| Frank Barchard | b0e4fae | 2020-05-04 15:27:51 -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/f16-igemm-minmax.yaml |
| 11 | // Generator: tools/generate-gemm-test.py |
| 12 | |
| 13 | |
| 14 | #include <gtest/gtest.h> |
| 15 | |
| 16 | #include <xnnpack/common.h> |
| 17 | #include <xnnpack/isa-checks.h> |
| 18 | |
| 19 | #include <xnnpack/gemm.h> |
| 20 | #include <xnnpack/igemm.h> |
| 21 | #include <xnnpack/ppmm.h> |
| 22 | #include "gemm-microkernel-tester.h" |
| 23 | |
| 24 | |
| 25 | #if XNN_ARCH_ARM64 |
| 26 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_eq_4) { |
| 27 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 28 | GemmMicrokernelTester() |
| 29 | .mr(1) |
| 30 | .nr(8) |
| 31 | .kr(1) |
| 32 | .sr(1) |
| 33 | .m(1) |
| 34 | .n(8) |
| 35 | .k(4) |
| 36 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 37 | } |
| 38 | |
| 39 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, strided_cn) { |
| 40 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 41 | GemmMicrokernelTester() |
| 42 | .mr(1) |
| 43 | .nr(8) |
| 44 | .kr(1) |
| 45 | .sr(1) |
| 46 | .m(1) |
| 47 | .n(8) |
| 48 | .k(4) |
| 49 | .cn_stride(11) |
| 50 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 51 | } |
| 52 | |
| 53 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 54 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 55 | for (uint32_t m = 1; m <= 1; m++) { |
| 56 | for (uint32_t n = 1; n <= 8; n++) { |
| 57 | GemmMicrokernelTester() |
| 58 | .mr(1) |
| 59 | .nr(8) |
| 60 | .kr(1) |
| 61 | .sr(1) |
| 62 | .m(m) |
| 63 | .n(n) |
| 64 | .k(4) |
| 65 | .iterations(1) |
| 66 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 67 | } |
| 68 | } |
| 69 | } |
| 70 | |
| 71 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 72 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 73 | for (uint32_t m = 1; m <= 1; m++) { |
| 74 | GemmMicrokernelTester() |
| 75 | .mr(1) |
| 76 | .nr(8) |
| 77 | .kr(1) |
| 78 | .sr(1) |
| 79 | .m(m) |
| 80 | .n(8) |
| 81 | .k(4) |
| 82 | .iterations(1) |
| 83 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 84 | } |
| 85 | } |
| 86 | |
| 87 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 88 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 89 | for (uint32_t n = 1; n <= 8; n++) { |
| 90 | GemmMicrokernelTester() |
| 91 | .mr(1) |
| 92 | .nr(8) |
| 93 | .kr(1) |
| 94 | .sr(1) |
| 95 | .m(1) |
| 96 | .n(n) |
| 97 | .k(4) |
| 98 | .iterations(1) |
| 99 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 100 | } |
| 101 | } |
| 102 | |
| 103 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_lt_4) { |
| 104 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 105 | for (size_t k = 1; k < 4; k++) { |
| 106 | GemmMicrokernelTester() |
| 107 | .mr(1) |
| 108 | .nr(8) |
| 109 | .kr(1) |
| 110 | .sr(1) |
| 111 | .m(1) |
| 112 | .n(8) |
| 113 | .k(k) |
| 114 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 119 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 120 | for (size_t k = 1; k < 4; k++) { |
| 121 | for (uint32_t m = 1; m <= 1; m++) { |
| 122 | for (uint32_t n = 1; n <= 8; n++) { |
| 123 | GemmMicrokernelTester() |
| 124 | .mr(1) |
| 125 | .nr(8) |
| 126 | .kr(1) |
| 127 | .sr(1) |
| 128 | .m(m) |
| 129 | .n(n) |
| 130 | .k(k) |
| 131 | .iterations(1) |
| 132 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 133 | } |
| 134 | } |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_gt_4) { |
| 139 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 140 | for (size_t k = 5; k < 8; k++) { |
| 141 | GemmMicrokernelTester() |
| 142 | .mr(1) |
| 143 | .nr(8) |
| 144 | .kr(1) |
| 145 | .sr(1) |
| 146 | .m(1) |
| 147 | .n(8) |
| 148 | .k(k) |
| 149 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 150 | } |
| 151 | } |
| 152 | |
| 153 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 154 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 155 | for (size_t k = 5; k < 8; k++) { |
| 156 | for (uint32_t m = 1; m <= 1; m++) { |
| 157 | for (uint32_t n = 1; n <= 8; n++) { |
| 158 | GemmMicrokernelTester() |
| 159 | .mr(1) |
| 160 | .nr(8) |
| 161 | .kr(1) |
| 162 | .sr(1) |
| 163 | .m(m) |
| 164 | .n(n) |
| 165 | .k(k) |
| 166 | .iterations(1) |
| 167 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 168 | } |
| 169 | } |
| 170 | } |
| 171 | } |
| 172 | |
| 173 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_div_4) { |
| 174 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 175 | for (size_t k = 8; k <= 40; k += 4) { |
| 176 | GemmMicrokernelTester() |
| 177 | .mr(1) |
| 178 | .nr(8) |
| 179 | .kr(1) |
| 180 | .sr(1) |
| 181 | .m(1) |
| 182 | .n(8) |
| 183 | .k(k) |
| 184 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 185 | } |
| 186 | } |
| 187 | |
| 188 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 189 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 190 | for (size_t k = 8; k <= 40; k += 4) { |
| 191 | for (uint32_t m = 1; m <= 1; m++) { |
| 192 | for (uint32_t n = 1; n <= 8; n++) { |
| 193 | GemmMicrokernelTester() |
| 194 | .mr(1) |
| 195 | .nr(8) |
| 196 | .kr(1) |
| 197 | .sr(1) |
| 198 | .m(m) |
| 199 | .n(n) |
| 200 | .k(k) |
| 201 | .iterations(1) |
| 202 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 203 | } |
| 204 | } |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, n_gt_8) { |
| 209 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 210 | for (uint32_t n = 9; n < 16; n++) { |
| 211 | for (size_t k = 1; k <= 20; k += 5) { |
| 212 | GemmMicrokernelTester() |
| 213 | .mr(1) |
| 214 | .nr(8) |
| 215 | .kr(1) |
| 216 | .sr(1) |
| 217 | .m(1) |
| 218 | .n(8) |
| 219 | .k(k) |
| 220 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 221 | } |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
| 226 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 227 | for (uint32_t n = 9; n < 16; n++) { |
| 228 | for (size_t k = 1; k <= 20; k += 5) { |
| 229 | GemmMicrokernelTester() |
| 230 | .mr(1) |
| 231 | .nr(8) |
| 232 | .kr(1) |
| 233 | .sr(1) |
| 234 | .m(1) |
| 235 | .n(8) |
| 236 | .k(k) |
| 237 | .cn_stride(11) |
| 238 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 239 | } |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
| 244 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 245 | for (uint32_t n = 9; n < 16; n++) { |
| 246 | for (size_t k = 1; k <= 20; k += 5) { |
| 247 | for (uint32_t m = 1; m <= 1; m++) { |
| 248 | GemmMicrokernelTester() |
| 249 | .mr(1) |
| 250 | .nr(8) |
| 251 | .kr(1) |
| 252 | .sr(1) |
| 253 | .m(m) |
| 254 | .n(n) |
| 255 | .k(k) |
| 256 | .iterations(1) |
| 257 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 258 | } |
| 259 | } |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, n_div_8) { |
| 264 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 265 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 266 | for (size_t k = 1; k <= 20; k += 5) { |
| 267 | GemmMicrokernelTester() |
| 268 | .mr(1) |
| 269 | .nr(8) |
| 270 | .kr(1) |
| 271 | .sr(1) |
| 272 | .m(1) |
| 273 | .n(8) |
| 274 | .k(k) |
| 275 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 276 | } |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
| 281 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 282 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 283 | for (size_t k = 1; k <= 20; k += 5) { |
| 284 | GemmMicrokernelTester() |
| 285 | .mr(1) |
| 286 | .nr(8) |
| 287 | .kr(1) |
| 288 | .sr(1) |
| 289 | .m(1) |
| 290 | .n(n) |
| 291 | .k(k) |
| 292 | .cn_stride(11) |
| 293 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 294 | } |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
| 299 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 300 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 301 | for (size_t k = 1; k <= 20; k += 5) { |
| 302 | for (uint32_t m = 1; m <= 1; m++) { |
| 303 | GemmMicrokernelTester() |
| 304 | .mr(1) |
| 305 | .nr(8) |
| 306 | .kr(1) |
| 307 | .sr(1) |
| 308 | .m(m) |
| 309 | .n(n) |
| 310 | .k(k) |
| 311 | .iterations(1) |
| 312 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 313 | } |
| 314 | } |
| 315 | } |
| 316 | } |
| 317 | |
| 318 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, small_kernel) { |
| 319 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 320 | for (size_t k = 1; k <= 20; k += 5) { |
| 321 | GemmMicrokernelTester() |
| 322 | .mr(1) |
| 323 | .nr(8) |
| 324 | .kr(1) |
| 325 | .sr(1) |
| 326 | .m(1) |
| 327 | .n(8) |
| 328 | .k(k) |
| 329 | .ks(3) |
| 330 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 331 | } |
| 332 | } |
| 333 | |
| 334 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, small_kernel_subtile) { |
| 335 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 336 | for (size_t k = 1; k <= 20; k += 5) { |
| 337 | for (uint32_t m = 1; m <= 1; m++) { |
| 338 | for (uint32_t n = 1; n <= 8; n++) { |
| 339 | GemmMicrokernelTester() |
| 340 | .mr(1) |
| 341 | .nr(8) |
| 342 | .kr(1) |
| 343 | .sr(1) |
| 344 | .m(m) |
| 345 | .n(n) |
| 346 | .k(k) |
| 347 | .ks(3) |
| 348 | .iterations(1) |
| 349 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 350 | } |
| 351 | } |
| 352 | } |
| 353 | } |
| 354 | |
| 355 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, n_gt_8_small_kernel) { |
| 356 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 357 | for (uint32_t n = 9; n < 16; n++) { |
| 358 | for (size_t k = 1; k <= 20; k += 5) { |
| 359 | GemmMicrokernelTester() |
| 360 | .mr(1) |
| 361 | .nr(8) |
| 362 | .kr(1) |
| 363 | .sr(1) |
| 364 | .m(1) |
| 365 | .n(8) |
| 366 | .k(k) |
| 367 | .ks(3) |
| 368 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 369 | } |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, n_div_8_small_kernel) { |
| 374 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 375 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 376 | for (size_t k = 1; k <= 20; k += 5) { |
| 377 | GemmMicrokernelTester() |
| 378 | .mr(1) |
| 379 | .nr(8) |
| 380 | .kr(1) |
| 381 | .sr(1) |
| 382 | .m(1) |
| 383 | .n(8) |
| 384 | .k(k) |
| 385 | .ks(3) |
| 386 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 387 | } |
| 388 | } |
| 389 | } |
| 390 | |
| 391 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 392 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 393 | for (size_t k = 1; k <= 20; k += 5) { |
| 394 | for (uint32_t m = 1; m <= 1; m++) { |
| 395 | for (uint32_t n = 1; n <= 8; n++) { |
| 396 | GemmMicrokernelTester() |
| 397 | .mr(1) |
| 398 | .nr(8) |
| 399 | .kr(1) |
| 400 | .sr(1) |
| 401 | .m(m) |
| 402 | .n(n) |
| 403 | .k(k) |
| 404 | .cm_stride(11) |
| 405 | .iterations(1) |
| 406 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 407 | } |
| 408 | } |
| 409 | } |
| 410 | } |
| 411 | |
| 412 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, a_offset) { |
| 413 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 414 | for (size_t k = 1; k <= 20; k += 5) { |
| 415 | GemmMicrokernelTester() |
| 416 | .mr(1) |
| 417 | .nr(8) |
| 418 | .kr(1) |
| 419 | .sr(1) |
| 420 | .m(1) |
| 421 | .n(8) |
| 422 | .k(k) |
| 423 | .ks(3) |
| 424 | .a_offset(23) |
| 425 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 426 | } |
| 427 | } |
| 428 | |
| 429 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, zero) { |
| 430 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 431 | for (uint32_t mz = 0; mz < 1; mz++) { |
| 432 | for (size_t k = 1; k <= 20; k += 5) { |
| 433 | GemmMicrokernelTester() |
| 434 | .mr(1) |
| 435 | .nr(8) |
| 436 | .kr(1) |
| 437 | .sr(1) |
| 438 | .m(1) |
| 439 | .n(8) |
| 440 | .k(k) |
| 441 | .ks(3) |
| 442 | .a_offset(23) |
| 443 | .zero_index(mz) |
| 444 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 445 | } |
| 446 | } |
| 447 | } |
| 448 | |
| 449 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, qmin) { |
| 450 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 451 | GemmMicrokernelTester() |
| 452 | .mr(1) |
| 453 | .nr(8) |
| 454 | .kr(1) |
| 455 | .sr(1) |
| 456 | .m(1) |
| 457 | .n(8) |
| 458 | .k(4) |
| 459 | .qmin(128) |
| 460 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 461 | } |
| 462 | |
| 463 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, qmax) { |
| 464 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 465 | GemmMicrokernelTester() |
| 466 | .mr(1) |
| 467 | .nr(8) |
| 468 | .kr(1) |
| 469 | .sr(1) |
| 470 | .m(1) |
| 471 | .n(8) |
| 472 | .k(4) |
| 473 | .qmax(128) |
| 474 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 475 | } |
| 476 | |
| 477 | TEST(F16_IGEMM_MINMAX_1X8__NEONFP16ARITH_LD64, strided_cm) { |
| 478 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 479 | GemmMicrokernelTester() |
| 480 | .mr(1) |
| 481 | .nr(8) |
| 482 | .kr(1) |
| 483 | .sr(1) |
| 484 | .m(1) |
| 485 | .n(8) |
| 486 | .k(4) |
| 487 | .cm_stride(11) |
| 488 | .Test(xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64); |
| 489 | } |
| 490 | #endif // XNN_ARCH_ARM64 |
| 491 | |
| 492 | |
| 493 | #if XNN_ARCH_ARM64 |
| 494 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_eq_4) { |
| 495 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 496 | GemmMicrokernelTester() |
| 497 | .mr(4) |
| 498 | .nr(8) |
| 499 | .kr(1) |
| 500 | .sr(1) |
| 501 | .m(4) |
| 502 | .n(8) |
| 503 | .k(4) |
| 504 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 505 | } |
| 506 | |
| 507 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, strided_cn) { |
| 508 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 509 | GemmMicrokernelTester() |
| 510 | .mr(4) |
| 511 | .nr(8) |
| 512 | .kr(1) |
| 513 | .sr(1) |
| 514 | .m(4) |
| 515 | .n(8) |
| 516 | .k(4) |
| 517 | .cn_stride(11) |
| 518 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 519 | } |
| 520 | |
| 521 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 522 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 523 | for (uint32_t m = 1; m <= 4; m++) { |
| 524 | for (uint32_t n = 1; n <= 8; n++) { |
| 525 | GemmMicrokernelTester() |
| 526 | .mr(4) |
| 527 | .nr(8) |
| 528 | .kr(1) |
| 529 | .sr(1) |
| 530 | .m(m) |
| 531 | .n(n) |
| 532 | .k(4) |
| 533 | .iterations(1) |
| 534 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 535 | } |
| 536 | } |
| 537 | } |
| 538 | |
| 539 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 540 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 541 | for (uint32_t m = 1; m <= 4; m++) { |
| 542 | GemmMicrokernelTester() |
| 543 | .mr(4) |
| 544 | .nr(8) |
| 545 | .kr(1) |
| 546 | .sr(1) |
| 547 | .m(m) |
| 548 | .n(8) |
| 549 | .k(4) |
| 550 | .iterations(1) |
| 551 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 552 | } |
| 553 | } |
| 554 | |
| 555 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 556 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 557 | for (uint32_t n = 1; n <= 8; n++) { |
| 558 | GemmMicrokernelTester() |
| 559 | .mr(4) |
| 560 | .nr(8) |
| 561 | .kr(1) |
| 562 | .sr(1) |
| 563 | .m(4) |
| 564 | .n(n) |
| 565 | .k(4) |
| 566 | .iterations(1) |
| 567 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 568 | } |
| 569 | } |
| 570 | |
| 571 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_lt_4) { |
| 572 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 573 | for (size_t k = 1; k < 4; k++) { |
| 574 | GemmMicrokernelTester() |
| 575 | .mr(4) |
| 576 | .nr(8) |
| 577 | .kr(1) |
| 578 | .sr(1) |
| 579 | .m(4) |
| 580 | .n(8) |
| 581 | .k(k) |
| 582 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 583 | } |
| 584 | } |
| 585 | |
| 586 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 587 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 588 | for (size_t k = 1; k < 4; k++) { |
| 589 | for (uint32_t m = 1; m <= 4; m++) { |
| 590 | for (uint32_t n = 1; n <= 8; n++) { |
| 591 | GemmMicrokernelTester() |
| 592 | .mr(4) |
| 593 | .nr(8) |
| 594 | .kr(1) |
| 595 | .sr(1) |
| 596 | .m(m) |
| 597 | .n(n) |
| 598 | .k(k) |
| 599 | .iterations(1) |
| 600 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 601 | } |
| 602 | } |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_gt_4) { |
| 607 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 608 | for (size_t k = 5; k < 8; k++) { |
| 609 | GemmMicrokernelTester() |
| 610 | .mr(4) |
| 611 | .nr(8) |
| 612 | .kr(1) |
| 613 | .sr(1) |
| 614 | .m(4) |
| 615 | .n(8) |
| 616 | .k(k) |
| 617 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 618 | } |
| 619 | } |
| 620 | |
| 621 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 622 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 623 | for (size_t k = 5; k < 8; k++) { |
| 624 | for (uint32_t m = 1; m <= 4; m++) { |
| 625 | for (uint32_t n = 1; n <= 8; n++) { |
| 626 | GemmMicrokernelTester() |
| 627 | .mr(4) |
| 628 | .nr(8) |
| 629 | .kr(1) |
| 630 | .sr(1) |
| 631 | .m(m) |
| 632 | .n(n) |
| 633 | .k(k) |
| 634 | .iterations(1) |
| 635 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 636 | } |
| 637 | } |
| 638 | } |
| 639 | } |
| 640 | |
| 641 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_div_4) { |
| 642 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 643 | for (size_t k = 8; k <= 40; k += 4) { |
| 644 | GemmMicrokernelTester() |
| 645 | .mr(4) |
| 646 | .nr(8) |
| 647 | .kr(1) |
| 648 | .sr(1) |
| 649 | .m(4) |
| 650 | .n(8) |
| 651 | .k(k) |
| 652 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 653 | } |
| 654 | } |
| 655 | |
| 656 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 657 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 658 | for (size_t k = 8; k <= 40; k += 4) { |
| 659 | for (uint32_t m = 1; m <= 4; m++) { |
| 660 | for (uint32_t n = 1; n <= 8; n++) { |
| 661 | GemmMicrokernelTester() |
| 662 | .mr(4) |
| 663 | .nr(8) |
| 664 | .kr(1) |
| 665 | .sr(1) |
| 666 | .m(m) |
| 667 | .n(n) |
| 668 | .k(k) |
| 669 | .iterations(1) |
| 670 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 671 | } |
| 672 | } |
| 673 | } |
| 674 | } |
| 675 | |
| 676 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_gt_8) { |
| 677 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 678 | for (uint32_t n = 9; n < 16; n++) { |
| 679 | for (size_t k = 1; k <= 20; k += 5) { |
| 680 | GemmMicrokernelTester() |
| 681 | .mr(4) |
| 682 | .nr(8) |
| 683 | .kr(1) |
| 684 | .sr(1) |
| 685 | .m(4) |
| 686 | .n(8) |
| 687 | .k(k) |
| 688 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 689 | } |
| 690 | } |
| 691 | } |
| 692 | |
| 693 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
| 694 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 695 | for (uint32_t n = 9; n < 16; n++) { |
| 696 | for (size_t k = 1; k <= 20; k += 5) { |
| 697 | GemmMicrokernelTester() |
| 698 | .mr(4) |
| 699 | .nr(8) |
| 700 | .kr(1) |
| 701 | .sr(1) |
| 702 | .m(4) |
| 703 | .n(8) |
| 704 | .k(k) |
| 705 | .cn_stride(11) |
| 706 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 707 | } |
| 708 | } |
| 709 | } |
| 710 | |
| 711 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
| 712 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 713 | for (uint32_t n = 9; n < 16; n++) { |
| 714 | for (size_t k = 1; k <= 20; k += 5) { |
| 715 | for (uint32_t m = 1; m <= 4; m++) { |
| 716 | GemmMicrokernelTester() |
| 717 | .mr(4) |
| 718 | .nr(8) |
| 719 | .kr(1) |
| 720 | .sr(1) |
| 721 | .m(m) |
| 722 | .n(n) |
| 723 | .k(k) |
| 724 | .iterations(1) |
| 725 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 726 | } |
| 727 | } |
| 728 | } |
| 729 | } |
| 730 | |
| 731 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_div_8) { |
| 732 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 733 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 734 | for (size_t k = 1; k <= 20; k += 5) { |
| 735 | GemmMicrokernelTester() |
| 736 | .mr(4) |
| 737 | .nr(8) |
| 738 | .kr(1) |
| 739 | .sr(1) |
| 740 | .m(4) |
| 741 | .n(8) |
| 742 | .k(k) |
| 743 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 744 | } |
| 745 | } |
| 746 | } |
| 747 | |
| 748 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
| 749 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 750 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 751 | for (size_t k = 1; k <= 20; k += 5) { |
| 752 | GemmMicrokernelTester() |
| 753 | .mr(4) |
| 754 | .nr(8) |
| 755 | .kr(1) |
| 756 | .sr(1) |
| 757 | .m(4) |
| 758 | .n(n) |
| 759 | .k(k) |
| 760 | .cn_stride(11) |
| 761 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 762 | } |
| 763 | } |
| 764 | } |
| 765 | |
| 766 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
| 767 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 768 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 769 | for (size_t k = 1; k <= 20; k += 5) { |
| 770 | for (uint32_t m = 1; m <= 4; m++) { |
| 771 | GemmMicrokernelTester() |
| 772 | .mr(4) |
| 773 | .nr(8) |
| 774 | .kr(1) |
| 775 | .sr(1) |
| 776 | .m(m) |
| 777 | .n(n) |
| 778 | .k(k) |
| 779 | .iterations(1) |
| 780 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 781 | } |
| 782 | } |
| 783 | } |
| 784 | } |
| 785 | |
| 786 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, small_kernel) { |
| 787 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 788 | for (size_t k = 1; k <= 20; k += 5) { |
| 789 | GemmMicrokernelTester() |
| 790 | .mr(4) |
| 791 | .nr(8) |
| 792 | .kr(1) |
| 793 | .sr(1) |
| 794 | .m(4) |
| 795 | .n(8) |
| 796 | .k(k) |
| 797 | .ks(3) |
| 798 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 799 | } |
| 800 | } |
| 801 | |
| 802 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, small_kernel_subtile) { |
| 803 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 804 | for (size_t k = 1; k <= 20; k += 5) { |
| 805 | for (uint32_t m = 1; m <= 4; m++) { |
| 806 | for (uint32_t n = 1; n <= 8; n++) { |
| 807 | GemmMicrokernelTester() |
| 808 | .mr(4) |
| 809 | .nr(8) |
| 810 | .kr(1) |
| 811 | .sr(1) |
| 812 | .m(m) |
| 813 | .n(n) |
| 814 | .k(k) |
| 815 | .ks(3) |
| 816 | .iterations(1) |
| 817 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 818 | } |
| 819 | } |
| 820 | } |
| 821 | } |
| 822 | |
| 823 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_gt_8_small_kernel) { |
| 824 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 825 | for (uint32_t n = 9; n < 16; n++) { |
| 826 | for (size_t k = 1; k <= 20; k += 5) { |
| 827 | GemmMicrokernelTester() |
| 828 | .mr(4) |
| 829 | .nr(8) |
| 830 | .kr(1) |
| 831 | .sr(1) |
| 832 | .m(4) |
| 833 | .n(8) |
| 834 | .k(k) |
| 835 | .ks(3) |
| 836 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 837 | } |
| 838 | } |
| 839 | } |
| 840 | |
| 841 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_div_8_small_kernel) { |
| 842 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 843 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 844 | for (size_t k = 1; k <= 20; k += 5) { |
| 845 | GemmMicrokernelTester() |
| 846 | .mr(4) |
| 847 | .nr(8) |
| 848 | .kr(1) |
| 849 | .sr(1) |
| 850 | .m(4) |
| 851 | .n(8) |
| 852 | .k(k) |
| 853 | .ks(3) |
| 854 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 855 | } |
| 856 | } |
| 857 | } |
| 858 | |
| 859 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 860 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 861 | for (size_t k = 1; k <= 20; k += 5) { |
| 862 | for (uint32_t m = 1; m <= 4; m++) { |
| 863 | for (uint32_t n = 1; n <= 8; n++) { |
| 864 | GemmMicrokernelTester() |
| 865 | .mr(4) |
| 866 | .nr(8) |
| 867 | .kr(1) |
| 868 | .sr(1) |
| 869 | .m(m) |
| 870 | .n(n) |
| 871 | .k(k) |
| 872 | .cm_stride(11) |
| 873 | .iterations(1) |
| 874 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 875 | } |
| 876 | } |
| 877 | } |
| 878 | } |
| 879 | |
| 880 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, a_offset) { |
| 881 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 882 | for (size_t k = 1; k <= 20; k += 5) { |
| 883 | GemmMicrokernelTester() |
| 884 | .mr(4) |
| 885 | .nr(8) |
| 886 | .kr(1) |
| 887 | .sr(1) |
| 888 | .m(4) |
| 889 | .n(8) |
| 890 | .k(k) |
| 891 | .ks(3) |
| 892 | .a_offset(83) |
| 893 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 894 | } |
| 895 | } |
| 896 | |
| 897 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, zero) { |
| 898 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 899 | for (uint32_t mz = 0; mz < 4; mz++) { |
| 900 | for (size_t k = 1; k <= 20; k += 5) { |
| 901 | GemmMicrokernelTester() |
| 902 | .mr(4) |
| 903 | .nr(8) |
| 904 | .kr(1) |
| 905 | .sr(1) |
| 906 | .m(4) |
| 907 | .n(8) |
| 908 | .k(k) |
| 909 | .ks(3) |
| 910 | .a_offset(83) |
| 911 | .zero_index(mz) |
| 912 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 913 | } |
| 914 | } |
| 915 | } |
| 916 | |
| 917 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, qmin) { |
| 918 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 919 | GemmMicrokernelTester() |
| 920 | .mr(4) |
| 921 | .nr(8) |
| 922 | .kr(1) |
| 923 | .sr(1) |
| 924 | .m(4) |
| 925 | .n(8) |
| 926 | .k(4) |
| 927 | .qmin(128) |
| 928 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 929 | } |
| 930 | |
| 931 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, qmax) { |
| 932 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 933 | GemmMicrokernelTester() |
| 934 | .mr(4) |
| 935 | .nr(8) |
| 936 | .kr(1) |
| 937 | .sr(1) |
| 938 | .m(4) |
| 939 | .n(8) |
| 940 | .k(4) |
| 941 | .qmax(128) |
| 942 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 943 | } |
| 944 | |
| 945 | TEST(F16_IGEMM_MINMAX_4X8__NEONFP16ARITH_LD64, strided_cm) { |
| 946 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 947 | GemmMicrokernelTester() |
| 948 | .mr(4) |
| 949 | .nr(8) |
| 950 | .kr(1) |
| 951 | .sr(1) |
| 952 | .m(4) |
| 953 | .n(8) |
| 954 | .k(4) |
| 955 | .cm_stride(11) |
| 956 | .Test(xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
| 957 | } |
| 958 | #endif // XNN_ARCH_ARM64 |
| 959 | |
| 960 | |
| 961 | #if XNN_ARCH_ARM64 |
| 962 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_eq_4) { |
| 963 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 964 | GemmMicrokernelTester() |
| 965 | .mr(6) |
| 966 | .nr(8) |
| 967 | .kr(1) |
| 968 | .sr(1) |
| 969 | .m(6) |
| 970 | .n(8) |
| 971 | .k(4) |
| 972 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 973 | } |
| 974 | |
| 975 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, strided_cn) { |
| 976 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 977 | GemmMicrokernelTester() |
| 978 | .mr(6) |
| 979 | .nr(8) |
| 980 | .kr(1) |
| 981 | .sr(1) |
| 982 | .m(6) |
| 983 | .n(8) |
| 984 | .k(4) |
| 985 | .cn_stride(11) |
| 986 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 987 | } |
| 988 | |
| 989 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 990 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 991 | for (uint32_t m = 1; m <= 6; m++) { |
| 992 | for (uint32_t n = 1; n <= 8; n++) { |
| 993 | GemmMicrokernelTester() |
| 994 | .mr(6) |
| 995 | .nr(8) |
| 996 | .kr(1) |
| 997 | .sr(1) |
| 998 | .m(m) |
| 999 | .n(n) |
| 1000 | .k(4) |
| 1001 | .iterations(1) |
| 1002 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1003 | } |
| 1004 | } |
| 1005 | } |
| 1006 | |
| 1007 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 1008 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1009 | for (uint32_t m = 1; m <= 6; m++) { |
| 1010 | GemmMicrokernelTester() |
| 1011 | .mr(6) |
| 1012 | .nr(8) |
| 1013 | .kr(1) |
| 1014 | .sr(1) |
| 1015 | .m(m) |
| 1016 | .n(8) |
| 1017 | .k(4) |
| 1018 | .iterations(1) |
| 1019 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1020 | } |
| 1021 | } |
| 1022 | |
| 1023 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 1024 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1025 | for (uint32_t n = 1; n <= 8; n++) { |
| 1026 | GemmMicrokernelTester() |
| 1027 | .mr(6) |
| 1028 | .nr(8) |
| 1029 | .kr(1) |
| 1030 | .sr(1) |
| 1031 | .m(6) |
| 1032 | .n(n) |
| 1033 | .k(4) |
| 1034 | .iterations(1) |
| 1035 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1036 | } |
| 1037 | } |
| 1038 | |
| 1039 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_lt_4) { |
| 1040 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1041 | for (size_t k = 1; k < 4; k++) { |
| 1042 | GemmMicrokernelTester() |
| 1043 | .mr(6) |
| 1044 | .nr(8) |
| 1045 | .kr(1) |
| 1046 | .sr(1) |
| 1047 | .m(6) |
| 1048 | .n(8) |
| 1049 | .k(k) |
| 1050 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1051 | } |
| 1052 | } |
| 1053 | |
| 1054 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 1055 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1056 | for (size_t k = 1; k < 4; k++) { |
| 1057 | for (uint32_t m = 1; m <= 6; m++) { |
| 1058 | for (uint32_t n = 1; n <= 8; n++) { |
| 1059 | GemmMicrokernelTester() |
| 1060 | .mr(6) |
| 1061 | .nr(8) |
| 1062 | .kr(1) |
| 1063 | .sr(1) |
| 1064 | .m(m) |
| 1065 | .n(n) |
| 1066 | .k(k) |
| 1067 | .iterations(1) |
| 1068 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1069 | } |
| 1070 | } |
| 1071 | } |
| 1072 | } |
| 1073 | |
| 1074 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_gt_4) { |
| 1075 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1076 | for (size_t k = 5; k < 8; k++) { |
| 1077 | GemmMicrokernelTester() |
| 1078 | .mr(6) |
| 1079 | .nr(8) |
| 1080 | .kr(1) |
| 1081 | .sr(1) |
| 1082 | .m(6) |
| 1083 | .n(8) |
| 1084 | .k(k) |
| 1085 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1086 | } |
| 1087 | } |
| 1088 | |
| 1089 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 1090 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1091 | for (size_t k = 5; k < 8; k++) { |
| 1092 | for (uint32_t m = 1; m <= 6; m++) { |
| 1093 | for (uint32_t n = 1; n <= 8; n++) { |
| 1094 | GemmMicrokernelTester() |
| 1095 | .mr(6) |
| 1096 | .nr(8) |
| 1097 | .kr(1) |
| 1098 | .sr(1) |
| 1099 | .m(m) |
| 1100 | .n(n) |
| 1101 | .k(k) |
| 1102 | .iterations(1) |
| 1103 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1104 | } |
| 1105 | } |
| 1106 | } |
| 1107 | } |
| 1108 | |
| 1109 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_div_4) { |
| 1110 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1111 | for (size_t k = 8; k <= 40; k += 4) { |
| 1112 | GemmMicrokernelTester() |
| 1113 | .mr(6) |
| 1114 | .nr(8) |
| 1115 | .kr(1) |
| 1116 | .sr(1) |
| 1117 | .m(6) |
| 1118 | .n(8) |
| 1119 | .k(k) |
| 1120 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1121 | } |
| 1122 | } |
| 1123 | |
| 1124 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 1125 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1126 | for (size_t k = 8; k <= 40; k += 4) { |
| 1127 | for (uint32_t m = 1; m <= 6; m++) { |
| 1128 | for (uint32_t n = 1; n <= 8; n++) { |
| 1129 | GemmMicrokernelTester() |
| 1130 | .mr(6) |
| 1131 | .nr(8) |
| 1132 | .kr(1) |
| 1133 | .sr(1) |
| 1134 | .m(m) |
| 1135 | .n(n) |
| 1136 | .k(k) |
| 1137 | .iterations(1) |
| 1138 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1139 | } |
| 1140 | } |
| 1141 | } |
| 1142 | } |
| 1143 | |
| 1144 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_gt_8) { |
| 1145 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1146 | for (uint32_t n = 9; n < 16; n++) { |
| 1147 | for (size_t k = 1; k <= 20; k += 5) { |
| 1148 | GemmMicrokernelTester() |
| 1149 | .mr(6) |
| 1150 | .nr(8) |
| 1151 | .kr(1) |
| 1152 | .sr(1) |
| 1153 | .m(6) |
| 1154 | .n(8) |
| 1155 | .k(k) |
| 1156 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1157 | } |
| 1158 | } |
| 1159 | } |
| 1160 | |
| 1161 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
| 1162 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1163 | for (uint32_t n = 9; n < 16; n++) { |
| 1164 | for (size_t k = 1; k <= 20; k += 5) { |
| 1165 | GemmMicrokernelTester() |
| 1166 | .mr(6) |
| 1167 | .nr(8) |
| 1168 | .kr(1) |
| 1169 | .sr(1) |
| 1170 | .m(6) |
| 1171 | .n(8) |
| 1172 | .k(k) |
| 1173 | .cn_stride(11) |
| 1174 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1175 | } |
| 1176 | } |
| 1177 | } |
| 1178 | |
| 1179 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
| 1180 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1181 | for (uint32_t n = 9; n < 16; n++) { |
| 1182 | for (size_t k = 1; k <= 20; k += 5) { |
| 1183 | for (uint32_t m = 1; m <= 6; m++) { |
| 1184 | GemmMicrokernelTester() |
| 1185 | .mr(6) |
| 1186 | .nr(8) |
| 1187 | .kr(1) |
| 1188 | .sr(1) |
| 1189 | .m(m) |
| 1190 | .n(n) |
| 1191 | .k(k) |
| 1192 | .iterations(1) |
| 1193 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1194 | } |
| 1195 | } |
| 1196 | } |
| 1197 | } |
| 1198 | |
| 1199 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_div_8) { |
| 1200 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1201 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1202 | for (size_t k = 1; k <= 20; k += 5) { |
| 1203 | GemmMicrokernelTester() |
| 1204 | .mr(6) |
| 1205 | .nr(8) |
| 1206 | .kr(1) |
| 1207 | .sr(1) |
| 1208 | .m(6) |
| 1209 | .n(8) |
| 1210 | .k(k) |
| 1211 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1212 | } |
| 1213 | } |
| 1214 | } |
| 1215 | |
| 1216 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
| 1217 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1218 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1219 | for (size_t k = 1; k <= 20; k += 5) { |
| 1220 | GemmMicrokernelTester() |
| 1221 | .mr(6) |
| 1222 | .nr(8) |
| 1223 | .kr(1) |
| 1224 | .sr(1) |
| 1225 | .m(6) |
| 1226 | .n(n) |
| 1227 | .k(k) |
| 1228 | .cn_stride(11) |
| 1229 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1230 | } |
| 1231 | } |
| 1232 | } |
| 1233 | |
| 1234 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
| 1235 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1236 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1237 | for (size_t k = 1; k <= 20; k += 5) { |
| 1238 | for (uint32_t m = 1; m <= 6; m++) { |
| 1239 | GemmMicrokernelTester() |
| 1240 | .mr(6) |
| 1241 | .nr(8) |
| 1242 | .kr(1) |
| 1243 | .sr(1) |
| 1244 | .m(m) |
| 1245 | .n(n) |
| 1246 | .k(k) |
| 1247 | .iterations(1) |
| 1248 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1249 | } |
| 1250 | } |
| 1251 | } |
| 1252 | } |
| 1253 | |
| 1254 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, small_kernel) { |
| 1255 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1256 | for (size_t k = 1; k <= 20; k += 5) { |
| 1257 | GemmMicrokernelTester() |
| 1258 | .mr(6) |
| 1259 | .nr(8) |
| 1260 | .kr(1) |
| 1261 | .sr(1) |
| 1262 | .m(6) |
| 1263 | .n(8) |
| 1264 | .k(k) |
| 1265 | .ks(3) |
| 1266 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1267 | } |
| 1268 | } |
| 1269 | |
| 1270 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, small_kernel_subtile) { |
| 1271 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1272 | for (size_t k = 1; k <= 20; k += 5) { |
| 1273 | for (uint32_t m = 1; m <= 6; m++) { |
| 1274 | for (uint32_t n = 1; n <= 8; n++) { |
| 1275 | GemmMicrokernelTester() |
| 1276 | .mr(6) |
| 1277 | .nr(8) |
| 1278 | .kr(1) |
| 1279 | .sr(1) |
| 1280 | .m(m) |
| 1281 | .n(n) |
| 1282 | .k(k) |
| 1283 | .ks(3) |
| 1284 | .iterations(1) |
| 1285 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1286 | } |
| 1287 | } |
| 1288 | } |
| 1289 | } |
| 1290 | |
| 1291 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_gt_8_small_kernel) { |
| 1292 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1293 | for (uint32_t n = 9; n < 16; n++) { |
| 1294 | for (size_t k = 1; k <= 20; k += 5) { |
| 1295 | GemmMicrokernelTester() |
| 1296 | .mr(6) |
| 1297 | .nr(8) |
| 1298 | .kr(1) |
| 1299 | .sr(1) |
| 1300 | .m(6) |
| 1301 | .n(8) |
| 1302 | .k(k) |
| 1303 | .ks(3) |
| 1304 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1305 | } |
| 1306 | } |
| 1307 | } |
| 1308 | |
| 1309 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_div_8_small_kernel) { |
| 1310 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1311 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1312 | for (size_t k = 1; k <= 20; k += 5) { |
| 1313 | GemmMicrokernelTester() |
| 1314 | .mr(6) |
| 1315 | .nr(8) |
| 1316 | .kr(1) |
| 1317 | .sr(1) |
| 1318 | .m(6) |
| 1319 | .n(8) |
| 1320 | .k(k) |
| 1321 | .ks(3) |
| 1322 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1323 | } |
| 1324 | } |
| 1325 | } |
| 1326 | |
| 1327 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 1328 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1329 | for (size_t k = 1; k <= 20; k += 5) { |
| 1330 | for (uint32_t m = 1; m <= 6; m++) { |
| 1331 | for (uint32_t n = 1; n <= 8; n++) { |
| 1332 | GemmMicrokernelTester() |
| 1333 | .mr(6) |
| 1334 | .nr(8) |
| 1335 | .kr(1) |
| 1336 | .sr(1) |
| 1337 | .m(m) |
| 1338 | .n(n) |
| 1339 | .k(k) |
| 1340 | .cm_stride(11) |
| 1341 | .iterations(1) |
| 1342 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1343 | } |
| 1344 | } |
| 1345 | } |
| 1346 | } |
| 1347 | |
| 1348 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, a_offset) { |
| 1349 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1350 | for (size_t k = 1; k <= 20; k += 5) { |
| 1351 | GemmMicrokernelTester() |
| 1352 | .mr(6) |
| 1353 | .nr(8) |
| 1354 | .kr(1) |
| 1355 | .sr(1) |
| 1356 | .m(6) |
| 1357 | .n(8) |
| 1358 | .k(k) |
| 1359 | .ks(3) |
| 1360 | .a_offset(127) |
| 1361 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1362 | } |
| 1363 | } |
| 1364 | |
| 1365 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, zero) { |
| 1366 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1367 | for (uint32_t mz = 0; mz < 6; mz++) { |
| 1368 | for (size_t k = 1; k <= 20; k += 5) { |
| 1369 | GemmMicrokernelTester() |
| 1370 | .mr(6) |
| 1371 | .nr(8) |
| 1372 | .kr(1) |
| 1373 | .sr(1) |
| 1374 | .m(6) |
| 1375 | .n(8) |
| 1376 | .k(k) |
| 1377 | .ks(3) |
| 1378 | .a_offset(127) |
| 1379 | .zero_index(mz) |
| 1380 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1381 | } |
| 1382 | } |
| 1383 | } |
| 1384 | |
| 1385 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, qmin) { |
| 1386 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1387 | GemmMicrokernelTester() |
| 1388 | .mr(6) |
| 1389 | .nr(8) |
| 1390 | .kr(1) |
| 1391 | .sr(1) |
| 1392 | .m(6) |
| 1393 | .n(8) |
| 1394 | .k(4) |
| 1395 | .qmin(128) |
| 1396 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1397 | } |
| 1398 | |
| 1399 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, qmax) { |
| 1400 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1401 | GemmMicrokernelTester() |
| 1402 | .mr(6) |
| 1403 | .nr(8) |
| 1404 | .kr(1) |
| 1405 | .sr(1) |
| 1406 | .m(6) |
| 1407 | .n(8) |
| 1408 | .k(4) |
| 1409 | .qmax(128) |
| 1410 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1411 | } |
| 1412 | |
| 1413 | TEST(F16_IGEMM_MINMAX_6X8__NEONFP16ARITH_LD64, strided_cm) { |
| 1414 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1415 | GemmMicrokernelTester() |
| 1416 | .mr(6) |
| 1417 | .nr(8) |
| 1418 | .kr(1) |
| 1419 | .sr(1) |
| 1420 | .m(6) |
| 1421 | .n(8) |
| 1422 | .k(4) |
| 1423 | .cm_stride(11) |
| 1424 | .Test(xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
| 1425 | } |
| 1426 | #endif // XNN_ARCH_ARM64 |
| 1427 | |
| 1428 | |
| 1429 | #if XNN_ARCH_ARM64 |
| 1430 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_eq_4) { |
| 1431 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1432 | GemmMicrokernelTester() |
| 1433 | .mr(8) |
| 1434 | .nr(8) |
| 1435 | .kr(1) |
| 1436 | .sr(1) |
| 1437 | .m(8) |
| 1438 | .n(8) |
| 1439 | .k(4) |
| 1440 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1441 | } |
| 1442 | |
| 1443 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, strided_cn) { |
| 1444 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1445 | GemmMicrokernelTester() |
| 1446 | .mr(8) |
| 1447 | .nr(8) |
| 1448 | .kr(1) |
| 1449 | .sr(1) |
| 1450 | .m(8) |
| 1451 | .n(8) |
| 1452 | .k(4) |
| 1453 | .cn_stride(11) |
| 1454 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1455 | } |
| 1456 | |
| 1457 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 1458 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1459 | for (uint32_t m = 1; m <= 8; m++) { |
| 1460 | for (uint32_t n = 1; n <= 8; n++) { |
| 1461 | GemmMicrokernelTester() |
| 1462 | .mr(8) |
| 1463 | .nr(8) |
| 1464 | .kr(1) |
| 1465 | .sr(1) |
| 1466 | .m(m) |
| 1467 | .n(n) |
| 1468 | .k(4) |
| 1469 | .iterations(1) |
| 1470 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1471 | } |
| 1472 | } |
| 1473 | } |
| 1474 | |
| 1475 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 1476 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1477 | for (uint32_t m = 1; m <= 8; m++) { |
| 1478 | GemmMicrokernelTester() |
| 1479 | .mr(8) |
| 1480 | .nr(8) |
| 1481 | .kr(1) |
| 1482 | .sr(1) |
| 1483 | .m(m) |
| 1484 | .n(8) |
| 1485 | .k(4) |
| 1486 | .iterations(1) |
| 1487 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1488 | } |
| 1489 | } |
| 1490 | |
| 1491 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 1492 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1493 | for (uint32_t n = 1; n <= 8; n++) { |
| 1494 | GemmMicrokernelTester() |
| 1495 | .mr(8) |
| 1496 | .nr(8) |
| 1497 | .kr(1) |
| 1498 | .sr(1) |
| 1499 | .m(8) |
| 1500 | .n(n) |
| 1501 | .k(4) |
| 1502 | .iterations(1) |
| 1503 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1504 | } |
| 1505 | } |
| 1506 | |
| 1507 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_lt_4) { |
| 1508 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1509 | for (size_t k = 1; k < 4; k++) { |
| 1510 | GemmMicrokernelTester() |
| 1511 | .mr(8) |
| 1512 | .nr(8) |
| 1513 | .kr(1) |
| 1514 | .sr(1) |
| 1515 | .m(8) |
| 1516 | .n(8) |
| 1517 | .k(k) |
| 1518 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1519 | } |
| 1520 | } |
| 1521 | |
| 1522 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 1523 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1524 | for (size_t k = 1; k < 4; k++) { |
| 1525 | for (uint32_t m = 1; m <= 8; m++) { |
| 1526 | for (uint32_t n = 1; n <= 8; n++) { |
| 1527 | GemmMicrokernelTester() |
| 1528 | .mr(8) |
| 1529 | .nr(8) |
| 1530 | .kr(1) |
| 1531 | .sr(1) |
| 1532 | .m(m) |
| 1533 | .n(n) |
| 1534 | .k(k) |
| 1535 | .iterations(1) |
| 1536 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1537 | } |
| 1538 | } |
| 1539 | } |
| 1540 | } |
| 1541 | |
| 1542 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_gt_4) { |
| 1543 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1544 | for (size_t k = 5; k < 8; k++) { |
| 1545 | GemmMicrokernelTester() |
| 1546 | .mr(8) |
| 1547 | .nr(8) |
| 1548 | .kr(1) |
| 1549 | .sr(1) |
| 1550 | .m(8) |
| 1551 | .n(8) |
| 1552 | .k(k) |
| 1553 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1554 | } |
| 1555 | } |
| 1556 | |
| 1557 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 1558 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1559 | for (size_t k = 5; k < 8; k++) { |
| 1560 | for (uint32_t m = 1; m <= 8; m++) { |
| 1561 | for (uint32_t n = 1; n <= 8; n++) { |
| 1562 | GemmMicrokernelTester() |
| 1563 | .mr(8) |
| 1564 | .nr(8) |
| 1565 | .kr(1) |
| 1566 | .sr(1) |
| 1567 | .m(m) |
| 1568 | .n(n) |
| 1569 | .k(k) |
| 1570 | .iterations(1) |
| 1571 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1572 | } |
| 1573 | } |
| 1574 | } |
| 1575 | } |
| 1576 | |
| 1577 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_div_4) { |
| 1578 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1579 | for (size_t k = 8; k <= 40; k += 4) { |
| 1580 | GemmMicrokernelTester() |
| 1581 | .mr(8) |
| 1582 | .nr(8) |
| 1583 | .kr(1) |
| 1584 | .sr(1) |
| 1585 | .m(8) |
| 1586 | .n(8) |
| 1587 | .k(k) |
| 1588 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1589 | } |
| 1590 | } |
| 1591 | |
| 1592 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 1593 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1594 | for (size_t k = 8; k <= 40; k += 4) { |
| 1595 | for (uint32_t m = 1; m <= 8; m++) { |
| 1596 | for (uint32_t n = 1; n <= 8; n++) { |
| 1597 | GemmMicrokernelTester() |
| 1598 | .mr(8) |
| 1599 | .nr(8) |
| 1600 | .kr(1) |
| 1601 | .sr(1) |
| 1602 | .m(m) |
| 1603 | .n(n) |
| 1604 | .k(k) |
| 1605 | .iterations(1) |
| 1606 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1607 | } |
| 1608 | } |
| 1609 | } |
| 1610 | } |
| 1611 | |
| 1612 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_gt_8) { |
| 1613 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1614 | for (uint32_t n = 9; n < 16; n++) { |
| 1615 | for (size_t k = 1; k <= 20; k += 5) { |
| 1616 | GemmMicrokernelTester() |
| 1617 | .mr(8) |
| 1618 | .nr(8) |
| 1619 | .kr(1) |
| 1620 | .sr(1) |
| 1621 | .m(8) |
| 1622 | .n(8) |
| 1623 | .k(k) |
| 1624 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1625 | } |
| 1626 | } |
| 1627 | } |
| 1628 | |
| 1629 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
| 1630 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1631 | for (uint32_t n = 9; n < 16; n++) { |
| 1632 | for (size_t k = 1; k <= 20; k += 5) { |
| 1633 | GemmMicrokernelTester() |
| 1634 | .mr(8) |
| 1635 | .nr(8) |
| 1636 | .kr(1) |
| 1637 | .sr(1) |
| 1638 | .m(8) |
| 1639 | .n(8) |
| 1640 | .k(k) |
| 1641 | .cn_stride(11) |
| 1642 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1643 | } |
| 1644 | } |
| 1645 | } |
| 1646 | |
| 1647 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
| 1648 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1649 | for (uint32_t n = 9; n < 16; n++) { |
| 1650 | for (size_t k = 1; k <= 20; k += 5) { |
| 1651 | for (uint32_t m = 1; m <= 8; m++) { |
| 1652 | GemmMicrokernelTester() |
| 1653 | .mr(8) |
| 1654 | .nr(8) |
| 1655 | .kr(1) |
| 1656 | .sr(1) |
| 1657 | .m(m) |
| 1658 | .n(n) |
| 1659 | .k(k) |
| 1660 | .iterations(1) |
| 1661 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1662 | } |
| 1663 | } |
| 1664 | } |
| 1665 | } |
| 1666 | |
| 1667 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_div_8) { |
| 1668 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1669 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1670 | for (size_t k = 1; k <= 20; k += 5) { |
| 1671 | GemmMicrokernelTester() |
| 1672 | .mr(8) |
| 1673 | .nr(8) |
| 1674 | .kr(1) |
| 1675 | .sr(1) |
| 1676 | .m(8) |
| 1677 | .n(8) |
| 1678 | .k(k) |
| 1679 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1680 | } |
| 1681 | } |
| 1682 | } |
| 1683 | |
| 1684 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
| 1685 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1686 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1687 | for (size_t k = 1; k <= 20; k += 5) { |
| 1688 | GemmMicrokernelTester() |
| 1689 | .mr(8) |
| 1690 | .nr(8) |
| 1691 | .kr(1) |
| 1692 | .sr(1) |
| 1693 | .m(8) |
| 1694 | .n(n) |
| 1695 | .k(k) |
| 1696 | .cn_stride(11) |
| 1697 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1698 | } |
| 1699 | } |
| 1700 | } |
| 1701 | |
| 1702 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
| 1703 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1704 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1705 | for (size_t k = 1; k <= 20; k += 5) { |
| 1706 | for (uint32_t m = 1; m <= 8; m++) { |
| 1707 | GemmMicrokernelTester() |
| 1708 | .mr(8) |
| 1709 | .nr(8) |
| 1710 | .kr(1) |
| 1711 | .sr(1) |
| 1712 | .m(m) |
| 1713 | .n(n) |
| 1714 | .k(k) |
| 1715 | .iterations(1) |
| 1716 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1717 | } |
| 1718 | } |
| 1719 | } |
| 1720 | } |
| 1721 | |
| 1722 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, small_kernel) { |
| 1723 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1724 | for (size_t k = 1; k <= 20; k += 5) { |
| 1725 | GemmMicrokernelTester() |
| 1726 | .mr(8) |
| 1727 | .nr(8) |
| 1728 | .kr(1) |
| 1729 | .sr(1) |
| 1730 | .m(8) |
| 1731 | .n(8) |
| 1732 | .k(k) |
| 1733 | .ks(3) |
| 1734 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1735 | } |
| 1736 | } |
| 1737 | |
| 1738 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, small_kernel_subtile) { |
| 1739 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1740 | for (size_t k = 1; k <= 20; k += 5) { |
| 1741 | for (uint32_t m = 1; m <= 8; m++) { |
| 1742 | for (uint32_t n = 1; n <= 8; n++) { |
| 1743 | GemmMicrokernelTester() |
| 1744 | .mr(8) |
| 1745 | .nr(8) |
| 1746 | .kr(1) |
| 1747 | .sr(1) |
| 1748 | .m(m) |
| 1749 | .n(n) |
| 1750 | .k(k) |
| 1751 | .ks(3) |
| 1752 | .iterations(1) |
| 1753 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1754 | } |
| 1755 | } |
| 1756 | } |
| 1757 | } |
| 1758 | |
| 1759 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_gt_8_small_kernel) { |
| 1760 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1761 | for (uint32_t n = 9; n < 16; n++) { |
| 1762 | for (size_t k = 1; k <= 20; k += 5) { |
| 1763 | GemmMicrokernelTester() |
| 1764 | .mr(8) |
| 1765 | .nr(8) |
| 1766 | .kr(1) |
| 1767 | .sr(1) |
| 1768 | .m(8) |
| 1769 | .n(8) |
| 1770 | .k(k) |
| 1771 | .ks(3) |
| 1772 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1773 | } |
| 1774 | } |
| 1775 | } |
| 1776 | |
| 1777 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_div_8_small_kernel) { |
| 1778 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1779 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1780 | for (size_t k = 1; k <= 20; k += 5) { |
| 1781 | GemmMicrokernelTester() |
| 1782 | .mr(8) |
| 1783 | .nr(8) |
| 1784 | .kr(1) |
| 1785 | .sr(1) |
| 1786 | .m(8) |
| 1787 | .n(8) |
| 1788 | .k(k) |
| 1789 | .ks(3) |
| 1790 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1791 | } |
| 1792 | } |
| 1793 | } |
| 1794 | |
| 1795 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 1796 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1797 | for (size_t k = 1; k <= 20; k += 5) { |
| 1798 | for (uint32_t m = 1; m <= 8; m++) { |
| 1799 | for (uint32_t n = 1; n <= 8; n++) { |
| 1800 | GemmMicrokernelTester() |
| 1801 | .mr(8) |
| 1802 | .nr(8) |
| 1803 | .kr(1) |
| 1804 | .sr(1) |
| 1805 | .m(m) |
| 1806 | .n(n) |
| 1807 | .k(k) |
| 1808 | .cm_stride(11) |
| 1809 | .iterations(1) |
| 1810 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1811 | } |
| 1812 | } |
| 1813 | } |
| 1814 | } |
| 1815 | |
| 1816 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, a_offset) { |
| 1817 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1818 | for (size_t k = 1; k <= 20; k += 5) { |
| 1819 | GemmMicrokernelTester() |
| 1820 | .mr(8) |
| 1821 | .nr(8) |
| 1822 | .kr(1) |
| 1823 | .sr(1) |
| 1824 | .m(8) |
| 1825 | .n(8) |
| 1826 | .k(k) |
| 1827 | .ks(3) |
| 1828 | .a_offset(163) |
| 1829 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1830 | } |
| 1831 | } |
| 1832 | |
| 1833 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, zero) { |
| 1834 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1835 | for (uint32_t mz = 0; mz < 8; mz++) { |
| 1836 | for (size_t k = 1; k <= 20; k += 5) { |
| 1837 | GemmMicrokernelTester() |
| 1838 | .mr(8) |
| 1839 | .nr(8) |
| 1840 | .kr(1) |
| 1841 | .sr(1) |
| 1842 | .m(8) |
| 1843 | .n(8) |
| 1844 | .k(k) |
| 1845 | .ks(3) |
| 1846 | .a_offset(163) |
| 1847 | .zero_index(mz) |
| 1848 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1849 | } |
| 1850 | } |
| 1851 | } |
| 1852 | |
| 1853 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, qmin) { |
| 1854 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1855 | GemmMicrokernelTester() |
| 1856 | .mr(8) |
| 1857 | .nr(8) |
| 1858 | .kr(1) |
| 1859 | .sr(1) |
| 1860 | .m(8) |
| 1861 | .n(8) |
| 1862 | .k(4) |
| 1863 | .qmin(128) |
| 1864 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1865 | } |
| 1866 | |
| 1867 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, qmax) { |
| 1868 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1869 | GemmMicrokernelTester() |
| 1870 | .mr(8) |
| 1871 | .nr(8) |
| 1872 | .kr(1) |
| 1873 | .sr(1) |
| 1874 | .m(8) |
| 1875 | .n(8) |
| 1876 | .k(4) |
| 1877 | .qmax(128) |
| 1878 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1879 | } |
| 1880 | |
| 1881 | TEST(F16_IGEMM_MINMAX_8X8__NEONFP16ARITH_LD64, strided_cm) { |
| 1882 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1883 | GemmMicrokernelTester() |
| 1884 | .mr(8) |
| 1885 | .nr(8) |
| 1886 | .kr(1) |
| 1887 | .sr(1) |
| 1888 | .m(8) |
| 1889 | .n(8) |
| 1890 | .k(4) |
| 1891 | .cm_stride(11) |
| 1892 | .Test(xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
| 1893 | } |
| 1894 | #endif // XNN_ARCH_ARM64 |
| Frank Barchard | 3f9f99f | 2020-05-06 01:12:04 -0700 | [diff] [blame] | 1895 | |
| 1896 | |
| 1897 | #if XNN_ARCH_ARM64 |
| 1898 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_eq_4) { |
| 1899 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1900 | GemmMicrokernelTester() |
| 1901 | .mr(1) |
| 1902 | .nr(16) |
| 1903 | .kr(1) |
| 1904 | .sr(1) |
| 1905 | .m(1) |
| 1906 | .n(16) |
| 1907 | .k(4) |
| 1908 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 1909 | } |
| 1910 | |
| 1911 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, strided_cn) { |
| 1912 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1913 | GemmMicrokernelTester() |
| 1914 | .mr(1) |
| 1915 | .nr(16) |
| 1916 | .kr(1) |
| 1917 | .sr(1) |
| 1918 | .m(1) |
| 1919 | .n(16) |
| 1920 | .k(4) |
| 1921 | .cn_stride(19) |
| 1922 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 1923 | } |
| 1924 | |
| 1925 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 1926 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1927 | for (uint32_t m = 1; m <= 1; m++) { |
| 1928 | for (uint32_t n = 1; n <= 16; n++) { |
| 1929 | GemmMicrokernelTester() |
| 1930 | .mr(1) |
| 1931 | .nr(16) |
| 1932 | .kr(1) |
| 1933 | .sr(1) |
| 1934 | .m(m) |
| 1935 | .n(n) |
| 1936 | .k(4) |
| 1937 | .iterations(1) |
| 1938 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 1939 | } |
| 1940 | } |
| 1941 | } |
| 1942 | |
| 1943 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 1944 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1945 | for (uint32_t m = 1; m <= 1; m++) { |
| 1946 | GemmMicrokernelTester() |
| 1947 | .mr(1) |
| 1948 | .nr(16) |
| 1949 | .kr(1) |
| 1950 | .sr(1) |
| 1951 | .m(m) |
| 1952 | .n(16) |
| 1953 | .k(4) |
| 1954 | .iterations(1) |
| 1955 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 1956 | } |
| 1957 | } |
| 1958 | |
| 1959 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 1960 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1961 | for (uint32_t n = 1; n <= 16; n++) { |
| 1962 | GemmMicrokernelTester() |
| 1963 | .mr(1) |
| 1964 | .nr(16) |
| 1965 | .kr(1) |
| 1966 | .sr(1) |
| 1967 | .m(1) |
| 1968 | .n(n) |
| 1969 | .k(4) |
| 1970 | .iterations(1) |
| 1971 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 1972 | } |
| 1973 | } |
| 1974 | |
| 1975 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_lt_4) { |
| 1976 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1977 | for (size_t k = 1; k < 4; k++) { |
| 1978 | GemmMicrokernelTester() |
| 1979 | .mr(1) |
| 1980 | .nr(16) |
| 1981 | .kr(1) |
| 1982 | .sr(1) |
| 1983 | .m(1) |
| 1984 | .n(16) |
| 1985 | .k(k) |
| 1986 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 1987 | } |
| 1988 | } |
| 1989 | |
| 1990 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 1991 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1992 | for (size_t k = 1; k < 4; k++) { |
| 1993 | for (uint32_t m = 1; m <= 1; m++) { |
| 1994 | for (uint32_t n = 1; n <= 16; n++) { |
| 1995 | GemmMicrokernelTester() |
| 1996 | .mr(1) |
| 1997 | .nr(16) |
| 1998 | .kr(1) |
| 1999 | .sr(1) |
| 2000 | .m(m) |
| 2001 | .n(n) |
| 2002 | .k(k) |
| 2003 | .iterations(1) |
| 2004 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2005 | } |
| 2006 | } |
| 2007 | } |
| 2008 | } |
| 2009 | |
| 2010 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_gt_4) { |
| 2011 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2012 | for (size_t k = 5; k < 8; k++) { |
| 2013 | GemmMicrokernelTester() |
| 2014 | .mr(1) |
| 2015 | .nr(16) |
| 2016 | .kr(1) |
| 2017 | .sr(1) |
| 2018 | .m(1) |
| 2019 | .n(16) |
| 2020 | .k(k) |
| 2021 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2022 | } |
| 2023 | } |
| 2024 | |
| 2025 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 2026 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2027 | for (size_t k = 5; k < 8; k++) { |
| 2028 | for (uint32_t m = 1; m <= 1; m++) { |
| 2029 | for (uint32_t n = 1; n <= 16; n++) { |
| 2030 | GemmMicrokernelTester() |
| 2031 | .mr(1) |
| 2032 | .nr(16) |
| 2033 | .kr(1) |
| 2034 | .sr(1) |
| 2035 | .m(m) |
| 2036 | .n(n) |
| 2037 | .k(k) |
| 2038 | .iterations(1) |
| 2039 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2040 | } |
| 2041 | } |
| 2042 | } |
| 2043 | } |
| 2044 | |
| 2045 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_div_4) { |
| 2046 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2047 | for (size_t k = 8; k <= 40; k += 4) { |
| 2048 | GemmMicrokernelTester() |
| 2049 | .mr(1) |
| 2050 | .nr(16) |
| 2051 | .kr(1) |
| 2052 | .sr(1) |
| 2053 | .m(1) |
| 2054 | .n(16) |
| 2055 | .k(k) |
| 2056 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2057 | } |
| 2058 | } |
| 2059 | |
| 2060 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 2061 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2062 | for (size_t k = 8; k <= 40; k += 4) { |
| 2063 | for (uint32_t m = 1; m <= 1; m++) { |
| 2064 | for (uint32_t n = 1; n <= 16; n++) { |
| 2065 | GemmMicrokernelTester() |
| 2066 | .mr(1) |
| 2067 | .nr(16) |
| 2068 | .kr(1) |
| 2069 | .sr(1) |
| 2070 | .m(m) |
| 2071 | .n(n) |
| 2072 | .k(k) |
| 2073 | .iterations(1) |
| 2074 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2075 | } |
| 2076 | } |
| 2077 | } |
| 2078 | } |
| 2079 | |
| 2080 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, n_gt_16) { |
| 2081 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2082 | for (uint32_t n = 17; n < 32; n++) { |
| 2083 | for (size_t k = 1; k <= 20; k += 5) { |
| 2084 | GemmMicrokernelTester() |
| 2085 | .mr(1) |
| 2086 | .nr(16) |
| 2087 | .kr(1) |
| 2088 | .sr(1) |
| 2089 | .m(1) |
| 2090 | .n(16) |
| 2091 | .k(k) |
| 2092 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2093 | } |
| 2094 | } |
| 2095 | } |
| 2096 | |
| 2097 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, n_gt_16_strided_cn) { |
| 2098 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2099 | for (uint32_t n = 17; n < 32; n++) { |
| 2100 | for (size_t k = 1; k <= 20; k += 5) { |
| 2101 | GemmMicrokernelTester() |
| 2102 | .mr(1) |
| 2103 | .nr(16) |
| 2104 | .kr(1) |
| 2105 | .sr(1) |
| 2106 | .m(1) |
| 2107 | .n(16) |
| 2108 | .k(k) |
| 2109 | .cn_stride(19) |
| 2110 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2111 | } |
| 2112 | } |
| 2113 | } |
| 2114 | |
| 2115 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, n_gt_16_subtile) { |
| 2116 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2117 | for (uint32_t n = 17; n < 32; n++) { |
| 2118 | for (size_t k = 1; k <= 20; k += 5) { |
| 2119 | for (uint32_t m = 1; m <= 1; m++) { |
| 2120 | GemmMicrokernelTester() |
| 2121 | .mr(1) |
| 2122 | .nr(16) |
| 2123 | .kr(1) |
| 2124 | .sr(1) |
| 2125 | .m(m) |
| 2126 | .n(n) |
| 2127 | .k(k) |
| 2128 | .iterations(1) |
| 2129 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2130 | } |
| 2131 | } |
| 2132 | } |
| 2133 | } |
| 2134 | |
| 2135 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, n_div_16) { |
| 2136 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2137 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2138 | for (size_t k = 1; k <= 20; k += 5) { |
| 2139 | GemmMicrokernelTester() |
| 2140 | .mr(1) |
| 2141 | .nr(16) |
| 2142 | .kr(1) |
| 2143 | .sr(1) |
| 2144 | .m(1) |
| 2145 | .n(16) |
| 2146 | .k(k) |
| 2147 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2148 | } |
| 2149 | } |
| 2150 | } |
| 2151 | |
| 2152 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, n_div_16_strided_cn) { |
| 2153 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2154 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2155 | for (size_t k = 1; k <= 20; k += 5) { |
| 2156 | GemmMicrokernelTester() |
| 2157 | .mr(1) |
| 2158 | .nr(16) |
| 2159 | .kr(1) |
| 2160 | .sr(1) |
| 2161 | .m(1) |
| 2162 | .n(n) |
| 2163 | .k(k) |
| 2164 | .cn_stride(19) |
| 2165 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2166 | } |
| 2167 | } |
| 2168 | } |
| 2169 | |
| 2170 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, n_div_16_subtile) { |
| 2171 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2172 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2173 | for (size_t k = 1; k <= 20; k += 5) { |
| 2174 | for (uint32_t m = 1; m <= 1; m++) { |
| 2175 | GemmMicrokernelTester() |
| 2176 | .mr(1) |
| 2177 | .nr(16) |
| 2178 | .kr(1) |
| 2179 | .sr(1) |
| 2180 | .m(m) |
| 2181 | .n(n) |
| 2182 | .k(k) |
| 2183 | .iterations(1) |
| 2184 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2185 | } |
| 2186 | } |
| 2187 | } |
| 2188 | } |
| 2189 | |
| 2190 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, small_kernel) { |
| 2191 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2192 | for (size_t k = 1; k <= 20; k += 5) { |
| 2193 | GemmMicrokernelTester() |
| 2194 | .mr(1) |
| 2195 | .nr(16) |
| 2196 | .kr(1) |
| 2197 | .sr(1) |
| 2198 | .m(1) |
| 2199 | .n(16) |
| 2200 | .k(k) |
| 2201 | .ks(3) |
| 2202 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2203 | } |
| 2204 | } |
| 2205 | |
| 2206 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, small_kernel_subtile) { |
| 2207 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2208 | for (size_t k = 1; k <= 20; k += 5) { |
| 2209 | for (uint32_t m = 1; m <= 1; m++) { |
| 2210 | for (uint32_t n = 1; n <= 16; n++) { |
| 2211 | GemmMicrokernelTester() |
| 2212 | .mr(1) |
| 2213 | .nr(16) |
| 2214 | .kr(1) |
| 2215 | .sr(1) |
| 2216 | .m(m) |
| 2217 | .n(n) |
| 2218 | .k(k) |
| 2219 | .ks(3) |
| 2220 | .iterations(1) |
| 2221 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2222 | } |
| 2223 | } |
| 2224 | } |
| 2225 | } |
| 2226 | |
| 2227 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, n_gt_16_small_kernel) { |
| 2228 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2229 | for (uint32_t n = 17; n < 32; n++) { |
| 2230 | for (size_t k = 1; k <= 20; k += 5) { |
| 2231 | GemmMicrokernelTester() |
| 2232 | .mr(1) |
| 2233 | .nr(16) |
| 2234 | .kr(1) |
| 2235 | .sr(1) |
| 2236 | .m(1) |
| 2237 | .n(16) |
| 2238 | .k(k) |
| 2239 | .ks(3) |
| 2240 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2241 | } |
| 2242 | } |
| 2243 | } |
| 2244 | |
| 2245 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, n_div_16_small_kernel) { |
| 2246 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2247 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2248 | for (size_t k = 1; k <= 20; k += 5) { |
| 2249 | GemmMicrokernelTester() |
| 2250 | .mr(1) |
| 2251 | .nr(16) |
| 2252 | .kr(1) |
| 2253 | .sr(1) |
| 2254 | .m(1) |
| 2255 | .n(16) |
| 2256 | .k(k) |
| 2257 | .ks(3) |
| 2258 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2259 | } |
| 2260 | } |
| 2261 | } |
| 2262 | |
| 2263 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 2264 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2265 | for (size_t k = 1; k <= 20; k += 5) { |
| 2266 | for (uint32_t m = 1; m <= 1; m++) { |
| 2267 | for (uint32_t n = 1; n <= 16; n++) { |
| 2268 | GemmMicrokernelTester() |
| 2269 | .mr(1) |
| 2270 | .nr(16) |
| 2271 | .kr(1) |
| 2272 | .sr(1) |
| 2273 | .m(m) |
| 2274 | .n(n) |
| 2275 | .k(k) |
| 2276 | .cm_stride(19) |
| 2277 | .iterations(1) |
| 2278 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2279 | } |
| 2280 | } |
| 2281 | } |
| 2282 | } |
| 2283 | |
| 2284 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, a_offset) { |
| 2285 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2286 | for (size_t k = 1; k <= 20; k += 5) { |
| 2287 | GemmMicrokernelTester() |
| 2288 | .mr(1) |
| 2289 | .nr(16) |
| 2290 | .kr(1) |
| 2291 | .sr(1) |
| 2292 | .m(1) |
| 2293 | .n(16) |
| 2294 | .k(k) |
| 2295 | .ks(3) |
| 2296 | .a_offset(23) |
| 2297 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2298 | } |
| 2299 | } |
| 2300 | |
| 2301 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, zero) { |
| 2302 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2303 | for (uint32_t mz = 0; mz < 1; mz++) { |
| 2304 | for (size_t k = 1; k <= 20; k += 5) { |
| 2305 | GemmMicrokernelTester() |
| 2306 | .mr(1) |
| 2307 | .nr(16) |
| 2308 | .kr(1) |
| 2309 | .sr(1) |
| 2310 | .m(1) |
| 2311 | .n(16) |
| 2312 | .k(k) |
| 2313 | .ks(3) |
| 2314 | .a_offset(23) |
| 2315 | .zero_index(mz) |
| 2316 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2317 | } |
| 2318 | } |
| 2319 | } |
| 2320 | |
| 2321 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, qmin) { |
| 2322 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2323 | GemmMicrokernelTester() |
| 2324 | .mr(1) |
| 2325 | .nr(16) |
| 2326 | .kr(1) |
| 2327 | .sr(1) |
| 2328 | .m(1) |
| 2329 | .n(16) |
| 2330 | .k(4) |
| 2331 | .qmin(128) |
| 2332 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2333 | } |
| 2334 | |
| 2335 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, qmax) { |
| 2336 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2337 | GemmMicrokernelTester() |
| 2338 | .mr(1) |
| 2339 | .nr(16) |
| 2340 | .kr(1) |
| 2341 | .sr(1) |
| 2342 | .m(1) |
| 2343 | .n(16) |
| 2344 | .k(4) |
| 2345 | .qmax(128) |
| 2346 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2347 | } |
| 2348 | |
| 2349 | TEST(F16_IGEMM_MINMAX_1X16__NEONFP16ARITH_LD64, strided_cm) { |
| 2350 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2351 | GemmMicrokernelTester() |
| 2352 | .mr(1) |
| 2353 | .nr(16) |
| 2354 | .kr(1) |
| 2355 | .sr(1) |
| 2356 | .m(1) |
| 2357 | .n(16) |
| 2358 | .k(4) |
| 2359 | .cm_stride(19) |
| 2360 | .Test(xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 2361 | } |
| 2362 | #endif // XNN_ARCH_ARM64 |
| 2363 | |
| 2364 | |
| 2365 | #if XNN_ARCH_ARM64 |
| 2366 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_eq_4) { |
| 2367 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2368 | GemmMicrokernelTester() |
| 2369 | .mr(4) |
| 2370 | .nr(16) |
| 2371 | .kr(1) |
| 2372 | .sr(1) |
| 2373 | .m(4) |
| 2374 | .n(16) |
| 2375 | .k(4) |
| 2376 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2377 | } |
| 2378 | |
| 2379 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, strided_cn) { |
| 2380 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2381 | GemmMicrokernelTester() |
| 2382 | .mr(4) |
| 2383 | .nr(16) |
| 2384 | .kr(1) |
| 2385 | .sr(1) |
| 2386 | .m(4) |
| 2387 | .n(16) |
| 2388 | .k(4) |
| 2389 | .cn_stride(19) |
| 2390 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2391 | } |
| 2392 | |
| 2393 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 2394 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2395 | for (uint32_t m = 1; m <= 4; m++) { |
| 2396 | for (uint32_t n = 1; n <= 16; n++) { |
| 2397 | GemmMicrokernelTester() |
| 2398 | .mr(4) |
| 2399 | .nr(16) |
| 2400 | .kr(1) |
| 2401 | .sr(1) |
| 2402 | .m(m) |
| 2403 | .n(n) |
| 2404 | .k(4) |
| 2405 | .iterations(1) |
| 2406 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2407 | } |
| 2408 | } |
| 2409 | } |
| 2410 | |
| 2411 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 2412 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2413 | for (uint32_t m = 1; m <= 4; m++) { |
| 2414 | GemmMicrokernelTester() |
| 2415 | .mr(4) |
| 2416 | .nr(16) |
| 2417 | .kr(1) |
| 2418 | .sr(1) |
| 2419 | .m(m) |
| 2420 | .n(16) |
| 2421 | .k(4) |
| 2422 | .iterations(1) |
| 2423 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2424 | } |
| 2425 | } |
| 2426 | |
| 2427 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 2428 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2429 | for (uint32_t n = 1; n <= 16; n++) { |
| 2430 | GemmMicrokernelTester() |
| 2431 | .mr(4) |
| 2432 | .nr(16) |
| 2433 | .kr(1) |
| 2434 | .sr(1) |
| 2435 | .m(4) |
| 2436 | .n(n) |
| 2437 | .k(4) |
| 2438 | .iterations(1) |
| 2439 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2440 | } |
| 2441 | } |
| 2442 | |
| 2443 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_lt_4) { |
| 2444 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2445 | for (size_t k = 1; k < 4; k++) { |
| 2446 | GemmMicrokernelTester() |
| 2447 | .mr(4) |
| 2448 | .nr(16) |
| 2449 | .kr(1) |
| 2450 | .sr(1) |
| 2451 | .m(4) |
| 2452 | .n(16) |
| 2453 | .k(k) |
| 2454 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2455 | } |
| 2456 | } |
| 2457 | |
| 2458 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 2459 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2460 | for (size_t k = 1; k < 4; k++) { |
| 2461 | for (uint32_t m = 1; m <= 4; m++) { |
| 2462 | for (uint32_t n = 1; n <= 16; n++) { |
| 2463 | GemmMicrokernelTester() |
| 2464 | .mr(4) |
| 2465 | .nr(16) |
| 2466 | .kr(1) |
| 2467 | .sr(1) |
| 2468 | .m(m) |
| 2469 | .n(n) |
| 2470 | .k(k) |
| 2471 | .iterations(1) |
| 2472 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2473 | } |
| 2474 | } |
| 2475 | } |
| 2476 | } |
| 2477 | |
| 2478 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_gt_4) { |
| 2479 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2480 | for (size_t k = 5; k < 8; k++) { |
| 2481 | GemmMicrokernelTester() |
| 2482 | .mr(4) |
| 2483 | .nr(16) |
| 2484 | .kr(1) |
| 2485 | .sr(1) |
| 2486 | .m(4) |
| 2487 | .n(16) |
| 2488 | .k(k) |
| 2489 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2490 | } |
| 2491 | } |
| 2492 | |
| 2493 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 2494 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2495 | for (size_t k = 5; k < 8; k++) { |
| 2496 | for (uint32_t m = 1; m <= 4; m++) { |
| 2497 | for (uint32_t n = 1; n <= 16; n++) { |
| 2498 | GemmMicrokernelTester() |
| 2499 | .mr(4) |
| 2500 | .nr(16) |
| 2501 | .kr(1) |
| 2502 | .sr(1) |
| 2503 | .m(m) |
| 2504 | .n(n) |
| 2505 | .k(k) |
| 2506 | .iterations(1) |
| 2507 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2508 | } |
| 2509 | } |
| 2510 | } |
| 2511 | } |
| 2512 | |
| 2513 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_div_4) { |
| 2514 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2515 | for (size_t k = 8; k <= 40; k += 4) { |
| 2516 | GemmMicrokernelTester() |
| 2517 | .mr(4) |
| 2518 | .nr(16) |
| 2519 | .kr(1) |
| 2520 | .sr(1) |
| 2521 | .m(4) |
| 2522 | .n(16) |
| 2523 | .k(k) |
| 2524 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2525 | } |
| 2526 | } |
| 2527 | |
| 2528 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 2529 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2530 | for (size_t k = 8; k <= 40; k += 4) { |
| 2531 | for (uint32_t m = 1; m <= 4; m++) { |
| 2532 | for (uint32_t n = 1; n <= 16; n++) { |
| 2533 | GemmMicrokernelTester() |
| 2534 | .mr(4) |
| 2535 | .nr(16) |
| 2536 | .kr(1) |
| 2537 | .sr(1) |
| 2538 | .m(m) |
| 2539 | .n(n) |
| 2540 | .k(k) |
| 2541 | .iterations(1) |
| 2542 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2543 | } |
| 2544 | } |
| 2545 | } |
| 2546 | } |
| 2547 | |
| 2548 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, n_gt_16) { |
| 2549 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2550 | for (uint32_t n = 17; n < 32; n++) { |
| 2551 | for (size_t k = 1; k <= 20; k += 5) { |
| 2552 | GemmMicrokernelTester() |
| 2553 | .mr(4) |
| 2554 | .nr(16) |
| 2555 | .kr(1) |
| 2556 | .sr(1) |
| 2557 | .m(4) |
| 2558 | .n(16) |
| 2559 | .k(k) |
| 2560 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2561 | } |
| 2562 | } |
| 2563 | } |
| 2564 | |
| 2565 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, n_gt_16_strided_cn) { |
| 2566 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2567 | for (uint32_t n = 17; n < 32; n++) { |
| 2568 | for (size_t k = 1; k <= 20; k += 5) { |
| 2569 | GemmMicrokernelTester() |
| 2570 | .mr(4) |
| 2571 | .nr(16) |
| 2572 | .kr(1) |
| 2573 | .sr(1) |
| 2574 | .m(4) |
| 2575 | .n(16) |
| 2576 | .k(k) |
| 2577 | .cn_stride(19) |
| 2578 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2579 | } |
| 2580 | } |
| 2581 | } |
| 2582 | |
| 2583 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, n_gt_16_subtile) { |
| 2584 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2585 | for (uint32_t n = 17; n < 32; n++) { |
| 2586 | for (size_t k = 1; k <= 20; k += 5) { |
| 2587 | for (uint32_t m = 1; m <= 4; m++) { |
| 2588 | GemmMicrokernelTester() |
| 2589 | .mr(4) |
| 2590 | .nr(16) |
| 2591 | .kr(1) |
| 2592 | .sr(1) |
| 2593 | .m(m) |
| 2594 | .n(n) |
| 2595 | .k(k) |
| 2596 | .iterations(1) |
| 2597 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2598 | } |
| 2599 | } |
| 2600 | } |
| 2601 | } |
| 2602 | |
| 2603 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, n_div_16) { |
| 2604 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2605 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2606 | for (size_t k = 1; k <= 20; k += 5) { |
| 2607 | GemmMicrokernelTester() |
| 2608 | .mr(4) |
| 2609 | .nr(16) |
| 2610 | .kr(1) |
| 2611 | .sr(1) |
| 2612 | .m(4) |
| 2613 | .n(16) |
| 2614 | .k(k) |
| 2615 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2616 | } |
| 2617 | } |
| 2618 | } |
| 2619 | |
| 2620 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, n_div_16_strided_cn) { |
| 2621 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2622 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2623 | for (size_t k = 1; k <= 20; k += 5) { |
| 2624 | GemmMicrokernelTester() |
| 2625 | .mr(4) |
| 2626 | .nr(16) |
| 2627 | .kr(1) |
| 2628 | .sr(1) |
| 2629 | .m(4) |
| 2630 | .n(n) |
| 2631 | .k(k) |
| 2632 | .cn_stride(19) |
| 2633 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2634 | } |
| 2635 | } |
| 2636 | } |
| 2637 | |
| 2638 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, n_div_16_subtile) { |
| 2639 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2640 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2641 | for (size_t k = 1; k <= 20; k += 5) { |
| 2642 | for (uint32_t m = 1; m <= 4; m++) { |
| 2643 | GemmMicrokernelTester() |
| 2644 | .mr(4) |
| 2645 | .nr(16) |
| 2646 | .kr(1) |
| 2647 | .sr(1) |
| 2648 | .m(m) |
| 2649 | .n(n) |
| 2650 | .k(k) |
| 2651 | .iterations(1) |
| 2652 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2653 | } |
| 2654 | } |
| 2655 | } |
| 2656 | } |
| 2657 | |
| 2658 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, small_kernel) { |
| 2659 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2660 | for (size_t k = 1; k <= 20; k += 5) { |
| 2661 | GemmMicrokernelTester() |
| 2662 | .mr(4) |
| 2663 | .nr(16) |
| 2664 | .kr(1) |
| 2665 | .sr(1) |
| 2666 | .m(4) |
| 2667 | .n(16) |
| 2668 | .k(k) |
| 2669 | .ks(3) |
| 2670 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2671 | } |
| 2672 | } |
| 2673 | |
| 2674 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, small_kernel_subtile) { |
| 2675 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2676 | for (size_t k = 1; k <= 20; k += 5) { |
| 2677 | for (uint32_t m = 1; m <= 4; m++) { |
| 2678 | for (uint32_t n = 1; n <= 16; n++) { |
| 2679 | GemmMicrokernelTester() |
| 2680 | .mr(4) |
| 2681 | .nr(16) |
| 2682 | .kr(1) |
| 2683 | .sr(1) |
| 2684 | .m(m) |
| 2685 | .n(n) |
| 2686 | .k(k) |
| 2687 | .ks(3) |
| 2688 | .iterations(1) |
| 2689 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2690 | } |
| 2691 | } |
| 2692 | } |
| 2693 | } |
| 2694 | |
| 2695 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, n_gt_16_small_kernel) { |
| 2696 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2697 | for (uint32_t n = 17; n < 32; n++) { |
| 2698 | for (size_t k = 1; k <= 20; k += 5) { |
| 2699 | GemmMicrokernelTester() |
| 2700 | .mr(4) |
| 2701 | .nr(16) |
| 2702 | .kr(1) |
| 2703 | .sr(1) |
| 2704 | .m(4) |
| 2705 | .n(16) |
| 2706 | .k(k) |
| 2707 | .ks(3) |
| 2708 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2709 | } |
| 2710 | } |
| 2711 | } |
| 2712 | |
| 2713 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, n_div_16_small_kernel) { |
| 2714 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2715 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2716 | for (size_t k = 1; k <= 20; k += 5) { |
| 2717 | GemmMicrokernelTester() |
| 2718 | .mr(4) |
| 2719 | .nr(16) |
| 2720 | .kr(1) |
| 2721 | .sr(1) |
| 2722 | .m(4) |
| 2723 | .n(16) |
| 2724 | .k(k) |
| 2725 | .ks(3) |
| 2726 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2727 | } |
| 2728 | } |
| 2729 | } |
| 2730 | |
| 2731 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 2732 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2733 | for (size_t k = 1; k <= 20; k += 5) { |
| 2734 | for (uint32_t m = 1; m <= 4; m++) { |
| 2735 | for (uint32_t n = 1; n <= 16; n++) { |
| 2736 | GemmMicrokernelTester() |
| 2737 | .mr(4) |
| 2738 | .nr(16) |
| 2739 | .kr(1) |
| 2740 | .sr(1) |
| 2741 | .m(m) |
| 2742 | .n(n) |
| 2743 | .k(k) |
| 2744 | .cm_stride(19) |
| 2745 | .iterations(1) |
| 2746 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2747 | } |
| 2748 | } |
| 2749 | } |
| 2750 | } |
| 2751 | |
| 2752 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, a_offset) { |
| 2753 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2754 | for (size_t k = 1; k <= 20; k += 5) { |
| 2755 | GemmMicrokernelTester() |
| 2756 | .mr(4) |
| 2757 | .nr(16) |
| 2758 | .kr(1) |
| 2759 | .sr(1) |
| 2760 | .m(4) |
| 2761 | .n(16) |
| 2762 | .k(k) |
| 2763 | .ks(3) |
| 2764 | .a_offset(83) |
| 2765 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2766 | } |
| 2767 | } |
| 2768 | |
| 2769 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, zero) { |
| 2770 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2771 | for (uint32_t mz = 0; mz < 4; mz++) { |
| 2772 | for (size_t k = 1; k <= 20; k += 5) { |
| 2773 | GemmMicrokernelTester() |
| 2774 | .mr(4) |
| 2775 | .nr(16) |
| 2776 | .kr(1) |
| 2777 | .sr(1) |
| 2778 | .m(4) |
| 2779 | .n(16) |
| 2780 | .k(k) |
| 2781 | .ks(3) |
| 2782 | .a_offset(83) |
| 2783 | .zero_index(mz) |
| 2784 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2785 | } |
| 2786 | } |
| 2787 | } |
| 2788 | |
| 2789 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, qmin) { |
| 2790 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2791 | GemmMicrokernelTester() |
| 2792 | .mr(4) |
| 2793 | .nr(16) |
| 2794 | .kr(1) |
| 2795 | .sr(1) |
| 2796 | .m(4) |
| 2797 | .n(16) |
| 2798 | .k(4) |
| 2799 | .qmin(128) |
| 2800 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2801 | } |
| 2802 | |
| 2803 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, qmax) { |
| 2804 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2805 | GemmMicrokernelTester() |
| 2806 | .mr(4) |
| 2807 | .nr(16) |
| 2808 | .kr(1) |
| 2809 | .sr(1) |
| 2810 | .m(4) |
| 2811 | .n(16) |
| 2812 | .k(4) |
| 2813 | .qmax(128) |
| 2814 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2815 | } |
| 2816 | |
| 2817 | TEST(F16_IGEMM_MINMAX_4X16__NEONFP16ARITH_LD64, strided_cm) { |
| 2818 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2819 | GemmMicrokernelTester() |
| 2820 | .mr(4) |
| 2821 | .nr(16) |
| 2822 | .kr(1) |
| 2823 | .sr(1) |
| 2824 | .m(4) |
| 2825 | .n(16) |
| 2826 | .k(4) |
| 2827 | .cm_stride(19) |
| 2828 | .Test(xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64); |
| 2829 | } |
| 2830 | #endif // XNN_ARCH_ARM64 |
| 2831 | |
| 2832 | |
| 2833 | #if XNN_ARCH_ARM64 |
| 2834 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_eq_4) { |
| 2835 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2836 | GemmMicrokernelTester() |
| 2837 | .mr(6) |
| 2838 | .nr(16) |
| 2839 | .kr(1) |
| 2840 | .sr(1) |
| 2841 | .m(6) |
| 2842 | .n(16) |
| 2843 | .k(4) |
| 2844 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2845 | } |
| 2846 | |
| 2847 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, strided_cn) { |
| 2848 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2849 | GemmMicrokernelTester() |
| 2850 | .mr(6) |
| 2851 | .nr(16) |
| 2852 | .kr(1) |
| 2853 | .sr(1) |
| 2854 | .m(6) |
| 2855 | .n(16) |
| 2856 | .k(4) |
| 2857 | .cn_stride(19) |
| 2858 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2859 | } |
| 2860 | |
| 2861 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 2862 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2863 | for (uint32_t m = 1; m <= 6; m++) { |
| 2864 | for (uint32_t n = 1; n <= 16; n++) { |
| 2865 | GemmMicrokernelTester() |
| 2866 | .mr(6) |
| 2867 | .nr(16) |
| 2868 | .kr(1) |
| 2869 | .sr(1) |
| 2870 | .m(m) |
| 2871 | .n(n) |
| 2872 | .k(4) |
| 2873 | .iterations(1) |
| 2874 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2875 | } |
| 2876 | } |
| 2877 | } |
| 2878 | |
| 2879 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 2880 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2881 | for (uint32_t m = 1; m <= 6; m++) { |
| 2882 | GemmMicrokernelTester() |
| 2883 | .mr(6) |
| 2884 | .nr(16) |
| 2885 | .kr(1) |
| 2886 | .sr(1) |
| 2887 | .m(m) |
| 2888 | .n(16) |
| 2889 | .k(4) |
| 2890 | .iterations(1) |
| 2891 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2892 | } |
| 2893 | } |
| 2894 | |
| 2895 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 2896 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2897 | for (uint32_t n = 1; n <= 16; n++) { |
| 2898 | GemmMicrokernelTester() |
| 2899 | .mr(6) |
| 2900 | .nr(16) |
| 2901 | .kr(1) |
| 2902 | .sr(1) |
| 2903 | .m(6) |
| 2904 | .n(n) |
| 2905 | .k(4) |
| 2906 | .iterations(1) |
| 2907 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2908 | } |
| 2909 | } |
| 2910 | |
| 2911 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_lt_4) { |
| 2912 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2913 | for (size_t k = 1; k < 4; k++) { |
| 2914 | GemmMicrokernelTester() |
| 2915 | .mr(6) |
| 2916 | .nr(16) |
| 2917 | .kr(1) |
| 2918 | .sr(1) |
| 2919 | .m(6) |
| 2920 | .n(16) |
| 2921 | .k(k) |
| 2922 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2923 | } |
| 2924 | } |
| 2925 | |
| 2926 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 2927 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2928 | for (size_t k = 1; k < 4; k++) { |
| 2929 | for (uint32_t m = 1; m <= 6; m++) { |
| 2930 | for (uint32_t n = 1; n <= 16; n++) { |
| 2931 | GemmMicrokernelTester() |
| 2932 | .mr(6) |
| 2933 | .nr(16) |
| 2934 | .kr(1) |
| 2935 | .sr(1) |
| 2936 | .m(m) |
| 2937 | .n(n) |
| 2938 | .k(k) |
| 2939 | .iterations(1) |
| 2940 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2941 | } |
| 2942 | } |
| 2943 | } |
| 2944 | } |
| 2945 | |
| 2946 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_gt_4) { |
| 2947 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2948 | for (size_t k = 5; k < 8; k++) { |
| 2949 | GemmMicrokernelTester() |
| 2950 | .mr(6) |
| 2951 | .nr(16) |
| 2952 | .kr(1) |
| 2953 | .sr(1) |
| 2954 | .m(6) |
| 2955 | .n(16) |
| 2956 | .k(k) |
| 2957 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2958 | } |
| 2959 | } |
| 2960 | |
| 2961 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 2962 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2963 | for (size_t k = 5; k < 8; k++) { |
| 2964 | for (uint32_t m = 1; m <= 6; m++) { |
| 2965 | for (uint32_t n = 1; n <= 16; n++) { |
| 2966 | GemmMicrokernelTester() |
| 2967 | .mr(6) |
| 2968 | .nr(16) |
| 2969 | .kr(1) |
| 2970 | .sr(1) |
| 2971 | .m(m) |
| 2972 | .n(n) |
| 2973 | .k(k) |
| 2974 | .iterations(1) |
| 2975 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2976 | } |
| 2977 | } |
| 2978 | } |
| 2979 | } |
| 2980 | |
| 2981 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_div_4) { |
| 2982 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2983 | for (size_t k = 8; k <= 40; k += 4) { |
| 2984 | GemmMicrokernelTester() |
| 2985 | .mr(6) |
| 2986 | .nr(16) |
| 2987 | .kr(1) |
| 2988 | .sr(1) |
| 2989 | .m(6) |
| 2990 | .n(16) |
| 2991 | .k(k) |
| 2992 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 2993 | } |
| 2994 | } |
| 2995 | |
| 2996 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 2997 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2998 | for (size_t k = 8; k <= 40; k += 4) { |
| 2999 | for (uint32_t m = 1; m <= 6; m++) { |
| 3000 | for (uint32_t n = 1; n <= 16; n++) { |
| 3001 | GemmMicrokernelTester() |
| 3002 | .mr(6) |
| 3003 | .nr(16) |
| 3004 | .kr(1) |
| 3005 | .sr(1) |
| 3006 | .m(m) |
| 3007 | .n(n) |
| 3008 | .k(k) |
| 3009 | .iterations(1) |
| 3010 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3011 | } |
| 3012 | } |
| 3013 | } |
| 3014 | } |
| 3015 | |
| 3016 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, n_gt_16) { |
| 3017 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3018 | for (uint32_t n = 17; n < 32; n++) { |
| 3019 | for (size_t k = 1; k <= 20; k += 5) { |
| 3020 | GemmMicrokernelTester() |
| 3021 | .mr(6) |
| 3022 | .nr(16) |
| 3023 | .kr(1) |
| 3024 | .sr(1) |
| 3025 | .m(6) |
| 3026 | .n(16) |
| 3027 | .k(k) |
| 3028 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3029 | } |
| 3030 | } |
| 3031 | } |
| 3032 | |
| 3033 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, n_gt_16_strided_cn) { |
| 3034 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3035 | for (uint32_t n = 17; n < 32; n++) { |
| 3036 | for (size_t k = 1; k <= 20; k += 5) { |
| 3037 | GemmMicrokernelTester() |
| 3038 | .mr(6) |
| 3039 | .nr(16) |
| 3040 | .kr(1) |
| 3041 | .sr(1) |
| 3042 | .m(6) |
| 3043 | .n(16) |
| 3044 | .k(k) |
| 3045 | .cn_stride(19) |
| 3046 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3047 | } |
| 3048 | } |
| 3049 | } |
| 3050 | |
| 3051 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, n_gt_16_subtile) { |
| 3052 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3053 | for (uint32_t n = 17; n < 32; n++) { |
| 3054 | for (size_t k = 1; k <= 20; k += 5) { |
| 3055 | for (uint32_t m = 1; m <= 6; m++) { |
| 3056 | GemmMicrokernelTester() |
| 3057 | .mr(6) |
| 3058 | .nr(16) |
| 3059 | .kr(1) |
| 3060 | .sr(1) |
| 3061 | .m(m) |
| 3062 | .n(n) |
| 3063 | .k(k) |
| 3064 | .iterations(1) |
| 3065 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3066 | } |
| 3067 | } |
| 3068 | } |
| 3069 | } |
| 3070 | |
| 3071 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, n_div_16) { |
| 3072 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3073 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3074 | for (size_t k = 1; k <= 20; k += 5) { |
| 3075 | GemmMicrokernelTester() |
| 3076 | .mr(6) |
| 3077 | .nr(16) |
| 3078 | .kr(1) |
| 3079 | .sr(1) |
| 3080 | .m(6) |
| 3081 | .n(16) |
| 3082 | .k(k) |
| 3083 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3084 | } |
| 3085 | } |
| 3086 | } |
| 3087 | |
| 3088 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, n_div_16_strided_cn) { |
| 3089 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3090 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3091 | for (size_t k = 1; k <= 20; k += 5) { |
| 3092 | GemmMicrokernelTester() |
| 3093 | .mr(6) |
| 3094 | .nr(16) |
| 3095 | .kr(1) |
| 3096 | .sr(1) |
| 3097 | .m(6) |
| 3098 | .n(n) |
| 3099 | .k(k) |
| 3100 | .cn_stride(19) |
| 3101 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3102 | } |
| 3103 | } |
| 3104 | } |
| 3105 | |
| 3106 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, n_div_16_subtile) { |
| 3107 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3108 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3109 | for (size_t k = 1; k <= 20; k += 5) { |
| 3110 | for (uint32_t m = 1; m <= 6; m++) { |
| 3111 | GemmMicrokernelTester() |
| 3112 | .mr(6) |
| 3113 | .nr(16) |
| 3114 | .kr(1) |
| 3115 | .sr(1) |
| 3116 | .m(m) |
| 3117 | .n(n) |
| 3118 | .k(k) |
| 3119 | .iterations(1) |
| 3120 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3121 | } |
| 3122 | } |
| 3123 | } |
| 3124 | } |
| 3125 | |
| 3126 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, small_kernel) { |
| 3127 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3128 | for (size_t k = 1; k <= 20; k += 5) { |
| 3129 | GemmMicrokernelTester() |
| 3130 | .mr(6) |
| 3131 | .nr(16) |
| 3132 | .kr(1) |
| 3133 | .sr(1) |
| 3134 | .m(6) |
| 3135 | .n(16) |
| 3136 | .k(k) |
| 3137 | .ks(3) |
| 3138 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3139 | } |
| 3140 | } |
| 3141 | |
| 3142 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, small_kernel_subtile) { |
| 3143 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3144 | for (size_t k = 1; k <= 20; k += 5) { |
| 3145 | for (uint32_t m = 1; m <= 6; m++) { |
| 3146 | for (uint32_t n = 1; n <= 16; n++) { |
| 3147 | GemmMicrokernelTester() |
| 3148 | .mr(6) |
| 3149 | .nr(16) |
| 3150 | .kr(1) |
| 3151 | .sr(1) |
| 3152 | .m(m) |
| 3153 | .n(n) |
| 3154 | .k(k) |
| 3155 | .ks(3) |
| 3156 | .iterations(1) |
| 3157 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3158 | } |
| 3159 | } |
| 3160 | } |
| 3161 | } |
| 3162 | |
| 3163 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, n_gt_16_small_kernel) { |
| 3164 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3165 | for (uint32_t n = 17; n < 32; n++) { |
| 3166 | for (size_t k = 1; k <= 20; k += 5) { |
| 3167 | GemmMicrokernelTester() |
| 3168 | .mr(6) |
| 3169 | .nr(16) |
| 3170 | .kr(1) |
| 3171 | .sr(1) |
| 3172 | .m(6) |
| 3173 | .n(16) |
| 3174 | .k(k) |
| 3175 | .ks(3) |
| 3176 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3177 | } |
| 3178 | } |
| 3179 | } |
| 3180 | |
| 3181 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, n_div_16_small_kernel) { |
| 3182 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3183 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3184 | for (size_t k = 1; k <= 20; k += 5) { |
| 3185 | GemmMicrokernelTester() |
| 3186 | .mr(6) |
| 3187 | .nr(16) |
| 3188 | .kr(1) |
| 3189 | .sr(1) |
| 3190 | .m(6) |
| 3191 | .n(16) |
| 3192 | .k(k) |
| 3193 | .ks(3) |
| 3194 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3195 | } |
| 3196 | } |
| 3197 | } |
| 3198 | |
| 3199 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 3200 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3201 | for (size_t k = 1; k <= 20; k += 5) { |
| 3202 | for (uint32_t m = 1; m <= 6; m++) { |
| 3203 | for (uint32_t n = 1; n <= 16; n++) { |
| 3204 | GemmMicrokernelTester() |
| 3205 | .mr(6) |
| 3206 | .nr(16) |
| 3207 | .kr(1) |
| 3208 | .sr(1) |
| 3209 | .m(m) |
| 3210 | .n(n) |
| 3211 | .k(k) |
| 3212 | .cm_stride(19) |
| 3213 | .iterations(1) |
| 3214 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3215 | } |
| 3216 | } |
| 3217 | } |
| 3218 | } |
| 3219 | |
| 3220 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, a_offset) { |
| 3221 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3222 | for (size_t k = 1; k <= 20; k += 5) { |
| 3223 | GemmMicrokernelTester() |
| 3224 | .mr(6) |
| 3225 | .nr(16) |
| 3226 | .kr(1) |
| 3227 | .sr(1) |
| 3228 | .m(6) |
| 3229 | .n(16) |
| 3230 | .k(k) |
| 3231 | .ks(3) |
| 3232 | .a_offset(127) |
| 3233 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3234 | } |
| 3235 | } |
| 3236 | |
| 3237 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, zero) { |
| 3238 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3239 | for (uint32_t mz = 0; mz < 6; mz++) { |
| 3240 | for (size_t k = 1; k <= 20; k += 5) { |
| 3241 | GemmMicrokernelTester() |
| 3242 | .mr(6) |
| 3243 | .nr(16) |
| 3244 | .kr(1) |
| 3245 | .sr(1) |
| 3246 | .m(6) |
| 3247 | .n(16) |
| 3248 | .k(k) |
| 3249 | .ks(3) |
| 3250 | .a_offset(127) |
| 3251 | .zero_index(mz) |
| 3252 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3253 | } |
| 3254 | } |
| 3255 | } |
| 3256 | |
| 3257 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, qmin) { |
| 3258 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3259 | GemmMicrokernelTester() |
| 3260 | .mr(6) |
| 3261 | .nr(16) |
| 3262 | .kr(1) |
| 3263 | .sr(1) |
| 3264 | .m(6) |
| 3265 | .n(16) |
| 3266 | .k(4) |
| 3267 | .qmin(128) |
| 3268 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3269 | } |
| 3270 | |
| 3271 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, qmax) { |
| 3272 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3273 | GemmMicrokernelTester() |
| 3274 | .mr(6) |
| 3275 | .nr(16) |
| 3276 | .kr(1) |
| 3277 | .sr(1) |
| 3278 | .m(6) |
| 3279 | .n(16) |
| 3280 | .k(4) |
| 3281 | .qmax(128) |
| 3282 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3283 | } |
| 3284 | |
| 3285 | TEST(F16_IGEMM_MINMAX_6X16__NEONFP16ARITH_LD64, strided_cm) { |
| 3286 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3287 | GemmMicrokernelTester() |
| 3288 | .mr(6) |
| 3289 | .nr(16) |
| 3290 | .kr(1) |
| 3291 | .sr(1) |
| 3292 | .m(6) |
| 3293 | .n(16) |
| 3294 | .k(4) |
| 3295 | .cm_stride(19) |
| 3296 | .Test(xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 3297 | } |
| 3298 | #endif // XNN_ARCH_ARM64 |
| 3299 | |
| 3300 | |
| 3301 | #if XNN_ARCH_ARM64 |
| 3302 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_eq_4) { |
| 3303 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3304 | GemmMicrokernelTester() |
| 3305 | .mr(8) |
| 3306 | .nr(16) |
| 3307 | .kr(1) |
| 3308 | .sr(1) |
| 3309 | .m(8) |
| 3310 | .n(16) |
| 3311 | .k(4) |
| 3312 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3313 | } |
| 3314 | |
| 3315 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, strided_cn) { |
| 3316 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3317 | GemmMicrokernelTester() |
| 3318 | .mr(8) |
| 3319 | .nr(16) |
| 3320 | .kr(1) |
| 3321 | .sr(1) |
| 3322 | .m(8) |
| 3323 | .n(16) |
| 3324 | .k(4) |
| 3325 | .cn_stride(19) |
| 3326 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3327 | } |
| 3328 | |
| 3329 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 3330 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3331 | for (uint32_t m = 1; m <= 8; m++) { |
| 3332 | for (uint32_t n = 1; n <= 16; n++) { |
| 3333 | GemmMicrokernelTester() |
| 3334 | .mr(8) |
| 3335 | .nr(16) |
| 3336 | .kr(1) |
| 3337 | .sr(1) |
| 3338 | .m(m) |
| 3339 | .n(n) |
| 3340 | .k(4) |
| 3341 | .iterations(1) |
| 3342 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3343 | } |
| 3344 | } |
| 3345 | } |
| 3346 | |
| 3347 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 3348 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3349 | for (uint32_t m = 1; m <= 8; m++) { |
| 3350 | GemmMicrokernelTester() |
| 3351 | .mr(8) |
| 3352 | .nr(16) |
| 3353 | .kr(1) |
| 3354 | .sr(1) |
| 3355 | .m(m) |
| 3356 | .n(16) |
| 3357 | .k(4) |
| 3358 | .iterations(1) |
| 3359 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3360 | } |
| 3361 | } |
| 3362 | |
| 3363 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 3364 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3365 | for (uint32_t n = 1; n <= 16; n++) { |
| 3366 | GemmMicrokernelTester() |
| 3367 | .mr(8) |
| 3368 | .nr(16) |
| 3369 | .kr(1) |
| 3370 | .sr(1) |
| 3371 | .m(8) |
| 3372 | .n(n) |
| 3373 | .k(4) |
| 3374 | .iterations(1) |
| 3375 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3376 | } |
| 3377 | } |
| 3378 | |
| 3379 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_lt_4) { |
| 3380 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3381 | for (size_t k = 1; k < 4; k++) { |
| 3382 | GemmMicrokernelTester() |
| 3383 | .mr(8) |
| 3384 | .nr(16) |
| 3385 | .kr(1) |
| 3386 | .sr(1) |
| 3387 | .m(8) |
| 3388 | .n(16) |
| 3389 | .k(k) |
| 3390 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3391 | } |
| 3392 | } |
| 3393 | |
| 3394 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 3395 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3396 | for (size_t k = 1; k < 4; k++) { |
| 3397 | for (uint32_t m = 1; m <= 8; m++) { |
| 3398 | for (uint32_t n = 1; n <= 16; n++) { |
| 3399 | GemmMicrokernelTester() |
| 3400 | .mr(8) |
| 3401 | .nr(16) |
| 3402 | .kr(1) |
| 3403 | .sr(1) |
| 3404 | .m(m) |
| 3405 | .n(n) |
| 3406 | .k(k) |
| 3407 | .iterations(1) |
| 3408 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3409 | } |
| 3410 | } |
| 3411 | } |
| 3412 | } |
| 3413 | |
| 3414 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_gt_4) { |
| 3415 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3416 | for (size_t k = 5; k < 8; k++) { |
| 3417 | GemmMicrokernelTester() |
| 3418 | .mr(8) |
| 3419 | .nr(16) |
| 3420 | .kr(1) |
| 3421 | .sr(1) |
| 3422 | .m(8) |
| 3423 | .n(16) |
| 3424 | .k(k) |
| 3425 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3426 | } |
| 3427 | } |
| 3428 | |
| 3429 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 3430 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3431 | for (size_t k = 5; k < 8; k++) { |
| 3432 | for (uint32_t m = 1; m <= 8; m++) { |
| 3433 | for (uint32_t n = 1; n <= 16; n++) { |
| 3434 | GemmMicrokernelTester() |
| 3435 | .mr(8) |
| 3436 | .nr(16) |
| 3437 | .kr(1) |
| 3438 | .sr(1) |
| 3439 | .m(m) |
| 3440 | .n(n) |
| 3441 | .k(k) |
| 3442 | .iterations(1) |
| 3443 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3444 | } |
| 3445 | } |
| 3446 | } |
| 3447 | } |
| 3448 | |
| 3449 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_div_4) { |
| 3450 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3451 | for (size_t k = 8; k <= 40; k += 4) { |
| 3452 | GemmMicrokernelTester() |
| 3453 | .mr(8) |
| 3454 | .nr(16) |
| 3455 | .kr(1) |
| 3456 | .sr(1) |
| 3457 | .m(8) |
| 3458 | .n(16) |
| 3459 | .k(k) |
| 3460 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3461 | } |
| 3462 | } |
| 3463 | |
| 3464 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 3465 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3466 | for (size_t k = 8; k <= 40; k += 4) { |
| 3467 | for (uint32_t m = 1; m <= 8; m++) { |
| 3468 | for (uint32_t n = 1; n <= 16; n++) { |
| 3469 | GemmMicrokernelTester() |
| 3470 | .mr(8) |
| 3471 | .nr(16) |
| 3472 | .kr(1) |
| 3473 | .sr(1) |
| 3474 | .m(m) |
| 3475 | .n(n) |
| 3476 | .k(k) |
| 3477 | .iterations(1) |
| 3478 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3479 | } |
| 3480 | } |
| 3481 | } |
| 3482 | } |
| 3483 | |
| 3484 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, n_gt_16) { |
| 3485 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3486 | for (uint32_t n = 17; n < 32; n++) { |
| 3487 | for (size_t k = 1; k <= 20; k += 5) { |
| 3488 | GemmMicrokernelTester() |
| 3489 | .mr(8) |
| 3490 | .nr(16) |
| 3491 | .kr(1) |
| 3492 | .sr(1) |
| 3493 | .m(8) |
| 3494 | .n(16) |
| 3495 | .k(k) |
| 3496 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3497 | } |
| 3498 | } |
| 3499 | } |
| 3500 | |
| 3501 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, n_gt_16_strided_cn) { |
| 3502 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3503 | for (uint32_t n = 17; n < 32; n++) { |
| 3504 | for (size_t k = 1; k <= 20; k += 5) { |
| 3505 | GemmMicrokernelTester() |
| 3506 | .mr(8) |
| 3507 | .nr(16) |
| 3508 | .kr(1) |
| 3509 | .sr(1) |
| 3510 | .m(8) |
| 3511 | .n(16) |
| 3512 | .k(k) |
| 3513 | .cn_stride(19) |
| 3514 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3515 | } |
| 3516 | } |
| 3517 | } |
| 3518 | |
| 3519 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, n_gt_16_subtile) { |
| 3520 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3521 | for (uint32_t n = 17; n < 32; n++) { |
| 3522 | for (size_t k = 1; k <= 20; k += 5) { |
| 3523 | for (uint32_t m = 1; m <= 8; m++) { |
| 3524 | GemmMicrokernelTester() |
| 3525 | .mr(8) |
| 3526 | .nr(16) |
| 3527 | .kr(1) |
| 3528 | .sr(1) |
| 3529 | .m(m) |
| 3530 | .n(n) |
| 3531 | .k(k) |
| 3532 | .iterations(1) |
| 3533 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3534 | } |
| 3535 | } |
| 3536 | } |
| 3537 | } |
| 3538 | |
| 3539 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, n_div_16) { |
| 3540 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3541 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3542 | for (size_t k = 1; k <= 20; k += 5) { |
| 3543 | GemmMicrokernelTester() |
| 3544 | .mr(8) |
| 3545 | .nr(16) |
| 3546 | .kr(1) |
| 3547 | .sr(1) |
| 3548 | .m(8) |
| 3549 | .n(16) |
| 3550 | .k(k) |
| 3551 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3552 | } |
| 3553 | } |
| 3554 | } |
| 3555 | |
| 3556 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, n_div_16_strided_cn) { |
| 3557 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3558 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3559 | for (size_t k = 1; k <= 20; k += 5) { |
| 3560 | GemmMicrokernelTester() |
| 3561 | .mr(8) |
| 3562 | .nr(16) |
| 3563 | .kr(1) |
| 3564 | .sr(1) |
| 3565 | .m(8) |
| 3566 | .n(n) |
| 3567 | .k(k) |
| 3568 | .cn_stride(19) |
| 3569 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3570 | } |
| 3571 | } |
| 3572 | } |
| 3573 | |
| 3574 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, n_div_16_subtile) { |
| 3575 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3576 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3577 | for (size_t k = 1; k <= 20; k += 5) { |
| 3578 | for (uint32_t m = 1; m <= 8; m++) { |
| 3579 | GemmMicrokernelTester() |
| 3580 | .mr(8) |
| 3581 | .nr(16) |
| 3582 | .kr(1) |
| 3583 | .sr(1) |
| 3584 | .m(m) |
| 3585 | .n(n) |
| 3586 | .k(k) |
| 3587 | .iterations(1) |
| 3588 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3589 | } |
| 3590 | } |
| 3591 | } |
| 3592 | } |
| 3593 | |
| 3594 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, small_kernel) { |
| 3595 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3596 | for (size_t k = 1; k <= 20; k += 5) { |
| 3597 | GemmMicrokernelTester() |
| 3598 | .mr(8) |
| 3599 | .nr(16) |
| 3600 | .kr(1) |
| 3601 | .sr(1) |
| 3602 | .m(8) |
| 3603 | .n(16) |
| 3604 | .k(k) |
| 3605 | .ks(3) |
| 3606 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3607 | } |
| 3608 | } |
| 3609 | |
| 3610 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, small_kernel_subtile) { |
| 3611 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3612 | for (size_t k = 1; k <= 20; k += 5) { |
| 3613 | for (uint32_t m = 1; m <= 8; m++) { |
| 3614 | for (uint32_t n = 1; n <= 16; n++) { |
| 3615 | GemmMicrokernelTester() |
| 3616 | .mr(8) |
| 3617 | .nr(16) |
| 3618 | .kr(1) |
| 3619 | .sr(1) |
| 3620 | .m(m) |
| 3621 | .n(n) |
| 3622 | .k(k) |
| 3623 | .ks(3) |
| 3624 | .iterations(1) |
| 3625 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3626 | } |
| 3627 | } |
| 3628 | } |
| 3629 | } |
| 3630 | |
| 3631 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, n_gt_16_small_kernel) { |
| 3632 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3633 | for (uint32_t n = 17; n < 32; n++) { |
| 3634 | for (size_t k = 1; k <= 20; k += 5) { |
| 3635 | GemmMicrokernelTester() |
| 3636 | .mr(8) |
| 3637 | .nr(16) |
| 3638 | .kr(1) |
| 3639 | .sr(1) |
| 3640 | .m(8) |
| 3641 | .n(16) |
| 3642 | .k(k) |
| 3643 | .ks(3) |
| 3644 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3645 | } |
| 3646 | } |
| 3647 | } |
| 3648 | |
| 3649 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, n_div_16_small_kernel) { |
| 3650 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3651 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3652 | for (size_t k = 1; k <= 20; k += 5) { |
| 3653 | GemmMicrokernelTester() |
| 3654 | .mr(8) |
| 3655 | .nr(16) |
| 3656 | .kr(1) |
| 3657 | .sr(1) |
| 3658 | .m(8) |
| 3659 | .n(16) |
| 3660 | .k(k) |
| 3661 | .ks(3) |
| 3662 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3663 | } |
| 3664 | } |
| 3665 | } |
| 3666 | |
| 3667 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 3668 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3669 | for (size_t k = 1; k <= 20; k += 5) { |
| 3670 | for (uint32_t m = 1; m <= 8; m++) { |
| 3671 | for (uint32_t n = 1; n <= 16; n++) { |
| 3672 | GemmMicrokernelTester() |
| 3673 | .mr(8) |
| 3674 | .nr(16) |
| 3675 | .kr(1) |
| 3676 | .sr(1) |
| 3677 | .m(m) |
| 3678 | .n(n) |
| 3679 | .k(k) |
| 3680 | .cm_stride(19) |
| 3681 | .iterations(1) |
| 3682 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3683 | } |
| 3684 | } |
| 3685 | } |
| 3686 | } |
| 3687 | |
| 3688 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, a_offset) { |
| 3689 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3690 | for (size_t k = 1; k <= 20; k += 5) { |
| 3691 | GemmMicrokernelTester() |
| 3692 | .mr(8) |
| 3693 | .nr(16) |
| 3694 | .kr(1) |
| 3695 | .sr(1) |
| 3696 | .m(8) |
| 3697 | .n(16) |
| 3698 | .k(k) |
| 3699 | .ks(3) |
| 3700 | .a_offset(163) |
| 3701 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3702 | } |
| 3703 | } |
| 3704 | |
| 3705 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, zero) { |
| 3706 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3707 | for (uint32_t mz = 0; mz < 8; mz++) { |
| 3708 | for (size_t k = 1; k <= 20; k += 5) { |
| 3709 | GemmMicrokernelTester() |
| 3710 | .mr(8) |
| 3711 | .nr(16) |
| 3712 | .kr(1) |
| 3713 | .sr(1) |
| 3714 | .m(8) |
| 3715 | .n(16) |
| 3716 | .k(k) |
| 3717 | .ks(3) |
| 3718 | .a_offset(163) |
| 3719 | .zero_index(mz) |
| 3720 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3721 | } |
| 3722 | } |
| 3723 | } |
| 3724 | |
| 3725 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, qmin) { |
| 3726 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3727 | GemmMicrokernelTester() |
| 3728 | .mr(8) |
| 3729 | .nr(16) |
| 3730 | .kr(1) |
| 3731 | .sr(1) |
| 3732 | .m(8) |
| 3733 | .n(16) |
| 3734 | .k(4) |
| 3735 | .qmin(128) |
| 3736 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3737 | } |
| 3738 | |
| 3739 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, qmax) { |
| 3740 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3741 | GemmMicrokernelTester() |
| 3742 | .mr(8) |
| 3743 | .nr(16) |
| 3744 | .kr(1) |
| 3745 | .sr(1) |
| 3746 | .m(8) |
| 3747 | .n(16) |
| 3748 | .k(4) |
| 3749 | .qmax(128) |
| 3750 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3751 | } |
| 3752 | |
| 3753 | TEST(F16_IGEMM_MINMAX_8X16__NEONFP16ARITH_LD64, strided_cm) { |
| 3754 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 3755 | GemmMicrokernelTester() |
| 3756 | .mr(8) |
| 3757 | .nr(16) |
| 3758 | .kr(1) |
| 3759 | .sr(1) |
| 3760 | .m(8) |
| 3761 | .n(16) |
| 3762 | .k(4) |
| 3763 | .cm_stride(19) |
| 3764 | .Test(xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64); |
| 3765 | } |
| 3766 | #endif // XNN_ARCH_ARM64 |