Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 1 | // Copyright 2021 Google LLC |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 5 | // |
| 6 | // Auto-generated file. Do not edit! |
| 7 | // Specification: test/x8-lut.yaml |
| 8 | // Generator: tools/generate-lut-test.py |
| 9 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 10 | |
| 11 | #include <gtest/gtest.h> |
| 12 | |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 13 | #include <xnnpack/common.h> |
| 14 | #include <xnnpack/isa-checks.h> |
| 15 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 16 | #include <xnnpack/lut.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 17 | #include "lut-microkernel-tester.h" |
| 18 | |
| 19 | |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 20 | TEST(X8_LUT__SCALAR_X1, batch_eq_1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 21 | LUTMicrokernelTester() |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 22 | .batch_size(1) |
| 23 | .Test(xnn_x8_lut_ukernel__scalar_x1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | } |
| 25 | |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 26 | TEST(X8_LUT__SCALAR_X1, batch_gt_1) { |
| 27 | for (size_t batch_size = 2; batch_size < 10; batch_size++) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 28 | LUTMicrokernelTester() |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 29 | .batch_size(batch_size) |
| 30 | .Test(xnn_x8_lut_ukernel__scalar_x1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 31 | } |
| 32 | } |
| 33 | |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 34 | TEST(X8_LUT__SCALAR_X1, inplace) { |
| 35 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 36 | LUTMicrokernelTester() |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 37 | .batch_size(batch_size) |
| 38 | .inplace(true) |
| 39 | .Test(xnn_x8_lut_ukernel__scalar_x1); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 40 | } |
| 41 | } |
| 42 | |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 43 | TEST(X8_LUT__SCALAR_X2, batch_eq_2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 44 | LUTMicrokernelTester() |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 45 | .batch_size(2) |
| 46 | .Test(xnn_x8_lut_ukernel__scalar_x2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | } |
| 48 | |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 49 | TEST(X8_LUT__SCALAR_X2, batch_div_2) { |
| 50 | for (size_t batch_size = 4; batch_size < 20; batch_size += 2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 51 | LUTMicrokernelTester() |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 52 | .batch_size(batch_size) |
| 53 | .Test(xnn_x8_lut_ukernel__scalar_x2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 54 | } |
| 55 | } |
| 56 | |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 57 | TEST(X8_LUT__SCALAR_X2, batch_lt_2) { |
| 58 | for (size_t batch_size = 1; batch_size < 2; batch_size++) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 59 | LUTMicrokernelTester() |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 60 | .batch_size(batch_size) |
| 61 | .Test(xnn_x8_lut_ukernel__scalar_x2); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 62 | } |
| 63 | } |
Marat Dukhan | d67539d | 2021-09-08 23:06:03 -0700 | [diff] [blame] | 64 | |
| 65 | TEST(X8_LUT__SCALAR_X2, batch_gt_2) { |
| 66 | for (size_t batch_size = 3; batch_size < 4; batch_size++) { |
| 67 | LUTMicrokernelTester() |
| 68 | .batch_size(batch_size) |
| 69 | .Test(xnn_x8_lut_ukernel__scalar_x2); |
| 70 | } |
| 71 | } |
| 72 | |
| 73 | TEST(X8_LUT__SCALAR_X2, inplace) { |
| 74 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
| 75 | LUTMicrokernelTester() |
| 76 | .batch_size(batch_size) |
| 77 | .inplace(true) |
| 78 | .Test(xnn_x8_lut_ukernel__scalar_x2); |
| 79 | } |
| 80 | } |
| 81 | |
| 82 | TEST(X8_LUT__SCALAR_X4, batch_eq_4) { |
| 83 | LUTMicrokernelTester() |
| 84 | .batch_size(4) |
| 85 | .Test(xnn_x8_lut_ukernel__scalar_x4); |
| 86 | } |
| 87 | |
| 88 | TEST(X8_LUT__SCALAR_X4, batch_div_4) { |
| 89 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 90 | LUTMicrokernelTester() |
| 91 | .batch_size(batch_size) |
| 92 | .Test(xnn_x8_lut_ukernel__scalar_x4); |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | TEST(X8_LUT__SCALAR_X4, batch_lt_4) { |
| 97 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 98 | LUTMicrokernelTester() |
| 99 | .batch_size(batch_size) |
| 100 | .Test(xnn_x8_lut_ukernel__scalar_x4); |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | TEST(X8_LUT__SCALAR_X4, batch_gt_4) { |
| 105 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 106 | LUTMicrokernelTester() |
| 107 | .batch_size(batch_size) |
| 108 | .Test(xnn_x8_lut_ukernel__scalar_x4); |
| 109 | } |
| 110 | } |
| 111 | |
| 112 | TEST(X8_LUT__SCALAR_X4, inplace) { |
| 113 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 114 | LUTMicrokernelTester() |
| 115 | .batch_size(batch_size) |
| 116 | .inplace(true) |
| 117 | .Test(xnn_x8_lut_ukernel__scalar_x4); |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | TEST(X8_LUT__SCALAR_X8, batch_eq_8) { |
| 122 | LUTMicrokernelTester() |
| 123 | .batch_size(8) |
| 124 | .Test(xnn_x8_lut_ukernel__scalar_x8); |
| 125 | } |
| 126 | |
| 127 | TEST(X8_LUT__SCALAR_X8, batch_div_8) { |
| 128 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 129 | LUTMicrokernelTester() |
| 130 | .batch_size(batch_size) |
| 131 | .Test(xnn_x8_lut_ukernel__scalar_x8); |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | TEST(X8_LUT__SCALAR_X8, batch_lt_8) { |
| 136 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 137 | LUTMicrokernelTester() |
| 138 | .batch_size(batch_size) |
| 139 | .Test(xnn_x8_lut_ukernel__scalar_x8); |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | TEST(X8_LUT__SCALAR_X8, batch_gt_8) { |
| 144 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 145 | LUTMicrokernelTester() |
| 146 | .batch_size(batch_size) |
| 147 | .Test(xnn_x8_lut_ukernel__scalar_x8); |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | TEST(X8_LUT__SCALAR_X8, inplace) { |
| 152 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 153 | LUTMicrokernelTester() |
| 154 | .batch_size(batch_size) |
| 155 | .inplace(true) |
| 156 | .Test(xnn_x8_lut_ukernel__scalar_x8); |
| 157 | } |
| 158 | } |
| 159 | |
| 160 | TEST(X8_LUT__SCALAR_X16, batch_eq_16) { |
| 161 | LUTMicrokernelTester() |
| 162 | .batch_size(16) |
| 163 | .Test(xnn_x8_lut_ukernel__scalar_x16); |
| 164 | } |
| 165 | |
| 166 | TEST(X8_LUT__SCALAR_X16, batch_div_16) { |
| 167 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 168 | LUTMicrokernelTester() |
| 169 | .batch_size(batch_size) |
| 170 | .Test(xnn_x8_lut_ukernel__scalar_x16); |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | TEST(X8_LUT__SCALAR_X16, batch_lt_16) { |
| 175 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 176 | LUTMicrokernelTester() |
| 177 | .batch_size(batch_size) |
| 178 | .Test(xnn_x8_lut_ukernel__scalar_x16); |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | TEST(X8_LUT__SCALAR_X16, batch_gt_16) { |
| 183 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 184 | LUTMicrokernelTester() |
| 185 | .batch_size(batch_size) |
| 186 | .Test(xnn_x8_lut_ukernel__scalar_x16); |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | TEST(X8_LUT__SCALAR_X16, inplace) { |
| 191 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 192 | LUTMicrokernelTester() |
| 193 | .batch_size(batch_size) |
| 194 | .inplace(true) |
| 195 | .Test(xnn_x8_lut_ukernel__scalar_x16); |
| 196 | } |
Marat Dukhan | f718232 | 2021-09-09 18:53:46 -0700 | [diff] [blame] | 197 | } |
| 198 | |
| 199 | #if XNN_ARCH_ARM64 |
| 200 | TEST(X8_LUT__NEON_TBX128X4_X16, batch_eq_16) { |
| 201 | TEST_REQUIRES_ARM_NEON; |
| 202 | LUTMicrokernelTester() |
| 203 | .batch_size(16) |
| 204 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x16); |
| 205 | } |
| 206 | |
| 207 | TEST(X8_LUT__NEON_TBX128X4_X16, batch_div_16) { |
| 208 | TEST_REQUIRES_ARM_NEON; |
| 209 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 210 | LUTMicrokernelTester() |
| 211 | .batch_size(batch_size) |
| 212 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x16); |
| 213 | } |
| 214 | } |
| 215 | |
| 216 | TEST(X8_LUT__NEON_TBX128X4_X16, batch_lt_16) { |
| 217 | TEST_REQUIRES_ARM_NEON; |
| 218 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 219 | LUTMicrokernelTester() |
| 220 | .batch_size(batch_size) |
| 221 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x16); |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | TEST(X8_LUT__NEON_TBX128X4_X16, batch_gt_16) { |
| 226 | TEST_REQUIRES_ARM_NEON; |
| 227 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 228 | LUTMicrokernelTester() |
| 229 | .batch_size(batch_size) |
| 230 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x16); |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | TEST(X8_LUT__NEON_TBX128X4_X16, inplace) { |
| 235 | TEST_REQUIRES_ARM_NEON; |
| 236 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 237 | LUTMicrokernelTester() |
| 238 | .batch_size(batch_size) |
| 239 | .inplace(true) |
| 240 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x16); |
| 241 | } |
| 242 | } |
| 243 | #endif // XNN_ARCH_ARM64 |
| 244 | |
| 245 | |
| 246 | #if XNN_ARCH_ARM64 |
| 247 | TEST(X8_LUT__NEON_TBX128X4_X32, batch_eq_32) { |
| 248 | TEST_REQUIRES_ARM_NEON; |
| 249 | LUTMicrokernelTester() |
| 250 | .batch_size(32) |
| 251 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x32); |
| 252 | } |
| 253 | |
| 254 | TEST(X8_LUT__NEON_TBX128X4_X32, batch_div_32) { |
| 255 | TEST_REQUIRES_ARM_NEON; |
| 256 | for (size_t batch_size = 64; batch_size < 320; batch_size += 32) { |
| 257 | LUTMicrokernelTester() |
| 258 | .batch_size(batch_size) |
| 259 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x32); |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | TEST(X8_LUT__NEON_TBX128X4_X32, batch_lt_32) { |
| 264 | TEST_REQUIRES_ARM_NEON; |
| 265 | for (size_t batch_size = 1; batch_size < 32; batch_size++) { |
| 266 | LUTMicrokernelTester() |
| 267 | .batch_size(batch_size) |
| 268 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x32); |
| 269 | } |
| 270 | } |
| 271 | |
| 272 | TEST(X8_LUT__NEON_TBX128X4_X32, batch_gt_32) { |
| 273 | TEST_REQUIRES_ARM_NEON; |
| 274 | for (size_t batch_size = 33; batch_size < 64; batch_size++) { |
| 275 | LUTMicrokernelTester() |
| 276 | .batch_size(batch_size) |
| 277 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x32); |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | TEST(X8_LUT__NEON_TBX128X4_X32, inplace) { |
| 282 | TEST_REQUIRES_ARM_NEON; |
| 283 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
| 284 | LUTMicrokernelTester() |
| 285 | .batch_size(batch_size) |
| 286 | .inplace(true) |
| 287 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x32); |
| 288 | } |
| 289 | } |
| 290 | #endif // XNN_ARCH_ARM64 |
| 291 | |
| 292 | |
| 293 | #if XNN_ARCH_ARM64 |
| 294 | TEST(X8_LUT__NEON_TBX128X4_X48, batch_eq_48) { |
| 295 | TEST_REQUIRES_ARM_NEON; |
| 296 | LUTMicrokernelTester() |
| 297 | .batch_size(48) |
| 298 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x48); |
| 299 | } |
| 300 | |
| 301 | TEST(X8_LUT__NEON_TBX128X4_X48, batch_div_48) { |
| 302 | TEST_REQUIRES_ARM_NEON; |
| 303 | for (size_t batch_size = 96; batch_size < 480; batch_size += 48) { |
| 304 | LUTMicrokernelTester() |
| 305 | .batch_size(batch_size) |
| 306 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x48); |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | TEST(X8_LUT__NEON_TBX128X4_X48, batch_lt_48) { |
| 311 | TEST_REQUIRES_ARM_NEON; |
| 312 | for (size_t batch_size = 1; batch_size < 48; batch_size++) { |
| 313 | LUTMicrokernelTester() |
| 314 | .batch_size(batch_size) |
| 315 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x48); |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | TEST(X8_LUT__NEON_TBX128X4_X48, batch_gt_48) { |
| 320 | TEST_REQUIRES_ARM_NEON; |
| 321 | for (size_t batch_size = 49; batch_size < 96; batch_size++) { |
| 322 | LUTMicrokernelTester() |
| 323 | .batch_size(batch_size) |
| 324 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x48); |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | TEST(X8_LUT__NEON_TBX128X4_X48, inplace) { |
| 329 | TEST_REQUIRES_ARM_NEON; |
| 330 | for (size_t batch_size = 1; batch_size <= 240; batch_size += 47) { |
| 331 | LUTMicrokernelTester() |
| 332 | .batch_size(batch_size) |
| 333 | .inplace(true) |
| 334 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x48); |
| 335 | } |
| 336 | } |
| 337 | #endif // XNN_ARCH_ARM64 |
| 338 | |
| 339 | |
| 340 | #if XNN_ARCH_ARM64 |
| 341 | TEST(X8_LUT__NEON_TBX128X4_X64, batch_eq_64) { |
| 342 | TEST_REQUIRES_ARM_NEON; |
| 343 | LUTMicrokernelTester() |
| 344 | .batch_size(64) |
| 345 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x64); |
| 346 | } |
| 347 | |
| 348 | TEST(X8_LUT__NEON_TBX128X4_X64, batch_div_64) { |
| 349 | TEST_REQUIRES_ARM_NEON; |
| 350 | for (size_t batch_size = 128; batch_size < 640; batch_size += 64) { |
| 351 | LUTMicrokernelTester() |
| 352 | .batch_size(batch_size) |
| 353 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x64); |
| 354 | } |
| 355 | } |
| 356 | |
| 357 | TEST(X8_LUT__NEON_TBX128X4_X64, batch_lt_64) { |
| 358 | TEST_REQUIRES_ARM_NEON; |
| 359 | for (size_t batch_size = 1; batch_size < 64; batch_size++) { |
| 360 | LUTMicrokernelTester() |
| 361 | .batch_size(batch_size) |
| 362 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x64); |
| 363 | } |
| 364 | } |
| 365 | |
| 366 | TEST(X8_LUT__NEON_TBX128X4_X64, batch_gt_64) { |
| 367 | TEST_REQUIRES_ARM_NEON; |
| 368 | for (size_t batch_size = 65; batch_size < 128; batch_size++) { |
| 369 | LUTMicrokernelTester() |
| 370 | .batch_size(batch_size) |
| 371 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x64); |
| 372 | } |
| 373 | } |
| 374 | |
| 375 | TEST(X8_LUT__NEON_TBX128X4_X64, inplace) { |
| 376 | TEST_REQUIRES_ARM_NEON; |
| 377 | for (size_t batch_size = 1; batch_size <= 320; batch_size += 63) { |
| 378 | LUTMicrokernelTester() |
| 379 | .batch_size(batch_size) |
| 380 | .inplace(true) |
| 381 | .Test(xnn_x8_lut_ukernel__neon_tbx128x4_x64); |
| 382 | } |
| 383 | } |
| 384 | #endif // XNN_ARCH_ARM64 |