XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | // |
| 6 | // Auto-generated file. Do not edit! |
| 7 | // Specification: test/f32-spmm.yaml |
| 8 | // Generator: tools/generate-spmm-test.py |
| 9 | |
| 10 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 11 | #include <gtest/gtest.h> |
| 12 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 13 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | #include <xnnpack/isa-checks.h> |
| 15 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 16 | #include <xnnpack/spmm.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 17 | #include "spmm-microkernel-tester.h" |
| 18 | |
| 19 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 20 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 21 | TEST(F32_SPMM_4X1__NEONFMA, k_eq_1) { |
| 22 | TEST_REQUIRES_ARM_NEON_FMA; |
| 23 | SpMMMicrokernelTester() |
| 24 | .mr(4) |
| 25 | .nr(1) |
| 26 | .m(4) |
| 27 | .n(1) |
| 28 | .k(1) |
| 29 | .sparsity(0.0f) |
| 30 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 31 | } |
| 32 | |
| 33 | TEST(F32_SPMM_4X1__NEONFMA, k_gt_1) { |
| 34 | TEST_REQUIRES_ARM_NEON_FMA; |
| 35 | for (size_t k = 2; k < 10; k++) { |
| 36 | SpMMMicrokernelTester() |
| 37 | .mr(4) |
| 38 | .nr(1) |
| 39 | .m(4) |
| 40 | .n(1) |
| 41 | .k(k) |
| 42 | .sparsity(0.0f) |
| 43 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 44 | } |
| 45 | } |
| 46 | |
| 47 | TEST(F32_SPMM_4X1__NEONFMA, n_gt_1) { |
| 48 | TEST_REQUIRES_ARM_NEON_FMA; |
| 49 | for (uint32_t n = 2; n < 10; n++) { |
| 50 | for (size_t k = 1; k <= 5; k += 2) { |
| 51 | SpMMMicrokernelTester() |
| 52 | .mr(4) |
| 53 | .nr(1) |
| 54 | .m(4) |
| 55 | .n(n) |
| 56 | .k(k) |
| 57 | .sparsity(0.0f) |
| 58 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 59 | } |
| 60 | } |
| 61 | } |
| 62 | |
| 63 | TEST(F32_SPMM_4X1__NEONFMA, m_lt_4) { |
| 64 | TEST_REQUIRES_ARM_NEON_FMA; |
| 65 | for (uint32_t m = 1; m < 4; m++) { |
| 66 | for (uint32_t n = 1; n < 10; n += 2) { |
| 67 | for (size_t k = 1; k <= 5; k += 2) { |
| 68 | SpMMMicrokernelTester() |
| 69 | .mr(4) |
| 70 | .nr(1) |
| 71 | .m(m) |
| 72 | .n(n) |
| 73 | .k(k) |
| 74 | .sparsity(0.0f) |
| 75 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 76 | } |
| 77 | } |
| 78 | } |
| 79 | } |
| 80 | |
| 81 | TEST(F32_SPMM_4X1__NEONFMA, m_div_4) { |
| 82 | TEST_REQUIRES_ARM_NEON_FMA; |
| 83 | for (uint32_t m = 8; m <= 12; m += 4) { |
| 84 | for (uint32_t n = 1; n < 10; n += 2) { |
| 85 | for (size_t k = 1; k <= 5; k += 2) { |
| 86 | SpMMMicrokernelTester() |
| 87 | .mr(4) |
| 88 | .nr(1) |
| 89 | .m(m) |
| 90 | .n(n) |
| 91 | .k(k) |
| 92 | .sparsity(0.0f) |
| 93 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 94 | } |
| 95 | } |
| 96 | } |
| 97 | } |
| 98 | |
| 99 | TEST(F32_SPMM_4X1__NEONFMA, m_gt_4) { |
| 100 | TEST_REQUIRES_ARM_NEON_FMA; |
| 101 | for (uint32_t m = 5; m < 8; m++) { |
| 102 | for (uint32_t n = 1; n < 10; n += 2) { |
| 103 | for (size_t k = 1; k <= 5; k += 2) { |
| 104 | SpMMMicrokernelTester() |
| 105 | .mr(4) |
| 106 | .nr(1) |
| 107 | .m(m) |
| 108 | .n(n) |
| 109 | .k(k) |
| 110 | .sparsity(0.0f) |
| 111 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 112 | } |
| 113 | } |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | TEST(F32_SPMM_4X1__NEONFMA, qmin) { |
| 118 | TEST_REQUIRES_ARM_NEON_FMA; |
| 119 | for (uint32_t n = 1; n < 10; n += 2) { |
| 120 | for (size_t k = 1; k <= 5; k += 2) { |
| 121 | SpMMMicrokernelTester() |
| 122 | .mr(4) |
| 123 | .nr(1) |
| 124 | .m(8) |
| 125 | .n(n) |
| 126 | .k(k) |
| 127 | .sparsity(0.0f) |
| 128 | .qmin(128) |
| 129 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 130 | } |
| 131 | } |
| 132 | } |
| 133 | |
| 134 | TEST(F32_SPMM_4X1__NEONFMA, qmax) { |
| 135 | TEST_REQUIRES_ARM_NEON_FMA; |
| 136 | for (uint32_t n = 1; n < 10; n += 2) { |
| 137 | for (size_t k = 1; k <= 5; k += 2) { |
| 138 | SpMMMicrokernelTester() |
| 139 | .mr(4) |
| 140 | .nr(1) |
| 141 | .m(8) |
| 142 | .n(n) |
| 143 | .k(k) |
| 144 | .sparsity(0.0f) |
| 145 | .qmax(128) |
| 146 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 147 | } |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | TEST(F32_SPMM_4X1__NEONFMA, half_sparse) { |
| 152 | TEST_REQUIRES_ARM_NEON_FMA; |
| 153 | for (uint32_t n = 1; n < 10; n += 2) { |
| 154 | for (size_t k = 1; k <= 5; k += 2) { |
| 155 | SpMMMicrokernelTester() |
| 156 | .mr(4) |
| 157 | .nr(1) |
| 158 | .m(8) |
| 159 | .n(n) |
| 160 | .k(k) |
| 161 | .sparsity(0.5f) |
| 162 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 163 | } |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | TEST(F32_SPMM_4X1__NEONFMA, zero_weights) { |
| 168 | TEST_REQUIRES_ARM_NEON_FMA; |
| 169 | for (uint32_t n = 1; n < 10; n += 2) { |
| 170 | for (size_t k = 1; k <= 5; k += 2) { |
| 171 | SpMMMicrokernelTester() |
| 172 | .mr(4) |
| 173 | .nr(1) |
| 174 | .m(8) |
| 175 | .n(n) |
| 176 | .k(k) |
| 177 | .sparsity(1.0f) |
| 178 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma); |
| 179 | } |
| 180 | } |
| 181 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 182 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 183 | |
| 184 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 185 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 186 | TEST(F32_SPMM_4X2__NEONFMA, k_eq_1) { |
| 187 | TEST_REQUIRES_ARM_NEON_FMA; |
| 188 | SpMMMicrokernelTester() |
| 189 | .mr(4) |
| 190 | .nr(2) |
| 191 | .m(4) |
| 192 | .n(2) |
| 193 | .k(1) |
| 194 | .sparsity(0.0f) |
| 195 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 196 | } |
| 197 | |
| 198 | TEST(F32_SPMM_4X2__NEONFMA, k_eq_1_subtile) { |
| 199 | TEST_REQUIRES_ARM_NEON_FMA; |
| 200 | for (uint32_t n = 1; n <= 2; n++) { |
| 201 | SpMMMicrokernelTester() |
| 202 | .mr(4) |
| 203 | .nr(2) |
| 204 | .m(4) |
| 205 | .n(n) |
| 206 | .k(1) |
| 207 | .sparsity(0.0f) |
| 208 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | TEST(F32_SPMM_4X2__NEONFMA, k_gt_1) { |
| 213 | TEST_REQUIRES_ARM_NEON_FMA; |
| 214 | for (size_t k = 2; k < 10; k++) { |
| 215 | SpMMMicrokernelTester() |
| 216 | .mr(4) |
| 217 | .nr(2) |
| 218 | .m(4) |
| 219 | .n(2) |
| 220 | .k(k) |
| 221 | .sparsity(0.0f) |
| 222 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 223 | } |
| 224 | } |
| 225 | |
| 226 | TEST(F32_SPMM_4X2__NEONFMA, k_gt_1_subtile) { |
| 227 | TEST_REQUIRES_ARM_NEON_FMA; |
| 228 | for (size_t k = 2; k < 10; k++) { |
| 229 | for (uint32_t n = 1; n <= 2; n++) { |
| 230 | SpMMMicrokernelTester() |
| 231 | .mr(4) |
| 232 | .nr(2) |
| 233 | .m(4) |
| 234 | .n(n) |
| 235 | .k(k) |
| 236 | .sparsity(0.0f) |
| 237 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 238 | } |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | TEST(F32_SPMM_4X2__NEONFMA, n_gt_2) { |
| 243 | TEST_REQUIRES_ARM_NEON_FMA; |
| 244 | for (uint32_t n = 3; n < 10; n++) { |
| 245 | for (size_t k = 1; k <= 5; k += 2) { |
| 246 | SpMMMicrokernelTester() |
| 247 | .mr(4) |
| 248 | .nr(2) |
| 249 | .m(4) |
| 250 | .n(n) |
| 251 | .k(k) |
| 252 | .sparsity(0.0f) |
| 253 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 254 | } |
| 255 | } |
| 256 | } |
| 257 | |
| 258 | TEST(F32_SPMM_4X2__NEONFMA, n_div_2) { |
| 259 | TEST_REQUIRES_ARM_NEON_FMA; |
| 260 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 261 | for (size_t k = 1; k <= 5; k += 2) { |
| 262 | SpMMMicrokernelTester() |
| 263 | .mr(4) |
| 264 | .nr(2) |
| 265 | .m(4) |
| 266 | .n(n) |
| 267 | .k(k) |
| 268 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 269 | } |
| 270 | } |
| 271 | } |
| 272 | |
| 273 | TEST(F32_SPMM_4X2__NEONFMA, m_lt_4) { |
| 274 | TEST_REQUIRES_ARM_NEON_FMA; |
| 275 | for (uint32_t m = 1; m < 4; m++) { |
| 276 | for (uint32_t n = 1; n < 10; n += 3) { |
| 277 | for (size_t k = 1; k <= 5; k += 2) { |
| 278 | SpMMMicrokernelTester() |
| 279 | .mr(4) |
| 280 | .nr(2) |
| 281 | .m(m) |
| 282 | .n(n) |
| 283 | .k(k) |
| 284 | .sparsity(0.0f) |
| 285 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 286 | } |
| 287 | } |
| 288 | } |
| 289 | } |
| 290 | |
| 291 | TEST(F32_SPMM_4X2__NEONFMA, m_div_4) { |
| 292 | TEST_REQUIRES_ARM_NEON_FMA; |
| 293 | for (uint32_t m = 8; m <= 12; m += 4) { |
| 294 | for (uint32_t n = 1; n < 10; n += 3) { |
| 295 | for (size_t k = 1; k <= 5; k += 2) { |
| 296 | SpMMMicrokernelTester() |
| 297 | .mr(4) |
| 298 | .nr(2) |
| 299 | .m(m) |
| 300 | .n(n) |
| 301 | .k(k) |
| 302 | .sparsity(0.0f) |
| 303 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 304 | } |
| 305 | } |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | TEST(F32_SPMM_4X2__NEONFMA, m_gt_4) { |
| 310 | TEST_REQUIRES_ARM_NEON_FMA; |
| 311 | for (uint32_t m = 5; m < 8; m++) { |
| 312 | for (uint32_t n = 1; n < 10; n += 3) { |
| 313 | for (size_t k = 1; k <= 5; k += 2) { |
| 314 | SpMMMicrokernelTester() |
| 315 | .mr(4) |
| 316 | .nr(2) |
| 317 | .m(m) |
| 318 | .n(n) |
| 319 | .k(k) |
| 320 | .sparsity(0.0f) |
| 321 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 322 | } |
| 323 | } |
| 324 | } |
| 325 | } |
| 326 | |
| 327 | TEST(F32_SPMM_4X2__NEONFMA, qmin) { |
| 328 | TEST_REQUIRES_ARM_NEON_FMA; |
| 329 | for (uint32_t n = 1; n < 10; n += 3) { |
| 330 | for (size_t k = 1; k <= 5; k += 2) { |
| 331 | SpMMMicrokernelTester() |
| 332 | .mr(4) |
| 333 | .nr(2) |
| 334 | .m(8) |
| 335 | .n(n) |
| 336 | .k(k) |
| 337 | .sparsity(0.0f) |
| 338 | .qmin(128) |
| 339 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 340 | } |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | TEST(F32_SPMM_4X2__NEONFMA, qmax) { |
| 345 | TEST_REQUIRES_ARM_NEON_FMA; |
| 346 | for (uint32_t n = 1; n < 10; n += 3) { |
| 347 | for (size_t k = 1; k <= 5; k += 2) { |
| 348 | SpMMMicrokernelTester() |
| 349 | .mr(4) |
| 350 | .nr(2) |
| 351 | .m(8) |
| 352 | .n(n) |
| 353 | .k(k) |
| 354 | .sparsity(0.0f) |
| 355 | .qmax(128) |
| 356 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 357 | } |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | TEST(F32_SPMM_4X2__NEONFMA, half_sparse) { |
| 362 | TEST_REQUIRES_ARM_NEON_FMA; |
| 363 | for (uint32_t n = 1; n < 10; n += 3) { |
| 364 | for (size_t k = 1; k <= 5; k += 2) { |
| 365 | SpMMMicrokernelTester() |
| 366 | .mr(4) |
| 367 | .nr(2) |
| 368 | .m(8) |
| 369 | .n(n) |
| 370 | .k(k) |
| 371 | .sparsity(0.5f) |
| 372 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 373 | } |
| 374 | } |
| 375 | } |
| 376 | |
| 377 | TEST(F32_SPMM_4X2__NEONFMA, zero_weights) { |
| 378 | TEST_REQUIRES_ARM_NEON_FMA; |
| 379 | for (uint32_t n = 1; n < 10; n += 3) { |
| 380 | for (size_t k = 1; k <= 5; k += 2) { |
| 381 | SpMMMicrokernelTester() |
| 382 | .mr(4) |
| 383 | .nr(2) |
| 384 | .m(8) |
| 385 | .n(n) |
| 386 | .k(k) |
| 387 | .sparsity(1.0f) |
| 388 | .Test(xnn_f32_spmm_ukernel_4x2__neonfma); |
| 389 | } |
| 390 | } |
| 391 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 392 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 393 | |
| 394 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 395 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 396 | TEST(F32_SPMM_4X4__NEONFMA, k_eq_1) { |
| 397 | TEST_REQUIRES_ARM_NEON_FMA; |
| 398 | SpMMMicrokernelTester() |
| 399 | .mr(4) |
| 400 | .nr(4) |
| 401 | .m(4) |
| 402 | .n(4) |
| 403 | .k(1) |
| 404 | .sparsity(0.0f) |
| 405 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 406 | } |
| 407 | |
| 408 | TEST(F32_SPMM_4X4__NEONFMA, k_eq_1_subtile) { |
| 409 | TEST_REQUIRES_ARM_NEON_FMA; |
| 410 | for (uint32_t n = 1; n <= 4; n++) { |
| 411 | SpMMMicrokernelTester() |
| 412 | .mr(4) |
| 413 | .nr(4) |
| 414 | .m(4) |
| 415 | .n(n) |
| 416 | .k(1) |
| 417 | .sparsity(0.0f) |
| 418 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 419 | } |
| 420 | } |
| 421 | |
| 422 | TEST(F32_SPMM_4X4__NEONFMA, k_gt_1) { |
| 423 | TEST_REQUIRES_ARM_NEON_FMA; |
| 424 | for (size_t k = 2; k < 10; k++) { |
| 425 | SpMMMicrokernelTester() |
| 426 | .mr(4) |
| 427 | .nr(4) |
| 428 | .m(4) |
| 429 | .n(4) |
| 430 | .k(k) |
| 431 | .sparsity(0.0f) |
| 432 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 433 | } |
| 434 | } |
| 435 | |
| 436 | TEST(F32_SPMM_4X4__NEONFMA, k_gt_1_subtile) { |
| 437 | TEST_REQUIRES_ARM_NEON_FMA; |
| 438 | for (size_t k = 2; k < 10; k++) { |
| 439 | for (uint32_t n = 1; n <= 4; n++) { |
| 440 | SpMMMicrokernelTester() |
| 441 | .mr(4) |
| 442 | .nr(4) |
| 443 | .m(4) |
| 444 | .n(n) |
| 445 | .k(k) |
| 446 | .sparsity(0.0f) |
| 447 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 448 | } |
| 449 | } |
| 450 | } |
| 451 | |
| 452 | TEST(F32_SPMM_4X4__NEONFMA, n_gt_4) { |
| 453 | TEST_REQUIRES_ARM_NEON_FMA; |
| 454 | for (uint32_t n = 5; n < 10; n++) { |
| 455 | for (size_t k = 1; k <= 5; k += 2) { |
| 456 | SpMMMicrokernelTester() |
| 457 | .mr(4) |
| 458 | .nr(4) |
| 459 | .m(4) |
| 460 | .n(n) |
| 461 | .k(k) |
| 462 | .sparsity(0.0f) |
| 463 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 464 | } |
| 465 | } |
| 466 | } |
| 467 | |
| 468 | TEST(F32_SPMM_4X4__NEONFMA, n_div_4) { |
| 469 | TEST_REQUIRES_ARM_NEON_FMA; |
| 470 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 471 | for (size_t k = 1; k <= 5; k += 2) { |
| 472 | SpMMMicrokernelTester() |
| 473 | .mr(4) |
| 474 | .nr(4) |
| 475 | .m(4) |
| 476 | .n(n) |
| 477 | .k(k) |
| 478 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 479 | } |
| 480 | } |
| 481 | } |
| 482 | |
| 483 | TEST(F32_SPMM_4X4__NEONFMA, m_lt_4) { |
| 484 | TEST_REQUIRES_ARM_NEON_FMA; |
| 485 | for (uint32_t m = 1; m < 4; m++) { |
| 486 | for (uint32_t n = 1; n < 20; n += 5) { |
| 487 | for (size_t k = 1; k <= 5; k += 2) { |
| 488 | SpMMMicrokernelTester() |
| 489 | .mr(4) |
| 490 | .nr(4) |
| 491 | .m(m) |
| 492 | .n(n) |
| 493 | .k(k) |
| 494 | .sparsity(0.0f) |
| 495 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 496 | } |
| 497 | } |
| 498 | } |
| 499 | } |
| 500 | |
| 501 | TEST(F32_SPMM_4X4__NEONFMA, m_div_4) { |
| 502 | TEST_REQUIRES_ARM_NEON_FMA; |
| 503 | for (uint32_t m = 8; m <= 12; m += 4) { |
| 504 | for (uint32_t n = 1; n < 20; n += 5) { |
| 505 | for (size_t k = 1; k <= 5; k += 2) { |
| 506 | SpMMMicrokernelTester() |
| 507 | .mr(4) |
| 508 | .nr(4) |
| 509 | .m(m) |
| 510 | .n(n) |
| 511 | .k(k) |
| 512 | .sparsity(0.0f) |
| 513 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 514 | } |
| 515 | } |
| 516 | } |
| 517 | } |
| 518 | |
| 519 | TEST(F32_SPMM_4X4__NEONFMA, m_gt_4) { |
| 520 | TEST_REQUIRES_ARM_NEON_FMA; |
| 521 | for (uint32_t m = 5; m < 8; m++) { |
| 522 | for (uint32_t n = 1; n < 20; n += 5) { |
| 523 | for (size_t k = 1; k <= 5; k += 2) { |
| 524 | SpMMMicrokernelTester() |
| 525 | .mr(4) |
| 526 | .nr(4) |
| 527 | .m(m) |
| 528 | .n(n) |
| 529 | .k(k) |
| 530 | .sparsity(0.0f) |
| 531 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 532 | } |
| 533 | } |
| 534 | } |
| 535 | } |
| 536 | |
| 537 | TEST(F32_SPMM_4X4__NEONFMA, qmin) { |
| 538 | TEST_REQUIRES_ARM_NEON_FMA; |
| 539 | for (uint32_t n = 1; n < 20; n += 5) { |
| 540 | for (size_t k = 1; k <= 5; k += 2) { |
| 541 | SpMMMicrokernelTester() |
| 542 | .mr(4) |
| 543 | .nr(4) |
| 544 | .m(8) |
| 545 | .n(n) |
| 546 | .k(k) |
| 547 | .sparsity(0.0f) |
| 548 | .qmin(128) |
| 549 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 550 | } |
| 551 | } |
| 552 | } |
| 553 | |
| 554 | TEST(F32_SPMM_4X4__NEONFMA, qmax) { |
| 555 | TEST_REQUIRES_ARM_NEON_FMA; |
| 556 | for (uint32_t n = 1; n < 20; n += 5) { |
| 557 | for (size_t k = 1; k <= 5; k += 2) { |
| 558 | SpMMMicrokernelTester() |
| 559 | .mr(4) |
| 560 | .nr(4) |
| 561 | .m(8) |
| 562 | .n(n) |
| 563 | .k(k) |
| 564 | .sparsity(0.0f) |
| 565 | .qmax(128) |
| 566 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 567 | } |
| 568 | } |
| 569 | } |
| 570 | |
| 571 | TEST(F32_SPMM_4X4__NEONFMA, half_sparse) { |
| 572 | TEST_REQUIRES_ARM_NEON_FMA; |
| 573 | for (uint32_t n = 1; n < 20; n += 5) { |
| 574 | for (size_t k = 1; k <= 5; k += 2) { |
| 575 | SpMMMicrokernelTester() |
| 576 | .mr(4) |
| 577 | .nr(4) |
| 578 | .m(8) |
| 579 | .n(n) |
| 580 | .k(k) |
| 581 | .sparsity(0.5f) |
| 582 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 583 | } |
| 584 | } |
| 585 | } |
| 586 | |
| 587 | TEST(F32_SPMM_4X4__NEONFMA, zero_weights) { |
| 588 | TEST_REQUIRES_ARM_NEON_FMA; |
| 589 | for (uint32_t n = 1; n < 20; n += 5) { |
| 590 | for (size_t k = 1; k <= 5; k += 2) { |
| 591 | SpMMMicrokernelTester() |
| 592 | .mr(4) |
| 593 | .nr(4) |
| 594 | .m(8) |
| 595 | .n(n) |
| 596 | .k(k) |
| 597 | .sparsity(1.0f) |
| 598 | .Test(xnn_f32_spmm_ukernel_4x4__neonfma); |
| 599 | } |
| 600 | } |
| 601 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 602 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 603 | |
| 604 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 605 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 606 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, k_eq_1) { |
| 607 | TEST_REQUIRES_ARM_NEON_FMA; |
| 608 | SpMMMicrokernelTester() |
| 609 | .mr(4) |
| 610 | .nr(1) |
| 611 | .m(4) |
| 612 | .n(1) |
| 613 | .k(1) |
| 614 | .sparsity(0.0f) |
| 615 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 616 | } |
| 617 | |
| 618 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, k_gt_1) { |
| 619 | TEST_REQUIRES_ARM_NEON_FMA; |
| 620 | for (size_t k = 2; k < 10; k++) { |
| 621 | SpMMMicrokernelTester() |
| 622 | .mr(4) |
| 623 | .nr(1) |
| 624 | .m(4) |
| 625 | .n(1) |
| 626 | .k(k) |
| 627 | .sparsity(0.0f) |
| 628 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 629 | } |
| 630 | } |
| 631 | |
| 632 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, n_gt_1) { |
| 633 | TEST_REQUIRES_ARM_NEON_FMA; |
| 634 | for (uint32_t n = 2; n < 10; n++) { |
| 635 | for (size_t k = 1; k <= 5; k += 2) { |
| 636 | SpMMMicrokernelTester() |
| 637 | .mr(4) |
| 638 | .nr(1) |
| 639 | .m(4) |
| 640 | .n(n) |
| 641 | .k(k) |
| 642 | .sparsity(0.0f) |
| 643 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 644 | } |
| 645 | } |
| 646 | } |
| 647 | |
| 648 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, m_lt_4) { |
| 649 | TEST_REQUIRES_ARM_NEON_FMA; |
| 650 | for (uint32_t m = 1; m < 4; m++) { |
| 651 | for (uint32_t n = 1; n < 10; n += 2) { |
| 652 | for (size_t k = 1; k <= 5; k += 2) { |
| 653 | SpMMMicrokernelTester() |
| 654 | .mr(4) |
| 655 | .nr(1) |
| 656 | .m(m) |
| 657 | .n(n) |
| 658 | .k(k) |
| 659 | .sparsity(0.0f) |
| 660 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 661 | } |
| 662 | } |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, m_div_4) { |
| 667 | TEST_REQUIRES_ARM_NEON_FMA; |
| 668 | for (uint32_t m = 8; m <= 12; m += 4) { |
| 669 | for (uint32_t n = 1; n < 10; n += 2) { |
| 670 | for (size_t k = 1; k <= 5; k += 2) { |
| 671 | SpMMMicrokernelTester() |
| 672 | .mr(4) |
| 673 | .nr(1) |
| 674 | .m(m) |
| 675 | .n(n) |
| 676 | .k(k) |
| 677 | .sparsity(0.0f) |
| 678 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 679 | } |
| 680 | } |
| 681 | } |
| 682 | } |
| 683 | |
| 684 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, m_gt_4) { |
| 685 | TEST_REQUIRES_ARM_NEON_FMA; |
| 686 | for (uint32_t m = 5; m < 8; m++) { |
| 687 | for (uint32_t n = 1; n < 10; n += 2) { |
| 688 | for (size_t k = 1; k <= 5; k += 2) { |
| 689 | SpMMMicrokernelTester() |
| 690 | .mr(4) |
| 691 | .nr(1) |
| 692 | .m(m) |
| 693 | .n(n) |
| 694 | .k(k) |
| 695 | .sparsity(0.0f) |
| 696 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 697 | } |
| 698 | } |
| 699 | } |
| 700 | } |
| 701 | |
| 702 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, qmin) { |
| 703 | TEST_REQUIRES_ARM_NEON_FMA; |
| 704 | for (uint32_t n = 1; n < 10; n += 2) { |
| 705 | for (size_t k = 1; k <= 5; k += 2) { |
| 706 | SpMMMicrokernelTester() |
| 707 | .mr(4) |
| 708 | .nr(1) |
| 709 | .m(8) |
| 710 | .n(n) |
| 711 | .k(k) |
| 712 | .sparsity(0.0f) |
| 713 | .qmin(128) |
| 714 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 715 | } |
| 716 | } |
| 717 | } |
| 718 | |
| 719 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, qmax) { |
| 720 | TEST_REQUIRES_ARM_NEON_FMA; |
| 721 | for (uint32_t n = 1; n < 10; n += 2) { |
| 722 | for (size_t k = 1; k <= 5; k += 2) { |
| 723 | SpMMMicrokernelTester() |
| 724 | .mr(4) |
| 725 | .nr(1) |
| 726 | .m(8) |
| 727 | .n(n) |
| 728 | .k(k) |
| 729 | .sparsity(0.0f) |
| 730 | .qmax(128) |
| 731 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 732 | } |
| 733 | } |
| 734 | } |
| 735 | |
| 736 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, half_sparse) { |
| 737 | TEST_REQUIRES_ARM_NEON_FMA; |
| 738 | for (uint32_t n = 1; n < 10; n += 2) { |
| 739 | for (size_t k = 1; k <= 5; k += 2) { |
| 740 | SpMMMicrokernelTester() |
| 741 | .mr(4) |
| 742 | .nr(1) |
| 743 | .m(8) |
| 744 | .n(n) |
| 745 | .k(k) |
| 746 | .sparsity(0.5f) |
| 747 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 748 | } |
| 749 | } |
| 750 | } |
| 751 | |
| 752 | TEST(F32_SPMM_4X1__NEONFMA_PIPELINED, zero_weights) { |
| 753 | TEST_REQUIRES_ARM_NEON_FMA; |
| 754 | for (uint32_t n = 1; n < 10; n += 2) { |
| 755 | for (size_t k = 1; k <= 5; k += 2) { |
| 756 | SpMMMicrokernelTester() |
| 757 | .mr(4) |
| 758 | .nr(1) |
| 759 | .m(8) |
| 760 | .n(n) |
| 761 | .k(k) |
| 762 | .sparsity(1.0f) |
| 763 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_pipelined); |
| 764 | } |
| 765 | } |
| 766 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 767 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 768 | |
| 769 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 770 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 771 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, k_eq_2) { |
| 772 | TEST_REQUIRES_ARM_NEON_FMA; |
| 773 | SpMMMicrokernelTester() |
| 774 | .mr(4) |
| 775 | .nr(1) |
| 776 | .m(4) |
| 777 | .n(1) |
| 778 | .k(2) |
| 779 | .sparsity(0.0f) |
| 780 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 781 | } |
| 782 | |
| 783 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, k_lt_2) { |
| 784 | TEST_REQUIRES_ARM_NEON_FMA; |
| 785 | for (size_t k = 1; k < 2; k++) { |
| 786 | SpMMMicrokernelTester() |
| 787 | .mr(4) |
| 788 | .nr(1) |
| 789 | .m(4) |
| 790 | .n(1) |
| 791 | .k(k) |
| 792 | .sparsity(0.0f) |
| 793 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 794 | } |
| 795 | } |
| 796 | |
| 797 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, k_gt_2) { |
| 798 | TEST_REQUIRES_ARM_NEON_FMA; |
| 799 | for (size_t k = 3; k < 4; k++) { |
| 800 | SpMMMicrokernelTester() |
| 801 | .mr(4) |
| 802 | .nr(1) |
| 803 | .m(4) |
| 804 | .n(1) |
| 805 | .k(k) |
| 806 | .sparsity(0.0f) |
| 807 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 808 | } |
| 809 | } |
| 810 | |
| 811 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, k_div_2) { |
| 812 | TEST_REQUIRES_ARM_NEON_FMA; |
| 813 | for (size_t k = 4; k <= 20; k += 2) { |
| 814 | SpMMMicrokernelTester() |
| 815 | .mr(4) |
| 816 | .nr(1) |
| 817 | .m(4) |
| 818 | .n(1) |
| 819 | .k(k) |
| 820 | .sparsity(0.0f) |
| 821 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 822 | } |
| 823 | } |
| 824 | |
| 825 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, n_gt_1) { |
| 826 | TEST_REQUIRES_ARM_NEON_FMA; |
| 827 | for (uint32_t n = 2; n < 10; n++) { |
| 828 | for (size_t k = 1; k <= 10; k += 3) { |
| 829 | SpMMMicrokernelTester() |
| 830 | .mr(4) |
| 831 | .nr(1) |
| 832 | .m(4) |
| 833 | .n(n) |
| 834 | .k(k) |
| 835 | .sparsity(0.0f) |
| 836 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 837 | } |
| 838 | } |
| 839 | } |
| 840 | |
| 841 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, m_lt_4) { |
| 842 | TEST_REQUIRES_ARM_NEON_FMA; |
| 843 | for (uint32_t m = 1; m < 4; m++) { |
| 844 | for (uint32_t n = 1; n < 10; n += 2) { |
| 845 | for (size_t k = 1; k <= 10; k += 3) { |
| 846 | SpMMMicrokernelTester() |
| 847 | .mr(4) |
| 848 | .nr(1) |
| 849 | .m(m) |
| 850 | .n(n) |
| 851 | .k(k) |
| 852 | .sparsity(0.0f) |
| 853 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 854 | } |
| 855 | } |
| 856 | } |
| 857 | } |
| 858 | |
| 859 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, m_div_4) { |
| 860 | TEST_REQUIRES_ARM_NEON_FMA; |
| 861 | for (uint32_t m = 8; m <= 12; m += 4) { |
| 862 | for (uint32_t n = 1; n < 10; n += 2) { |
| 863 | for (size_t k = 1; k <= 10; k += 3) { |
| 864 | SpMMMicrokernelTester() |
| 865 | .mr(4) |
| 866 | .nr(1) |
| 867 | .m(m) |
| 868 | .n(n) |
| 869 | .k(k) |
| 870 | .sparsity(0.0f) |
| 871 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 872 | } |
| 873 | } |
| 874 | } |
| 875 | } |
| 876 | |
| 877 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, m_gt_4) { |
| 878 | TEST_REQUIRES_ARM_NEON_FMA; |
| 879 | for (uint32_t m = 5; m < 8; m++) { |
| 880 | for (uint32_t n = 1; n < 10; n += 2) { |
| 881 | for (size_t k = 1; k <= 10; k += 3) { |
| 882 | SpMMMicrokernelTester() |
| 883 | .mr(4) |
| 884 | .nr(1) |
| 885 | .m(m) |
| 886 | .n(n) |
| 887 | .k(k) |
| 888 | .sparsity(0.0f) |
| 889 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 890 | } |
| 891 | } |
| 892 | } |
| 893 | } |
| 894 | |
| 895 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, qmin) { |
| 896 | TEST_REQUIRES_ARM_NEON_FMA; |
| 897 | for (uint32_t n = 1; n < 10; n += 2) { |
| 898 | for (size_t k = 1; k <= 10; k += 3) { |
| 899 | SpMMMicrokernelTester() |
| 900 | .mr(4) |
| 901 | .nr(1) |
| 902 | .m(8) |
| 903 | .n(n) |
| 904 | .k(k) |
| 905 | .sparsity(0.0f) |
| 906 | .qmin(128) |
| 907 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 908 | } |
| 909 | } |
| 910 | } |
| 911 | |
| 912 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, qmax) { |
| 913 | TEST_REQUIRES_ARM_NEON_FMA; |
| 914 | for (uint32_t n = 1; n < 10; n += 2) { |
| 915 | for (size_t k = 1; k <= 10; k += 3) { |
| 916 | SpMMMicrokernelTester() |
| 917 | .mr(4) |
| 918 | .nr(1) |
| 919 | .m(8) |
| 920 | .n(n) |
| 921 | .k(k) |
| 922 | .sparsity(0.0f) |
| 923 | .qmax(128) |
| 924 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 925 | } |
| 926 | } |
| 927 | } |
| 928 | |
| 929 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, half_sparse) { |
| 930 | TEST_REQUIRES_ARM_NEON_FMA; |
| 931 | for (uint32_t n = 1; n < 10; n += 2) { |
| 932 | for (size_t k = 1; k <= 10; k += 3) { |
| 933 | SpMMMicrokernelTester() |
| 934 | .mr(4) |
| 935 | .nr(1) |
| 936 | .m(8) |
| 937 | .n(n) |
| 938 | .k(k) |
| 939 | .sparsity(0.5f) |
| 940 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 941 | } |
| 942 | } |
| 943 | } |
| 944 | |
| 945 | TEST(F32_SPMM_4X1__NEONFMA_UNROLL2, zero_weights) { |
| 946 | TEST_REQUIRES_ARM_NEON_FMA; |
| 947 | for (uint32_t n = 1; n < 10; n += 2) { |
| 948 | for (size_t k = 1; k <= 10; k += 3) { |
| 949 | SpMMMicrokernelTester() |
| 950 | .mr(4) |
| 951 | .nr(1) |
| 952 | .m(8) |
| 953 | .n(n) |
| 954 | .k(k) |
| 955 | .sparsity(1.0f) |
| 956 | .Test(xnn_f32_spmm_ukernel_4x1__neonfma_unroll2); |
| 957 | } |
| 958 | } |
| 959 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 960 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 961 | |
| 962 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 963 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 964 | TEST(F32_SPMM_8X1__NEONFMA, k_eq_1) { |
| 965 | TEST_REQUIRES_ARM_NEON_FMA; |
| 966 | SpMMMicrokernelTester() |
| 967 | .mr(8) |
| 968 | .nr(1) |
| 969 | .m(8) |
| 970 | .n(1) |
| 971 | .k(1) |
| 972 | .sparsity(0.0f) |
| 973 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 974 | } |
| 975 | |
| 976 | TEST(F32_SPMM_8X1__NEONFMA, k_gt_1) { |
| 977 | TEST_REQUIRES_ARM_NEON_FMA; |
| 978 | for (size_t k = 2; k < 10; k++) { |
| 979 | SpMMMicrokernelTester() |
| 980 | .mr(8) |
| 981 | .nr(1) |
| 982 | .m(8) |
| 983 | .n(1) |
| 984 | .k(k) |
| 985 | .sparsity(0.0f) |
| 986 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 987 | } |
| 988 | } |
| 989 | |
| 990 | TEST(F32_SPMM_8X1__NEONFMA, n_gt_1) { |
| 991 | TEST_REQUIRES_ARM_NEON_FMA; |
| 992 | for (uint32_t n = 2; n < 10; n++) { |
| 993 | for (size_t k = 1; k <= 5; k += 2) { |
| 994 | SpMMMicrokernelTester() |
| 995 | .mr(8) |
| 996 | .nr(1) |
| 997 | .m(8) |
| 998 | .n(n) |
| 999 | .k(k) |
| 1000 | .sparsity(0.0f) |
| 1001 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 1002 | } |
| 1003 | } |
| 1004 | } |
| 1005 | |
| 1006 | TEST(F32_SPMM_8X1__NEONFMA, m_lt_8) { |
| 1007 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1008 | for (uint32_t m = 1; m < 8; m++) { |
| 1009 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1010 | for (size_t k = 1; k <= 5; k += 2) { |
| 1011 | SpMMMicrokernelTester() |
| 1012 | .mr(8) |
| 1013 | .nr(1) |
| 1014 | .m(m) |
| 1015 | .n(n) |
| 1016 | .k(k) |
| 1017 | .sparsity(0.0f) |
| 1018 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 1019 | } |
| 1020 | } |
| 1021 | } |
| 1022 | } |
| 1023 | |
| 1024 | TEST(F32_SPMM_8X1__NEONFMA, m_div_8) { |
| 1025 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1026 | for (uint32_t m = 16; m <= 24; m += 8) { |
| 1027 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1028 | for (size_t k = 1; k <= 5; k += 2) { |
| 1029 | SpMMMicrokernelTester() |
| 1030 | .mr(8) |
| 1031 | .nr(1) |
| 1032 | .m(m) |
| 1033 | .n(n) |
| 1034 | .k(k) |
| 1035 | .sparsity(0.0f) |
| 1036 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 1037 | } |
| 1038 | } |
| 1039 | } |
| 1040 | } |
| 1041 | |
| 1042 | TEST(F32_SPMM_8X1__NEONFMA, m_gt_8) { |
| 1043 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1044 | for (uint32_t m = 9; m < 16; m++) { |
| 1045 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1046 | for (size_t k = 1; k <= 5; k += 2) { |
| 1047 | SpMMMicrokernelTester() |
| 1048 | .mr(8) |
| 1049 | .nr(1) |
| 1050 | .m(m) |
| 1051 | .n(n) |
| 1052 | .k(k) |
| 1053 | .sparsity(0.0f) |
| 1054 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 1055 | } |
| 1056 | } |
| 1057 | } |
| 1058 | } |
| 1059 | |
| 1060 | TEST(F32_SPMM_8X1__NEONFMA, qmin) { |
| 1061 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1062 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1063 | for (size_t k = 1; k <= 5; k += 2) { |
| 1064 | SpMMMicrokernelTester() |
| 1065 | .mr(8) |
| 1066 | .nr(1) |
| 1067 | .m(16) |
| 1068 | .n(n) |
| 1069 | .k(k) |
| 1070 | .sparsity(0.0f) |
| 1071 | .qmin(128) |
| 1072 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 1073 | } |
| 1074 | } |
| 1075 | } |
| 1076 | |
| 1077 | TEST(F32_SPMM_8X1__NEONFMA, qmax) { |
| 1078 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1079 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1080 | for (size_t k = 1; k <= 5; k += 2) { |
| 1081 | SpMMMicrokernelTester() |
| 1082 | .mr(8) |
| 1083 | .nr(1) |
| 1084 | .m(16) |
| 1085 | .n(n) |
| 1086 | .k(k) |
| 1087 | .sparsity(0.0f) |
| 1088 | .qmax(128) |
| 1089 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 1090 | } |
| 1091 | } |
| 1092 | } |
| 1093 | |
| 1094 | TEST(F32_SPMM_8X1__NEONFMA, half_sparse) { |
| 1095 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1096 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1097 | for (size_t k = 1; k <= 5; k += 2) { |
| 1098 | SpMMMicrokernelTester() |
| 1099 | .mr(8) |
| 1100 | .nr(1) |
| 1101 | .m(16) |
| 1102 | .n(n) |
| 1103 | .k(k) |
| 1104 | .sparsity(0.5f) |
| 1105 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 1106 | } |
| 1107 | } |
| 1108 | } |
| 1109 | |
| 1110 | TEST(F32_SPMM_8X1__NEONFMA, zero_weights) { |
| 1111 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1112 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1113 | for (size_t k = 1; k <= 5; k += 2) { |
| 1114 | SpMMMicrokernelTester() |
| 1115 | .mr(8) |
| 1116 | .nr(1) |
| 1117 | .m(16) |
| 1118 | .n(n) |
| 1119 | .k(k) |
| 1120 | .sparsity(1.0f) |
| 1121 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma); |
| 1122 | } |
| 1123 | } |
| 1124 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1125 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1126 | |
| 1127 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1128 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1129 | TEST(F32_SPMM_8X2__NEONFMA, k_eq_1) { |
| 1130 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1131 | SpMMMicrokernelTester() |
| 1132 | .mr(8) |
| 1133 | .nr(2) |
| 1134 | .m(8) |
| 1135 | .n(2) |
| 1136 | .k(1) |
| 1137 | .sparsity(0.0f) |
| 1138 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1139 | } |
| 1140 | |
| 1141 | TEST(F32_SPMM_8X2__NEONFMA, k_eq_1_subtile) { |
| 1142 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1143 | for (uint32_t n = 1; n <= 2; n++) { |
| 1144 | SpMMMicrokernelTester() |
| 1145 | .mr(8) |
| 1146 | .nr(2) |
| 1147 | .m(8) |
| 1148 | .n(n) |
| 1149 | .k(1) |
| 1150 | .sparsity(0.0f) |
| 1151 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1152 | } |
| 1153 | } |
| 1154 | |
| 1155 | TEST(F32_SPMM_8X2__NEONFMA, k_gt_1) { |
| 1156 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1157 | for (size_t k = 2; k < 10; k++) { |
| 1158 | SpMMMicrokernelTester() |
| 1159 | .mr(8) |
| 1160 | .nr(2) |
| 1161 | .m(8) |
| 1162 | .n(2) |
| 1163 | .k(k) |
| 1164 | .sparsity(0.0f) |
| 1165 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1166 | } |
| 1167 | } |
| 1168 | |
| 1169 | TEST(F32_SPMM_8X2__NEONFMA, k_gt_1_subtile) { |
| 1170 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1171 | for (size_t k = 2; k < 10; k++) { |
| 1172 | for (uint32_t n = 1; n <= 2; n++) { |
| 1173 | SpMMMicrokernelTester() |
| 1174 | .mr(8) |
| 1175 | .nr(2) |
| 1176 | .m(8) |
| 1177 | .n(n) |
| 1178 | .k(k) |
| 1179 | .sparsity(0.0f) |
| 1180 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1181 | } |
| 1182 | } |
| 1183 | } |
| 1184 | |
| 1185 | TEST(F32_SPMM_8X2__NEONFMA, n_gt_2) { |
| 1186 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1187 | for (uint32_t n = 3; n < 10; n++) { |
| 1188 | for (size_t k = 1; k <= 5; k += 2) { |
| 1189 | SpMMMicrokernelTester() |
| 1190 | .mr(8) |
| 1191 | .nr(2) |
| 1192 | .m(8) |
| 1193 | .n(n) |
| 1194 | .k(k) |
| 1195 | .sparsity(0.0f) |
| 1196 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1197 | } |
| 1198 | } |
| 1199 | } |
| 1200 | |
| 1201 | TEST(F32_SPMM_8X2__NEONFMA, n_div_2) { |
| 1202 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1203 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 1204 | for (size_t k = 1; k <= 5; k += 2) { |
| 1205 | SpMMMicrokernelTester() |
| 1206 | .mr(8) |
| 1207 | .nr(2) |
| 1208 | .m(8) |
| 1209 | .n(n) |
| 1210 | .k(k) |
| 1211 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1212 | } |
| 1213 | } |
| 1214 | } |
| 1215 | |
| 1216 | TEST(F32_SPMM_8X2__NEONFMA, m_lt_8) { |
| 1217 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1218 | for (uint32_t m = 1; m < 8; m++) { |
| 1219 | for (uint32_t n = 1; n < 10; n += 3) { |
| 1220 | for (size_t k = 1; k <= 5; k += 2) { |
| 1221 | SpMMMicrokernelTester() |
| 1222 | .mr(8) |
| 1223 | .nr(2) |
| 1224 | .m(m) |
| 1225 | .n(n) |
| 1226 | .k(k) |
| 1227 | .sparsity(0.0f) |
| 1228 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1229 | } |
| 1230 | } |
| 1231 | } |
| 1232 | } |
| 1233 | |
| 1234 | TEST(F32_SPMM_8X2__NEONFMA, m_div_8) { |
| 1235 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1236 | for (uint32_t m = 16; m <= 24; m += 8) { |
| 1237 | for (uint32_t n = 1; n < 10; n += 3) { |
| 1238 | for (size_t k = 1; k <= 5; k += 2) { |
| 1239 | SpMMMicrokernelTester() |
| 1240 | .mr(8) |
| 1241 | .nr(2) |
| 1242 | .m(m) |
| 1243 | .n(n) |
| 1244 | .k(k) |
| 1245 | .sparsity(0.0f) |
| 1246 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1247 | } |
| 1248 | } |
| 1249 | } |
| 1250 | } |
| 1251 | |
| 1252 | TEST(F32_SPMM_8X2__NEONFMA, m_gt_8) { |
| 1253 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1254 | for (uint32_t m = 9; m < 16; m++) { |
| 1255 | for (uint32_t n = 1; n < 10; n += 3) { |
| 1256 | for (size_t k = 1; k <= 5; k += 2) { |
| 1257 | SpMMMicrokernelTester() |
| 1258 | .mr(8) |
| 1259 | .nr(2) |
| 1260 | .m(m) |
| 1261 | .n(n) |
| 1262 | .k(k) |
| 1263 | .sparsity(0.0f) |
| 1264 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1265 | } |
| 1266 | } |
| 1267 | } |
| 1268 | } |
| 1269 | |
| 1270 | TEST(F32_SPMM_8X2__NEONFMA, qmin) { |
| 1271 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1272 | for (uint32_t n = 1; n < 10; n += 3) { |
| 1273 | for (size_t k = 1; k <= 5; k += 2) { |
| 1274 | SpMMMicrokernelTester() |
| 1275 | .mr(8) |
| 1276 | .nr(2) |
| 1277 | .m(16) |
| 1278 | .n(n) |
| 1279 | .k(k) |
| 1280 | .sparsity(0.0f) |
| 1281 | .qmin(128) |
| 1282 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1283 | } |
| 1284 | } |
| 1285 | } |
| 1286 | |
| 1287 | TEST(F32_SPMM_8X2__NEONFMA, qmax) { |
| 1288 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1289 | for (uint32_t n = 1; n < 10; n += 3) { |
| 1290 | for (size_t k = 1; k <= 5; k += 2) { |
| 1291 | SpMMMicrokernelTester() |
| 1292 | .mr(8) |
| 1293 | .nr(2) |
| 1294 | .m(16) |
| 1295 | .n(n) |
| 1296 | .k(k) |
| 1297 | .sparsity(0.0f) |
| 1298 | .qmax(128) |
| 1299 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1300 | } |
| 1301 | } |
| 1302 | } |
| 1303 | |
| 1304 | TEST(F32_SPMM_8X2__NEONFMA, half_sparse) { |
| 1305 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1306 | for (uint32_t n = 1; n < 10; n += 3) { |
| 1307 | for (size_t k = 1; k <= 5; k += 2) { |
| 1308 | SpMMMicrokernelTester() |
| 1309 | .mr(8) |
| 1310 | .nr(2) |
| 1311 | .m(16) |
| 1312 | .n(n) |
| 1313 | .k(k) |
| 1314 | .sparsity(0.5f) |
| 1315 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1316 | } |
| 1317 | } |
| 1318 | } |
| 1319 | |
| 1320 | TEST(F32_SPMM_8X2__NEONFMA, zero_weights) { |
| 1321 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1322 | for (uint32_t n = 1; n < 10; n += 3) { |
| 1323 | for (size_t k = 1; k <= 5; k += 2) { |
| 1324 | SpMMMicrokernelTester() |
| 1325 | .mr(8) |
| 1326 | .nr(2) |
| 1327 | .m(16) |
| 1328 | .n(n) |
| 1329 | .k(k) |
| 1330 | .sparsity(1.0f) |
| 1331 | .Test(xnn_f32_spmm_ukernel_8x2__neonfma); |
| 1332 | } |
| 1333 | } |
| 1334 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1335 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1336 | |
| 1337 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1338 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1339 | TEST(F32_SPMM_8X4__NEONFMA, k_eq_1) { |
| 1340 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1341 | SpMMMicrokernelTester() |
| 1342 | .mr(8) |
| 1343 | .nr(4) |
| 1344 | .m(8) |
| 1345 | .n(4) |
| 1346 | .k(1) |
| 1347 | .sparsity(0.0f) |
| 1348 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1349 | } |
| 1350 | |
| 1351 | TEST(F32_SPMM_8X4__NEONFMA, k_eq_1_subtile) { |
| 1352 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1353 | for (uint32_t n = 1; n <= 4; n++) { |
| 1354 | SpMMMicrokernelTester() |
| 1355 | .mr(8) |
| 1356 | .nr(4) |
| 1357 | .m(8) |
| 1358 | .n(n) |
| 1359 | .k(1) |
| 1360 | .sparsity(0.0f) |
| 1361 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1362 | } |
| 1363 | } |
| 1364 | |
| 1365 | TEST(F32_SPMM_8X4__NEONFMA, k_gt_1) { |
| 1366 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1367 | for (size_t k = 2; k < 10; k++) { |
| 1368 | SpMMMicrokernelTester() |
| 1369 | .mr(8) |
| 1370 | .nr(4) |
| 1371 | .m(8) |
| 1372 | .n(4) |
| 1373 | .k(k) |
| 1374 | .sparsity(0.0f) |
| 1375 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1376 | } |
| 1377 | } |
| 1378 | |
| 1379 | TEST(F32_SPMM_8X4__NEONFMA, k_gt_1_subtile) { |
| 1380 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1381 | for (size_t k = 2; k < 10; k++) { |
| 1382 | for (uint32_t n = 1; n <= 4; n++) { |
| 1383 | SpMMMicrokernelTester() |
| 1384 | .mr(8) |
| 1385 | .nr(4) |
| 1386 | .m(8) |
| 1387 | .n(n) |
| 1388 | .k(k) |
| 1389 | .sparsity(0.0f) |
| 1390 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1391 | } |
| 1392 | } |
| 1393 | } |
| 1394 | |
| 1395 | TEST(F32_SPMM_8X4__NEONFMA, n_gt_4) { |
| 1396 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1397 | for (uint32_t n = 5; n < 10; n++) { |
| 1398 | for (size_t k = 1; k <= 5; k += 2) { |
| 1399 | SpMMMicrokernelTester() |
| 1400 | .mr(8) |
| 1401 | .nr(4) |
| 1402 | .m(8) |
| 1403 | .n(n) |
| 1404 | .k(k) |
| 1405 | .sparsity(0.0f) |
| 1406 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1407 | } |
| 1408 | } |
| 1409 | } |
| 1410 | |
| 1411 | TEST(F32_SPMM_8X4__NEONFMA, n_div_4) { |
| 1412 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1413 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 1414 | for (size_t k = 1; k <= 5; k += 2) { |
| 1415 | SpMMMicrokernelTester() |
| 1416 | .mr(8) |
| 1417 | .nr(4) |
| 1418 | .m(8) |
| 1419 | .n(n) |
| 1420 | .k(k) |
| 1421 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1422 | } |
| 1423 | } |
| 1424 | } |
| 1425 | |
| 1426 | TEST(F32_SPMM_8X4__NEONFMA, m_lt_8) { |
| 1427 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1428 | for (uint32_t m = 1; m < 8; m++) { |
| 1429 | for (uint32_t n = 1; n < 20; n += 5) { |
| 1430 | for (size_t k = 1; k <= 5; k += 2) { |
| 1431 | SpMMMicrokernelTester() |
| 1432 | .mr(8) |
| 1433 | .nr(4) |
| 1434 | .m(m) |
| 1435 | .n(n) |
| 1436 | .k(k) |
| 1437 | .sparsity(0.0f) |
| 1438 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1439 | } |
| 1440 | } |
| 1441 | } |
| 1442 | } |
| 1443 | |
| 1444 | TEST(F32_SPMM_8X4__NEONFMA, m_div_8) { |
| 1445 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1446 | for (uint32_t m = 16; m <= 24; m += 8) { |
| 1447 | for (uint32_t n = 1; n < 20; n += 5) { |
| 1448 | for (size_t k = 1; k <= 5; k += 2) { |
| 1449 | SpMMMicrokernelTester() |
| 1450 | .mr(8) |
| 1451 | .nr(4) |
| 1452 | .m(m) |
| 1453 | .n(n) |
| 1454 | .k(k) |
| 1455 | .sparsity(0.0f) |
| 1456 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1457 | } |
| 1458 | } |
| 1459 | } |
| 1460 | } |
| 1461 | |
| 1462 | TEST(F32_SPMM_8X4__NEONFMA, m_gt_8) { |
| 1463 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1464 | for (uint32_t m = 9; m < 16; m++) { |
| 1465 | for (uint32_t n = 1; n < 20; n += 5) { |
| 1466 | for (size_t k = 1; k <= 5; k += 2) { |
| 1467 | SpMMMicrokernelTester() |
| 1468 | .mr(8) |
| 1469 | .nr(4) |
| 1470 | .m(m) |
| 1471 | .n(n) |
| 1472 | .k(k) |
| 1473 | .sparsity(0.0f) |
| 1474 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1475 | } |
| 1476 | } |
| 1477 | } |
| 1478 | } |
| 1479 | |
| 1480 | TEST(F32_SPMM_8X4__NEONFMA, qmin) { |
| 1481 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1482 | for (uint32_t n = 1; n < 20; n += 5) { |
| 1483 | for (size_t k = 1; k <= 5; k += 2) { |
| 1484 | SpMMMicrokernelTester() |
| 1485 | .mr(8) |
| 1486 | .nr(4) |
| 1487 | .m(16) |
| 1488 | .n(n) |
| 1489 | .k(k) |
| 1490 | .sparsity(0.0f) |
| 1491 | .qmin(128) |
| 1492 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1493 | } |
| 1494 | } |
| 1495 | } |
| 1496 | |
| 1497 | TEST(F32_SPMM_8X4__NEONFMA, qmax) { |
| 1498 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1499 | for (uint32_t n = 1; n < 20; n += 5) { |
| 1500 | for (size_t k = 1; k <= 5; k += 2) { |
| 1501 | SpMMMicrokernelTester() |
| 1502 | .mr(8) |
| 1503 | .nr(4) |
| 1504 | .m(16) |
| 1505 | .n(n) |
| 1506 | .k(k) |
| 1507 | .sparsity(0.0f) |
| 1508 | .qmax(128) |
| 1509 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1510 | } |
| 1511 | } |
| 1512 | } |
| 1513 | |
| 1514 | TEST(F32_SPMM_8X4__NEONFMA, half_sparse) { |
| 1515 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1516 | for (uint32_t n = 1; n < 20; n += 5) { |
| 1517 | for (size_t k = 1; k <= 5; k += 2) { |
| 1518 | SpMMMicrokernelTester() |
| 1519 | .mr(8) |
| 1520 | .nr(4) |
| 1521 | .m(16) |
| 1522 | .n(n) |
| 1523 | .k(k) |
| 1524 | .sparsity(0.5f) |
| 1525 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1526 | } |
| 1527 | } |
| 1528 | } |
| 1529 | |
| 1530 | TEST(F32_SPMM_8X4__NEONFMA, zero_weights) { |
| 1531 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1532 | for (uint32_t n = 1; n < 20; n += 5) { |
| 1533 | for (size_t k = 1; k <= 5; k += 2) { |
| 1534 | SpMMMicrokernelTester() |
| 1535 | .mr(8) |
| 1536 | .nr(4) |
| 1537 | .m(16) |
| 1538 | .n(n) |
| 1539 | .k(k) |
| 1540 | .sparsity(1.0f) |
| 1541 | .Test(xnn_f32_spmm_ukernel_8x4__neonfma); |
| 1542 | } |
| 1543 | } |
| 1544 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1545 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1546 | |
| 1547 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1548 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1549 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, k_eq_1) { |
| 1550 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1551 | SpMMMicrokernelTester() |
| 1552 | .mr(8) |
| 1553 | .nr(1) |
| 1554 | .m(8) |
| 1555 | .n(1) |
| 1556 | .k(1) |
| 1557 | .sparsity(0.0f) |
| 1558 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1559 | } |
| 1560 | |
| 1561 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, k_gt_1) { |
| 1562 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1563 | for (size_t k = 2; k < 10; k++) { |
| 1564 | SpMMMicrokernelTester() |
| 1565 | .mr(8) |
| 1566 | .nr(1) |
| 1567 | .m(8) |
| 1568 | .n(1) |
| 1569 | .k(k) |
| 1570 | .sparsity(0.0f) |
| 1571 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1572 | } |
| 1573 | } |
| 1574 | |
| 1575 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, n_gt_1) { |
| 1576 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1577 | for (uint32_t n = 2; n < 10; n++) { |
| 1578 | for (size_t k = 1; k <= 5; k += 2) { |
| 1579 | SpMMMicrokernelTester() |
| 1580 | .mr(8) |
| 1581 | .nr(1) |
| 1582 | .m(8) |
| 1583 | .n(n) |
| 1584 | .k(k) |
| 1585 | .sparsity(0.0f) |
| 1586 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1587 | } |
| 1588 | } |
| 1589 | } |
| 1590 | |
| 1591 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, m_lt_8) { |
| 1592 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1593 | for (uint32_t m = 1; m < 8; m++) { |
| 1594 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1595 | for (size_t k = 1; k <= 5; k += 2) { |
| 1596 | SpMMMicrokernelTester() |
| 1597 | .mr(8) |
| 1598 | .nr(1) |
| 1599 | .m(m) |
| 1600 | .n(n) |
| 1601 | .k(k) |
| 1602 | .sparsity(0.0f) |
| 1603 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1604 | } |
| 1605 | } |
| 1606 | } |
| 1607 | } |
| 1608 | |
| 1609 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, m_div_8) { |
| 1610 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1611 | for (uint32_t m = 16; m <= 24; m += 8) { |
| 1612 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1613 | for (size_t k = 1; k <= 5; k += 2) { |
| 1614 | SpMMMicrokernelTester() |
| 1615 | .mr(8) |
| 1616 | .nr(1) |
| 1617 | .m(m) |
| 1618 | .n(n) |
| 1619 | .k(k) |
| 1620 | .sparsity(0.0f) |
| 1621 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1622 | } |
| 1623 | } |
| 1624 | } |
| 1625 | } |
| 1626 | |
| 1627 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, m_gt_8) { |
| 1628 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1629 | for (uint32_t m = 9; m < 16; m++) { |
| 1630 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1631 | for (size_t k = 1; k <= 5; k += 2) { |
| 1632 | SpMMMicrokernelTester() |
| 1633 | .mr(8) |
| 1634 | .nr(1) |
| 1635 | .m(m) |
| 1636 | .n(n) |
| 1637 | .k(k) |
| 1638 | .sparsity(0.0f) |
| 1639 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1640 | } |
| 1641 | } |
| 1642 | } |
| 1643 | } |
| 1644 | |
| 1645 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, qmin) { |
| 1646 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1647 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1648 | for (size_t k = 1; k <= 5; k += 2) { |
| 1649 | SpMMMicrokernelTester() |
| 1650 | .mr(8) |
| 1651 | .nr(1) |
| 1652 | .m(16) |
| 1653 | .n(n) |
| 1654 | .k(k) |
| 1655 | .sparsity(0.0f) |
| 1656 | .qmin(128) |
| 1657 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1658 | } |
| 1659 | } |
| 1660 | } |
| 1661 | |
| 1662 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, qmax) { |
| 1663 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1664 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1665 | for (size_t k = 1; k <= 5; k += 2) { |
| 1666 | SpMMMicrokernelTester() |
| 1667 | .mr(8) |
| 1668 | .nr(1) |
| 1669 | .m(16) |
| 1670 | .n(n) |
| 1671 | .k(k) |
| 1672 | .sparsity(0.0f) |
| 1673 | .qmax(128) |
| 1674 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1675 | } |
| 1676 | } |
| 1677 | } |
| 1678 | |
| 1679 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, half_sparse) { |
| 1680 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1681 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1682 | for (size_t k = 1; k <= 5; k += 2) { |
| 1683 | SpMMMicrokernelTester() |
| 1684 | .mr(8) |
| 1685 | .nr(1) |
| 1686 | .m(16) |
| 1687 | .n(n) |
| 1688 | .k(k) |
| 1689 | .sparsity(0.5f) |
| 1690 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1691 | } |
| 1692 | } |
| 1693 | } |
| 1694 | |
| 1695 | TEST(F32_SPMM_8X1__NEONFMA_PIPELINED, zero_weights) { |
| 1696 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1697 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1698 | for (size_t k = 1; k <= 5; k += 2) { |
| 1699 | SpMMMicrokernelTester() |
| 1700 | .mr(8) |
| 1701 | .nr(1) |
| 1702 | .m(16) |
| 1703 | .n(n) |
| 1704 | .k(k) |
| 1705 | .sparsity(1.0f) |
| 1706 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_pipelined); |
| 1707 | } |
| 1708 | } |
| 1709 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1710 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1711 | |
| 1712 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1713 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1714 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, k_eq_2) { |
| 1715 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1716 | SpMMMicrokernelTester() |
| 1717 | .mr(8) |
| 1718 | .nr(1) |
| 1719 | .m(8) |
| 1720 | .n(1) |
| 1721 | .k(2) |
| 1722 | .sparsity(0.0f) |
| 1723 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1724 | } |
| 1725 | |
| 1726 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, k_lt_2) { |
| 1727 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1728 | for (size_t k = 1; k < 2; k++) { |
| 1729 | SpMMMicrokernelTester() |
| 1730 | .mr(8) |
| 1731 | .nr(1) |
| 1732 | .m(8) |
| 1733 | .n(1) |
| 1734 | .k(k) |
| 1735 | .sparsity(0.0f) |
| 1736 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1737 | } |
| 1738 | } |
| 1739 | |
| 1740 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, k_gt_2) { |
| 1741 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1742 | for (size_t k = 3; k < 4; k++) { |
| 1743 | SpMMMicrokernelTester() |
| 1744 | .mr(8) |
| 1745 | .nr(1) |
| 1746 | .m(8) |
| 1747 | .n(1) |
| 1748 | .k(k) |
| 1749 | .sparsity(0.0f) |
| 1750 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1751 | } |
| 1752 | } |
| 1753 | |
| 1754 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, k_div_2) { |
| 1755 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1756 | for (size_t k = 4; k <= 20; k += 2) { |
| 1757 | SpMMMicrokernelTester() |
| 1758 | .mr(8) |
| 1759 | .nr(1) |
| 1760 | .m(8) |
| 1761 | .n(1) |
| 1762 | .k(k) |
| 1763 | .sparsity(0.0f) |
| 1764 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1765 | } |
| 1766 | } |
| 1767 | |
| 1768 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, n_gt_1) { |
| 1769 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1770 | for (uint32_t n = 2; n < 10; n++) { |
| 1771 | for (size_t k = 1; k <= 10; k += 3) { |
| 1772 | SpMMMicrokernelTester() |
| 1773 | .mr(8) |
| 1774 | .nr(1) |
| 1775 | .m(8) |
| 1776 | .n(n) |
| 1777 | .k(k) |
| 1778 | .sparsity(0.0f) |
| 1779 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1780 | } |
| 1781 | } |
| 1782 | } |
| 1783 | |
| 1784 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, m_lt_8) { |
| 1785 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1786 | for (uint32_t m = 1; m < 8; m++) { |
| 1787 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1788 | for (size_t k = 1; k <= 10; k += 3) { |
| 1789 | SpMMMicrokernelTester() |
| 1790 | .mr(8) |
| 1791 | .nr(1) |
| 1792 | .m(m) |
| 1793 | .n(n) |
| 1794 | .k(k) |
| 1795 | .sparsity(0.0f) |
| 1796 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1797 | } |
| 1798 | } |
| 1799 | } |
| 1800 | } |
| 1801 | |
| 1802 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, m_div_8) { |
| 1803 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1804 | for (uint32_t m = 16; m <= 24; m += 8) { |
| 1805 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1806 | for (size_t k = 1; k <= 10; k += 3) { |
| 1807 | SpMMMicrokernelTester() |
| 1808 | .mr(8) |
| 1809 | .nr(1) |
| 1810 | .m(m) |
| 1811 | .n(n) |
| 1812 | .k(k) |
| 1813 | .sparsity(0.0f) |
| 1814 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1815 | } |
| 1816 | } |
| 1817 | } |
| 1818 | } |
| 1819 | |
| 1820 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, m_gt_8) { |
| 1821 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1822 | for (uint32_t m = 9; m < 16; m++) { |
| 1823 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1824 | for (size_t k = 1; k <= 10; k += 3) { |
| 1825 | SpMMMicrokernelTester() |
| 1826 | .mr(8) |
| 1827 | .nr(1) |
| 1828 | .m(m) |
| 1829 | .n(n) |
| 1830 | .k(k) |
| 1831 | .sparsity(0.0f) |
| 1832 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1833 | } |
| 1834 | } |
| 1835 | } |
| 1836 | } |
| 1837 | |
| 1838 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, qmin) { |
| 1839 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1840 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1841 | for (size_t k = 1; k <= 10; k += 3) { |
| 1842 | SpMMMicrokernelTester() |
| 1843 | .mr(8) |
| 1844 | .nr(1) |
| 1845 | .m(16) |
| 1846 | .n(n) |
| 1847 | .k(k) |
| 1848 | .sparsity(0.0f) |
| 1849 | .qmin(128) |
| 1850 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1851 | } |
| 1852 | } |
| 1853 | } |
| 1854 | |
| 1855 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, qmax) { |
| 1856 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1857 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1858 | for (size_t k = 1; k <= 10; k += 3) { |
| 1859 | SpMMMicrokernelTester() |
| 1860 | .mr(8) |
| 1861 | .nr(1) |
| 1862 | .m(16) |
| 1863 | .n(n) |
| 1864 | .k(k) |
| 1865 | .sparsity(0.0f) |
| 1866 | .qmax(128) |
| 1867 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1868 | } |
| 1869 | } |
| 1870 | } |
| 1871 | |
| 1872 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, half_sparse) { |
| 1873 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1874 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1875 | for (size_t k = 1; k <= 10; k += 3) { |
| 1876 | SpMMMicrokernelTester() |
| 1877 | .mr(8) |
| 1878 | .nr(1) |
| 1879 | .m(16) |
| 1880 | .n(n) |
| 1881 | .k(k) |
| 1882 | .sparsity(0.5f) |
| 1883 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1884 | } |
| 1885 | } |
| 1886 | } |
| 1887 | |
| 1888 | TEST(F32_SPMM_8X1__NEONFMA_UNROLL2, zero_weights) { |
| 1889 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1890 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1891 | for (size_t k = 1; k <= 10; k += 3) { |
| 1892 | SpMMMicrokernelTester() |
| 1893 | .mr(8) |
| 1894 | .nr(1) |
| 1895 | .m(16) |
| 1896 | .n(n) |
| 1897 | .k(k) |
| 1898 | .sparsity(1.0f) |
| 1899 | .Test(xnn_f32_spmm_ukernel_8x1__neonfma_unroll2); |
| 1900 | } |
| 1901 | } |
| 1902 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1903 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1904 | |
| 1905 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1906 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1907 | TEST(F32_SPMM_12X1__NEONFMA, k_eq_1) { |
| 1908 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1909 | SpMMMicrokernelTester() |
| 1910 | .mr(12) |
| 1911 | .nr(1) |
| 1912 | .m(12) |
| 1913 | .n(1) |
| 1914 | .k(1) |
| 1915 | .sparsity(0.0f) |
| 1916 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 1917 | } |
| 1918 | |
| 1919 | TEST(F32_SPMM_12X1__NEONFMA, k_gt_1) { |
| 1920 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1921 | for (size_t k = 2; k < 10; k++) { |
| 1922 | SpMMMicrokernelTester() |
| 1923 | .mr(12) |
| 1924 | .nr(1) |
| 1925 | .m(12) |
| 1926 | .n(1) |
| 1927 | .k(k) |
| 1928 | .sparsity(0.0f) |
| 1929 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 1930 | } |
| 1931 | } |
| 1932 | |
| 1933 | TEST(F32_SPMM_12X1__NEONFMA, n_gt_1) { |
| 1934 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1935 | for (uint32_t n = 2; n < 10; n++) { |
| 1936 | for (size_t k = 1; k <= 5; k += 2) { |
| 1937 | SpMMMicrokernelTester() |
| 1938 | .mr(12) |
| 1939 | .nr(1) |
| 1940 | .m(12) |
| 1941 | .n(n) |
| 1942 | .k(k) |
| 1943 | .sparsity(0.0f) |
| 1944 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 1945 | } |
| 1946 | } |
| 1947 | } |
| 1948 | |
| 1949 | TEST(F32_SPMM_12X1__NEONFMA, m_lt_12) { |
| 1950 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1951 | for (uint32_t m = 1; m < 12; m++) { |
| 1952 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1953 | for (size_t k = 1; k <= 5; k += 2) { |
| 1954 | SpMMMicrokernelTester() |
| 1955 | .mr(12) |
| 1956 | .nr(1) |
| 1957 | .m(m) |
| 1958 | .n(n) |
| 1959 | .k(k) |
| 1960 | .sparsity(0.0f) |
| 1961 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 1962 | } |
| 1963 | } |
| 1964 | } |
| 1965 | } |
| 1966 | |
| 1967 | TEST(F32_SPMM_12X1__NEONFMA, m_div_12) { |
| 1968 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1969 | for (uint32_t m = 24; m <= 36; m += 12) { |
| 1970 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1971 | for (size_t k = 1; k <= 5; k += 2) { |
| 1972 | SpMMMicrokernelTester() |
| 1973 | .mr(12) |
| 1974 | .nr(1) |
| 1975 | .m(m) |
| 1976 | .n(n) |
| 1977 | .k(k) |
| 1978 | .sparsity(0.0f) |
| 1979 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 1980 | } |
| 1981 | } |
| 1982 | } |
| 1983 | } |
| 1984 | |
| 1985 | TEST(F32_SPMM_12X1__NEONFMA, m_gt_12) { |
| 1986 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1987 | for (uint32_t m = 13; m < 24; m++) { |
| 1988 | for (uint32_t n = 1; n < 10; n += 2) { |
| 1989 | for (size_t k = 1; k <= 5; k += 2) { |
| 1990 | SpMMMicrokernelTester() |
| 1991 | .mr(12) |
| 1992 | .nr(1) |
| 1993 | .m(m) |
| 1994 | .n(n) |
| 1995 | .k(k) |
| 1996 | .sparsity(0.0f) |
| 1997 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 1998 | } |
| 1999 | } |
| 2000 | } |
| 2001 | } |
| 2002 | |
| 2003 | TEST(F32_SPMM_12X1__NEONFMA, qmin) { |
| 2004 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2005 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2006 | for (size_t k = 1; k <= 5; k += 2) { |
| 2007 | SpMMMicrokernelTester() |
| 2008 | .mr(12) |
| 2009 | .nr(1) |
| 2010 | .m(24) |
| 2011 | .n(n) |
| 2012 | .k(k) |
| 2013 | .sparsity(0.0f) |
| 2014 | .qmin(128) |
| 2015 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 2016 | } |
| 2017 | } |
| 2018 | } |
| 2019 | |
| 2020 | TEST(F32_SPMM_12X1__NEONFMA, qmax) { |
| 2021 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2022 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2023 | for (size_t k = 1; k <= 5; k += 2) { |
| 2024 | SpMMMicrokernelTester() |
| 2025 | .mr(12) |
| 2026 | .nr(1) |
| 2027 | .m(24) |
| 2028 | .n(n) |
| 2029 | .k(k) |
| 2030 | .sparsity(0.0f) |
| 2031 | .qmax(128) |
| 2032 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 2033 | } |
| 2034 | } |
| 2035 | } |
| 2036 | |
| 2037 | TEST(F32_SPMM_12X1__NEONFMA, half_sparse) { |
| 2038 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2039 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2040 | for (size_t k = 1; k <= 5; k += 2) { |
| 2041 | SpMMMicrokernelTester() |
| 2042 | .mr(12) |
| 2043 | .nr(1) |
| 2044 | .m(24) |
| 2045 | .n(n) |
| 2046 | .k(k) |
| 2047 | .sparsity(0.5f) |
| 2048 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 2049 | } |
| 2050 | } |
| 2051 | } |
| 2052 | |
| 2053 | TEST(F32_SPMM_12X1__NEONFMA, zero_weights) { |
| 2054 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2055 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2056 | for (size_t k = 1; k <= 5; k += 2) { |
| 2057 | SpMMMicrokernelTester() |
| 2058 | .mr(12) |
| 2059 | .nr(1) |
| 2060 | .m(24) |
| 2061 | .n(n) |
| 2062 | .k(k) |
| 2063 | .sparsity(1.0f) |
| 2064 | .Test(xnn_f32_spmm_ukernel_12x1__neonfma); |
| 2065 | } |
| 2066 | } |
| 2067 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2068 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2069 | |
| 2070 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2071 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2072 | TEST(F32_SPMM_12X2__NEONFMA, k_eq_1) { |
| 2073 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2074 | SpMMMicrokernelTester() |
| 2075 | .mr(12) |
| 2076 | .nr(2) |
| 2077 | .m(12) |
| 2078 | .n(2) |
| 2079 | .k(1) |
| 2080 | .sparsity(0.0f) |
| 2081 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2082 | } |
| 2083 | |
| 2084 | TEST(F32_SPMM_12X2__NEONFMA, k_eq_1_subtile) { |
| 2085 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2086 | for (uint32_t n = 1; n <= 2; n++) { |
| 2087 | SpMMMicrokernelTester() |
| 2088 | .mr(12) |
| 2089 | .nr(2) |
| 2090 | .m(12) |
| 2091 | .n(n) |
| 2092 | .k(1) |
| 2093 | .sparsity(0.0f) |
| 2094 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2095 | } |
| 2096 | } |
| 2097 | |
| 2098 | TEST(F32_SPMM_12X2__NEONFMA, k_gt_1) { |
| 2099 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2100 | for (size_t k = 2; k < 10; k++) { |
| 2101 | SpMMMicrokernelTester() |
| 2102 | .mr(12) |
| 2103 | .nr(2) |
| 2104 | .m(12) |
| 2105 | .n(2) |
| 2106 | .k(k) |
| 2107 | .sparsity(0.0f) |
| 2108 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2109 | } |
| 2110 | } |
| 2111 | |
| 2112 | TEST(F32_SPMM_12X2__NEONFMA, k_gt_1_subtile) { |
| 2113 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2114 | for (size_t k = 2; k < 10; k++) { |
| 2115 | for (uint32_t n = 1; n <= 2; n++) { |
| 2116 | SpMMMicrokernelTester() |
| 2117 | .mr(12) |
| 2118 | .nr(2) |
| 2119 | .m(12) |
| 2120 | .n(n) |
| 2121 | .k(k) |
| 2122 | .sparsity(0.0f) |
| 2123 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2124 | } |
| 2125 | } |
| 2126 | } |
| 2127 | |
| 2128 | TEST(F32_SPMM_12X2__NEONFMA, n_gt_2) { |
| 2129 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2130 | for (uint32_t n = 3; n < 10; n++) { |
| 2131 | for (size_t k = 1; k <= 5; k += 2) { |
| 2132 | SpMMMicrokernelTester() |
| 2133 | .mr(12) |
| 2134 | .nr(2) |
| 2135 | .m(12) |
| 2136 | .n(n) |
| 2137 | .k(k) |
| 2138 | .sparsity(0.0f) |
| 2139 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2140 | } |
| 2141 | } |
| 2142 | } |
| 2143 | |
| 2144 | TEST(F32_SPMM_12X2__NEONFMA, n_div_2) { |
| 2145 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2146 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2147 | for (size_t k = 1; k <= 5; k += 2) { |
| 2148 | SpMMMicrokernelTester() |
| 2149 | .mr(12) |
| 2150 | .nr(2) |
| 2151 | .m(12) |
| 2152 | .n(n) |
| 2153 | .k(k) |
| 2154 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2155 | } |
| 2156 | } |
| 2157 | } |
| 2158 | |
| 2159 | TEST(F32_SPMM_12X2__NEONFMA, m_lt_12) { |
| 2160 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2161 | for (uint32_t m = 1; m < 12; m++) { |
| 2162 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2163 | for (size_t k = 1; k <= 5; k += 2) { |
| 2164 | SpMMMicrokernelTester() |
| 2165 | .mr(12) |
| 2166 | .nr(2) |
| 2167 | .m(m) |
| 2168 | .n(n) |
| 2169 | .k(k) |
| 2170 | .sparsity(0.0f) |
| 2171 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2172 | } |
| 2173 | } |
| 2174 | } |
| 2175 | } |
| 2176 | |
| 2177 | TEST(F32_SPMM_12X2__NEONFMA, m_div_12) { |
| 2178 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2179 | for (uint32_t m = 24; m <= 36; m += 12) { |
| 2180 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2181 | for (size_t k = 1; k <= 5; k += 2) { |
| 2182 | SpMMMicrokernelTester() |
| 2183 | .mr(12) |
| 2184 | .nr(2) |
| 2185 | .m(m) |
| 2186 | .n(n) |
| 2187 | .k(k) |
| 2188 | .sparsity(0.0f) |
| 2189 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2190 | } |
| 2191 | } |
| 2192 | } |
| 2193 | } |
| 2194 | |
| 2195 | TEST(F32_SPMM_12X2__NEONFMA, m_gt_12) { |
| 2196 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2197 | for (uint32_t m = 13; m < 24; m++) { |
| 2198 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2199 | for (size_t k = 1; k <= 5; k += 2) { |
| 2200 | SpMMMicrokernelTester() |
| 2201 | .mr(12) |
| 2202 | .nr(2) |
| 2203 | .m(m) |
| 2204 | .n(n) |
| 2205 | .k(k) |
| 2206 | .sparsity(0.0f) |
| 2207 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2208 | } |
| 2209 | } |
| 2210 | } |
| 2211 | } |
| 2212 | |
| 2213 | TEST(F32_SPMM_12X2__NEONFMA, qmin) { |
| 2214 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2215 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2216 | for (size_t k = 1; k <= 5; k += 2) { |
| 2217 | SpMMMicrokernelTester() |
| 2218 | .mr(12) |
| 2219 | .nr(2) |
| 2220 | .m(24) |
| 2221 | .n(n) |
| 2222 | .k(k) |
| 2223 | .sparsity(0.0f) |
| 2224 | .qmin(128) |
| 2225 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2226 | } |
| 2227 | } |
| 2228 | } |
| 2229 | |
| 2230 | TEST(F32_SPMM_12X2__NEONFMA, qmax) { |
| 2231 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2232 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2233 | for (size_t k = 1; k <= 5; k += 2) { |
| 2234 | SpMMMicrokernelTester() |
| 2235 | .mr(12) |
| 2236 | .nr(2) |
| 2237 | .m(24) |
| 2238 | .n(n) |
| 2239 | .k(k) |
| 2240 | .sparsity(0.0f) |
| 2241 | .qmax(128) |
| 2242 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2243 | } |
| 2244 | } |
| 2245 | } |
| 2246 | |
| 2247 | TEST(F32_SPMM_12X2__NEONFMA, half_sparse) { |
| 2248 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2249 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2250 | for (size_t k = 1; k <= 5; k += 2) { |
| 2251 | SpMMMicrokernelTester() |
| 2252 | .mr(12) |
| 2253 | .nr(2) |
| 2254 | .m(24) |
| 2255 | .n(n) |
| 2256 | .k(k) |
| 2257 | .sparsity(0.5f) |
| 2258 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2259 | } |
| 2260 | } |
| 2261 | } |
| 2262 | |
| 2263 | TEST(F32_SPMM_12X2__NEONFMA, zero_weights) { |
| 2264 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2265 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2266 | for (size_t k = 1; k <= 5; k += 2) { |
| 2267 | SpMMMicrokernelTester() |
| 2268 | .mr(12) |
| 2269 | .nr(2) |
| 2270 | .m(24) |
| 2271 | .n(n) |
| 2272 | .k(k) |
| 2273 | .sparsity(1.0f) |
| 2274 | .Test(xnn_f32_spmm_ukernel_12x2__neonfma); |
| 2275 | } |
| 2276 | } |
| 2277 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2278 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2279 | |
| 2280 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2281 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2282 | TEST(F32_SPMM_12X4__NEONFMA, k_eq_1) { |
| 2283 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2284 | SpMMMicrokernelTester() |
| 2285 | .mr(12) |
| 2286 | .nr(4) |
| 2287 | .m(12) |
| 2288 | .n(4) |
| 2289 | .k(1) |
| 2290 | .sparsity(0.0f) |
| 2291 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2292 | } |
| 2293 | |
| 2294 | TEST(F32_SPMM_12X4__NEONFMA, k_eq_1_subtile) { |
| 2295 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2296 | for (uint32_t n = 1; n <= 4; n++) { |
| 2297 | SpMMMicrokernelTester() |
| 2298 | .mr(12) |
| 2299 | .nr(4) |
| 2300 | .m(12) |
| 2301 | .n(n) |
| 2302 | .k(1) |
| 2303 | .sparsity(0.0f) |
| 2304 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2305 | } |
| 2306 | } |
| 2307 | |
| 2308 | TEST(F32_SPMM_12X4__NEONFMA, k_gt_1) { |
| 2309 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2310 | for (size_t k = 2; k < 10; k++) { |
| 2311 | SpMMMicrokernelTester() |
| 2312 | .mr(12) |
| 2313 | .nr(4) |
| 2314 | .m(12) |
| 2315 | .n(4) |
| 2316 | .k(k) |
| 2317 | .sparsity(0.0f) |
| 2318 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2319 | } |
| 2320 | } |
| 2321 | |
| 2322 | TEST(F32_SPMM_12X4__NEONFMA, k_gt_1_subtile) { |
| 2323 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2324 | for (size_t k = 2; k < 10; k++) { |
| 2325 | for (uint32_t n = 1; n <= 4; n++) { |
| 2326 | SpMMMicrokernelTester() |
| 2327 | .mr(12) |
| 2328 | .nr(4) |
| 2329 | .m(12) |
| 2330 | .n(n) |
| 2331 | .k(k) |
| 2332 | .sparsity(0.0f) |
| 2333 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2334 | } |
| 2335 | } |
| 2336 | } |
| 2337 | |
| 2338 | TEST(F32_SPMM_12X4__NEONFMA, n_gt_4) { |
| 2339 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2340 | for (uint32_t n = 5; n < 10; n++) { |
| 2341 | for (size_t k = 1; k <= 5; k += 2) { |
| 2342 | SpMMMicrokernelTester() |
| 2343 | .mr(12) |
| 2344 | .nr(4) |
| 2345 | .m(12) |
| 2346 | .n(n) |
| 2347 | .k(k) |
| 2348 | .sparsity(0.0f) |
| 2349 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2350 | } |
| 2351 | } |
| 2352 | } |
| 2353 | |
| 2354 | TEST(F32_SPMM_12X4__NEONFMA, n_div_4) { |
| 2355 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2356 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2357 | for (size_t k = 1; k <= 5; k += 2) { |
| 2358 | SpMMMicrokernelTester() |
| 2359 | .mr(12) |
| 2360 | .nr(4) |
| 2361 | .m(12) |
| 2362 | .n(n) |
| 2363 | .k(k) |
| 2364 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2365 | } |
| 2366 | } |
| 2367 | } |
| 2368 | |
| 2369 | TEST(F32_SPMM_12X4__NEONFMA, m_lt_12) { |
| 2370 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2371 | for (uint32_t m = 1; m < 12; m++) { |
| 2372 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2373 | for (size_t k = 1; k <= 5; k += 2) { |
| 2374 | SpMMMicrokernelTester() |
| 2375 | .mr(12) |
| 2376 | .nr(4) |
| 2377 | .m(m) |
| 2378 | .n(n) |
| 2379 | .k(k) |
| 2380 | .sparsity(0.0f) |
| 2381 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2382 | } |
| 2383 | } |
| 2384 | } |
| 2385 | } |
| 2386 | |
| 2387 | TEST(F32_SPMM_12X4__NEONFMA, m_div_12) { |
| 2388 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2389 | for (uint32_t m = 24; m <= 36; m += 12) { |
| 2390 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2391 | for (size_t k = 1; k <= 5; k += 2) { |
| 2392 | SpMMMicrokernelTester() |
| 2393 | .mr(12) |
| 2394 | .nr(4) |
| 2395 | .m(m) |
| 2396 | .n(n) |
| 2397 | .k(k) |
| 2398 | .sparsity(0.0f) |
| 2399 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2400 | } |
| 2401 | } |
| 2402 | } |
| 2403 | } |
| 2404 | |
| 2405 | TEST(F32_SPMM_12X4__NEONFMA, m_gt_12) { |
| 2406 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2407 | for (uint32_t m = 13; m < 24; m++) { |
| 2408 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2409 | for (size_t k = 1; k <= 5; k += 2) { |
| 2410 | SpMMMicrokernelTester() |
| 2411 | .mr(12) |
| 2412 | .nr(4) |
| 2413 | .m(m) |
| 2414 | .n(n) |
| 2415 | .k(k) |
| 2416 | .sparsity(0.0f) |
| 2417 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2418 | } |
| 2419 | } |
| 2420 | } |
| 2421 | } |
| 2422 | |
| 2423 | TEST(F32_SPMM_12X4__NEONFMA, qmin) { |
| 2424 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2425 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2426 | for (size_t k = 1; k <= 5; k += 2) { |
| 2427 | SpMMMicrokernelTester() |
| 2428 | .mr(12) |
| 2429 | .nr(4) |
| 2430 | .m(24) |
| 2431 | .n(n) |
| 2432 | .k(k) |
| 2433 | .sparsity(0.0f) |
| 2434 | .qmin(128) |
| 2435 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2436 | } |
| 2437 | } |
| 2438 | } |
| 2439 | |
| 2440 | TEST(F32_SPMM_12X4__NEONFMA, qmax) { |
| 2441 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2442 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2443 | for (size_t k = 1; k <= 5; k += 2) { |
| 2444 | SpMMMicrokernelTester() |
| 2445 | .mr(12) |
| 2446 | .nr(4) |
| 2447 | .m(24) |
| 2448 | .n(n) |
| 2449 | .k(k) |
| 2450 | .sparsity(0.0f) |
| 2451 | .qmax(128) |
| 2452 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2453 | } |
| 2454 | } |
| 2455 | } |
| 2456 | |
| 2457 | TEST(F32_SPMM_12X4__NEONFMA, half_sparse) { |
| 2458 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2459 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2460 | for (size_t k = 1; k <= 5; k += 2) { |
| 2461 | SpMMMicrokernelTester() |
| 2462 | .mr(12) |
| 2463 | .nr(4) |
| 2464 | .m(24) |
| 2465 | .n(n) |
| 2466 | .k(k) |
| 2467 | .sparsity(0.5f) |
| 2468 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2469 | } |
| 2470 | } |
| 2471 | } |
| 2472 | |
| 2473 | TEST(F32_SPMM_12X4__NEONFMA, zero_weights) { |
| 2474 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2475 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2476 | for (size_t k = 1; k <= 5; k += 2) { |
| 2477 | SpMMMicrokernelTester() |
| 2478 | .mr(12) |
| 2479 | .nr(4) |
| 2480 | .m(24) |
| 2481 | .n(n) |
| 2482 | .k(k) |
| 2483 | .sparsity(1.0f) |
| 2484 | .Test(xnn_f32_spmm_ukernel_12x4__neonfma); |
| 2485 | } |
| 2486 | } |
| 2487 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2488 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2489 | |
| 2490 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2491 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2492 | TEST(F32_SPMM_16X1__NEONFMA, k_eq_1) { |
| 2493 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2494 | SpMMMicrokernelTester() |
| 2495 | .mr(16) |
| 2496 | .nr(1) |
| 2497 | .m(16) |
| 2498 | .n(1) |
| 2499 | .k(1) |
| 2500 | .sparsity(0.0f) |
| 2501 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2502 | } |
| 2503 | |
| 2504 | TEST(F32_SPMM_16X1__NEONFMA, k_gt_1) { |
| 2505 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2506 | for (size_t k = 2; k < 10; k++) { |
| 2507 | SpMMMicrokernelTester() |
| 2508 | .mr(16) |
| 2509 | .nr(1) |
| 2510 | .m(16) |
| 2511 | .n(1) |
| 2512 | .k(k) |
| 2513 | .sparsity(0.0f) |
| 2514 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2515 | } |
| 2516 | } |
| 2517 | |
| 2518 | TEST(F32_SPMM_16X1__NEONFMA, n_gt_1) { |
| 2519 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2520 | for (uint32_t n = 2; n < 10; n++) { |
| 2521 | for (size_t k = 1; k <= 5; k += 2) { |
| 2522 | SpMMMicrokernelTester() |
| 2523 | .mr(16) |
| 2524 | .nr(1) |
| 2525 | .m(16) |
| 2526 | .n(n) |
| 2527 | .k(k) |
| 2528 | .sparsity(0.0f) |
| 2529 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2530 | } |
| 2531 | } |
| 2532 | } |
| 2533 | |
| 2534 | TEST(F32_SPMM_16X1__NEONFMA, m_lt_16) { |
| 2535 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2536 | for (uint32_t m = 1; m < 16; m++) { |
| 2537 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2538 | for (size_t k = 1; k <= 5; k += 2) { |
| 2539 | SpMMMicrokernelTester() |
| 2540 | .mr(16) |
| 2541 | .nr(1) |
| 2542 | .m(m) |
| 2543 | .n(n) |
| 2544 | .k(k) |
| 2545 | .sparsity(0.0f) |
| 2546 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2547 | } |
| 2548 | } |
| 2549 | } |
| 2550 | } |
| 2551 | |
| 2552 | TEST(F32_SPMM_16X1__NEONFMA, m_div_16) { |
| 2553 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2554 | for (uint32_t m = 32; m <= 48; m += 16) { |
| 2555 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2556 | for (size_t k = 1; k <= 5; k += 2) { |
| 2557 | SpMMMicrokernelTester() |
| 2558 | .mr(16) |
| 2559 | .nr(1) |
| 2560 | .m(m) |
| 2561 | .n(n) |
| 2562 | .k(k) |
| 2563 | .sparsity(0.0f) |
| 2564 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2565 | } |
| 2566 | } |
| 2567 | } |
| 2568 | } |
| 2569 | |
| 2570 | TEST(F32_SPMM_16X1__NEONFMA, m_gt_16) { |
| 2571 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2572 | for (uint32_t m = 17; m < 32; m++) { |
| 2573 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2574 | for (size_t k = 1; k <= 5; k += 2) { |
| 2575 | SpMMMicrokernelTester() |
| 2576 | .mr(16) |
| 2577 | .nr(1) |
| 2578 | .m(m) |
| 2579 | .n(n) |
| 2580 | .k(k) |
| 2581 | .sparsity(0.0f) |
| 2582 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2583 | } |
| 2584 | } |
| 2585 | } |
| 2586 | } |
| 2587 | |
| 2588 | TEST(F32_SPMM_16X1__NEONFMA, qmin) { |
| 2589 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2590 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2591 | for (size_t k = 1; k <= 5; k += 2) { |
| 2592 | SpMMMicrokernelTester() |
| 2593 | .mr(16) |
| 2594 | .nr(1) |
| 2595 | .m(32) |
| 2596 | .n(n) |
| 2597 | .k(k) |
| 2598 | .sparsity(0.0f) |
| 2599 | .qmin(128) |
| 2600 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2601 | } |
| 2602 | } |
| 2603 | } |
| 2604 | |
| 2605 | TEST(F32_SPMM_16X1__NEONFMA, qmax) { |
| 2606 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2607 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2608 | for (size_t k = 1; k <= 5; k += 2) { |
| 2609 | SpMMMicrokernelTester() |
| 2610 | .mr(16) |
| 2611 | .nr(1) |
| 2612 | .m(32) |
| 2613 | .n(n) |
| 2614 | .k(k) |
| 2615 | .sparsity(0.0f) |
| 2616 | .qmax(128) |
| 2617 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2618 | } |
| 2619 | } |
| 2620 | } |
| 2621 | |
| 2622 | TEST(F32_SPMM_16X1__NEONFMA, half_sparse) { |
| 2623 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2624 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2625 | for (size_t k = 1; k <= 5; k += 2) { |
| 2626 | SpMMMicrokernelTester() |
| 2627 | .mr(16) |
| 2628 | .nr(1) |
| 2629 | .m(32) |
| 2630 | .n(n) |
| 2631 | .k(k) |
| 2632 | .sparsity(0.5f) |
| 2633 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2634 | } |
| 2635 | } |
| 2636 | } |
| 2637 | |
| 2638 | TEST(F32_SPMM_16X1__NEONFMA, zero_weights) { |
| 2639 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2640 | for (uint32_t n = 1; n < 10; n += 2) { |
| 2641 | for (size_t k = 1; k <= 5; k += 2) { |
| 2642 | SpMMMicrokernelTester() |
| 2643 | .mr(16) |
| 2644 | .nr(1) |
| 2645 | .m(32) |
| 2646 | .n(n) |
| 2647 | .k(k) |
| 2648 | .sparsity(1.0f) |
| 2649 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma); |
| 2650 | } |
| 2651 | } |
| 2652 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2653 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2654 | |
| 2655 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2656 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2657 | TEST(F32_SPMM_16X2__NEONFMA, k_eq_1) { |
| 2658 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2659 | SpMMMicrokernelTester() |
| 2660 | .mr(16) |
| 2661 | .nr(2) |
| 2662 | .m(16) |
| 2663 | .n(2) |
| 2664 | .k(1) |
| 2665 | .sparsity(0.0f) |
| 2666 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2667 | } |
| 2668 | |
| 2669 | TEST(F32_SPMM_16X2__NEONFMA, k_eq_1_subtile) { |
| 2670 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2671 | for (uint32_t n = 1; n <= 2; n++) { |
| 2672 | SpMMMicrokernelTester() |
| 2673 | .mr(16) |
| 2674 | .nr(2) |
| 2675 | .m(16) |
| 2676 | .n(n) |
| 2677 | .k(1) |
| 2678 | .sparsity(0.0f) |
| 2679 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2680 | } |
| 2681 | } |
| 2682 | |
| 2683 | TEST(F32_SPMM_16X2__NEONFMA, k_gt_1) { |
| 2684 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2685 | for (size_t k = 2; k < 10; k++) { |
| 2686 | SpMMMicrokernelTester() |
| 2687 | .mr(16) |
| 2688 | .nr(2) |
| 2689 | .m(16) |
| 2690 | .n(2) |
| 2691 | .k(k) |
| 2692 | .sparsity(0.0f) |
| 2693 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2694 | } |
| 2695 | } |
| 2696 | |
| 2697 | TEST(F32_SPMM_16X2__NEONFMA, k_gt_1_subtile) { |
| 2698 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2699 | for (size_t k = 2; k < 10; k++) { |
| 2700 | for (uint32_t n = 1; n <= 2; n++) { |
| 2701 | SpMMMicrokernelTester() |
| 2702 | .mr(16) |
| 2703 | .nr(2) |
| 2704 | .m(16) |
| 2705 | .n(n) |
| 2706 | .k(k) |
| 2707 | .sparsity(0.0f) |
| 2708 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2709 | } |
| 2710 | } |
| 2711 | } |
| 2712 | |
| 2713 | TEST(F32_SPMM_16X2__NEONFMA, n_gt_2) { |
| 2714 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2715 | for (uint32_t n = 3; n < 10; n++) { |
| 2716 | for (size_t k = 1; k <= 5; k += 2) { |
| 2717 | SpMMMicrokernelTester() |
| 2718 | .mr(16) |
| 2719 | .nr(2) |
| 2720 | .m(16) |
| 2721 | .n(n) |
| 2722 | .k(k) |
| 2723 | .sparsity(0.0f) |
| 2724 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2725 | } |
| 2726 | } |
| 2727 | } |
| 2728 | |
| 2729 | TEST(F32_SPMM_16X2__NEONFMA, n_div_2) { |
| 2730 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2731 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 2732 | for (size_t k = 1; k <= 5; k += 2) { |
| 2733 | SpMMMicrokernelTester() |
| 2734 | .mr(16) |
| 2735 | .nr(2) |
| 2736 | .m(16) |
| 2737 | .n(n) |
| 2738 | .k(k) |
| 2739 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2740 | } |
| 2741 | } |
| 2742 | } |
| 2743 | |
| 2744 | TEST(F32_SPMM_16X2__NEONFMA, m_lt_16) { |
| 2745 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2746 | for (uint32_t m = 1; m < 16; m++) { |
| 2747 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2748 | for (size_t k = 1; k <= 5; k += 2) { |
| 2749 | SpMMMicrokernelTester() |
| 2750 | .mr(16) |
| 2751 | .nr(2) |
| 2752 | .m(m) |
| 2753 | .n(n) |
| 2754 | .k(k) |
| 2755 | .sparsity(0.0f) |
| 2756 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2757 | } |
| 2758 | } |
| 2759 | } |
| 2760 | } |
| 2761 | |
| 2762 | TEST(F32_SPMM_16X2__NEONFMA, m_div_16) { |
| 2763 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2764 | for (uint32_t m = 32; m <= 48; m += 16) { |
| 2765 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2766 | for (size_t k = 1; k <= 5; k += 2) { |
| 2767 | SpMMMicrokernelTester() |
| 2768 | .mr(16) |
| 2769 | .nr(2) |
| 2770 | .m(m) |
| 2771 | .n(n) |
| 2772 | .k(k) |
| 2773 | .sparsity(0.0f) |
| 2774 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2775 | } |
| 2776 | } |
| 2777 | } |
| 2778 | } |
| 2779 | |
| 2780 | TEST(F32_SPMM_16X2__NEONFMA, m_gt_16) { |
| 2781 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2782 | for (uint32_t m = 17; m < 32; m++) { |
| 2783 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2784 | for (size_t k = 1; k <= 5; k += 2) { |
| 2785 | SpMMMicrokernelTester() |
| 2786 | .mr(16) |
| 2787 | .nr(2) |
| 2788 | .m(m) |
| 2789 | .n(n) |
| 2790 | .k(k) |
| 2791 | .sparsity(0.0f) |
| 2792 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2793 | } |
| 2794 | } |
| 2795 | } |
| 2796 | } |
| 2797 | |
| 2798 | TEST(F32_SPMM_16X2__NEONFMA, qmin) { |
| 2799 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2800 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2801 | for (size_t k = 1; k <= 5; k += 2) { |
| 2802 | SpMMMicrokernelTester() |
| 2803 | .mr(16) |
| 2804 | .nr(2) |
| 2805 | .m(32) |
| 2806 | .n(n) |
| 2807 | .k(k) |
| 2808 | .sparsity(0.0f) |
| 2809 | .qmin(128) |
| 2810 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2811 | } |
| 2812 | } |
| 2813 | } |
| 2814 | |
| 2815 | TEST(F32_SPMM_16X2__NEONFMA, qmax) { |
| 2816 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2817 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2818 | for (size_t k = 1; k <= 5; k += 2) { |
| 2819 | SpMMMicrokernelTester() |
| 2820 | .mr(16) |
| 2821 | .nr(2) |
| 2822 | .m(32) |
| 2823 | .n(n) |
| 2824 | .k(k) |
| 2825 | .sparsity(0.0f) |
| 2826 | .qmax(128) |
| 2827 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2828 | } |
| 2829 | } |
| 2830 | } |
| 2831 | |
| 2832 | TEST(F32_SPMM_16X2__NEONFMA, half_sparse) { |
| 2833 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2834 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2835 | for (size_t k = 1; k <= 5; k += 2) { |
| 2836 | SpMMMicrokernelTester() |
| 2837 | .mr(16) |
| 2838 | .nr(2) |
| 2839 | .m(32) |
| 2840 | .n(n) |
| 2841 | .k(k) |
| 2842 | .sparsity(0.5f) |
| 2843 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2844 | } |
| 2845 | } |
| 2846 | } |
| 2847 | |
| 2848 | TEST(F32_SPMM_16X2__NEONFMA, zero_weights) { |
| 2849 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2850 | for (uint32_t n = 1; n < 10; n += 3) { |
| 2851 | for (size_t k = 1; k <= 5; k += 2) { |
| 2852 | SpMMMicrokernelTester() |
| 2853 | .mr(16) |
| 2854 | .nr(2) |
| 2855 | .m(32) |
| 2856 | .n(n) |
| 2857 | .k(k) |
| 2858 | .sparsity(1.0f) |
| 2859 | .Test(xnn_f32_spmm_ukernel_16x2__neonfma); |
| 2860 | } |
| 2861 | } |
| 2862 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2863 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2864 | |
| 2865 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2866 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2867 | TEST(F32_SPMM_16X4__NEONFMA, k_eq_1) { |
| 2868 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2869 | SpMMMicrokernelTester() |
| 2870 | .mr(16) |
| 2871 | .nr(4) |
| 2872 | .m(16) |
| 2873 | .n(4) |
| 2874 | .k(1) |
| 2875 | .sparsity(0.0f) |
| 2876 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 2877 | } |
| 2878 | |
| 2879 | TEST(F32_SPMM_16X4__NEONFMA, k_eq_1_subtile) { |
| 2880 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2881 | for (uint32_t n = 1; n <= 4; n++) { |
| 2882 | SpMMMicrokernelTester() |
| 2883 | .mr(16) |
| 2884 | .nr(4) |
| 2885 | .m(16) |
| 2886 | .n(n) |
| 2887 | .k(1) |
| 2888 | .sparsity(0.0f) |
| 2889 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 2890 | } |
| 2891 | } |
| 2892 | |
| 2893 | TEST(F32_SPMM_16X4__NEONFMA, k_gt_1) { |
| 2894 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2895 | for (size_t k = 2; k < 10; k++) { |
| 2896 | SpMMMicrokernelTester() |
| 2897 | .mr(16) |
| 2898 | .nr(4) |
| 2899 | .m(16) |
| 2900 | .n(4) |
| 2901 | .k(k) |
| 2902 | .sparsity(0.0f) |
| 2903 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 2904 | } |
| 2905 | } |
| 2906 | |
| 2907 | TEST(F32_SPMM_16X4__NEONFMA, k_gt_1_subtile) { |
| 2908 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2909 | for (size_t k = 2; k < 10; k++) { |
| 2910 | for (uint32_t n = 1; n <= 4; n++) { |
| 2911 | SpMMMicrokernelTester() |
| 2912 | .mr(16) |
| 2913 | .nr(4) |
| 2914 | .m(16) |
| 2915 | .n(n) |
| 2916 | .k(k) |
| 2917 | .sparsity(0.0f) |
| 2918 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 2919 | } |
| 2920 | } |
| 2921 | } |
| 2922 | |
| 2923 | TEST(F32_SPMM_16X4__NEONFMA, n_gt_4) { |
| 2924 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2925 | for (uint32_t n = 5; n < 10; n++) { |
| 2926 | for (size_t k = 1; k <= 5; k += 2) { |
| 2927 | SpMMMicrokernelTester() |
| 2928 | .mr(16) |
| 2929 | .nr(4) |
| 2930 | .m(16) |
| 2931 | .n(n) |
| 2932 | .k(k) |
| 2933 | .sparsity(0.0f) |
| 2934 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 2935 | } |
| 2936 | } |
| 2937 | } |
| 2938 | |
| 2939 | TEST(F32_SPMM_16X4__NEONFMA, n_div_4) { |
| 2940 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2941 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 2942 | for (size_t k = 1; k <= 5; k += 2) { |
| 2943 | SpMMMicrokernelTester() |
| 2944 | .mr(16) |
| 2945 | .nr(4) |
| 2946 | .m(16) |
| 2947 | .n(n) |
| 2948 | .k(k) |
| 2949 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 2950 | } |
| 2951 | } |
| 2952 | } |
| 2953 | |
| 2954 | TEST(F32_SPMM_16X4__NEONFMA, m_lt_16) { |
| 2955 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2956 | for (uint32_t m = 1; m < 16; m++) { |
| 2957 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2958 | for (size_t k = 1; k <= 5; k += 2) { |
| 2959 | SpMMMicrokernelTester() |
| 2960 | .mr(16) |
| 2961 | .nr(4) |
| 2962 | .m(m) |
| 2963 | .n(n) |
| 2964 | .k(k) |
| 2965 | .sparsity(0.0f) |
| 2966 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 2967 | } |
| 2968 | } |
| 2969 | } |
| 2970 | } |
| 2971 | |
| 2972 | TEST(F32_SPMM_16X4__NEONFMA, m_div_16) { |
| 2973 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2974 | for (uint32_t m = 32; m <= 48; m += 16) { |
| 2975 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2976 | for (size_t k = 1; k <= 5; k += 2) { |
| 2977 | SpMMMicrokernelTester() |
| 2978 | .mr(16) |
| 2979 | .nr(4) |
| 2980 | .m(m) |
| 2981 | .n(n) |
| 2982 | .k(k) |
| 2983 | .sparsity(0.0f) |
| 2984 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 2985 | } |
| 2986 | } |
| 2987 | } |
| 2988 | } |
| 2989 | |
| 2990 | TEST(F32_SPMM_16X4__NEONFMA, m_gt_16) { |
| 2991 | TEST_REQUIRES_ARM_NEON_FMA; |
| 2992 | for (uint32_t m = 17; m < 32; m++) { |
| 2993 | for (uint32_t n = 1; n < 20; n += 5) { |
| 2994 | for (size_t k = 1; k <= 5; k += 2) { |
| 2995 | SpMMMicrokernelTester() |
| 2996 | .mr(16) |
| 2997 | .nr(4) |
| 2998 | .m(m) |
| 2999 | .n(n) |
| 3000 | .k(k) |
| 3001 | .sparsity(0.0f) |
| 3002 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 3003 | } |
| 3004 | } |
| 3005 | } |
| 3006 | } |
| 3007 | |
| 3008 | TEST(F32_SPMM_16X4__NEONFMA, qmin) { |
| 3009 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3010 | for (uint32_t n = 1; n < 20; n += 5) { |
| 3011 | for (size_t k = 1; k <= 5; k += 2) { |
| 3012 | SpMMMicrokernelTester() |
| 3013 | .mr(16) |
| 3014 | .nr(4) |
| 3015 | .m(32) |
| 3016 | .n(n) |
| 3017 | .k(k) |
| 3018 | .sparsity(0.0f) |
| 3019 | .qmin(128) |
| 3020 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 3021 | } |
| 3022 | } |
| 3023 | } |
| 3024 | |
| 3025 | TEST(F32_SPMM_16X4__NEONFMA, qmax) { |
| 3026 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3027 | for (uint32_t n = 1; n < 20; n += 5) { |
| 3028 | for (size_t k = 1; k <= 5; k += 2) { |
| 3029 | SpMMMicrokernelTester() |
| 3030 | .mr(16) |
| 3031 | .nr(4) |
| 3032 | .m(32) |
| 3033 | .n(n) |
| 3034 | .k(k) |
| 3035 | .sparsity(0.0f) |
| 3036 | .qmax(128) |
| 3037 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 3038 | } |
| 3039 | } |
| 3040 | } |
| 3041 | |
| 3042 | TEST(F32_SPMM_16X4__NEONFMA, half_sparse) { |
| 3043 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3044 | for (uint32_t n = 1; n < 20; n += 5) { |
| 3045 | for (size_t k = 1; k <= 5; k += 2) { |
| 3046 | SpMMMicrokernelTester() |
| 3047 | .mr(16) |
| 3048 | .nr(4) |
| 3049 | .m(32) |
| 3050 | .n(n) |
| 3051 | .k(k) |
| 3052 | .sparsity(0.5f) |
| 3053 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 3054 | } |
| 3055 | } |
| 3056 | } |
| 3057 | |
| 3058 | TEST(F32_SPMM_16X4__NEONFMA, zero_weights) { |
| 3059 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3060 | for (uint32_t n = 1; n < 20; n += 5) { |
| 3061 | for (size_t k = 1; k <= 5; k += 2) { |
| 3062 | SpMMMicrokernelTester() |
| 3063 | .mr(16) |
| 3064 | .nr(4) |
| 3065 | .m(32) |
| 3066 | .n(n) |
| 3067 | .k(k) |
| 3068 | .sparsity(1.0f) |
| 3069 | .Test(xnn_f32_spmm_ukernel_16x4__neonfma); |
| 3070 | } |
| 3071 | } |
| 3072 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 3073 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 3074 | |
| 3075 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 3076 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 3077 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, k_eq_1) { |
| 3078 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3079 | SpMMMicrokernelTester() |
| 3080 | .mr(16) |
| 3081 | .nr(1) |
| 3082 | .m(16) |
| 3083 | .n(1) |
| 3084 | .k(1) |
| 3085 | .sparsity(0.0f) |
| 3086 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3087 | } |
| 3088 | |
| 3089 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, k_gt_1) { |
| 3090 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3091 | for (size_t k = 2; k < 10; k++) { |
| 3092 | SpMMMicrokernelTester() |
| 3093 | .mr(16) |
| 3094 | .nr(1) |
| 3095 | .m(16) |
| 3096 | .n(1) |
| 3097 | .k(k) |
| 3098 | .sparsity(0.0f) |
| 3099 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3100 | } |
| 3101 | } |
| 3102 | |
| 3103 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, n_gt_1) { |
| 3104 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3105 | for (uint32_t n = 2; n < 10; n++) { |
| 3106 | for (size_t k = 1; k <= 5; k += 2) { |
| 3107 | SpMMMicrokernelTester() |
| 3108 | .mr(16) |
| 3109 | .nr(1) |
| 3110 | .m(16) |
| 3111 | .n(n) |
| 3112 | .k(k) |
| 3113 | .sparsity(0.0f) |
| 3114 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3115 | } |
| 3116 | } |
| 3117 | } |
| 3118 | |
| 3119 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, m_lt_16) { |
| 3120 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3121 | for (uint32_t m = 1; m < 16; m++) { |
| 3122 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3123 | for (size_t k = 1; k <= 5; k += 2) { |
| 3124 | SpMMMicrokernelTester() |
| 3125 | .mr(16) |
| 3126 | .nr(1) |
| 3127 | .m(m) |
| 3128 | .n(n) |
| 3129 | .k(k) |
| 3130 | .sparsity(0.0f) |
| 3131 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3132 | } |
| 3133 | } |
| 3134 | } |
| 3135 | } |
| 3136 | |
| 3137 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, m_div_16) { |
| 3138 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3139 | for (uint32_t m = 32; m <= 48; m += 16) { |
| 3140 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3141 | for (size_t k = 1; k <= 5; k += 2) { |
| 3142 | SpMMMicrokernelTester() |
| 3143 | .mr(16) |
| 3144 | .nr(1) |
| 3145 | .m(m) |
| 3146 | .n(n) |
| 3147 | .k(k) |
| 3148 | .sparsity(0.0f) |
| 3149 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3150 | } |
| 3151 | } |
| 3152 | } |
| 3153 | } |
| 3154 | |
| 3155 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, m_gt_16) { |
| 3156 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3157 | for (uint32_t m = 17; m < 32; m++) { |
| 3158 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3159 | for (size_t k = 1; k <= 5; k += 2) { |
| 3160 | SpMMMicrokernelTester() |
| 3161 | .mr(16) |
| 3162 | .nr(1) |
| 3163 | .m(m) |
| 3164 | .n(n) |
| 3165 | .k(k) |
| 3166 | .sparsity(0.0f) |
| 3167 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3168 | } |
| 3169 | } |
| 3170 | } |
| 3171 | } |
| 3172 | |
| 3173 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, qmin) { |
| 3174 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3175 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3176 | for (size_t k = 1; k <= 5; k += 2) { |
| 3177 | SpMMMicrokernelTester() |
| 3178 | .mr(16) |
| 3179 | .nr(1) |
| 3180 | .m(32) |
| 3181 | .n(n) |
| 3182 | .k(k) |
| 3183 | .sparsity(0.0f) |
| 3184 | .qmin(128) |
| 3185 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3186 | } |
| 3187 | } |
| 3188 | } |
| 3189 | |
| 3190 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, qmax) { |
| 3191 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3192 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3193 | for (size_t k = 1; k <= 5; k += 2) { |
| 3194 | SpMMMicrokernelTester() |
| 3195 | .mr(16) |
| 3196 | .nr(1) |
| 3197 | .m(32) |
| 3198 | .n(n) |
| 3199 | .k(k) |
| 3200 | .sparsity(0.0f) |
| 3201 | .qmax(128) |
| 3202 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3203 | } |
| 3204 | } |
| 3205 | } |
| 3206 | |
| 3207 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, half_sparse) { |
| 3208 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3209 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3210 | for (size_t k = 1; k <= 5; k += 2) { |
| 3211 | SpMMMicrokernelTester() |
| 3212 | .mr(16) |
| 3213 | .nr(1) |
| 3214 | .m(32) |
| 3215 | .n(n) |
| 3216 | .k(k) |
| 3217 | .sparsity(0.5f) |
| 3218 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3219 | } |
| 3220 | } |
| 3221 | } |
| 3222 | |
| 3223 | TEST(F32_SPMM_16X1__NEONFMA_PIPELINED, zero_weights) { |
| 3224 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3225 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3226 | for (size_t k = 1; k <= 5; k += 2) { |
| 3227 | SpMMMicrokernelTester() |
| 3228 | .mr(16) |
| 3229 | .nr(1) |
| 3230 | .m(32) |
| 3231 | .n(n) |
| 3232 | .k(k) |
| 3233 | .sparsity(1.0f) |
| 3234 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_pipelined); |
| 3235 | } |
| 3236 | } |
| 3237 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 3238 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 3239 | |
| 3240 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 3241 | #if XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 3242 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, k_eq_2) { |
| 3243 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3244 | SpMMMicrokernelTester() |
| 3245 | .mr(16) |
| 3246 | .nr(1) |
| 3247 | .m(16) |
| 3248 | .n(1) |
| 3249 | .k(2) |
| 3250 | .sparsity(0.0f) |
| 3251 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3252 | } |
| 3253 | |
| 3254 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, k_lt_2) { |
| 3255 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3256 | for (size_t k = 1; k < 2; k++) { |
| 3257 | SpMMMicrokernelTester() |
| 3258 | .mr(16) |
| 3259 | .nr(1) |
| 3260 | .m(16) |
| 3261 | .n(1) |
| 3262 | .k(k) |
| 3263 | .sparsity(0.0f) |
| 3264 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3265 | } |
| 3266 | } |
| 3267 | |
| 3268 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, k_gt_2) { |
| 3269 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3270 | for (size_t k = 3; k < 4; k++) { |
| 3271 | SpMMMicrokernelTester() |
| 3272 | .mr(16) |
| 3273 | .nr(1) |
| 3274 | .m(16) |
| 3275 | .n(1) |
| 3276 | .k(k) |
| 3277 | .sparsity(0.0f) |
| 3278 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3279 | } |
| 3280 | } |
| 3281 | |
| 3282 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, k_div_2) { |
| 3283 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3284 | for (size_t k = 4; k <= 20; k += 2) { |
| 3285 | SpMMMicrokernelTester() |
| 3286 | .mr(16) |
| 3287 | .nr(1) |
| 3288 | .m(16) |
| 3289 | .n(1) |
| 3290 | .k(k) |
| 3291 | .sparsity(0.0f) |
| 3292 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3293 | } |
| 3294 | } |
| 3295 | |
| 3296 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, n_gt_1) { |
| 3297 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3298 | for (uint32_t n = 2; n < 10; n++) { |
| 3299 | for (size_t k = 1; k <= 10; k += 3) { |
| 3300 | SpMMMicrokernelTester() |
| 3301 | .mr(16) |
| 3302 | .nr(1) |
| 3303 | .m(16) |
| 3304 | .n(n) |
| 3305 | .k(k) |
| 3306 | .sparsity(0.0f) |
| 3307 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3308 | } |
| 3309 | } |
| 3310 | } |
| 3311 | |
| 3312 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, m_lt_16) { |
| 3313 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3314 | for (uint32_t m = 1; m < 16; m++) { |
| 3315 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3316 | for (size_t k = 1; k <= 10; k += 3) { |
| 3317 | SpMMMicrokernelTester() |
| 3318 | .mr(16) |
| 3319 | .nr(1) |
| 3320 | .m(m) |
| 3321 | .n(n) |
| 3322 | .k(k) |
| 3323 | .sparsity(0.0f) |
| 3324 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3325 | } |
| 3326 | } |
| 3327 | } |
| 3328 | } |
| 3329 | |
| 3330 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, m_div_16) { |
| 3331 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3332 | for (uint32_t m = 32; m <= 48; m += 16) { |
| 3333 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3334 | for (size_t k = 1; k <= 10; k += 3) { |
| 3335 | SpMMMicrokernelTester() |
| 3336 | .mr(16) |
| 3337 | .nr(1) |
| 3338 | .m(m) |
| 3339 | .n(n) |
| 3340 | .k(k) |
| 3341 | .sparsity(0.0f) |
| 3342 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3343 | } |
| 3344 | } |
| 3345 | } |
| 3346 | } |
| 3347 | |
| 3348 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, m_gt_16) { |
| 3349 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3350 | for (uint32_t m = 17; m < 32; m++) { |
| 3351 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3352 | for (size_t k = 1; k <= 10; k += 3) { |
| 3353 | SpMMMicrokernelTester() |
| 3354 | .mr(16) |
| 3355 | .nr(1) |
| 3356 | .m(m) |
| 3357 | .n(n) |
| 3358 | .k(k) |
| 3359 | .sparsity(0.0f) |
| 3360 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3361 | } |
| 3362 | } |
| 3363 | } |
| 3364 | } |
| 3365 | |
| 3366 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, qmin) { |
| 3367 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3368 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3369 | for (size_t k = 1; k <= 10; k += 3) { |
| 3370 | SpMMMicrokernelTester() |
| 3371 | .mr(16) |
| 3372 | .nr(1) |
| 3373 | .m(32) |
| 3374 | .n(n) |
| 3375 | .k(k) |
| 3376 | .sparsity(0.0f) |
| 3377 | .qmin(128) |
| 3378 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3379 | } |
| 3380 | } |
| 3381 | } |
| 3382 | |
| 3383 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, qmax) { |
| 3384 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3385 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3386 | for (size_t k = 1; k <= 10; k += 3) { |
| 3387 | SpMMMicrokernelTester() |
| 3388 | .mr(16) |
| 3389 | .nr(1) |
| 3390 | .m(32) |
| 3391 | .n(n) |
| 3392 | .k(k) |
| 3393 | .sparsity(0.0f) |
| 3394 | .qmax(128) |
| 3395 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3396 | } |
| 3397 | } |
| 3398 | } |
| 3399 | |
| 3400 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, half_sparse) { |
| 3401 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3402 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3403 | for (size_t k = 1; k <= 10; k += 3) { |
| 3404 | SpMMMicrokernelTester() |
| 3405 | .mr(16) |
| 3406 | .nr(1) |
| 3407 | .m(32) |
| 3408 | .n(n) |
| 3409 | .k(k) |
| 3410 | .sparsity(0.5f) |
| 3411 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3412 | } |
| 3413 | } |
| 3414 | } |
| 3415 | |
| 3416 | TEST(F32_SPMM_16X1__NEONFMA_UNROLL2, zero_weights) { |
| 3417 | TEST_REQUIRES_ARM_NEON_FMA; |
| 3418 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3419 | for (size_t k = 1; k <= 10; k += 3) { |
| 3420 | SpMMMicrokernelTester() |
| 3421 | .mr(16) |
| 3422 | .nr(1) |
| 3423 | .m(32) |
| 3424 | .n(n) |
| 3425 | .k(k) |
| 3426 | .sparsity(1.0f) |
| 3427 | .Test(xnn_f32_spmm_ukernel_16x1__neonfma_unroll2); |
| 3428 | } |
| 3429 | } |
| 3430 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 3431 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 3432 | |
| 3433 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 3434 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 3435 | TEST(F32_SPMM_4X1__SSE, k_eq_1) { |
| 3436 | TEST_REQUIRES_X86_SSE; |
| 3437 | SpMMMicrokernelTester() |
| 3438 | .mr(4) |
| 3439 | .nr(1) |
| 3440 | .m(4) |
| 3441 | .n(1) |
| 3442 | .k(1) |
| 3443 | .sparsity(0.0f) |
| 3444 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3445 | } |
| 3446 | |
| 3447 | TEST(F32_SPMM_4X1__SSE, k_gt_1) { |
| 3448 | TEST_REQUIRES_X86_SSE; |
| 3449 | for (size_t k = 2; k < 10; k++) { |
| 3450 | SpMMMicrokernelTester() |
| 3451 | .mr(4) |
| 3452 | .nr(1) |
| 3453 | .m(4) |
| 3454 | .n(1) |
| 3455 | .k(k) |
| 3456 | .sparsity(0.0f) |
| 3457 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3458 | } |
| 3459 | } |
| 3460 | |
| 3461 | TEST(F32_SPMM_4X1__SSE, n_gt_1) { |
| 3462 | TEST_REQUIRES_X86_SSE; |
| 3463 | for (uint32_t n = 2; n < 10; n++) { |
| 3464 | for (size_t k = 1; k <= 5; k += 2) { |
| 3465 | SpMMMicrokernelTester() |
| 3466 | .mr(4) |
| 3467 | .nr(1) |
| 3468 | .m(4) |
| 3469 | .n(n) |
| 3470 | .k(k) |
| 3471 | .sparsity(0.0f) |
| 3472 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3473 | } |
| 3474 | } |
| 3475 | } |
| 3476 | |
| 3477 | TEST(F32_SPMM_4X1__SSE, m_lt_4) { |
| 3478 | TEST_REQUIRES_X86_SSE; |
| 3479 | for (uint32_t m = 1; m < 4; m++) { |
| 3480 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3481 | for (size_t k = 1; k <= 5; k += 2) { |
| 3482 | SpMMMicrokernelTester() |
| 3483 | .mr(4) |
| 3484 | .nr(1) |
| 3485 | .m(m) |
| 3486 | .n(n) |
| 3487 | .k(k) |
| 3488 | .sparsity(0.0f) |
| 3489 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3490 | } |
| 3491 | } |
| 3492 | } |
| 3493 | } |
| 3494 | |
| 3495 | TEST(F32_SPMM_4X1__SSE, m_div_4) { |
| 3496 | TEST_REQUIRES_X86_SSE; |
| 3497 | for (uint32_t m = 8; m <= 12; m += 4) { |
| 3498 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3499 | for (size_t k = 1; k <= 5; k += 2) { |
| 3500 | SpMMMicrokernelTester() |
| 3501 | .mr(4) |
| 3502 | .nr(1) |
| 3503 | .m(m) |
| 3504 | .n(n) |
| 3505 | .k(k) |
| 3506 | .sparsity(0.0f) |
| 3507 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3508 | } |
| 3509 | } |
| 3510 | } |
| 3511 | } |
| 3512 | |
| 3513 | TEST(F32_SPMM_4X1__SSE, m_gt_4) { |
| 3514 | TEST_REQUIRES_X86_SSE; |
| 3515 | for (uint32_t m = 5; m < 8; m++) { |
| 3516 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3517 | for (size_t k = 1; k <= 5; k += 2) { |
| 3518 | SpMMMicrokernelTester() |
| 3519 | .mr(4) |
| 3520 | .nr(1) |
| 3521 | .m(m) |
| 3522 | .n(n) |
| 3523 | .k(k) |
| 3524 | .sparsity(0.0f) |
| 3525 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3526 | } |
| 3527 | } |
| 3528 | } |
| 3529 | } |
| 3530 | |
| 3531 | TEST(F32_SPMM_4X1__SSE, qmin) { |
| 3532 | TEST_REQUIRES_X86_SSE; |
| 3533 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3534 | for (size_t k = 1; k <= 5; k += 2) { |
| 3535 | SpMMMicrokernelTester() |
| 3536 | .mr(4) |
| 3537 | .nr(1) |
| 3538 | .m(8) |
| 3539 | .n(n) |
| 3540 | .k(k) |
| 3541 | .sparsity(0.0f) |
| 3542 | .qmin(128) |
| 3543 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3544 | } |
| 3545 | } |
| 3546 | } |
| 3547 | |
| 3548 | TEST(F32_SPMM_4X1__SSE, qmax) { |
| 3549 | TEST_REQUIRES_X86_SSE; |
| 3550 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3551 | for (size_t k = 1; k <= 5; k += 2) { |
| 3552 | SpMMMicrokernelTester() |
| 3553 | .mr(4) |
| 3554 | .nr(1) |
| 3555 | .m(8) |
| 3556 | .n(n) |
| 3557 | .k(k) |
| 3558 | .sparsity(0.0f) |
| 3559 | .qmax(128) |
| 3560 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3561 | } |
| 3562 | } |
| 3563 | } |
| 3564 | |
| 3565 | TEST(F32_SPMM_4X1__SSE, half_sparse) { |
| 3566 | TEST_REQUIRES_X86_SSE; |
| 3567 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3568 | for (size_t k = 1; k <= 5; k += 2) { |
| 3569 | SpMMMicrokernelTester() |
| 3570 | .mr(4) |
| 3571 | .nr(1) |
| 3572 | .m(8) |
| 3573 | .n(n) |
| 3574 | .k(k) |
| 3575 | .sparsity(0.5f) |
| 3576 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3577 | } |
| 3578 | } |
| 3579 | } |
| 3580 | |
| 3581 | TEST(F32_SPMM_4X1__SSE, zero_weights) { |
| 3582 | TEST_REQUIRES_X86_SSE; |
| 3583 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3584 | for (size_t k = 1; k <= 5; k += 2) { |
| 3585 | SpMMMicrokernelTester() |
| 3586 | .mr(4) |
| 3587 | .nr(1) |
| 3588 | .m(8) |
| 3589 | .n(n) |
| 3590 | .k(k) |
| 3591 | .sparsity(1.0f) |
| 3592 | .Test(xnn_f32_spmm_ukernel_4x1__sse); |
| 3593 | } |
| 3594 | } |
| 3595 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 3596 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 3597 | |
| 3598 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 3599 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 3600 | TEST(F32_SPMM_8X1__SSE, k_eq_1) { |
| 3601 | TEST_REQUIRES_X86_SSE; |
| 3602 | SpMMMicrokernelTester() |
| 3603 | .mr(8) |
| 3604 | .nr(1) |
| 3605 | .m(8) |
| 3606 | .n(1) |
| 3607 | .k(1) |
| 3608 | .sparsity(0.0f) |
| 3609 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3610 | } |
| 3611 | |
| 3612 | TEST(F32_SPMM_8X1__SSE, k_gt_1) { |
| 3613 | TEST_REQUIRES_X86_SSE; |
| 3614 | for (size_t k = 2; k < 10; k++) { |
| 3615 | SpMMMicrokernelTester() |
| 3616 | .mr(8) |
| 3617 | .nr(1) |
| 3618 | .m(8) |
| 3619 | .n(1) |
| 3620 | .k(k) |
| 3621 | .sparsity(0.0f) |
| 3622 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3623 | } |
| 3624 | } |
| 3625 | |
| 3626 | TEST(F32_SPMM_8X1__SSE, n_gt_1) { |
| 3627 | TEST_REQUIRES_X86_SSE; |
| 3628 | for (uint32_t n = 2; n < 10; n++) { |
| 3629 | for (size_t k = 1; k <= 5; k += 2) { |
| 3630 | SpMMMicrokernelTester() |
| 3631 | .mr(8) |
| 3632 | .nr(1) |
| 3633 | .m(8) |
| 3634 | .n(n) |
| 3635 | .k(k) |
| 3636 | .sparsity(0.0f) |
| 3637 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3638 | } |
| 3639 | } |
| 3640 | } |
| 3641 | |
| 3642 | TEST(F32_SPMM_8X1__SSE, m_lt_8) { |
| 3643 | TEST_REQUIRES_X86_SSE; |
| 3644 | for (uint32_t m = 1; m < 8; m++) { |
| 3645 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3646 | for (size_t k = 1; k <= 5; k += 2) { |
| 3647 | SpMMMicrokernelTester() |
| 3648 | .mr(8) |
| 3649 | .nr(1) |
| 3650 | .m(m) |
| 3651 | .n(n) |
| 3652 | .k(k) |
| 3653 | .sparsity(0.0f) |
| 3654 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3655 | } |
| 3656 | } |
| 3657 | } |
| 3658 | } |
| 3659 | |
| 3660 | TEST(F32_SPMM_8X1__SSE, m_div_8) { |
| 3661 | TEST_REQUIRES_X86_SSE; |
| 3662 | for (uint32_t m = 16; m <= 24; m += 8) { |
| 3663 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3664 | for (size_t k = 1; k <= 5; k += 2) { |
| 3665 | SpMMMicrokernelTester() |
| 3666 | .mr(8) |
| 3667 | .nr(1) |
| 3668 | .m(m) |
| 3669 | .n(n) |
| 3670 | .k(k) |
| 3671 | .sparsity(0.0f) |
| 3672 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3673 | } |
| 3674 | } |
| 3675 | } |
| 3676 | } |
| 3677 | |
| 3678 | TEST(F32_SPMM_8X1__SSE, m_gt_8) { |
| 3679 | TEST_REQUIRES_X86_SSE; |
| 3680 | for (uint32_t m = 9; m < 16; m++) { |
| 3681 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3682 | for (size_t k = 1; k <= 5; k += 2) { |
| 3683 | SpMMMicrokernelTester() |
| 3684 | .mr(8) |
| 3685 | .nr(1) |
| 3686 | .m(m) |
| 3687 | .n(n) |
| 3688 | .k(k) |
| 3689 | .sparsity(0.0f) |
| 3690 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3691 | } |
| 3692 | } |
| 3693 | } |
| 3694 | } |
| 3695 | |
| 3696 | TEST(F32_SPMM_8X1__SSE, qmin) { |
| 3697 | TEST_REQUIRES_X86_SSE; |
| 3698 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3699 | for (size_t k = 1; k <= 5; k += 2) { |
| 3700 | SpMMMicrokernelTester() |
| 3701 | .mr(8) |
| 3702 | .nr(1) |
| 3703 | .m(16) |
| 3704 | .n(n) |
| 3705 | .k(k) |
| 3706 | .sparsity(0.0f) |
| 3707 | .qmin(128) |
| 3708 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3709 | } |
| 3710 | } |
| 3711 | } |
| 3712 | |
| 3713 | TEST(F32_SPMM_8X1__SSE, qmax) { |
| 3714 | TEST_REQUIRES_X86_SSE; |
| 3715 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3716 | for (size_t k = 1; k <= 5; k += 2) { |
| 3717 | SpMMMicrokernelTester() |
| 3718 | .mr(8) |
| 3719 | .nr(1) |
| 3720 | .m(16) |
| 3721 | .n(n) |
| 3722 | .k(k) |
| 3723 | .sparsity(0.0f) |
| 3724 | .qmax(128) |
| 3725 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3726 | } |
| 3727 | } |
| 3728 | } |
| 3729 | |
| 3730 | TEST(F32_SPMM_8X1__SSE, half_sparse) { |
| 3731 | TEST_REQUIRES_X86_SSE; |
| 3732 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3733 | for (size_t k = 1; k <= 5; k += 2) { |
| 3734 | SpMMMicrokernelTester() |
| 3735 | .mr(8) |
| 3736 | .nr(1) |
| 3737 | .m(16) |
| 3738 | .n(n) |
| 3739 | .k(k) |
| 3740 | .sparsity(0.5f) |
| 3741 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3742 | } |
| 3743 | } |
| 3744 | } |
| 3745 | |
| 3746 | TEST(F32_SPMM_8X1__SSE, zero_weights) { |
| 3747 | TEST_REQUIRES_X86_SSE; |
| 3748 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3749 | for (size_t k = 1; k <= 5; k += 2) { |
| 3750 | SpMMMicrokernelTester() |
| 3751 | .mr(8) |
| 3752 | .nr(1) |
| 3753 | .m(16) |
| 3754 | .n(n) |
| 3755 | .k(k) |
| 3756 | .sparsity(1.0f) |
| 3757 | .Test(xnn_f32_spmm_ukernel_8x1__sse); |
| 3758 | } |
| 3759 | } |
| 3760 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 3761 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 3762 | |
| 3763 | |
| 3764 | TEST(F32_SPMM_1X1__SCALAR, k_eq_1) { |
| 3765 | SpMMMicrokernelTester() |
| 3766 | .mr(1) |
| 3767 | .nr(1) |
| 3768 | .m(1) |
| 3769 | .n(1) |
| 3770 | .k(1) |
| 3771 | .sparsity(0.0f) |
| 3772 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3773 | } |
| 3774 | |
| 3775 | TEST(F32_SPMM_1X1__SCALAR, k_gt_1) { |
| 3776 | for (size_t k = 2; k < 10; k++) { |
| 3777 | SpMMMicrokernelTester() |
| 3778 | .mr(1) |
| 3779 | .nr(1) |
| 3780 | .m(1) |
| 3781 | .n(1) |
| 3782 | .k(k) |
| 3783 | .sparsity(0.0f) |
| 3784 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3785 | } |
| 3786 | } |
| 3787 | |
| 3788 | TEST(F32_SPMM_1X1__SCALAR, n_gt_1) { |
| 3789 | for (uint32_t n = 2; n < 10; n++) { |
| 3790 | for (size_t k = 1; k <= 5; k += 2) { |
| 3791 | SpMMMicrokernelTester() |
| 3792 | .mr(1) |
| 3793 | .nr(1) |
| 3794 | .m(1) |
| 3795 | .n(n) |
| 3796 | .k(k) |
| 3797 | .sparsity(0.0f) |
| 3798 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3799 | } |
| 3800 | } |
| 3801 | } |
| 3802 | |
| 3803 | TEST(F32_SPMM_1X1__SCALAR, m_lt_1) { |
| 3804 | for (uint32_t m = 1; m < 1; m++) { |
| 3805 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3806 | for (size_t k = 1; k <= 5; k += 2) { |
| 3807 | SpMMMicrokernelTester() |
| 3808 | .mr(1) |
| 3809 | .nr(1) |
| 3810 | .m(m) |
| 3811 | .n(n) |
| 3812 | .k(k) |
| 3813 | .sparsity(0.0f) |
| 3814 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3815 | } |
| 3816 | } |
| 3817 | } |
| 3818 | } |
| 3819 | |
| 3820 | TEST(F32_SPMM_1X1__SCALAR, m_div_1) { |
| 3821 | for (uint32_t m = 2; m <= 3; m += 1) { |
| 3822 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3823 | for (size_t k = 1; k <= 5; k += 2) { |
| 3824 | SpMMMicrokernelTester() |
| 3825 | .mr(1) |
| 3826 | .nr(1) |
| 3827 | .m(m) |
| 3828 | .n(n) |
| 3829 | .k(k) |
| 3830 | .sparsity(0.0f) |
| 3831 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3832 | } |
| 3833 | } |
| 3834 | } |
| 3835 | } |
| 3836 | |
| 3837 | TEST(F32_SPMM_1X1__SCALAR, m_gt_1) { |
| 3838 | for (uint32_t m = 2; m < 2; m++) { |
| 3839 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3840 | for (size_t k = 1; k <= 5; k += 2) { |
| 3841 | SpMMMicrokernelTester() |
| 3842 | .mr(1) |
| 3843 | .nr(1) |
| 3844 | .m(m) |
| 3845 | .n(n) |
| 3846 | .k(k) |
| 3847 | .sparsity(0.0f) |
| 3848 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3849 | } |
| 3850 | } |
| 3851 | } |
| 3852 | } |
| 3853 | |
| 3854 | TEST(F32_SPMM_1X1__SCALAR, qmin) { |
| 3855 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3856 | for (size_t k = 1; k <= 5; k += 2) { |
| 3857 | SpMMMicrokernelTester() |
| 3858 | .mr(1) |
| 3859 | .nr(1) |
| 3860 | .m(2) |
| 3861 | .n(n) |
| 3862 | .k(k) |
| 3863 | .sparsity(0.0f) |
| 3864 | .qmin(128) |
| 3865 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3866 | } |
| 3867 | } |
| 3868 | } |
| 3869 | |
| 3870 | TEST(F32_SPMM_1X1__SCALAR, qmax) { |
| 3871 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3872 | for (size_t k = 1; k <= 5; k += 2) { |
| 3873 | SpMMMicrokernelTester() |
| 3874 | .mr(1) |
| 3875 | .nr(1) |
| 3876 | .m(2) |
| 3877 | .n(n) |
| 3878 | .k(k) |
| 3879 | .sparsity(0.0f) |
| 3880 | .qmax(128) |
| 3881 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3882 | } |
| 3883 | } |
| 3884 | } |
| 3885 | |
| 3886 | TEST(F32_SPMM_1X1__SCALAR, half_sparse) { |
| 3887 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3888 | for (size_t k = 1; k <= 5; k += 2) { |
| 3889 | SpMMMicrokernelTester() |
| 3890 | .mr(1) |
| 3891 | .nr(1) |
| 3892 | .m(2) |
| 3893 | .n(n) |
| 3894 | .k(k) |
| 3895 | .sparsity(0.5f) |
| 3896 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3897 | } |
| 3898 | } |
| 3899 | } |
| 3900 | |
| 3901 | TEST(F32_SPMM_1X1__SCALAR, zero_weights) { |
| 3902 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3903 | for (size_t k = 1; k <= 5; k += 2) { |
| 3904 | SpMMMicrokernelTester() |
| 3905 | .mr(1) |
| 3906 | .nr(1) |
| 3907 | .m(2) |
| 3908 | .n(n) |
| 3909 | .k(k) |
| 3910 | .sparsity(1.0f) |
| 3911 | .Test(xnn_f32_spmm_ukernel_1x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 3912 | } |
| 3913 | } |
| 3914 | } |
| 3915 | |
| 3916 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, k_eq_1) { |
| 3917 | SpMMMicrokernelTester() |
| 3918 | .mr(1) |
| 3919 | .nr(1) |
| 3920 | .m(1) |
| 3921 | .n(1) |
| 3922 | .k(1) |
| 3923 | .sparsity(0.0f) |
| 3924 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 3925 | } |
| 3926 | |
| 3927 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, k_gt_1) { |
| 3928 | for (size_t k = 2; k < 10; k++) { |
| 3929 | SpMMMicrokernelTester() |
| 3930 | .mr(1) |
| 3931 | .nr(1) |
| 3932 | .m(1) |
| 3933 | .n(1) |
| 3934 | .k(k) |
| 3935 | .sparsity(0.0f) |
| 3936 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 3937 | } |
| 3938 | } |
| 3939 | |
| 3940 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, n_gt_1) { |
| 3941 | for (uint32_t n = 2; n < 10; n++) { |
| 3942 | for (size_t k = 1; k <= 5; k += 2) { |
| 3943 | SpMMMicrokernelTester() |
| 3944 | .mr(1) |
| 3945 | .nr(1) |
| 3946 | .m(1) |
| 3947 | .n(n) |
| 3948 | .k(k) |
| 3949 | .sparsity(0.0f) |
| 3950 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 3951 | } |
| 3952 | } |
| 3953 | } |
| 3954 | |
| 3955 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, m_lt_1) { |
| 3956 | for (uint32_t m = 1; m < 1; m++) { |
| 3957 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3958 | for (size_t k = 1; k <= 5; k += 2) { |
| 3959 | SpMMMicrokernelTester() |
| 3960 | .mr(1) |
| 3961 | .nr(1) |
| 3962 | .m(m) |
| 3963 | .n(n) |
| 3964 | .k(k) |
| 3965 | .sparsity(0.0f) |
| 3966 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 3967 | } |
| 3968 | } |
| 3969 | } |
| 3970 | } |
| 3971 | |
| 3972 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, m_div_1) { |
| 3973 | for (uint32_t m = 2; m <= 3; m += 1) { |
| 3974 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3975 | for (size_t k = 1; k <= 5; k += 2) { |
| 3976 | SpMMMicrokernelTester() |
| 3977 | .mr(1) |
| 3978 | .nr(1) |
| 3979 | .m(m) |
| 3980 | .n(n) |
| 3981 | .k(k) |
| 3982 | .sparsity(0.0f) |
| 3983 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 3984 | } |
| 3985 | } |
| 3986 | } |
| 3987 | } |
| 3988 | |
| 3989 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, m_gt_1) { |
| 3990 | for (uint32_t m = 2; m < 2; m++) { |
| 3991 | for (uint32_t n = 1; n < 10; n += 2) { |
| 3992 | for (size_t k = 1; k <= 5; k += 2) { |
| 3993 | SpMMMicrokernelTester() |
| 3994 | .mr(1) |
| 3995 | .nr(1) |
| 3996 | .m(m) |
| 3997 | .n(n) |
| 3998 | .k(k) |
| 3999 | .sparsity(0.0f) |
| 4000 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4001 | } |
| 4002 | } |
| 4003 | } |
| 4004 | } |
| 4005 | |
| 4006 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, qmin) { |
| 4007 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4008 | for (size_t k = 1; k <= 5; k += 2) { |
| 4009 | SpMMMicrokernelTester() |
| 4010 | .mr(1) |
| 4011 | .nr(1) |
| 4012 | .m(2) |
| 4013 | .n(n) |
| 4014 | .k(k) |
| 4015 | .sparsity(0.0f) |
| 4016 | .qmin(128) |
| 4017 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4018 | } |
| 4019 | } |
| 4020 | } |
| 4021 | |
| 4022 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, qmax) { |
| 4023 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4024 | for (size_t k = 1; k <= 5; k += 2) { |
| 4025 | SpMMMicrokernelTester() |
| 4026 | .mr(1) |
| 4027 | .nr(1) |
| 4028 | .m(2) |
| 4029 | .n(n) |
| 4030 | .k(k) |
| 4031 | .sparsity(0.0f) |
| 4032 | .qmax(128) |
| 4033 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4034 | } |
| 4035 | } |
| 4036 | } |
| 4037 | |
| 4038 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, half_sparse) { |
| 4039 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4040 | for (size_t k = 1; k <= 5; k += 2) { |
| 4041 | SpMMMicrokernelTester() |
| 4042 | .mr(1) |
| 4043 | .nr(1) |
| 4044 | .m(2) |
| 4045 | .n(n) |
| 4046 | .k(k) |
| 4047 | .sparsity(0.5f) |
| 4048 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4049 | } |
| 4050 | } |
| 4051 | } |
| 4052 | |
| 4053 | TEST(F32_SPMM_1X1__SCALAR_PIPELINED, zero_weights) { |
| 4054 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4055 | for (size_t k = 1; k <= 5; k += 2) { |
| 4056 | SpMMMicrokernelTester() |
| 4057 | .mr(1) |
| 4058 | .nr(1) |
| 4059 | .m(2) |
| 4060 | .n(n) |
| 4061 | .k(k) |
| 4062 | .sparsity(1.0f) |
| 4063 | .Test(xnn_f32_spmm_ukernel_1x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4064 | } |
| 4065 | } |
| 4066 | } |
| 4067 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4068 | TEST(F32_SPMM_2X1__SCALAR, k_eq_1) { |
| 4069 | SpMMMicrokernelTester() |
| 4070 | .mr(2) |
| 4071 | .nr(1) |
| 4072 | .m(2) |
| 4073 | .n(1) |
| 4074 | .k(1) |
| 4075 | .sparsity(0.0f) |
| 4076 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4077 | } |
| 4078 | |
| 4079 | TEST(F32_SPMM_2X1__SCALAR, k_gt_1) { |
| 4080 | for (size_t k = 2; k < 10; k++) { |
| 4081 | SpMMMicrokernelTester() |
| 4082 | .mr(2) |
| 4083 | .nr(1) |
| 4084 | .m(2) |
| 4085 | .n(1) |
| 4086 | .k(k) |
| 4087 | .sparsity(0.0f) |
| 4088 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4089 | } |
| 4090 | } |
| 4091 | |
| 4092 | TEST(F32_SPMM_2X1__SCALAR, n_gt_1) { |
| 4093 | for (uint32_t n = 2; n < 10; n++) { |
| 4094 | for (size_t k = 1; k <= 5; k += 2) { |
| 4095 | SpMMMicrokernelTester() |
| 4096 | .mr(2) |
| 4097 | .nr(1) |
| 4098 | .m(2) |
| 4099 | .n(n) |
| 4100 | .k(k) |
| 4101 | .sparsity(0.0f) |
| 4102 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4103 | } |
| 4104 | } |
| 4105 | } |
| 4106 | |
| 4107 | TEST(F32_SPMM_2X1__SCALAR, m_lt_2) { |
| 4108 | for (uint32_t m = 1; m < 2; m++) { |
| 4109 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4110 | for (size_t k = 1; k <= 5; k += 2) { |
| 4111 | SpMMMicrokernelTester() |
| 4112 | .mr(2) |
| 4113 | .nr(1) |
| 4114 | .m(m) |
| 4115 | .n(n) |
| 4116 | .k(k) |
| 4117 | .sparsity(0.0f) |
| 4118 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4119 | } |
| 4120 | } |
| 4121 | } |
| 4122 | } |
| 4123 | |
| 4124 | TEST(F32_SPMM_2X1__SCALAR, m_div_2) { |
| 4125 | for (uint32_t m = 4; m <= 6; m += 2) { |
| 4126 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4127 | for (size_t k = 1; k <= 5; k += 2) { |
| 4128 | SpMMMicrokernelTester() |
| 4129 | .mr(2) |
| 4130 | .nr(1) |
| 4131 | .m(m) |
| 4132 | .n(n) |
| 4133 | .k(k) |
| 4134 | .sparsity(0.0f) |
| 4135 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4136 | } |
| 4137 | } |
| 4138 | } |
| 4139 | } |
| 4140 | |
| 4141 | TEST(F32_SPMM_2X1__SCALAR, m_gt_2) { |
| 4142 | for (uint32_t m = 3; m < 4; m++) { |
| 4143 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4144 | for (size_t k = 1; k <= 5; k += 2) { |
| 4145 | SpMMMicrokernelTester() |
| 4146 | .mr(2) |
| 4147 | .nr(1) |
| 4148 | .m(m) |
| 4149 | .n(n) |
| 4150 | .k(k) |
| 4151 | .sparsity(0.0f) |
| 4152 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4153 | } |
| 4154 | } |
| 4155 | } |
| 4156 | } |
| 4157 | |
| 4158 | TEST(F32_SPMM_2X1__SCALAR, qmin) { |
| 4159 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4160 | for (size_t k = 1; k <= 5; k += 2) { |
| 4161 | SpMMMicrokernelTester() |
| 4162 | .mr(2) |
| 4163 | .nr(1) |
| 4164 | .m(4) |
| 4165 | .n(n) |
| 4166 | .k(k) |
| 4167 | .sparsity(0.0f) |
| 4168 | .qmin(128) |
| 4169 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4170 | } |
| 4171 | } |
| 4172 | } |
| 4173 | |
| 4174 | TEST(F32_SPMM_2X1__SCALAR, qmax) { |
| 4175 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4176 | for (size_t k = 1; k <= 5; k += 2) { |
| 4177 | SpMMMicrokernelTester() |
| 4178 | .mr(2) |
| 4179 | .nr(1) |
| 4180 | .m(4) |
| 4181 | .n(n) |
| 4182 | .k(k) |
| 4183 | .sparsity(0.0f) |
| 4184 | .qmax(128) |
| 4185 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4186 | } |
| 4187 | } |
| 4188 | } |
| 4189 | |
| 4190 | TEST(F32_SPMM_2X1__SCALAR, half_sparse) { |
| 4191 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4192 | for (size_t k = 1; k <= 5; k += 2) { |
| 4193 | SpMMMicrokernelTester() |
| 4194 | .mr(2) |
| 4195 | .nr(1) |
| 4196 | .m(4) |
| 4197 | .n(n) |
| 4198 | .k(k) |
| 4199 | .sparsity(0.5f) |
| 4200 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4201 | } |
| 4202 | } |
| 4203 | } |
| 4204 | |
| 4205 | TEST(F32_SPMM_2X1__SCALAR, zero_weights) { |
| 4206 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4207 | for (size_t k = 1; k <= 5; k += 2) { |
| 4208 | SpMMMicrokernelTester() |
| 4209 | .mr(2) |
| 4210 | .nr(1) |
| 4211 | .m(4) |
| 4212 | .n(n) |
| 4213 | .k(k) |
| 4214 | .sparsity(1.0f) |
| 4215 | .Test(xnn_f32_spmm_ukernel_2x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4216 | } |
| 4217 | } |
| 4218 | } |
| 4219 | |
| 4220 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, k_eq_1) { |
| 4221 | SpMMMicrokernelTester() |
| 4222 | .mr(2) |
| 4223 | .nr(1) |
| 4224 | .m(2) |
| 4225 | .n(1) |
| 4226 | .k(1) |
| 4227 | .sparsity(0.0f) |
| 4228 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4229 | } |
| 4230 | |
| 4231 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, k_gt_1) { |
| 4232 | for (size_t k = 2; k < 10; k++) { |
| 4233 | SpMMMicrokernelTester() |
| 4234 | .mr(2) |
| 4235 | .nr(1) |
| 4236 | .m(2) |
| 4237 | .n(1) |
| 4238 | .k(k) |
| 4239 | .sparsity(0.0f) |
| 4240 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4241 | } |
| 4242 | } |
| 4243 | |
| 4244 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, n_gt_1) { |
| 4245 | for (uint32_t n = 2; n < 10; n++) { |
| 4246 | for (size_t k = 1; k <= 5; k += 2) { |
| 4247 | SpMMMicrokernelTester() |
| 4248 | .mr(2) |
| 4249 | .nr(1) |
| 4250 | .m(2) |
| 4251 | .n(n) |
| 4252 | .k(k) |
| 4253 | .sparsity(0.0f) |
| 4254 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4255 | } |
| 4256 | } |
| 4257 | } |
| 4258 | |
| 4259 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, m_lt_2) { |
| 4260 | for (uint32_t m = 1; m < 2; m++) { |
| 4261 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4262 | for (size_t k = 1; k <= 5; k += 2) { |
| 4263 | SpMMMicrokernelTester() |
| 4264 | .mr(2) |
| 4265 | .nr(1) |
| 4266 | .m(m) |
| 4267 | .n(n) |
| 4268 | .k(k) |
| 4269 | .sparsity(0.0f) |
| 4270 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4271 | } |
| 4272 | } |
| 4273 | } |
| 4274 | } |
| 4275 | |
| 4276 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, m_div_2) { |
| 4277 | for (uint32_t m = 4; m <= 6; m += 2) { |
| 4278 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4279 | for (size_t k = 1; k <= 5; k += 2) { |
| 4280 | SpMMMicrokernelTester() |
| 4281 | .mr(2) |
| 4282 | .nr(1) |
| 4283 | .m(m) |
| 4284 | .n(n) |
| 4285 | .k(k) |
| 4286 | .sparsity(0.0f) |
| 4287 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4288 | } |
| 4289 | } |
| 4290 | } |
| 4291 | } |
| 4292 | |
| 4293 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, m_gt_2) { |
| 4294 | for (uint32_t m = 3; m < 4; m++) { |
| 4295 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4296 | for (size_t k = 1; k <= 5; k += 2) { |
| 4297 | SpMMMicrokernelTester() |
| 4298 | .mr(2) |
| 4299 | .nr(1) |
| 4300 | .m(m) |
| 4301 | .n(n) |
| 4302 | .k(k) |
| 4303 | .sparsity(0.0f) |
| 4304 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4305 | } |
| 4306 | } |
| 4307 | } |
| 4308 | } |
| 4309 | |
| 4310 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, qmin) { |
| 4311 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4312 | for (size_t k = 1; k <= 5; k += 2) { |
| 4313 | SpMMMicrokernelTester() |
| 4314 | .mr(2) |
| 4315 | .nr(1) |
| 4316 | .m(4) |
| 4317 | .n(n) |
| 4318 | .k(k) |
| 4319 | .sparsity(0.0f) |
| 4320 | .qmin(128) |
| 4321 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4322 | } |
| 4323 | } |
| 4324 | } |
| 4325 | |
| 4326 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, qmax) { |
| 4327 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4328 | for (size_t k = 1; k <= 5; k += 2) { |
| 4329 | SpMMMicrokernelTester() |
| 4330 | .mr(2) |
| 4331 | .nr(1) |
| 4332 | .m(4) |
| 4333 | .n(n) |
| 4334 | .k(k) |
| 4335 | .sparsity(0.0f) |
| 4336 | .qmax(128) |
| 4337 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4338 | } |
| 4339 | } |
| 4340 | } |
| 4341 | |
| 4342 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, half_sparse) { |
| 4343 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4344 | for (size_t k = 1; k <= 5; k += 2) { |
| 4345 | SpMMMicrokernelTester() |
| 4346 | .mr(2) |
| 4347 | .nr(1) |
| 4348 | .m(4) |
| 4349 | .n(n) |
| 4350 | .k(k) |
| 4351 | .sparsity(0.5f) |
| 4352 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4353 | } |
| 4354 | } |
| 4355 | } |
| 4356 | |
| 4357 | TEST(F32_SPMM_2X1__SCALAR_PIPELINED, zero_weights) { |
| 4358 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4359 | for (size_t k = 1; k <= 5; k += 2) { |
| 4360 | SpMMMicrokernelTester() |
| 4361 | .mr(2) |
| 4362 | .nr(1) |
| 4363 | .m(4) |
| 4364 | .n(n) |
| 4365 | .k(k) |
| 4366 | .sparsity(1.0f) |
| 4367 | .Test(xnn_f32_spmm_ukernel_2x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4368 | } |
| 4369 | } |
| 4370 | } |
| 4371 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4372 | TEST(F32_SPMM_4X1__SCALAR, k_eq_1) { |
| 4373 | SpMMMicrokernelTester() |
| 4374 | .mr(4) |
| 4375 | .nr(1) |
| 4376 | .m(4) |
| 4377 | .n(1) |
| 4378 | .k(1) |
| 4379 | .sparsity(0.0f) |
| 4380 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4381 | } |
| 4382 | |
| 4383 | TEST(F32_SPMM_4X1__SCALAR, k_gt_1) { |
| 4384 | for (size_t k = 2; k < 10; k++) { |
| 4385 | SpMMMicrokernelTester() |
| 4386 | .mr(4) |
| 4387 | .nr(1) |
| 4388 | .m(4) |
| 4389 | .n(1) |
| 4390 | .k(k) |
| 4391 | .sparsity(0.0f) |
| 4392 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4393 | } |
| 4394 | } |
| 4395 | |
| 4396 | TEST(F32_SPMM_4X1__SCALAR, n_gt_1) { |
| 4397 | for (uint32_t n = 2; n < 10; n++) { |
| 4398 | for (size_t k = 1; k <= 5; k += 2) { |
| 4399 | SpMMMicrokernelTester() |
| 4400 | .mr(4) |
| 4401 | .nr(1) |
| 4402 | .m(4) |
| 4403 | .n(n) |
| 4404 | .k(k) |
| 4405 | .sparsity(0.0f) |
| 4406 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4407 | } |
| 4408 | } |
| 4409 | } |
| 4410 | |
| 4411 | TEST(F32_SPMM_4X1__SCALAR, m_lt_4) { |
| 4412 | for (uint32_t m = 1; m < 4; m++) { |
| 4413 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4414 | for (size_t k = 1; k <= 5; k += 2) { |
| 4415 | SpMMMicrokernelTester() |
| 4416 | .mr(4) |
| 4417 | .nr(1) |
| 4418 | .m(m) |
| 4419 | .n(n) |
| 4420 | .k(k) |
| 4421 | .sparsity(0.0f) |
| 4422 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4423 | } |
| 4424 | } |
| 4425 | } |
| 4426 | } |
| 4427 | |
| 4428 | TEST(F32_SPMM_4X1__SCALAR, m_div_4) { |
| 4429 | for (uint32_t m = 8; m <= 12; m += 4) { |
| 4430 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4431 | for (size_t k = 1; k <= 5; k += 2) { |
| 4432 | SpMMMicrokernelTester() |
| 4433 | .mr(4) |
| 4434 | .nr(1) |
| 4435 | .m(m) |
| 4436 | .n(n) |
| 4437 | .k(k) |
| 4438 | .sparsity(0.0f) |
| 4439 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4440 | } |
| 4441 | } |
| 4442 | } |
| 4443 | } |
| 4444 | |
| 4445 | TEST(F32_SPMM_4X1__SCALAR, m_gt_4) { |
| 4446 | for (uint32_t m = 5; m < 8; m++) { |
| 4447 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4448 | for (size_t k = 1; k <= 5; k += 2) { |
| 4449 | SpMMMicrokernelTester() |
| 4450 | .mr(4) |
| 4451 | .nr(1) |
| 4452 | .m(m) |
| 4453 | .n(n) |
| 4454 | .k(k) |
| 4455 | .sparsity(0.0f) |
| 4456 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4457 | } |
| 4458 | } |
| 4459 | } |
| 4460 | } |
| 4461 | |
| 4462 | TEST(F32_SPMM_4X1__SCALAR, qmin) { |
| 4463 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4464 | for (size_t k = 1; k <= 5; k += 2) { |
| 4465 | SpMMMicrokernelTester() |
| 4466 | .mr(4) |
| 4467 | .nr(1) |
| 4468 | .m(8) |
| 4469 | .n(n) |
| 4470 | .k(k) |
| 4471 | .sparsity(0.0f) |
| 4472 | .qmin(128) |
| 4473 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4474 | } |
| 4475 | } |
| 4476 | } |
| 4477 | |
| 4478 | TEST(F32_SPMM_4X1__SCALAR, qmax) { |
| 4479 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4480 | for (size_t k = 1; k <= 5; k += 2) { |
| 4481 | SpMMMicrokernelTester() |
| 4482 | .mr(4) |
| 4483 | .nr(1) |
| 4484 | .m(8) |
| 4485 | .n(n) |
| 4486 | .k(k) |
| 4487 | .sparsity(0.0f) |
| 4488 | .qmax(128) |
| 4489 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4490 | } |
| 4491 | } |
| 4492 | } |
| 4493 | |
| 4494 | TEST(F32_SPMM_4X1__SCALAR, half_sparse) { |
| 4495 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4496 | for (size_t k = 1; k <= 5; k += 2) { |
| 4497 | SpMMMicrokernelTester() |
| 4498 | .mr(4) |
| 4499 | .nr(1) |
| 4500 | .m(8) |
| 4501 | .n(n) |
| 4502 | .k(k) |
| 4503 | .sparsity(0.5f) |
| 4504 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4505 | } |
| 4506 | } |
| 4507 | } |
| 4508 | |
| 4509 | TEST(F32_SPMM_4X1__SCALAR, zero_weights) { |
| 4510 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4511 | for (size_t k = 1; k <= 5; k += 2) { |
| 4512 | SpMMMicrokernelTester() |
| 4513 | .mr(4) |
| 4514 | .nr(1) |
| 4515 | .m(8) |
| 4516 | .n(n) |
| 4517 | .k(k) |
| 4518 | .sparsity(1.0f) |
| 4519 | .Test(xnn_f32_spmm_ukernel_4x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4520 | } |
| 4521 | } |
| 4522 | } |
| 4523 | |
| 4524 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, k_eq_1) { |
| 4525 | SpMMMicrokernelTester() |
| 4526 | .mr(4) |
| 4527 | .nr(1) |
| 4528 | .m(4) |
| 4529 | .n(1) |
| 4530 | .k(1) |
| 4531 | .sparsity(0.0f) |
| 4532 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4533 | } |
| 4534 | |
| 4535 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, k_gt_1) { |
| 4536 | for (size_t k = 2; k < 10; k++) { |
| 4537 | SpMMMicrokernelTester() |
| 4538 | .mr(4) |
| 4539 | .nr(1) |
| 4540 | .m(4) |
| 4541 | .n(1) |
| 4542 | .k(k) |
| 4543 | .sparsity(0.0f) |
| 4544 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4545 | } |
| 4546 | } |
| 4547 | |
| 4548 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, n_gt_1) { |
| 4549 | for (uint32_t n = 2; n < 10; n++) { |
| 4550 | for (size_t k = 1; k <= 5; k += 2) { |
| 4551 | SpMMMicrokernelTester() |
| 4552 | .mr(4) |
| 4553 | .nr(1) |
| 4554 | .m(4) |
| 4555 | .n(n) |
| 4556 | .k(k) |
| 4557 | .sparsity(0.0f) |
| 4558 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4559 | } |
| 4560 | } |
| 4561 | } |
| 4562 | |
| 4563 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, m_lt_4) { |
| 4564 | for (uint32_t m = 1; m < 4; m++) { |
| 4565 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4566 | for (size_t k = 1; k <= 5; k += 2) { |
| 4567 | SpMMMicrokernelTester() |
| 4568 | .mr(4) |
| 4569 | .nr(1) |
| 4570 | .m(m) |
| 4571 | .n(n) |
| 4572 | .k(k) |
| 4573 | .sparsity(0.0f) |
| 4574 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4575 | } |
| 4576 | } |
| 4577 | } |
| 4578 | } |
| 4579 | |
| 4580 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, m_div_4) { |
| 4581 | for (uint32_t m = 8; m <= 12; m += 4) { |
| 4582 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4583 | for (size_t k = 1; k <= 5; k += 2) { |
| 4584 | SpMMMicrokernelTester() |
| 4585 | .mr(4) |
| 4586 | .nr(1) |
| 4587 | .m(m) |
| 4588 | .n(n) |
| 4589 | .k(k) |
| 4590 | .sparsity(0.0f) |
| 4591 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4592 | } |
| 4593 | } |
| 4594 | } |
| 4595 | } |
| 4596 | |
| 4597 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, m_gt_4) { |
| 4598 | for (uint32_t m = 5; m < 8; m++) { |
| 4599 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4600 | for (size_t k = 1; k <= 5; k += 2) { |
| 4601 | SpMMMicrokernelTester() |
| 4602 | .mr(4) |
| 4603 | .nr(1) |
| 4604 | .m(m) |
| 4605 | .n(n) |
| 4606 | .k(k) |
| 4607 | .sparsity(0.0f) |
| 4608 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4609 | } |
| 4610 | } |
| 4611 | } |
| 4612 | } |
| 4613 | |
| 4614 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, qmin) { |
| 4615 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4616 | for (size_t k = 1; k <= 5; k += 2) { |
| 4617 | SpMMMicrokernelTester() |
| 4618 | .mr(4) |
| 4619 | .nr(1) |
| 4620 | .m(8) |
| 4621 | .n(n) |
| 4622 | .k(k) |
| 4623 | .sparsity(0.0f) |
| 4624 | .qmin(128) |
| 4625 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4626 | } |
| 4627 | } |
| 4628 | } |
| 4629 | |
| 4630 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, qmax) { |
| 4631 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4632 | for (size_t k = 1; k <= 5; k += 2) { |
| 4633 | SpMMMicrokernelTester() |
| 4634 | .mr(4) |
| 4635 | .nr(1) |
| 4636 | .m(8) |
| 4637 | .n(n) |
| 4638 | .k(k) |
| 4639 | .sparsity(0.0f) |
| 4640 | .qmax(128) |
| 4641 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4642 | } |
| 4643 | } |
| 4644 | } |
| 4645 | |
| 4646 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, half_sparse) { |
| 4647 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4648 | for (size_t k = 1; k <= 5; k += 2) { |
| 4649 | SpMMMicrokernelTester() |
| 4650 | .mr(4) |
| 4651 | .nr(1) |
| 4652 | .m(8) |
| 4653 | .n(n) |
| 4654 | .k(k) |
| 4655 | .sparsity(0.5f) |
| 4656 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4657 | } |
| 4658 | } |
| 4659 | } |
| 4660 | |
| 4661 | TEST(F32_SPMM_4X1__SCALAR_PIPELINED, zero_weights) { |
| 4662 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4663 | for (size_t k = 1; k <= 5; k += 2) { |
| 4664 | SpMMMicrokernelTester() |
| 4665 | .mr(4) |
| 4666 | .nr(1) |
| 4667 | .m(8) |
| 4668 | .n(n) |
| 4669 | .k(k) |
| 4670 | .sparsity(1.0f) |
| 4671 | .Test(xnn_f32_spmm_ukernel_4x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4672 | } |
| 4673 | } |
| 4674 | } |
| 4675 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4676 | TEST(F32_SPMM_8X1__SCALAR, k_eq_1) { |
| 4677 | SpMMMicrokernelTester() |
| 4678 | .mr(8) |
| 4679 | .nr(1) |
| 4680 | .m(8) |
| 4681 | .n(1) |
| 4682 | .k(1) |
| 4683 | .sparsity(0.0f) |
| 4684 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4685 | } |
| 4686 | |
| 4687 | TEST(F32_SPMM_8X1__SCALAR, k_gt_1) { |
| 4688 | for (size_t k = 2; k < 10; k++) { |
| 4689 | SpMMMicrokernelTester() |
| 4690 | .mr(8) |
| 4691 | .nr(1) |
| 4692 | .m(8) |
| 4693 | .n(1) |
| 4694 | .k(k) |
| 4695 | .sparsity(0.0f) |
| 4696 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4697 | } |
| 4698 | } |
| 4699 | |
| 4700 | TEST(F32_SPMM_8X1__SCALAR, n_gt_1) { |
| 4701 | for (uint32_t n = 2; n < 10; n++) { |
| 4702 | for (size_t k = 1; k <= 5; k += 2) { |
| 4703 | SpMMMicrokernelTester() |
| 4704 | .mr(8) |
| 4705 | .nr(1) |
| 4706 | .m(8) |
| 4707 | .n(n) |
| 4708 | .k(k) |
| 4709 | .sparsity(0.0f) |
| 4710 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4711 | } |
| 4712 | } |
| 4713 | } |
| 4714 | |
| 4715 | TEST(F32_SPMM_8X1__SCALAR, m_lt_8) { |
| 4716 | for (uint32_t m = 1; m < 8; m++) { |
| 4717 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4718 | for (size_t k = 1; k <= 5; k += 2) { |
| 4719 | SpMMMicrokernelTester() |
| 4720 | .mr(8) |
| 4721 | .nr(1) |
| 4722 | .m(m) |
| 4723 | .n(n) |
| 4724 | .k(k) |
| 4725 | .sparsity(0.0f) |
| 4726 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4727 | } |
| 4728 | } |
| 4729 | } |
| 4730 | } |
| 4731 | |
| 4732 | TEST(F32_SPMM_8X1__SCALAR, m_div_8) { |
| 4733 | for (uint32_t m = 16; m <= 24; m += 8) { |
| 4734 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4735 | for (size_t k = 1; k <= 5; k += 2) { |
| 4736 | SpMMMicrokernelTester() |
| 4737 | .mr(8) |
| 4738 | .nr(1) |
| 4739 | .m(m) |
| 4740 | .n(n) |
| 4741 | .k(k) |
| 4742 | .sparsity(0.0f) |
| 4743 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4744 | } |
| 4745 | } |
| 4746 | } |
| 4747 | } |
| 4748 | |
| 4749 | TEST(F32_SPMM_8X1__SCALAR, m_gt_8) { |
| 4750 | for (uint32_t m = 9; m < 16; m++) { |
| 4751 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4752 | for (size_t k = 1; k <= 5; k += 2) { |
| 4753 | SpMMMicrokernelTester() |
| 4754 | .mr(8) |
| 4755 | .nr(1) |
| 4756 | .m(m) |
| 4757 | .n(n) |
| 4758 | .k(k) |
| 4759 | .sparsity(0.0f) |
| 4760 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4761 | } |
| 4762 | } |
| 4763 | } |
| 4764 | } |
| 4765 | |
| 4766 | TEST(F32_SPMM_8X1__SCALAR, qmin) { |
| 4767 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4768 | for (size_t k = 1; k <= 5; k += 2) { |
| 4769 | SpMMMicrokernelTester() |
| 4770 | .mr(8) |
| 4771 | .nr(1) |
| 4772 | .m(16) |
| 4773 | .n(n) |
| 4774 | .k(k) |
| 4775 | .sparsity(0.0f) |
| 4776 | .qmin(128) |
| 4777 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4778 | } |
| 4779 | } |
| 4780 | } |
| 4781 | |
| 4782 | TEST(F32_SPMM_8X1__SCALAR, qmax) { |
| 4783 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4784 | for (size_t k = 1; k <= 5; k += 2) { |
| 4785 | SpMMMicrokernelTester() |
| 4786 | .mr(8) |
| 4787 | .nr(1) |
| 4788 | .m(16) |
| 4789 | .n(n) |
| 4790 | .k(k) |
| 4791 | .sparsity(0.0f) |
| 4792 | .qmax(128) |
| 4793 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4794 | } |
| 4795 | } |
| 4796 | } |
| 4797 | |
| 4798 | TEST(F32_SPMM_8X1__SCALAR, half_sparse) { |
| 4799 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4800 | for (size_t k = 1; k <= 5; k += 2) { |
| 4801 | SpMMMicrokernelTester() |
| 4802 | .mr(8) |
| 4803 | .nr(1) |
| 4804 | .m(16) |
| 4805 | .n(n) |
| 4806 | .k(k) |
| 4807 | .sparsity(0.5f) |
| 4808 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4809 | } |
| 4810 | } |
| 4811 | } |
| 4812 | |
| 4813 | TEST(F32_SPMM_8X1__SCALAR, zero_weights) { |
| 4814 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4815 | for (size_t k = 1; k <= 5; k += 2) { |
| 4816 | SpMMMicrokernelTester() |
| 4817 | .mr(8) |
| 4818 | .nr(1) |
| 4819 | .m(16) |
| 4820 | .n(n) |
| 4821 | .k(k) |
| 4822 | .sparsity(1.0f) |
| 4823 | .Test(xnn_f32_spmm_ukernel_8x1__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 4824 | } |
| 4825 | } |
| 4826 | } |
| 4827 | |
| 4828 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, k_eq_1) { |
| 4829 | SpMMMicrokernelTester() |
| 4830 | .mr(8) |
| 4831 | .nr(1) |
| 4832 | .m(8) |
| 4833 | .n(1) |
| 4834 | .k(1) |
| 4835 | .sparsity(0.0f) |
| 4836 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4837 | } |
| 4838 | |
| 4839 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, k_gt_1) { |
| 4840 | for (size_t k = 2; k < 10; k++) { |
| 4841 | SpMMMicrokernelTester() |
| 4842 | .mr(8) |
| 4843 | .nr(1) |
| 4844 | .m(8) |
| 4845 | .n(1) |
| 4846 | .k(k) |
| 4847 | .sparsity(0.0f) |
| 4848 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4849 | } |
| 4850 | } |
| 4851 | |
| 4852 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, n_gt_1) { |
| 4853 | for (uint32_t n = 2; n < 10; n++) { |
| 4854 | for (size_t k = 1; k <= 5; k += 2) { |
| 4855 | SpMMMicrokernelTester() |
| 4856 | .mr(8) |
| 4857 | .nr(1) |
| 4858 | .m(8) |
| 4859 | .n(n) |
| 4860 | .k(k) |
| 4861 | .sparsity(0.0f) |
| 4862 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4863 | } |
| 4864 | } |
| 4865 | } |
| 4866 | |
| 4867 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, m_lt_8) { |
| 4868 | for (uint32_t m = 1; m < 8; m++) { |
| 4869 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4870 | for (size_t k = 1; k <= 5; k += 2) { |
| 4871 | SpMMMicrokernelTester() |
| 4872 | .mr(8) |
| 4873 | .nr(1) |
| 4874 | .m(m) |
| 4875 | .n(n) |
| 4876 | .k(k) |
| 4877 | .sparsity(0.0f) |
| 4878 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4879 | } |
| 4880 | } |
| 4881 | } |
| 4882 | } |
| 4883 | |
| 4884 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, m_div_8) { |
| 4885 | for (uint32_t m = 16; m <= 24; m += 8) { |
| 4886 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4887 | for (size_t k = 1; k <= 5; k += 2) { |
| 4888 | SpMMMicrokernelTester() |
| 4889 | .mr(8) |
| 4890 | .nr(1) |
| 4891 | .m(m) |
| 4892 | .n(n) |
| 4893 | .k(k) |
| 4894 | .sparsity(0.0f) |
| 4895 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4896 | } |
| 4897 | } |
| 4898 | } |
| 4899 | } |
| 4900 | |
| 4901 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, m_gt_8) { |
| 4902 | for (uint32_t m = 9; m < 16; m++) { |
| 4903 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4904 | for (size_t k = 1; k <= 5; k += 2) { |
| 4905 | SpMMMicrokernelTester() |
| 4906 | .mr(8) |
| 4907 | .nr(1) |
| 4908 | .m(m) |
| 4909 | .n(n) |
| 4910 | .k(k) |
| 4911 | .sparsity(0.0f) |
| 4912 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4913 | } |
| 4914 | } |
| 4915 | } |
| 4916 | } |
| 4917 | |
| 4918 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, qmin) { |
| 4919 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4920 | for (size_t k = 1; k <= 5; k += 2) { |
| 4921 | SpMMMicrokernelTester() |
| 4922 | .mr(8) |
| 4923 | .nr(1) |
| 4924 | .m(16) |
| 4925 | .n(n) |
| 4926 | .k(k) |
| 4927 | .sparsity(0.0f) |
| 4928 | .qmin(128) |
| 4929 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4930 | } |
| 4931 | } |
| 4932 | } |
| 4933 | |
| 4934 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, qmax) { |
| 4935 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4936 | for (size_t k = 1; k <= 5; k += 2) { |
| 4937 | SpMMMicrokernelTester() |
| 4938 | .mr(8) |
| 4939 | .nr(1) |
| 4940 | .m(16) |
| 4941 | .n(n) |
| 4942 | .k(k) |
| 4943 | .sparsity(0.0f) |
| 4944 | .qmax(128) |
| 4945 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4946 | } |
| 4947 | } |
| 4948 | } |
| 4949 | |
| 4950 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, half_sparse) { |
| 4951 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4952 | for (size_t k = 1; k <= 5; k += 2) { |
| 4953 | SpMMMicrokernelTester() |
| 4954 | .mr(8) |
| 4955 | .nr(1) |
| 4956 | .m(16) |
| 4957 | .n(n) |
| 4958 | .k(k) |
| 4959 | .sparsity(0.5f) |
| 4960 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4961 | } |
| 4962 | } |
| 4963 | } |
| 4964 | |
| 4965 | TEST(F32_SPMM_8X1__SCALAR_PIPELINED, zero_weights) { |
| 4966 | for (uint32_t n = 1; n < 10; n += 2) { |
| 4967 | for (size_t k = 1; k <= 5; k += 2) { |
| 4968 | SpMMMicrokernelTester() |
| 4969 | .mr(8) |
| 4970 | .nr(1) |
| 4971 | .m(16) |
| 4972 | .n(n) |
| 4973 | .k(k) |
| 4974 | .sparsity(1.0f) |
| 4975 | .Test(xnn_f32_spmm_ukernel_8x1__scalar_pipelined, SpMMMicrokernelTester::Variant::Scalar); |
| 4976 | } |
| 4977 | } |
| 4978 | } |
| 4979 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 4980 | TEST(F32_SPMM_8X2__SCALAR, k_eq_1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4981 | SpMMMicrokernelTester() |
| 4982 | .mr(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 4983 | .nr(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4984 | .m(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 4985 | .n(2) |
| 4986 | .k(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4987 | .sparsity(0.0f) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 4988 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4989 | } |
| 4990 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 4991 | TEST(F32_SPMM_8X2__SCALAR, k_eq_1_subtile) { |
| 4992 | for (uint32_t n = 1; n <= 2; n++) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4993 | SpMMMicrokernelTester() |
| 4994 | .mr(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 4995 | .nr(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4996 | .m(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 4997 | .n(n) |
| 4998 | .k(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 4999 | .sparsity(0.0f) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5000 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5001 | } |
| 5002 | } |
| 5003 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5004 | TEST(F32_SPMM_8X2__SCALAR, k_gt_1) { |
| 5005 | for (size_t k = 2; k < 10; k++) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5006 | SpMMMicrokernelTester() |
| 5007 | .mr(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5008 | .nr(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5009 | .m(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5010 | .n(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5011 | .k(k) |
| 5012 | .sparsity(0.0f) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5013 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5014 | } |
| 5015 | } |
| 5016 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5017 | TEST(F32_SPMM_8X2__SCALAR, k_gt_1_subtile) { |
| 5018 | for (size_t k = 2; k < 10; k++) { |
| 5019 | for (uint32_t n = 1; n <= 2; n++) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5020 | SpMMMicrokernelTester() |
| 5021 | .mr(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5022 | .nr(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5023 | .m(8) |
| 5024 | .n(n) |
| 5025 | .k(k) |
| 5026 | .sparsity(0.0f) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5027 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5028 | } |
| 5029 | } |
| 5030 | } |
| 5031 | |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5032 | TEST(F32_SPMM_8X2__SCALAR, n_gt_2) { |
| 5033 | for (uint32_t n = 3; n < 10; n++) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5034 | for (size_t k = 1; k <= 5; k += 2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5035 | SpMMMicrokernelTester() |
| 5036 | .mr(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5037 | .nr(2) |
| 5038 | .m(8) |
| 5039 | .n(n) |
| 5040 | .k(k) |
| 5041 | .sparsity(0.0f) |
| 5042 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5043 | } |
| 5044 | } |
| 5045 | } |
| 5046 | |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5047 | TEST(F32_SPMM_8X2__SCALAR, n_div_2) { |
| 5048 | for (uint32_t n = 4; n <= 6; n += 2) { |
| 5049 | for (size_t k = 1; k <= 5; k += 2) { |
| 5050 | SpMMMicrokernelTester() |
| 5051 | .mr(8) |
| 5052 | .nr(2) |
| 5053 | .m(8) |
| 5054 | .n(n) |
| 5055 | .k(k) |
| 5056 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5057 | } |
| 5058 | } |
| 5059 | } |
| 5060 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5061 | TEST(F32_SPMM_8X2__SCALAR, m_lt_8) { |
| 5062 | for (uint32_t m = 1; m < 8; m++) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5063 | for (uint32_t n = 1; n < 10; n += 3) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5064 | for (size_t k = 1; k <= 5; k += 2) { |
| 5065 | SpMMMicrokernelTester() |
| 5066 | .mr(8) |
| 5067 | .nr(2) |
| 5068 | .m(m) |
| 5069 | .n(n) |
| 5070 | .k(k) |
| 5071 | .sparsity(0.0f) |
| 5072 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5073 | } |
| 5074 | } |
| 5075 | } |
| 5076 | } |
| 5077 | |
| 5078 | TEST(F32_SPMM_8X2__SCALAR, m_div_8) { |
| 5079 | for (uint32_t m = 16; m <= 24; m += 8) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5080 | for (uint32_t n = 1; n < 10; n += 3) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5081 | for (size_t k = 1; k <= 5; k += 2) { |
| 5082 | SpMMMicrokernelTester() |
| 5083 | .mr(8) |
| 5084 | .nr(2) |
| 5085 | .m(m) |
| 5086 | .n(n) |
| 5087 | .k(k) |
| 5088 | .sparsity(0.0f) |
| 5089 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5090 | } |
| 5091 | } |
| 5092 | } |
| 5093 | } |
| 5094 | |
| 5095 | TEST(F32_SPMM_8X2__SCALAR, m_gt_8) { |
| 5096 | for (uint32_t m = 9; m < 16; m++) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5097 | for (uint32_t n = 1; n < 10; n += 3) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5098 | for (size_t k = 1; k <= 5; k += 2) { |
| 5099 | SpMMMicrokernelTester() |
| 5100 | .mr(8) |
| 5101 | .nr(2) |
| 5102 | .m(m) |
| 5103 | .n(n) |
| 5104 | .k(k) |
| 5105 | .sparsity(0.0f) |
| 5106 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5107 | } |
| 5108 | } |
| 5109 | } |
| 5110 | } |
| 5111 | |
| 5112 | TEST(F32_SPMM_8X2__SCALAR, qmin) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5113 | for (uint32_t n = 1; n < 10; n += 3) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5114 | for (size_t k = 1; k <= 5; k += 2) { |
| 5115 | SpMMMicrokernelTester() |
| 5116 | .mr(8) |
| 5117 | .nr(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5118 | .m(16) |
| 5119 | .n(n) |
| 5120 | .k(k) |
| 5121 | .sparsity(0.0f) |
| 5122 | .qmin(128) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5123 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5124 | } |
| 5125 | } |
| 5126 | } |
| 5127 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5128 | TEST(F32_SPMM_8X2__SCALAR, qmax) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5129 | for (uint32_t n = 1; n < 10; n += 3) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5130 | for (size_t k = 1; k <= 5; k += 2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5131 | SpMMMicrokernelTester() |
| 5132 | .mr(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5133 | .nr(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5134 | .m(16) |
| 5135 | .n(n) |
| 5136 | .k(k) |
| 5137 | .sparsity(0.0f) |
| 5138 | .qmax(128) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5139 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5140 | } |
| 5141 | } |
| 5142 | } |
| 5143 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5144 | TEST(F32_SPMM_8X2__SCALAR, half_sparse) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5145 | for (uint32_t n = 1; n < 10; n += 3) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5146 | for (size_t k = 1; k <= 5; k += 2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5147 | SpMMMicrokernelTester() |
| 5148 | .mr(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5149 | .nr(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5150 | .m(16) |
| 5151 | .n(n) |
| 5152 | .k(k) |
| 5153 | .sparsity(0.5f) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5154 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5155 | } |
| 5156 | } |
| 5157 | } |
| 5158 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5159 | TEST(F32_SPMM_8X2__SCALAR, zero_weights) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5160 | for (uint32_t n = 1; n < 10; n += 3) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5161 | for (size_t k = 1; k <= 5; k += 2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5162 | SpMMMicrokernelTester() |
| 5163 | .mr(8) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5164 | .nr(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5165 | .m(16) |
| 5166 | .n(n) |
| 5167 | .k(k) |
| 5168 | .sparsity(1.0f) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5169 | .Test(xnn_f32_spmm_ukernel_8x2__scalar, SpMMMicrokernelTester::Variant::Scalar); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5170 | } |
| 5171 | } |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5172 | } |
| 5173 | |
| 5174 | TEST(F32_SPMM_8X4__SCALAR, k_eq_1) { |
| 5175 | SpMMMicrokernelTester() |
| 5176 | .mr(8) |
| 5177 | .nr(4) |
| 5178 | .m(8) |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5179 | .n(4) |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5180 | .k(1) |
| 5181 | .sparsity(0.0f) |
| 5182 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5183 | } |
| 5184 | |
| 5185 | TEST(F32_SPMM_8X4__SCALAR, k_eq_1_subtile) { |
| 5186 | for (uint32_t n = 1; n <= 4; n++) { |
| 5187 | SpMMMicrokernelTester() |
| 5188 | .mr(8) |
| 5189 | .nr(4) |
| 5190 | .m(8) |
| 5191 | .n(n) |
| 5192 | .k(1) |
| 5193 | .sparsity(0.0f) |
| 5194 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5195 | } |
| 5196 | } |
| 5197 | |
| 5198 | TEST(F32_SPMM_8X4__SCALAR, k_gt_1) { |
| 5199 | for (size_t k = 2; k < 10; k++) { |
| 5200 | SpMMMicrokernelTester() |
| 5201 | .mr(8) |
| 5202 | .nr(4) |
| 5203 | .m(8) |
| 5204 | .n(4) |
| 5205 | .k(k) |
| 5206 | .sparsity(0.0f) |
| 5207 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5208 | } |
| 5209 | } |
| 5210 | |
| 5211 | TEST(F32_SPMM_8X4__SCALAR, k_gt_1_subtile) { |
| 5212 | for (size_t k = 2; k < 10; k++) { |
| 5213 | for (uint32_t n = 1; n <= 4; n++) { |
| 5214 | SpMMMicrokernelTester() |
| 5215 | .mr(8) |
| 5216 | .nr(4) |
| 5217 | .m(8) |
| 5218 | .n(n) |
| 5219 | .k(k) |
| 5220 | .sparsity(0.0f) |
| 5221 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5222 | } |
| 5223 | } |
| 5224 | } |
| 5225 | |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5226 | TEST(F32_SPMM_8X4__SCALAR, n_gt_4) { |
| 5227 | for (uint32_t n = 5; n < 10; n++) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5228 | for (size_t k = 1; k <= 5; k += 2) { |
| 5229 | SpMMMicrokernelTester() |
| 5230 | .mr(8) |
| 5231 | .nr(4) |
| 5232 | .m(8) |
| 5233 | .n(n) |
| 5234 | .k(k) |
| 5235 | .sparsity(0.0f) |
| 5236 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5237 | } |
| 5238 | } |
| 5239 | } |
| 5240 | |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5241 | TEST(F32_SPMM_8X4__SCALAR, n_div_4) { |
| 5242 | for (uint32_t n = 8; n <= 12; n += 4) { |
| 5243 | for (size_t k = 1; k <= 5; k += 2) { |
| 5244 | SpMMMicrokernelTester() |
| 5245 | .mr(8) |
| 5246 | .nr(4) |
| 5247 | .m(8) |
| 5248 | .n(n) |
| 5249 | .k(k) |
| 5250 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5251 | } |
| 5252 | } |
| 5253 | } |
| 5254 | |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5255 | TEST(F32_SPMM_8X4__SCALAR, m_lt_8) { |
| 5256 | for (uint32_t m = 1; m < 8; m++) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5257 | for (uint32_t n = 1; n < 20; n += 5) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5258 | for (size_t k = 1; k <= 5; k += 2) { |
| 5259 | SpMMMicrokernelTester() |
| 5260 | .mr(8) |
| 5261 | .nr(4) |
| 5262 | .m(m) |
| 5263 | .n(n) |
| 5264 | .k(k) |
| 5265 | .sparsity(0.0f) |
| 5266 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5267 | } |
| 5268 | } |
| 5269 | } |
| 5270 | } |
| 5271 | |
| 5272 | TEST(F32_SPMM_8X4__SCALAR, m_div_8) { |
| 5273 | for (uint32_t m = 16; m <= 24; m += 8) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5274 | for (uint32_t n = 1; n < 20; n += 5) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5275 | for (size_t k = 1; k <= 5; k += 2) { |
| 5276 | SpMMMicrokernelTester() |
| 5277 | .mr(8) |
| 5278 | .nr(4) |
| 5279 | .m(m) |
| 5280 | .n(n) |
| 5281 | .k(k) |
| 5282 | .sparsity(0.0f) |
| 5283 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5284 | } |
| 5285 | } |
| 5286 | } |
| 5287 | } |
| 5288 | |
| 5289 | TEST(F32_SPMM_8X4__SCALAR, m_gt_8) { |
| 5290 | for (uint32_t m = 9; m < 16; m++) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5291 | for (uint32_t n = 1; n < 20; n += 5) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5292 | for (size_t k = 1; k <= 5; k += 2) { |
| 5293 | SpMMMicrokernelTester() |
| 5294 | .mr(8) |
| 5295 | .nr(4) |
| 5296 | .m(m) |
| 5297 | .n(n) |
| 5298 | .k(k) |
| 5299 | .sparsity(0.0f) |
| 5300 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5301 | } |
| 5302 | } |
| 5303 | } |
| 5304 | } |
| 5305 | |
| 5306 | TEST(F32_SPMM_8X4__SCALAR, qmin) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5307 | for (uint32_t n = 1; n < 20; n += 5) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5308 | for (size_t k = 1; k <= 5; k += 2) { |
| 5309 | SpMMMicrokernelTester() |
| 5310 | .mr(8) |
| 5311 | .nr(4) |
| 5312 | .m(16) |
| 5313 | .n(n) |
| 5314 | .k(k) |
| 5315 | .sparsity(0.0f) |
| 5316 | .qmin(128) |
| 5317 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5318 | } |
| 5319 | } |
| 5320 | } |
| 5321 | |
| 5322 | TEST(F32_SPMM_8X4__SCALAR, qmax) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5323 | for (uint32_t n = 1; n < 20; n += 5) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5324 | for (size_t k = 1; k <= 5; k += 2) { |
| 5325 | SpMMMicrokernelTester() |
| 5326 | .mr(8) |
| 5327 | .nr(4) |
| 5328 | .m(16) |
| 5329 | .n(n) |
| 5330 | .k(k) |
| 5331 | .sparsity(0.0f) |
| 5332 | .qmax(128) |
| 5333 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5334 | } |
| 5335 | } |
| 5336 | } |
| 5337 | |
| 5338 | TEST(F32_SPMM_8X4__SCALAR, half_sparse) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5339 | for (uint32_t n = 1; n < 20; n += 5) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5340 | for (size_t k = 1; k <= 5; k += 2) { |
| 5341 | SpMMMicrokernelTester() |
| 5342 | .mr(8) |
| 5343 | .nr(4) |
| 5344 | .m(16) |
| 5345 | .n(n) |
| 5346 | .k(k) |
| 5347 | .sparsity(0.5f) |
| 5348 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5349 | } |
| 5350 | } |
| 5351 | } |
| 5352 | |
| 5353 | TEST(F32_SPMM_8X4__SCALAR, zero_weights) { |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5354 | for (uint32_t n = 1; n < 20; n += 5) { |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 5355 | for (size_t k = 1; k <= 5; k += 2) { |
| 5356 | SpMMMicrokernelTester() |
| 5357 | .mr(8) |
| 5358 | .nr(4) |
| 5359 | .m(16) |
| 5360 | .n(n) |
| 5361 | .k(k) |
| 5362 | .sparsity(1.0f) |
| 5363 | .Test(xnn_f32_spmm_ukernel_8x4__scalar, SpMMMicrokernelTester::Variant::Scalar); |
| 5364 | } |
| 5365 | } |
Marat Dukhan | c452eb1 | 2019-10-25 19:19:34 -0700 | [diff] [blame] | 5366 | } |