Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | // |
| 6 | // Auto-generated file. Do not edit! |
| 7 | // Specification: test/f32-vrndne.yaml |
| 8 | // Generator: tools/generate-vunary-test.py |
| 9 | |
| 10 | |
| 11 | #include <gtest/gtest.h> |
| 12 | |
| 13 | #include <xnnpack/common.h> |
| 14 | #include <xnnpack/isa-checks.h> |
| 15 | |
| 16 | #include <xnnpack/vunary.h> |
| 17 | #include "vunary-microkernel-tester.h" |
| 18 | |
| 19 | |
| 20 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 21 | TEST(F32_VRNDNE__NEON_X4, batch_eq_4) { |
| 22 | TEST_REQUIRES_ARM_NEON; |
| 23 | VUnOpMicrokernelTester() |
| 24 | .batch_size(4) |
| 25 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 26 | } |
| 27 | |
| 28 | TEST(F32_VRNDNE__NEON_X4, batch_div_4) { |
| 29 | TEST_REQUIRES_ARM_NEON; |
| 30 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 31 | VUnOpMicrokernelTester() |
| 32 | .batch_size(batch_size) |
| 33 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 34 | } |
| 35 | } |
| 36 | |
| 37 | TEST(F32_VRNDNE__NEON_X4, batch_lt_4) { |
| 38 | TEST_REQUIRES_ARM_NEON; |
| 39 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 40 | VUnOpMicrokernelTester() |
| 41 | .batch_size(batch_size) |
| 42 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 43 | } |
| 44 | } |
| 45 | |
| 46 | TEST(F32_VRNDNE__NEON_X4, batch_gt_4) { |
| 47 | TEST_REQUIRES_ARM_NEON; |
| 48 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 49 | VUnOpMicrokernelTester() |
| 50 | .batch_size(batch_size) |
| 51 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 52 | } |
| 53 | } |
| 54 | |
| 55 | TEST(F32_VRNDNE__NEON_X4, inplace) { |
| 56 | TEST_REQUIRES_ARM_NEON; |
| 57 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 58 | VUnOpMicrokernelTester() |
| 59 | .batch_size(batch_size) |
| 60 | .inplace(true) |
| 61 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 62 | } |
| 63 | } |
| 64 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 65 | |
| 66 | |
| 67 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 68 | TEST(F32_VRNDNE__NEON_X8, batch_eq_8) { |
| 69 | TEST_REQUIRES_ARM_NEON; |
| 70 | VUnOpMicrokernelTester() |
| 71 | .batch_size(8) |
| 72 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 73 | } |
| 74 | |
| 75 | TEST(F32_VRNDNE__NEON_X8, batch_div_8) { |
| 76 | TEST_REQUIRES_ARM_NEON; |
| 77 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 78 | VUnOpMicrokernelTester() |
| 79 | .batch_size(batch_size) |
| 80 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | TEST(F32_VRNDNE__NEON_X8, batch_lt_8) { |
| 85 | TEST_REQUIRES_ARM_NEON; |
| 86 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 87 | VUnOpMicrokernelTester() |
| 88 | .batch_size(batch_size) |
| 89 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | TEST(F32_VRNDNE__NEON_X8, batch_gt_8) { |
| 94 | TEST_REQUIRES_ARM_NEON; |
| 95 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 96 | VUnOpMicrokernelTester() |
| 97 | .batch_size(batch_size) |
| 98 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | TEST(F32_VRNDNE__NEON_X8, inplace) { |
| 103 | TEST_REQUIRES_ARM_NEON; |
| 104 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 105 | VUnOpMicrokernelTester() |
| 106 | .batch_size(batch_size) |
| 107 | .inplace(true) |
| 108 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neon_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 109 | } |
| 110 | } |
| 111 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 112 | |
| 113 | |
| 114 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 115 | TEST(F32_VRNDNE__NEONV8_X4, batch_eq_4) { |
| 116 | TEST_REQUIRES_ARM_NEON_V8; |
| 117 | VUnOpMicrokernelTester() |
| 118 | .batch_size(4) |
| 119 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 120 | } |
| 121 | |
| 122 | TEST(F32_VRNDNE__NEONV8_X4, batch_div_4) { |
| 123 | TEST_REQUIRES_ARM_NEON_V8; |
| 124 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 125 | VUnOpMicrokernelTester() |
| 126 | .batch_size(batch_size) |
| 127 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | TEST(F32_VRNDNE__NEONV8_X4, batch_lt_4) { |
| 132 | TEST_REQUIRES_ARM_NEON_V8; |
| 133 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 134 | VUnOpMicrokernelTester() |
| 135 | .batch_size(batch_size) |
| 136 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 137 | } |
| 138 | } |
| 139 | |
| 140 | TEST(F32_VRNDNE__NEONV8_X4, batch_gt_4) { |
| 141 | TEST_REQUIRES_ARM_NEON_V8; |
| 142 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 143 | VUnOpMicrokernelTester() |
| 144 | .batch_size(batch_size) |
| 145 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | TEST(F32_VRNDNE__NEONV8_X4, inplace) { |
| 150 | TEST_REQUIRES_ARM_NEON_V8; |
| 151 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 152 | VUnOpMicrokernelTester() |
| 153 | .batch_size(batch_size) |
| 154 | .inplace(true) |
| 155 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 156 | } |
| 157 | } |
| 158 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 159 | |
| 160 | |
| 161 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 162 | TEST(F32_VRNDNE__NEONV8_X8, batch_eq_8) { |
| 163 | TEST_REQUIRES_ARM_NEON_V8; |
| 164 | VUnOpMicrokernelTester() |
| 165 | .batch_size(8) |
| 166 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 167 | } |
| 168 | |
| 169 | TEST(F32_VRNDNE__NEONV8_X8, batch_div_8) { |
| 170 | TEST_REQUIRES_ARM_NEON_V8; |
| 171 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 172 | VUnOpMicrokernelTester() |
| 173 | .batch_size(batch_size) |
| 174 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | TEST(F32_VRNDNE__NEONV8_X8, batch_lt_8) { |
| 179 | TEST_REQUIRES_ARM_NEON_V8; |
| 180 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 181 | VUnOpMicrokernelTester() |
| 182 | .batch_size(batch_size) |
| 183 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 184 | } |
| 185 | } |
| 186 | |
| 187 | TEST(F32_VRNDNE__NEONV8_X8, batch_gt_8) { |
| 188 | TEST_REQUIRES_ARM_NEON_V8; |
| 189 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 190 | VUnOpMicrokernelTester() |
| 191 | .batch_size(batch_size) |
| 192 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 193 | } |
| 194 | } |
| 195 | |
| 196 | TEST(F32_VRNDNE__NEONV8_X8, inplace) { |
| 197 | TEST_REQUIRES_ARM_NEON_V8; |
| 198 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 199 | VUnOpMicrokernelTester() |
| 200 | .batch_size(batch_size) |
| 201 | .inplace(true) |
| 202 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__neonv8_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 203 | } |
| 204 | } |
| 205 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 206 | |
| 207 | |
| 208 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 209 | TEST(F32_VRNDNE__SSE2_X4, batch_eq_4) { |
| 210 | TEST_REQUIRES_X86_SSE2; |
| 211 | VUnOpMicrokernelTester() |
| 212 | .batch_size(4) |
| 213 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 214 | } |
| 215 | |
| 216 | TEST(F32_VRNDNE__SSE2_X4, batch_div_4) { |
| 217 | TEST_REQUIRES_X86_SSE2; |
| 218 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 219 | VUnOpMicrokernelTester() |
| 220 | .batch_size(batch_size) |
| 221 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | TEST(F32_VRNDNE__SSE2_X4, batch_lt_4) { |
| 226 | TEST_REQUIRES_X86_SSE2; |
| 227 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 228 | VUnOpMicrokernelTester() |
| 229 | .batch_size(batch_size) |
| 230 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | TEST(F32_VRNDNE__SSE2_X4, batch_gt_4) { |
| 235 | TEST_REQUIRES_X86_SSE2; |
| 236 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 237 | VUnOpMicrokernelTester() |
| 238 | .batch_size(batch_size) |
| 239 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | TEST(F32_VRNDNE__SSE2_X4, inplace) { |
| 244 | TEST_REQUIRES_X86_SSE2; |
| 245 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 246 | VUnOpMicrokernelTester() |
| 247 | .batch_size(batch_size) |
| 248 | .inplace(true) |
| 249 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 250 | } |
| 251 | } |
| 252 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 253 | |
| 254 | |
| 255 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 256 | TEST(F32_VRNDNE__SSE2_X8, batch_eq_8) { |
| 257 | TEST_REQUIRES_X86_SSE2; |
| 258 | VUnOpMicrokernelTester() |
| 259 | .batch_size(8) |
| 260 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 261 | } |
| 262 | |
| 263 | TEST(F32_VRNDNE__SSE2_X8, batch_div_8) { |
| 264 | TEST_REQUIRES_X86_SSE2; |
| 265 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 266 | VUnOpMicrokernelTester() |
| 267 | .batch_size(batch_size) |
| 268 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 269 | } |
| 270 | } |
| 271 | |
| 272 | TEST(F32_VRNDNE__SSE2_X8, batch_lt_8) { |
| 273 | TEST_REQUIRES_X86_SSE2; |
| 274 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 275 | VUnOpMicrokernelTester() |
| 276 | .batch_size(batch_size) |
| 277 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | TEST(F32_VRNDNE__SSE2_X8, batch_gt_8) { |
| 282 | TEST_REQUIRES_X86_SSE2; |
| 283 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 284 | VUnOpMicrokernelTester() |
| 285 | .batch_size(batch_size) |
| 286 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | TEST(F32_VRNDNE__SSE2_X8, inplace) { |
| 291 | TEST_REQUIRES_X86_SSE2; |
| 292 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 293 | VUnOpMicrokernelTester() |
| 294 | .batch_size(batch_size) |
| 295 | .inplace(true) |
| 296 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse2_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 297 | } |
| 298 | } |
| 299 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 300 | |
| 301 | |
| 302 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 303 | TEST(F32_VRNDNE__SSE41_X4, batch_eq_4) { |
| 304 | TEST_REQUIRES_X86_SSE41; |
| 305 | VUnOpMicrokernelTester() |
| 306 | .batch_size(4) |
| 307 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 308 | } |
| 309 | |
| 310 | TEST(F32_VRNDNE__SSE41_X4, batch_div_4) { |
| 311 | TEST_REQUIRES_X86_SSE41; |
| 312 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 313 | VUnOpMicrokernelTester() |
| 314 | .batch_size(batch_size) |
| 315 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | TEST(F32_VRNDNE__SSE41_X4, batch_lt_4) { |
| 320 | TEST_REQUIRES_X86_SSE41; |
| 321 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 322 | VUnOpMicrokernelTester() |
| 323 | .batch_size(batch_size) |
| 324 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | TEST(F32_VRNDNE__SSE41_X4, batch_gt_4) { |
| 329 | TEST_REQUIRES_X86_SSE41; |
| 330 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 331 | VUnOpMicrokernelTester() |
| 332 | .batch_size(batch_size) |
| 333 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | TEST(F32_VRNDNE__SSE41_X4, inplace) { |
| 338 | TEST_REQUIRES_X86_SSE41; |
| 339 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 340 | VUnOpMicrokernelTester() |
| 341 | .batch_size(batch_size) |
| 342 | .inplace(true) |
| 343 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 344 | } |
| 345 | } |
| 346 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 347 | |
| 348 | |
| 349 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 350 | TEST(F32_VRNDNE__SSE41_X8, batch_eq_8) { |
| 351 | TEST_REQUIRES_X86_SSE41; |
| 352 | VUnOpMicrokernelTester() |
| 353 | .batch_size(8) |
| 354 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 355 | } |
| 356 | |
| 357 | TEST(F32_VRNDNE__SSE41_X8, batch_div_8) { |
| 358 | TEST_REQUIRES_X86_SSE41; |
| 359 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 360 | VUnOpMicrokernelTester() |
| 361 | .batch_size(batch_size) |
| 362 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 363 | } |
| 364 | } |
| 365 | |
| 366 | TEST(F32_VRNDNE__SSE41_X8, batch_lt_8) { |
| 367 | TEST_REQUIRES_X86_SSE41; |
| 368 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 369 | VUnOpMicrokernelTester() |
| 370 | .batch_size(batch_size) |
| 371 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 372 | } |
| 373 | } |
| 374 | |
| 375 | TEST(F32_VRNDNE__SSE41_X8, batch_gt_8) { |
| 376 | TEST_REQUIRES_X86_SSE41; |
| 377 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 378 | VUnOpMicrokernelTester() |
| 379 | .batch_size(batch_size) |
| 380 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 381 | } |
| 382 | } |
| 383 | |
| 384 | TEST(F32_VRNDNE__SSE41_X8, inplace) { |
| 385 | TEST_REQUIRES_X86_SSE41; |
| 386 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 387 | VUnOpMicrokernelTester() |
| 388 | .batch_size(batch_size) |
| 389 | .inplace(true) |
| 390 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__sse41_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 391 | } |
| 392 | } |
| 393 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 394 | |
| 395 | |
| 396 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 397 | TEST(F32_VRNDNE__AVX_X8, batch_eq_8) { |
| 398 | TEST_REQUIRES_X86_AVX; |
| 399 | VUnOpMicrokernelTester() |
| 400 | .batch_size(8) |
| 401 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 402 | } |
| 403 | |
| 404 | TEST(F32_VRNDNE__AVX_X8, batch_div_8) { |
| 405 | TEST_REQUIRES_X86_AVX; |
| 406 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 407 | VUnOpMicrokernelTester() |
| 408 | .batch_size(batch_size) |
| 409 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 410 | } |
| 411 | } |
| 412 | |
| 413 | TEST(F32_VRNDNE__AVX_X8, batch_lt_8) { |
| 414 | TEST_REQUIRES_X86_AVX; |
| 415 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 416 | VUnOpMicrokernelTester() |
| 417 | .batch_size(batch_size) |
| 418 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 419 | } |
| 420 | } |
| 421 | |
| 422 | TEST(F32_VRNDNE__AVX_X8, batch_gt_8) { |
| 423 | TEST_REQUIRES_X86_AVX; |
| 424 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 425 | VUnOpMicrokernelTester() |
| 426 | .batch_size(batch_size) |
| 427 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 428 | } |
| 429 | } |
| 430 | |
| 431 | TEST(F32_VRNDNE__AVX_X8, inplace) { |
| 432 | TEST_REQUIRES_X86_AVX; |
| 433 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 434 | VUnOpMicrokernelTester() |
| 435 | .batch_size(batch_size) |
| 436 | .inplace(true) |
| 437 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 438 | } |
| 439 | } |
| 440 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 441 | |
| 442 | |
| 443 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 444 | TEST(F32_VRNDNE__AVX_X16, batch_eq_16) { |
| 445 | TEST_REQUIRES_X86_AVX; |
| 446 | VUnOpMicrokernelTester() |
| 447 | .batch_size(16) |
| 448 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 449 | } |
| 450 | |
| 451 | TEST(F32_VRNDNE__AVX_X16, batch_div_16) { |
| 452 | TEST_REQUIRES_X86_AVX; |
| 453 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 454 | VUnOpMicrokernelTester() |
| 455 | .batch_size(batch_size) |
| 456 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 457 | } |
| 458 | } |
| 459 | |
| 460 | TEST(F32_VRNDNE__AVX_X16, batch_lt_16) { |
| 461 | TEST_REQUIRES_X86_AVX; |
| 462 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 463 | VUnOpMicrokernelTester() |
| 464 | .batch_size(batch_size) |
| 465 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 466 | } |
| 467 | } |
| 468 | |
| 469 | TEST(F32_VRNDNE__AVX_X16, batch_gt_16) { |
| 470 | TEST_REQUIRES_X86_AVX; |
| 471 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 472 | VUnOpMicrokernelTester() |
| 473 | .batch_size(batch_size) |
| 474 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 475 | } |
| 476 | } |
| 477 | |
| 478 | TEST(F32_VRNDNE__AVX_X16, inplace) { |
| 479 | TEST_REQUIRES_X86_AVX; |
| 480 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 481 | VUnOpMicrokernelTester() |
| 482 | .batch_size(batch_size) |
| 483 | .inplace(true) |
| 484 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 485 | } |
| 486 | } |
| 487 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 488 | |
| 489 | |
| 490 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 491 | TEST(F32_VRNDNE__AVX512F_X16, batch_eq_16) { |
| 492 | TEST_REQUIRES_X86_AVX512F; |
| 493 | VUnOpMicrokernelTester() |
| 494 | .batch_size(16) |
| 495 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 496 | } |
| 497 | |
| 498 | TEST(F32_VRNDNE__AVX512F_X16, batch_div_16) { |
| 499 | TEST_REQUIRES_X86_AVX512F; |
| 500 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 501 | VUnOpMicrokernelTester() |
| 502 | .batch_size(batch_size) |
| 503 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 504 | } |
| 505 | } |
| 506 | |
| 507 | TEST(F32_VRNDNE__AVX512F_X16, batch_lt_16) { |
| 508 | TEST_REQUIRES_X86_AVX512F; |
| 509 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 510 | VUnOpMicrokernelTester() |
| 511 | .batch_size(batch_size) |
| 512 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 513 | } |
| 514 | } |
| 515 | |
| 516 | TEST(F32_VRNDNE__AVX512F_X16, batch_gt_16) { |
| 517 | TEST_REQUIRES_X86_AVX512F; |
| 518 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 519 | VUnOpMicrokernelTester() |
| 520 | .batch_size(batch_size) |
| 521 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 522 | } |
| 523 | } |
| 524 | |
| 525 | TEST(F32_VRNDNE__AVX512F_X16, inplace) { |
| 526 | TEST_REQUIRES_X86_AVX512F; |
| 527 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 528 | VUnOpMicrokernelTester() |
| 529 | .batch_size(batch_size) |
| 530 | .inplace(true) |
| 531 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x16), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 532 | } |
| 533 | } |
| 534 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 535 | |
| 536 | |
| 537 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 538 | TEST(F32_VRNDNE__AVX512F_X32, batch_eq_32) { |
| 539 | TEST_REQUIRES_X86_AVX512F; |
| 540 | VUnOpMicrokernelTester() |
| 541 | .batch_size(32) |
| 542 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 543 | } |
| 544 | |
| 545 | TEST(F32_VRNDNE__AVX512F_X32, batch_div_32) { |
| 546 | TEST_REQUIRES_X86_AVX512F; |
| 547 | for (size_t batch_size = 64; batch_size < 320; batch_size += 32) { |
| 548 | VUnOpMicrokernelTester() |
| 549 | .batch_size(batch_size) |
| 550 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 551 | } |
| 552 | } |
| 553 | |
| 554 | TEST(F32_VRNDNE__AVX512F_X32, batch_lt_32) { |
| 555 | TEST_REQUIRES_X86_AVX512F; |
| 556 | for (size_t batch_size = 1; batch_size < 32; batch_size++) { |
| 557 | VUnOpMicrokernelTester() |
| 558 | .batch_size(batch_size) |
| 559 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 560 | } |
| 561 | } |
| 562 | |
| 563 | TEST(F32_VRNDNE__AVX512F_X32, batch_gt_32) { |
| 564 | TEST_REQUIRES_X86_AVX512F; |
| 565 | for (size_t batch_size = 33; batch_size < 64; batch_size++) { |
| 566 | VUnOpMicrokernelTester() |
| 567 | .batch_size(batch_size) |
| 568 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 569 | } |
| 570 | } |
| 571 | |
| 572 | TEST(F32_VRNDNE__AVX512F_X32, inplace) { |
| 573 | TEST_REQUIRES_X86_AVX512F; |
| 574 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
| 575 | VUnOpMicrokernelTester() |
| 576 | .batch_size(batch_size) |
| 577 | .inplace(true) |
| 578 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__avx512f_x32), VUnOpMicrokernelTester::OpType::RoundToNearestEven); |
| 579 | } |
| 580 | } |
| 581 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 582 | |
| 583 | |
| 584 | #if !XNN_ARCH_ASMJS && !XNN_ARCH_WASM && !XNN_COMPILER_MSVC && !XNN_COMPILER_ICC |
| 585 | TEST(F32_VRNDNE__PSIMD_X4, batch_eq_4) { |
| 586 | TEST_REQUIRES_PSIMD; |
| 587 | VUnOpMicrokernelTester() |
| 588 | .batch_size(4) |
| 589 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 590 | } |
| 591 | |
| 592 | TEST(F32_VRNDNE__PSIMD_X4, batch_div_4) { |
| 593 | TEST_REQUIRES_PSIMD; |
| 594 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 595 | VUnOpMicrokernelTester() |
| 596 | .batch_size(batch_size) |
| 597 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 598 | } |
| 599 | } |
| 600 | |
| 601 | TEST(F32_VRNDNE__PSIMD_X4, batch_lt_4) { |
| 602 | TEST_REQUIRES_PSIMD; |
| 603 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 604 | VUnOpMicrokernelTester() |
| 605 | .batch_size(batch_size) |
| 606 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 607 | } |
| 608 | } |
| 609 | |
| 610 | TEST(F32_VRNDNE__PSIMD_X4, batch_gt_4) { |
| 611 | TEST_REQUIRES_PSIMD; |
| 612 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 613 | VUnOpMicrokernelTester() |
| 614 | .batch_size(batch_size) |
| 615 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 616 | } |
| 617 | } |
| 618 | |
| 619 | TEST(F32_VRNDNE__PSIMD_X4, inplace) { |
| 620 | TEST_REQUIRES_PSIMD; |
| 621 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 622 | VUnOpMicrokernelTester() |
| 623 | .batch_size(batch_size) |
| 624 | .inplace(true) |
| 625 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 626 | } |
| 627 | } |
| 628 | #endif // !XNN_ARCH_ASMJS && !XNN_ARCH_WASM && !XNN_COMPILER_MSVC && !XNN_COMPILER_ICC |
| 629 | |
| 630 | |
| 631 | #if !XNN_ARCH_ASMJS && !XNN_ARCH_WASM && !XNN_COMPILER_MSVC && !XNN_COMPILER_ICC |
| 632 | TEST(F32_VRNDNE__PSIMD_X8, batch_eq_8) { |
| 633 | TEST_REQUIRES_PSIMD; |
| 634 | VUnOpMicrokernelTester() |
| 635 | .batch_size(8) |
| 636 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 637 | } |
| 638 | |
| 639 | TEST(F32_VRNDNE__PSIMD_X8, batch_div_8) { |
| 640 | TEST_REQUIRES_PSIMD; |
| 641 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 642 | VUnOpMicrokernelTester() |
| 643 | .batch_size(batch_size) |
| 644 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 645 | } |
| 646 | } |
| 647 | |
| 648 | TEST(F32_VRNDNE__PSIMD_X8, batch_lt_8) { |
| 649 | TEST_REQUIRES_PSIMD; |
| 650 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 651 | VUnOpMicrokernelTester() |
| 652 | .batch_size(batch_size) |
| 653 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 654 | } |
| 655 | } |
| 656 | |
| 657 | TEST(F32_VRNDNE__PSIMD_X8, batch_gt_8) { |
| 658 | TEST_REQUIRES_PSIMD; |
| 659 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 660 | VUnOpMicrokernelTester() |
| 661 | .batch_size(batch_size) |
| 662 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | TEST(F32_VRNDNE__PSIMD_X8, inplace) { |
| 667 | TEST_REQUIRES_PSIMD; |
| 668 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 669 | VUnOpMicrokernelTester() |
| 670 | .batch_size(batch_size) |
| 671 | .inplace(true) |
| 672 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__psimd_x8), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 673 | } |
| 674 | } |
| 675 | #endif // !XNN_ARCH_ASMJS && !XNN_ARCH_WASM && !XNN_COMPILER_MSVC && !XNN_COMPILER_ICC |
| 676 | |
| 677 | |
| 678 | TEST(F32_VRNDNE__SCALAR_LIBM_X1, batch_eq_1) { |
| 679 | VUnOpMicrokernelTester() |
| 680 | .batch_size(1) |
| 681 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x1), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 682 | } |
| 683 | |
| 684 | TEST(F32_VRNDNE__SCALAR_LIBM_X1, batch_gt_1) { |
| 685 | for (size_t batch_size = 2; batch_size < 10; batch_size++) { |
| 686 | VUnOpMicrokernelTester() |
| 687 | .batch_size(batch_size) |
| 688 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x1), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 689 | } |
| 690 | } |
| 691 | |
| 692 | TEST(F32_VRNDNE__SCALAR_LIBM_X1, inplace) { |
| 693 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
| 694 | VUnOpMicrokernelTester() |
| 695 | .batch_size(batch_size) |
| 696 | .inplace(true) |
| 697 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x1), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 698 | } |
| 699 | } |
| 700 | |
Marat Dukhan | af45248 | 2020-06-10 22:17:00 -0700 | [diff] [blame^] | 701 | |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 702 | TEST(F32_VRNDNE__SCALAR_LIBM_X2, batch_eq_2) { |
| 703 | VUnOpMicrokernelTester() |
| 704 | .batch_size(2) |
| 705 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x2), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 706 | } |
| 707 | |
| 708 | TEST(F32_VRNDNE__SCALAR_LIBM_X2, batch_div_2) { |
| 709 | for (size_t batch_size = 4; batch_size < 20; batch_size += 2) { |
| 710 | VUnOpMicrokernelTester() |
| 711 | .batch_size(batch_size) |
| 712 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x2), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 713 | } |
| 714 | } |
| 715 | |
| 716 | TEST(F32_VRNDNE__SCALAR_LIBM_X2, batch_lt_2) { |
| 717 | for (size_t batch_size = 1; batch_size < 2; batch_size++) { |
| 718 | VUnOpMicrokernelTester() |
| 719 | .batch_size(batch_size) |
| 720 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x2), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 721 | } |
| 722 | } |
| 723 | |
| 724 | TEST(F32_VRNDNE__SCALAR_LIBM_X2, batch_gt_2) { |
| 725 | for (size_t batch_size = 3; batch_size < 4; batch_size++) { |
| 726 | VUnOpMicrokernelTester() |
| 727 | .batch_size(batch_size) |
| 728 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x2), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 729 | } |
| 730 | } |
| 731 | |
| 732 | TEST(F32_VRNDNE__SCALAR_LIBM_X2, inplace) { |
| 733 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
| 734 | VUnOpMicrokernelTester() |
| 735 | .batch_size(batch_size) |
| 736 | .inplace(true) |
| 737 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x2), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 738 | } |
| 739 | } |
| 740 | |
Marat Dukhan | af45248 | 2020-06-10 22:17:00 -0700 | [diff] [blame^] | 741 | |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 742 | TEST(F32_VRNDNE__SCALAR_LIBM_X4, batch_eq_4) { |
| 743 | VUnOpMicrokernelTester() |
| 744 | .batch_size(4) |
| 745 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 746 | } |
| 747 | |
| 748 | TEST(F32_VRNDNE__SCALAR_LIBM_X4, batch_div_4) { |
| 749 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 750 | VUnOpMicrokernelTester() |
| 751 | .batch_size(batch_size) |
| 752 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 753 | } |
| 754 | } |
| 755 | |
| 756 | TEST(F32_VRNDNE__SCALAR_LIBM_X4, batch_lt_4) { |
| 757 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 758 | VUnOpMicrokernelTester() |
| 759 | .batch_size(batch_size) |
| 760 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 761 | } |
| 762 | } |
| 763 | |
| 764 | TEST(F32_VRNDNE__SCALAR_LIBM_X4, batch_gt_4) { |
| 765 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 766 | VUnOpMicrokernelTester() |
| 767 | .batch_size(batch_size) |
| 768 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 769 | } |
| 770 | } |
| 771 | |
| 772 | TEST(F32_VRNDNE__SCALAR_LIBM_X4, inplace) { |
| 773 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 774 | VUnOpMicrokernelTester() |
| 775 | .batch_size(batch_size) |
| 776 | .inplace(true) |
| 777 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vrndne_ukernel__scalar_libm_x4), VUnOpMicrokernelTester::OpType::RoundToNearestEven, VUnOpMicrokernelTester::Variant::Scalar); |
| 778 | } |
Marat Dukhan | af45248 | 2020-06-10 22:17:00 -0700 | [diff] [blame^] | 779 | } |