| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 9 | #include <gtest/gtest.h> |
| 10 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 11 | #include <xnnpack/common.h> |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 12 | #include <xnnpack/isa-checks.h> |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 13 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 14 | #include <xnnpack/avgpool.h> |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 15 | #include "avgpool-microkernel-tester.h" |
| 16 | |
| 17 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 18 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 19 | TEST(Q8_AVGPOOL_UP9__NEON, kc_eq_8_fulltile) { |
| 20 | TEST_REQUIRES_ARM_NEON; |
| 21 | auto tester = AvgPoolMicrokernelTester() |
| 22 | .mr(9) |
| 23 | .kc(8); |
| 24 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 25 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 26 | if (kh * kw == tester.mr()) { |
| 27 | tester |
| 28 | .kh(kh) |
| 29 | .kw(kw) |
| 30 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 31 | } |
| 32 | } |
| 33 | } |
| 34 | } |
| 35 | |
| 36 | TEST(Q8_AVGPOOL_UP9__NEON, kc_eq_8_subtile) { |
| 37 | TEST_REQUIRES_ARM_NEON; |
| 38 | auto tester = AvgPoolMicrokernelTester() |
| 39 | .mr(9) |
| 40 | .kc(8); |
| 41 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 42 | for (size_t kh = 1; kh <= ks; kh++) { |
| 43 | for (size_t kw = 1; kw <= ks; kw++) { |
| 44 | if (kh * kw == ks) { |
| 45 | tester |
| 46 | .kh(kh) |
| 47 | .kw(kw) |
| 48 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 49 | } |
| 50 | } |
| 51 | } |
| 52 | } |
| 53 | } |
| 54 | |
| 55 | TEST(Q8_AVGPOOL_UP9__NEON, kc_div_8_fulltile) { |
| 56 | TEST_REQUIRES_ARM_NEON; |
| 57 | auto tester = AvgPoolMicrokernelTester() |
| 58 | .mr(9); |
| 59 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 60 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 61 | if (kh * kw == tester.mr()) { |
| 62 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 63 | tester |
| 64 | .kh(kh) |
| 65 | .kw(kw) |
| 66 | .kc(kc) |
| 67 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 68 | } |
| 69 | } |
| 70 | } |
| 71 | } |
| 72 | } |
| 73 | |
| 74 | TEST(Q8_AVGPOOL_UP9__NEON, kc_div_8_subtile) { |
| 75 | TEST_REQUIRES_ARM_NEON; |
| 76 | auto tester = AvgPoolMicrokernelTester() |
| 77 | .mr(9) |
| 78 | .iterations(3); |
| 79 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 80 | for (size_t kh = 1; kh <= ks; kh++) { |
| 81 | for (size_t kw = 1; kw <= ks; kw++) { |
| 82 | if (kh * kw == ks) { |
| 83 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 84 | tester |
| 85 | .kh(kh) |
| 86 | .kw(kw) |
| 87 | .kc(kc) |
| 88 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 89 | } |
| 90 | } |
| 91 | } |
| 92 | } |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | TEST(Q8_AVGPOOL_UP9__NEON, kc_div_8_fulltile_with_x_stride) { |
| 97 | TEST_REQUIRES_ARM_NEON; |
| 98 | auto tester = AvgPoolMicrokernelTester() |
| 99 | .mr(9) |
| 100 | .iterations(3); |
| 101 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 102 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 103 | if (kh * kw == tester.mr()) { |
| 104 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 105 | tester |
| 106 | .kh(kh) |
| 107 | .kw(kw) |
| 108 | .kc(kc) |
| 109 | .x_stride(131) |
| 110 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 111 | } |
| 112 | } |
| 113 | } |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | TEST(Q8_AVGPOOL_UP9__NEON, kc_lt_8_fulltile) { |
| 118 | TEST_REQUIRES_ARM_NEON; |
| 119 | auto tester = AvgPoolMicrokernelTester() |
| 120 | .mr(9); |
| 121 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 122 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 123 | if (kh * kw == tester.mr()) { |
| 124 | for (size_t kc = 1; kc < 8; kc++) { |
| 125 | tester |
| 126 | .kh(kh) |
| 127 | .kw(kw) |
| 128 | .kc(kc) |
| 129 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 130 | } |
| 131 | } |
| 132 | } |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | TEST(Q8_AVGPOOL_UP9__NEON, kc_lt_8_subtile) { |
| 137 | TEST_REQUIRES_ARM_NEON; |
| 138 | auto tester = AvgPoolMicrokernelTester() |
| 139 | .mr(9) |
| 140 | .iterations(3); |
| 141 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 142 | for (size_t kh = 1; kh <= ks; kh++) { |
| 143 | for (size_t kw = 1; kw <= ks; kw++) { |
| 144 | if (kh * kw == ks) { |
| 145 | for (size_t kc = 1; kc < 8; kc++) { |
| 146 | tester |
| 147 | .kh(kh) |
| 148 | .kw(kw) |
| 149 | .kc(kc) |
| 150 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 151 | } |
| 152 | } |
| 153 | } |
| 154 | } |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | TEST(Q8_AVGPOOL_UP9__NEON, kc_lt_8_fulltile_with_x_stride) { |
| 159 | TEST_REQUIRES_ARM_NEON; |
| 160 | auto tester = AvgPoolMicrokernelTester() |
| 161 | .mr(9) |
| 162 | .iterations(3); |
| 163 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 164 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 165 | if (kh * kw == tester.mr()) { |
| 166 | for (size_t kc = 1; kc < 8; kc++) { |
| 167 | tester |
| 168 | .kh(kh) |
| 169 | .kw(kw) |
| 170 | .kc(kc) |
| 171 | .x_stride(23) |
| 172 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 173 | } |
| 174 | } |
| 175 | } |
| 176 | } |
| 177 | } |
| 178 | |
| 179 | TEST(Q8_AVGPOOL_UP9__NEON, kc_gt_8_fulltile) { |
| 180 | TEST_REQUIRES_ARM_NEON; |
| 181 | auto tester = AvgPoolMicrokernelTester() |
| 182 | .mr(9); |
| 183 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 184 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 185 | if (kh * kw == tester.mr()) { |
| 186 | for (size_t kc = 9; kc < 16; kc++) { |
| 187 | tester |
| 188 | .kh(kh) |
| 189 | .kw(kw) |
| 190 | .kc(kc) |
| 191 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 192 | } |
| 193 | } |
| 194 | } |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | TEST(Q8_AVGPOOL_UP9__NEON, kc_gt_8_subtile) { |
| 199 | TEST_REQUIRES_ARM_NEON; |
| 200 | auto tester = AvgPoolMicrokernelTester() |
| 201 | .mr(9) |
| 202 | .iterations(3); |
| 203 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 204 | for (size_t kh = 1; kh <= ks; kh++) { |
| 205 | for (size_t kw = 1; kw <= ks; kw++) { |
| 206 | if (kh * kw == ks) { |
| 207 | for (size_t kc = 9; kc < 16; kc++) { |
| 208 | tester |
| 209 | .kh(kh) |
| 210 | .kw(kw) |
| 211 | .kc(kc) |
| 212 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 213 | } |
| 214 | } |
| 215 | } |
| 216 | } |
| 217 | } |
| 218 | } |
| 219 | |
| 220 | TEST(Q8_AVGPOOL_UP9__NEON, kc_gt_8_fulltile_with_x_stride) { |
| 221 | TEST_REQUIRES_ARM_NEON; |
| 222 | auto tester = AvgPoolMicrokernelTester() |
| 223 | .mr(9) |
| 224 | .iterations(3); |
| 225 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 226 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 227 | if (kh * kw == tester.mr()) { |
| 228 | for (size_t kc = 9; kc < 16; kc++) { |
| 229 | tester |
| 230 | .kh(kh) |
| 231 | .kw(kw) |
| 232 | .kc(kc) |
| 233 | .x_stride(23) |
| 234 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 235 | } |
| 236 | } |
| 237 | } |
| 238 | } |
| 239 | } |
| 240 | |
| 241 | TEST(Q8_AVGPOOL_UP9__NEON, kc_div_8_with_x_scale) { |
| 242 | TEST_REQUIRES_ARM_NEON; |
| 243 | for (size_t n = 1; n <= 5; n += 2) { |
| 244 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 245 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 246 | AvgPoolMicrokernelTester() |
| 247 | .mr(9) |
| 248 | .n(n) |
| 249 | .kh(3) |
| 250 | .kw(3) |
| 251 | .kc(kc) |
| 252 | .x_scale(x_scale) |
| 253 | .iterations(2) |
| 254 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 255 | } |
| 256 | } |
| 257 | } |
| 258 | } |
| 259 | |
| 260 | TEST(Q8_AVGPOOL_UP9__NEON, kc_div_8_with_x_zero_point) { |
| 261 | TEST_REQUIRES_ARM_NEON; |
| 262 | for (size_t n = 1; n <= 5; n += 2) { |
| 263 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 264 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 265 | AvgPoolMicrokernelTester() |
| 266 | .mr(9) |
| 267 | .n(n) |
| 268 | .kh(3) |
| 269 | .kw(3) |
| 270 | .kc(kc) |
| 271 | .x_zero_point(uint8_t(x_zero_point)) |
| 272 | .iterations(3) |
| 273 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 274 | } |
| 275 | } |
| 276 | } |
| 277 | } |
| 278 | |
| 279 | TEST(Q8_AVGPOOL_UP9__NEON, kc_div_8_with_y_scale) { |
| 280 | TEST_REQUIRES_ARM_NEON; |
| 281 | for (size_t n = 1; n <= 5; n += 2) { |
| 282 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 283 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 284 | AvgPoolMicrokernelTester() |
| 285 | .mr(9) |
| 286 | .n(n) |
| 287 | .kh(3) |
| 288 | .kw(3) |
| 289 | .kc(kc) |
| 290 | .y_scale(y_scale) |
| 291 | .iterations(2) |
| 292 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 293 | } |
| 294 | } |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | TEST(Q8_AVGPOOL_UP9__NEON, kc_div_8_with_y_zero_point) { |
| 299 | TEST_REQUIRES_ARM_NEON; |
| 300 | for (size_t n = 1; n <= 5; n += 2) { |
| 301 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 302 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 303 | AvgPoolMicrokernelTester() |
| 304 | .mr(9) |
| 305 | .n(n) |
| 306 | .kh(3) |
| 307 | .kw(3) |
| 308 | .kc(kc) |
| 309 | .y_zero_point(uint8_t(y_zero_point)) |
| 310 | .iterations(3) |
| 311 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 312 | } |
| 313 | } |
| 314 | } |
| 315 | } |
| 316 | |
| 317 | TEST(Q8_AVGPOOL_UP9__NEON, kc_div_8_with_qmax) { |
| 318 | TEST_REQUIRES_ARM_NEON; |
| 319 | for (size_t n = 1; n <= 5; n += 2) { |
| 320 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 321 | AvgPoolMicrokernelTester() |
| 322 | .mr(9) |
| 323 | .n(n) |
| 324 | .kh(3) |
| 325 | .kw(3) |
| 326 | .kc(kc) |
| 327 | .x_zero_point(128) |
| 328 | .y_zero_point(128) |
| 329 | .x_scale(1.0f) |
| 330 | .y_scale(1.0f) |
| 331 | .qmax(128) |
| 332 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 333 | } |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | TEST(Q8_AVGPOOL_UP9__NEON, kc_div_8_with_qmin) { |
| 338 | TEST_REQUIRES_ARM_NEON; |
| 339 | for (size_t n = 1; n <= 5; n += 2) { |
| 340 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 341 | AvgPoolMicrokernelTester() |
| 342 | .mr(9) |
| 343 | .n(n) |
| 344 | .kh(3) |
| 345 | .kw(3) |
| 346 | .kc(kc) |
| 347 | .x_zero_point(128) |
| 348 | .y_zero_point(128) |
| 349 | .x_scale(1.0f) |
| 350 | .y_scale(1.0f) |
| 351 | .qmin(128) |
| 352 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 353 | } |
| 354 | } |
| 355 | } |
| 356 | |
| 357 | TEST(Q8_AVGPOOL_UP9__NEON, small_n) { |
| 358 | TEST_REQUIRES_ARM_NEON; |
| 359 | for (size_t n = 2; n < 5; n++) { |
| 360 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 361 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 362 | AvgPoolMicrokernelTester() |
| 363 | .mr(9) |
| 364 | .n(n) |
| 365 | .kh(ks) |
| 366 | .kw(ks) |
| 367 | .kc(kc) |
| 368 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 369 | } |
| 370 | } |
| 371 | } |
| 372 | } |
| 373 | |
| 374 | TEST(Q8_AVGPOOL_UP9__NEON, small_n_with_x_stride) { |
| 375 | TEST_REQUIRES_ARM_NEON; |
| 376 | for (size_t n = 2; n < 5; n++) { |
| 377 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 378 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 379 | AvgPoolMicrokernelTester() |
| 380 | .mr(9) |
| 381 | .n(n) |
| 382 | .kh(ks) |
| 383 | .kw(ks) |
| 384 | .kc(kc) |
| 385 | .x_stride(29) |
| 386 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 387 | } |
| 388 | } |
| 389 | } |
| 390 | } |
| 391 | |
| 392 | TEST(Q8_AVGPOOL_UP9__NEON, small_n_with_y_stride) { |
| 393 | TEST_REQUIRES_ARM_NEON; |
| 394 | for (size_t n = 2; n < 5; n++) { |
| 395 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 396 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 397 | AvgPoolMicrokernelTester() |
| 398 | .mr(9) |
| 399 | .n(n) |
| 400 | .kh(ks) |
| 401 | .kw(ks) |
| 402 | .kc(kc) |
| 403 | .y_stride(31) |
| 404 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 405 | } |
| 406 | } |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | TEST(Q8_AVGPOOL_UP9__NEON, small_n_with_s) { |
| 411 | TEST_REQUIRES_ARM_NEON; |
| 412 | for (size_t n = 2; n < 5; n++) { |
| 413 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 414 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 415 | for (size_t s = 2; s <= ks; s++) { |
| 416 | AvgPoolMicrokernelTester() |
| 417 | .mr(9) |
| 418 | .n(n) |
| 419 | .kh(ks) |
| 420 | .kw(ks) |
| 421 | .kc(kc) |
| 422 | .s(s) |
| 423 | .Test(xnn_q8_avgpool_ukernel_up9__neon); |
| 424 | } |
| 425 | } |
| 426 | } |
| 427 | } |
| 428 | } |
| 429 | |
| 430 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_eq_8_twopass_fulltile) { |
| 431 | TEST_REQUIRES_ARM_NEON; |
| 432 | auto tester = AvgPoolMicrokernelTester() |
| 433 | .mr(9) |
| 434 | .qr(8) |
| 435 | .kc(8); |
| 436 | const size_t ks = tester.mr() + tester.qr(); |
| 437 | for (size_t kh = 1; kh <= ks; kh++) { |
| 438 | for (size_t kw = 1; kw <= ks; kw++) { |
| 439 | if (kh * kw == ks) { |
| 440 | tester |
| 441 | .kh(kh) |
| 442 | .kw(kw) |
| 443 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 444 | } |
| 445 | } |
| 446 | } |
| 447 | } |
| 448 | |
| 449 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_eq_8_twopass_subtile) { |
| 450 | TEST_REQUIRES_ARM_NEON; |
| 451 | auto tester = AvgPoolMicrokernelTester() |
| 452 | .mr(9) |
| 453 | .qr(8) |
| 454 | .kc(8); |
| 455 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 456 | tester |
| 457 | .kh(ks) |
| 458 | .kw(1) |
| 459 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 460 | tester |
| 461 | .kh(1) |
| 462 | .kw(ks) |
| 463 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 464 | } |
| 465 | } |
| 466 | |
| 467 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_eq_8_multipass_fulltile) { |
| 468 | TEST_REQUIRES_ARM_NEON; |
| 469 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 470 | auto tester = AvgPoolMicrokernelTester() |
| 471 | .mr(9) |
| 472 | .qr(8) |
| 473 | .kc(8); |
| 474 | for (size_t kh = 1; kh <= ks; kh++) { |
| 475 | for (size_t kw = 1; kw <= ks; kw++) { |
| 476 | if (kh * kw == ks) { |
| 477 | tester |
| 478 | .kh(kh) |
| 479 | .kw(kw) |
| 480 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 481 | } |
| 482 | } |
| 483 | } |
| 484 | } |
| 485 | } |
| 486 | |
| 487 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_eq_8_multipass_subtile) { |
| 488 | TEST_REQUIRES_ARM_NEON; |
| 489 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 490 | auto tester = AvgPoolMicrokernelTester() |
| 491 | .mr(9) |
| 492 | .qr(8) |
| 493 | .kc(8); |
| 494 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 495 | tester |
| 496 | .kh(ks) |
| 497 | .kw(1) |
| 498 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 499 | tester |
| 500 | .kh(1) |
| 501 | .kw(ks) |
| 502 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 503 | } |
| 504 | } |
| 505 | } |
| 506 | |
| 507 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_twopass_fulltile) { |
| 508 | TEST_REQUIRES_ARM_NEON; |
| 509 | auto tester = AvgPoolMicrokernelTester() |
| 510 | .mr(9) |
| 511 | .qr(8) |
| 512 | .iterations(3); |
| 513 | const size_t ks = 17; |
| 514 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 515 | tester |
| 516 | .kc(kc) |
| 517 | .kh(ks) |
| 518 | .kw(1) |
| 519 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 520 | tester |
| 521 | .kc(kc) |
| 522 | .kh(1) |
| 523 | .kw(ks) |
| 524 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 525 | } |
| 526 | } |
| 527 | |
| 528 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_twopass_subtile) { |
| 529 | TEST_REQUIRES_ARM_NEON; |
| 530 | auto tester = AvgPoolMicrokernelTester() |
| 531 | .mr(9) |
| 532 | .qr(8) |
| 533 | .iterations(3); |
| 534 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 535 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 536 | tester |
| 537 | .kc(kc) |
| 538 | .kh(ks) |
| 539 | .kw(1) |
| 540 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 541 | tester |
| 542 | .kc(kc) |
| 543 | .kh(1) |
| 544 | .kw(ks) |
| 545 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 546 | } |
| 547 | } |
| 548 | } |
| 549 | |
| 550 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_twopass_fulltile_with_x_stride) { |
| 551 | TEST_REQUIRES_ARM_NEON; |
| 552 | auto tester = AvgPoolMicrokernelTester() |
| 553 | .mr(9) |
| 554 | .qr(8) |
| 555 | .iterations(3); |
| 556 | const size_t ks = tester.mr() + tester.qr(); |
| 557 | for (size_t kh = 1; kh <= ks; kh++) { |
| 558 | for (size_t kw = 1; kw <= ks; kw++) { |
| 559 | if (kh * kw == ks) { |
| 560 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 561 | tester |
| 562 | .kh(kh) |
| 563 | .kw(kw) |
| 564 | .kc(kc) |
| 565 | .x_stride(131) |
| 566 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 567 | } |
| 568 | } |
| 569 | } |
| 570 | } |
| 571 | } |
| 572 | |
| 573 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_multipass_fulltile) { |
| 574 | TEST_REQUIRES_ARM_NEON; |
| 575 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 576 | auto tester = AvgPoolMicrokernelTester() |
| 577 | .mr(9) |
| 578 | .qr(8) |
| 579 | .iterations(3); |
| 580 | for (size_t kh = 1; kh <= ks; kh++) { |
| 581 | for (size_t kw = 1; kw <= ks; kw++) { |
| 582 | if (kh * kw == ks) { |
| 583 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 584 | tester |
| 585 | .kh(kh) |
| 586 | .kw(kw) |
| 587 | .kc(kc) |
| 588 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 589 | } |
| 590 | } |
| 591 | } |
| 592 | } |
| 593 | } |
| 594 | } |
| 595 | |
| 596 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_multipass_subtile) { |
| 597 | TEST_REQUIRES_ARM_NEON; |
| 598 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 599 | auto tester = AvgPoolMicrokernelTester() |
| 600 | .mr(9) |
| 601 | .qr(8) |
| 602 | .iterations(3); |
| 603 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 604 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 605 | tester |
| 606 | .kc(kc) |
| 607 | .kh(ks) |
| 608 | .kw(1) |
| 609 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 610 | tester |
| 611 | .kc(kc) |
| 612 | .kh(1) |
| 613 | .kw(ks) |
| 614 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 615 | } |
| 616 | } |
| 617 | } |
| 618 | } |
| 619 | |
| 620 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_multipass_fulltile_with_x_stride) { |
| 621 | TEST_REQUIRES_ARM_NEON; |
| 622 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 623 | auto tester = AvgPoolMicrokernelTester() |
| 624 | .mr(9) |
| 625 | .qr(8) |
| 626 | .iterations(3); |
| 627 | for (size_t kh = 1; kh <= ks; kh++) { |
| 628 | for (size_t kw = 1; kw <= ks; kw++) { |
| 629 | if (kh * kw == ks) { |
| 630 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 631 | tester |
| 632 | .kh(kh) |
| 633 | .kw(kw) |
| 634 | .kc(kc) |
| 635 | .x_stride(131) |
| 636 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 637 | } |
| 638 | } |
| 639 | } |
| 640 | } |
| 641 | } |
| 642 | } |
| 643 | |
| 644 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_lt_8_twopass_fulltile) { |
| 645 | TEST_REQUIRES_ARM_NEON; |
| 646 | auto tester = AvgPoolMicrokernelTester() |
| 647 | .mr(9) |
| 648 | .qr(8) |
| 649 | .iterations(3); |
| 650 | const size_t ks = tester.mr() + tester.qr(); |
| 651 | for (size_t kh = 1; kh <= ks; kh++) { |
| 652 | for (size_t kw = 1; kw <= ks; kw++) { |
| 653 | if (kh * kw == ks) { |
| 654 | for (size_t kc = 1; kc < 8; kc++) { |
| 655 | tester |
| 656 | .kh(kh) |
| 657 | .kw(kw) |
| 658 | .kc(kc) |
| 659 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 660 | } |
| 661 | } |
| 662 | } |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_lt_8_twopass_subtile) { |
| 667 | TEST_REQUIRES_ARM_NEON; |
| 668 | auto tester = AvgPoolMicrokernelTester() |
| 669 | .mr(9) |
| 670 | .qr(8) |
| 671 | .iterations(3); |
| 672 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 673 | for (size_t kc = 1; kc < 8; kc++) { |
| 674 | tester |
| 675 | .kc(kc) |
| 676 | .kh(ks) |
| 677 | .kw(1) |
| 678 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 679 | tester |
| 680 | .kc(kc) |
| 681 | .kh(1) |
| 682 | .kw(ks) |
| 683 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 684 | } |
| 685 | } |
| 686 | } |
| 687 | |
| 688 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_lt_8_twopass_fulltile_with_x_stride) { |
| 689 | TEST_REQUIRES_ARM_NEON; |
| 690 | auto tester = AvgPoolMicrokernelTester() |
| 691 | .mr(9) |
| 692 | .qr(8) |
| 693 | .iterations(3); |
| 694 | const size_t ks = tester.mr() + tester.qr(); |
| 695 | for (size_t kh = 1; kh <= ks; kh++) { |
| 696 | for (size_t kw = 1; kw <= ks; kw++) { |
| 697 | if (kh * kw == ks) { |
| 698 | for (size_t kc = 1; kc < 8; kc++) { |
| 699 | tester |
| 700 | .kh(kh) |
| 701 | .kw(kw) |
| 702 | .kc(kc) |
| 703 | .x_stride(23) |
| 704 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 705 | } |
| 706 | } |
| 707 | } |
| 708 | } |
| 709 | } |
| 710 | |
| 711 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_lt_8_multipass_fulltile) { |
| 712 | TEST_REQUIRES_ARM_NEON; |
| 713 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 714 | auto tester = AvgPoolMicrokernelTester() |
| 715 | .mr(9) |
| 716 | .qr(8) |
| 717 | .iterations(3); |
| 718 | for (size_t kh = 1; kh <= ks; kh++) { |
| 719 | for (size_t kw = 1; kw <= ks; kw++) { |
| 720 | if (kh * kw == ks) { |
| 721 | for (size_t kc = 1; kc < 8; kc++) { |
| 722 | tester |
| 723 | .kh(kh) |
| 724 | .kw(kw) |
| 725 | .kc(kc) |
| 726 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 727 | } |
| 728 | } |
| 729 | } |
| 730 | } |
| 731 | } |
| 732 | } |
| 733 | |
| 734 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_lt_8_multipass_subtile) { |
| 735 | TEST_REQUIRES_ARM_NEON; |
| 736 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 737 | auto tester = AvgPoolMicrokernelTester() |
| 738 | .mr(9) |
| 739 | .qr(8) |
| 740 | .iterations(3); |
| 741 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 742 | for (size_t kc = 1; kc < 8; kc++) { |
| 743 | tester |
| 744 | .kc(kc) |
| 745 | .kh(ks) |
| 746 | .kw(1) |
| 747 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 748 | tester |
| 749 | .kc(kc) |
| 750 | .kh(1) |
| 751 | .kw(ks) |
| 752 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 753 | } |
| 754 | } |
| 755 | } |
| 756 | } |
| 757 | |
| 758 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_lt_8_multipass_fulltile_with_x_stride) { |
| 759 | TEST_REQUIRES_ARM_NEON; |
| 760 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 761 | auto tester = AvgPoolMicrokernelTester() |
| 762 | .mr(9) |
| 763 | .qr(8) |
| 764 | .iterations(3); |
| 765 | for (size_t kh = 1; kh <= ks; kh++) { |
| 766 | for (size_t kw = 1; kw <= ks; kw++) { |
| 767 | if (kh * kw == ks) { |
| 768 | for (size_t kc = 1; kc < 8; kc++) { |
| 769 | tester |
| 770 | .kh(kh) |
| 771 | .kw(kw) |
| 772 | .kc(kc) |
| 773 | .x_stride(23) |
| 774 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 775 | } |
| 776 | } |
| 777 | } |
| 778 | } |
| 779 | } |
| 780 | } |
| 781 | |
| 782 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_gt_8_twopass_fulltile) { |
| 783 | TEST_REQUIRES_ARM_NEON; |
| 784 | auto tester = AvgPoolMicrokernelTester() |
| 785 | .mr(9) |
| 786 | .qr(8) |
| 787 | .iterations(3); |
| 788 | const size_t ks = tester.mr() + tester.qr(); |
| 789 | for (size_t kh = 1; kh <= ks; kh++) { |
| 790 | for (size_t kw = 1; kw <= ks; kw++) { |
| 791 | if (kh * kw == ks) { |
| 792 | for (size_t kc = 9; kc < 16; kc++) { |
| 793 | tester |
| 794 | .kh(kh) |
| 795 | .kw(kw) |
| 796 | .kc(kc) |
| 797 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 798 | } |
| 799 | } |
| 800 | } |
| 801 | } |
| 802 | } |
| 803 | |
| 804 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_gt_8_twopass_subtile) { |
| 805 | TEST_REQUIRES_ARM_NEON; |
| 806 | auto tester = AvgPoolMicrokernelTester() |
| 807 | .mr(9) |
| 808 | .qr(8) |
| 809 | .iterations(3); |
| 810 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 811 | for (size_t kc = 9; kc < 16; kc++) { |
| 812 | tester |
| 813 | .kc(kc) |
| 814 | .kh(ks) |
| 815 | .kw(1) |
| 816 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 817 | tester |
| 818 | .kc(kc) |
| 819 | .kh(1) |
| 820 | .kw(ks) |
| 821 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 822 | } |
| 823 | } |
| 824 | } |
| 825 | |
| 826 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_gt_8_twopass_fulltile_with_x_stride) { |
| 827 | TEST_REQUIRES_ARM_NEON; |
| 828 | auto tester = AvgPoolMicrokernelTester() |
| 829 | .mr(9) |
| 830 | .qr(8) |
| 831 | .iterations(3); |
| 832 | const size_t ks = tester.mr() + tester.qr(); |
| 833 | for (size_t kh = 1; kh <= ks; kh++) { |
| 834 | for (size_t kw = 1; kw <= ks; kw++) { |
| 835 | if (kh * kw == ks) { |
| 836 | for (size_t kc = 9; kc < 16; kc++) { |
| 837 | tester |
| 838 | .kh(kh) |
| 839 | .kw(kw) |
| 840 | .kc(kc) |
| 841 | .x_stride(23) |
| 842 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 843 | } |
| 844 | } |
| 845 | } |
| 846 | } |
| 847 | } |
| 848 | |
| 849 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_gt_8_multipass_fulltile) { |
| 850 | TEST_REQUIRES_ARM_NEON; |
| 851 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 852 | auto tester = AvgPoolMicrokernelTester() |
| 853 | .mr(9) |
| 854 | .qr(8) |
| 855 | .iterations(3); |
| 856 | for (size_t kh = 1; kh <= ks; kh++) { |
| 857 | for (size_t kw = 1; kw <= ks; kw++) { |
| 858 | if (kh * kw == ks) { |
| 859 | for (size_t kc = 9; kc < 16; kc++) { |
| 860 | tester |
| 861 | .kh(kh) |
| 862 | .kw(kw) |
| 863 | .kc(kc) |
| 864 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 865 | } |
| 866 | } |
| 867 | } |
| 868 | } |
| 869 | } |
| 870 | } |
| 871 | |
| 872 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_gt_8_multipass_subtile) { |
| 873 | TEST_REQUIRES_ARM_NEON; |
| 874 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 875 | auto tester = AvgPoolMicrokernelTester() |
| 876 | .mr(9) |
| 877 | .qr(8) |
| 878 | .iterations(3); |
| 879 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 880 | for (size_t kc = 9; kc < 16; kc++) { |
| 881 | tester |
| 882 | .kc(kc) |
| 883 | .kh(ks) |
| 884 | .kw(1) |
| 885 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 886 | tester |
| 887 | .kc(kc) |
| 888 | .kh(1) |
| 889 | .kw(ks) |
| 890 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 891 | } |
| 892 | } |
| 893 | } |
| 894 | } |
| 895 | |
| 896 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_gt_8_multipass_fulltile_with_x_stride) { |
| 897 | TEST_REQUIRES_ARM_NEON; |
| 898 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 899 | auto tester = AvgPoolMicrokernelTester() |
| 900 | .mr(9) |
| 901 | .qr(8) |
| 902 | .iterations(3); |
| 903 | for (size_t kh = 1; kh <= ks; kh++) { |
| 904 | for (size_t kw = 1; kw <= ks; kw++) { |
| 905 | if (kh * kw == ks) { |
| 906 | for (size_t kc = 9; kc < 16; kc++) { |
| 907 | tester |
| 908 | .kh(kh) |
| 909 | .kw(kw) |
| 910 | .kc(kc) |
| 911 | .x_stride(23) |
| 912 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 913 | } |
| 914 | } |
| 915 | } |
| 916 | } |
| 917 | } |
| 918 | } |
| 919 | |
| 920 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_with_x_scale) { |
| 921 | TEST_REQUIRES_ARM_NEON; |
| 922 | for (size_t n = 1; n <= 5; n += 2) { |
| 923 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 924 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 925 | AvgPoolMicrokernelTester() |
| 926 | .mr(9) |
| 927 | .qr(8) |
| 928 | .n(n) |
| 929 | .kh(5) |
| 930 | .kw(5) |
| 931 | .kc(kc) |
| 932 | .x_scale(x_scale) |
| 933 | .iterations(1) |
| 934 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 935 | } |
| 936 | } |
| 937 | } |
| 938 | } |
| 939 | |
| 940 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_with_x_zero_point) { |
| 941 | TEST_REQUIRES_ARM_NEON; |
| 942 | for (size_t n = 1; n <= 5; n += 2) { |
| 943 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 944 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 945 | AvgPoolMicrokernelTester() |
| 946 | .mr(9) |
| 947 | .qr(8) |
| 948 | .n(n) |
| 949 | .kh(5) |
| 950 | .kw(5) |
| 951 | .kc(kc) |
| 952 | .x_zero_point(uint8_t(x_zero_point)) |
| 953 | .iterations(1) |
| 954 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 955 | } |
| 956 | } |
| 957 | } |
| 958 | } |
| 959 | |
| 960 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_with_y_scale) { |
| 961 | TEST_REQUIRES_ARM_NEON; |
| 962 | for (size_t n = 1; n <= 5; n += 2) { |
| 963 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 964 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 965 | AvgPoolMicrokernelTester() |
| 966 | .mr(9) |
| 967 | .qr(8) |
| 968 | .n(n) |
| 969 | .kh(5) |
| 970 | .kw(5) |
| 971 | .kc(kc) |
| 972 | .y_scale(y_scale) |
| 973 | .iterations(1) |
| 974 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 975 | } |
| 976 | } |
| 977 | } |
| 978 | } |
| 979 | |
| 980 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_with_y_zero_point) { |
| 981 | TEST_REQUIRES_ARM_NEON; |
| 982 | for (size_t n = 1; n <= 5; n += 2) { |
| 983 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 984 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 985 | AvgPoolMicrokernelTester() |
| 986 | .mr(9) |
| 987 | .qr(8) |
| 988 | .n(n) |
| 989 | .kh(5) |
| 990 | .kw(5) |
| 991 | .kc(kc) |
| 992 | .y_zero_point(uint8_t(y_zero_point)) |
| 993 | .iterations(1) |
| 994 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 995 | } |
| 996 | } |
| 997 | } |
| 998 | } |
| 999 | |
| 1000 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_with_qmax) { |
| 1001 | TEST_REQUIRES_ARM_NEON; |
| 1002 | for (size_t n = 1; n <= 5; n += 2) { |
| 1003 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1004 | AvgPoolMicrokernelTester() |
| 1005 | .mr(9) |
| 1006 | .qr(8) |
| 1007 | .n(n) |
| 1008 | .kh(5) |
| 1009 | .kw(5) |
| 1010 | .kc(kc) |
| 1011 | .x_zero_point(128) |
| 1012 | .y_zero_point(128) |
| 1013 | .x_scale(1.0f) |
| 1014 | .y_scale(1.0f) |
| 1015 | .qmax(128) |
| 1016 | .iterations(3) |
| 1017 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 1018 | } |
| 1019 | } |
| 1020 | } |
| 1021 | |
| 1022 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, kc_div_8_with_qmin) { |
| 1023 | TEST_REQUIRES_ARM_NEON; |
| 1024 | for (size_t n = 1; n <= 5; n += 2) { |
| 1025 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1026 | AvgPoolMicrokernelTester() |
| 1027 | .mr(9) |
| 1028 | .qr(8) |
| 1029 | .n(n) |
| 1030 | .kh(5) |
| 1031 | .kw(5) |
| 1032 | .kc(kc) |
| 1033 | .x_zero_point(128) |
| 1034 | .y_zero_point(128) |
| 1035 | .x_scale(1.0f) |
| 1036 | .y_scale(1.0f) |
| 1037 | .qmin(128) |
| 1038 | .iterations(3) |
| 1039 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 1040 | } |
| 1041 | } |
| 1042 | } |
| 1043 | |
| 1044 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, small_n) { |
| 1045 | TEST_REQUIRES_ARM_NEON; |
| 1046 | for (size_t n = 2; n < 5; n++) { |
| 1047 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 1048 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 1049 | AvgPoolMicrokernelTester() |
| 1050 | .mr(9) |
| 1051 | .qr(8) |
| 1052 | .n(n) |
| 1053 | .kh(ks) |
| 1054 | .kw(ks) |
| 1055 | .kc(kc) |
| 1056 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 1057 | } |
| 1058 | } |
| 1059 | } |
| 1060 | } |
| 1061 | |
| 1062 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, small_n_with_x_stride) { |
| 1063 | TEST_REQUIRES_ARM_NEON; |
| 1064 | for (size_t n = 2; n < 5; n++) { |
| 1065 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 1066 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 1067 | AvgPoolMicrokernelTester() |
| 1068 | .mr(9) |
| 1069 | .qr(8) |
| 1070 | .n(n) |
| 1071 | .kh(ks) |
| 1072 | .kw(ks) |
| 1073 | .kc(kc) |
| 1074 | .x_stride(29) |
| 1075 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 1076 | } |
| 1077 | } |
| 1078 | } |
| 1079 | } |
| 1080 | |
| 1081 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, small_n_with_y_stride) { |
| 1082 | TEST_REQUIRES_ARM_NEON; |
| 1083 | for (size_t n = 2; n < 5; n++) { |
| 1084 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 1085 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 1086 | AvgPoolMicrokernelTester() |
| 1087 | .mr(9) |
| 1088 | .qr(8) |
| 1089 | .n(n) |
| 1090 | .kh(ks) |
| 1091 | .kw(ks) |
| 1092 | .kc(kc) |
| 1093 | .y_stride(31) |
| 1094 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 1095 | } |
| 1096 | } |
| 1097 | } |
| 1098 | } |
| 1099 | |
| 1100 | TEST(Q8_AVGPOOL_MP9P8Q__NEON, small_n_with_s) { |
| 1101 | TEST_REQUIRES_ARM_NEON; |
| 1102 | for (size_t n = 2; n < 5; n++) { |
| 1103 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 1104 | for (size_t s = 2; s <= 5; s++) { |
| 1105 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 1106 | AvgPoolMicrokernelTester() |
| 1107 | .mr(9) |
| 1108 | .qr(8) |
| 1109 | .n(n) |
| 1110 | .kh(ks) |
| 1111 | .kw(ks) |
| 1112 | .kc(kc) |
| 1113 | .s(s) |
| 1114 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__neon); |
| 1115 | } |
| 1116 | } |
| 1117 | } |
| 1118 | } |
| 1119 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1120 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1121 | |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1122 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1123 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_eq_8_fulltile) { |
| 1124 | TEST_REQUIRES_X86_SSE2; |
| 1125 | auto tester = AvgPoolMicrokernelTester() |
| 1126 | .mr(9) |
| 1127 | .kc(8); |
| 1128 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 1129 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 1130 | if (kh * kw == tester.mr()) { |
| 1131 | tester |
| 1132 | .kh(kh) |
| 1133 | .kw(kw) |
| 1134 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1135 | } |
| 1136 | } |
| 1137 | } |
| 1138 | } |
| 1139 | |
| 1140 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_eq_8_subtile) { |
| 1141 | TEST_REQUIRES_X86_SSE2; |
| 1142 | auto tester = AvgPoolMicrokernelTester() |
| 1143 | .mr(9) |
| 1144 | .kc(8); |
| 1145 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 1146 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1147 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1148 | if (kh * kw == ks) { |
| 1149 | tester |
| 1150 | .kh(kh) |
| 1151 | .kw(kw) |
| 1152 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1153 | } |
| 1154 | } |
| 1155 | } |
| 1156 | } |
| 1157 | } |
| 1158 | |
| 1159 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_div_8_fulltile) { |
| 1160 | TEST_REQUIRES_X86_SSE2; |
| 1161 | auto tester = AvgPoolMicrokernelTester() |
| 1162 | .mr(9); |
| 1163 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 1164 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 1165 | if (kh * kw == tester.mr()) { |
| 1166 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1167 | tester |
| 1168 | .kh(kh) |
| 1169 | .kw(kw) |
| 1170 | .kc(kc) |
| 1171 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1172 | } |
| 1173 | } |
| 1174 | } |
| 1175 | } |
| 1176 | } |
| 1177 | |
| 1178 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_div_8_subtile) { |
| 1179 | TEST_REQUIRES_X86_SSE2; |
| 1180 | auto tester = AvgPoolMicrokernelTester() |
| 1181 | .mr(9) |
| 1182 | .iterations(3); |
| 1183 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 1184 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1185 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1186 | if (kh * kw == ks) { |
| 1187 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1188 | tester |
| 1189 | .kh(kh) |
| 1190 | .kw(kw) |
| 1191 | .kc(kc) |
| 1192 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1193 | } |
| 1194 | } |
| 1195 | } |
| 1196 | } |
| 1197 | } |
| 1198 | } |
| 1199 | |
| 1200 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_div_8_fulltile_with_x_stride) { |
| 1201 | TEST_REQUIRES_X86_SSE2; |
| 1202 | auto tester = AvgPoolMicrokernelTester() |
| 1203 | .mr(9) |
| 1204 | .iterations(3); |
| 1205 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 1206 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 1207 | if (kh * kw == tester.mr()) { |
| 1208 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1209 | tester |
| 1210 | .kh(kh) |
| 1211 | .kw(kw) |
| 1212 | .kc(kc) |
| 1213 | .x_stride(131) |
| 1214 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1215 | } |
| 1216 | } |
| 1217 | } |
| 1218 | } |
| 1219 | } |
| 1220 | |
| 1221 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_lt_8_fulltile) { |
| 1222 | TEST_REQUIRES_X86_SSE2; |
| 1223 | auto tester = AvgPoolMicrokernelTester() |
| 1224 | .mr(9); |
| 1225 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 1226 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 1227 | if (kh * kw == tester.mr()) { |
| 1228 | for (size_t kc = 1; kc < 8; kc++) { |
| 1229 | tester |
| 1230 | .kh(kh) |
| 1231 | .kw(kw) |
| 1232 | .kc(kc) |
| 1233 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1234 | } |
| 1235 | } |
| 1236 | } |
| 1237 | } |
| 1238 | } |
| 1239 | |
| 1240 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_lt_8_subtile) { |
| 1241 | TEST_REQUIRES_X86_SSE2; |
| 1242 | auto tester = AvgPoolMicrokernelTester() |
| 1243 | .mr(9) |
| 1244 | .iterations(3); |
| 1245 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 1246 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1247 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1248 | if (kh * kw == ks) { |
| 1249 | for (size_t kc = 1; kc < 8; kc++) { |
| 1250 | tester |
| 1251 | .kh(kh) |
| 1252 | .kw(kw) |
| 1253 | .kc(kc) |
| 1254 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1255 | } |
| 1256 | } |
| 1257 | } |
| 1258 | } |
| 1259 | } |
| 1260 | } |
| 1261 | |
| 1262 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_lt_8_fulltile_with_x_stride) { |
| 1263 | TEST_REQUIRES_X86_SSE2; |
| 1264 | auto tester = AvgPoolMicrokernelTester() |
| 1265 | .mr(9) |
| 1266 | .iterations(3); |
| 1267 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 1268 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 1269 | if (kh * kw == tester.mr()) { |
| 1270 | for (size_t kc = 1; kc < 8; kc++) { |
| 1271 | tester |
| 1272 | .kh(kh) |
| 1273 | .kw(kw) |
| 1274 | .kc(kc) |
| 1275 | .x_stride(23) |
| 1276 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1277 | } |
| 1278 | } |
| 1279 | } |
| 1280 | } |
| 1281 | } |
| 1282 | |
| 1283 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_gt_8_fulltile) { |
| 1284 | TEST_REQUIRES_X86_SSE2; |
| 1285 | auto tester = AvgPoolMicrokernelTester() |
| 1286 | .mr(9); |
| 1287 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 1288 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 1289 | if (kh * kw == tester.mr()) { |
| 1290 | for (size_t kc = 9; kc < 16; kc++) { |
| 1291 | tester |
| 1292 | .kh(kh) |
| 1293 | .kw(kw) |
| 1294 | .kc(kc) |
| 1295 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1296 | } |
| 1297 | } |
| 1298 | } |
| 1299 | } |
| 1300 | } |
| 1301 | |
| 1302 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_gt_8_subtile) { |
| 1303 | TEST_REQUIRES_X86_SSE2; |
| 1304 | auto tester = AvgPoolMicrokernelTester() |
| 1305 | .mr(9) |
| 1306 | .iterations(3); |
| 1307 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 1308 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1309 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1310 | if (kh * kw == ks) { |
| 1311 | for (size_t kc = 9; kc < 16; kc++) { |
| 1312 | tester |
| 1313 | .kh(kh) |
| 1314 | .kw(kw) |
| 1315 | .kc(kc) |
| 1316 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1317 | } |
| 1318 | } |
| 1319 | } |
| 1320 | } |
| 1321 | } |
| 1322 | } |
| 1323 | |
| 1324 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_gt_8_fulltile_with_x_stride) { |
| 1325 | TEST_REQUIRES_X86_SSE2; |
| 1326 | auto tester = AvgPoolMicrokernelTester() |
| 1327 | .mr(9) |
| 1328 | .iterations(3); |
| 1329 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 1330 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 1331 | if (kh * kw == tester.mr()) { |
| 1332 | for (size_t kc = 9; kc < 16; kc++) { |
| 1333 | tester |
| 1334 | .kh(kh) |
| 1335 | .kw(kw) |
| 1336 | .kc(kc) |
| 1337 | .x_stride(23) |
| 1338 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1339 | } |
| 1340 | } |
| 1341 | } |
| 1342 | } |
| 1343 | } |
| 1344 | |
| 1345 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_div_8_with_x_scale) { |
| 1346 | TEST_REQUIRES_X86_SSE2; |
| 1347 | for (size_t n = 1; n <= 5; n += 2) { |
| 1348 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1349 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 1350 | AvgPoolMicrokernelTester() |
| 1351 | .mr(9) |
| 1352 | .n(n) |
| 1353 | .kh(3) |
| 1354 | .kw(3) |
| 1355 | .kc(kc) |
| 1356 | .x_scale(x_scale) |
| 1357 | .iterations(2) |
| 1358 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1359 | } |
| 1360 | } |
| 1361 | } |
| 1362 | } |
| 1363 | |
| 1364 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_div_8_with_x_zero_point) { |
| 1365 | TEST_REQUIRES_X86_SSE2; |
| 1366 | for (size_t n = 1; n <= 5; n += 2) { |
| 1367 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1368 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 1369 | AvgPoolMicrokernelTester() |
| 1370 | .mr(9) |
| 1371 | .n(n) |
| 1372 | .kh(3) |
| 1373 | .kw(3) |
| 1374 | .kc(kc) |
| 1375 | .x_zero_point(uint8_t(x_zero_point)) |
| 1376 | .iterations(3) |
| 1377 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1378 | } |
| 1379 | } |
| 1380 | } |
| 1381 | } |
| 1382 | |
| 1383 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_div_8_with_y_scale) { |
| 1384 | TEST_REQUIRES_X86_SSE2; |
| 1385 | for (size_t n = 1; n <= 5; n += 2) { |
| 1386 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1387 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 1388 | AvgPoolMicrokernelTester() |
| 1389 | .mr(9) |
| 1390 | .n(n) |
| 1391 | .kh(3) |
| 1392 | .kw(3) |
| 1393 | .kc(kc) |
| 1394 | .y_scale(y_scale) |
| 1395 | .iterations(2) |
| 1396 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1397 | } |
| 1398 | } |
| 1399 | } |
| 1400 | } |
| 1401 | |
| 1402 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_div_8_with_y_zero_point) { |
| 1403 | TEST_REQUIRES_X86_SSE2; |
| 1404 | for (size_t n = 1; n <= 5; n += 2) { |
| 1405 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1406 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 1407 | AvgPoolMicrokernelTester() |
| 1408 | .mr(9) |
| 1409 | .n(n) |
| 1410 | .kh(3) |
| 1411 | .kw(3) |
| 1412 | .kc(kc) |
| 1413 | .y_zero_point(uint8_t(y_zero_point)) |
| 1414 | .iterations(3) |
| 1415 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1416 | } |
| 1417 | } |
| 1418 | } |
| 1419 | } |
| 1420 | |
| 1421 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_div_8_with_qmax) { |
| 1422 | TEST_REQUIRES_X86_SSE2; |
| 1423 | for (size_t n = 1; n <= 5; n += 2) { |
| 1424 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1425 | AvgPoolMicrokernelTester() |
| 1426 | .mr(9) |
| 1427 | .n(n) |
| 1428 | .kh(3) |
| 1429 | .kw(3) |
| 1430 | .kc(kc) |
| 1431 | .x_zero_point(128) |
| 1432 | .y_zero_point(128) |
| 1433 | .x_scale(1.0f) |
| 1434 | .y_scale(1.0f) |
| 1435 | .qmax(128) |
| 1436 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1437 | } |
| 1438 | } |
| 1439 | } |
| 1440 | |
| 1441 | TEST(Q8_AVGPOOL_UP9__SSE2, kc_div_8_with_qmin) { |
| 1442 | TEST_REQUIRES_X86_SSE2; |
| 1443 | for (size_t n = 1; n <= 5; n += 2) { |
| 1444 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1445 | AvgPoolMicrokernelTester() |
| 1446 | .mr(9) |
| 1447 | .n(n) |
| 1448 | .kh(3) |
| 1449 | .kw(3) |
| 1450 | .kc(kc) |
| 1451 | .x_zero_point(128) |
| 1452 | .y_zero_point(128) |
| 1453 | .x_scale(1.0f) |
| 1454 | .y_scale(1.0f) |
| 1455 | .qmin(128) |
| 1456 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1457 | } |
| 1458 | } |
| 1459 | } |
| 1460 | |
| 1461 | TEST(Q8_AVGPOOL_UP9__SSE2, small_n) { |
| 1462 | TEST_REQUIRES_X86_SSE2; |
| 1463 | for (size_t n = 2; n < 5; n++) { |
| 1464 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 1465 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 1466 | AvgPoolMicrokernelTester() |
| 1467 | .mr(9) |
| 1468 | .n(n) |
| 1469 | .kh(ks) |
| 1470 | .kw(ks) |
| 1471 | .kc(kc) |
| 1472 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1473 | } |
| 1474 | } |
| 1475 | } |
| 1476 | } |
| 1477 | |
| 1478 | TEST(Q8_AVGPOOL_UP9__SSE2, small_n_with_x_stride) { |
| 1479 | TEST_REQUIRES_X86_SSE2; |
| 1480 | for (size_t n = 2; n < 5; n++) { |
| 1481 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 1482 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 1483 | AvgPoolMicrokernelTester() |
| 1484 | .mr(9) |
| 1485 | .n(n) |
| 1486 | .kh(ks) |
| 1487 | .kw(ks) |
| 1488 | .kc(kc) |
| 1489 | .x_stride(29) |
| 1490 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1491 | } |
| 1492 | } |
| 1493 | } |
| 1494 | } |
| 1495 | |
| 1496 | TEST(Q8_AVGPOOL_UP9__SSE2, small_n_with_y_stride) { |
| 1497 | TEST_REQUIRES_X86_SSE2; |
| 1498 | for (size_t n = 2; n < 5; n++) { |
| 1499 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 1500 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 1501 | AvgPoolMicrokernelTester() |
| 1502 | .mr(9) |
| 1503 | .n(n) |
| 1504 | .kh(ks) |
| 1505 | .kw(ks) |
| 1506 | .kc(kc) |
| 1507 | .y_stride(31) |
| 1508 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1509 | } |
| 1510 | } |
| 1511 | } |
| 1512 | } |
| 1513 | |
| 1514 | TEST(Q8_AVGPOOL_UP9__SSE2, small_n_with_s) { |
| 1515 | TEST_REQUIRES_X86_SSE2; |
| 1516 | for (size_t n = 2; n < 5; n++) { |
| 1517 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 1518 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 1519 | for (size_t s = 2; s <= ks; s++) { |
| 1520 | AvgPoolMicrokernelTester() |
| 1521 | .mr(9) |
| 1522 | .n(n) |
| 1523 | .kh(ks) |
| 1524 | .kw(ks) |
| 1525 | .kc(kc) |
| 1526 | .s(s) |
| 1527 | .Test(xnn_q8_avgpool_ukernel_up9__sse2); |
| 1528 | } |
| 1529 | } |
| 1530 | } |
| 1531 | } |
| 1532 | } |
| 1533 | |
| 1534 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_eq_8_twopass_fulltile) { |
| 1535 | TEST_REQUIRES_X86_SSE2; |
| 1536 | auto tester = AvgPoolMicrokernelTester() |
| 1537 | .mr(9) |
| 1538 | .qr(8) |
| 1539 | .kc(8); |
| 1540 | const size_t ks = tester.mr() + tester.qr(); |
| 1541 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1542 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1543 | if (kh * kw == ks) { |
| 1544 | tester |
| 1545 | .kh(kh) |
| 1546 | .kw(kw) |
| 1547 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1548 | } |
| 1549 | } |
| 1550 | } |
| 1551 | } |
| 1552 | |
| 1553 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_eq_8_twopass_subtile) { |
| 1554 | TEST_REQUIRES_X86_SSE2; |
| 1555 | auto tester = AvgPoolMicrokernelTester() |
| 1556 | .mr(9) |
| 1557 | .qr(8) |
| 1558 | .kc(8); |
| 1559 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 1560 | tester |
| 1561 | .kh(ks) |
| 1562 | .kw(1) |
| 1563 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1564 | tester |
| 1565 | .kh(1) |
| 1566 | .kw(ks) |
| 1567 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1568 | } |
| 1569 | } |
| 1570 | |
| 1571 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_eq_8_multipass_fulltile) { |
| 1572 | TEST_REQUIRES_X86_SSE2; |
| 1573 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 1574 | auto tester = AvgPoolMicrokernelTester() |
| 1575 | .mr(9) |
| 1576 | .qr(8) |
| 1577 | .kc(8); |
| 1578 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1579 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1580 | if (kh * kw == ks) { |
| 1581 | tester |
| 1582 | .kh(kh) |
| 1583 | .kw(kw) |
| 1584 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1585 | } |
| 1586 | } |
| 1587 | } |
| 1588 | } |
| 1589 | } |
| 1590 | |
| 1591 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_eq_8_multipass_subtile) { |
| 1592 | TEST_REQUIRES_X86_SSE2; |
| 1593 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 1594 | auto tester = AvgPoolMicrokernelTester() |
| 1595 | .mr(9) |
| 1596 | .qr(8) |
| 1597 | .kc(8); |
| 1598 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 1599 | tester |
| 1600 | .kh(ks) |
| 1601 | .kw(1) |
| 1602 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1603 | tester |
| 1604 | .kh(1) |
| 1605 | .kw(ks) |
| 1606 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1607 | } |
| 1608 | } |
| 1609 | } |
| 1610 | |
| 1611 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_twopass_fulltile) { |
| 1612 | TEST_REQUIRES_X86_SSE2; |
| 1613 | auto tester = AvgPoolMicrokernelTester() |
| 1614 | .mr(9) |
| 1615 | .qr(8) |
| 1616 | .iterations(3); |
| 1617 | const size_t ks = 17; |
| 1618 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1619 | tester |
| 1620 | .kc(kc) |
| 1621 | .kh(ks) |
| 1622 | .kw(1) |
| 1623 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1624 | tester |
| 1625 | .kc(kc) |
| 1626 | .kh(1) |
| 1627 | .kw(ks) |
| 1628 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1629 | } |
| 1630 | } |
| 1631 | |
| 1632 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_twopass_subtile) { |
| 1633 | TEST_REQUIRES_X86_SSE2; |
| 1634 | auto tester = AvgPoolMicrokernelTester() |
| 1635 | .mr(9) |
| 1636 | .qr(8) |
| 1637 | .iterations(3); |
| 1638 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 1639 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1640 | tester |
| 1641 | .kc(kc) |
| 1642 | .kh(ks) |
| 1643 | .kw(1) |
| 1644 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1645 | tester |
| 1646 | .kc(kc) |
| 1647 | .kh(1) |
| 1648 | .kw(ks) |
| 1649 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1650 | } |
| 1651 | } |
| 1652 | } |
| 1653 | |
| 1654 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_twopass_fulltile_with_x_stride) { |
| 1655 | TEST_REQUIRES_X86_SSE2; |
| 1656 | auto tester = AvgPoolMicrokernelTester() |
| 1657 | .mr(9) |
| 1658 | .qr(8) |
| 1659 | .iterations(3); |
| 1660 | const size_t ks = tester.mr() + tester.qr(); |
| 1661 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1662 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1663 | if (kh * kw == ks) { |
| 1664 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1665 | tester |
| 1666 | .kh(kh) |
| 1667 | .kw(kw) |
| 1668 | .kc(kc) |
| 1669 | .x_stride(131) |
| 1670 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1671 | } |
| 1672 | } |
| 1673 | } |
| 1674 | } |
| 1675 | } |
| 1676 | |
| 1677 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_multipass_fulltile) { |
| 1678 | TEST_REQUIRES_X86_SSE2; |
| 1679 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 1680 | auto tester = AvgPoolMicrokernelTester() |
| 1681 | .mr(9) |
| 1682 | .qr(8) |
| 1683 | .iterations(3); |
| 1684 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1685 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1686 | if (kh * kw == ks) { |
| 1687 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1688 | tester |
| 1689 | .kh(kh) |
| 1690 | .kw(kw) |
| 1691 | .kc(kc) |
| 1692 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1693 | } |
| 1694 | } |
| 1695 | } |
| 1696 | } |
| 1697 | } |
| 1698 | } |
| 1699 | |
| 1700 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_multipass_subtile) { |
| 1701 | TEST_REQUIRES_X86_SSE2; |
| 1702 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 1703 | auto tester = AvgPoolMicrokernelTester() |
| 1704 | .mr(9) |
| 1705 | .qr(8) |
| 1706 | .iterations(3); |
| 1707 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 1708 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1709 | tester |
| 1710 | .kc(kc) |
| 1711 | .kh(ks) |
| 1712 | .kw(1) |
| 1713 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1714 | tester |
| 1715 | .kc(kc) |
| 1716 | .kh(1) |
| 1717 | .kw(ks) |
| 1718 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1719 | } |
| 1720 | } |
| 1721 | } |
| 1722 | } |
| 1723 | |
| 1724 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_multipass_fulltile_with_x_stride) { |
| 1725 | TEST_REQUIRES_X86_SSE2; |
| 1726 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 1727 | auto tester = AvgPoolMicrokernelTester() |
| 1728 | .mr(9) |
| 1729 | .qr(8) |
| 1730 | .iterations(3); |
| 1731 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1732 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1733 | if (kh * kw == ks) { |
| 1734 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 1735 | tester |
| 1736 | .kh(kh) |
| 1737 | .kw(kw) |
| 1738 | .kc(kc) |
| 1739 | .x_stride(131) |
| 1740 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1741 | } |
| 1742 | } |
| 1743 | } |
| 1744 | } |
| 1745 | } |
| 1746 | } |
| 1747 | |
| 1748 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_lt_8_twopass_fulltile) { |
| 1749 | TEST_REQUIRES_X86_SSE2; |
| 1750 | auto tester = AvgPoolMicrokernelTester() |
| 1751 | .mr(9) |
| 1752 | .qr(8) |
| 1753 | .iterations(3); |
| 1754 | const size_t ks = tester.mr() + tester.qr(); |
| 1755 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1756 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1757 | if (kh * kw == ks) { |
| 1758 | for (size_t kc = 1; kc < 8; kc++) { |
| 1759 | tester |
| 1760 | .kh(kh) |
| 1761 | .kw(kw) |
| 1762 | .kc(kc) |
| 1763 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1764 | } |
| 1765 | } |
| 1766 | } |
| 1767 | } |
| 1768 | } |
| 1769 | |
| 1770 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_lt_8_twopass_subtile) { |
| 1771 | TEST_REQUIRES_X86_SSE2; |
| 1772 | auto tester = AvgPoolMicrokernelTester() |
| 1773 | .mr(9) |
| 1774 | .qr(8) |
| 1775 | .iterations(3); |
| 1776 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 1777 | for (size_t kc = 1; kc < 8; kc++) { |
| 1778 | tester |
| 1779 | .kc(kc) |
| 1780 | .kh(ks) |
| 1781 | .kw(1) |
| 1782 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1783 | tester |
| 1784 | .kc(kc) |
| 1785 | .kh(1) |
| 1786 | .kw(ks) |
| 1787 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1788 | } |
| 1789 | } |
| 1790 | } |
| 1791 | |
| 1792 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_lt_8_twopass_fulltile_with_x_stride) { |
| 1793 | TEST_REQUIRES_X86_SSE2; |
| 1794 | auto tester = AvgPoolMicrokernelTester() |
| 1795 | .mr(9) |
| 1796 | .qr(8) |
| 1797 | .iterations(3); |
| 1798 | const size_t ks = tester.mr() + tester.qr(); |
| 1799 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1800 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1801 | if (kh * kw == ks) { |
| 1802 | for (size_t kc = 1; kc < 8; kc++) { |
| 1803 | tester |
| 1804 | .kh(kh) |
| 1805 | .kw(kw) |
| 1806 | .kc(kc) |
| 1807 | .x_stride(23) |
| 1808 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1809 | } |
| 1810 | } |
| 1811 | } |
| 1812 | } |
| 1813 | } |
| 1814 | |
| 1815 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_lt_8_multipass_fulltile) { |
| 1816 | TEST_REQUIRES_X86_SSE2; |
| 1817 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 1818 | auto tester = AvgPoolMicrokernelTester() |
| 1819 | .mr(9) |
| 1820 | .qr(8) |
| 1821 | .iterations(3); |
| 1822 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1823 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1824 | if (kh * kw == ks) { |
| 1825 | for (size_t kc = 1; kc < 8; kc++) { |
| 1826 | tester |
| 1827 | .kh(kh) |
| 1828 | .kw(kw) |
| 1829 | .kc(kc) |
| 1830 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1831 | } |
| 1832 | } |
| 1833 | } |
| 1834 | } |
| 1835 | } |
| 1836 | } |
| 1837 | |
| 1838 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_lt_8_multipass_subtile) { |
| 1839 | TEST_REQUIRES_X86_SSE2; |
| 1840 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 1841 | auto tester = AvgPoolMicrokernelTester() |
| 1842 | .mr(9) |
| 1843 | .qr(8) |
| 1844 | .iterations(3); |
| 1845 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 1846 | for (size_t kc = 1; kc < 8; kc++) { |
| 1847 | tester |
| 1848 | .kc(kc) |
| 1849 | .kh(ks) |
| 1850 | .kw(1) |
| 1851 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1852 | tester |
| 1853 | .kc(kc) |
| 1854 | .kh(1) |
| 1855 | .kw(ks) |
| 1856 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1857 | } |
| 1858 | } |
| 1859 | } |
| 1860 | } |
| 1861 | |
| 1862 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_lt_8_multipass_fulltile_with_x_stride) { |
| 1863 | TEST_REQUIRES_X86_SSE2; |
| 1864 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 1865 | auto tester = AvgPoolMicrokernelTester() |
| 1866 | .mr(9) |
| 1867 | .qr(8) |
| 1868 | .iterations(3); |
| 1869 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1870 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1871 | if (kh * kw == ks) { |
| 1872 | for (size_t kc = 1; kc < 8; kc++) { |
| 1873 | tester |
| 1874 | .kh(kh) |
| 1875 | .kw(kw) |
| 1876 | .kc(kc) |
| 1877 | .x_stride(23) |
| 1878 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1879 | } |
| 1880 | } |
| 1881 | } |
| 1882 | } |
| 1883 | } |
| 1884 | } |
| 1885 | |
| 1886 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_gt_8_twopass_fulltile) { |
| 1887 | TEST_REQUIRES_X86_SSE2; |
| 1888 | auto tester = AvgPoolMicrokernelTester() |
| 1889 | .mr(9) |
| 1890 | .qr(8) |
| 1891 | .iterations(3); |
| 1892 | const size_t ks = tester.mr() + tester.qr(); |
| 1893 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1894 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1895 | if (kh * kw == ks) { |
| 1896 | for (size_t kc = 9; kc < 16; kc++) { |
| 1897 | tester |
| 1898 | .kh(kh) |
| 1899 | .kw(kw) |
| 1900 | .kc(kc) |
| 1901 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1902 | } |
| 1903 | } |
| 1904 | } |
| 1905 | } |
| 1906 | } |
| 1907 | |
| 1908 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_gt_8_twopass_subtile) { |
| 1909 | TEST_REQUIRES_X86_SSE2; |
| 1910 | auto tester = AvgPoolMicrokernelTester() |
| 1911 | .mr(9) |
| 1912 | .qr(8) |
| 1913 | .iterations(3); |
| 1914 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 1915 | for (size_t kc = 9; kc < 16; kc++) { |
| 1916 | tester |
| 1917 | .kc(kc) |
| 1918 | .kh(ks) |
| 1919 | .kw(1) |
| 1920 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1921 | tester |
| 1922 | .kc(kc) |
| 1923 | .kh(1) |
| 1924 | .kw(ks) |
| 1925 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1926 | } |
| 1927 | } |
| 1928 | } |
| 1929 | |
| 1930 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_gt_8_twopass_fulltile_with_x_stride) { |
| 1931 | TEST_REQUIRES_X86_SSE2; |
| 1932 | auto tester = AvgPoolMicrokernelTester() |
| 1933 | .mr(9) |
| 1934 | .qr(8) |
| 1935 | .iterations(3); |
| 1936 | const size_t ks = tester.mr() + tester.qr(); |
| 1937 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1938 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1939 | if (kh * kw == ks) { |
| 1940 | for (size_t kc = 9; kc < 16; kc++) { |
| 1941 | tester |
| 1942 | .kh(kh) |
| 1943 | .kw(kw) |
| 1944 | .kc(kc) |
| 1945 | .x_stride(23) |
| 1946 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1947 | } |
| 1948 | } |
| 1949 | } |
| 1950 | } |
| 1951 | } |
| 1952 | |
| 1953 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_gt_8_multipass_fulltile) { |
| 1954 | TEST_REQUIRES_X86_SSE2; |
| 1955 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 1956 | auto tester = AvgPoolMicrokernelTester() |
| 1957 | .mr(9) |
| 1958 | .qr(8) |
| 1959 | .iterations(3); |
| 1960 | for (size_t kh = 1; kh <= ks; kh++) { |
| 1961 | for (size_t kw = 1; kw <= ks; kw++) { |
| 1962 | if (kh * kw == ks) { |
| 1963 | for (size_t kc = 9; kc < 16; kc++) { |
| 1964 | tester |
| 1965 | .kh(kh) |
| 1966 | .kw(kw) |
| 1967 | .kc(kc) |
| 1968 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1969 | } |
| 1970 | } |
| 1971 | } |
| 1972 | } |
| 1973 | } |
| 1974 | } |
| 1975 | |
| 1976 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_gt_8_multipass_subtile) { |
| 1977 | TEST_REQUIRES_X86_SSE2; |
| 1978 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 1979 | auto tester = AvgPoolMicrokernelTester() |
| 1980 | .mr(9) |
| 1981 | .qr(8) |
| 1982 | .iterations(3); |
| 1983 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 1984 | for (size_t kc = 9; kc < 16; kc++) { |
| 1985 | tester |
| 1986 | .kc(kc) |
| 1987 | .kh(ks) |
| 1988 | .kw(1) |
| 1989 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1990 | tester |
| 1991 | .kc(kc) |
| 1992 | .kh(1) |
| 1993 | .kw(ks) |
| 1994 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 1995 | } |
| 1996 | } |
| 1997 | } |
| 1998 | } |
| 1999 | |
| 2000 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_gt_8_multipass_fulltile_with_x_stride) { |
| 2001 | TEST_REQUIRES_X86_SSE2; |
| 2002 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 2003 | auto tester = AvgPoolMicrokernelTester() |
| 2004 | .mr(9) |
| 2005 | .qr(8) |
| 2006 | .iterations(3); |
| 2007 | for (size_t kh = 1; kh <= ks; kh++) { |
| 2008 | for (size_t kw = 1; kw <= ks; kw++) { |
| 2009 | if (kh * kw == ks) { |
| 2010 | for (size_t kc = 9; kc < 16; kc++) { |
| 2011 | tester |
| 2012 | .kh(kh) |
| 2013 | .kw(kw) |
| 2014 | .kc(kc) |
| 2015 | .x_stride(23) |
| 2016 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2017 | } |
| 2018 | } |
| 2019 | } |
| 2020 | } |
| 2021 | } |
| 2022 | } |
| 2023 | |
| 2024 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_with_x_scale) { |
| 2025 | TEST_REQUIRES_X86_SSE2; |
| 2026 | for (size_t n = 1; n <= 5; n += 2) { |
| 2027 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 2028 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 2029 | AvgPoolMicrokernelTester() |
| 2030 | .mr(9) |
| 2031 | .qr(8) |
| 2032 | .n(n) |
| 2033 | .kh(5) |
| 2034 | .kw(5) |
| 2035 | .kc(kc) |
| 2036 | .x_scale(x_scale) |
| 2037 | .iterations(1) |
| 2038 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2039 | } |
| 2040 | } |
| 2041 | } |
| 2042 | } |
| 2043 | |
| 2044 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_with_x_zero_point) { |
| 2045 | TEST_REQUIRES_X86_SSE2; |
| 2046 | for (size_t n = 1; n <= 5; n += 2) { |
| 2047 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 2048 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 2049 | AvgPoolMicrokernelTester() |
| 2050 | .mr(9) |
| 2051 | .qr(8) |
| 2052 | .n(n) |
| 2053 | .kh(5) |
| 2054 | .kw(5) |
| 2055 | .kc(kc) |
| 2056 | .x_zero_point(uint8_t(x_zero_point)) |
| 2057 | .iterations(1) |
| 2058 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2059 | } |
| 2060 | } |
| 2061 | } |
| 2062 | } |
| 2063 | |
| 2064 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_with_y_scale) { |
| 2065 | TEST_REQUIRES_X86_SSE2; |
| 2066 | for (size_t n = 1; n <= 5; n += 2) { |
| 2067 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 2068 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 2069 | AvgPoolMicrokernelTester() |
| 2070 | .mr(9) |
| 2071 | .qr(8) |
| 2072 | .n(n) |
| 2073 | .kh(5) |
| 2074 | .kw(5) |
| 2075 | .kc(kc) |
| 2076 | .y_scale(y_scale) |
| 2077 | .iterations(1) |
| 2078 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2079 | } |
| 2080 | } |
| 2081 | } |
| 2082 | } |
| 2083 | |
| 2084 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_with_y_zero_point) { |
| 2085 | TEST_REQUIRES_X86_SSE2; |
| 2086 | for (size_t n = 1; n <= 5; n += 2) { |
| 2087 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 2088 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 2089 | AvgPoolMicrokernelTester() |
| 2090 | .mr(9) |
| 2091 | .qr(8) |
| 2092 | .n(n) |
| 2093 | .kh(5) |
| 2094 | .kw(5) |
| 2095 | .kc(kc) |
| 2096 | .y_zero_point(uint8_t(y_zero_point)) |
| 2097 | .iterations(1) |
| 2098 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2099 | } |
| 2100 | } |
| 2101 | } |
| 2102 | } |
| 2103 | |
| 2104 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_with_qmax) { |
| 2105 | TEST_REQUIRES_X86_SSE2; |
| 2106 | for (size_t n = 1; n <= 5; n += 2) { |
| 2107 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 2108 | AvgPoolMicrokernelTester() |
| 2109 | .mr(9) |
| 2110 | .qr(8) |
| 2111 | .n(n) |
| 2112 | .kh(5) |
| 2113 | .kw(5) |
| 2114 | .kc(kc) |
| 2115 | .x_zero_point(128) |
| 2116 | .y_zero_point(128) |
| 2117 | .x_scale(1.0f) |
| 2118 | .y_scale(1.0f) |
| 2119 | .qmax(128) |
| 2120 | .iterations(3) |
| 2121 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2122 | } |
| 2123 | } |
| 2124 | } |
| 2125 | |
| 2126 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, kc_div_8_with_qmin) { |
| 2127 | TEST_REQUIRES_X86_SSE2; |
| 2128 | for (size_t n = 1; n <= 5; n += 2) { |
| 2129 | for (size_t kc = 8; kc < 128; kc += 24) { |
| 2130 | AvgPoolMicrokernelTester() |
| 2131 | .mr(9) |
| 2132 | .qr(8) |
| 2133 | .n(n) |
| 2134 | .kh(5) |
| 2135 | .kw(5) |
| 2136 | .kc(kc) |
| 2137 | .x_zero_point(128) |
| 2138 | .y_zero_point(128) |
| 2139 | .x_scale(1.0f) |
| 2140 | .y_scale(1.0f) |
| 2141 | .qmin(128) |
| 2142 | .iterations(3) |
| 2143 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2144 | } |
| 2145 | } |
| 2146 | } |
| 2147 | |
| 2148 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, small_n) { |
| 2149 | TEST_REQUIRES_X86_SSE2; |
| 2150 | for (size_t n = 2; n < 5; n++) { |
| 2151 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 2152 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 2153 | AvgPoolMicrokernelTester() |
| 2154 | .mr(9) |
| 2155 | .qr(8) |
| 2156 | .n(n) |
| 2157 | .kh(ks) |
| 2158 | .kw(ks) |
| 2159 | .kc(kc) |
| 2160 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2161 | } |
| 2162 | } |
| 2163 | } |
| 2164 | } |
| 2165 | |
| 2166 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, small_n_with_x_stride) { |
| 2167 | TEST_REQUIRES_X86_SSE2; |
| 2168 | for (size_t n = 2; n < 5; n++) { |
| 2169 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 2170 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 2171 | AvgPoolMicrokernelTester() |
| 2172 | .mr(9) |
| 2173 | .qr(8) |
| 2174 | .n(n) |
| 2175 | .kh(ks) |
| 2176 | .kw(ks) |
| 2177 | .kc(kc) |
| 2178 | .x_stride(29) |
| 2179 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2180 | } |
| 2181 | } |
| 2182 | } |
| 2183 | } |
| 2184 | |
| 2185 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, small_n_with_y_stride) { |
| 2186 | TEST_REQUIRES_X86_SSE2; |
| 2187 | for (size_t n = 2; n < 5; n++) { |
| 2188 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 2189 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 2190 | AvgPoolMicrokernelTester() |
| 2191 | .mr(9) |
| 2192 | .qr(8) |
| 2193 | .n(n) |
| 2194 | .kh(ks) |
| 2195 | .kw(ks) |
| 2196 | .kc(kc) |
| 2197 | .y_stride(31) |
| 2198 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2199 | } |
| 2200 | } |
| 2201 | } |
| 2202 | } |
| 2203 | |
| 2204 | TEST(Q8_AVGPOOL_MP9P8Q__SSE2, small_n_with_s) { |
| 2205 | TEST_REQUIRES_X86_SSE2; |
| 2206 | for (size_t n = 2; n < 5; n++) { |
| 2207 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 2208 | for (size_t s = 2; s <= 5; s++) { |
| 2209 | for (size_t kc = 8; kc < 25; kc += 5) { |
| 2210 | AvgPoolMicrokernelTester() |
| 2211 | .mr(9) |
| 2212 | .qr(8) |
| 2213 | .n(n) |
| 2214 | .kh(ks) |
| 2215 | .kw(ks) |
| 2216 | .kc(kc) |
| 2217 | .s(s) |
| 2218 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__sse2); |
| 2219 | } |
| 2220 | } |
| 2221 | } |
| 2222 | } |
| 2223 | } |
| Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 2224 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2225 | |
| 2226 | TEST(Q8_AVGPOOL_UP9__SCALAR, kc_eq_1_fulltile) { |
| 2227 | auto tester = AvgPoolMicrokernelTester() |
| 2228 | .mr(9) |
| 2229 | .kc(1); |
| 2230 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 2231 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 2232 | if (kh * kw == tester.mr()) { |
| 2233 | tester |
| 2234 | .kh(kh) |
| 2235 | .kw(kw) |
| 2236 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2237 | } |
| 2238 | } |
| 2239 | } |
| 2240 | } |
| 2241 | |
| 2242 | TEST(Q8_AVGPOOL_UP9__SCALAR, kc_eq_1_subtile) { |
| 2243 | auto tester = AvgPoolMicrokernelTester() |
| 2244 | .mr(9) |
| 2245 | .kc(1); |
| 2246 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 2247 | for (size_t kh = 1; kh <= ks; kh++) { |
| 2248 | for (size_t kw = 1; kw <= ks; kw++) { |
| 2249 | if (kh * kw == ks) { |
| 2250 | tester |
| 2251 | .kh(kh) |
| 2252 | .kw(kw) |
| 2253 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2254 | } |
| 2255 | } |
| 2256 | } |
| 2257 | } |
| 2258 | } |
| 2259 | |
| 2260 | TEST(Q8_AVGPOOL_UP9__SCALAR, kc_gt_1_fulltile) { |
| 2261 | auto tester = AvgPoolMicrokernelTester() |
| 2262 | .mr(9); |
| 2263 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 2264 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 2265 | if (kh * kw == tester.mr()) { |
| 2266 | for (size_t kc = 2; kc < 8; kc++) { |
| 2267 | tester |
| 2268 | .kh(kh) |
| 2269 | .kw(kw) |
| 2270 | .kc(kc) |
| 2271 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2272 | } |
| 2273 | } |
| 2274 | } |
| 2275 | } |
| 2276 | } |
| 2277 | |
| 2278 | TEST(Q8_AVGPOOL_UP9__SCALAR, kc_gt_1_subtile) { |
| 2279 | auto tester = AvgPoolMicrokernelTester() |
| 2280 | .mr(9) |
| 2281 | .iterations(3); |
| 2282 | for (size_t ks = 2; ks < tester.mr(); ks++) { |
| 2283 | for (size_t kh = 1; kh <= ks; kh++) { |
| 2284 | for (size_t kw = 1; kw <= ks; kw++) { |
| 2285 | if (kh * kw == ks) { |
| 2286 | for (size_t kc = 2; kc < 8; kc++) { |
| 2287 | tester |
| 2288 | .kh(kh) |
| 2289 | .kw(kw) |
| 2290 | .kc(kc) |
| 2291 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2292 | } |
| 2293 | } |
| 2294 | } |
| 2295 | } |
| 2296 | } |
| 2297 | } |
| 2298 | |
| 2299 | TEST(Q8_AVGPOOL_UP9__SCALAR, kc_gt_1_fulltile_with_x_stride) { |
| 2300 | auto tester = AvgPoolMicrokernelTester() |
| 2301 | .mr(9) |
| 2302 | .iterations(3); |
| 2303 | for (size_t kh = 1; kh <= tester.mr(); kh++) { |
| 2304 | for (size_t kw = 1; kw <= tester.mr(); kw++) { |
| 2305 | if (kh * kw == tester.mr()) { |
| 2306 | for (size_t kc = 2; kc < 8; kc++) { |
| 2307 | tester |
| 2308 | .kh(kh) |
| 2309 | .kw(kw) |
| 2310 | .kc(kc) |
| 2311 | .x_stride(23) |
| 2312 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2313 | } |
| 2314 | } |
| 2315 | } |
| 2316 | } |
| 2317 | } |
| 2318 | |
| 2319 | TEST(Q8_AVGPOOL_UP9__SCALAR, x_scale) { |
| 2320 | for (size_t n = 1; n <= 5; n += 2) { |
| 2321 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2322 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 2323 | AvgPoolMicrokernelTester() |
| 2324 | .mr(9) |
| 2325 | .n(n) |
| 2326 | .kh(3) |
| 2327 | .kw(3) |
| 2328 | .kc(kc) |
| 2329 | .x_scale(x_scale) |
| 2330 | .iterations(2) |
| 2331 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2332 | } |
| 2333 | } |
| 2334 | } |
| 2335 | } |
| 2336 | |
| 2337 | TEST(Q8_AVGPOOL_UP9__SCALAR, x_zero_point) { |
| 2338 | for (size_t n = 1; n <= 5; n += 2) { |
| 2339 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2340 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 2341 | AvgPoolMicrokernelTester() |
| 2342 | .mr(9) |
| 2343 | .n(n) |
| 2344 | .kh(3) |
| 2345 | .kw(3) |
| 2346 | .kc(kc) |
| 2347 | .x_zero_point(uint8_t(x_zero_point)) |
| 2348 | .iterations(3) |
| 2349 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2350 | } |
| 2351 | } |
| 2352 | } |
| 2353 | } |
| 2354 | |
| 2355 | TEST(Q8_AVGPOOL_UP9__SCALAR, y_scale) { |
| 2356 | for (size_t n = 1; n <= 5; n += 2) { |
| 2357 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2358 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 2359 | AvgPoolMicrokernelTester() |
| 2360 | .mr(9) |
| 2361 | .n(n) |
| 2362 | .kh(3) |
| 2363 | .kw(3) |
| 2364 | .kc(kc) |
| 2365 | .y_scale(y_scale) |
| 2366 | .iterations(2) |
| 2367 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2368 | } |
| 2369 | } |
| 2370 | } |
| 2371 | } |
| 2372 | |
| 2373 | TEST(Q8_AVGPOOL_UP9__SCALAR, y_zero_point) { |
| 2374 | for (size_t n = 1; n <= 5; n += 2) { |
| 2375 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2376 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 2377 | AvgPoolMicrokernelTester() |
| 2378 | .mr(9) |
| 2379 | .n(n) |
| 2380 | .kh(3) |
| 2381 | .kw(3) |
| 2382 | .kc(kc) |
| 2383 | .y_zero_point(uint8_t(y_zero_point)) |
| 2384 | .iterations(3) |
| 2385 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2386 | } |
| 2387 | } |
| 2388 | } |
| 2389 | } |
| 2390 | |
| 2391 | TEST(Q8_AVGPOOL_UP9__SCALAR, qmax) { |
| 2392 | for (size_t n = 1; n <= 5; n += 2) { |
| 2393 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2394 | AvgPoolMicrokernelTester() |
| 2395 | .mr(9) |
| 2396 | .n(n) |
| 2397 | .kh(3) |
| 2398 | .kw(3) |
| 2399 | .kc(kc) |
| 2400 | .x_zero_point(128) |
| 2401 | .y_zero_point(128) |
| 2402 | .x_scale(1.0f) |
| 2403 | .y_scale(1.0f) |
| 2404 | .qmax(128) |
| 2405 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2406 | } |
| 2407 | } |
| 2408 | } |
| 2409 | |
| 2410 | TEST(Q8_AVGPOOL_UP9__SCALAR, qmin) { |
| 2411 | for (size_t n = 1; n <= 5; n += 2) { |
| 2412 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2413 | AvgPoolMicrokernelTester() |
| 2414 | .mr(9) |
| 2415 | .n(n) |
| 2416 | .kh(3) |
| 2417 | .kw(3) |
| 2418 | .kc(kc) |
| 2419 | .x_zero_point(128) |
| 2420 | .y_zero_point(128) |
| 2421 | .x_scale(1.0f) |
| 2422 | .y_scale(1.0f) |
| 2423 | .qmin(128) |
| 2424 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2425 | } |
| 2426 | } |
| 2427 | } |
| 2428 | |
| 2429 | TEST(Q8_AVGPOOL_UP9__SCALAR, small_n) { |
| 2430 | for (size_t n = 2; n < 5; n++) { |
| 2431 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 2432 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2433 | AvgPoolMicrokernelTester() |
| 2434 | .mr(9) |
| 2435 | .n(n) |
| 2436 | .kh(ks) |
| 2437 | .kw(ks) |
| 2438 | .kc(kc) |
| 2439 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2440 | } |
| 2441 | } |
| 2442 | } |
| 2443 | } |
| 2444 | |
| 2445 | TEST(Q8_AVGPOOL_UP9__SCALAR, small_n_with_x_stride) { |
| 2446 | for (size_t n = 2; n < 5; n++) { |
| 2447 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 2448 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2449 | AvgPoolMicrokernelTester() |
| 2450 | .mr(9) |
| 2451 | .n(n) |
| 2452 | .kh(ks) |
| 2453 | .kw(ks) |
| 2454 | .kc(kc) |
| 2455 | .x_stride(29) |
| 2456 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2457 | } |
| 2458 | } |
| 2459 | } |
| 2460 | } |
| 2461 | |
| 2462 | TEST(Q8_AVGPOOL_UP9__SCALAR, small_n_with_y_stride) { |
| 2463 | for (size_t n = 2; n < 5; n++) { |
| 2464 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 2465 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2466 | AvgPoolMicrokernelTester() |
| 2467 | .mr(9) |
| 2468 | .n(n) |
| 2469 | .kh(ks) |
| 2470 | .kw(ks) |
| 2471 | .kc(kc) |
| 2472 | .y_stride(31) |
| 2473 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2474 | } |
| 2475 | } |
| 2476 | } |
| 2477 | } |
| 2478 | |
| 2479 | TEST(Q8_AVGPOOL_UP9__SCALAR, small_n_with_s) { |
| 2480 | for (size_t n = 2; n < 5; n++) { |
| 2481 | for (size_t ks : std::vector<size_t>{{2, 3}}) { |
| 2482 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2483 | for (size_t s = 2; s <= ks; s++) { |
| 2484 | AvgPoolMicrokernelTester() |
| 2485 | .mr(9) |
| 2486 | .n(n) |
| 2487 | .kh(ks) |
| 2488 | .kw(ks) |
| 2489 | .kc(kc) |
| 2490 | .s(s) |
| 2491 | .Test(xnn_q8_avgpool_ukernel_up9__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2492 | } |
| 2493 | } |
| 2494 | } |
| 2495 | } |
| 2496 | } |
| 2497 | |
| 2498 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_eq_1_twopass_fulltile) { |
| 2499 | auto tester = AvgPoolMicrokernelTester() |
| 2500 | .mr(9) |
| 2501 | .qr(8) |
| 2502 | .kc(1); |
| 2503 | const size_t ks = tester.mr() + tester.qr(); |
| 2504 | for (size_t kh = 1; kh <= ks; kh++) { |
| 2505 | for (size_t kw = 1; kw <= ks; kw++) { |
| 2506 | if (kh * kw == ks) { |
| 2507 | tester |
| 2508 | .kh(kh) |
| 2509 | .kw(kw) |
| 2510 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2511 | } |
| 2512 | } |
| 2513 | } |
| 2514 | } |
| 2515 | |
| 2516 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_eq_1_twopass_subtile) { |
| 2517 | auto tester = AvgPoolMicrokernelTester() |
| 2518 | .mr(9) |
| 2519 | .qr(8) |
| 2520 | .kc(1); |
| 2521 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 2522 | tester |
| 2523 | .kh(ks) |
| 2524 | .kw(1) |
| 2525 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2526 | tester |
| 2527 | .kh(1) |
| 2528 | .kw(ks) |
| 2529 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2530 | } |
| 2531 | } |
| 2532 | |
| 2533 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_eq_1_multipass_fulltile) { |
| 2534 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 2535 | auto tester = AvgPoolMicrokernelTester() |
| 2536 | .mr(9) |
| 2537 | .qr(8) |
| 2538 | .kc(1); |
| 2539 | for (size_t kh = 1; kh <= ks; kh++) { |
| 2540 | for (size_t kw = 1; kw <= ks; kw++) { |
| 2541 | if (kh * kw == ks) { |
| 2542 | tester |
| 2543 | .kh(kh) |
| 2544 | .kw(kw) |
| 2545 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2546 | } |
| 2547 | } |
| 2548 | } |
| 2549 | } |
| 2550 | } |
| 2551 | |
| 2552 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_eq_1_multipass_subtile) { |
| 2553 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 2554 | auto tester = AvgPoolMicrokernelTester() |
| 2555 | .mr(9) |
| 2556 | .qr(8) |
| 2557 | .kc(1); |
| 2558 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 2559 | tester |
| 2560 | .kh(ks) |
| 2561 | .kw(1) |
| 2562 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2563 | tester |
| 2564 | .kh(1) |
| 2565 | .kw(ks) |
| 2566 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2567 | } |
| 2568 | } |
| 2569 | } |
| 2570 | |
| 2571 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_gt_1_twopass_fulltile) { |
| 2572 | auto tester = AvgPoolMicrokernelTester() |
| 2573 | .mr(9) |
| 2574 | .qr(8) |
| 2575 | .iterations(3); |
| 2576 | const size_t ks = tester.mr() + tester.qr(); |
| 2577 | for (size_t kh = 1; kh <= ks; kh++) { |
| 2578 | for (size_t kw = 1; kw <= ks; kw++) { |
| 2579 | if (kh * kw == ks) { |
| 2580 | for (size_t kc = 2; kc < 8; kc++) { |
| 2581 | tester |
| 2582 | .kh(kh) |
| 2583 | .kw(kw) |
| 2584 | .kc(kc) |
| 2585 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2586 | } |
| 2587 | } |
| 2588 | } |
| 2589 | } |
| 2590 | } |
| 2591 | |
| 2592 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_gt_1_twopass_subtile) { |
| 2593 | auto tester = AvgPoolMicrokernelTester() |
| 2594 | .mr(9) |
| 2595 | .qr(8) |
| 2596 | .iterations(3); |
| 2597 | for (size_t ks = 10; ks < tester.mr() + tester.qr(); ks++) { |
| 2598 | for (size_t kc = 2; kc < 8; kc++) { |
| 2599 | tester |
| 2600 | .kc(kc) |
| 2601 | .kh(ks) |
| 2602 | .kw(1) |
| 2603 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2604 | tester |
| 2605 | .kc(kc) |
| 2606 | .kh(1) |
| 2607 | .kw(ks) |
| 2608 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2609 | } |
| 2610 | } |
| 2611 | } |
| 2612 | |
| 2613 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_gt_1_twopass_fulltile_with_x_stride) { |
| 2614 | auto tester = AvgPoolMicrokernelTester() |
| 2615 | .mr(9) |
| 2616 | .qr(8) |
| 2617 | .iterations(3); |
| 2618 | const size_t ks = tester.mr() + tester.qr(); |
| 2619 | for (size_t kh = 1; kh <= ks; kh++) { |
| 2620 | for (size_t kw = 1; kw <= ks; kw++) { |
| 2621 | if (kh * kw == ks) { |
| 2622 | for (size_t kc = 2; kc < 8; kc++) { |
| 2623 | tester |
| 2624 | .kh(kh) |
| 2625 | .kw(kw) |
| 2626 | .kc(kc) |
| 2627 | .x_stride(23) |
| 2628 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2629 | } |
| 2630 | } |
| 2631 | } |
| 2632 | } |
| 2633 | } |
| 2634 | |
| 2635 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_gt_1_multipass_fulltile) { |
| 2636 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 2637 | auto tester = AvgPoolMicrokernelTester() |
| 2638 | .mr(9) |
| 2639 | .qr(8) |
| 2640 | .iterations(3); |
| 2641 | for (size_t kh = 1; kh <= ks; kh++) { |
| 2642 | for (size_t kw = 1; kw <= ks; kw++) { |
| 2643 | if (kh * kw == ks) { |
| 2644 | for (size_t kc = 2; kc < 8; kc++) { |
| 2645 | tester |
| 2646 | .kh(kh) |
| 2647 | .kw(kw) |
| 2648 | .kc(kc) |
| 2649 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2650 | } |
| 2651 | } |
| 2652 | } |
| 2653 | } |
| 2654 | } |
| 2655 | } |
| 2656 | |
| 2657 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_gt_1_multipass_subtile) { |
| 2658 | for (size_t ks_max : std::vector<size_t>{{25, 49}}) { |
| 2659 | auto tester = AvgPoolMicrokernelTester() |
| 2660 | .mr(9) |
| 2661 | .qr(8) |
| 2662 | .iterations(3); |
| 2663 | for (size_t ks = ks_max - tester.qr() + 1; ks < ks_max; ks++) { |
| 2664 | for (size_t kc = 2; kc < 8; kc++) { |
| 2665 | tester |
| 2666 | .kc(kc) |
| 2667 | .kh(ks) |
| 2668 | .kw(1) |
| 2669 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2670 | tester |
| 2671 | .kc(kc) |
| 2672 | .kh(1) |
| 2673 | .kw(ks) |
| 2674 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2675 | } |
| 2676 | } |
| 2677 | } |
| 2678 | } |
| 2679 | |
| 2680 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, kc_gt_1_multipass_fulltile_with_x_stride) { |
| 2681 | for (size_t ks : std::vector<size_t>{{25, 49}}) { |
| 2682 | auto tester = AvgPoolMicrokernelTester() |
| 2683 | .mr(9) |
| 2684 | .qr(8) |
| 2685 | .iterations(3); |
| 2686 | for (size_t kh = 1; kh <= ks; kh++) { |
| 2687 | for (size_t kw = 1; kw <= ks; kw++) { |
| 2688 | if (kh * kw == ks) { |
| 2689 | for (size_t kc = 2; kc < 8; kc++) { |
| 2690 | tester |
| 2691 | .kh(kh) |
| 2692 | .kw(kw) |
| 2693 | .kc(kc) |
| 2694 | .x_stride(23) |
| 2695 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2696 | } |
| 2697 | } |
| 2698 | } |
| 2699 | } |
| 2700 | } |
| 2701 | } |
| 2702 | |
| 2703 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, x_scale) { |
| 2704 | for (size_t n = 1; n <= 5; n += 2) { |
| 2705 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2706 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 2707 | AvgPoolMicrokernelTester() |
| 2708 | .mr(9) |
| 2709 | .qr(8) |
| 2710 | .n(n) |
| 2711 | .kh(5) |
| 2712 | .kw(5) |
| 2713 | .kc(kc) |
| 2714 | .x_scale(x_scale) |
| 2715 | .iterations(1) |
| 2716 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2717 | } |
| 2718 | } |
| 2719 | } |
| 2720 | } |
| 2721 | |
| 2722 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, x_zero_point) { |
| 2723 | for (size_t n = 1; n <= 5; n += 2) { |
| 2724 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2725 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 2726 | AvgPoolMicrokernelTester() |
| 2727 | .mr(9) |
| 2728 | .qr(8) |
| 2729 | .n(n) |
| 2730 | .kh(5) |
| 2731 | .kw(5) |
| 2732 | .kc(kc) |
| 2733 | .x_zero_point(uint8_t(x_zero_point)) |
| 2734 | .iterations(1) |
| 2735 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2736 | } |
| 2737 | } |
| 2738 | } |
| 2739 | } |
| 2740 | |
| 2741 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, y_scale) { |
| 2742 | for (size_t n = 1; n <= 5; n += 2) { |
| 2743 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2744 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 2745 | AvgPoolMicrokernelTester() |
| 2746 | .mr(9) |
| 2747 | .qr(8) |
| 2748 | .n(n) |
| 2749 | .kh(5) |
| 2750 | .kw(5) |
| 2751 | .kc(kc) |
| 2752 | .y_scale(y_scale) |
| 2753 | .iterations(1) |
| 2754 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2755 | } |
| 2756 | } |
| 2757 | } |
| 2758 | } |
| 2759 | |
| 2760 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, y_zero_point) { |
| 2761 | for (size_t n = 1; n <= 5; n += 2) { |
| 2762 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2763 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 2764 | AvgPoolMicrokernelTester() |
| 2765 | .mr(9) |
| 2766 | .qr(8) |
| 2767 | .n(n) |
| 2768 | .kh(5) |
| 2769 | .kw(5) |
| 2770 | .kc(kc) |
| 2771 | .y_zero_point(uint8_t(y_zero_point)) |
| 2772 | .iterations(1) |
| 2773 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2774 | } |
| 2775 | } |
| 2776 | } |
| 2777 | } |
| 2778 | |
| 2779 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, qmax) { |
| 2780 | for (size_t n = 1; n <= 5; n += 2) { |
| 2781 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2782 | AvgPoolMicrokernelTester() |
| 2783 | .mr(9) |
| 2784 | .qr(8) |
| 2785 | .n(n) |
| 2786 | .kh(5) |
| 2787 | .kw(5) |
| 2788 | .kc(kc) |
| 2789 | .x_zero_point(128) |
| 2790 | .y_zero_point(128) |
| 2791 | .x_scale(1.0f) |
| 2792 | .y_scale(1.0f) |
| 2793 | .qmax(128) |
| 2794 | .iterations(3) |
| 2795 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2796 | } |
| 2797 | } |
| 2798 | } |
| 2799 | |
| 2800 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, qmin) { |
| 2801 | for (size_t n = 1; n <= 5; n += 2) { |
| 2802 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2803 | AvgPoolMicrokernelTester() |
| 2804 | .mr(9) |
| 2805 | .qr(8) |
| 2806 | .n(n) |
| 2807 | .kh(5) |
| 2808 | .kw(5) |
| 2809 | .kc(kc) |
| 2810 | .x_zero_point(128) |
| 2811 | .y_zero_point(128) |
| 2812 | .x_scale(1.0f) |
| 2813 | .y_scale(1.0f) |
| 2814 | .qmin(128) |
| 2815 | .iterations(3) |
| 2816 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2817 | } |
| 2818 | } |
| 2819 | } |
| 2820 | |
| 2821 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, small_n) { |
| 2822 | for (size_t n = 2; n < 5; n++) { |
| 2823 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 2824 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2825 | AvgPoolMicrokernelTester() |
| 2826 | .mr(9) |
| 2827 | .qr(8) |
| 2828 | .n(n) |
| 2829 | .kh(ks) |
| 2830 | .kw(ks) |
| 2831 | .kc(kc) |
| 2832 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2833 | } |
| 2834 | } |
| 2835 | } |
| 2836 | } |
| 2837 | |
| 2838 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, small_n_with_x_stride) { |
| 2839 | for (size_t n = 2; n < 5; n++) { |
| 2840 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 2841 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2842 | AvgPoolMicrokernelTester() |
| 2843 | .mr(9) |
| 2844 | .qr(8) |
| 2845 | .n(n) |
| 2846 | .kh(ks) |
| 2847 | .kw(ks) |
| 2848 | .kc(kc) |
| 2849 | .x_stride(29) |
| 2850 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2851 | } |
| 2852 | } |
| 2853 | } |
| 2854 | } |
| 2855 | |
| 2856 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, small_n_with_y_stride) { |
| 2857 | for (size_t n = 2; n < 5; n++) { |
| 2858 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 2859 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2860 | AvgPoolMicrokernelTester() |
| 2861 | .mr(9) |
| 2862 | .qr(8) |
| 2863 | .n(n) |
| 2864 | .kh(ks) |
| 2865 | .kw(ks) |
| 2866 | .kc(kc) |
| 2867 | .y_stride(31) |
| 2868 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2869 | } |
| 2870 | } |
| 2871 | } |
| 2872 | } |
| 2873 | |
| 2874 | TEST(Q8_AVGPOOL_MP9P8Q__SCALAR, small_n_with_s) { |
| 2875 | for (size_t n = 2; n < 5; n++) { |
| 2876 | for (size_t ks : std::vector<size_t>{{5, 7}}) { |
| 2877 | for (size_t s = 2; s <= 5; s++) { |
| 2878 | for (size_t kc = 1; kc < 8; kc += 3) { |
| 2879 | AvgPoolMicrokernelTester() |
| 2880 | .mr(9) |
| 2881 | .qr(8) |
| 2882 | .n(n) |
| 2883 | .kh(ks) |
| 2884 | .kw(ks) |
| 2885 | .kc(kc) |
| 2886 | .s(s) |
| 2887 | .Test(xnn_q8_avgpool_ukernel_mp9p8q__scalar, AvgPoolMicrokernelTester::Variant::Scalar); |
| 2888 | } |
| 2889 | } |
| 2890 | } |
| 2891 | } |
| 2892 | } |