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/gavgpool.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 15 | #include "gavgpool-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_GAVGPOOL_UP7__NEON, n_eq_8_fulltile) { |
| 20 | TEST_REQUIRES_ARM_NEON; |
| 21 | GAvgPoolMicrokernelTester() |
| 22 | .m(7) |
| 23 | .n(8) |
| 24 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 25 | } |
| 26 | |
| 27 | TEST(Q8_GAVGPOOL_UP7__NEON, n_eq_8_subtile) { |
| 28 | TEST_REQUIRES_ARM_NEON; |
| 29 | for (size_t m = 1; m < 7; m++) { |
| 30 | GAvgPoolMicrokernelTester() |
| 31 | .m(m) |
| 32 | .n(8) |
| 33 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 34 | } |
| 35 | } |
| 36 | |
| 37 | TEST(Q8_GAVGPOOL_UP7__NEON, n_eq_8_fulltile_with_x_stride) { |
| 38 | TEST_REQUIRES_ARM_NEON; |
| 39 | GAvgPoolMicrokernelTester() |
| 40 | .m(7) |
| 41 | .n(8) |
| 42 | .x_stride(11) |
| 43 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 44 | } |
| 45 | |
| 46 | TEST(Q8_GAVGPOOL_UP7__NEON, n_eq_8_fulltile_with_x_scale) { |
| 47 | TEST_REQUIRES_ARM_NEON; |
| 48 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 49 | GAvgPoolMicrokernelTester() |
| 50 | .m(7) |
| 51 | .n(8) |
| 52 | .x_scale(x_scale) |
| 53 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 54 | } |
| 55 | } |
| 56 | |
| 57 | TEST(Q8_GAVGPOOL_UP7__NEON, n_eq_8_fulltile_with_x_zero_point) { |
| 58 | TEST_REQUIRES_ARM_NEON; |
| 59 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 60 | GAvgPoolMicrokernelTester() |
| 61 | .m(7) |
| 62 | .n(8) |
| 63 | .x_zero_point(x_zero_point) |
| 64 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | TEST(Q8_GAVGPOOL_UP7__NEON, n_eq_8_fulltile_with_y_scale) { |
| 69 | TEST_REQUIRES_ARM_NEON; |
| 70 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 71 | GAvgPoolMicrokernelTester() |
| 72 | .m(7) |
| 73 | .n(8) |
| 74 | .y_scale(y_scale) |
| 75 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 76 | } |
| 77 | } |
| 78 | |
| 79 | TEST(Q8_GAVGPOOL_UP7__NEON, n_eq_8_fulltile_with_y_zero_point) { |
| 80 | TEST_REQUIRES_ARM_NEON; |
| 81 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 82 | GAvgPoolMicrokernelTester() |
| 83 | .m(7) |
| 84 | .n(8) |
| 85 | .y_zero_point(y_zero_point) |
| 86 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | TEST(Q8_GAVGPOOL_UP7__NEON, n_eq_8_fulltile_with_qmax) { |
| 91 | TEST_REQUIRES_ARM_NEON; |
| 92 | GAvgPoolMicrokernelTester() |
| 93 | .m(7) |
| 94 | .n(8) |
| 95 | .x_zero_point(128) |
| 96 | .y_zero_point(128) |
| 97 | .x_scale(1.0f) |
| 98 | .y_scale(1.0f) |
| 99 | .qmax(128) |
| 100 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 101 | } |
| 102 | |
| 103 | TEST(Q8_GAVGPOOL_UP7__NEON, n_eq_8_fulltile_with_qmin) { |
| 104 | TEST_REQUIRES_ARM_NEON; |
| 105 | GAvgPoolMicrokernelTester() |
| 106 | .m(7) |
| 107 | .n(8) |
| 108 | .x_zero_point(128) |
| 109 | .y_zero_point(128) |
| 110 | .x_scale(1.0f) |
| 111 | .y_scale(1.0f) |
| 112 | .qmin(128) |
| 113 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 114 | } |
| 115 | |
| 116 | TEST(Q8_GAVGPOOL_UP7__NEON, n_div_8_fulltile) { |
| 117 | TEST_REQUIRES_ARM_NEON; |
| 118 | for (size_t n = 8; n < 128; n += 24) { |
| 119 | GAvgPoolMicrokernelTester() |
| 120 | .m(7) |
| 121 | .n(n) |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 122 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 123 | } |
| 124 | } |
| 125 | |
| 126 | TEST(Q8_GAVGPOOL_UP7__NEON, n_div_8_subtile) { |
| 127 | TEST_REQUIRES_ARM_NEON; |
| 128 | for (size_t n = 8; n < 128; n += 24) { |
| 129 | for (size_t m = 1; m < 7; m++) { |
| 130 | GAvgPoolMicrokernelTester() |
| 131 | .m(m) |
| 132 | .n(n) |
| 133 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 134 | } |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | TEST(Q8_GAVGPOOL_UP7__NEON, n_lt_8_fulltile) { |
| 139 | TEST_REQUIRES_ARM_NEON; |
| 140 | for (size_t n = 1; n < 8; n++) { |
| 141 | GAvgPoolMicrokernelTester() |
| 142 | .m(7) |
| 143 | .n(n) |
| 144 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 145 | } |
| 146 | } |
| 147 | |
| 148 | TEST(Q8_GAVGPOOL_UP7__NEON, n_lt_8_subtile) { |
| 149 | TEST_REQUIRES_ARM_NEON; |
| 150 | for (size_t n = 1; n < 8; n++) { |
| 151 | for (size_t m = 1; m < 7; m++) { |
| 152 | GAvgPoolMicrokernelTester() |
| 153 | .m(m) |
| 154 | .n(n) |
| 155 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 156 | } |
| 157 | } |
| 158 | } |
| 159 | |
| 160 | TEST(Q8_GAVGPOOL_UP7__NEON, n_lt_8_fulltile_with_x_scale) { |
| 161 | TEST_REQUIRES_ARM_NEON; |
| 162 | for (size_t n = 1; n < 8; n++) { |
| 163 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 164 | GAvgPoolMicrokernelTester() |
| 165 | .m(7) |
| 166 | .n(n) |
| 167 | .x_scale(x_scale) |
| 168 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 169 | } |
| 170 | } |
| 171 | } |
| 172 | |
| 173 | TEST(Q8_GAVGPOOL_UP7__NEON, n_lt_8_fulltile_with_x_zero_point) { |
| 174 | TEST_REQUIRES_ARM_NEON; |
| 175 | for (size_t n = 1; n < 8; n++) { |
| 176 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 177 | GAvgPoolMicrokernelTester() |
| 178 | .m(7) |
| 179 | .n(n) |
| 180 | .x_zero_point(x_zero_point) |
| 181 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 182 | } |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | TEST(Q8_GAVGPOOL_UP7__NEON, n_lt_8_fulltile_with_y_scale) { |
| 187 | TEST_REQUIRES_ARM_NEON; |
| 188 | for (size_t n = 1; n < 8; n++) { |
| 189 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 190 | GAvgPoolMicrokernelTester() |
| 191 | .m(7) |
| 192 | .n(n) |
| 193 | .y_scale(y_scale) |
| 194 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 195 | } |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | TEST(Q8_GAVGPOOL_UP7__NEON, n_lt_8_fulltile_with_y_zero_point) { |
| 200 | TEST_REQUIRES_ARM_NEON; |
| 201 | for (size_t n = 1; n < 8; n++) { |
| 202 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 203 | GAvgPoolMicrokernelTester() |
| 204 | .m(7) |
| 205 | .n(n) |
| 206 | .y_zero_point(y_zero_point) |
| 207 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 208 | } |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | TEST(Q8_GAVGPOOL_UP7__NEON, n_lt_8_fulltile_with_qmax) { |
| 213 | TEST_REQUIRES_ARM_NEON; |
| 214 | for (size_t n = 1; n < 8; n++) { |
| 215 | GAvgPoolMicrokernelTester() |
| 216 | .m(7) |
| 217 | .n(n) |
| 218 | .x_zero_point(128) |
| 219 | .y_zero_point(128) |
| 220 | .x_scale(1.0f) |
| 221 | .y_scale(1.0f) |
| 222 | .qmax(128) |
| 223 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 224 | } |
| 225 | } |
| 226 | |
| 227 | TEST(Q8_GAVGPOOL_UP7__NEON, n_lt_8_fulltile_with_qmin) { |
| 228 | TEST_REQUIRES_ARM_NEON; |
| 229 | for (size_t n = 1; n < 8; n++) { |
| 230 | GAvgPoolMicrokernelTester() |
| 231 | .m(7) |
| 232 | .n(n) |
| 233 | .x_zero_point(128) |
| 234 | .y_zero_point(128) |
| 235 | .x_scale(1.0f) |
| 236 | .y_scale(1.0f) |
| 237 | .qmin(128) |
| 238 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | TEST(Q8_GAVGPOOL_UP7__NEON, n_gt_8_fulltile) { |
| 243 | TEST_REQUIRES_ARM_NEON; |
| 244 | for (size_t n = 9; n < 16; n++) { |
| 245 | GAvgPoolMicrokernelTester() |
| 246 | .m(7) |
| 247 | .n(n) |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 248 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 249 | } |
| 250 | } |
| 251 | |
| 252 | TEST(Q8_GAVGPOOL_UP7__NEON, n_gt_8_subtile) { |
| 253 | TEST_REQUIRES_ARM_NEON; |
| 254 | for (size_t n = 9; n < 16; n++) { |
| 255 | for (size_t m = 1; m < 7; m++) { |
| 256 | GAvgPoolMicrokernelTester() |
| 257 | .m(m) |
| 258 | .n(n) |
| 259 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 260 | } |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | TEST(Q8_GAVGPOOL_UP7__NEON, n_gt_8_fulltile_with_x_scale) { |
| 265 | TEST_REQUIRES_ARM_NEON; |
| 266 | for (size_t n = 9; n < 16; n++) { |
| 267 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 268 | GAvgPoolMicrokernelTester() |
| 269 | .m(7) |
| 270 | .n(n) |
| 271 | .x_scale(x_scale) |
| 272 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 273 | } |
| 274 | } |
| 275 | } |
| 276 | |
| 277 | TEST(Q8_GAVGPOOL_UP7__NEON, n_gt_8_fulltile_with_x_zero_point) { |
| 278 | TEST_REQUIRES_ARM_NEON; |
| 279 | for (size_t n = 9; n < 16; n++) { |
| 280 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 281 | GAvgPoolMicrokernelTester() |
| 282 | .m(7) |
| 283 | .n(n) |
| 284 | .x_zero_point(x_zero_point) |
| 285 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 286 | } |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | TEST(Q8_GAVGPOOL_UP7__NEON, n_gt_8_fulltile_with_y_scale) { |
| 291 | TEST_REQUIRES_ARM_NEON; |
| 292 | for (size_t n = 9; n < 16; n++) { |
| 293 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 294 | GAvgPoolMicrokernelTester() |
| 295 | .m(7) |
| 296 | .n(n) |
| 297 | .y_scale(y_scale) |
| 298 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 299 | } |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | TEST(Q8_GAVGPOOL_UP7__NEON, n_gt_8_fulltile_with_y_zero_point) { |
| 304 | TEST_REQUIRES_ARM_NEON; |
| 305 | for (size_t n = 9; n < 16; n++) { |
| 306 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 307 | GAvgPoolMicrokernelTester() |
| 308 | .m(7) |
| 309 | .n(n) |
| 310 | .y_zero_point(y_zero_point) |
| 311 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 312 | } |
| 313 | } |
| 314 | } |
| 315 | |
| 316 | TEST(Q8_GAVGPOOL_UP7__NEON, n_gt_8_fulltile_with_qmax) { |
| 317 | TEST_REQUIRES_ARM_NEON; |
| 318 | for (size_t n = 9; n < 16; n++) { |
| 319 | GAvgPoolMicrokernelTester() |
| 320 | .m(7) |
| 321 | .n(n) |
| 322 | .x_zero_point(128) |
| 323 | .y_zero_point(128) |
| 324 | .x_scale(1.0f) |
| 325 | .y_scale(1.0f) |
| 326 | .qmax(128) |
| 327 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | TEST(Q8_GAVGPOOL_UP7__NEON, n_gt_8_fulltile_with_qmin) { |
| 332 | TEST_REQUIRES_ARM_NEON; |
| 333 | for (size_t n = 9; n < 16; n++) { |
| 334 | GAvgPoolMicrokernelTester() |
| 335 | .m(7) |
| 336 | .n(n) |
| 337 | .x_zero_point(128) |
| 338 | .y_zero_point(128) |
| 339 | .x_scale(1.0f) |
| 340 | .y_scale(1.0f) |
| 341 | .qmin(128) |
| 342 | .Test(xnn_q8_gavgpool_ukernel_up7__neon); |
| 343 | } |
| 344 | } |
| 345 | |
| 346 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_fulltile) { |
| 347 | TEST_REQUIRES_ARM_NEON; |
| 348 | GAvgPoolMicrokernelTester() |
| 349 | .m(14) |
| 350 | .n(8) |
| 351 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 352 | } |
| 353 | |
| 354 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_fulltile_with_x_stride) { |
| 355 | TEST_REQUIRES_ARM_NEON; |
| 356 | GAvgPoolMicrokernelTester() |
| 357 | .m(14) |
| 358 | .n(8) |
| 359 | .x_stride(11) |
| 360 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 361 | } |
| 362 | |
| 363 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_fulltile_with_x_scale) { |
| 364 | TEST_REQUIRES_ARM_NEON; |
| 365 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 366 | GAvgPoolMicrokernelTester() |
| 367 | .m(14) |
| 368 | .n(8) |
| 369 | .x_scale(x_scale) |
| 370 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 371 | } |
| 372 | } |
| 373 | |
| 374 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_fulltile_with_x_zero_point) { |
| 375 | TEST_REQUIRES_ARM_NEON; |
| 376 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 377 | GAvgPoolMicrokernelTester() |
| 378 | .m(14) |
| 379 | .n(8) |
| 380 | .x_zero_point(x_zero_point) |
| 381 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 382 | } |
| 383 | } |
| 384 | |
| 385 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_fulltile_with_y_scale) { |
| 386 | TEST_REQUIRES_ARM_NEON; |
| 387 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 388 | GAvgPoolMicrokernelTester() |
| 389 | .m(14) |
| 390 | .n(8) |
| 391 | .y_scale(y_scale) |
| 392 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 393 | } |
| 394 | } |
| 395 | |
| 396 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_fulltile_with_y_zero_point) { |
| 397 | TEST_REQUIRES_ARM_NEON; |
| 398 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 399 | GAvgPoolMicrokernelTester() |
| 400 | .m(14) |
| 401 | .n(8) |
| 402 | .y_zero_point(y_zero_point) |
| 403 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 404 | } |
| 405 | } |
| 406 | |
| 407 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_fulltile_with_qmax) { |
| 408 | TEST_REQUIRES_ARM_NEON; |
| 409 | GAvgPoolMicrokernelTester() |
| 410 | .m(14) |
| 411 | .n(8) |
| 412 | .x_zero_point(128) |
| 413 | .y_zero_point(128) |
| 414 | .x_scale(1.0f) |
| 415 | .y_scale(1.0f) |
| 416 | .qmax(128) |
| 417 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 418 | } |
| 419 | |
| 420 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_fulltile_with_qmin) { |
| 421 | TEST_REQUIRES_ARM_NEON; |
| 422 | GAvgPoolMicrokernelTester() |
| 423 | .m(14) |
| 424 | .n(8) |
| 425 | .x_zero_point(128) |
| 426 | .y_zero_point(128) |
| 427 | .x_scale(1.0f) |
| 428 | .y_scale(1.0f) |
| 429 | .qmin(128) |
| 430 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 431 | } |
| 432 | |
| 433 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_subtile) { |
| 434 | TEST_REQUIRES_ARM_NEON; |
| 435 | for (size_t m = 1; m < 7; m++) { |
| 436 | GAvgPoolMicrokernelTester() |
| 437 | .m(7 + m) |
| 438 | .n(8) |
| 439 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 440 | } |
| 441 | } |
| 442 | |
| 443 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_2pass_subtile_with_x_stride) { |
| 444 | TEST_REQUIRES_ARM_NEON; |
| 445 | for (size_t m = 1; m < 7; m++) { |
| 446 | GAvgPoolMicrokernelTester() |
| 447 | .m(7 + m) |
| 448 | .n(8) |
| 449 | .x_stride(11) |
| 450 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 451 | } |
| 452 | } |
| 453 | |
| 454 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_multipass_fulltile) { |
| 455 | TEST_REQUIRES_ARM_NEON; |
| 456 | for (size_t m = 14; m <= 35; m += 7) { |
| 457 | GAvgPoolMicrokernelTester() |
| 458 | .m(m) |
| 459 | .n(8) |
| 460 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 461 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 462 | } |
| 463 | |
| 464 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_eq_8_multipass_fulltile_with_x_stride) { |
| 465 | TEST_REQUIRES_ARM_NEON; |
| 466 | for (size_t m = 14; m <= 35; m += 7) { |
| 467 | GAvgPoolMicrokernelTester() |
| 468 | .m(m) |
| 469 | .n(8) |
| 470 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 471 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 472 | } |
| 473 | |
| 474 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_div_8_2pass_fulltile) { |
| 475 | TEST_REQUIRES_ARM_NEON; |
| 476 | for (size_t n = 8; n < 128; n += 24) { |
| 477 | GAvgPoolMicrokernelTester() |
| 478 | .m(14) |
| 479 | .n(n) |
| 480 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 481 | } |
| 482 | } |
| 483 | |
| 484 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_div_8_2pass_subtile) { |
| 485 | TEST_REQUIRES_ARM_NEON; |
| 486 | for (size_t n = 8; n < 128; n += 24) { |
| 487 | for (size_t m = 1; m < 7; m++) { |
| 488 | GAvgPoolMicrokernelTester() |
| 489 | .m(7 + m) |
| 490 | .n(n) |
| 491 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 492 | } |
| 493 | } |
| 494 | } |
| 495 | |
| 496 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_div_8_multipass_fulltile) { |
| 497 | TEST_REQUIRES_ARM_NEON; |
| 498 | for (size_t n = 8; n < 128; n += 24) { |
| 499 | for (size_t m = 14; m <= 35; m += 7) { |
| 500 | GAvgPoolMicrokernelTester() |
| 501 | .m(m) |
| 502 | .n(n) |
| 503 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 504 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 505 | } |
| 506 | } |
| 507 | |
| 508 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_div_8_multipass_fulltile_with_x_stride) { |
| 509 | TEST_REQUIRES_ARM_NEON; |
| 510 | for (size_t n = 8; n < 128; n += 24) { |
| 511 | for (size_t m = 14; m <= 35; m += 7) { |
| 512 | GAvgPoolMicrokernelTester() |
| 513 | .m(m) |
| 514 | .n(n) |
| 515 | .x_stride(131) |
| 516 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 517 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 518 | } |
| 519 | } |
| 520 | |
| 521 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_2pass_fulltile) { |
| 522 | TEST_REQUIRES_ARM_NEON; |
| 523 | for (size_t n = 1; n < 8; n++) { |
| 524 | GAvgPoolMicrokernelTester() |
| 525 | .m(14) |
| 526 | .n(n) |
| 527 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 528 | } |
| 529 | } |
| 530 | |
| 531 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_2pass_fulltile_with_x_scale) { |
| 532 | TEST_REQUIRES_ARM_NEON; |
| 533 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 534 | for (size_t n = 1; n < 8; n++) { |
| 535 | GAvgPoolMicrokernelTester() |
| 536 | .m(14) |
| 537 | .n(n) |
| 538 | .x_scale(x_scale) |
| 539 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 540 | } |
| 541 | } |
| 542 | } |
| 543 | |
| 544 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_2pass_fulltile_with_x_zero_point) { |
| 545 | TEST_REQUIRES_ARM_NEON; |
| 546 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 547 | for (size_t n = 1; n < 8; n++) { |
| 548 | GAvgPoolMicrokernelTester() |
| 549 | .m(14) |
| 550 | .n(n) |
| 551 | .x_zero_point(x_zero_point) |
| 552 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 553 | } |
| 554 | } |
| 555 | } |
| 556 | |
| 557 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_2pass_fulltile_with_y_scale) { |
| 558 | TEST_REQUIRES_ARM_NEON; |
| 559 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 560 | for (size_t n = 1; n < 8; n++) { |
| 561 | GAvgPoolMicrokernelTester() |
| 562 | .m(14) |
| 563 | .n(n) |
| 564 | .y_scale(y_scale) |
| 565 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 566 | } |
| 567 | } |
| 568 | } |
| 569 | |
| 570 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_2pass_fulltile_with_y_zero_point) { |
| 571 | TEST_REQUIRES_ARM_NEON; |
| 572 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 573 | for (size_t n = 1; n < 8; n++) { |
| 574 | GAvgPoolMicrokernelTester() |
| 575 | .m(14) |
| 576 | .n(n) |
| 577 | .y_zero_point(y_zero_point) |
| 578 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 579 | } |
| 580 | } |
| 581 | } |
| 582 | |
| 583 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_2pass_fulltile_with_qmax) { |
| 584 | TEST_REQUIRES_ARM_NEON; |
| 585 | for (size_t n = 1; n < 8; n++) { |
| 586 | GAvgPoolMicrokernelTester() |
| 587 | .m(14) |
| 588 | .n(n) |
| 589 | .x_zero_point(128) |
| 590 | .y_zero_point(128) |
| 591 | .x_scale(1.0f) |
| 592 | .y_scale(1.0f) |
| 593 | .qmax(128) |
| 594 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 595 | } |
| 596 | } |
| 597 | |
| 598 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_2pass_fulltile_with_qmin) { |
| 599 | TEST_REQUIRES_ARM_NEON; |
| 600 | for (size_t n = 1; n < 8; n++) { |
| 601 | GAvgPoolMicrokernelTester() |
| 602 | .m(14) |
| 603 | .n(n) |
| 604 | .x_zero_point(128) |
| 605 | .y_zero_point(128) |
| 606 | .x_scale(1.0f) |
| 607 | .y_scale(1.0f) |
| 608 | .qmin(128) |
| 609 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 610 | } |
| 611 | } |
| 612 | |
| 613 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_2pass_subtile) { |
| 614 | TEST_REQUIRES_ARM_NEON; |
| 615 | for (size_t n = 1; n < 8; n++) { |
| 616 | for (size_t m = 1; m < 7; m++) { |
| 617 | GAvgPoolMicrokernelTester() |
| 618 | .m(7 + m) |
| 619 | .n(n) |
| 620 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 621 | } |
| 622 | } |
| 623 | } |
| 624 | |
| 625 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_multipass_fulltile) { |
| 626 | TEST_REQUIRES_ARM_NEON; |
| 627 | for (size_t n = 1; n < 8; n++) { |
| 628 | for (size_t m = 14; m <= 35; m += 7) { |
| 629 | GAvgPoolMicrokernelTester() |
| 630 | .m(m) |
| 631 | .n(n) |
| 632 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 633 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 634 | } |
| 635 | } |
| 636 | |
| 637 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_lt_8_multipass_fulltile_with_x_stride) { |
| 638 | TEST_REQUIRES_ARM_NEON; |
| 639 | for (size_t n = 1; n < 8; n++) { |
| 640 | for (size_t m = 14; m <= 35; m += 7) { |
| 641 | GAvgPoolMicrokernelTester() |
| 642 | .m(m) |
| 643 | .n(n) |
| 644 | .x_stride(23) |
| 645 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 646 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 647 | } |
| 648 | } |
| 649 | |
| 650 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_2pass_fulltile) { |
| 651 | TEST_REQUIRES_ARM_NEON; |
| 652 | for (size_t n = 9; n < 16; n++) { |
| 653 | GAvgPoolMicrokernelTester() |
| 654 | .m(14) |
| 655 | .n(n) |
| 656 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 657 | } |
| 658 | } |
| 659 | |
| 660 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_2pass_fulltile_with_x_scale) { |
| 661 | TEST_REQUIRES_ARM_NEON; |
| 662 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 663 | for (size_t n = 9; n < 16; n++) { |
| 664 | GAvgPoolMicrokernelTester() |
| 665 | .m(14) |
| 666 | .n(n) |
| 667 | .x_scale(x_scale) |
| 668 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 669 | } |
| 670 | } |
| 671 | } |
| 672 | |
| 673 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_2pass_fulltile_with_x_zero_point) { |
| 674 | TEST_REQUIRES_ARM_NEON; |
| 675 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 676 | for (size_t n = 9; n < 16; n++) { |
| 677 | GAvgPoolMicrokernelTester() |
| 678 | .m(14) |
| 679 | .n(n) |
| 680 | .x_zero_point(x_zero_point) |
| 681 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 682 | } |
| 683 | } |
| 684 | } |
| 685 | |
| 686 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_2pass_fulltile_with_y_scale) { |
| 687 | TEST_REQUIRES_ARM_NEON; |
| 688 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 689 | for (size_t n = 9; n < 16; n++) { |
| 690 | GAvgPoolMicrokernelTester() |
| 691 | .m(14) |
| 692 | .n(n) |
| 693 | .y_scale(y_scale) |
| 694 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 695 | } |
| 696 | } |
| 697 | } |
| 698 | |
| 699 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_2pass_fulltile_with_y_zero_point) { |
| 700 | TEST_REQUIRES_ARM_NEON; |
| 701 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 702 | for (size_t n = 9; n < 16; n++) { |
| 703 | GAvgPoolMicrokernelTester() |
| 704 | .m(14) |
| 705 | .n(n) |
| 706 | .y_zero_point(y_zero_point) |
| 707 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 708 | } |
| 709 | } |
| 710 | } |
| 711 | |
| 712 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_2pass_fulltile_with_qmax) { |
| 713 | TEST_REQUIRES_ARM_NEON; |
| 714 | for (size_t n = 9; n < 16; n++) { |
| 715 | GAvgPoolMicrokernelTester() |
| 716 | .m(14) |
| 717 | .n(n) |
| 718 | .x_zero_point(128) |
| 719 | .y_zero_point(128) |
| 720 | .x_scale(1.0f) |
| 721 | .y_scale(1.0f) |
| 722 | .qmax(128) |
| 723 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 724 | } |
| 725 | } |
| 726 | |
| 727 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_2pass_fulltile_with_qmin) { |
| 728 | TEST_REQUIRES_ARM_NEON; |
| 729 | for (size_t n = 9; n < 16; n++) { |
| 730 | GAvgPoolMicrokernelTester() |
| 731 | .m(14) |
| 732 | .n(n) |
| 733 | .x_zero_point(128) |
| 734 | .y_zero_point(128) |
| 735 | .x_scale(1.0f) |
| 736 | .y_scale(1.0f) |
| 737 | .qmin(128) |
| 738 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 739 | } |
| 740 | } |
| 741 | |
| 742 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_2pass_subtile) { |
| 743 | TEST_REQUIRES_ARM_NEON; |
| 744 | for (size_t n = 9; n < 16; n++) { |
| 745 | for (size_t m = 1; m < 7; m++) { |
| 746 | GAvgPoolMicrokernelTester() |
| 747 | .m(7 + m) |
| 748 | .n(n) |
| 749 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
| 750 | } |
| 751 | } |
| 752 | } |
| 753 | |
| 754 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_multipass_fulltile) { |
| 755 | TEST_REQUIRES_ARM_NEON; |
| 756 | for (size_t n = 9; n < 16; n++) { |
| 757 | for (size_t m = 14; m <= 35; m += 7) { |
| 758 | GAvgPoolMicrokernelTester() |
| 759 | .m(m) |
| 760 | .n(n) |
| 761 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 762 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 763 | } |
| 764 | } |
| 765 | |
| 766 | TEST(Q8_GAVGPOOL_MP7p7q__NEON, n_gt_8_multipass_fulltile_with_x_stride) { |
| 767 | TEST_REQUIRES_ARM_NEON; |
| 768 | for (size_t n = 9; n < 16; n++) { |
| 769 | for (size_t m = 14; m <= 35; m += 7) { |
| 770 | GAvgPoolMicrokernelTester() |
| 771 | .m(m) |
| 772 | .n(n) |
| 773 | .x_stride(23) |
| 774 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__neon); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 775 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 776 | } |
| 777 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 778 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 779 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 780 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 781 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_eq_8_fulltile) { |
| 782 | TEST_REQUIRES_X86_SSE2; |
| 783 | GAvgPoolMicrokernelTester() |
| 784 | .m(7) |
| 785 | .n(8) |
| 786 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 787 | } |
| 788 | |
| 789 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_eq_8_subtile) { |
| 790 | TEST_REQUIRES_X86_SSE2; |
| 791 | for (size_t m = 1; m < 7; m++) { |
| 792 | GAvgPoolMicrokernelTester() |
| 793 | .m(m) |
| 794 | .n(8) |
| 795 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 796 | } |
| 797 | } |
| 798 | |
| 799 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_eq_8_fulltile_with_x_stride) { |
| 800 | TEST_REQUIRES_X86_SSE2; |
| 801 | GAvgPoolMicrokernelTester() |
| 802 | .m(7) |
| 803 | .n(8) |
| 804 | .x_stride(11) |
| 805 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 806 | } |
| 807 | |
| 808 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_eq_8_fulltile_with_x_scale) { |
| 809 | TEST_REQUIRES_X86_SSE2; |
| 810 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 811 | GAvgPoolMicrokernelTester() |
| 812 | .m(7) |
| 813 | .n(8) |
| 814 | .x_scale(x_scale) |
| 815 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 816 | } |
| 817 | } |
| 818 | |
| 819 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_eq_8_fulltile_with_x_zero_point) { |
| 820 | TEST_REQUIRES_X86_SSE2; |
| 821 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 822 | GAvgPoolMicrokernelTester() |
| 823 | .m(7) |
| 824 | .n(8) |
| 825 | .x_zero_point(x_zero_point) |
| 826 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 827 | } |
| 828 | } |
| 829 | |
| 830 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_eq_8_fulltile_with_y_scale) { |
| 831 | TEST_REQUIRES_X86_SSE2; |
| 832 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 833 | GAvgPoolMicrokernelTester() |
| 834 | .m(7) |
| 835 | .n(8) |
| 836 | .y_scale(y_scale) |
| 837 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 838 | } |
| 839 | } |
| 840 | |
| 841 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_eq_8_fulltile_with_y_zero_point) { |
| 842 | TEST_REQUIRES_X86_SSE2; |
| 843 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 844 | GAvgPoolMicrokernelTester() |
| 845 | .m(7) |
| 846 | .n(8) |
| 847 | .y_zero_point(y_zero_point) |
| 848 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 849 | } |
| 850 | } |
| 851 | |
| 852 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_eq_8_fulltile_with_qmax) { |
| 853 | TEST_REQUIRES_X86_SSE2; |
| 854 | GAvgPoolMicrokernelTester() |
| 855 | .m(7) |
| 856 | .n(8) |
| 857 | .x_zero_point(128) |
| 858 | .y_zero_point(128) |
| 859 | .x_scale(1.0f) |
| 860 | .y_scale(1.0f) |
| 861 | .qmax(128) |
| 862 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 863 | } |
| 864 | |
| 865 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_eq_8_fulltile_with_qmin) { |
| 866 | TEST_REQUIRES_X86_SSE2; |
| 867 | GAvgPoolMicrokernelTester() |
| 868 | .m(7) |
| 869 | .n(8) |
| 870 | .x_zero_point(128) |
| 871 | .y_zero_point(128) |
| 872 | .x_scale(1.0f) |
| 873 | .y_scale(1.0f) |
| 874 | .qmin(128) |
| 875 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 876 | } |
| 877 | |
| 878 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_div_8_fulltile) { |
| 879 | TEST_REQUIRES_X86_SSE2; |
| 880 | for (size_t n = 8; n < 128; n += 24) { |
| 881 | GAvgPoolMicrokernelTester() |
| 882 | .m(7) |
| 883 | .n(n) |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 884 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 885 | } |
| 886 | } |
| 887 | |
| 888 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_div_8_subtile) { |
| 889 | TEST_REQUIRES_X86_SSE2; |
| 890 | for (size_t n = 8; n < 128; n += 24) { |
| 891 | for (size_t m = 1; m < 7; m++) { |
| 892 | GAvgPoolMicrokernelTester() |
| 893 | .m(m) |
| 894 | .n(n) |
| 895 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 896 | } |
| 897 | } |
| 898 | } |
| 899 | |
| 900 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_lt_8_fulltile) { |
| 901 | TEST_REQUIRES_X86_SSE2; |
| 902 | for (size_t n = 1; n < 8; n++) { |
| 903 | GAvgPoolMicrokernelTester() |
| 904 | .m(7) |
| 905 | .n(n) |
| 906 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 907 | } |
| 908 | } |
| 909 | |
| 910 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_lt_8_subtile) { |
| 911 | TEST_REQUIRES_X86_SSE2; |
| 912 | for (size_t n = 1; n < 8; n++) { |
| 913 | for (size_t m = 1; m < 7; m++) { |
| 914 | GAvgPoolMicrokernelTester() |
| 915 | .m(m) |
| 916 | .n(n) |
| 917 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 918 | } |
| 919 | } |
| 920 | } |
| 921 | |
| 922 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_lt_8_fulltile_with_x_scale) { |
| 923 | TEST_REQUIRES_X86_SSE2; |
| 924 | for (size_t n = 1; n < 8; n++) { |
| 925 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 926 | GAvgPoolMicrokernelTester() |
| 927 | .m(7) |
| 928 | .n(n) |
| 929 | .x_scale(x_scale) |
| 930 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 931 | } |
| 932 | } |
| 933 | } |
| 934 | |
| 935 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_lt_8_fulltile_with_x_zero_point) { |
| 936 | TEST_REQUIRES_X86_SSE2; |
| 937 | for (size_t n = 1; n < 8; n++) { |
| 938 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 939 | GAvgPoolMicrokernelTester() |
| 940 | .m(7) |
| 941 | .n(n) |
| 942 | .x_zero_point(x_zero_point) |
| 943 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 944 | } |
| 945 | } |
| 946 | } |
| 947 | |
| 948 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_lt_8_fulltile_with_y_scale) { |
| 949 | TEST_REQUIRES_X86_SSE2; |
| 950 | for (size_t n = 1; n < 8; n++) { |
| 951 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 952 | GAvgPoolMicrokernelTester() |
| 953 | .m(7) |
| 954 | .n(n) |
| 955 | .y_scale(y_scale) |
| 956 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 957 | } |
| 958 | } |
| 959 | } |
| 960 | |
| 961 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_lt_8_fulltile_with_y_zero_point) { |
| 962 | TEST_REQUIRES_X86_SSE2; |
| 963 | for (size_t n = 1; n < 8; n++) { |
| 964 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 965 | GAvgPoolMicrokernelTester() |
| 966 | .m(7) |
| 967 | .n(n) |
| 968 | .y_zero_point(y_zero_point) |
| 969 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 970 | } |
| 971 | } |
| 972 | } |
| 973 | |
| 974 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_lt_8_fulltile_with_qmax) { |
| 975 | TEST_REQUIRES_X86_SSE2; |
| 976 | for (size_t n = 1; n < 8; n++) { |
| 977 | GAvgPoolMicrokernelTester() |
| 978 | .m(7) |
| 979 | .n(n) |
| 980 | .x_zero_point(128) |
| 981 | .y_zero_point(128) |
| 982 | .x_scale(1.0f) |
| 983 | .y_scale(1.0f) |
| 984 | .qmax(128) |
| 985 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 986 | } |
| 987 | } |
| 988 | |
| 989 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_lt_8_fulltile_with_qmin) { |
| 990 | TEST_REQUIRES_X86_SSE2; |
| 991 | for (size_t n = 1; n < 8; n++) { |
| 992 | GAvgPoolMicrokernelTester() |
| 993 | .m(7) |
| 994 | .n(n) |
| 995 | .x_zero_point(128) |
| 996 | .y_zero_point(128) |
| 997 | .x_scale(1.0f) |
| 998 | .y_scale(1.0f) |
| 999 | .qmin(128) |
| 1000 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 1001 | } |
| 1002 | } |
| 1003 | |
| 1004 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_gt_8_fulltile) { |
| 1005 | TEST_REQUIRES_X86_SSE2; |
| 1006 | for (size_t n = 9; n < 16; n++) { |
| 1007 | GAvgPoolMicrokernelTester() |
| 1008 | .m(7) |
| 1009 | .n(n) |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1010 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1011 | } |
| 1012 | } |
| 1013 | |
| 1014 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_gt_8_subtile) { |
| 1015 | TEST_REQUIRES_X86_SSE2; |
| 1016 | for (size_t n = 9; n < 16; n++) { |
| 1017 | for (size_t m = 1; m < 7; m++) { |
| 1018 | GAvgPoolMicrokernelTester() |
| 1019 | .m(m) |
| 1020 | .n(n) |
| 1021 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 1022 | } |
| 1023 | } |
| 1024 | } |
| 1025 | |
| 1026 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_gt_8_fulltile_with_x_scale) { |
| 1027 | TEST_REQUIRES_X86_SSE2; |
| 1028 | for (size_t n = 9; n < 16; n++) { |
| 1029 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 1030 | GAvgPoolMicrokernelTester() |
| 1031 | .m(7) |
| 1032 | .n(n) |
| 1033 | .x_scale(x_scale) |
| 1034 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 1035 | } |
| 1036 | } |
| 1037 | } |
| 1038 | |
| 1039 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_gt_8_fulltile_with_x_zero_point) { |
| 1040 | TEST_REQUIRES_X86_SSE2; |
| 1041 | for (size_t n = 9; n < 16; n++) { |
| 1042 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 1043 | GAvgPoolMicrokernelTester() |
| 1044 | .m(7) |
| 1045 | .n(n) |
| 1046 | .x_zero_point(x_zero_point) |
| 1047 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 1048 | } |
| 1049 | } |
| 1050 | } |
| 1051 | |
| 1052 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_gt_8_fulltile_with_y_scale) { |
| 1053 | TEST_REQUIRES_X86_SSE2; |
| 1054 | for (size_t n = 9; n < 16; n++) { |
| 1055 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 1056 | GAvgPoolMicrokernelTester() |
| 1057 | .m(7) |
| 1058 | .n(n) |
| 1059 | .y_scale(y_scale) |
| 1060 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 1061 | } |
| 1062 | } |
| 1063 | } |
| 1064 | |
| 1065 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_gt_8_fulltile_with_y_zero_point) { |
| 1066 | TEST_REQUIRES_X86_SSE2; |
| 1067 | for (size_t n = 9; n < 16; n++) { |
| 1068 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 1069 | GAvgPoolMicrokernelTester() |
| 1070 | .m(7) |
| 1071 | .n(n) |
| 1072 | .y_zero_point(y_zero_point) |
| 1073 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 1074 | } |
| 1075 | } |
| 1076 | } |
| 1077 | |
| 1078 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_gt_8_fulltile_with_qmax) { |
| 1079 | TEST_REQUIRES_X86_SSE2; |
| 1080 | for (size_t n = 9; n < 16; n++) { |
| 1081 | GAvgPoolMicrokernelTester() |
| 1082 | .m(7) |
| 1083 | .n(n) |
| 1084 | .x_zero_point(128) |
| 1085 | .y_zero_point(128) |
| 1086 | .x_scale(1.0f) |
| 1087 | .y_scale(1.0f) |
| 1088 | .qmax(128) |
| 1089 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 1090 | } |
| 1091 | } |
| 1092 | |
| 1093 | TEST(Q8_GAVGPOOL_UP7__SSE2, n_gt_8_fulltile_with_qmin) { |
| 1094 | TEST_REQUIRES_X86_SSE2; |
| 1095 | for (size_t n = 9; n < 16; n++) { |
| 1096 | GAvgPoolMicrokernelTester() |
| 1097 | .m(7) |
| 1098 | .n(n) |
| 1099 | .x_zero_point(128) |
| 1100 | .y_zero_point(128) |
| 1101 | .x_scale(1.0f) |
| 1102 | .y_scale(1.0f) |
| 1103 | .qmin(128) |
| 1104 | .Test(xnn_q8_gavgpool_ukernel_up7__sse2); |
| 1105 | } |
| 1106 | } |
| 1107 | |
| 1108 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_fulltile) { |
| 1109 | TEST_REQUIRES_X86_SSE2; |
| 1110 | GAvgPoolMicrokernelTester() |
| 1111 | .m(14) |
| 1112 | .n(8) |
| 1113 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1114 | } |
| 1115 | |
| 1116 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_fulltile_with_x_stride) { |
| 1117 | TEST_REQUIRES_X86_SSE2; |
| 1118 | GAvgPoolMicrokernelTester() |
| 1119 | .m(14) |
| 1120 | .n(8) |
| 1121 | .x_stride(11) |
| 1122 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1123 | } |
| 1124 | |
| 1125 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_fulltile_with_x_scale) { |
| 1126 | TEST_REQUIRES_X86_SSE2; |
| 1127 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 1128 | GAvgPoolMicrokernelTester() |
| 1129 | .m(14) |
| 1130 | .n(8) |
| 1131 | .x_scale(x_scale) |
| 1132 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1133 | } |
| 1134 | } |
| 1135 | |
| 1136 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_fulltile_with_x_zero_point) { |
| 1137 | TEST_REQUIRES_X86_SSE2; |
| 1138 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 1139 | GAvgPoolMicrokernelTester() |
| 1140 | .m(14) |
| 1141 | .n(8) |
| 1142 | .x_zero_point(x_zero_point) |
| 1143 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1144 | } |
| 1145 | } |
| 1146 | |
| 1147 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_fulltile_with_y_scale) { |
| 1148 | TEST_REQUIRES_X86_SSE2; |
| 1149 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 1150 | GAvgPoolMicrokernelTester() |
| 1151 | .m(14) |
| 1152 | .n(8) |
| 1153 | .y_scale(y_scale) |
| 1154 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1155 | } |
| 1156 | } |
| 1157 | |
| 1158 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_fulltile_with_y_zero_point) { |
| 1159 | TEST_REQUIRES_X86_SSE2; |
| 1160 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 1161 | GAvgPoolMicrokernelTester() |
| 1162 | .m(14) |
| 1163 | .n(8) |
| 1164 | .y_zero_point(y_zero_point) |
| 1165 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1166 | } |
| 1167 | } |
| 1168 | |
| 1169 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_fulltile_with_qmax) { |
| 1170 | TEST_REQUIRES_X86_SSE2; |
| 1171 | GAvgPoolMicrokernelTester() |
| 1172 | .m(14) |
| 1173 | .n(8) |
| 1174 | .x_zero_point(128) |
| 1175 | .y_zero_point(128) |
| 1176 | .x_scale(1.0f) |
| 1177 | .y_scale(1.0f) |
| 1178 | .qmax(128) |
| 1179 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1180 | } |
| 1181 | |
| 1182 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_fulltile_with_qmin) { |
| 1183 | TEST_REQUIRES_X86_SSE2; |
| 1184 | GAvgPoolMicrokernelTester() |
| 1185 | .m(14) |
| 1186 | .n(8) |
| 1187 | .x_zero_point(128) |
| 1188 | .y_zero_point(128) |
| 1189 | .x_scale(1.0f) |
| 1190 | .y_scale(1.0f) |
| 1191 | .qmin(128) |
| 1192 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1193 | } |
| 1194 | |
| 1195 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_subtile) { |
| 1196 | TEST_REQUIRES_X86_SSE2; |
| 1197 | for (size_t m = 1; m < 7; m++) { |
| 1198 | GAvgPoolMicrokernelTester() |
| 1199 | .m(7 + m) |
| 1200 | .n(8) |
| 1201 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1202 | } |
| 1203 | } |
| 1204 | |
| 1205 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_2pass_subtile_with_x_stride) { |
| 1206 | TEST_REQUIRES_X86_SSE2; |
| 1207 | for (size_t m = 1; m < 7; m++) { |
| 1208 | GAvgPoolMicrokernelTester() |
| 1209 | .m(7 + m) |
| 1210 | .n(8) |
| 1211 | .x_stride(11) |
| 1212 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1213 | } |
| 1214 | } |
| 1215 | |
| 1216 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_multipass_fulltile) { |
| 1217 | TEST_REQUIRES_X86_SSE2; |
| 1218 | for (size_t m = 14; m <= 35; m += 7) { |
| 1219 | GAvgPoolMicrokernelTester() |
| 1220 | .m(m) |
| 1221 | .n(8) |
| 1222 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1223 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1224 | } |
| 1225 | |
| 1226 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_eq_8_multipass_fulltile_with_x_stride) { |
| 1227 | TEST_REQUIRES_X86_SSE2; |
| 1228 | for (size_t m = 14; m <= 35; m += 7) { |
| 1229 | GAvgPoolMicrokernelTester() |
| 1230 | .m(m) |
| 1231 | .n(8) |
| 1232 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1233 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1234 | } |
| 1235 | |
| 1236 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_div_8_2pass_fulltile) { |
| 1237 | TEST_REQUIRES_X86_SSE2; |
| 1238 | for (size_t n = 8; n < 128; n += 24) { |
| 1239 | GAvgPoolMicrokernelTester() |
| 1240 | .m(14) |
| 1241 | .n(n) |
| 1242 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1243 | } |
| 1244 | } |
| 1245 | |
| 1246 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_div_8_2pass_subtile) { |
| 1247 | TEST_REQUIRES_X86_SSE2; |
| 1248 | for (size_t n = 8; n < 128; n += 24) { |
| 1249 | for (size_t m = 1; m < 7; m++) { |
| 1250 | GAvgPoolMicrokernelTester() |
| 1251 | .m(7 + m) |
| 1252 | .n(n) |
| 1253 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1254 | } |
| 1255 | } |
| 1256 | } |
| 1257 | |
| 1258 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_div_8_multipass_fulltile) { |
| 1259 | TEST_REQUIRES_X86_SSE2; |
| 1260 | for (size_t n = 8; n < 128; n += 24) { |
| 1261 | for (size_t m = 14; m <= 35; m += 7) { |
| 1262 | GAvgPoolMicrokernelTester() |
| 1263 | .m(m) |
| 1264 | .n(n) |
| 1265 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1266 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1267 | } |
| 1268 | } |
| 1269 | |
| 1270 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_div_8_multipass_fulltile_with_x_stride) { |
| 1271 | TEST_REQUIRES_X86_SSE2; |
| 1272 | for (size_t n = 8; n < 128; n += 24) { |
| 1273 | for (size_t m = 14; m <= 35; m += 7) { |
| 1274 | GAvgPoolMicrokernelTester() |
| 1275 | .m(m) |
| 1276 | .n(n) |
| 1277 | .x_stride(131) |
| 1278 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1279 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1280 | } |
| 1281 | } |
| 1282 | |
| 1283 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_2pass_fulltile) { |
| 1284 | TEST_REQUIRES_X86_SSE2; |
| 1285 | for (size_t n = 1; n < 8; n++) { |
| 1286 | GAvgPoolMicrokernelTester() |
| 1287 | .m(14) |
| 1288 | .n(n) |
| 1289 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1290 | } |
| 1291 | } |
| 1292 | |
| 1293 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_2pass_fulltile_with_x_scale) { |
| 1294 | TEST_REQUIRES_X86_SSE2; |
| 1295 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 1296 | for (size_t n = 1; n < 8; n++) { |
| 1297 | GAvgPoolMicrokernelTester() |
| 1298 | .m(14) |
| 1299 | .n(n) |
| 1300 | .x_scale(x_scale) |
| 1301 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1302 | } |
| 1303 | } |
| 1304 | } |
| 1305 | |
| 1306 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_2pass_fulltile_with_x_zero_point) { |
| 1307 | TEST_REQUIRES_X86_SSE2; |
| 1308 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 1309 | for (size_t n = 1; n < 8; n++) { |
| 1310 | GAvgPoolMicrokernelTester() |
| 1311 | .m(14) |
| 1312 | .n(n) |
| 1313 | .x_zero_point(x_zero_point) |
| 1314 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1315 | } |
| 1316 | } |
| 1317 | } |
| 1318 | |
| 1319 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_2pass_fulltile_with_y_scale) { |
| 1320 | TEST_REQUIRES_X86_SSE2; |
| 1321 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 1322 | for (size_t n = 1; n < 8; n++) { |
| 1323 | GAvgPoolMicrokernelTester() |
| 1324 | .m(14) |
| 1325 | .n(n) |
| 1326 | .y_scale(y_scale) |
| 1327 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1328 | } |
| 1329 | } |
| 1330 | } |
| 1331 | |
| 1332 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_2pass_fulltile_with_y_zero_point) { |
| 1333 | TEST_REQUIRES_X86_SSE2; |
| 1334 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 1335 | for (size_t n = 1; n < 8; n++) { |
| 1336 | GAvgPoolMicrokernelTester() |
| 1337 | .m(14) |
| 1338 | .n(n) |
| 1339 | .y_zero_point(y_zero_point) |
| 1340 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1341 | } |
| 1342 | } |
| 1343 | } |
| 1344 | |
| 1345 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_2pass_fulltile_with_qmax) { |
| 1346 | TEST_REQUIRES_X86_SSE2; |
| 1347 | for (size_t n = 1; n < 8; n++) { |
| 1348 | GAvgPoolMicrokernelTester() |
| 1349 | .m(14) |
| 1350 | .n(n) |
| 1351 | .x_zero_point(128) |
| 1352 | .y_zero_point(128) |
| 1353 | .x_scale(1.0f) |
| 1354 | .y_scale(1.0f) |
| 1355 | .qmax(128) |
| 1356 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1357 | } |
| 1358 | } |
| 1359 | |
| 1360 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_2pass_fulltile_with_qmin) { |
| 1361 | TEST_REQUIRES_X86_SSE2; |
| 1362 | for (size_t n = 1; n < 8; n++) { |
| 1363 | GAvgPoolMicrokernelTester() |
| 1364 | .m(14) |
| 1365 | .n(n) |
| 1366 | .x_zero_point(128) |
| 1367 | .y_zero_point(128) |
| 1368 | .x_scale(1.0f) |
| 1369 | .y_scale(1.0f) |
| 1370 | .qmin(128) |
| 1371 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1372 | } |
| 1373 | } |
| 1374 | |
| 1375 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_2pass_subtile) { |
| 1376 | TEST_REQUIRES_X86_SSE2; |
| 1377 | for (size_t n = 1; n < 8; n++) { |
| 1378 | for (size_t m = 1; m < 7; m++) { |
| 1379 | GAvgPoolMicrokernelTester() |
| 1380 | .m(7 + m) |
| 1381 | .n(n) |
| 1382 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1383 | } |
| 1384 | } |
| 1385 | } |
| 1386 | |
| 1387 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_multipass_fulltile) { |
| 1388 | TEST_REQUIRES_X86_SSE2; |
| 1389 | for (size_t n = 1; n < 8; n++) { |
| 1390 | for (size_t m = 14; m <= 35; m += 7) { |
| 1391 | GAvgPoolMicrokernelTester() |
| 1392 | .m(m) |
| 1393 | .n(n) |
| 1394 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1395 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1396 | } |
| 1397 | } |
| 1398 | |
| 1399 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_lt_8_multipass_fulltile_with_x_stride) { |
| 1400 | TEST_REQUIRES_X86_SSE2; |
| 1401 | for (size_t n = 1; n < 8; n++) { |
| 1402 | for (size_t m = 14; m <= 35; m += 7) { |
| 1403 | GAvgPoolMicrokernelTester() |
| 1404 | .m(m) |
| 1405 | .n(n) |
| 1406 | .x_stride(23) |
| 1407 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1408 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1409 | } |
| 1410 | } |
| 1411 | |
| 1412 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_2pass_fulltile) { |
| 1413 | TEST_REQUIRES_X86_SSE2; |
| 1414 | for (size_t n = 9; n < 16; n++) { |
| 1415 | GAvgPoolMicrokernelTester() |
| 1416 | .m(14) |
| 1417 | .n(n) |
| 1418 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1419 | } |
| 1420 | } |
| 1421 | |
| 1422 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_2pass_fulltile_with_x_scale) { |
| 1423 | TEST_REQUIRES_X86_SSE2; |
| 1424 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 1425 | for (size_t n = 9; n < 16; n++) { |
| 1426 | GAvgPoolMicrokernelTester() |
| 1427 | .m(14) |
| 1428 | .n(n) |
| 1429 | .x_scale(x_scale) |
| 1430 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1431 | } |
| 1432 | } |
| 1433 | } |
| 1434 | |
| 1435 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_2pass_fulltile_with_x_zero_point) { |
| 1436 | TEST_REQUIRES_X86_SSE2; |
| 1437 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 1438 | for (size_t n = 9; n < 16; n++) { |
| 1439 | GAvgPoolMicrokernelTester() |
| 1440 | .m(14) |
| 1441 | .n(n) |
| 1442 | .x_zero_point(x_zero_point) |
| 1443 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1444 | } |
| 1445 | } |
| 1446 | } |
| 1447 | |
| 1448 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_2pass_fulltile_with_y_scale) { |
| 1449 | TEST_REQUIRES_X86_SSE2; |
| 1450 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 1451 | for (size_t n = 9; n < 16; n++) { |
| 1452 | GAvgPoolMicrokernelTester() |
| 1453 | .m(14) |
| 1454 | .n(n) |
| 1455 | .y_scale(y_scale) |
| 1456 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1457 | } |
| 1458 | } |
| 1459 | } |
| 1460 | |
| 1461 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_2pass_fulltile_with_y_zero_point) { |
| 1462 | TEST_REQUIRES_X86_SSE2; |
| 1463 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 1464 | for (size_t n = 9; n < 16; n++) { |
| 1465 | GAvgPoolMicrokernelTester() |
| 1466 | .m(14) |
| 1467 | .n(n) |
| 1468 | .y_zero_point(y_zero_point) |
| 1469 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1470 | } |
| 1471 | } |
| 1472 | } |
| 1473 | |
| 1474 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_2pass_fulltile_with_qmax) { |
| 1475 | TEST_REQUIRES_X86_SSE2; |
| 1476 | for (size_t n = 9; n < 16; n++) { |
| 1477 | GAvgPoolMicrokernelTester() |
| 1478 | .m(14) |
| 1479 | .n(n) |
| 1480 | .x_zero_point(128) |
| 1481 | .y_zero_point(128) |
| 1482 | .x_scale(1.0f) |
| 1483 | .y_scale(1.0f) |
| 1484 | .qmax(128) |
| 1485 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1486 | } |
| 1487 | } |
| 1488 | |
| 1489 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_2pass_fulltile_with_qmin) { |
| 1490 | TEST_REQUIRES_X86_SSE2; |
| 1491 | for (size_t n = 9; n < 16; n++) { |
| 1492 | GAvgPoolMicrokernelTester() |
| 1493 | .m(14) |
| 1494 | .n(n) |
| 1495 | .x_zero_point(128) |
| 1496 | .y_zero_point(128) |
| 1497 | .x_scale(1.0f) |
| 1498 | .y_scale(1.0f) |
| 1499 | .qmin(128) |
| 1500 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1501 | } |
| 1502 | } |
| 1503 | |
| 1504 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_2pass_subtile) { |
| 1505 | TEST_REQUIRES_X86_SSE2; |
| 1506 | for (size_t n = 9; n < 16; n++) { |
| 1507 | for (size_t m = 1; m < 7; m++) { |
| 1508 | GAvgPoolMicrokernelTester() |
| 1509 | .m(7 + m) |
| 1510 | .n(n) |
| 1511 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
| 1512 | } |
| 1513 | } |
| 1514 | } |
| 1515 | |
| 1516 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_multipass_fulltile) { |
| 1517 | TEST_REQUIRES_X86_SSE2; |
| 1518 | for (size_t n = 9; n < 16; n++) { |
| 1519 | for (size_t m = 14; m <= 35; m += 7) { |
| 1520 | GAvgPoolMicrokernelTester() |
| 1521 | .m(m) |
| 1522 | .n(n) |
| 1523 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1524 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1525 | } |
| 1526 | } |
| 1527 | |
| 1528 | TEST(Q8_GAVGPOOL_MP7p7q__SSE2, n_gt_8_multipass_fulltile_with_x_stride) { |
| 1529 | TEST_REQUIRES_X86_SSE2; |
| 1530 | for (size_t n = 9; n < 16; n++) { |
| 1531 | for (size_t m = 14; m <= 35; m += 7) { |
| 1532 | GAvgPoolMicrokernelTester() |
| 1533 | .m(m) |
| 1534 | .n(n) |
| 1535 | .x_stride(23) |
| 1536 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__sse2); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1537 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1538 | } |
| 1539 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1540 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1541 | |
| 1542 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_eq_1_fulltile) { |
| 1543 | GAvgPoolMicrokernelTester() |
| 1544 | .m(7) |
| 1545 | .n(1) |
| 1546 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1547 | } |
| 1548 | |
| 1549 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_eq_1_subtile) { |
| 1550 | for (size_t m = 1; m < 7; m++) { |
| 1551 | GAvgPoolMicrokernelTester() |
| 1552 | .m(m) |
| 1553 | .n(1) |
| 1554 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1555 | } |
| 1556 | } |
| 1557 | |
| 1558 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_eq_1_fulltile_with_x_stride) { |
| 1559 | GAvgPoolMicrokernelTester() |
| 1560 | .m(7) |
| 1561 | .n(1) |
| 1562 | .x_stride(11) |
| 1563 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1564 | } |
| 1565 | |
| 1566 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_eq_1_fulltile_with_x_scale) { |
| 1567 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 1568 | GAvgPoolMicrokernelTester() |
| 1569 | .m(7) |
| 1570 | .n(1) |
| 1571 | .x_scale(x_scale) |
| 1572 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1573 | } |
| 1574 | } |
| 1575 | |
| 1576 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_eq_1_fulltile_with_x_zero_point) { |
| 1577 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 1578 | GAvgPoolMicrokernelTester() |
| 1579 | .m(7) |
| 1580 | .n(1) |
| 1581 | .x_zero_point(x_zero_point) |
| 1582 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1583 | } |
| 1584 | } |
| 1585 | |
| 1586 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_eq_1_fulltile_with_y_scale) { |
| 1587 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 1588 | GAvgPoolMicrokernelTester() |
| 1589 | .m(7) |
| 1590 | .n(1) |
| 1591 | .y_scale(y_scale) |
| 1592 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1593 | } |
| 1594 | } |
| 1595 | |
| 1596 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_eq_1_fulltile_with_y_zero_point) { |
| 1597 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 1598 | GAvgPoolMicrokernelTester() |
| 1599 | .m(7) |
| 1600 | .n(1) |
| 1601 | .y_zero_point(y_zero_point) |
| 1602 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1603 | } |
| 1604 | } |
| 1605 | |
| 1606 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_eq_1_fulltile_with_qmax) { |
| 1607 | GAvgPoolMicrokernelTester() |
| 1608 | .m(7) |
| 1609 | .n(1) |
| 1610 | .x_zero_point(128) |
| 1611 | .y_zero_point(128) |
| 1612 | .x_scale(1.0f) |
| 1613 | .y_scale(1.0f) |
| 1614 | .qmax(128) |
| 1615 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1616 | } |
| 1617 | |
| 1618 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_eq_1_fulltile_with_qmin) { |
| 1619 | GAvgPoolMicrokernelTester() |
| 1620 | .m(7) |
| 1621 | .n(1) |
| 1622 | .x_zero_point(128) |
| 1623 | .y_zero_point(128) |
| 1624 | .x_scale(1.0f) |
| 1625 | .y_scale(1.0f) |
| 1626 | .qmin(128) |
| 1627 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1628 | } |
| 1629 | |
| 1630 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_gt_1_fulltile) { |
| 1631 | for (size_t n = 2; n < 8; n++) { |
| 1632 | GAvgPoolMicrokernelTester() |
| 1633 | .m(7) |
| 1634 | .n(n) |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1635 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1636 | } |
| 1637 | } |
| 1638 | |
| 1639 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_gt_1_subtile) { |
| 1640 | for (size_t n = 2; n < 8; n++) { |
| 1641 | for (size_t m = 1; m < 7; m++) { |
| 1642 | GAvgPoolMicrokernelTester() |
| 1643 | .m(m) |
| 1644 | .n(n) |
| 1645 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1646 | } |
| 1647 | } |
| 1648 | } |
| 1649 | |
| 1650 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_gt_1_fulltile_with_x_scale) { |
| 1651 | for (size_t n = 2; n < 8; n++) { |
| 1652 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 1653 | GAvgPoolMicrokernelTester() |
| 1654 | .m(7) |
| 1655 | .n(n) |
| 1656 | .x_scale(x_scale) |
| 1657 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1658 | } |
| 1659 | } |
| 1660 | } |
| 1661 | |
| 1662 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_gt_1_fulltile_with_x_zero_point) { |
| 1663 | for (size_t n = 2; n < 8; n++) { |
| 1664 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 1665 | GAvgPoolMicrokernelTester() |
| 1666 | .m(7) |
| 1667 | .n(n) |
| 1668 | .x_zero_point(x_zero_point) |
| 1669 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1670 | } |
| 1671 | } |
| 1672 | } |
| 1673 | |
| 1674 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_gt_1_fulltile_with_y_scale) { |
| 1675 | for (size_t n = 2; n < 8; n++) { |
| 1676 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 1677 | GAvgPoolMicrokernelTester() |
| 1678 | .m(7) |
| 1679 | .n(n) |
| 1680 | .y_scale(y_scale) |
| 1681 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1682 | } |
| 1683 | } |
| 1684 | } |
| 1685 | |
| 1686 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_gt_1_fulltile_with_y_zero_point) { |
| 1687 | for (size_t n = 2; n < 8; n++) { |
| 1688 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 1689 | GAvgPoolMicrokernelTester() |
| 1690 | .m(7) |
| 1691 | .n(n) |
| 1692 | .y_zero_point(y_zero_point) |
| 1693 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1694 | } |
| 1695 | } |
| 1696 | } |
| 1697 | |
| 1698 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_gt_1_fulltile_with_qmax) { |
| 1699 | for (size_t n = 2; n < 8; n++) { |
| 1700 | GAvgPoolMicrokernelTester() |
| 1701 | .m(7) |
| 1702 | .n(n) |
| 1703 | .x_zero_point(128) |
| 1704 | .y_zero_point(128) |
| 1705 | .x_scale(1.0f) |
| 1706 | .y_scale(1.0f) |
| 1707 | .qmax(128) |
| 1708 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1709 | } |
| 1710 | } |
| 1711 | |
| 1712 | TEST(Q8_GAVGPOOL_UP7__SCALAR, n_gt_1_fulltile_with_qmin) { |
| 1713 | for (size_t n = 2; n < 8; n++) { |
| 1714 | GAvgPoolMicrokernelTester() |
| 1715 | .m(7) |
| 1716 | .n(n) |
| 1717 | .x_zero_point(128) |
| 1718 | .y_zero_point(128) |
| 1719 | .x_scale(1.0f) |
| 1720 | .y_scale(1.0f) |
| 1721 | .qmin(128) |
| 1722 | .Test(xnn_q8_gavgpool_ukernel_up7__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1723 | } |
| 1724 | } |
| 1725 | |
| 1726 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_fulltile) { |
| 1727 | GAvgPoolMicrokernelTester() |
| 1728 | .m(14) |
| 1729 | .n(1) |
| 1730 | .nr(8) |
| 1731 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1732 | } |
| 1733 | |
| 1734 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_fulltile_with_x_stride) { |
| 1735 | GAvgPoolMicrokernelTester() |
| 1736 | .m(14) |
| 1737 | .n(1) |
| 1738 | .nr(8) |
| 1739 | .x_stride(11) |
| 1740 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1741 | } |
| 1742 | |
| 1743 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_fulltile_with_x_scale) { |
| 1744 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 1745 | GAvgPoolMicrokernelTester() |
| 1746 | .m(14) |
| 1747 | .n(1) |
| 1748 | .x_scale(x_scale) |
| 1749 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1750 | } |
| 1751 | } |
| 1752 | |
| 1753 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_fulltile_with_x_zero_point) { |
| 1754 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 1755 | GAvgPoolMicrokernelTester() |
| 1756 | .m(14) |
| 1757 | .n(1) |
| 1758 | .x_zero_point(x_zero_point) |
| 1759 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1760 | } |
| 1761 | } |
| 1762 | |
| 1763 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_fulltile_with_y_scale) { |
| 1764 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 1765 | GAvgPoolMicrokernelTester() |
| 1766 | .m(14) |
| 1767 | .n(1) |
| 1768 | .y_scale(y_scale) |
| 1769 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1770 | } |
| 1771 | } |
| 1772 | |
| 1773 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_fulltile_with_y_zero_point) { |
| 1774 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 1775 | GAvgPoolMicrokernelTester() |
| 1776 | .m(14) |
| 1777 | .n(1) |
| 1778 | .y_zero_point(y_zero_point) |
| 1779 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1780 | } |
| 1781 | } |
| 1782 | |
| 1783 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_fulltile_with_qmax) { |
| 1784 | GAvgPoolMicrokernelTester() |
| 1785 | .m(14) |
| 1786 | .n(1) |
| 1787 | .nr(8) |
| 1788 | .x_zero_point(128) |
| 1789 | .y_zero_point(128) |
| 1790 | .x_scale(1.0f) |
| 1791 | .y_scale(1.0f) |
| 1792 | .qmax(128) |
| 1793 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1794 | } |
| 1795 | |
| 1796 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_fulltile_with_qmin) { |
| 1797 | GAvgPoolMicrokernelTester() |
| 1798 | .m(14) |
| 1799 | .n(1) |
| 1800 | .nr(8) |
| 1801 | .x_zero_point(128) |
| 1802 | .y_zero_point(128) |
| 1803 | .x_scale(1.0f) |
| 1804 | .y_scale(1.0f) |
| 1805 | .qmin(128) |
| 1806 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1807 | } |
| 1808 | |
| 1809 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_subtile) { |
| 1810 | for (size_t m = 1; m < 7; m++) { |
| 1811 | GAvgPoolMicrokernelTester() |
| 1812 | .m(7 + m) |
| 1813 | .n(1) |
| 1814 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1815 | } |
| 1816 | } |
| 1817 | |
| 1818 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_2pass_subtile_with_x_stride) { |
| 1819 | for (size_t m = 1; m < 7; m++) { |
| 1820 | GAvgPoolMicrokernelTester() |
| 1821 | .m(7 + m) |
| 1822 | .n(1) |
| 1823 | .x_stride(11) |
| 1824 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1825 | } |
| 1826 | } |
| 1827 | |
| 1828 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_multipass_fulltile) { |
| 1829 | for (size_t m = 14; m <= 35; m += 7) { |
| 1830 | GAvgPoolMicrokernelTester() |
| 1831 | .m(m) |
| 1832 | .n(1) |
| 1833 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1834 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1835 | } |
| 1836 | |
| 1837 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_eq_1_multipass_fulltile_with_x_stride) { |
| 1838 | for (size_t m = 14; m <= 35; m += 7) { |
| 1839 | GAvgPoolMicrokernelTester() |
| 1840 | .m(m) |
| 1841 | .n(1) |
| 1842 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1843 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1844 | } |
| 1845 | |
| 1846 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_2pass_fulltile) { |
| 1847 | for (size_t n = 2; n < 8; n++) { |
| 1848 | GAvgPoolMicrokernelTester() |
| 1849 | .m(14) |
| 1850 | .n(n) |
| 1851 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1852 | } |
| 1853 | } |
| 1854 | |
| 1855 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_2pass_fulltile_with_x_scale) { |
| 1856 | for (float x_scale = 0.01f; x_scale < 100.0f; x_scale *= 3.14159265f) { |
| 1857 | for (size_t n = 2; n < 8; n++) { |
| 1858 | GAvgPoolMicrokernelTester() |
| 1859 | .m(14) |
| 1860 | .n(n) |
| 1861 | .x_scale(x_scale) |
| 1862 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1863 | } |
| 1864 | } |
| 1865 | } |
| 1866 | |
| 1867 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_2pass_fulltile_with_x_zero_point) { |
| 1868 | for (int32_t x_zero_point = 0; x_zero_point <= 255; x_zero_point += 51) { |
| 1869 | for (size_t n = 2; n < 8; n++) { |
| 1870 | GAvgPoolMicrokernelTester() |
| 1871 | .m(14) |
| 1872 | .n(n) |
| 1873 | .x_zero_point(x_zero_point) |
| 1874 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1875 | } |
| 1876 | } |
| 1877 | } |
| 1878 | |
| 1879 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_2pass_fulltile_with_y_scale) { |
| 1880 | for (float y_scale = 0.01f; y_scale < 100.0f; y_scale *= 3.14159265f) { |
| 1881 | for (size_t n = 2; n < 8; n++) { |
| 1882 | GAvgPoolMicrokernelTester() |
| 1883 | .m(14) |
| 1884 | .n(n) |
| 1885 | .y_scale(y_scale) |
| 1886 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1887 | } |
| 1888 | } |
| 1889 | } |
| 1890 | |
| 1891 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_2pass_fulltile_with_y_zero_point) { |
| 1892 | for (int32_t y_zero_point = 0; y_zero_point <= 255; y_zero_point += 51) { |
| 1893 | for (size_t n = 2; n < 8; n++) { |
| 1894 | GAvgPoolMicrokernelTester() |
| 1895 | .m(14) |
| 1896 | .n(n) |
| 1897 | .y_zero_point(y_zero_point) |
| 1898 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1899 | } |
| 1900 | } |
| 1901 | } |
| 1902 | |
| 1903 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_2pass_fulltile_with_qmax) { |
| 1904 | for (size_t n = 2; n < 8; n++) { |
| 1905 | GAvgPoolMicrokernelTester() |
| 1906 | .m(14) |
| 1907 | .n(n) |
| 1908 | .x_zero_point(128) |
| 1909 | .y_zero_point(128) |
| 1910 | .x_scale(1.0f) |
| 1911 | .y_scale(1.0f) |
| 1912 | .qmax(128) |
| 1913 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1914 | } |
| 1915 | } |
| 1916 | |
| 1917 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_2pass_fulltile_with_qmin) { |
| 1918 | for (size_t n = 2; n < 8; n++) { |
| 1919 | GAvgPoolMicrokernelTester() |
| 1920 | .m(14) |
| 1921 | .n(n) |
| 1922 | .x_zero_point(128) |
| 1923 | .y_zero_point(128) |
| 1924 | .x_scale(1.0f) |
| 1925 | .y_scale(1.0f) |
| 1926 | .qmin(128) |
| 1927 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1928 | } |
| 1929 | } |
| 1930 | |
| 1931 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_2pass_subtile) { |
| 1932 | for (size_t n = 2; n < 8; n++) { |
| 1933 | for (size_t m = 1; m < 7; m++) { |
| 1934 | GAvgPoolMicrokernelTester() |
| 1935 | .m(7 + m) |
| 1936 | .n(n) |
| 1937 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
| 1938 | } |
| 1939 | } |
| 1940 | } |
| 1941 | |
| 1942 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_multipass_fulltile) { |
| 1943 | for (size_t n = 2; n < 8; n++) { |
| 1944 | for (size_t m = 14; m <= 35; m += 7) { |
| 1945 | GAvgPoolMicrokernelTester() |
| 1946 | .m(m) |
| 1947 | .n(n) |
| 1948 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1949 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1950 | } |
| 1951 | } |
| 1952 | |
| 1953 | TEST(Q8_GAVGPOOL_MP7p7q__SCALAR, n_gt_1_multipass_fulltile_with_x_stride) { |
| 1954 | for (size_t n = 2; n < 8; n++) { |
| 1955 | for (size_t m = 14; m <= 35; m += 7) { |
| 1956 | GAvgPoolMicrokernelTester() |
| 1957 | .m(m) |
| 1958 | .n(n) |
| 1959 | .x_stride(23) |
| 1960 | .Test(xnn_q8_gavgpool_ukernel_mp7p7q__scalar, GAvgPoolMicrokernelTester::Variant::Scalar); |
Marat Dukhan | e0df831 | 2019-10-22 18:16:56 -0700 | [diff] [blame] | 1961 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1962 | } |
| 1963 | } |