| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | // |
| 9 | // Auto-generated file. Do not edit! |
| 10 | // Specification: test/f16-gemm.yaml |
| 11 | // Generator: tools/generate-gemm-test.py |
| 12 | |
| 13 | |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | #include <gtest/gtest.h> |
| 15 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 16 | #include <xnnpack/common.h> |
| 17 | #include <xnnpack/isa-checks.h> |
| 18 | |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 19 | #include <xnnpack/gemm.h> |
| 20 | #include <xnnpack/igemm.h> |
| 21 | #include <xnnpack/ppmm.h> |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 22 | #include "gemm-microkernel-tester.h" |
| 23 | |
| 24 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 25 | #if XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_eq_4) { |
| 27 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 28 | GemmMicrokernelTester() |
| 29 | .mr(4) |
| 30 | .nr(8) |
| 31 | .kr(1) |
| 32 | .sr(1) |
| 33 | .m(4) |
| 34 | .n(8) |
| 35 | .k(4) |
| 36 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 37 | } |
| 38 | |
| 39 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, strided_cn) { |
| 40 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 41 | GemmMicrokernelTester() |
| 42 | .mr(4) |
| 43 | .nr(8) |
| 44 | .kr(1) |
| 45 | .sr(1) |
| 46 | .m(4) |
| 47 | .n(8) |
| 48 | .k(4) |
| 49 | .cn_stride(11) |
| 50 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 51 | } |
| 52 | |
| 53 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_eq_4_strided_a) { |
| 54 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 55 | GemmMicrokernelTester() |
| 56 | .mr(4) |
| 57 | .nr(8) |
| 58 | .kr(1) |
| 59 | .sr(1) |
| 60 | .m(4) |
| 61 | .n(8) |
| 62 | .k(4) |
| 63 | .a_stride(7) |
| 64 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 65 | } |
| 66 | |
| 67 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 68 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 69 | for (uint32_t m = 1; m <= 4; m++) { |
| 70 | for (uint32_t n = 1; n <= 8; n++) { |
| 71 | GemmMicrokernelTester() |
| 72 | .mr(4) |
| 73 | .nr(8) |
| 74 | .kr(1) |
| 75 | .sr(1) |
| 76 | .m(m) |
| 77 | .n(n) |
| 78 | .k(4) |
| 79 | .iterations(1) |
| 80 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 81 | } |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 86 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 87 | for (uint32_t m = 1; m <= 4; m++) { |
| 88 | GemmMicrokernelTester() |
| 89 | .mr(4) |
| 90 | .nr(8) |
| 91 | .kr(1) |
| 92 | .sr(1) |
| 93 | .m(m) |
| 94 | .n(8) |
| 95 | .k(4) |
| 96 | .iterations(1) |
| 97 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 102 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 103 | for (uint32_t n = 1; n <= 8; n++) { |
| 104 | GemmMicrokernelTester() |
| 105 | .mr(4) |
| 106 | .nr(8) |
| 107 | .kr(1) |
| 108 | .sr(1) |
| 109 | .m(4) |
| 110 | .n(n) |
| 111 | .k(4) |
| 112 | .iterations(1) |
| 113 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_lt_4) { |
| 118 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 119 | for (size_t k = 1; k < 4; k++) { |
| 120 | GemmMicrokernelTester() |
| 121 | .mr(4) |
| 122 | .nr(8) |
| 123 | .kr(1) |
| 124 | .sr(1) |
| 125 | .m(4) |
| 126 | .n(8) |
| 127 | .k(k) |
| 128 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_lt_4_strided_a) { |
| 133 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 134 | for (size_t k = 1; k < 4; k++) { |
| 135 | GemmMicrokernelTester() |
| 136 | .mr(4) |
| 137 | .nr(8) |
| 138 | .kr(1) |
| 139 | .sr(1) |
| 140 | .m(4) |
| 141 | .n(8) |
| 142 | .k(k) |
| 143 | .a_stride(7) |
| 144 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 145 | } |
| 146 | } |
| 147 | |
| 148 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 149 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 150 | for (size_t k = 1; k < 4; k++) { |
| 151 | for (uint32_t m = 1; m <= 4; m++) { |
| 152 | for (uint32_t n = 1; n <= 8; n++) { |
| 153 | GemmMicrokernelTester() |
| 154 | .mr(4) |
| 155 | .nr(8) |
| 156 | .kr(1) |
| 157 | .sr(1) |
| 158 | .m(m) |
| 159 | .n(n) |
| 160 | .k(k) |
| 161 | .iterations(1) |
| 162 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 163 | } |
| 164 | } |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_gt_4) { |
| 169 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 170 | for (size_t k = 5; k < 8; k++) { |
| 171 | GemmMicrokernelTester() |
| 172 | .mr(4) |
| 173 | .nr(8) |
| 174 | .kr(1) |
| 175 | .sr(1) |
| 176 | .m(4) |
| 177 | .n(8) |
| 178 | .k(k) |
| 179 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_gt_4_strided_a) { |
| 184 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 185 | for (size_t k = 5; k < 8; k++) { |
| 186 | GemmMicrokernelTester() |
| 187 | .mr(4) |
| 188 | .nr(8) |
| 189 | .kr(1) |
| 190 | .sr(1) |
| 191 | .m(4) |
| 192 | .n(8) |
| 193 | .k(k) |
| 194 | .a_stride(11) |
| 195 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 200 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 201 | for (size_t k = 5; k < 8; k++) { |
| 202 | for (uint32_t m = 1; m <= 4; m++) { |
| 203 | for (uint32_t n = 1; n <= 8; n++) { |
| 204 | GemmMicrokernelTester() |
| 205 | .mr(4) |
| 206 | .nr(8) |
| 207 | .kr(1) |
| 208 | .sr(1) |
| 209 | .m(m) |
| 210 | .n(n) |
| 211 | .k(k) |
| 212 | .iterations(1) |
| 213 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 214 | } |
| 215 | } |
| 216 | } |
| 217 | } |
| 218 | |
| 219 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_div_4) { |
| 220 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 221 | for (size_t k = 8; k <= 40; k += 4) { |
| 222 | GemmMicrokernelTester() |
| 223 | .mr(4) |
| 224 | .nr(8) |
| 225 | .kr(1) |
| 226 | .sr(1) |
| 227 | .m(4) |
| 228 | .n(8) |
| 229 | .k(k) |
| 230 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_div_4_strided_a) { |
| 235 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 236 | for (size_t k = 8; k <= 40; k += 4) { |
| 237 | GemmMicrokernelTester() |
| 238 | .mr(4) |
| 239 | .nr(8) |
| 240 | .kr(1) |
| 241 | .sr(1) |
| 242 | .m(4) |
| 243 | .n(8) |
| 244 | .k(k) |
| 245 | .a_stride(43) |
| 246 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 247 | } |
| 248 | } |
| 249 | |
| 250 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 251 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 252 | for (size_t k = 8; k <= 40; k += 4) { |
| 253 | for (uint32_t m = 1; m <= 4; m++) { |
| 254 | for (uint32_t n = 1; n <= 8; n++) { |
| 255 | GemmMicrokernelTester() |
| 256 | .mr(4) |
| 257 | .nr(8) |
| 258 | .kr(1) |
| 259 | .sr(1) |
| 260 | .m(m) |
| 261 | .n(n) |
| 262 | .k(k) |
| 263 | .iterations(1) |
| 264 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 265 | } |
| 266 | } |
| 267 | } |
| 268 | } |
| 269 | |
| 270 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, n_gt_8) { |
| 271 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 272 | for (uint32_t n = 9; n < 16; n++) { |
| 273 | for (size_t k = 1; k <= 20; k += 5) { |
| 274 | GemmMicrokernelTester() |
| 275 | .mr(4) |
| 276 | .nr(8) |
| 277 | .kr(1) |
| 278 | .sr(1) |
| 279 | .m(4) |
| 280 | .n(8) |
| 281 | .k(k) |
| 282 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 283 | } |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
| 288 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 289 | for (uint32_t n = 9; n < 16; n++) { |
| 290 | for (size_t k = 1; k <= 20; k += 5) { |
| 291 | GemmMicrokernelTester() |
| 292 | .mr(4) |
| 293 | .nr(8) |
| 294 | .kr(1) |
| 295 | .sr(1) |
| 296 | .m(4) |
| 297 | .n(8) |
| 298 | .k(k) |
| 299 | .cn_stride(11) |
| 300 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 301 | } |
| 302 | } |
| 303 | } |
| 304 | |
| 305 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, n_gt_8_strided_a) { |
| 306 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 307 | for (uint32_t n = 9; n < 16; n++) { |
| 308 | for (size_t k = 1; k <= 20; k += 5) { |
| 309 | GemmMicrokernelTester() |
| 310 | .mr(4) |
| 311 | .nr(8) |
| 312 | .kr(1) |
| 313 | .sr(1) |
| 314 | .m(4) |
| 315 | .n(n) |
| 316 | .k(k) |
| 317 | .a_stride(23) |
| 318 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 319 | } |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
| 324 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 325 | for (uint32_t n = 9; n < 16; n++) { |
| 326 | for (size_t k = 1; k <= 20; k += 5) { |
| 327 | for (uint32_t m = 1; m <= 4; m++) { |
| 328 | GemmMicrokernelTester() |
| 329 | .mr(4) |
| 330 | .nr(8) |
| 331 | .kr(1) |
| 332 | .sr(1) |
| 333 | .m(m) |
| 334 | .n(n) |
| 335 | .k(k) |
| 336 | .iterations(1) |
| 337 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 338 | } |
| 339 | } |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, n_div_8) { |
| 344 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 345 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 346 | for (size_t k = 1; k <= 20; k += 5) { |
| 347 | GemmMicrokernelTester() |
| 348 | .mr(4) |
| 349 | .nr(8) |
| 350 | .kr(1) |
| 351 | .sr(1) |
| 352 | .m(4) |
| 353 | .n(8) |
| 354 | .k(k) |
| 355 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 356 | } |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
| 361 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 362 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 363 | for (size_t k = 1; k <= 20; k += 5) { |
| 364 | GemmMicrokernelTester() |
| 365 | .mr(4) |
| 366 | .nr(8) |
| 367 | .kr(1) |
| 368 | .sr(1) |
| 369 | .m(4) |
| 370 | .n(n) |
| 371 | .k(k) |
| 372 | .cn_stride(11) |
| 373 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 374 | } |
| 375 | } |
| 376 | } |
| 377 | |
| 378 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, n_div_8_strided_a) { |
| 379 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 380 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 381 | for (size_t k = 1; k <= 20; k += 5) { |
| 382 | GemmMicrokernelTester() |
| 383 | .mr(4) |
| 384 | .nr(8) |
| 385 | .kr(1) |
| 386 | .sr(1) |
| 387 | .m(4) |
| 388 | .n(n) |
| 389 | .k(k) |
| 390 | .a_stride(23) |
| 391 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 392 | } |
| 393 | } |
| 394 | } |
| 395 | |
| 396 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
| 397 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 398 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 399 | for (size_t k = 1; k <= 20; k += 5) { |
| 400 | for (uint32_t m = 1; m <= 4; m++) { |
| 401 | GemmMicrokernelTester() |
| 402 | .mr(4) |
| 403 | .nr(8) |
| 404 | .kr(1) |
| 405 | .sr(1) |
| 406 | .m(m) |
| 407 | .n(n) |
| 408 | .k(k) |
| 409 | .iterations(1) |
| 410 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 411 | } |
| 412 | } |
| 413 | } |
| 414 | } |
| 415 | |
| 416 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 417 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 418 | for (size_t k = 1; k <= 20; k += 5) { |
| 419 | for (uint32_t m = 1; m <= 4; m++) { |
| 420 | for (uint32_t n = 1; n <= 8; n++) { |
| 421 | GemmMicrokernelTester() |
| 422 | .mr(4) |
| 423 | .nr(8) |
| 424 | .kr(1) |
| 425 | .sr(1) |
| 426 | .m(m) |
| 427 | .n(n) |
| 428 | .k(k) |
| 429 | .cm_stride(11) |
| 430 | .iterations(1) |
| 431 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 432 | } |
| 433 | } |
| 434 | } |
| 435 | } |
| 436 | |
| 437 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, qmin) { |
| 438 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 439 | GemmMicrokernelTester() |
| 440 | .mr(4) |
| 441 | .nr(8) |
| 442 | .kr(1) |
| 443 | .sr(1) |
| 444 | .m(4) |
| 445 | .n(8) |
| 446 | .k(4) |
| 447 | .qmin(128) |
| 448 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 449 | } |
| 450 | |
| 451 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, qmax) { |
| 452 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 453 | GemmMicrokernelTester() |
| 454 | .mr(4) |
| 455 | .nr(8) |
| 456 | .kr(1) |
| 457 | .sr(1) |
| 458 | .m(4) |
| 459 | .n(8) |
| 460 | .k(4) |
| 461 | .qmax(128) |
| 462 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 463 | } |
| 464 | |
| 465 | TEST(F16_GEMM_4X8__NEONFP16ARITH_LD64, strided_cm) { |
| 466 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 467 | GemmMicrokernelTester() |
| 468 | .mr(4) |
| 469 | .nr(8) |
| 470 | .kr(1) |
| 471 | .sr(1) |
| 472 | .m(4) |
| 473 | .n(8) |
| 474 | .k(4) |
| 475 | .cm_stride(11) |
| 476 | .Test(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64); |
| 477 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 478 | #endif // XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 479 | |
| 480 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 481 | #if XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 482 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_eq_4) { |
| 483 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 484 | GemmMicrokernelTester() |
| 485 | .mr(6) |
| 486 | .nr(8) |
| 487 | .kr(1) |
| 488 | .sr(1) |
| 489 | .m(6) |
| 490 | .n(8) |
| 491 | .k(4) |
| 492 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 493 | } |
| 494 | |
| 495 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, strided_cn) { |
| 496 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 497 | GemmMicrokernelTester() |
| 498 | .mr(6) |
| 499 | .nr(8) |
| 500 | .kr(1) |
| 501 | .sr(1) |
| 502 | .m(6) |
| 503 | .n(8) |
| 504 | .k(4) |
| 505 | .cn_stride(11) |
| 506 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 507 | } |
| 508 | |
| 509 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_eq_4_strided_a) { |
| 510 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 511 | GemmMicrokernelTester() |
| 512 | .mr(6) |
| 513 | .nr(8) |
| 514 | .kr(1) |
| 515 | .sr(1) |
| 516 | .m(6) |
| 517 | .n(8) |
| 518 | .k(4) |
| 519 | .a_stride(7) |
| 520 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 521 | } |
| 522 | |
| 523 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 524 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 525 | for (uint32_t m = 1; m <= 6; m++) { |
| 526 | for (uint32_t n = 1; n <= 8; n++) { |
| 527 | GemmMicrokernelTester() |
| 528 | .mr(6) |
| 529 | .nr(8) |
| 530 | .kr(1) |
| 531 | .sr(1) |
| 532 | .m(m) |
| 533 | .n(n) |
| 534 | .k(4) |
| 535 | .iterations(1) |
| 536 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 537 | } |
| 538 | } |
| 539 | } |
| 540 | |
| 541 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 542 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 543 | for (uint32_t m = 1; m <= 6; m++) { |
| 544 | GemmMicrokernelTester() |
| 545 | .mr(6) |
| 546 | .nr(8) |
| 547 | .kr(1) |
| 548 | .sr(1) |
| 549 | .m(m) |
| 550 | .n(8) |
| 551 | .k(4) |
| 552 | .iterations(1) |
| 553 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 554 | } |
| 555 | } |
| 556 | |
| 557 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 558 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 559 | for (uint32_t n = 1; n <= 8; n++) { |
| 560 | GemmMicrokernelTester() |
| 561 | .mr(6) |
| 562 | .nr(8) |
| 563 | .kr(1) |
| 564 | .sr(1) |
| 565 | .m(6) |
| 566 | .n(n) |
| 567 | .k(4) |
| 568 | .iterations(1) |
| 569 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 570 | } |
| 571 | } |
| 572 | |
| 573 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_lt_4) { |
| 574 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 575 | for (size_t k = 1; k < 4; k++) { |
| 576 | GemmMicrokernelTester() |
| 577 | .mr(6) |
| 578 | .nr(8) |
| 579 | .kr(1) |
| 580 | .sr(1) |
| 581 | .m(6) |
| 582 | .n(8) |
| 583 | .k(k) |
| 584 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 585 | } |
| 586 | } |
| 587 | |
| 588 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_lt_4_strided_a) { |
| 589 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 590 | for (size_t k = 1; k < 4; k++) { |
| 591 | GemmMicrokernelTester() |
| 592 | .mr(6) |
| 593 | .nr(8) |
| 594 | .kr(1) |
| 595 | .sr(1) |
| 596 | .m(6) |
| 597 | .n(8) |
| 598 | .k(k) |
| 599 | .a_stride(7) |
| 600 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 601 | } |
| 602 | } |
| 603 | |
| 604 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 605 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 606 | for (size_t k = 1; k < 4; k++) { |
| 607 | for (uint32_t m = 1; m <= 6; m++) { |
| 608 | for (uint32_t n = 1; n <= 8; n++) { |
| 609 | GemmMicrokernelTester() |
| 610 | .mr(6) |
| 611 | .nr(8) |
| 612 | .kr(1) |
| 613 | .sr(1) |
| 614 | .m(m) |
| 615 | .n(n) |
| 616 | .k(k) |
| 617 | .iterations(1) |
| 618 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 619 | } |
| 620 | } |
| 621 | } |
| 622 | } |
| 623 | |
| 624 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_gt_4) { |
| 625 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 626 | for (size_t k = 5; k < 8; k++) { |
| 627 | GemmMicrokernelTester() |
| 628 | .mr(6) |
| 629 | .nr(8) |
| 630 | .kr(1) |
| 631 | .sr(1) |
| 632 | .m(6) |
| 633 | .n(8) |
| 634 | .k(k) |
| 635 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 636 | } |
| 637 | } |
| 638 | |
| 639 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_gt_4_strided_a) { |
| 640 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 641 | for (size_t k = 5; k < 8; k++) { |
| 642 | GemmMicrokernelTester() |
| 643 | .mr(6) |
| 644 | .nr(8) |
| 645 | .kr(1) |
| 646 | .sr(1) |
| 647 | .m(6) |
| 648 | .n(8) |
| 649 | .k(k) |
| 650 | .a_stride(11) |
| 651 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 652 | } |
| 653 | } |
| 654 | |
| 655 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 656 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 657 | for (size_t k = 5; k < 8; k++) { |
| 658 | for (uint32_t m = 1; m <= 6; m++) { |
| 659 | for (uint32_t n = 1; n <= 8; n++) { |
| 660 | GemmMicrokernelTester() |
| 661 | .mr(6) |
| 662 | .nr(8) |
| 663 | .kr(1) |
| 664 | .sr(1) |
| 665 | .m(m) |
| 666 | .n(n) |
| 667 | .k(k) |
| 668 | .iterations(1) |
| 669 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 670 | } |
| 671 | } |
| 672 | } |
| 673 | } |
| 674 | |
| 675 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_div_4) { |
| 676 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 677 | for (size_t k = 8; k <= 40; k += 4) { |
| 678 | GemmMicrokernelTester() |
| 679 | .mr(6) |
| 680 | .nr(8) |
| 681 | .kr(1) |
| 682 | .sr(1) |
| 683 | .m(6) |
| 684 | .n(8) |
| 685 | .k(k) |
| 686 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 687 | } |
| 688 | } |
| 689 | |
| 690 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_div_4_strided_a) { |
| 691 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 692 | for (size_t k = 8; k <= 40; k += 4) { |
| 693 | GemmMicrokernelTester() |
| 694 | .mr(6) |
| 695 | .nr(8) |
| 696 | .kr(1) |
| 697 | .sr(1) |
| 698 | .m(6) |
| 699 | .n(8) |
| 700 | .k(k) |
| 701 | .a_stride(43) |
| 702 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 703 | } |
| 704 | } |
| 705 | |
| 706 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 707 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 708 | for (size_t k = 8; k <= 40; k += 4) { |
| 709 | for (uint32_t m = 1; m <= 6; m++) { |
| 710 | for (uint32_t n = 1; n <= 8; n++) { |
| 711 | GemmMicrokernelTester() |
| 712 | .mr(6) |
| 713 | .nr(8) |
| 714 | .kr(1) |
| 715 | .sr(1) |
| 716 | .m(m) |
| 717 | .n(n) |
| 718 | .k(k) |
| 719 | .iterations(1) |
| 720 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 721 | } |
| 722 | } |
| 723 | } |
| 724 | } |
| 725 | |
| 726 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, n_gt_8) { |
| 727 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 728 | for (uint32_t n = 9; n < 16; n++) { |
| 729 | for (size_t k = 1; k <= 20; k += 5) { |
| 730 | GemmMicrokernelTester() |
| 731 | .mr(6) |
| 732 | .nr(8) |
| 733 | .kr(1) |
| 734 | .sr(1) |
| 735 | .m(6) |
| 736 | .n(8) |
| 737 | .k(k) |
| 738 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 739 | } |
| 740 | } |
| 741 | } |
| 742 | |
| 743 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
| 744 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 745 | for (uint32_t n = 9; n < 16; n++) { |
| 746 | for (size_t k = 1; k <= 20; k += 5) { |
| 747 | GemmMicrokernelTester() |
| 748 | .mr(6) |
| 749 | .nr(8) |
| 750 | .kr(1) |
| 751 | .sr(1) |
| 752 | .m(6) |
| 753 | .n(8) |
| 754 | .k(k) |
| 755 | .cn_stride(11) |
| 756 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 757 | } |
| 758 | } |
| 759 | } |
| 760 | |
| 761 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, n_gt_8_strided_a) { |
| 762 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 763 | for (uint32_t n = 9; n < 16; n++) { |
| 764 | for (size_t k = 1; k <= 20; k += 5) { |
| 765 | GemmMicrokernelTester() |
| 766 | .mr(6) |
| 767 | .nr(8) |
| 768 | .kr(1) |
| 769 | .sr(1) |
| 770 | .m(6) |
| 771 | .n(n) |
| 772 | .k(k) |
| 773 | .a_stride(23) |
| 774 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 775 | } |
| 776 | } |
| 777 | } |
| 778 | |
| 779 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
| 780 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 781 | for (uint32_t n = 9; n < 16; n++) { |
| 782 | for (size_t k = 1; k <= 20; k += 5) { |
| 783 | for (uint32_t m = 1; m <= 6; m++) { |
| 784 | GemmMicrokernelTester() |
| 785 | .mr(6) |
| 786 | .nr(8) |
| 787 | .kr(1) |
| 788 | .sr(1) |
| 789 | .m(m) |
| 790 | .n(n) |
| 791 | .k(k) |
| 792 | .iterations(1) |
| 793 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 794 | } |
| 795 | } |
| 796 | } |
| 797 | } |
| 798 | |
| 799 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, n_div_8) { |
| 800 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 801 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 802 | for (size_t k = 1; k <= 20; k += 5) { |
| 803 | GemmMicrokernelTester() |
| 804 | .mr(6) |
| 805 | .nr(8) |
| 806 | .kr(1) |
| 807 | .sr(1) |
| 808 | .m(6) |
| 809 | .n(8) |
| 810 | .k(k) |
| 811 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 812 | } |
| 813 | } |
| 814 | } |
| 815 | |
| 816 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
| 817 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 818 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 819 | for (size_t k = 1; k <= 20; k += 5) { |
| 820 | GemmMicrokernelTester() |
| 821 | .mr(6) |
| 822 | .nr(8) |
| 823 | .kr(1) |
| 824 | .sr(1) |
| 825 | .m(6) |
| 826 | .n(n) |
| 827 | .k(k) |
| 828 | .cn_stride(11) |
| 829 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 830 | } |
| 831 | } |
| 832 | } |
| 833 | |
| 834 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, n_div_8_strided_a) { |
| 835 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 836 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 837 | for (size_t k = 1; k <= 20; k += 5) { |
| 838 | GemmMicrokernelTester() |
| 839 | .mr(6) |
| 840 | .nr(8) |
| 841 | .kr(1) |
| 842 | .sr(1) |
| 843 | .m(6) |
| 844 | .n(n) |
| 845 | .k(k) |
| 846 | .a_stride(23) |
| 847 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 848 | } |
| 849 | } |
| 850 | } |
| 851 | |
| 852 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
| 853 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 854 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 855 | for (size_t k = 1; k <= 20; k += 5) { |
| 856 | for (uint32_t m = 1; m <= 6; m++) { |
| 857 | GemmMicrokernelTester() |
| 858 | .mr(6) |
| 859 | .nr(8) |
| 860 | .kr(1) |
| 861 | .sr(1) |
| 862 | .m(m) |
| 863 | .n(n) |
| 864 | .k(k) |
| 865 | .iterations(1) |
| 866 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 867 | } |
| 868 | } |
| 869 | } |
| 870 | } |
| 871 | |
| 872 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 873 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 874 | for (size_t k = 1; k <= 20; k += 5) { |
| 875 | for (uint32_t m = 1; m <= 6; m++) { |
| 876 | for (uint32_t n = 1; n <= 8; n++) { |
| 877 | GemmMicrokernelTester() |
| 878 | .mr(6) |
| 879 | .nr(8) |
| 880 | .kr(1) |
| 881 | .sr(1) |
| 882 | .m(m) |
| 883 | .n(n) |
| 884 | .k(k) |
| 885 | .cm_stride(11) |
| 886 | .iterations(1) |
| 887 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 888 | } |
| 889 | } |
| 890 | } |
| 891 | } |
| 892 | |
| 893 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, qmin) { |
| 894 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 895 | GemmMicrokernelTester() |
| 896 | .mr(6) |
| 897 | .nr(8) |
| 898 | .kr(1) |
| 899 | .sr(1) |
| 900 | .m(6) |
| 901 | .n(8) |
| 902 | .k(4) |
| 903 | .qmin(128) |
| 904 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 905 | } |
| 906 | |
| 907 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, qmax) { |
| 908 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 909 | GemmMicrokernelTester() |
| 910 | .mr(6) |
| 911 | .nr(8) |
| 912 | .kr(1) |
| 913 | .sr(1) |
| 914 | .m(6) |
| 915 | .n(8) |
| 916 | .k(4) |
| 917 | .qmax(128) |
| 918 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 919 | } |
| 920 | |
| 921 | TEST(F16_GEMM_6X8__NEONFP16ARITH_LD64, strided_cm) { |
| 922 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 923 | GemmMicrokernelTester() |
| 924 | .mr(6) |
| 925 | .nr(8) |
| 926 | .kr(1) |
| 927 | .sr(1) |
| 928 | .m(6) |
| 929 | .n(8) |
| 930 | .k(4) |
| 931 | .cm_stride(11) |
| 932 | .Test(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64); |
| 933 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 934 | #endif // XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 935 | |
| 936 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 937 | #if XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 938 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_eq_4) { |
| 939 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 940 | GemmMicrokernelTester() |
| 941 | .mr(8) |
| 942 | .nr(8) |
| 943 | .kr(1) |
| 944 | .sr(1) |
| 945 | .m(8) |
| 946 | .n(8) |
| 947 | .k(4) |
| 948 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 949 | } |
| 950 | |
| 951 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, strided_cn) { |
| 952 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 953 | GemmMicrokernelTester() |
| 954 | .mr(8) |
| 955 | .nr(8) |
| 956 | .kr(1) |
| 957 | .sr(1) |
| 958 | .m(8) |
| 959 | .n(8) |
| 960 | .k(4) |
| 961 | .cn_stride(11) |
| 962 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 963 | } |
| 964 | |
| 965 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_eq_4_strided_a) { |
| 966 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 967 | GemmMicrokernelTester() |
| 968 | .mr(8) |
| 969 | .nr(8) |
| 970 | .kr(1) |
| 971 | .sr(1) |
| 972 | .m(8) |
| 973 | .n(8) |
| 974 | .k(4) |
| 975 | .a_stride(7) |
| 976 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 977 | } |
| 978 | |
| 979 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
| 980 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 981 | for (uint32_t m = 1; m <= 8; m++) { |
| 982 | for (uint32_t n = 1; n <= 8; n++) { |
| 983 | GemmMicrokernelTester() |
| 984 | .mr(8) |
| 985 | .nr(8) |
| 986 | .kr(1) |
| 987 | .sr(1) |
| 988 | .m(m) |
| 989 | .n(n) |
| 990 | .k(4) |
| 991 | .iterations(1) |
| 992 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 993 | } |
| 994 | } |
| 995 | } |
| 996 | |
| 997 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
| 998 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 999 | for (uint32_t m = 1; m <= 8; m++) { |
| 1000 | GemmMicrokernelTester() |
| 1001 | .mr(8) |
| 1002 | .nr(8) |
| 1003 | .kr(1) |
| 1004 | .sr(1) |
| 1005 | .m(m) |
| 1006 | .n(8) |
| 1007 | .k(4) |
| 1008 | .iterations(1) |
| 1009 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1010 | } |
| 1011 | } |
| 1012 | |
| 1013 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
| 1014 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1015 | for (uint32_t n = 1; n <= 8; n++) { |
| 1016 | GemmMicrokernelTester() |
| 1017 | .mr(8) |
| 1018 | .nr(8) |
| 1019 | .kr(1) |
| 1020 | .sr(1) |
| 1021 | .m(8) |
| 1022 | .n(n) |
| 1023 | .k(4) |
| 1024 | .iterations(1) |
| 1025 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1026 | } |
| 1027 | } |
| 1028 | |
| 1029 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_lt_4) { |
| 1030 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1031 | for (size_t k = 1; k < 4; k++) { |
| 1032 | GemmMicrokernelTester() |
| 1033 | .mr(8) |
| 1034 | .nr(8) |
| 1035 | .kr(1) |
| 1036 | .sr(1) |
| 1037 | .m(8) |
| 1038 | .n(8) |
| 1039 | .k(k) |
| 1040 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1041 | } |
| 1042 | } |
| 1043 | |
| 1044 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_lt_4_strided_a) { |
| 1045 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1046 | for (size_t k = 1; k < 4; k++) { |
| 1047 | GemmMicrokernelTester() |
| 1048 | .mr(8) |
| 1049 | .nr(8) |
| 1050 | .kr(1) |
| 1051 | .sr(1) |
| 1052 | .m(8) |
| 1053 | .n(8) |
| 1054 | .k(k) |
| 1055 | .a_stride(7) |
| 1056 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1057 | } |
| 1058 | } |
| 1059 | |
| 1060 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
| 1061 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1062 | for (size_t k = 1; k < 4; k++) { |
| 1063 | for (uint32_t m = 1; m <= 8; m++) { |
| 1064 | for (uint32_t n = 1; n <= 8; n++) { |
| 1065 | GemmMicrokernelTester() |
| 1066 | .mr(8) |
| 1067 | .nr(8) |
| 1068 | .kr(1) |
| 1069 | .sr(1) |
| 1070 | .m(m) |
| 1071 | .n(n) |
| 1072 | .k(k) |
| 1073 | .iterations(1) |
| 1074 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1075 | } |
| 1076 | } |
| 1077 | } |
| 1078 | } |
| 1079 | |
| 1080 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_gt_4) { |
| 1081 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1082 | for (size_t k = 5; k < 8; k++) { |
| 1083 | GemmMicrokernelTester() |
| 1084 | .mr(8) |
| 1085 | .nr(8) |
| 1086 | .kr(1) |
| 1087 | .sr(1) |
| 1088 | .m(8) |
| 1089 | .n(8) |
| 1090 | .k(k) |
| 1091 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1092 | } |
| 1093 | } |
| 1094 | |
| 1095 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_gt_4_strided_a) { |
| 1096 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1097 | for (size_t k = 5; k < 8; k++) { |
| 1098 | GemmMicrokernelTester() |
| 1099 | .mr(8) |
| 1100 | .nr(8) |
| 1101 | .kr(1) |
| 1102 | .sr(1) |
| 1103 | .m(8) |
| 1104 | .n(8) |
| 1105 | .k(k) |
| 1106 | .a_stride(11) |
| 1107 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1108 | } |
| 1109 | } |
| 1110 | |
| 1111 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
| 1112 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1113 | for (size_t k = 5; k < 8; k++) { |
| 1114 | for (uint32_t m = 1; m <= 8; m++) { |
| 1115 | for (uint32_t n = 1; n <= 8; n++) { |
| 1116 | GemmMicrokernelTester() |
| 1117 | .mr(8) |
| 1118 | .nr(8) |
| 1119 | .kr(1) |
| 1120 | .sr(1) |
| 1121 | .m(m) |
| 1122 | .n(n) |
| 1123 | .k(k) |
| 1124 | .iterations(1) |
| 1125 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1126 | } |
| 1127 | } |
| 1128 | } |
| 1129 | } |
| 1130 | |
| 1131 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_div_4) { |
| 1132 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1133 | for (size_t k = 8; k <= 40; k += 4) { |
| 1134 | GemmMicrokernelTester() |
| 1135 | .mr(8) |
| 1136 | .nr(8) |
| 1137 | .kr(1) |
| 1138 | .sr(1) |
| 1139 | .m(8) |
| 1140 | .n(8) |
| 1141 | .k(k) |
| 1142 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1143 | } |
| 1144 | } |
| 1145 | |
| 1146 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_div_4_strided_a) { |
| 1147 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1148 | for (size_t k = 8; k <= 40; k += 4) { |
| 1149 | GemmMicrokernelTester() |
| 1150 | .mr(8) |
| 1151 | .nr(8) |
| 1152 | .kr(1) |
| 1153 | .sr(1) |
| 1154 | .m(8) |
| 1155 | .n(8) |
| 1156 | .k(k) |
| 1157 | .a_stride(43) |
| 1158 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1159 | } |
| 1160 | } |
| 1161 | |
| 1162 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
| 1163 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1164 | for (size_t k = 8; k <= 40; k += 4) { |
| 1165 | for (uint32_t m = 1; m <= 8; m++) { |
| 1166 | for (uint32_t n = 1; n <= 8; n++) { |
| 1167 | GemmMicrokernelTester() |
| 1168 | .mr(8) |
| 1169 | .nr(8) |
| 1170 | .kr(1) |
| 1171 | .sr(1) |
| 1172 | .m(m) |
| 1173 | .n(n) |
| 1174 | .k(k) |
| 1175 | .iterations(1) |
| 1176 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1177 | } |
| 1178 | } |
| 1179 | } |
| 1180 | } |
| 1181 | |
| 1182 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, n_gt_8) { |
| 1183 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1184 | for (uint32_t n = 9; n < 16; n++) { |
| 1185 | for (size_t k = 1; k <= 20; k += 5) { |
| 1186 | GemmMicrokernelTester() |
| 1187 | .mr(8) |
| 1188 | .nr(8) |
| 1189 | .kr(1) |
| 1190 | .sr(1) |
| 1191 | .m(8) |
| 1192 | .n(8) |
| 1193 | .k(k) |
| 1194 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1195 | } |
| 1196 | } |
| 1197 | } |
| 1198 | |
| 1199 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
| 1200 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1201 | for (uint32_t n = 9; n < 16; n++) { |
| 1202 | for (size_t k = 1; k <= 20; k += 5) { |
| 1203 | GemmMicrokernelTester() |
| 1204 | .mr(8) |
| 1205 | .nr(8) |
| 1206 | .kr(1) |
| 1207 | .sr(1) |
| 1208 | .m(8) |
| 1209 | .n(8) |
| 1210 | .k(k) |
| 1211 | .cn_stride(11) |
| 1212 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1213 | } |
| 1214 | } |
| 1215 | } |
| 1216 | |
| 1217 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, n_gt_8_strided_a) { |
| 1218 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1219 | for (uint32_t n = 9; n < 16; n++) { |
| 1220 | for (size_t k = 1; k <= 20; k += 5) { |
| 1221 | GemmMicrokernelTester() |
| 1222 | .mr(8) |
| 1223 | .nr(8) |
| 1224 | .kr(1) |
| 1225 | .sr(1) |
| 1226 | .m(8) |
| 1227 | .n(n) |
| 1228 | .k(k) |
| 1229 | .a_stride(23) |
| 1230 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1231 | } |
| 1232 | } |
| 1233 | } |
| 1234 | |
| 1235 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
| 1236 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1237 | for (uint32_t n = 9; n < 16; n++) { |
| 1238 | for (size_t k = 1; k <= 20; k += 5) { |
| 1239 | for (uint32_t m = 1; m <= 8; m++) { |
| 1240 | GemmMicrokernelTester() |
| 1241 | .mr(8) |
| 1242 | .nr(8) |
| 1243 | .kr(1) |
| 1244 | .sr(1) |
| 1245 | .m(m) |
| 1246 | .n(n) |
| 1247 | .k(k) |
| 1248 | .iterations(1) |
| 1249 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1250 | } |
| 1251 | } |
| 1252 | } |
| 1253 | } |
| 1254 | |
| 1255 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, n_div_8) { |
| 1256 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1257 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1258 | for (size_t k = 1; k <= 20; k += 5) { |
| 1259 | GemmMicrokernelTester() |
| 1260 | .mr(8) |
| 1261 | .nr(8) |
| 1262 | .kr(1) |
| 1263 | .sr(1) |
| 1264 | .m(8) |
| 1265 | .n(8) |
| 1266 | .k(k) |
| 1267 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1268 | } |
| 1269 | } |
| 1270 | } |
| 1271 | |
| 1272 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
| 1273 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1274 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1275 | for (size_t k = 1; k <= 20; k += 5) { |
| 1276 | GemmMicrokernelTester() |
| 1277 | .mr(8) |
| 1278 | .nr(8) |
| 1279 | .kr(1) |
| 1280 | .sr(1) |
| 1281 | .m(8) |
| 1282 | .n(n) |
| 1283 | .k(k) |
| 1284 | .cn_stride(11) |
| 1285 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1286 | } |
| 1287 | } |
| 1288 | } |
| 1289 | |
| 1290 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, n_div_8_strided_a) { |
| 1291 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1292 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1293 | for (size_t k = 1; k <= 20; k += 5) { |
| 1294 | GemmMicrokernelTester() |
| 1295 | .mr(8) |
| 1296 | .nr(8) |
| 1297 | .kr(1) |
| 1298 | .sr(1) |
| 1299 | .m(8) |
| 1300 | .n(n) |
| 1301 | .k(k) |
| 1302 | .a_stride(23) |
| 1303 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1304 | } |
| 1305 | } |
| 1306 | } |
| 1307 | |
| 1308 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
| 1309 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1310 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1311 | for (size_t k = 1; k <= 20; k += 5) { |
| 1312 | for (uint32_t m = 1; m <= 8; m++) { |
| 1313 | GemmMicrokernelTester() |
| 1314 | .mr(8) |
| 1315 | .nr(8) |
| 1316 | .kr(1) |
| 1317 | .sr(1) |
| 1318 | .m(m) |
| 1319 | .n(n) |
| 1320 | .k(k) |
| 1321 | .iterations(1) |
| 1322 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1323 | } |
| 1324 | } |
| 1325 | } |
| 1326 | } |
| 1327 | |
| 1328 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
| 1329 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1330 | for (size_t k = 1; k <= 20; k += 5) { |
| 1331 | for (uint32_t m = 1; m <= 8; m++) { |
| 1332 | for (uint32_t n = 1; n <= 8; n++) { |
| 1333 | GemmMicrokernelTester() |
| 1334 | .mr(8) |
| 1335 | .nr(8) |
| 1336 | .kr(1) |
| 1337 | .sr(1) |
| 1338 | .m(m) |
| 1339 | .n(n) |
| 1340 | .k(k) |
| 1341 | .cm_stride(11) |
| 1342 | .iterations(1) |
| 1343 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1344 | } |
| 1345 | } |
| 1346 | } |
| 1347 | } |
| 1348 | |
| 1349 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, qmin) { |
| 1350 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1351 | GemmMicrokernelTester() |
| 1352 | .mr(8) |
| 1353 | .nr(8) |
| 1354 | .kr(1) |
| 1355 | .sr(1) |
| 1356 | .m(8) |
| 1357 | .n(8) |
| 1358 | .k(4) |
| 1359 | .qmin(128) |
| 1360 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1361 | } |
| 1362 | |
| 1363 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, qmax) { |
| 1364 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1365 | GemmMicrokernelTester() |
| 1366 | .mr(8) |
| 1367 | .nr(8) |
| 1368 | .kr(1) |
| 1369 | .sr(1) |
| 1370 | .m(8) |
| 1371 | .n(8) |
| 1372 | .k(4) |
| 1373 | .qmax(128) |
| 1374 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1375 | } |
| 1376 | |
| 1377 | TEST(F16_GEMM_8X8__NEONFP16ARITH_LD64, strided_cm) { |
| 1378 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1379 | GemmMicrokernelTester() |
| 1380 | .mr(8) |
| 1381 | .nr(8) |
| 1382 | .kr(1) |
| 1383 | .sr(1) |
| 1384 | .m(8) |
| 1385 | .n(8) |
| 1386 | .k(4) |
| 1387 | .cm_stride(11) |
| 1388 | .Test(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64); |
| 1389 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1390 | #endif // XNN_ARCH_ARM64 |