| 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/q8-gemm.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(Q8_GEMM_4X8__NEON, k_eq_8) { |
| 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(8) |
| 36 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 37 | } |
| 38 | |
| 39 | TEST(Q8_GEMM_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(8) |
| 49 | .cn_stride(11) |
| 50 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 51 | } |
| 52 | |
| 53 | TEST(Q8_GEMM_4X8__NEON, k_eq_8_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(8) |
| 63 | .a_stride(11) |
| 64 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 65 | } |
| 66 | |
| 67 | TEST(Q8_GEMM_4X8__NEON, k_eq_8_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(8) |
| 79 | .iterations(1) |
| 80 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 81 | } |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | TEST(Q8_GEMM_4X8__NEON, k_eq_8_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(8) |
| 96 | .iterations(1) |
| 97 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | TEST(Q8_GEMM_4X8__NEON, k_eq_8_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(8) |
| 112 | .iterations(1) |
| 113 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | TEST(Q8_GEMM_4X8__NEON, k_lt_8) { |
| 118 | TEST_REQUIRES_ARM_NEON; |
| 119 | for (size_t k = 1; k < 8; 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_q8_gemm_ukernel_4x8__neon); |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | TEST(Q8_GEMM_4X8__NEON, k_lt_8_strided_a) { |
| 133 | TEST_REQUIRES_ARM_NEON; |
| 134 | for (size_t k = 1; k < 8; k++) { |
| 135 | GemmMicrokernelTester() |
| 136 | .mr(4) |
| 137 | .nr(8) |
| 138 | .kr(1) |
| 139 | .sr(1) |
| 140 | .m(4) |
| 141 | .n(8) |
| 142 | .k(k) |
| 143 | .a_stride(11) |
| 144 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 145 | } |
| 146 | } |
| 147 | |
| 148 | TEST(Q8_GEMM_4X8__NEON, k_lt_8_subtile) { |
| 149 | TEST_REQUIRES_ARM_NEON; |
| 150 | for (size_t k = 1; k < 8; k++) { |
| 151 | for (uint32_t m = 1; m <= 4; m++) { |
| 152 | for (uint32_t n = 1; n <= 8; n++) { |
| 153 | GemmMicrokernelTester() |
| 154 | .mr(4) |
| 155 | .nr(8) |
| 156 | .kr(1) |
| 157 | .sr(1) |
| 158 | .m(m) |
| 159 | .n(n) |
| 160 | .k(k) |
| 161 | .iterations(1) |
| 162 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 163 | } |
| 164 | } |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | TEST(Q8_GEMM_4X8__NEON, k_gt_8) { |
| 169 | TEST_REQUIRES_ARM_NEON; |
| 170 | for (size_t k = 9; k < 16; k++) { |
| 171 | GemmMicrokernelTester() |
| 172 | .mr(4) |
| 173 | .nr(8) |
| 174 | .kr(1) |
| 175 | .sr(1) |
| 176 | .m(4) |
| 177 | .n(8) |
| 178 | .k(k) |
| 179 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | TEST(Q8_GEMM_4X8__NEON, k_gt_8_strided_a) { |
| 184 | TEST_REQUIRES_ARM_NEON; |
| 185 | for (size_t k = 9; k < 16; k++) { |
| 186 | GemmMicrokernelTester() |
| 187 | .mr(4) |
| 188 | .nr(8) |
| 189 | .kr(1) |
| 190 | .sr(1) |
| 191 | .m(4) |
| 192 | .n(8) |
| 193 | .k(k) |
| 194 | .a_stride(19) |
| 195 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | TEST(Q8_GEMM_4X8__NEON, k_gt_8_subtile) { |
| 200 | TEST_REQUIRES_ARM_NEON; |
| 201 | for (size_t k = 9; k < 16; k++) { |
| 202 | for (uint32_t m = 1; m <= 4; m++) { |
| 203 | for (uint32_t n = 1; n <= 8; n++) { |
| 204 | GemmMicrokernelTester() |
| 205 | .mr(4) |
| 206 | .nr(8) |
| 207 | .kr(1) |
| 208 | .sr(1) |
| 209 | .m(m) |
| 210 | .n(n) |
| 211 | .k(k) |
| 212 | .iterations(1) |
| 213 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 214 | } |
| 215 | } |
| 216 | } |
| 217 | } |
| 218 | |
| 219 | TEST(Q8_GEMM_4X8__NEON, k_div_8) { |
| 220 | TEST_REQUIRES_ARM_NEON; |
| 221 | for (size_t k = 16; k <= 80; k += 8) { |
| 222 | GemmMicrokernelTester() |
| 223 | .mr(4) |
| 224 | .nr(8) |
| 225 | .kr(1) |
| 226 | .sr(1) |
| 227 | .m(4) |
| 228 | .n(8) |
| 229 | .k(k) |
| 230 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | TEST(Q8_GEMM_4X8__NEON, k_div_8_strided_a) { |
| 235 | TEST_REQUIRES_ARM_NEON; |
| 236 | for (size_t k = 16; k <= 80; k += 8) { |
| 237 | GemmMicrokernelTester() |
| 238 | .mr(4) |
| 239 | .nr(8) |
| 240 | .kr(1) |
| 241 | .sr(1) |
| 242 | .m(4) |
| 243 | .n(8) |
| 244 | .k(k) |
| 245 | .a_stride(83) |
| 246 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 247 | } |
| 248 | } |
| 249 | |
| 250 | TEST(Q8_GEMM_4X8__NEON, k_div_8_subtile) { |
| 251 | TEST_REQUIRES_ARM_NEON; |
| 252 | for (size_t k = 16; k <= 80; k += 8) { |
| 253 | for (uint32_t m = 1; m <= 4; m++) { |
| 254 | for (uint32_t n = 1; n <= 8; n++) { |
| 255 | GemmMicrokernelTester() |
| 256 | .mr(4) |
| 257 | .nr(8) |
| 258 | .kr(1) |
| 259 | .sr(1) |
| 260 | .m(m) |
| 261 | .n(n) |
| 262 | .k(k) |
| 263 | .iterations(1) |
| 264 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 265 | } |
| 266 | } |
| 267 | } |
| 268 | } |
| 269 | |
| 270 | TEST(Q8_GEMM_4X8__NEON, n_gt_8) { |
| 271 | TEST_REQUIRES_ARM_NEON; |
| 272 | for (uint32_t n = 9; n < 16; n++) { |
| 273 | for (size_t k = 1; k <= 40; k += 9) { |
| 274 | GemmMicrokernelTester() |
| 275 | .mr(4) |
| 276 | .nr(8) |
| 277 | .kr(1) |
| 278 | .sr(1) |
| 279 | .m(4) |
| 280 | .n(8) |
| 281 | .k(k) |
| 282 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 283 | } |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | TEST(Q8_GEMM_4X8__NEON, n_gt_8_strided_cn) { |
| 288 | TEST_REQUIRES_ARM_NEON; |
| 289 | for (uint32_t n = 9; n < 16; n++) { |
| 290 | for (size_t k = 1; k <= 40; k += 9) { |
| 291 | GemmMicrokernelTester() |
| 292 | .mr(4) |
| 293 | .nr(8) |
| 294 | .kr(1) |
| 295 | .sr(1) |
| 296 | .m(4) |
| 297 | .n(8) |
| 298 | .k(k) |
| 299 | .cn_stride(11) |
| 300 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 301 | } |
| 302 | } |
| 303 | } |
| 304 | |
| 305 | TEST(Q8_GEMM_4X8__NEON, n_gt_8_strided_a) { |
| 306 | TEST_REQUIRES_ARM_NEON; |
| 307 | for (uint32_t n = 9; n < 16; n++) { |
| 308 | for (size_t k = 1; k <= 40; k += 9) { |
| 309 | GemmMicrokernelTester() |
| 310 | .mr(4) |
| 311 | .nr(8) |
| 312 | .kr(1) |
| 313 | .sr(1) |
| 314 | .m(4) |
| 315 | .n(n) |
| 316 | .k(k) |
| 317 | .a_stride(43) |
| 318 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 319 | } |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | TEST(Q8_GEMM_4X8__NEON, n_gt_8_subtile) { |
| 324 | TEST_REQUIRES_ARM_NEON; |
| 325 | for (uint32_t n = 9; n < 16; n++) { |
| 326 | for (size_t k = 1; k <= 40; k += 9) { |
| 327 | for (uint32_t m = 1; m <= 4; m++) { |
| 328 | GemmMicrokernelTester() |
| 329 | .mr(4) |
| 330 | .nr(8) |
| 331 | .kr(1) |
| 332 | .sr(1) |
| 333 | .m(m) |
| 334 | .n(n) |
| 335 | .k(k) |
| 336 | .iterations(1) |
| 337 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 338 | } |
| 339 | } |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | TEST(Q8_GEMM_4X8__NEON, n_div_8) { |
| 344 | TEST_REQUIRES_ARM_NEON; |
| 345 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 346 | for (size_t k = 1; k <= 40; k += 9) { |
| 347 | GemmMicrokernelTester() |
| 348 | .mr(4) |
| 349 | .nr(8) |
| 350 | .kr(1) |
| 351 | .sr(1) |
| 352 | .m(4) |
| 353 | .n(8) |
| 354 | .k(k) |
| 355 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 356 | } |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | TEST(Q8_GEMM_4X8__NEON, n_div_8_strided_cn) { |
| 361 | TEST_REQUIRES_ARM_NEON; |
| 362 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 363 | for (size_t k = 1; k <= 40; k += 9) { |
| 364 | GemmMicrokernelTester() |
| 365 | .mr(4) |
| 366 | .nr(8) |
| 367 | .kr(1) |
| 368 | .sr(1) |
| 369 | .m(4) |
| 370 | .n(n) |
| 371 | .k(k) |
| 372 | .cn_stride(11) |
| 373 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 374 | } |
| 375 | } |
| 376 | } |
| 377 | |
| 378 | TEST(Q8_GEMM_4X8__NEON, n_div_8_strided_a) { |
| 379 | TEST_REQUIRES_ARM_NEON; |
| 380 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 381 | for (size_t k = 1; k <= 40; k += 9) { |
| 382 | GemmMicrokernelTester() |
| 383 | .mr(4) |
| 384 | .nr(8) |
| 385 | .kr(1) |
| 386 | .sr(1) |
| 387 | .m(4) |
| 388 | .n(n) |
| 389 | .k(k) |
| 390 | .a_stride(43) |
| 391 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 392 | } |
| 393 | } |
| 394 | } |
| 395 | |
| 396 | TEST(Q8_GEMM_4X8__NEON, n_div_8_subtile) { |
| 397 | TEST_REQUIRES_ARM_NEON; |
| 398 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 399 | for (size_t k = 1; k <= 40; k += 9) { |
| 400 | for (uint32_t m = 1; m <= 4; m++) { |
| 401 | GemmMicrokernelTester() |
| 402 | .mr(4) |
| 403 | .nr(8) |
| 404 | .kr(1) |
| 405 | .sr(1) |
| 406 | .m(m) |
| 407 | .n(n) |
| 408 | .k(k) |
| 409 | .iterations(1) |
| 410 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 411 | } |
| 412 | } |
| 413 | } |
| 414 | } |
| 415 | |
| 416 | TEST(Q8_GEMM_4X8__NEON, strided_cm_subtile) { |
| 417 | TEST_REQUIRES_ARM_NEON; |
| 418 | for (size_t k = 1; k <= 40; k += 9) { |
| 419 | for (uint32_t m = 1; m <= 4; m++) { |
| 420 | for (uint32_t n = 1; n <= 8; n++) { |
| 421 | GemmMicrokernelTester() |
| 422 | .mr(4) |
| 423 | .nr(8) |
| 424 | .kr(1) |
| 425 | .sr(1) |
| 426 | .m(m) |
| 427 | .n(n) |
| 428 | .k(k) |
| 429 | .cm_stride(11) |
| 430 | .iterations(1) |
| 431 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 432 | } |
| 433 | } |
| 434 | } |
| 435 | } |
| 436 | |
| 437 | TEST(Q8_GEMM_4X8__NEON, qmin) { |
| 438 | TEST_REQUIRES_ARM_NEON; |
| 439 | GemmMicrokernelTester() |
| 440 | .mr(4) |
| 441 | .nr(8) |
| 442 | .kr(1) |
| 443 | .sr(1) |
| 444 | .m(4) |
| 445 | .n(8) |
| 446 | .k(8) |
| 447 | .qmin(128) |
| 448 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 449 | } |
| 450 | |
| 451 | TEST(Q8_GEMM_4X8__NEON, qmax) { |
| 452 | TEST_REQUIRES_ARM_NEON; |
| 453 | GemmMicrokernelTester() |
| 454 | .mr(4) |
| 455 | .nr(8) |
| 456 | .kr(1) |
| 457 | .sr(1) |
| 458 | .m(4) |
| 459 | .n(8) |
| 460 | .k(8) |
| 461 | .qmax(128) |
| 462 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 463 | } |
| 464 | |
| 465 | TEST(Q8_GEMM_4X8__NEON, strided_cm) { |
| 466 | TEST_REQUIRES_ARM_NEON; |
| 467 | GemmMicrokernelTester() |
| 468 | .mr(4) |
| 469 | .nr(8) |
| 470 | .kr(1) |
| 471 | .sr(1) |
| 472 | .m(4) |
| 473 | .n(8) |
| 474 | .k(8) |
| 475 | .cm_stride(11) |
| 476 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 477 | } |
| 478 | |
| 479 | TEST(Q8_GEMM_4X8__NEON, no_a_zero_point) { |
| 480 | TEST_REQUIRES_ARM_NEON; |
| 481 | for (size_t k = 1; k <= 40; k += 9) { |
| 482 | GemmMicrokernelTester() |
| 483 | .mr(4) |
| 484 | .nr(8) |
| 485 | .kr(1) |
| 486 | .sr(1) |
| 487 | .m(4) |
| 488 | .n(8) |
| 489 | .k(k) |
| 490 | .a_zero_point(0) |
| 491 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 492 | } |
| 493 | } |
| 494 | |
| 495 | TEST(Q8_GEMM_4X8__NEON, no_b_zero_point) { |
| 496 | TEST_REQUIRES_ARM_NEON; |
| 497 | for (size_t k = 1; k <= 40; k += 9) { |
| 498 | GemmMicrokernelTester() |
| 499 | .mr(4) |
| 500 | .nr(8) |
| 501 | .kr(1) |
| 502 | .sr(1) |
| 503 | .m(4) |
| 504 | .n(8) |
| 505 | .k(k) |
| 506 | .b_zero_point(0) |
| 507 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 508 | } |
| 509 | } |
| 510 | |
| 511 | TEST(Q8_GEMM_4X8__NEON, no_zero_point) { |
| 512 | TEST_REQUIRES_ARM_NEON; |
| 513 | for (size_t k = 1; k <= 40; k += 9) { |
| 514 | GemmMicrokernelTester() |
| 515 | .mr(4) |
| 516 | .nr(8) |
| 517 | .kr(1) |
| 518 | .sr(1) |
| 519 | .m(4) |
| 520 | .n(8) |
| 521 | .k(k) |
| 522 | .a_zero_point(0) |
| 523 | .b_zero_point(0) |
| 524 | .Test(xnn_q8_gemm_ukernel_4x8__neon); |
| 525 | } |
| 526 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 527 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 528 | |
| 529 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 530 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 531 | TEST(Q8_GEMM_8X8__NEON, k_eq_8) { |
| 532 | TEST_REQUIRES_ARM_NEON; |
| 533 | GemmMicrokernelTester() |
| 534 | .mr(8) |
| 535 | .nr(8) |
| 536 | .kr(1) |
| 537 | .sr(1) |
| 538 | .m(8) |
| 539 | .n(8) |
| 540 | .k(8) |
| 541 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 542 | } |
| 543 | |
| 544 | TEST(Q8_GEMM_8X8__NEON, strided_cn) { |
| 545 | TEST_REQUIRES_ARM_NEON; |
| 546 | GemmMicrokernelTester() |
| 547 | .mr(8) |
| 548 | .nr(8) |
| 549 | .kr(1) |
| 550 | .sr(1) |
| 551 | .m(8) |
| 552 | .n(8) |
| 553 | .k(8) |
| 554 | .cn_stride(11) |
| 555 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 556 | } |
| 557 | |
| 558 | TEST(Q8_GEMM_8X8__NEON, k_eq_8_strided_a) { |
| 559 | TEST_REQUIRES_ARM_NEON; |
| 560 | GemmMicrokernelTester() |
| 561 | .mr(8) |
| 562 | .nr(8) |
| 563 | .kr(1) |
| 564 | .sr(1) |
| 565 | .m(8) |
| 566 | .n(8) |
| 567 | .k(8) |
| 568 | .a_stride(11) |
| 569 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 570 | } |
| 571 | |
| 572 | TEST(Q8_GEMM_8X8__NEON, k_eq_8_subtile) { |
| 573 | TEST_REQUIRES_ARM_NEON; |
| 574 | for (uint32_t m = 1; m <= 8; m++) { |
| 575 | for (uint32_t n = 1; n <= 8; n++) { |
| 576 | GemmMicrokernelTester() |
| 577 | .mr(8) |
| 578 | .nr(8) |
| 579 | .kr(1) |
| 580 | .sr(1) |
| 581 | .m(m) |
| 582 | .n(n) |
| 583 | .k(8) |
| 584 | .iterations(1) |
| 585 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 586 | } |
| 587 | } |
| 588 | } |
| 589 | |
| 590 | TEST(Q8_GEMM_8X8__NEON, k_eq_8_subtile_m) { |
| 591 | TEST_REQUIRES_ARM_NEON; |
| 592 | for (uint32_t m = 1; m <= 8; m++) { |
| 593 | GemmMicrokernelTester() |
| 594 | .mr(8) |
| 595 | .nr(8) |
| 596 | .kr(1) |
| 597 | .sr(1) |
| 598 | .m(m) |
| 599 | .n(8) |
| 600 | .k(8) |
| 601 | .iterations(1) |
| 602 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | TEST(Q8_GEMM_8X8__NEON, k_eq_8_subtile_n) { |
| 607 | TEST_REQUIRES_ARM_NEON; |
| 608 | for (uint32_t n = 1; n <= 8; n++) { |
| 609 | GemmMicrokernelTester() |
| 610 | .mr(8) |
| 611 | .nr(8) |
| 612 | .kr(1) |
| 613 | .sr(1) |
| 614 | .m(8) |
| 615 | .n(n) |
| 616 | .k(8) |
| 617 | .iterations(1) |
| 618 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 619 | } |
| 620 | } |
| 621 | |
| 622 | TEST(Q8_GEMM_8X8__NEON, k_lt_8) { |
| 623 | TEST_REQUIRES_ARM_NEON; |
| 624 | for (size_t k = 1; k < 8; k++) { |
| 625 | GemmMicrokernelTester() |
| 626 | .mr(8) |
| 627 | .nr(8) |
| 628 | .kr(1) |
| 629 | .sr(1) |
| 630 | .m(8) |
| 631 | .n(8) |
| 632 | .k(k) |
| 633 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 634 | } |
| 635 | } |
| 636 | |
| 637 | TEST(Q8_GEMM_8X8__NEON, k_lt_8_strided_a) { |
| 638 | TEST_REQUIRES_ARM_NEON; |
| 639 | for (size_t k = 1; k < 8; k++) { |
| 640 | GemmMicrokernelTester() |
| 641 | .mr(8) |
| 642 | .nr(8) |
| 643 | .kr(1) |
| 644 | .sr(1) |
| 645 | .m(8) |
| 646 | .n(8) |
| 647 | .k(k) |
| 648 | .a_stride(11) |
| 649 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 650 | } |
| 651 | } |
| 652 | |
| 653 | TEST(Q8_GEMM_8X8__NEON, k_lt_8_subtile) { |
| 654 | TEST_REQUIRES_ARM_NEON; |
| 655 | for (size_t k = 1; k < 8; k++) { |
| 656 | for (uint32_t m = 1; m <= 8; m++) { |
| 657 | for (uint32_t n = 1; n <= 8; n++) { |
| 658 | GemmMicrokernelTester() |
| 659 | .mr(8) |
| 660 | .nr(8) |
| 661 | .kr(1) |
| 662 | .sr(1) |
| 663 | .m(m) |
| 664 | .n(n) |
| 665 | .k(k) |
| 666 | .iterations(1) |
| 667 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 668 | } |
| 669 | } |
| 670 | } |
| 671 | } |
| 672 | |
| 673 | TEST(Q8_GEMM_8X8__NEON, k_gt_8) { |
| 674 | TEST_REQUIRES_ARM_NEON; |
| 675 | for (size_t k = 9; k < 16; k++) { |
| 676 | GemmMicrokernelTester() |
| 677 | .mr(8) |
| 678 | .nr(8) |
| 679 | .kr(1) |
| 680 | .sr(1) |
| 681 | .m(8) |
| 682 | .n(8) |
| 683 | .k(k) |
| 684 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 685 | } |
| 686 | } |
| 687 | |
| 688 | TEST(Q8_GEMM_8X8__NEON, k_gt_8_strided_a) { |
| 689 | TEST_REQUIRES_ARM_NEON; |
| 690 | for (size_t k = 9; k < 16; k++) { |
| 691 | GemmMicrokernelTester() |
| 692 | .mr(8) |
| 693 | .nr(8) |
| 694 | .kr(1) |
| 695 | .sr(1) |
| 696 | .m(8) |
| 697 | .n(8) |
| 698 | .k(k) |
| 699 | .a_stride(19) |
| 700 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 701 | } |
| 702 | } |
| 703 | |
| 704 | TEST(Q8_GEMM_8X8__NEON, k_gt_8_subtile) { |
| 705 | TEST_REQUIRES_ARM_NEON; |
| 706 | for (size_t k = 9; k < 16; k++) { |
| 707 | for (uint32_t m = 1; m <= 8; m++) { |
| 708 | for (uint32_t n = 1; n <= 8; n++) { |
| 709 | GemmMicrokernelTester() |
| 710 | .mr(8) |
| 711 | .nr(8) |
| 712 | .kr(1) |
| 713 | .sr(1) |
| 714 | .m(m) |
| 715 | .n(n) |
| 716 | .k(k) |
| 717 | .iterations(1) |
| 718 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 719 | } |
| 720 | } |
| 721 | } |
| 722 | } |
| 723 | |
| 724 | TEST(Q8_GEMM_8X8__NEON, k_div_8) { |
| 725 | TEST_REQUIRES_ARM_NEON; |
| 726 | for (size_t k = 16; k <= 80; k += 8) { |
| 727 | GemmMicrokernelTester() |
| 728 | .mr(8) |
| 729 | .nr(8) |
| 730 | .kr(1) |
| 731 | .sr(1) |
| 732 | .m(8) |
| 733 | .n(8) |
| 734 | .k(k) |
| 735 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 736 | } |
| 737 | } |
| 738 | |
| 739 | TEST(Q8_GEMM_8X8__NEON, k_div_8_strided_a) { |
| 740 | TEST_REQUIRES_ARM_NEON; |
| 741 | for (size_t k = 16; k <= 80; k += 8) { |
| 742 | GemmMicrokernelTester() |
| 743 | .mr(8) |
| 744 | .nr(8) |
| 745 | .kr(1) |
| 746 | .sr(1) |
| 747 | .m(8) |
| 748 | .n(8) |
| 749 | .k(k) |
| 750 | .a_stride(83) |
| 751 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 752 | } |
| 753 | } |
| 754 | |
| 755 | TEST(Q8_GEMM_8X8__NEON, k_div_8_subtile) { |
| 756 | TEST_REQUIRES_ARM_NEON; |
| 757 | for (size_t k = 16; k <= 80; k += 8) { |
| 758 | for (uint32_t m = 1; m <= 8; m++) { |
| 759 | for (uint32_t n = 1; n <= 8; n++) { |
| 760 | GemmMicrokernelTester() |
| 761 | .mr(8) |
| 762 | .nr(8) |
| 763 | .kr(1) |
| 764 | .sr(1) |
| 765 | .m(m) |
| 766 | .n(n) |
| 767 | .k(k) |
| 768 | .iterations(1) |
| 769 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 770 | } |
| 771 | } |
| 772 | } |
| 773 | } |
| 774 | |
| 775 | TEST(Q8_GEMM_8X8__NEON, n_gt_8) { |
| 776 | TEST_REQUIRES_ARM_NEON; |
| 777 | for (uint32_t n = 9; n < 16; n++) { |
| 778 | for (size_t k = 1; k <= 40; k += 9) { |
| 779 | GemmMicrokernelTester() |
| 780 | .mr(8) |
| 781 | .nr(8) |
| 782 | .kr(1) |
| 783 | .sr(1) |
| 784 | .m(8) |
| 785 | .n(8) |
| 786 | .k(k) |
| 787 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 788 | } |
| 789 | } |
| 790 | } |
| 791 | |
| 792 | TEST(Q8_GEMM_8X8__NEON, n_gt_8_strided_cn) { |
| 793 | TEST_REQUIRES_ARM_NEON; |
| 794 | for (uint32_t n = 9; n < 16; n++) { |
| 795 | for (size_t k = 1; k <= 40; k += 9) { |
| 796 | GemmMicrokernelTester() |
| 797 | .mr(8) |
| 798 | .nr(8) |
| 799 | .kr(1) |
| 800 | .sr(1) |
| 801 | .m(8) |
| 802 | .n(8) |
| 803 | .k(k) |
| 804 | .cn_stride(11) |
| 805 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 806 | } |
| 807 | } |
| 808 | } |
| 809 | |
| 810 | TEST(Q8_GEMM_8X8__NEON, n_gt_8_strided_a) { |
| 811 | TEST_REQUIRES_ARM_NEON; |
| 812 | for (uint32_t n = 9; n < 16; n++) { |
| 813 | for (size_t k = 1; k <= 40; k += 9) { |
| 814 | GemmMicrokernelTester() |
| 815 | .mr(8) |
| 816 | .nr(8) |
| 817 | .kr(1) |
| 818 | .sr(1) |
| 819 | .m(8) |
| 820 | .n(n) |
| 821 | .k(k) |
| 822 | .a_stride(43) |
| 823 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 824 | } |
| 825 | } |
| 826 | } |
| 827 | |
| 828 | TEST(Q8_GEMM_8X8__NEON, n_gt_8_subtile) { |
| 829 | TEST_REQUIRES_ARM_NEON; |
| 830 | for (uint32_t n = 9; n < 16; n++) { |
| 831 | for (size_t k = 1; k <= 40; k += 9) { |
| 832 | for (uint32_t m = 1; m <= 8; m++) { |
| 833 | GemmMicrokernelTester() |
| 834 | .mr(8) |
| 835 | .nr(8) |
| 836 | .kr(1) |
| 837 | .sr(1) |
| 838 | .m(m) |
| 839 | .n(n) |
| 840 | .k(k) |
| 841 | .iterations(1) |
| 842 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 843 | } |
| 844 | } |
| 845 | } |
| 846 | } |
| 847 | |
| 848 | TEST(Q8_GEMM_8X8__NEON, n_div_8) { |
| 849 | TEST_REQUIRES_ARM_NEON; |
| 850 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 851 | for (size_t k = 1; k <= 40; k += 9) { |
| 852 | GemmMicrokernelTester() |
| 853 | .mr(8) |
| 854 | .nr(8) |
| 855 | .kr(1) |
| 856 | .sr(1) |
| 857 | .m(8) |
| 858 | .n(8) |
| 859 | .k(k) |
| 860 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 861 | } |
| 862 | } |
| 863 | } |
| 864 | |
| 865 | TEST(Q8_GEMM_8X8__NEON, n_div_8_strided_cn) { |
| 866 | TEST_REQUIRES_ARM_NEON; |
| 867 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 868 | for (size_t k = 1; k <= 40; k += 9) { |
| 869 | GemmMicrokernelTester() |
| 870 | .mr(8) |
| 871 | .nr(8) |
| 872 | .kr(1) |
| 873 | .sr(1) |
| 874 | .m(8) |
| 875 | .n(n) |
| 876 | .k(k) |
| 877 | .cn_stride(11) |
| 878 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 879 | } |
| 880 | } |
| 881 | } |
| 882 | |
| 883 | TEST(Q8_GEMM_8X8__NEON, n_div_8_strided_a) { |
| 884 | TEST_REQUIRES_ARM_NEON; |
| 885 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 886 | for (size_t k = 1; k <= 40; k += 9) { |
| 887 | GemmMicrokernelTester() |
| 888 | .mr(8) |
| 889 | .nr(8) |
| 890 | .kr(1) |
| 891 | .sr(1) |
| 892 | .m(8) |
| 893 | .n(n) |
| 894 | .k(k) |
| 895 | .a_stride(43) |
| 896 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 897 | } |
| 898 | } |
| 899 | } |
| 900 | |
| 901 | TEST(Q8_GEMM_8X8__NEON, n_div_8_subtile) { |
| 902 | TEST_REQUIRES_ARM_NEON; |
| 903 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 904 | for (size_t k = 1; k <= 40; k += 9) { |
| 905 | for (uint32_t m = 1; m <= 8; m++) { |
| 906 | GemmMicrokernelTester() |
| 907 | .mr(8) |
| 908 | .nr(8) |
| 909 | .kr(1) |
| 910 | .sr(1) |
| 911 | .m(m) |
| 912 | .n(n) |
| 913 | .k(k) |
| 914 | .iterations(1) |
| 915 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 916 | } |
| 917 | } |
| 918 | } |
| 919 | } |
| 920 | |
| 921 | TEST(Q8_GEMM_8X8__NEON, strided_cm_subtile) { |
| 922 | TEST_REQUIRES_ARM_NEON; |
| 923 | for (size_t k = 1; k <= 40; k += 9) { |
| 924 | for (uint32_t m = 1; m <= 8; m++) { |
| 925 | for (uint32_t n = 1; n <= 8; n++) { |
| 926 | GemmMicrokernelTester() |
| 927 | .mr(8) |
| 928 | .nr(8) |
| 929 | .kr(1) |
| 930 | .sr(1) |
| 931 | .m(m) |
| 932 | .n(n) |
| 933 | .k(k) |
| 934 | .cm_stride(11) |
| 935 | .iterations(1) |
| 936 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 937 | } |
| 938 | } |
| 939 | } |
| 940 | } |
| 941 | |
| 942 | TEST(Q8_GEMM_8X8__NEON, qmin) { |
| 943 | TEST_REQUIRES_ARM_NEON; |
| 944 | GemmMicrokernelTester() |
| 945 | .mr(8) |
| 946 | .nr(8) |
| 947 | .kr(1) |
| 948 | .sr(1) |
| 949 | .m(8) |
| 950 | .n(8) |
| 951 | .k(8) |
| 952 | .qmin(128) |
| 953 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 954 | } |
| 955 | |
| 956 | TEST(Q8_GEMM_8X8__NEON, qmax) { |
| 957 | TEST_REQUIRES_ARM_NEON; |
| 958 | GemmMicrokernelTester() |
| 959 | .mr(8) |
| 960 | .nr(8) |
| 961 | .kr(1) |
| 962 | .sr(1) |
| 963 | .m(8) |
| 964 | .n(8) |
| 965 | .k(8) |
| 966 | .qmax(128) |
| 967 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 968 | } |
| 969 | |
| 970 | TEST(Q8_GEMM_8X8__NEON, strided_cm) { |
| 971 | TEST_REQUIRES_ARM_NEON; |
| 972 | GemmMicrokernelTester() |
| 973 | .mr(8) |
| 974 | .nr(8) |
| 975 | .kr(1) |
| 976 | .sr(1) |
| 977 | .m(8) |
| 978 | .n(8) |
| 979 | .k(8) |
| 980 | .cm_stride(11) |
| 981 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 982 | } |
| 983 | |
| 984 | TEST(Q8_GEMM_8X8__NEON, no_a_zero_point) { |
| 985 | TEST_REQUIRES_ARM_NEON; |
| 986 | for (size_t k = 1; k <= 40; k += 9) { |
| 987 | GemmMicrokernelTester() |
| 988 | .mr(8) |
| 989 | .nr(8) |
| 990 | .kr(1) |
| 991 | .sr(1) |
| 992 | .m(8) |
| 993 | .n(8) |
| 994 | .k(k) |
| 995 | .a_zero_point(0) |
| 996 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 997 | } |
| 998 | } |
| 999 | |
| 1000 | TEST(Q8_GEMM_8X8__NEON, no_b_zero_point) { |
| 1001 | TEST_REQUIRES_ARM_NEON; |
| 1002 | for (size_t k = 1; k <= 40; k += 9) { |
| 1003 | GemmMicrokernelTester() |
| 1004 | .mr(8) |
| 1005 | .nr(8) |
| 1006 | .kr(1) |
| 1007 | .sr(1) |
| 1008 | .m(8) |
| 1009 | .n(8) |
| 1010 | .k(k) |
| 1011 | .b_zero_point(0) |
| 1012 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 1013 | } |
| 1014 | } |
| 1015 | |
| 1016 | TEST(Q8_GEMM_8X8__NEON, no_zero_point) { |
| 1017 | TEST_REQUIRES_ARM_NEON; |
| 1018 | for (size_t k = 1; k <= 40; k += 9) { |
| 1019 | GemmMicrokernelTester() |
| 1020 | .mr(8) |
| 1021 | .nr(8) |
| 1022 | .kr(1) |
| 1023 | .sr(1) |
| 1024 | .m(8) |
| 1025 | .n(8) |
| 1026 | .k(k) |
| 1027 | .a_zero_point(0) |
| 1028 | .b_zero_point(0) |
| 1029 | .Test(xnn_q8_gemm_ukernel_8x8__neon); |
| 1030 | } |
| 1031 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1032 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1033 | |
| 1034 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1035 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1036 | TEST(Q8_GEMM_2X4C8__SSE2, k_eq_8) { |
| 1037 | TEST_REQUIRES_X86_SSE2; |
| 1038 | GemmMicrokernelTester() |
| 1039 | .mr(2) |
| 1040 | .nr(4) |
| 1041 | .kr(8) |
| 1042 | .sr(1) |
| 1043 | .m(2) |
| 1044 | .n(4) |
| 1045 | .k(8) |
| 1046 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1047 | } |
| 1048 | |
| 1049 | TEST(Q8_GEMM_2X4C8__SSE2, strided_cn) { |
| 1050 | TEST_REQUIRES_X86_SSE2; |
| 1051 | GemmMicrokernelTester() |
| 1052 | .mr(2) |
| 1053 | .nr(4) |
| 1054 | .kr(8) |
| 1055 | .sr(1) |
| 1056 | .m(2) |
| 1057 | .n(4) |
| 1058 | .k(8) |
| 1059 | .cn_stride(7) |
| 1060 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1061 | } |
| 1062 | |
| 1063 | TEST(Q8_GEMM_2X4C8__SSE2, k_eq_8_strided_a) { |
| 1064 | TEST_REQUIRES_X86_SSE2; |
| 1065 | GemmMicrokernelTester() |
| 1066 | .mr(2) |
| 1067 | .nr(4) |
| 1068 | .kr(8) |
| 1069 | .sr(1) |
| 1070 | .m(2) |
| 1071 | .n(4) |
| 1072 | .k(8) |
| 1073 | .a_stride(11) |
| 1074 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1075 | } |
| 1076 | |
| 1077 | TEST(Q8_GEMM_2X4C8__SSE2, k_eq_8_subtile) { |
| 1078 | TEST_REQUIRES_X86_SSE2; |
| 1079 | for (uint32_t m = 1; m <= 2; m++) { |
| 1080 | for (uint32_t n = 1; n <= 4; n++) { |
| 1081 | GemmMicrokernelTester() |
| 1082 | .mr(2) |
| 1083 | .nr(4) |
| 1084 | .kr(8) |
| 1085 | .sr(1) |
| 1086 | .m(m) |
| 1087 | .n(n) |
| 1088 | .k(8) |
| 1089 | .iterations(1) |
| 1090 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1091 | } |
| 1092 | } |
| 1093 | } |
| 1094 | |
| 1095 | TEST(Q8_GEMM_2X4C8__SSE2, k_eq_8_subtile_m) { |
| 1096 | TEST_REQUIRES_X86_SSE2; |
| 1097 | for (uint32_t m = 1; m <= 2; m++) { |
| 1098 | GemmMicrokernelTester() |
| 1099 | .mr(2) |
| 1100 | .nr(4) |
| 1101 | .kr(8) |
| 1102 | .sr(1) |
| 1103 | .m(m) |
| 1104 | .n(4) |
| 1105 | .k(8) |
| 1106 | .iterations(1) |
| 1107 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1108 | } |
| 1109 | } |
| 1110 | |
| 1111 | TEST(Q8_GEMM_2X4C8__SSE2, k_eq_8_subtile_n) { |
| 1112 | TEST_REQUIRES_X86_SSE2; |
| 1113 | for (uint32_t n = 1; n <= 4; n++) { |
| 1114 | GemmMicrokernelTester() |
| 1115 | .mr(2) |
| 1116 | .nr(4) |
| 1117 | .kr(8) |
| 1118 | .sr(1) |
| 1119 | .m(2) |
| 1120 | .n(n) |
| 1121 | .k(8) |
| 1122 | .iterations(1) |
| 1123 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1124 | } |
| 1125 | } |
| 1126 | |
| 1127 | TEST(Q8_GEMM_2X4C8__SSE2, k_lt_8) { |
| 1128 | TEST_REQUIRES_X86_SSE2; |
| 1129 | for (size_t k = 1; k < 8; k++) { |
| 1130 | GemmMicrokernelTester() |
| 1131 | .mr(2) |
| 1132 | .nr(4) |
| 1133 | .kr(8) |
| 1134 | .sr(1) |
| 1135 | .m(2) |
| 1136 | .n(4) |
| 1137 | .k(k) |
| 1138 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1139 | } |
| 1140 | } |
| 1141 | |
| 1142 | TEST(Q8_GEMM_2X4C8__SSE2, k_lt_8_strided_a) { |
| 1143 | TEST_REQUIRES_X86_SSE2; |
| 1144 | for (size_t k = 1; k < 8; k++) { |
| 1145 | GemmMicrokernelTester() |
| 1146 | .mr(2) |
| 1147 | .nr(4) |
| 1148 | .kr(8) |
| 1149 | .sr(1) |
| 1150 | .m(2) |
| 1151 | .n(4) |
| 1152 | .k(k) |
| 1153 | .a_stride(11) |
| 1154 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1155 | } |
| 1156 | } |
| 1157 | |
| 1158 | TEST(Q8_GEMM_2X4C8__SSE2, k_lt_8_subtile) { |
| 1159 | TEST_REQUIRES_X86_SSE2; |
| 1160 | for (size_t k = 1; k < 8; k++) { |
| 1161 | for (uint32_t m = 1; m <= 2; m++) { |
| 1162 | for (uint32_t n = 1; n <= 4; n++) { |
| 1163 | GemmMicrokernelTester() |
| 1164 | .mr(2) |
| 1165 | .nr(4) |
| 1166 | .kr(8) |
| 1167 | .sr(1) |
| 1168 | .m(m) |
| 1169 | .n(n) |
| 1170 | .k(k) |
| 1171 | .iterations(1) |
| 1172 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1173 | } |
| 1174 | } |
| 1175 | } |
| 1176 | } |
| 1177 | |
| 1178 | TEST(Q8_GEMM_2X4C8__SSE2, k_gt_8) { |
| 1179 | TEST_REQUIRES_X86_SSE2; |
| 1180 | for (size_t k = 9; k < 16; k++) { |
| 1181 | GemmMicrokernelTester() |
| 1182 | .mr(2) |
| 1183 | .nr(4) |
| 1184 | .kr(8) |
| 1185 | .sr(1) |
| 1186 | .m(2) |
| 1187 | .n(4) |
| 1188 | .k(k) |
| 1189 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1190 | } |
| 1191 | } |
| 1192 | |
| 1193 | TEST(Q8_GEMM_2X4C8__SSE2, k_gt_8_strided_a) { |
| 1194 | TEST_REQUIRES_X86_SSE2; |
| 1195 | for (size_t k = 9; k < 16; k++) { |
| 1196 | GemmMicrokernelTester() |
| 1197 | .mr(2) |
| 1198 | .nr(4) |
| 1199 | .kr(8) |
| 1200 | .sr(1) |
| 1201 | .m(2) |
| 1202 | .n(4) |
| 1203 | .k(k) |
| 1204 | .a_stride(19) |
| 1205 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1206 | } |
| 1207 | } |
| 1208 | |
| 1209 | TEST(Q8_GEMM_2X4C8__SSE2, k_gt_8_subtile) { |
| 1210 | TEST_REQUIRES_X86_SSE2; |
| 1211 | for (size_t k = 9; k < 16; k++) { |
| 1212 | for (uint32_t m = 1; m <= 2; m++) { |
| 1213 | for (uint32_t n = 1; n <= 4; n++) { |
| 1214 | GemmMicrokernelTester() |
| 1215 | .mr(2) |
| 1216 | .nr(4) |
| 1217 | .kr(8) |
| 1218 | .sr(1) |
| 1219 | .m(m) |
| 1220 | .n(n) |
| 1221 | .k(k) |
| 1222 | .iterations(1) |
| 1223 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1224 | } |
| 1225 | } |
| 1226 | } |
| 1227 | } |
| 1228 | |
| 1229 | TEST(Q8_GEMM_2X4C8__SSE2, k_div_8) { |
| 1230 | TEST_REQUIRES_X86_SSE2; |
| 1231 | for (size_t k = 16; k <= 80; k += 8) { |
| 1232 | GemmMicrokernelTester() |
| 1233 | .mr(2) |
| 1234 | .nr(4) |
| 1235 | .kr(8) |
| 1236 | .sr(1) |
| 1237 | .m(2) |
| 1238 | .n(4) |
| 1239 | .k(k) |
| 1240 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1241 | } |
| 1242 | } |
| 1243 | |
| 1244 | TEST(Q8_GEMM_2X4C8__SSE2, k_div_8_strided_a) { |
| 1245 | TEST_REQUIRES_X86_SSE2; |
| 1246 | for (size_t k = 16; k <= 80; k += 8) { |
| 1247 | GemmMicrokernelTester() |
| 1248 | .mr(2) |
| 1249 | .nr(4) |
| 1250 | .kr(8) |
| 1251 | .sr(1) |
| 1252 | .m(2) |
| 1253 | .n(4) |
| 1254 | .k(k) |
| 1255 | .a_stride(83) |
| 1256 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1257 | } |
| 1258 | } |
| 1259 | |
| 1260 | TEST(Q8_GEMM_2X4C8__SSE2, k_div_8_subtile) { |
| 1261 | TEST_REQUIRES_X86_SSE2; |
| 1262 | for (size_t k = 16; k <= 80; k += 8) { |
| 1263 | for (uint32_t m = 1; m <= 2; m++) { |
| 1264 | for (uint32_t n = 1; n <= 4; n++) { |
| 1265 | GemmMicrokernelTester() |
| 1266 | .mr(2) |
| 1267 | .nr(4) |
| 1268 | .kr(8) |
| 1269 | .sr(1) |
| 1270 | .m(m) |
| 1271 | .n(n) |
| 1272 | .k(k) |
| 1273 | .iterations(1) |
| 1274 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1275 | } |
| 1276 | } |
| 1277 | } |
| 1278 | } |
| 1279 | |
| 1280 | TEST(Q8_GEMM_2X4C8__SSE2, n_gt_4) { |
| 1281 | TEST_REQUIRES_X86_SSE2; |
| 1282 | for (uint32_t n = 5; n < 8; n++) { |
| 1283 | for (size_t k = 1; k <= 40; k += 9) { |
| 1284 | GemmMicrokernelTester() |
| 1285 | .mr(2) |
| 1286 | .nr(4) |
| 1287 | .kr(8) |
| 1288 | .sr(1) |
| 1289 | .m(2) |
| 1290 | .n(4) |
| 1291 | .k(k) |
| 1292 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1293 | } |
| 1294 | } |
| 1295 | } |
| 1296 | |
| 1297 | TEST(Q8_GEMM_2X4C8__SSE2, n_gt_4_strided_cn) { |
| 1298 | TEST_REQUIRES_X86_SSE2; |
| 1299 | for (uint32_t n = 5; n < 8; n++) { |
| 1300 | for (size_t k = 1; k <= 40; k += 9) { |
| 1301 | GemmMicrokernelTester() |
| 1302 | .mr(2) |
| 1303 | .nr(4) |
| 1304 | .kr(8) |
| 1305 | .sr(1) |
| 1306 | .m(2) |
| 1307 | .n(4) |
| 1308 | .k(k) |
| 1309 | .cn_stride(7) |
| 1310 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1311 | } |
| 1312 | } |
| 1313 | } |
| 1314 | |
| 1315 | TEST(Q8_GEMM_2X4C8__SSE2, n_gt_4_strided_a) { |
| 1316 | TEST_REQUIRES_X86_SSE2; |
| 1317 | for (uint32_t n = 5; n < 8; n++) { |
| 1318 | for (size_t k = 1; k <= 40; k += 9) { |
| 1319 | GemmMicrokernelTester() |
| 1320 | .mr(2) |
| 1321 | .nr(4) |
| 1322 | .kr(8) |
| 1323 | .sr(1) |
| 1324 | .m(2) |
| 1325 | .n(n) |
| 1326 | .k(k) |
| 1327 | .a_stride(43) |
| 1328 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1329 | } |
| 1330 | } |
| 1331 | } |
| 1332 | |
| 1333 | TEST(Q8_GEMM_2X4C8__SSE2, n_gt_4_subtile) { |
| 1334 | TEST_REQUIRES_X86_SSE2; |
| 1335 | for (uint32_t n = 5; n < 8; n++) { |
| 1336 | for (size_t k = 1; k <= 40; k += 9) { |
| 1337 | for (uint32_t m = 1; m <= 2; m++) { |
| 1338 | GemmMicrokernelTester() |
| 1339 | .mr(2) |
| 1340 | .nr(4) |
| 1341 | .kr(8) |
| 1342 | .sr(1) |
| 1343 | .m(m) |
| 1344 | .n(n) |
| 1345 | .k(k) |
| 1346 | .iterations(1) |
| 1347 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1348 | } |
| 1349 | } |
| 1350 | } |
| 1351 | } |
| 1352 | |
| 1353 | TEST(Q8_GEMM_2X4C8__SSE2, n_div_4) { |
| 1354 | TEST_REQUIRES_X86_SSE2; |
| 1355 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 1356 | for (size_t k = 1; k <= 40; k += 9) { |
| 1357 | GemmMicrokernelTester() |
| 1358 | .mr(2) |
| 1359 | .nr(4) |
| 1360 | .kr(8) |
| 1361 | .sr(1) |
| 1362 | .m(2) |
| 1363 | .n(4) |
| 1364 | .k(k) |
| 1365 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1366 | } |
| 1367 | } |
| 1368 | } |
| 1369 | |
| 1370 | TEST(Q8_GEMM_2X4C8__SSE2, n_div_4_strided_cn) { |
| 1371 | TEST_REQUIRES_X86_SSE2; |
| 1372 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 1373 | for (size_t k = 1; k <= 40; k += 9) { |
| 1374 | GemmMicrokernelTester() |
| 1375 | .mr(2) |
| 1376 | .nr(4) |
| 1377 | .kr(8) |
| 1378 | .sr(1) |
| 1379 | .m(2) |
| 1380 | .n(n) |
| 1381 | .k(k) |
| 1382 | .cn_stride(7) |
| 1383 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1384 | } |
| 1385 | } |
| 1386 | } |
| 1387 | |
| 1388 | TEST(Q8_GEMM_2X4C8__SSE2, n_div_4_strided_a) { |
| 1389 | TEST_REQUIRES_X86_SSE2; |
| 1390 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 1391 | for (size_t k = 1; k <= 40; k += 9) { |
| 1392 | GemmMicrokernelTester() |
| 1393 | .mr(2) |
| 1394 | .nr(4) |
| 1395 | .kr(8) |
| 1396 | .sr(1) |
| 1397 | .m(2) |
| 1398 | .n(n) |
| 1399 | .k(k) |
| 1400 | .a_stride(43) |
| 1401 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1402 | } |
| 1403 | } |
| 1404 | } |
| 1405 | |
| 1406 | TEST(Q8_GEMM_2X4C8__SSE2, n_div_4_subtile) { |
| 1407 | TEST_REQUIRES_X86_SSE2; |
| 1408 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 1409 | for (size_t k = 1; k <= 40; k += 9) { |
| 1410 | for (uint32_t m = 1; m <= 2; m++) { |
| 1411 | GemmMicrokernelTester() |
| 1412 | .mr(2) |
| 1413 | .nr(4) |
| 1414 | .kr(8) |
| 1415 | .sr(1) |
| 1416 | .m(m) |
| 1417 | .n(n) |
| 1418 | .k(k) |
| 1419 | .iterations(1) |
| 1420 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1421 | } |
| 1422 | } |
| 1423 | } |
| 1424 | } |
| 1425 | |
| 1426 | TEST(Q8_GEMM_2X4C8__SSE2, strided_cm_subtile) { |
| 1427 | TEST_REQUIRES_X86_SSE2; |
| 1428 | for (size_t k = 1; k <= 40; k += 9) { |
| 1429 | for (uint32_t m = 1; m <= 2; m++) { |
| 1430 | for (uint32_t n = 1; n <= 4; n++) { |
| 1431 | GemmMicrokernelTester() |
| 1432 | .mr(2) |
| 1433 | .nr(4) |
| 1434 | .kr(8) |
| 1435 | .sr(1) |
| 1436 | .m(m) |
| 1437 | .n(n) |
| 1438 | .k(k) |
| 1439 | .cm_stride(7) |
| 1440 | .iterations(1) |
| 1441 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1442 | } |
| 1443 | } |
| 1444 | } |
| 1445 | } |
| 1446 | |
| 1447 | TEST(Q8_GEMM_2X4C8__SSE2, qmin) { |
| 1448 | TEST_REQUIRES_X86_SSE2; |
| 1449 | GemmMicrokernelTester() |
| 1450 | .mr(2) |
| 1451 | .nr(4) |
| 1452 | .kr(8) |
| 1453 | .sr(1) |
| 1454 | .m(2) |
| 1455 | .n(4) |
| 1456 | .k(8) |
| 1457 | .qmin(128) |
| 1458 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1459 | } |
| 1460 | |
| 1461 | TEST(Q8_GEMM_2X4C8__SSE2, qmax) { |
| 1462 | TEST_REQUIRES_X86_SSE2; |
| 1463 | GemmMicrokernelTester() |
| 1464 | .mr(2) |
| 1465 | .nr(4) |
| 1466 | .kr(8) |
| 1467 | .sr(1) |
| 1468 | .m(2) |
| 1469 | .n(4) |
| 1470 | .k(8) |
| 1471 | .qmax(128) |
| 1472 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1473 | } |
| 1474 | |
| 1475 | TEST(Q8_GEMM_2X4C8__SSE2, strided_cm) { |
| 1476 | TEST_REQUIRES_X86_SSE2; |
| 1477 | GemmMicrokernelTester() |
| 1478 | .mr(2) |
| 1479 | .nr(4) |
| 1480 | .kr(8) |
| 1481 | .sr(1) |
| 1482 | .m(2) |
| 1483 | .n(4) |
| 1484 | .k(8) |
| 1485 | .cm_stride(7) |
| 1486 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1487 | } |
| 1488 | |
| 1489 | TEST(Q8_GEMM_2X4C8__SSE2, no_a_zero_point) { |
| 1490 | TEST_REQUIRES_X86_SSE2; |
| 1491 | for (size_t k = 1; k <= 40; k += 9) { |
| 1492 | GemmMicrokernelTester() |
| 1493 | .mr(2) |
| 1494 | .nr(4) |
| 1495 | .kr(8) |
| 1496 | .sr(1) |
| 1497 | .m(2) |
| 1498 | .n(4) |
| 1499 | .k(k) |
| 1500 | .a_zero_point(0) |
| 1501 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1502 | } |
| 1503 | } |
| 1504 | |
| 1505 | TEST(Q8_GEMM_2X4C8__SSE2, no_b_zero_point) { |
| 1506 | TEST_REQUIRES_X86_SSE2; |
| 1507 | for (size_t k = 1; k <= 40; k += 9) { |
| 1508 | GemmMicrokernelTester() |
| 1509 | .mr(2) |
| 1510 | .nr(4) |
| 1511 | .kr(8) |
| 1512 | .sr(1) |
| 1513 | .m(2) |
| 1514 | .n(4) |
| 1515 | .k(k) |
| 1516 | .b_zero_point(0) |
| 1517 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1518 | } |
| 1519 | } |
| 1520 | |
| 1521 | TEST(Q8_GEMM_2X4C8__SSE2, no_zero_point) { |
| 1522 | TEST_REQUIRES_X86_SSE2; |
| 1523 | for (size_t k = 1; k <= 40; k += 9) { |
| 1524 | GemmMicrokernelTester() |
| 1525 | .mr(2) |
| 1526 | .nr(4) |
| 1527 | .kr(8) |
| 1528 | .sr(1) |
| 1529 | .m(2) |
| 1530 | .n(4) |
| 1531 | .k(k) |
| 1532 | .a_zero_point(0) |
| 1533 | .b_zero_point(0) |
| 1534 | .Test(xnn_q8_gemm_ukernel_2x4c8__sse2); |
| 1535 | } |
| 1536 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1537 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1538 | |
| 1539 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1540 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1541 | TEST(Q8_GEMM_4X4C2__SSE2, k_eq_8) { |
| 1542 | TEST_REQUIRES_X86_SSE2; |
| 1543 | GemmMicrokernelTester() |
| 1544 | .mr(4) |
| 1545 | .nr(4) |
| 1546 | .kr(2) |
| 1547 | .sr(1) |
| 1548 | .m(4) |
| 1549 | .n(4) |
| 1550 | .k(8) |
| 1551 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1552 | } |
| 1553 | |
| 1554 | TEST(Q8_GEMM_4X4C2__SSE2, strided_cn) { |
| 1555 | TEST_REQUIRES_X86_SSE2; |
| 1556 | GemmMicrokernelTester() |
| 1557 | .mr(4) |
| 1558 | .nr(4) |
| 1559 | .kr(2) |
| 1560 | .sr(1) |
| 1561 | .m(4) |
| 1562 | .n(4) |
| 1563 | .k(8) |
| 1564 | .cn_stride(7) |
| 1565 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1566 | } |
| 1567 | |
| 1568 | TEST(Q8_GEMM_4X4C2__SSE2, k_eq_8_strided_a) { |
| 1569 | TEST_REQUIRES_X86_SSE2; |
| 1570 | GemmMicrokernelTester() |
| 1571 | .mr(4) |
| 1572 | .nr(4) |
| 1573 | .kr(2) |
| 1574 | .sr(1) |
| 1575 | .m(4) |
| 1576 | .n(4) |
| 1577 | .k(8) |
| 1578 | .a_stride(11) |
| 1579 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1580 | } |
| 1581 | |
| 1582 | TEST(Q8_GEMM_4X4C2__SSE2, k_eq_8_subtile) { |
| 1583 | TEST_REQUIRES_X86_SSE2; |
| 1584 | for (uint32_t m = 1; m <= 4; m++) { |
| 1585 | for (uint32_t n = 1; n <= 4; n++) { |
| 1586 | GemmMicrokernelTester() |
| 1587 | .mr(4) |
| 1588 | .nr(4) |
| 1589 | .kr(2) |
| 1590 | .sr(1) |
| 1591 | .m(m) |
| 1592 | .n(n) |
| 1593 | .k(8) |
| 1594 | .iterations(1) |
| 1595 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1596 | } |
| 1597 | } |
| 1598 | } |
| 1599 | |
| 1600 | TEST(Q8_GEMM_4X4C2__SSE2, k_eq_8_subtile_m) { |
| 1601 | TEST_REQUIRES_X86_SSE2; |
| 1602 | for (uint32_t m = 1; m <= 4; m++) { |
| 1603 | GemmMicrokernelTester() |
| 1604 | .mr(4) |
| 1605 | .nr(4) |
| 1606 | .kr(2) |
| 1607 | .sr(1) |
| 1608 | .m(m) |
| 1609 | .n(4) |
| 1610 | .k(8) |
| 1611 | .iterations(1) |
| 1612 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1613 | } |
| 1614 | } |
| 1615 | |
| 1616 | TEST(Q8_GEMM_4X4C2__SSE2, k_eq_8_subtile_n) { |
| 1617 | TEST_REQUIRES_X86_SSE2; |
| 1618 | for (uint32_t n = 1; n <= 4; n++) { |
| 1619 | GemmMicrokernelTester() |
| 1620 | .mr(4) |
| 1621 | .nr(4) |
| 1622 | .kr(2) |
| 1623 | .sr(1) |
| 1624 | .m(4) |
| 1625 | .n(n) |
| 1626 | .k(8) |
| 1627 | .iterations(1) |
| 1628 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1629 | } |
| 1630 | } |
| 1631 | |
| 1632 | TEST(Q8_GEMM_4X4C2__SSE2, k_lt_8) { |
| 1633 | TEST_REQUIRES_X86_SSE2; |
| 1634 | for (size_t k = 1; k < 8; k++) { |
| 1635 | GemmMicrokernelTester() |
| 1636 | .mr(4) |
| 1637 | .nr(4) |
| 1638 | .kr(2) |
| 1639 | .sr(1) |
| 1640 | .m(4) |
| 1641 | .n(4) |
| 1642 | .k(k) |
| 1643 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1644 | } |
| 1645 | } |
| 1646 | |
| 1647 | TEST(Q8_GEMM_4X4C2__SSE2, k_lt_8_strided_a) { |
| 1648 | TEST_REQUIRES_X86_SSE2; |
| 1649 | for (size_t k = 1; k < 8; k++) { |
| 1650 | GemmMicrokernelTester() |
| 1651 | .mr(4) |
| 1652 | .nr(4) |
| 1653 | .kr(2) |
| 1654 | .sr(1) |
| 1655 | .m(4) |
| 1656 | .n(4) |
| 1657 | .k(k) |
| 1658 | .a_stride(11) |
| 1659 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1660 | } |
| 1661 | } |
| 1662 | |
| 1663 | TEST(Q8_GEMM_4X4C2__SSE2, k_lt_8_subtile) { |
| 1664 | TEST_REQUIRES_X86_SSE2; |
| 1665 | for (size_t k = 1; k < 8; k++) { |
| 1666 | for (uint32_t m = 1; m <= 4; m++) { |
| 1667 | for (uint32_t n = 1; n <= 4; n++) { |
| 1668 | GemmMicrokernelTester() |
| 1669 | .mr(4) |
| 1670 | .nr(4) |
| 1671 | .kr(2) |
| 1672 | .sr(1) |
| 1673 | .m(m) |
| 1674 | .n(n) |
| 1675 | .k(k) |
| 1676 | .iterations(1) |
| 1677 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1678 | } |
| 1679 | } |
| 1680 | } |
| 1681 | } |
| 1682 | |
| 1683 | TEST(Q8_GEMM_4X4C2__SSE2, k_gt_8) { |
| 1684 | TEST_REQUIRES_X86_SSE2; |
| 1685 | for (size_t k = 9; k < 16; k++) { |
| 1686 | GemmMicrokernelTester() |
| 1687 | .mr(4) |
| 1688 | .nr(4) |
| 1689 | .kr(2) |
| 1690 | .sr(1) |
| 1691 | .m(4) |
| 1692 | .n(4) |
| 1693 | .k(k) |
| 1694 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1695 | } |
| 1696 | } |
| 1697 | |
| 1698 | TEST(Q8_GEMM_4X4C2__SSE2, k_gt_8_strided_a) { |
| 1699 | TEST_REQUIRES_X86_SSE2; |
| 1700 | for (size_t k = 9; k < 16; k++) { |
| 1701 | GemmMicrokernelTester() |
| 1702 | .mr(4) |
| 1703 | .nr(4) |
| 1704 | .kr(2) |
| 1705 | .sr(1) |
| 1706 | .m(4) |
| 1707 | .n(4) |
| 1708 | .k(k) |
| 1709 | .a_stride(19) |
| 1710 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1711 | } |
| 1712 | } |
| 1713 | |
| 1714 | TEST(Q8_GEMM_4X4C2__SSE2, k_gt_8_subtile) { |
| 1715 | TEST_REQUIRES_X86_SSE2; |
| 1716 | for (size_t k = 9; k < 16; k++) { |
| 1717 | for (uint32_t m = 1; m <= 4; m++) { |
| 1718 | for (uint32_t n = 1; n <= 4; n++) { |
| 1719 | GemmMicrokernelTester() |
| 1720 | .mr(4) |
| 1721 | .nr(4) |
| 1722 | .kr(2) |
| 1723 | .sr(1) |
| 1724 | .m(m) |
| 1725 | .n(n) |
| 1726 | .k(k) |
| 1727 | .iterations(1) |
| 1728 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1729 | } |
| 1730 | } |
| 1731 | } |
| 1732 | } |
| 1733 | |
| 1734 | TEST(Q8_GEMM_4X4C2__SSE2, k_div_8) { |
| 1735 | TEST_REQUIRES_X86_SSE2; |
| 1736 | for (size_t k = 16; k <= 80; k += 8) { |
| 1737 | GemmMicrokernelTester() |
| 1738 | .mr(4) |
| 1739 | .nr(4) |
| 1740 | .kr(2) |
| 1741 | .sr(1) |
| 1742 | .m(4) |
| 1743 | .n(4) |
| 1744 | .k(k) |
| 1745 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1746 | } |
| 1747 | } |
| 1748 | |
| 1749 | TEST(Q8_GEMM_4X4C2__SSE2, k_div_8_strided_a) { |
| 1750 | TEST_REQUIRES_X86_SSE2; |
| 1751 | for (size_t k = 16; k <= 80; k += 8) { |
| 1752 | GemmMicrokernelTester() |
| 1753 | .mr(4) |
| 1754 | .nr(4) |
| 1755 | .kr(2) |
| 1756 | .sr(1) |
| 1757 | .m(4) |
| 1758 | .n(4) |
| 1759 | .k(k) |
| 1760 | .a_stride(83) |
| 1761 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1762 | } |
| 1763 | } |
| 1764 | |
| 1765 | TEST(Q8_GEMM_4X4C2__SSE2, k_div_8_subtile) { |
| 1766 | TEST_REQUIRES_X86_SSE2; |
| 1767 | for (size_t k = 16; k <= 80; k += 8) { |
| 1768 | for (uint32_t m = 1; m <= 4; m++) { |
| 1769 | for (uint32_t n = 1; n <= 4; n++) { |
| 1770 | GemmMicrokernelTester() |
| 1771 | .mr(4) |
| 1772 | .nr(4) |
| 1773 | .kr(2) |
| 1774 | .sr(1) |
| 1775 | .m(m) |
| 1776 | .n(n) |
| 1777 | .k(k) |
| 1778 | .iterations(1) |
| 1779 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1780 | } |
| 1781 | } |
| 1782 | } |
| 1783 | } |
| 1784 | |
| 1785 | TEST(Q8_GEMM_4X4C2__SSE2, n_gt_4) { |
| 1786 | TEST_REQUIRES_X86_SSE2; |
| 1787 | for (uint32_t n = 5; n < 8; n++) { |
| 1788 | for (size_t k = 1; k <= 40; k += 9) { |
| 1789 | GemmMicrokernelTester() |
| 1790 | .mr(4) |
| 1791 | .nr(4) |
| 1792 | .kr(2) |
| 1793 | .sr(1) |
| 1794 | .m(4) |
| 1795 | .n(4) |
| 1796 | .k(k) |
| 1797 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1798 | } |
| 1799 | } |
| 1800 | } |
| 1801 | |
| 1802 | TEST(Q8_GEMM_4X4C2__SSE2, n_gt_4_strided_cn) { |
| 1803 | TEST_REQUIRES_X86_SSE2; |
| 1804 | for (uint32_t n = 5; n < 8; n++) { |
| 1805 | for (size_t k = 1; k <= 40; k += 9) { |
| 1806 | GemmMicrokernelTester() |
| 1807 | .mr(4) |
| 1808 | .nr(4) |
| 1809 | .kr(2) |
| 1810 | .sr(1) |
| 1811 | .m(4) |
| 1812 | .n(4) |
| 1813 | .k(k) |
| 1814 | .cn_stride(7) |
| 1815 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1816 | } |
| 1817 | } |
| 1818 | } |
| 1819 | |
| 1820 | TEST(Q8_GEMM_4X4C2__SSE2, n_gt_4_strided_a) { |
| 1821 | TEST_REQUIRES_X86_SSE2; |
| 1822 | for (uint32_t n = 5; n < 8; n++) { |
| 1823 | for (size_t k = 1; k <= 40; k += 9) { |
| 1824 | GemmMicrokernelTester() |
| 1825 | .mr(4) |
| 1826 | .nr(4) |
| 1827 | .kr(2) |
| 1828 | .sr(1) |
| 1829 | .m(4) |
| 1830 | .n(n) |
| 1831 | .k(k) |
| 1832 | .a_stride(43) |
| 1833 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1834 | } |
| 1835 | } |
| 1836 | } |
| 1837 | |
| 1838 | TEST(Q8_GEMM_4X4C2__SSE2, n_gt_4_subtile) { |
| 1839 | TEST_REQUIRES_X86_SSE2; |
| 1840 | for (uint32_t n = 5; n < 8; n++) { |
| 1841 | for (size_t k = 1; k <= 40; k += 9) { |
| 1842 | for (uint32_t m = 1; m <= 4; m++) { |
| 1843 | GemmMicrokernelTester() |
| 1844 | .mr(4) |
| 1845 | .nr(4) |
| 1846 | .kr(2) |
| 1847 | .sr(1) |
| 1848 | .m(m) |
| 1849 | .n(n) |
| 1850 | .k(k) |
| 1851 | .iterations(1) |
| 1852 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1853 | } |
| 1854 | } |
| 1855 | } |
| 1856 | } |
| 1857 | |
| 1858 | TEST(Q8_GEMM_4X4C2__SSE2, n_div_4) { |
| 1859 | TEST_REQUIRES_X86_SSE2; |
| 1860 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 1861 | for (size_t k = 1; k <= 40; k += 9) { |
| 1862 | GemmMicrokernelTester() |
| 1863 | .mr(4) |
| 1864 | .nr(4) |
| 1865 | .kr(2) |
| 1866 | .sr(1) |
| 1867 | .m(4) |
| 1868 | .n(4) |
| 1869 | .k(k) |
| 1870 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1871 | } |
| 1872 | } |
| 1873 | } |
| 1874 | |
| 1875 | TEST(Q8_GEMM_4X4C2__SSE2, n_div_4_strided_cn) { |
| 1876 | TEST_REQUIRES_X86_SSE2; |
| 1877 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 1878 | for (size_t k = 1; k <= 40; k += 9) { |
| 1879 | GemmMicrokernelTester() |
| 1880 | .mr(4) |
| 1881 | .nr(4) |
| 1882 | .kr(2) |
| 1883 | .sr(1) |
| 1884 | .m(4) |
| 1885 | .n(n) |
| 1886 | .k(k) |
| 1887 | .cn_stride(7) |
| 1888 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1889 | } |
| 1890 | } |
| 1891 | } |
| 1892 | |
| 1893 | TEST(Q8_GEMM_4X4C2__SSE2, n_div_4_strided_a) { |
| 1894 | TEST_REQUIRES_X86_SSE2; |
| 1895 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 1896 | for (size_t k = 1; k <= 40; k += 9) { |
| 1897 | GemmMicrokernelTester() |
| 1898 | .mr(4) |
| 1899 | .nr(4) |
| 1900 | .kr(2) |
| 1901 | .sr(1) |
| 1902 | .m(4) |
| 1903 | .n(n) |
| 1904 | .k(k) |
| 1905 | .a_stride(43) |
| 1906 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1907 | } |
| 1908 | } |
| 1909 | } |
| 1910 | |
| 1911 | TEST(Q8_GEMM_4X4C2__SSE2, n_div_4_subtile) { |
| 1912 | TEST_REQUIRES_X86_SSE2; |
| 1913 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 1914 | for (size_t k = 1; k <= 40; k += 9) { |
| 1915 | for (uint32_t m = 1; m <= 4; m++) { |
| 1916 | GemmMicrokernelTester() |
| 1917 | .mr(4) |
| 1918 | .nr(4) |
| 1919 | .kr(2) |
| 1920 | .sr(1) |
| 1921 | .m(m) |
| 1922 | .n(n) |
| 1923 | .k(k) |
| 1924 | .iterations(1) |
| 1925 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1926 | } |
| 1927 | } |
| 1928 | } |
| 1929 | } |
| 1930 | |
| 1931 | TEST(Q8_GEMM_4X4C2__SSE2, strided_cm_subtile) { |
| 1932 | TEST_REQUIRES_X86_SSE2; |
| 1933 | for (size_t k = 1; k <= 40; k += 9) { |
| 1934 | for (uint32_t m = 1; m <= 4; m++) { |
| 1935 | for (uint32_t n = 1; n <= 4; n++) { |
| 1936 | GemmMicrokernelTester() |
| 1937 | .mr(4) |
| 1938 | .nr(4) |
| 1939 | .kr(2) |
| 1940 | .sr(1) |
| 1941 | .m(m) |
| 1942 | .n(n) |
| 1943 | .k(k) |
| 1944 | .cm_stride(7) |
| 1945 | .iterations(1) |
| 1946 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1947 | } |
| 1948 | } |
| 1949 | } |
| 1950 | } |
| 1951 | |
| 1952 | TEST(Q8_GEMM_4X4C2__SSE2, qmin) { |
| 1953 | TEST_REQUIRES_X86_SSE2; |
| 1954 | GemmMicrokernelTester() |
| 1955 | .mr(4) |
| 1956 | .nr(4) |
| 1957 | .kr(2) |
| 1958 | .sr(1) |
| 1959 | .m(4) |
| 1960 | .n(4) |
| 1961 | .k(8) |
| 1962 | .qmin(128) |
| 1963 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1964 | } |
| 1965 | |
| 1966 | TEST(Q8_GEMM_4X4C2__SSE2, qmax) { |
| 1967 | TEST_REQUIRES_X86_SSE2; |
| 1968 | GemmMicrokernelTester() |
| 1969 | .mr(4) |
| 1970 | .nr(4) |
| 1971 | .kr(2) |
| 1972 | .sr(1) |
| 1973 | .m(4) |
| 1974 | .n(4) |
| 1975 | .k(8) |
| 1976 | .qmax(128) |
| 1977 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1978 | } |
| 1979 | |
| 1980 | TEST(Q8_GEMM_4X4C2__SSE2, strided_cm) { |
| 1981 | TEST_REQUIRES_X86_SSE2; |
| 1982 | GemmMicrokernelTester() |
| 1983 | .mr(4) |
| 1984 | .nr(4) |
| 1985 | .kr(2) |
| 1986 | .sr(1) |
| 1987 | .m(4) |
| 1988 | .n(4) |
| 1989 | .k(8) |
| 1990 | .cm_stride(7) |
| 1991 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 1992 | } |
| 1993 | |
| 1994 | TEST(Q8_GEMM_4X4C2__SSE2, no_a_zero_point) { |
| 1995 | TEST_REQUIRES_X86_SSE2; |
| 1996 | for (size_t k = 1; k <= 40; k += 9) { |
| 1997 | GemmMicrokernelTester() |
| 1998 | .mr(4) |
| 1999 | .nr(4) |
| 2000 | .kr(2) |
| 2001 | .sr(1) |
| 2002 | .m(4) |
| 2003 | .n(4) |
| 2004 | .k(k) |
| 2005 | .a_zero_point(0) |
| 2006 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 2007 | } |
| 2008 | } |
| 2009 | |
| 2010 | TEST(Q8_GEMM_4X4C2__SSE2, no_b_zero_point) { |
| 2011 | TEST_REQUIRES_X86_SSE2; |
| 2012 | for (size_t k = 1; k <= 40; k += 9) { |
| 2013 | GemmMicrokernelTester() |
| 2014 | .mr(4) |
| 2015 | .nr(4) |
| 2016 | .kr(2) |
| 2017 | .sr(1) |
| 2018 | .m(4) |
| 2019 | .n(4) |
| 2020 | .k(k) |
| 2021 | .b_zero_point(0) |
| 2022 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 2023 | } |
| 2024 | } |
| 2025 | |
| 2026 | TEST(Q8_GEMM_4X4C2__SSE2, no_zero_point) { |
| 2027 | TEST_REQUIRES_X86_SSE2; |
| 2028 | for (size_t k = 1; k <= 40; k += 9) { |
| 2029 | GemmMicrokernelTester() |
| 2030 | .mr(4) |
| 2031 | .nr(4) |
| 2032 | .kr(2) |
| 2033 | .sr(1) |
| 2034 | .m(4) |
| 2035 | .n(4) |
| 2036 | .k(k) |
| 2037 | .a_zero_point(0) |
| 2038 | .b_zero_point(0) |
| 2039 | .Test(xnn_q8_gemm_ukernel_4x4c2__sse2); |
| 2040 | } |
| 2041 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2042 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2043 | |
| 2044 | |
| 2045 | TEST(Q8_GEMM_2X2__SCALAR, k_eq_1) { |
| 2046 | GemmMicrokernelTester() |
| 2047 | .mr(2) |
| 2048 | .nr(2) |
| 2049 | .kr(1) |
| 2050 | .sr(1) |
| 2051 | .m(2) |
| 2052 | .n(2) |
| 2053 | .k(1) |
| 2054 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2055 | } |
| 2056 | |
| 2057 | TEST(Q8_GEMM_2X2__SCALAR, strided_cn) { |
| 2058 | GemmMicrokernelTester() |
| 2059 | .mr(2) |
| 2060 | .nr(2) |
| 2061 | .kr(1) |
| 2062 | .sr(1) |
| 2063 | .m(2) |
| 2064 | .n(2) |
| 2065 | .k(1) |
| 2066 | .cn_stride(5) |
| 2067 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2068 | } |
| 2069 | |
| 2070 | TEST(Q8_GEMM_2X2__SCALAR, k_eq_1_strided_a) { |
| 2071 | GemmMicrokernelTester() |
| 2072 | .mr(2) |
| 2073 | .nr(2) |
| 2074 | .kr(1) |
| 2075 | .sr(1) |
| 2076 | .m(2) |
| 2077 | .n(2) |
| 2078 | .k(1) |
| 2079 | .a_stride(3) |
| 2080 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2081 | } |
| 2082 | |
| 2083 | TEST(Q8_GEMM_2X2__SCALAR, k_eq_1_subtile) { |
| 2084 | for (uint32_t m = 1; m <= 2; m++) { |
| 2085 | for (uint32_t n = 1; n <= 2; n++) { |
| 2086 | GemmMicrokernelTester() |
| 2087 | .mr(2) |
| 2088 | .nr(2) |
| 2089 | .kr(1) |
| 2090 | .sr(1) |
| 2091 | .m(m) |
| 2092 | .n(n) |
| 2093 | .k(1) |
| 2094 | .iterations(1) |
| 2095 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2096 | } |
| 2097 | } |
| 2098 | } |
| 2099 | |
| 2100 | TEST(Q8_GEMM_2X2__SCALAR, k_eq_1_subtile_m) { |
| 2101 | for (uint32_t m = 1; m <= 2; m++) { |
| 2102 | GemmMicrokernelTester() |
| 2103 | .mr(2) |
| 2104 | .nr(2) |
| 2105 | .kr(1) |
| 2106 | .sr(1) |
| 2107 | .m(m) |
| 2108 | .n(2) |
| 2109 | .k(1) |
| 2110 | .iterations(1) |
| 2111 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2112 | } |
| 2113 | } |
| 2114 | |
| 2115 | TEST(Q8_GEMM_2X2__SCALAR, k_eq_1_subtile_n) { |
| 2116 | for (uint32_t n = 1; n <= 2; n++) { |
| 2117 | GemmMicrokernelTester() |
| 2118 | .mr(2) |
| 2119 | .nr(2) |
| 2120 | .kr(1) |
| 2121 | .sr(1) |
| 2122 | .m(2) |
| 2123 | .n(n) |
| 2124 | .k(1) |
| 2125 | .iterations(1) |
| 2126 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2127 | } |
| 2128 | } |
| 2129 | |
| 2130 | TEST(Q8_GEMM_2X2__SCALAR, k_gt_1) { |
| 2131 | for (size_t k = 2; k < 10; k++) { |
| 2132 | GemmMicrokernelTester() |
| 2133 | .mr(2) |
| 2134 | .nr(2) |
| 2135 | .kr(1) |
| 2136 | .sr(1) |
| 2137 | .m(2) |
| 2138 | .n(2) |
| 2139 | .k(k) |
| 2140 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2141 | } |
| 2142 | } |
| 2143 | |
| 2144 | TEST(Q8_GEMM_2X2__SCALAR, k_gt_1_strided_a) { |
| 2145 | for (size_t k = 2; k < 10; k++) { |
| 2146 | GemmMicrokernelTester() |
| 2147 | .mr(2) |
| 2148 | .nr(2) |
| 2149 | .kr(1) |
| 2150 | .sr(1) |
| 2151 | .m(2) |
| 2152 | .n(2) |
| 2153 | .k(k) |
| 2154 | .a_stride(11) |
| 2155 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2156 | } |
| 2157 | } |
| 2158 | |
| 2159 | TEST(Q8_GEMM_2X2__SCALAR, k_gt_1_subtile) { |
| 2160 | for (size_t k = 2; k < 10; k++) { |
| 2161 | for (uint32_t m = 1; m <= 2; m++) { |
| 2162 | for (uint32_t n = 1; n <= 2; n++) { |
| 2163 | GemmMicrokernelTester() |
| 2164 | .mr(2) |
| 2165 | .nr(2) |
| 2166 | .kr(1) |
| 2167 | .sr(1) |
| 2168 | .m(m) |
| 2169 | .n(n) |
| 2170 | .k(k) |
| 2171 | .iterations(1) |
| 2172 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2173 | } |
| 2174 | } |
| 2175 | } |
| 2176 | } |
| 2177 | |
| 2178 | TEST(Q8_GEMM_2X2__SCALAR, n_gt_2) { |
| 2179 | for (uint32_t n = 3; n < 4; n++) { |
| 2180 | for (size_t k = 1; k <= 5; k += 2) { |
| 2181 | GemmMicrokernelTester() |
| 2182 | .mr(2) |
| 2183 | .nr(2) |
| 2184 | .kr(1) |
| 2185 | .sr(1) |
| 2186 | .m(2) |
| 2187 | .n(2) |
| 2188 | .k(k) |
| 2189 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2190 | } |
| 2191 | } |
| 2192 | } |
| 2193 | |
| 2194 | TEST(Q8_GEMM_2X2__SCALAR, n_gt_2_strided_cn) { |
| 2195 | for (uint32_t n = 3; n < 4; n++) { |
| 2196 | for (size_t k = 1; k <= 5; k += 2) { |
| 2197 | GemmMicrokernelTester() |
| 2198 | .mr(2) |
| 2199 | .nr(2) |
| 2200 | .kr(1) |
| 2201 | .sr(1) |
| 2202 | .m(2) |
| 2203 | .n(2) |
| 2204 | .k(k) |
| 2205 | .cn_stride(5) |
| 2206 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2207 | } |
| 2208 | } |
| 2209 | } |
| 2210 | |
| 2211 | TEST(Q8_GEMM_2X2__SCALAR, n_gt_2_strided_a) { |
| 2212 | for (uint32_t n = 3; n < 4; n++) { |
| 2213 | for (size_t k = 1; k <= 5; k += 2) { |
| 2214 | GemmMicrokernelTester() |
| 2215 | .mr(2) |
| 2216 | .nr(2) |
| 2217 | .kr(1) |
| 2218 | .sr(1) |
| 2219 | .m(2) |
| 2220 | .n(n) |
| 2221 | .k(k) |
| 2222 | .a_stride(7) |
| 2223 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2224 | } |
| 2225 | } |
| 2226 | } |
| 2227 | |
| 2228 | TEST(Q8_GEMM_2X2__SCALAR, n_gt_2_subtile) { |
| 2229 | for (uint32_t n = 3; n < 4; n++) { |
| 2230 | for (size_t k = 1; k <= 5; k += 2) { |
| 2231 | for (uint32_t m = 1; m <= 2; m++) { |
| 2232 | GemmMicrokernelTester() |
| 2233 | .mr(2) |
| 2234 | .nr(2) |
| 2235 | .kr(1) |
| 2236 | .sr(1) |
| 2237 | .m(m) |
| 2238 | .n(n) |
| 2239 | .k(k) |
| 2240 | .iterations(1) |
| 2241 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2242 | } |
| 2243 | } |
| 2244 | } |
| 2245 | } |
| 2246 | |
| 2247 | TEST(Q8_GEMM_2X2__SCALAR, n_div_2) { |
| 2248 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2249 | for (size_t k = 1; k <= 5; k += 2) { |
| 2250 | GemmMicrokernelTester() |
| 2251 | .mr(2) |
| 2252 | .nr(2) |
| 2253 | .kr(1) |
| 2254 | .sr(1) |
| 2255 | .m(2) |
| 2256 | .n(2) |
| 2257 | .k(k) |
| 2258 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2259 | } |
| 2260 | } |
| 2261 | } |
| 2262 | |
| 2263 | TEST(Q8_GEMM_2X2__SCALAR, n_div_2_strided_cn) { |
| 2264 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2265 | for (size_t k = 1; k <= 5; k += 2) { |
| 2266 | GemmMicrokernelTester() |
| 2267 | .mr(2) |
| 2268 | .nr(2) |
| 2269 | .kr(1) |
| 2270 | .sr(1) |
| 2271 | .m(2) |
| 2272 | .n(n) |
| 2273 | .k(k) |
| 2274 | .cn_stride(5) |
| 2275 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2276 | } |
| 2277 | } |
| 2278 | } |
| 2279 | |
| 2280 | TEST(Q8_GEMM_2X2__SCALAR, n_div_2_strided_a) { |
| 2281 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2282 | for (size_t k = 1; k <= 5; k += 2) { |
| 2283 | GemmMicrokernelTester() |
| 2284 | .mr(2) |
| 2285 | .nr(2) |
| 2286 | .kr(1) |
| 2287 | .sr(1) |
| 2288 | .m(2) |
| 2289 | .n(n) |
| 2290 | .k(k) |
| 2291 | .a_stride(7) |
| 2292 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2293 | } |
| 2294 | } |
| 2295 | } |
| 2296 | |
| 2297 | TEST(Q8_GEMM_2X2__SCALAR, n_div_2_subtile) { |
| 2298 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2299 | for (size_t k = 1; k <= 5; k += 2) { |
| 2300 | for (uint32_t m = 1; m <= 2; m++) { |
| 2301 | GemmMicrokernelTester() |
| 2302 | .mr(2) |
| 2303 | .nr(2) |
| 2304 | .kr(1) |
| 2305 | .sr(1) |
| 2306 | .m(m) |
| 2307 | .n(n) |
| 2308 | .k(k) |
| 2309 | .iterations(1) |
| 2310 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2311 | } |
| 2312 | } |
| 2313 | } |
| 2314 | } |
| 2315 | |
| 2316 | TEST(Q8_GEMM_2X2__SCALAR, strided_cm_subtile) { |
| 2317 | for (size_t k = 1; k <= 5; k += 2) { |
| 2318 | for (uint32_t m = 1; m <= 2; m++) { |
| 2319 | for (uint32_t n = 1; n <= 2; n++) { |
| 2320 | GemmMicrokernelTester() |
| 2321 | .mr(2) |
| 2322 | .nr(2) |
| 2323 | .kr(1) |
| 2324 | .sr(1) |
| 2325 | .m(m) |
| 2326 | .n(n) |
| 2327 | .k(k) |
| 2328 | .cm_stride(5) |
| 2329 | .iterations(1) |
| 2330 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2331 | } |
| 2332 | } |
| 2333 | } |
| 2334 | } |
| 2335 | |
| 2336 | TEST(Q8_GEMM_2X2__SCALAR, qmin) { |
| 2337 | GemmMicrokernelTester() |
| 2338 | .mr(2) |
| 2339 | .nr(2) |
| 2340 | .kr(1) |
| 2341 | .sr(1) |
| 2342 | .m(2) |
| 2343 | .n(2) |
| 2344 | .k(1) |
| 2345 | .qmin(128) |
| 2346 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2347 | } |
| 2348 | |
| 2349 | TEST(Q8_GEMM_2X2__SCALAR, qmax) { |
| 2350 | GemmMicrokernelTester() |
| 2351 | .mr(2) |
| 2352 | .nr(2) |
| 2353 | .kr(1) |
| 2354 | .sr(1) |
| 2355 | .m(2) |
| 2356 | .n(2) |
| 2357 | .k(1) |
| 2358 | .qmax(128) |
| 2359 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2360 | } |
| 2361 | |
| 2362 | TEST(Q8_GEMM_2X2__SCALAR, strided_cm) { |
| 2363 | GemmMicrokernelTester() |
| 2364 | .mr(2) |
| 2365 | .nr(2) |
| 2366 | .kr(1) |
| 2367 | .sr(1) |
| 2368 | .m(2) |
| 2369 | .n(2) |
| 2370 | .k(1) |
| 2371 | .cm_stride(5) |
| 2372 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2373 | } |
| 2374 | |
| 2375 | TEST(Q8_GEMM_2X2__SCALAR, no_a_zero_point) { |
| 2376 | for (size_t k = 1; k <= 5; k += 2) { |
| 2377 | GemmMicrokernelTester() |
| 2378 | .mr(2) |
| 2379 | .nr(2) |
| 2380 | .kr(1) |
| 2381 | .sr(1) |
| 2382 | .m(2) |
| 2383 | .n(2) |
| 2384 | .k(k) |
| 2385 | .a_zero_point(0) |
| 2386 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2387 | } |
| 2388 | } |
| 2389 | |
| 2390 | TEST(Q8_GEMM_2X2__SCALAR, no_b_zero_point) { |
| 2391 | for (size_t k = 1; k <= 5; k += 2) { |
| 2392 | GemmMicrokernelTester() |
| 2393 | .mr(2) |
| 2394 | .nr(2) |
| 2395 | .kr(1) |
| 2396 | .sr(1) |
| 2397 | .m(2) |
| 2398 | .n(2) |
| 2399 | .k(k) |
| 2400 | .b_zero_point(0) |
| 2401 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2402 | } |
| 2403 | } |
| 2404 | |
| 2405 | TEST(Q8_GEMM_2X2__SCALAR, no_zero_point) { |
| 2406 | for (size_t k = 1; k <= 5; k += 2) { |
| 2407 | GemmMicrokernelTester() |
| 2408 | .mr(2) |
| 2409 | .nr(2) |
| 2410 | .kr(1) |
| 2411 | .sr(1) |
| 2412 | .m(2) |
| 2413 | .n(2) |
| 2414 | .k(k) |
| 2415 | .a_zero_point(0) |
| 2416 | .b_zero_point(0) |
| 2417 | .Test(xnn_q8_gemm_ukernel_2x2__scalar, GemmMicrokernelTester::Variant::Scalar); |
| 2418 | } |
| 2419 | } |