 | .mr(2) |
| 2525 | .nr(4) |
| 2526 | .kr(1) |
| 2527 | .sr(1) |
| 2528 | .m(2) |
| 2529 | .n(n) |
| 2530 | .k(k) |
| 2531 | .a_stride(7) |
| 2532 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2533 | } |
| 2534 | } |
| 2535 | } |
| 2536 | |
| 2537 | TEST(F32_PPMM_2X4__SCALAR, n_gt_4_subtile) { |
| 2538 | for (uint32_t n = 5; n < 8; n++) { |
| 2539 | for (size_t k = 1; k <= 5; k += 2) { |
| 2540 | for (uint32_t m = 1; m <= 2; m++) { |
| 2541 | GemmMicrokernelTester() |
| 2542 | .mr(2) |
| 2543 | .nr(4) |
| 2544 | .kr(1) |
| 2545 | .sr(1) |
| 2546 | .m(m) |
| 2547 | .n(n) |
| 2548 | .k(k) |
| 2549 | .iterations(1) |
| 2550 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2551 | } |
| 2552 | } |
| 2553 | } |
| 2554 | } |
| 2555 | |
| 2556 | TEST(F32_PPMM_2X4__SCALAR, n_div_4) { |
| 2557 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2558 | for (size_t k = 1; k <= 5; k += 2) { |
| 2559 | GemmMicrokernelTester() |
| 2560 | .mr(2) |
| 2561 | .nr(4) |
| 2562 | .kr(1) |
| 2563 | .sr(1) |
| 2564 | .m(2) |
| 2565 | .n(4) |
| 2566 | .k(k) |
| 2567 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2568 | } |
| 2569 | } |
| 2570 | } |
| 2571 | |
| 2572 | TEST(F32_PPMM_2X4__SCALAR, n_div_4_strided_cn) { |
| 2573 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2574 | for (size_t k = 1; k <= 5; k += 2) { |
| 2575 | GemmMicrokernelTester() |
| 2576 | .mr(2) |
| 2577 | .nr(4) |
| 2578 | .kr(1) |
| 2579 | .sr(1) |
| 2580 | .m(2) |
| 2581 | .n(n) |
| 2582 | .k(k) |
| 2583 | .cn_stride(7) |
| 2584 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2585 | } |
| 2586 | } |
| 2587 | } |
| 2588 | |
| 2589 | TEST(F32_PPMM_2X4__SCALAR, n_div_4_strided_a) { |
| 2590 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2591 | for (size_t k = 1; k <= 5; k += 2) { |
| 2592 | GemmMicrokernelTester() |
| 2593 | .mr(2) |
| 2594 | .nr(4) |
| 2595 | .kr(1) |
| 2596 | .sr(1) |
| 2597 | .m(2) |
| 2598 | .n(n) |
| 2599 | .k(k) |
| 2600 | .a_stride(7) |
| 2601 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2602 | } |
| 2603 | } |
| 2604 | } |
| 2605 | |
| 2606 | TEST(F32_PPMM_2X4__SCALAR, n_div_4_subtile) { |
| 2607 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2608 | for (size_t k = 1; k <= 5; k += 2) { |
| 2609 | for (uint32_t m = 1; m <= 2; m++) { |
| 2610 | GemmMicrokernelTester() |
| 2611 | .mr(2) |
| 2612 | .nr(4) |
| 2613 | .kr(1) |
| 2614 | .sr(1) |
| 2615 | .m(m) |
| 2616 | .n(n) |
| 2617 | .k(k) |
| 2618 | .iterations(1) |
| 2619 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2620 | } |
| 2621 | } |
| 2622 | } |
| 2623 | } |
| 2624 | |
| 2625 | TEST(F32_PPMM_2X4__SCALAR, strided_cm_subtile) { |
| 2626 | for (size_t k = 1; k <= 5; k += 2) { |
| 2627 | for (uint32_t m = 1; m <= 2; m++) { |
| 2628 | for (uint32_t n = 1; n <= 4; n++) { |
| 2629 | GemmMicrokernelTester() |
| 2630 | .mr(2) |
| 2631 | .nr(4) |
| 2632 | .kr(1) |
| 2633 | .sr(1) |
| 2634 | .m(m) |
| 2635 | .n(n) |
| 2636 | .k(k) |
| 2637 | .cm_stride(7) |
| 2638 | .iterations(1) |
| 2639 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2640 | } |
| 2641 | } |
| 2642 | } |
| 2643 | } |
| 2644 | |
| 2645 | TEST(F32_PPMM_2X4__SCALAR, qmin) { |
| 2646 | GemmMicrokernelTester() |
| 2647 | .mr(2) |
| 2648 | .nr(4) |
| 2649 | .kr(1) |
| 2650 | .sr(1) |
| 2651 | .m(2) |
| 2652 | .n(4) |
| 2653 | .k(1) |
| 2654 | .qmin(128) |
| 2655 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2656 | } |
| 2657 | |
| 2658 | TEST(F32_PPMM_2X4__SCALAR, qmax) { |
| 2659 | GemmMicrokernelTester() |
| 2660 | .mr(2) |
| 2661 | .nr(4) |
| 2662 | .kr(1) |
| 2663 | .sr(1) |
| 2664 | .m(2) |
| 2665 | .n(4) |
| 2666 | .k(1) |
| 2667 | .qmax(128) |
| 2668 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2669 | } |
| 2670 | |
| 2671 | TEST(F32_PPMM_2X4__SCALAR, strided_cm) { |
| 2672 | GemmMicrokernelTester() |
| 2673 | .mr(2) |
| 2674 | .nr(4) |
| 2675 | .kr(1) |
| 2676 | .sr(1) |
| 2677 | .m(2) |
| 2678 | .n(4) |
| 2679 | .k(1) |
| 2680 | .cm_stride(7) |
| 2681 | .Test(xnn_f32_ppmm_ukernel_2x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2682 | } |
| 2683 | |
| 2684 | |
| 2685 | TEST(F32_PPMM_4X4__SCALAR, k_eq_1) { |
| 2686 | GemmMicrokernelTester() |
| 2687 | .mr(4) |
| 2688 | .nr(4) |
| 2689 | .kr(1) |
| 2690 | .sr(1) |
| 2691 | .m(4) |
| 2692 | .n(4) |
| 2693 | .k(1) |
| 2694 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2695 | } |
| 2696 | |
| 2697 | TEST(F32_PPMM_4X4__SCALAR, strided_cn) { |
| 2698 | GemmMicrokernelTester() |
| 2699 | .mr(4) |
| 2700 | .nr(4) |
| 2701 | .kr(1) |
| 2702 | .sr(1) |
| 2703 | .m(4) |
| 2704 | .n(4) |
| 2705 | .k(1) |
| 2706 | .cn_stride(7) |
| 2707 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2708 | } |
| 2709 | |
| 2710 | TEST(F32_PPMM_4X4__SCALAR, k_eq_1_strided_a) { |
| 2711 | GemmMicrokernelTester() |
| 2712 | .mr(4) |
| 2713 | .nr(4) |
| 2714 | .kr(1) |
| 2715 | .sr(1) |
| 2716 | .m(4) |
| 2717 | .n(4) |
| 2718 | .k(1) |
| 2719 | .a_stride(3) |
| 2720 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2721 | } |
| 2722 | |
| 2723 | TEST(F32_PPMM_4X4__SCALAR, k_eq_1_subtile) { |
| 2724 | for (uint32_t m = 1; m <= 4; m++) { |
| 2725 | for (uint32_t n = 1; n <= 4; n++) { |
| 2726 | GemmMicrokernelTester() |
| 2727 | .mr(4) |
| 2728 | .nr(4) |
| 2729 | .kr(1) |
| 2730 | .sr(1) |
| 2731 | .m(m) |
| 2732 | .n(n) |
| 2733 | .k(1) |
| 2734 | .iterations(1) |
| 2735 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2736 | } |
| 2737 | } |
| 2738 | } |
| 2739 | |
| 2740 | TEST(F32_PPMM_4X4__SCALAR, k_eq_1_subtile_m) { |
| 2741 | for (uint32_t m = 1; m <= 4; m++) { |
| 2742 | GemmMicrokernelTester() |
| 2743 | .mr(4) |
| 2744 | .nr(4) |
| 2745 | .kr(1) |
| 2746 | .sr(1) |
| 2747 | .m(m) |
| 2748 | .n(4) |
| 2749 | .k(1) |
| 2750 | .iterations(1) |
| 2751 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2752 | } |
| 2753 | } |
| 2754 | |
| 2755 | TEST(F32_PPMM_4X4__SCALAR, k_eq_1_subtile_n) { |
| 2756 | for (uint32_t n = 1; n <= 4; n++) { |
| 2757 | GemmMicrokernelTester() |
| 2758 | .mr(4) |
| 2759 | .nr(4) |
| 2760 | .kr(1) |
| 2761 | .sr(1) |
| 2762 | .m(4) |
| 2763 | .n(n) |
| 2764 | .k(1) |
| 2765 | .iterations(1) |
| 2766 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2767 | } |
| 2768 | } |
| 2769 | |
| 2770 | TEST(F32_PPMM_4X4__SCALAR, k_gt_1) { |
| 2771 | for (size_t k = 2; k < 10; k++) { |
| 2772 | GemmMicrokernelTester() |
| 2773 | .mr(4) |
| 2774 | .nr(4) |
| 2775 | .kr(1) |
| 2776 | .sr(1) |
| 2777 | .m(4) |
| 2778 | .n(4) |
| 2779 | .k(k) |
| 2780 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2781 | } |
| 2782 | } |
| 2783 | |
| 2784 | TEST(F32_PPMM_4X4__SCALAR, k_gt_1_subtile) { |
| 2785 | for (size_t k = 2; k < 10; k++) { |
| 2786 | for (uint32_t m = 1; m <= 4; m++) { |
| 2787 | for (uint32_t n = 1; n <= 4; n++) { |
| 2788 | GemmMicrokernelTester() |
| 2789 | .mr(4) |
| 2790 | .nr(4) |
| 2791 | .kr(1) |
| 2792 | .sr(1) |
| 2793 | .m(m) |
| 2794 | .n(n) |
| 2795 | .k(k) |
| 2796 | .iterations(1) |
| 2797 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2798 | } |
| 2799 | } |
| 2800 | } |
| 2801 | } |
| 2802 | |
| 2803 | TEST(F32_PPMM_4X4__SCALAR, n_gt_4) { |
| 2804 | for (uint32_t n = 5; n < 8; n++) { |
| 2805 | for (size_t k = 1; k <= 5; k += 2) { |
| 2806 | GemmMicrokernelTester() |
| 2807 | .mr(4) |
| 2808 | .nr(4) |
| 2809 | .kr(1) |
| 2810 | .sr(1) |
| 2811 | .m(4) |
| 2812 | .n(4) |
| 2813 | .k(k) |
| 2814 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2815 | } |
| 2816 | } |
| 2817 | } |
| 2818 | |
| 2819 | TEST(F32_PPMM_4X4__SCALAR, n_gt_4_strided_cn) { |
| 2820 | for (uint32_t n = 5; n < 8; n++) { |
| 2821 | for (size_t k = 1; k <= 5; k += 2) { |
| 2822 | GemmMicrokernelTester() |
| 2823 | .mr(4) |
| 2824 | .nr(4) |
| 2825 | .kr(1) |
| 2826 | .sr(1) |
| 2827 | .m(4) |
| 2828 | .n(4) |
| 2829 | .k(k) |
| 2830 | .cn_stride(7) |
| 2831 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2832 | } |
| 2833 | } |
| 2834 | } |
| 2835 | |
| 2836 | TEST(F32_PPMM_4X4__SCALAR, n_gt_4_strided_a) { |
| 2837 | for (uint32_t n = 5; n < 8; n++) { |
| 2838 | for (size_t k = 1; k <= 5; k += 2) { |
| 2839 | GemmMicrokernelTester() |
| 2840 | .mr(4) |
| 2841 | .nr(4) |
| 2842 | .kr(1) |
| 2843 | .sr(1) |
| 2844 | .m(4) |
| 2845 | .n(n) |
| 2846 | .k(k) |
| 2847 | .a_stride(7) |
| 2848 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2849 | } |
| 2850 | } |
| 2851 | } |
| 2852 | |
| 2853 | TEST(F32_PPMM_4X4__SCALAR, n_gt_4_subtile) { |
| 2854 | for (uint32_t n = 5; n < 8; n++) { |
| 2855 | for (size_t k = 1; k <= 5; k += 2) { |
| 2856 | for (uint32_t m = 1; m <= 4; m++) { |
| 2857 | GemmMicrokernelTester() |
| 2858 | .mr(4) |
| 2859 | .nr(4) |
| 2860 | .kr(1) |
| 2861 | .sr(1) |
| 2862 | .m(m) |
| 2863 | .n(n) |
| 2864 | .k(k) |
| 2865 | .iterations(1) |
| 2866 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2867 | } |
| 2868 | } |
| 2869 | } |
| 2870 | } |
| 2871 | |
| 2872 | TEST(F32_PPMM_4X4__SCALAR, n_div_4) { |
| 2873 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2874 | for (size_t k = 1; k <= 5; k += 2) { |
| 2875 | GemmMicrokernelTester() |
| 2876 | .mr(4) |
| 2877 | .nr(4) |
| 2878 | .kr(1) |
| 2879 | .sr(1) |
| 2880 | .m(4) |
| 2881 | .n(4) |
| 2882 | .k(k) |
| 2883 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2884 | } |
| 2885 | } |
| 2886 | } |
| 2887 | |
| 2888 | TEST(F32_PPMM_4X4__SCALAR, n_div_4_strided_cn) { |
| 2889 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2890 | for (size_t k = 1; k <= 5; k += 2) { |
| 2891 | GemmMicrokernelTester() |
| 2892 | .mr(4) |
| 2893 | .nr(4) |
| 2894 | .kr(1) |
| 2895 | .sr(1) |
| 2896 | .m(4) |
| 2897 | .n(n) |
| 2898 | .k(k) |
| 2899 | .cn_stride(7) |
| 2900 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2901 | } |
| 2902 | } |
| 2903 | } |
| 2904 | |
| 2905 | TEST(F32_PPMM_4X4__SCALAR, n_div_4_strided_a) { |
| 2906 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2907 | for (size_t k = 1; k <= 5; k += 2) { |
| 2908 | GemmMicrokernelTester() |
| 2909 | .mr(4) |
| 2910 | .nr(4) |
| 2911 | .kr(1) |
| 2912 | .sr(1) |
| 2913 | .m(4) |
| 2914 | .n(n) |
| 2915 | .k(k) |
| 2916 | .a_stride(7) |
| 2917 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2918 | } |
| 2919 | } |
| 2920 | } |
| 2921 | |
| 2922 | TEST(F32_PPMM_4X4__SCALAR, n_div_4_subtile) { |
| 2923 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2924 | for (size_t k = 1; k <= 5; k += 2) { |
| 2925 | for (uint32_t m = 1; m <= 4; m++) { |
| 2926 | GemmMicrokernelTester() |
| 2927 | .mr(4) |
| 2928 | .nr(4) |
| 2929 | .kr(1) |
| 2930 | .sr(1) |
| 2931 | .m(m) |
| 2932 | .n(n) |
| 2933 | .k(k) |
| 2934 | .iterations(1) |
| 2935 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2936 | } |
| 2937 | } |
| 2938 | } |
| 2939 | } |
| 2940 | |
| 2941 | TEST(F32_PPMM_4X4__SCALAR, strided_cm_subtile) { |
| 2942 | for (size_t k = 1; k <= 5; k += 2) { |
| 2943 | for (uint32_t m = 1; m <= 4; m++) { |
| 2944 | for (uint32_t n = 1; n <= 4; n++) { |
| 2945 | GemmMicrokernelTester() |
| 2946 | .mr(4) |
| 2947 | .nr(4) |
| 2948 | .kr(1) |
| 2949 | .sr(1) |
| 2950 | .m(m) |
| 2951 | .n(n) |
| 2952 | .k(k) |
| 2953 | .cm_stride(7) |
| 2954 | .iterations(1) |
| 2955 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2956 | } |
| 2957 | } |
| 2958 | } |
| 2959 | } |
| 2960 | |
| 2961 | TEST(F32_PPMM_4X4__SCALAR, qmin) { |
| 2962 | GemmMicrokernelTester() |
| 2963 | .mr(4) |
| 2964 | .nr(4) |
| 2965 | .kr(1) |
| 2966 | .sr(1) |
| 2967 | .m(4) |
| 2968 | .n(4) |
| 2969 | .k(1) |
| 2970 | .qmin(128) |
| 2971 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2972 | } |
| 2973 | |
| 2974 | TEST(F32_PPMM_4X4__SCALAR, qmax) { |
| 2975 | GemmMicrokernelTester() |
| 2976 | .mr(4) |
| 2977 | .nr(4) |
| 2978 | .kr(1) |
| 2979 | .sr(1) |
| 2980 | .m(4) |
| 2981 | .n(4) |
| 2982 | .k(1) |
| 2983 | .qmax(128) |
| 2984 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2985 | } |
| 2986 | |
| 2987 | TEST(F32_PPMM_4X4__SCALAR, strided_cm) { |
| 2988 | GemmMicrokernelTester() |
| 2989 | .mr(4) |
| 2990 | .nr(4) |
| 2991 | .kr(1) |
| 2992 | .sr(1) |
| 2993 | .m(4) |
| 2994 | .n(4) |
| 2995 | .k(1) |
| 2996 | .cm_stride(7) |
| 2997 | .Test(xnn_f32_ppmm_ukernel_4x4__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2998 | } |
| 2999 | |
| 3000 | |
| 3001 | TEST(F32_PPMM_3X3__SCALAR, k_eq_1) { |
| 3002 | GemmMicrokernelTester() |
| 3003 | .mr(3) |
| 3004 | .nr(3) |
| 3005 | .kr(1) |
| 3006 | .sr(1) |
| 3007 | .m(3) |
| 3008 | .n(3) |
| 3009 | .k(1) |
| 3010 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3011 | } |
| 3012 | |
| 3013 | TEST(F32_PPMM_3X3__SCALAR, strided_cn) { |
| 3014 | GemmMicrokernelTester() |
| 3015 | .mr(3) |
| 3016 | .nr(3) |
| 3017 | .kr(1) |
| 3018 | .sr(1) |
| 3019 | .m(3) |
| 3020 | .n(3) |
| 3021 | .k(1) |
| 3022 | .cn_stride(5) |
| 3023 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3024 | } |
| 3025 | |
| 3026 | TEST(F32_PPMM_3X3__SCALAR, k_eq_1_strided_a) { |
| 3027 | GemmMicrokernelTester() |
| 3028 | .mr(3) |
| 3029 | .nr(3) |
| 3030 | .kr(1) |
| 3031 | .sr(1) |
| 3032 | .m(3) |
| 3033 | .n(3) |
| 3034 | .k(1) |
| 3035 | .a_stride(3) |
| 3036 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3037 | } |
| 3038 | |
| 3039 | TEST(F32_PPMM_3X3__SCALAR, k_eq_1_subtile) { |
| 3040 | for (uint32_t m = 1; m <= 3; m++) { |
| 3041 | for (uint32_t n = 1; n <= 3; n++) { |
| 3042 | GemmMicrokernelTester() |
| 3043 | .mr(3) |
| 3044 | .nr(3) |
| 3045 | .kr(1) |
| 3046 | .sr(1) |
| 3047 | .m(m) |
| 3048 | .n(n) |
| 3049 | .k(1) |
| 3050 | .iterations(1) |
| 3051 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3052 | } |
| 3053 | } |
| 3054 | } |
| 3055 | |
| 3056 | TEST(F32_PPMM_3X3__SCALAR, k_eq_1_subtile_m) { |
| 3057 | for (uint32_t m = 1; m <= 3; m++) { |
| 3058 | GemmMicrokernelTester() |
| 3059 | .mr(3) |
| 3060 | .nr(3) |
| 3061 | .kr(1) |
| 3062 | .sr(1) |
| 3063 | .m(m) |
| 3064 | .n(3) |
| 3065 | .k(1) |
| 3066 | .iterations(1) |
| 3067 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3068 | } |
| 3069 | } |
| 3070 | |
| 3071 | TEST(F32_PPMM_3X3__SCALAR, k_eq_1_subtile_n) { |
| 3072 | for (uint32_t n = 1; n <= 3; n++) { |
| 3073 | GemmMicrokernelTester() |
| 3074 | .mr(3) |
| 3075 | .nr(3) |
| 3076 | .kr(1) |
| 3077 | .sr(1) |
| 3078 | .m(3) |
| 3079 | .n(n) |
| 3080 | .k(1) |
| 3081 | .iterations(1) |
| 3082 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3083 | } |
| 3084 | } |
| 3085 | |
| 3086 | TEST(F32_PPMM_3X3__SCALAR, k_gt_1) { |
| 3087 | for (size_t k = 2; k < 10; k++) { |
| 3088 | GemmMicrokernelTester() |
| 3089 | .mr(3) |
| 3090 | .nr(3) |
| 3091 | .kr(1) |
| 3092 | .sr(1) |
| 3093 | .m(3) |
| 3094 | .n(3) |
| 3095 | .k(k) |
| 3096 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3097 | } |
| 3098 | } |
| 3099 | |
| 3100 | TEST(F32_PPMM_3X3__SCALAR, k_gt_1_subtile) { |
| 3101 | for (size_t k = 2; k < 10; k++) { |
| 3102 | for (uint32_t m = 1; m <= 3; m++) { |
| 3103 | for (uint32_t n = 1; n <= 3; n++) { |
| 3104 | GemmMicrokernelTester() |
| 3105 | .mr(3) |
| 3106 | .nr(3) |
| 3107 | .kr(1) |
| 3108 | .sr(1) |
| 3109 | .m(m) |
| 3110 | .n(n) |
| 3111 | .k(k) |
| 3112 | .iterations(1) |
| 3113 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3114 | } |
| 3115 | } |
| 3116 | } |
| 3117 | } |
| 3118 | |
| 3119 | TEST(F32_PPMM_3X3__SCALAR, n_gt_3) { |
| 3120 | for (uint32_t n = 4; n < 6; n++) { |
| 3121 | for (size_t k = 1; k <= 5; k += 2) { |
| 3122 | GemmMicrokernelTester() |
| 3123 | .mr(3) |
| 3124 | .nr(3) |
| 3125 | .kr(1) |
| 3126 | .sr(1) |
| 3127 | .m(3) |
| 3128 | .n(3) |
| 3129 | .k(k) |
| 3130 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3131 | } |
| 3132 | } |
| 3133 | } |
| 3134 | |
| 3135 | TEST(F32_PPMM_3X3__SCALAR, n_gt_3_strided_cn) { |
| 3136 | for (uint32_t n = 4; n < 6; n++) { |
| 3137 | for (size_t k = 1; k <= 5; k += 2) { |
| 3138 | GemmMicrokernelTester() |
| 3139 | .mr(3) |
| 3140 | .nr(3) |
| 3141 | .kr(1) |
| 3142 | .sr(1) |
| 3143 | .m(3) |
| 3144 | .n(3) |
| 3145 | .k(k) |
| 3146 | .cn_stride(5) |
| 3147 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3148 | } |
| 3149 | } |
| 3150 | } |
| 3151 | |
| 3152 | TEST(F32_PPMM_3X3__SCALAR, n_gt_3_strided_a) { |
| 3153 | for (uint32_t n = 4; n < 6; n++) { |
| 3154 | for (size_t k = 1; k <= 5; k += 2) { |
| 3155 | GemmMicrokernelTester() |
| 3156 | .mr(3) |
| 3157 | .nr(3) |
| 3158 | .kr(1) |
| 3159 | .sr(1) |
| 3160 | .m(3) |
| 3161 | .n(n) |
| 3162 | .k(k) |
| 3163 | .a_stride(7) |
| 3164 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3165 | } |
| 3166 | } |
| 3167 | } |
| 3168 | |
| 3169 | TEST(F32_PPMM_3X3__SCALAR, n_gt_3_subtile) { |
| 3170 | for (uint32_t n = 4; n < 6; n++) { |
| 3171 | for (size_t k = 1; k <= 5; k += 2) { |
| 3172 | for (uint32_t m = 1; m <= 3; m++) { |
| 3173 | GemmMicrokernelTester() |
| 3174 | .mr(3) |
| 3175 | .nr(3) |
| 3176 | .kr(1) |
| 3177 | .sr(1) |
| 3178 | .m(m) |
| 3179 | .n(n) |
| 3180 | .k(k) |
| 3181 | .iterations(1) |
| 3182 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3183 | } |
| 3184 | } |
| 3185 | } |
| 3186 | } |
| 3187 | |
| 3188 | TEST(F32_PPMM_3X3__SCALAR, n_div_3) { |
| 3189 | for (uint32_t n = 6; n <= 9; n += 3) { |
| 3190 | for (size_t k = 1; k <= 5; k += 2) { |
| 3191 | GemmMicrokernelTester() |
| 3192 | .mr(3) |
| 3193 | .nr(3) |
| 3194 | .kr(1) |
| 3195 | .sr(1) |
| 3196 | .m(3) |
| 3197 | .n(3) |
| 3198 | .k(k) |
| 3199 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3200 | } |
| 3201 | } |
| 3202 | } |
| 3203 | |
| 3204 | TEST(F32_PPMM_3X3__SCALAR, n_div_3_strided_cn) { |
| 3205 | for (uint32_t n = 6; n <= 9; n += 3) { |
| 3206 | for (size_t k = 1; k <= 5; k += 2) { |
| 3207 | GemmMicrokernelTester() |
| 3208 | .mr(3) |
| 3209 | .nr(3) |
| 3210 | .kr(1) |
| 3211 | .sr(1) |
| 3212 | .m(3) |
| 3213 | .n(n) |
| 3214 | .k(k) |
| 3215 | .cn_stride(5) |
| 3216 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3217 | } |
| 3218 | } |
| 3219 | } |
| 3220 | |
| 3221 | TEST(F32_PPMM_3X3__SCALAR, n_div_3_strided_a) { |
| 3222 | for (uint32_t n = 6; n <= 9; n += 3) { |
| 3223 | for (size_t k = 1; k <= 5; k += 2) { |
| 3224 | GemmMicrokernelTester() |
| 3225 | .mr(3) |
| 3226 | .nr(3) |
| 3227 | .kr(1) |
| 3228 | .sr(1) |
| 3229 | .m(3) |
| 3230 | .n(n) |
| 3231 | .k(k) |
| 3232 | .a_stride(7) |
| 3233 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3234 | } |
| 3235 | } |
| 3236 | } |
| 3237 | |
| 3238 | TEST(F32_PPMM_3X3__SCALAR, n_div_3_subtile) { |
| 3239 | for (uint32_t n = 6; n <= 9; n += 3) { |
| 3240 | for (size_t k = 1; k <= 5; k += 2) { |
| 3241 | for (uint32_t m = 1; m <= 3; m++) { |
| 3242 | GemmMicrokernelTester() |
| 3243 | .mr(3) |
| 3244 | .nr(3) |
| 3245 | .kr(1) |
| 3246 | .sr(1) |
| 3247 | .m(m) |
| 3248 | .n(n) |
| 3249 | .k(k) |
| 3250 | .iterations(1) |
| 3251 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3252 | } |
| 3253 | } |
| 3254 | } |
| 3255 | } |
| 3256 | |
| 3257 | TEST(F32_PPMM_3X3__SCALAR, strided_cm_subtile) { |
| 3258 | for (size_t k = 1; k <= 5; k += 2) { |
| 3259 | for (uint32_t m = 1; m <= 3; m++) { |
| 3260 | for (uint32_t n = 1; n <= 3; n++) { |
| 3261 | GemmMicrokernelTester() |
| 3262 | .mr(3) |
| 3263 | .nr(3) |
| 3264 | .kr(1) |
| 3265 | .sr(1) |
| 3266 | .m(m) |
| 3267 | .n(n) |
| 3268 | .k(k) |
| 3269 | .cm_stride(5) |
| 3270 | .iterations(1) |
| 3271 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3272 | } |
| 3273 | } |
| 3274 | } |
| 3275 | } |
| 3276 | |
| 3277 | TEST(F32_PPMM_3X3__SCALAR, qmin) { |
| 3278 | GemmMicrokernelTester() |
| 3279 | .mr(3) |
| 3280 | .nr(3) |
| 3281 | .kr(1) |
| 3282 | .sr(1) |
| 3283 | .m(3) |
| 3284 | .n(3) |
| 3285 | .k(1) |
| 3286 | .qmin(128) |
| 3287 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3288 | } |
| 3289 | |
| 3290 | TEST(F32_PPMM_3X3__SCALAR, qmax) { |
| 3291 | GemmMicrokernelTester() |
| 3292 | .mr(3) |
| 3293 | .nr(3) |
| 3294 | .kr(1) |
| 3295 | .sr(1) |
| 3296 | .m(3) |
| 3297 | .n(3) |
| 3298 | .k(1) |
| 3299 | .qmax(128) |
| 3300 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3301 | } |
| 3302 | |
| 3303 | TEST(F32_PPMM_3X3__SCALAR, strided_cm) { |
| 3304 | GemmMicrokernelTester() |
| 3305 | .mr(3) |
| 3306 | .nr(3) |
| 3307 | .kr(1) |
| 3308 | .sr(1) |
| 3309 | .m(3) |
| 3310 | .n(3) |
| 3311 | .k(1) |
| 3312 | .cm_stride(5) |
| 3313 | .Test(xnn_f32_ppmm_ukernel_3x3__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 3314 | } |