Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -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-vsqrt.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_ARM64 |
| 21 | TEST(F32_VSQRT__NEON_SQRT_X4, batch_eq_4) { |
| 22 | TEST_REQUIRES_ARM_NEON; |
| 23 | VUnOpMicrokernelTester() |
| 24 | .batch_size(4) |
| 25 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neon_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 26 | } |
| 27 | |
| 28 | TEST(F32_VSQRT__NEON_SQRT_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_vsqrt_ukernel__neon_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 34 | } |
| 35 | } |
| 36 | |
| 37 | TEST(F32_VSQRT__NEON_SQRT_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_vsqrt_ukernel__neon_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 43 | } |
| 44 | } |
| 45 | |
| 46 | TEST(F32_VSQRT__NEON_SQRT_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_vsqrt_ukernel__neon_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 52 | } |
| 53 | } |
| 54 | |
| 55 | TEST(F32_VSQRT__NEON_SQRT_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_vsqrt_ukernel__neon_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 62 | } |
| 63 | } |
| 64 | #endif // XNN_ARCH_ARM64 |
| 65 | |
| 66 | |
| 67 | #if XNN_ARCH_ARM64 |
| 68 | TEST(F32_VSQRT__NEON_SQRT_X8, batch_eq_8) { |
| 69 | TEST_REQUIRES_ARM_NEON; |
| 70 | VUnOpMicrokernelTester() |
| 71 | .batch_size(8) |
| 72 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neon_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 73 | } |
| 74 | |
| 75 | TEST(F32_VSQRT__NEON_SQRT_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_vsqrt_ukernel__neon_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | TEST(F32_VSQRT__NEON_SQRT_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_vsqrt_ukernel__neon_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | TEST(F32_VSQRT__NEON_SQRT_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_vsqrt_ukernel__neon_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | TEST(F32_VSQRT__NEON_SQRT_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_vsqrt_ukernel__neon_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 109 | } |
| 110 | } |
| 111 | #endif // XNN_ARCH_ARM64 |
| 112 | |
| 113 | |
| 114 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 115 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X4, batch_eq_4) { |
| 116 | TEST_REQUIRES_ARM_NEON_FMA; |
| 117 | VUnOpMicrokernelTester() |
| 118 | .batch_size(4) |
| 119 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 120 | } |
| 121 | |
| 122 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X4, batch_div_4) { |
| 123 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X4, batch_lt_4) { |
| 132 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 137 | } |
| 138 | } |
| 139 | |
| 140 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X4, batch_gt_4) { |
| 141 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X4, inplace) { |
| 150 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 156 | } |
| 157 | } |
| 158 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 159 | |
| 160 | |
| 161 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 162 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X8, batch_eq_8) { |
| 163 | TEST_REQUIRES_ARM_NEON_FMA; |
| 164 | VUnOpMicrokernelTester() |
| 165 | .batch_size(8) |
| 166 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 167 | } |
| 168 | |
| 169 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X8, batch_div_8) { |
| 170 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X8, batch_lt_8) { |
| 179 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 184 | } |
| 185 | } |
| 186 | |
| 187 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X8, batch_gt_8) { |
| 188 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 193 | } |
| 194 | } |
| 195 | |
| 196 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X8, inplace) { |
| 197 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 203 | } |
| 204 | } |
| 205 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 206 | |
| 207 | |
| 208 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 209 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X12, batch_eq_12) { |
| 210 | TEST_REQUIRES_ARM_NEON_FMA; |
| 211 | VUnOpMicrokernelTester() |
| 212 | .batch_size(12) |
| 213 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 214 | } |
| 215 | |
| 216 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X12, batch_div_12) { |
| 217 | TEST_REQUIRES_ARM_NEON_FMA; |
| 218 | for (size_t batch_size = 24; batch_size < 120; batch_size += 12) { |
| 219 | VUnOpMicrokernelTester() |
| 220 | .batch_size(batch_size) |
| 221 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X12, batch_lt_12) { |
| 226 | TEST_REQUIRES_ARM_NEON_FMA; |
| 227 | for (size_t batch_size = 1; batch_size < 12; batch_size++) { |
| 228 | VUnOpMicrokernelTester() |
| 229 | .batch_size(batch_size) |
| 230 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X12, batch_gt_12) { |
| 235 | TEST_REQUIRES_ARM_NEON_FMA; |
| 236 | for (size_t batch_size = 13; batch_size < 24; batch_size++) { |
| 237 | VUnOpMicrokernelTester() |
| 238 | .batch_size(batch_size) |
| 239 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X12, inplace) { |
| 244 | TEST_REQUIRES_ARM_NEON_FMA; |
| 245 | for (size_t batch_size = 1; batch_size <= 60; batch_size += 11) { |
| 246 | VUnOpMicrokernelTester() |
| 247 | .batch_size(batch_size) |
| 248 | .inplace(true) |
| 249 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 250 | } |
| 251 | } |
| 252 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 253 | |
| 254 | |
| 255 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 256 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X16, batch_eq_16) { |
| 257 | TEST_REQUIRES_ARM_NEON_FMA; |
| 258 | VUnOpMicrokernelTester() |
| 259 | .batch_size(16) |
| 260 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 261 | } |
| 262 | |
| 263 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X16, batch_div_16) { |
| 264 | TEST_REQUIRES_ARM_NEON_FMA; |
| 265 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 266 | VUnOpMicrokernelTester() |
| 267 | .batch_size(batch_size) |
| 268 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 269 | } |
| 270 | } |
| 271 | |
| 272 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X16, batch_lt_16) { |
| 273 | TEST_REQUIRES_ARM_NEON_FMA; |
| 274 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 275 | VUnOpMicrokernelTester() |
| 276 | .batch_size(batch_size) |
| 277 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X16, batch_gt_16) { |
| 282 | TEST_REQUIRES_ARM_NEON_FMA; |
| 283 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 284 | VUnOpMicrokernelTester() |
| 285 | .batch_size(batch_size) |
| 286 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X16, inplace) { |
| 291 | TEST_REQUIRES_ARM_NEON_FMA; |
| 292 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 293 | VUnOpMicrokernelTester() |
| 294 | .batch_size(batch_size) |
| 295 | .inplace(true) |
| 296 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 297 | } |
| 298 | } |
| 299 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 300 | |
| 301 | |
| 302 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 303 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X20, batch_eq_20) { |
| 304 | TEST_REQUIRES_ARM_NEON_FMA; |
| 305 | VUnOpMicrokernelTester() |
| 306 | .batch_size(20) |
| 307 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 308 | } |
| 309 | |
| 310 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X20, batch_div_20) { |
| 311 | TEST_REQUIRES_ARM_NEON_FMA; |
| 312 | for (size_t batch_size = 40; batch_size < 200; batch_size += 20) { |
| 313 | VUnOpMicrokernelTester() |
| 314 | .batch_size(batch_size) |
| 315 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X20, batch_lt_20) { |
| 320 | TEST_REQUIRES_ARM_NEON_FMA; |
| 321 | for (size_t batch_size = 1; batch_size < 20; batch_size++) { |
| 322 | VUnOpMicrokernelTester() |
| 323 | .batch_size(batch_size) |
| 324 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X20, batch_gt_20) { |
| 329 | TEST_REQUIRES_ARM_NEON_FMA; |
| 330 | for (size_t batch_size = 21; batch_size < 40; batch_size++) { |
| 331 | VUnOpMicrokernelTester() |
| 332 | .batch_size(batch_size) |
| 333 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X20, inplace) { |
| 338 | TEST_REQUIRES_ARM_NEON_FMA; |
| 339 | for (size_t batch_size = 1; batch_size <= 100; batch_size += 19) { |
| 340 | VUnOpMicrokernelTester() |
| 341 | .batch_size(batch_size) |
| 342 | .inplace(true) |
| 343 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 344 | } |
| 345 | } |
| 346 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 347 | |
| 348 | |
| 349 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 350 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X24, batch_eq_24) { |
| 351 | TEST_REQUIRES_ARM_NEON_FMA; |
| 352 | VUnOpMicrokernelTester() |
| 353 | .batch_size(24) |
| 354 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 355 | } |
| 356 | |
| 357 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X24, batch_div_24) { |
| 358 | TEST_REQUIRES_ARM_NEON_FMA; |
| 359 | for (size_t batch_size = 48; batch_size < 240; batch_size += 24) { |
| 360 | VUnOpMicrokernelTester() |
| 361 | .batch_size(batch_size) |
| 362 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 363 | } |
| 364 | } |
| 365 | |
| 366 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X24, batch_lt_24) { |
| 367 | TEST_REQUIRES_ARM_NEON_FMA; |
| 368 | for (size_t batch_size = 1; batch_size < 24; batch_size++) { |
| 369 | VUnOpMicrokernelTester() |
| 370 | .batch_size(batch_size) |
| 371 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 372 | } |
| 373 | } |
| 374 | |
| 375 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X24, batch_gt_24) { |
| 376 | TEST_REQUIRES_ARM_NEON_FMA; |
| 377 | for (size_t batch_size = 25; batch_size < 48; batch_size++) { |
| 378 | VUnOpMicrokernelTester() |
| 379 | .batch_size(batch_size) |
| 380 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 381 | } |
| 382 | } |
| 383 | |
| 384 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X24, inplace) { |
| 385 | TEST_REQUIRES_ARM_NEON_FMA; |
| 386 | for (size_t batch_size = 1; batch_size <= 120; batch_size += 23) { |
| 387 | VUnOpMicrokernelTester() |
| 388 | .batch_size(batch_size) |
| 389 | .inplace(true) |
| 390 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 391 | } |
| 392 | } |
| 393 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 394 | |
| 395 | |
| 396 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 397 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X28, batch_eq_28) { |
| 398 | TEST_REQUIRES_ARM_NEON_FMA; |
| 399 | VUnOpMicrokernelTester() |
| 400 | .batch_size(28) |
| 401 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 402 | } |
| 403 | |
| 404 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X28, batch_div_28) { |
| 405 | TEST_REQUIRES_ARM_NEON_FMA; |
| 406 | for (size_t batch_size = 56; batch_size < 280; batch_size += 28) { |
| 407 | VUnOpMicrokernelTester() |
| 408 | .batch_size(batch_size) |
| 409 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 410 | } |
| 411 | } |
| 412 | |
| 413 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X28, batch_lt_28) { |
| 414 | TEST_REQUIRES_ARM_NEON_FMA; |
| 415 | for (size_t batch_size = 1; batch_size < 28; batch_size++) { |
| 416 | VUnOpMicrokernelTester() |
| 417 | .batch_size(batch_size) |
| 418 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 419 | } |
| 420 | } |
| 421 | |
| 422 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X28, batch_gt_28) { |
| 423 | TEST_REQUIRES_ARM_NEON_FMA; |
| 424 | for (size_t batch_size = 29; batch_size < 56; batch_size++) { |
| 425 | VUnOpMicrokernelTester() |
| 426 | .batch_size(batch_size) |
| 427 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 428 | } |
| 429 | } |
| 430 | |
| 431 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X28, inplace) { |
| 432 | TEST_REQUIRES_ARM_NEON_FMA; |
| 433 | for (size_t batch_size = 1; batch_size <= 140; batch_size += 27) { |
| 434 | VUnOpMicrokernelTester() |
| 435 | .batch_size(batch_size) |
| 436 | .inplace(true) |
| 437 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 438 | } |
| 439 | } |
| 440 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 441 | |
| 442 | |
| 443 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 444 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X32, batch_eq_32) { |
| 445 | TEST_REQUIRES_ARM_NEON_FMA; |
| 446 | VUnOpMicrokernelTester() |
| 447 | .batch_size(32) |
| 448 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 449 | } |
| 450 | |
| 451 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X32, batch_div_32) { |
| 452 | TEST_REQUIRES_ARM_NEON_FMA; |
| 453 | for (size_t batch_size = 64; batch_size < 320; batch_size += 32) { |
| 454 | VUnOpMicrokernelTester() |
| 455 | .batch_size(batch_size) |
| 456 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 457 | } |
| 458 | } |
| 459 | |
| 460 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X32, batch_lt_32) { |
| 461 | TEST_REQUIRES_ARM_NEON_FMA; |
| 462 | for (size_t batch_size = 1; batch_size < 32; batch_size++) { |
| 463 | VUnOpMicrokernelTester() |
| 464 | .batch_size(batch_size) |
| 465 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 466 | } |
| 467 | } |
| 468 | |
| 469 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X32, batch_gt_32) { |
| 470 | TEST_REQUIRES_ARM_NEON_FMA; |
| 471 | for (size_t batch_size = 33; batch_size < 64; batch_size++) { |
| 472 | VUnOpMicrokernelTester() |
| 473 | .batch_size(batch_size) |
| 474 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 475 | } |
| 476 | } |
| 477 | |
| 478 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X32, inplace) { |
| 479 | TEST_REQUIRES_ARM_NEON_FMA; |
| 480 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
| 481 | VUnOpMicrokernelTester() |
| 482 | .batch_size(batch_size) |
| 483 | .inplace(true) |
| 484 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 485 | } |
| 486 | } |
| 487 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 488 | |
| 489 | |
| 490 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 491 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X36, batch_eq_36) { |
| 492 | TEST_REQUIRES_ARM_NEON_FMA; |
| 493 | VUnOpMicrokernelTester() |
| 494 | .batch_size(36) |
| 495 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 496 | } |
| 497 | |
| 498 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X36, batch_div_36) { |
| 499 | TEST_REQUIRES_ARM_NEON_FMA; |
| 500 | for (size_t batch_size = 72; batch_size < 360; batch_size += 36) { |
| 501 | VUnOpMicrokernelTester() |
| 502 | .batch_size(batch_size) |
| 503 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 504 | } |
| 505 | } |
| 506 | |
| 507 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X36, batch_lt_36) { |
| 508 | TEST_REQUIRES_ARM_NEON_FMA; |
| 509 | for (size_t batch_size = 1; batch_size < 36; batch_size++) { |
| 510 | VUnOpMicrokernelTester() |
| 511 | .batch_size(batch_size) |
| 512 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 513 | } |
| 514 | } |
| 515 | |
| 516 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X36, batch_gt_36) { |
| 517 | TEST_REQUIRES_ARM_NEON_FMA; |
| 518 | for (size_t batch_size = 37; batch_size < 72; batch_size++) { |
| 519 | VUnOpMicrokernelTester() |
| 520 | .batch_size(batch_size) |
| 521 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 522 | } |
| 523 | } |
| 524 | |
| 525 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X36, inplace) { |
| 526 | TEST_REQUIRES_ARM_NEON_FMA; |
| 527 | for (size_t batch_size = 1; batch_size <= 180; batch_size += 35) { |
| 528 | VUnOpMicrokernelTester() |
| 529 | .batch_size(batch_size) |
| 530 | .inplace(true) |
| 531 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 532 | } |
| 533 | } |
| 534 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 535 | |
| 536 | |
| 537 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 538 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X40, batch_eq_40) { |
| 539 | TEST_REQUIRES_ARM_NEON_FMA; |
| 540 | VUnOpMicrokernelTester() |
| 541 | .batch_size(40) |
| 542 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 543 | } |
| 544 | |
| 545 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X40, batch_div_40) { |
| 546 | TEST_REQUIRES_ARM_NEON_FMA; |
| 547 | for (size_t batch_size = 80; batch_size < 400; batch_size += 40) { |
| 548 | VUnOpMicrokernelTester() |
| 549 | .batch_size(batch_size) |
| 550 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 551 | } |
| 552 | } |
| 553 | |
| 554 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X40, batch_lt_40) { |
| 555 | TEST_REQUIRES_ARM_NEON_FMA; |
| 556 | for (size_t batch_size = 1; batch_size < 40; batch_size++) { |
| 557 | VUnOpMicrokernelTester() |
| 558 | .batch_size(batch_size) |
| 559 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 560 | } |
| 561 | } |
| 562 | |
| 563 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X40, batch_gt_40) { |
| 564 | TEST_REQUIRES_ARM_NEON_FMA; |
| 565 | for (size_t batch_size = 41; batch_size < 80; batch_size++) { |
| 566 | VUnOpMicrokernelTester() |
| 567 | .batch_size(batch_size) |
| 568 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 569 | } |
| 570 | } |
| 571 | |
| 572 | TEST(F32_VSQRT__NEONFMA_NR1RSQRTS1FMA1ADJ_X40, inplace) { |
| 573 | TEST_REQUIRES_ARM_NEON_FMA; |
| 574 | for (size_t batch_size = 1; batch_size <= 200; batch_size += 39) { |
| 575 | VUnOpMicrokernelTester() |
| 576 | .batch_size(batch_size) |
| 577 | .inplace(true) |
| 578 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 579 | } |
| 580 | } |
| 581 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 582 | |
| 583 | |
| 584 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 585 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X4, batch_eq_4) { |
| 586 | TEST_REQUIRES_ARM_NEON_FMA; |
| 587 | VUnOpMicrokernelTester() |
| 588 | .batch_size(4) |
| 589 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 590 | } |
| 591 | |
| 592 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X4, batch_div_4) { |
| 593 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr2fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 598 | } |
| 599 | } |
| 600 | |
| 601 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X4, batch_lt_4) { |
| 602 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr2fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 607 | } |
| 608 | } |
| 609 | |
| 610 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X4, batch_gt_4) { |
| 611 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr2fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 616 | } |
| 617 | } |
| 618 | |
| 619 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X4, inplace) { |
| 620 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr2fma1adj_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 626 | } |
| 627 | } |
| 628 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 629 | |
| 630 | |
| 631 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 632 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X8, batch_eq_8) { |
| 633 | TEST_REQUIRES_ARM_NEON_FMA; |
| 634 | VUnOpMicrokernelTester() |
| 635 | .batch_size(8) |
| 636 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 637 | } |
| 638 | |
| 639 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X8, batch_div_8) { |
| 640 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr2fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 645 | } |
| 646 | } |
| 647 | |
| 648 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X8, batch_lt_8) { |
| 649 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr2fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 654 | } |
| 655 | } |
| 656 | |
| 657 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X8, batch_gt_8) { |
| 658 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr2fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X8, inplace) { |
| 667 | TEST_REQUIRES_ARM_NEON_FMA; |
| 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_vsqrt_ukernel__neonfma_nr2fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 673 | } |
| 674 | } |
| 675 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 676 | |
| 677 | |
| 678 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 679 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X12, batch_eq_12) { |
| 680 | TEST_REQUIRES_ARM_NEON_FMA; |
| 681 | VUnOpMicrokernelTester() |
| 682 | .batch_size(12) |
| 683 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 684 | } |
| 685 | |
| 686 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X12, batch_div_12) { |
| 687 | TEST_REQUIRES_ARM_NEON_FMA; |
| 688 | for (size_t batch_size = 24; batch_size < 120; batch_size += 12) { |
| 689 | VUnOpMicrokernelTester() |
| 690 | .batch_size(batch_size) |
| 691 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 692 | } |
| 693 | } |
| 694 | |
| 695 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X12, batch_lt_12) { |
| 696 | TEST_REQUIRES_ARM_NEON_FMA; |
| 697 | for (size_t batch_size = 1; batch_size < 12; batch_size++) { |
| 698 | VUnOpMicrokernelTester() |
| 699 | .batch_size(batch_size) |
| 700 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 701 | } |
| 702 | } |
| 703 | |
| 704 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X12, batch_gt_12) { |
| 705 | TEST_REQUIRES_ARM_NEON_FMA; |
| 706 | for (size_t batch_size = 13; batch_size < 24; batch_size++) { |
| 707 | VUnOpMicrokernelTester() |
| 708 | .batch_size(batch_size) |
| 709 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 710 | } |
| 711 | } |
| 712 | |
| 713 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X12, inplace) { |
| 714 | TEST_REQUIRES_ARM_NEON_FMA; |
| 715 | for (size_t batch_size = 1; batch_size <= 60; batch_size += 11) { |
| 716 | VUnOpMicrokernelTester() |
| 717 | .batch_size(batch_size) |
| 718 | .inplace(true) |
| 719 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 720 | } |
| 721 | } |
| 722 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 723 | |
| 724 | |
| 725 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 726 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X16, batch_eq_16) { |
| 727 | TEST_REQUIRES_ARM_NEON_FMA; |
| 728 | VUnOpMicrokernelTester() |
| 729 | .batch_size(16) |
| 730 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 731 | } |
| 732 | |
| 733 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X16, batch_div_16) { |
| 734 | TEST_REQUIRES_ARM_NEON_FMA; |
| 735 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 736 | VUnOpMicrokernelTester() |
| 737 | .batch_size(batch_size) |
| 738 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 739 | } |
| 740 | } |
| 741 | |
| 742 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X16, batch_lt_16) { |
| 743 | TEST_REQUIRES_ARM_NEON_FMA; |
| 744 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 745 | VUnOpMicrokernelTester() |
| 746 | .batch_size(batch_size) |
| 747 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 748 | } |
| 749 | } |
| 750 | |
| 751 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X16, batch_gt_16) { |
| 752 | TEST_REQUIRES_ARM_NEON_FMA; |
| 753 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 754 | VUnOpMicrokernelTester() |
| 755 | .batch_size(batch_size) |
| 756 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 757 | } |
| 758 | } |
| 759 | |
| 760 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X16, inplace) { |
| 761 | TEST_REQUIRES_ARM_NEON_FMA; |
| 762 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 763 | VUnOpMicrokernelTester() |
| 764 | .batch_size(batch_size) |
| 765 | .inplace(true) |
| 766 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 767 | } |
| 768 | } |
| 769 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 770 | |
| 771 | |
| 772 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 773 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X20, batch_eq_20) { |
| 774 | TEST_REQUIRES_ARM_NEON_FMA; |
| 775 | VUnOpMicrokernelTester() |
| 776 | .batch_size(20) |
| 777 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 778 | } |
| 779 | |
| 780 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X20, batch_div_20) { |
| 781 | TEST_REQUIRES_ARM_NEON_FMA; |
| 782 | for (size_t batch_size = 40; batch_size < 200; batch_size += 20) { |
| 783 | VUnOpMicrokernelTester() |
| 784 | .batch_size(batch_size) |
| 785 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 786 | } |
| 787 | } |
| 788 | |
| 789 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X20, batch_lt_20) { |
| 790 | TEST_REQUIRES_ARM_NEON_FMA; |
| 791 | for (size_t batch_size = 1; batch_size < 20; batch_size++) { |
| 792 | VUnOpMicrokernelTester() |
| 793 | .batch_size(batch_size) |
| 794 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 795 | } |
| 796 | } |
| 797 | |
| 798 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X20, batch_gt_20) { |
| 799 | TEST_REQUIRES_ARM_NEON_FMA; |
| 800 | for (size_t batch_size = 21; batch_size < 40; batch_size++) { |
| 801 | VUnOpMicrokernelTester() |
| 802 | .batch_size(batch_size) |
| 803 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 804 | } |
| 805 | } |
| 806 | |
| 807 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X20, inplace) { |
| 808 | TEST_REQUIRES_ARM_NEON_FMA; |
| 809 | for (size_t batch_size = 1; batch_size <= 100; batch_size += 19) { |
| 810 | VUnOpMicrokernelTester() |
| 811 | .batch_size(batch_size) |
| 812 | .inplace(true) |
| 813 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 814 | } |
| 815 | } |
| 816 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 817 | |
| 818 | |
| 819 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 820 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X24, batch_eq_24) { |
| 821 | TEST_REQUIRES_ARM_NEON_FMA; |
| 822 | VUnOpMicrokernelTester() |
| 823 | .batch_size(24) |
| 824 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 825 | } |
| 826 | |
| 827 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X24, batch_div_24) { |
| 828 | TEST_REQUIRES_ARM_NEON_FMA; |
| 829 | for (size_t batch_size = 48; batch_size < 240; batch_size += 24) { |
| 830 | VUnOpMicrokernelTester() |
| 831 | .batch_size(batch_size) |
| 832 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 833 | } |
| 834 | } |
| 835 | |
| 836 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X24, batch_lt_24) { |
| 837 | TEST_REQUIRES_ARM_NEON_FMA; |
| 838 | for (size_t batch_size = 1; batch_size < 24; batch_size++) { |
| 839 | VUnOpMicrokernelTester() |
| 840 | .batch_size(batch_size) |
| 841 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 842 | } |
| 843 | } |
| 844 | |
| 845 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X24, batch_gt_24) { |
| 846 | TEST_REQUIRES_ARM_NEON_FMA; |
| 847 | for (size_t batch_size = 25; batch_size < 48; batch_size++) { |
| 848 | VUnOpMicrokernelTester() |
| 849 | .batch_size(batch_size) |
| 850 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 851 | } |
| 852 | } |
| 853 | |
| 854 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X24, inplace) { |
| 855 | TEST_REQUIRES_ARM_NEON_FMA; |
| 856 | for (size_t batch_size = 1; batch_size <= 120; batch_size += 23) { |
| 857 | VUnOpMicrokernelTester() |
| 858 | .batch_size(batch_size) |
| 859 | .inplace(true) |
| 860 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 861 | } |
| 862 | } |
| 863 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 864 | |
| 865 | |
| 866 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 867 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X28, batch_eq_28) { |
| 868 | TEST_REQUIRES_ARM_NEON_FMA; |
| 869 | VUnOpMicrokernelTester() |
| 870 | .batch_size(28) |
| 871 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 872 | } |
| 873 | |
| 874 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X28, batch_div_28) { |
| 875 | TEST_REQUIRES_ARM_NEON_FMA; |
| 876 | for (size_t batch_size = 56; batch_size < 280; batch_size += 28) { |
| 877 | VUnOpMicrokernelTester() |
| 878 | .batch_size(batch_size) |
| 879 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 880 | } |
| 881 | } |
| 882 | |
| 883 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X28, batch_lt_28) { |
| 884 | TEST_REQUIRES_ARM_NEON_FMA; |
| 885 | for (size_t batch_size = 1; batch_size < 28; batch_size++) { |
| 886 | VUnOpMicrokernelTester() |
| 887 | .batch_size(batch_size) |
| 888 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 889 | } |
| 890 | } |
| 891 | |
| 892 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X28, batch_gt_28) { |
| 893 | TEST_REQUIRES_ARM_NEON_FMA; |
| 894 | for (size_t batch_size = 29; batch_size < 56; batch_size++) { |
| 895 | VUnOpMicrokernelTester() |
| 896 | .batch_size(batch_size) |
| 897 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 898 | } |
| 899 | } |
| 900 | |
| 901 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X28, inplace) { |
| 902 | TEST_REQUIRES_ARM_NEON_FMA; |
| 903 | for (size_t batch_size = 1; batch_size <= 140; batch_size += 27) { |
| 904 | VUnOpMicrokernelTester() |
| 905 | .batch_size(batch_size) |
| 906 | .inplace(true) |
| 907 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 908 | } |
| 909 | } |
| 910 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 911 | |
| 912 | |
| 913 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 914 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X32, batch_eq_32) { |
| 915 | TEST_REQUIRES_ARM_NEON_FMA; |
| 916 | VUnOpMicrokernelTester() |
| 917 | .batch_size(32) |
| 918 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 919 | } |
| 920 | |
| 921 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X32, batch_div_32) { |
| 922 | TEST_REQUIRES_ARM_NEON_FMA; |
| 923 | for (size_t batch_size = 64; batch_size < 320; batch_size += 32) { |
| 924 | VUnOpMicrokernelTester() |
| 925 | .batch_size(batch_size) |
| 926 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 927 | } |
| 928 | } |
| 929 | |
| 930 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X32, batch_lt_32) { |
| 931 | TEST_REQUIRES_ARM_NEON_FMA; |
| 932 | for (size_t batch_size = 1; batch_size < 32; batch_size++) { |
| 933 | VUnOpMicrokernelTester() |
| 934 | .batch_size(batch_size) |
| 935 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 936 | } |
| 937 | } |
| 938 | |
| 939 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X32, batch_gt_32) { |
| 940 | TEST_REQUIRES_ARM_NEON_FMA; |
| 941 | for (size_t batch_size = 33; batch_size < 64; batch_size++) { |
| 942 | VUnOpMicrokernelTester() |
| 943 | .batch_size(batch_size) |
| 944 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 945 | } |
| 946 | } |
| 947 | |
| 948 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X32, inplace) { |
| 949 | TEST_REQUIRES_ARM_NEON_FMA; |
| 950 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
| 951 | VUnOpMicrokernelTester() |
| 952 | .batch_size(batch_size) |
| 953 | .inplace(true) |
| 954 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 955 | } |
| 956 | } |
| 957 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 958 | |
| 959 | |
| 960 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 961 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X36, batch_eq_36) { |
| 962 | TEST_REQUIRES_ARM_NEON_FMA; |
| 963 | VUnOpMicrokernelTester() |
| 964 | .batch_size(36) |
| 965 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 966 | } |
| 967 | |
| 968 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X36, batch_div_36) { |
| 969 | TEST_REQUIRES_ARM_NEON_FMA; |
| 970 | for (size_t batch_size = 72; batch_size < 360; batch_size += 36) { |
| 971 | VUnOpMicrokernelTester() |
| 972 | .batch_size(batch_size) |
| 973 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 974 | } |
| 975 | } |
| 976 | |
| 977 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X36, batch_lt_36) { |
| 978 | TEST_REQUIRES_ARM_NEON_FMA; |
| 979 | for (size_t batch_size = 1; batch_size < 36; batch_size++) { |
| 980 | VUnOpMicrokernelTester() |
| 981 | .batch_size(batch_size) |
| 982 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 983 | } |
| 984 | } |
| 985 | |
| 986 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X36, batch_gt_36) { |
| 987 | TEST_REQUIRES_ARM_NEON_FMA; |
| 988 | for (size_t batch_size = 37; batch_size < 72; batch_size++) { |
| 989 | VUnOpMicrokernelTester() |
| 990 | .batch_size(batch_size) |
| 991 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 992 | } |
| 993 | } |
| 994 | |
| 995 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X36, inplace) { |
| 996 | TEST_REQUIRES_ARM_NEON_FMA; |
| 997 | for (size_t batch_size = 1; batch_size <= 180; batch_size += 35) { |
| 998 | VUnOpMicrokernelTester() |
| 999 | .batch_size(batch_size) |
| 1000 | .inplace(true) |
| 1001 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1002 | } |
| 1003 | } |
| 1004 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 1005 | |
| 1006 | |
| 1007 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 1008 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X40, batch_eq_40) { |
| 1009 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1010 | VUnOpMicrokernelTester() |
| 1011 | .batch_size(40) |
| 1012 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1013 | } |
| 1014 | |
| 1015 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X40, batch_div_40) { |
| 1016 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1017 | for (size_t batch_size = 80; batch_size < 400; batch_size += 40) { |
| 1018 | VUnOpMicrokernelTester() |
| 1019 | .batch_size(batch_size) |
| 1020 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1021 | } |
| 1022 | } |
| 1023 | |
| 1024 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X40, batch_lt_40) { |
| 1025 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1026 | for (size_t batch_size = 1; batch_size < 40; batch_size++) { |
| 1027 | VUnOpMicrokernelTester() |
| 1028 | .batch_size(batch_size) |
| 1029 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1030 | } |
| 1031 | } |
| 1032 | |
| 1033 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X40, batch_gt_40) { |
| 1034 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1035 | for (size_t batch_size = 41; batch_size < 80; batch_size++) { |
| 1036 | VUnOpMicrokernelTester() |
| 1037 | .batch_size(batch_size) |
| 1038 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1039 | } |
| 1040 | } |
| 1041 | |
| 1042 | TEST(F32_VSQRT__NEONFMA_NR2FMA1ADJ_X40, inplace) { |
| 1043 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1044 | for (size_t batch_size = 1; batch_size <= 200; batch_size += 39) { |
| 1045 | VUnOpMicrokernelTester() |
| 1046 | .batch_size(batch_size) |
| 1047 | .inplace(true) |
| 1048 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1049 | } |
| 1050 | } |
| 1051 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 1052 | |
| 1053 | |
| 1054 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1055 | TEST(F32_VSQRT__SSE_SQRT_X4, batch_eq_4) { |
| 1056 | TEST_REQUIRES_X86_SSE; |
| 1057 | VUnOpMicrokernelTester() |
| 1058 | .batch_size(4) |
| 1059 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1060 | } |
| 1061 | |
| 1062 | TEST(F32_VSQRT__SSE_SQRT_X4, batch_div_4) { |
| 1063 | TEST_REQUIRES_X86_SSE; |
| 1064 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 1065 | VUnOpMicrokernelTester() |
| 1066 | .batch_size(batch_size) |
| 1067 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1068 | } |
| 1069 | } |
| 1070 | |
| 1071 | TEST(F32_VSQRT__SSE_SQRT_X4, batch_lt_4) { |
| 1072 | TEST_REQUIRES_X86_SSE; |
| 1073 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 1074 | VUnOpMicrokernelTester() |
| 1075 | .batch_size(batch_size) |
| 1076 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1077 | } |
| 1078 | } |
| 1079 | |
| 1080 | TEST(F32_VSQRT__SSE_SQRT_X4, batch_gt_4) { |
| 1081 | TEST_REQUIRES_X86_SSE; |
| 1082 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 1083 | VUnOpMicrokernelTester() |
| 1084 | .batch_size(batch_size) |
| 1085 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1086 | } |
| 1087 | } |
| 1088 | |
| 1089 | TEST(F32_VSQRT__SSE_SQRT_X4, inplace) { |
| 1090 | TEST_REQUIRES_X86_SSE; |
| 1091 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 1092 | VUnOpMicrokernelTester() |
| 1093 | .batch_size(batch_size) |
| 1094 | .inplace(true) |
| 1095 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1096 | } |
| 1097 | } |
| 1098 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1099 | |
| 1100 | |
| 1101 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1102 | TEST(F32_VSQRT__SSE_SQRT_X8, batch_eq_8) { |
| 1103 | TEST_REQUIRES_X86_SSE; |
| 1104 | VUnOpMicrokernelTester() |
| 1105 | .batch_size(8) |
| 1106 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1107 | } |
| 1108 | |
| 1109 | TEST(F32_VSQRT__SSE_SQRT_X8, batch_div_8) { |
| 1110 | TEST_REQUIRES_X86_SSE; |
| 1111 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 1112 | VUnOpMicrokernelTester() |
| 1113 | .batch_size(batch_size) |
| 1114 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1115 | } |
| 1116 | } |
| 1117 | |
| 1118 | TEST(F32_VSQRT__SSE_SQRT_X8, batch_lt_8) { |
| 1119 | TEST_REQUIRES_X86_SSE; |
| 1120 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 1121 | VUnOpMicrokernelTester() |
| 1122 | .batch_size(batch_size) |
| 1123 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1124 | } |
| 1125 | } |
| 1126 | |
| 1127 | TEST(F32_VSQRT__SSE_SQRT_X8, batch_gt_8) { |
| 1128 | TEST_REQUIRES_X86_SSE; |
| 1129 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 1130 | VUnOpMicrokernelTester() |
| 1131 | .batch_size(batch_size) |
| 1132 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1133 | } |
| 1134 | } |
| 1135 | |
| 1136 | TEST(F32_VSQRT__SSE_SQRT_X8, inplace) { |
| 1137 | TEST_REQUIRES_X86_SSE; |
| 1138 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 1139 | VUnOpMicrokernelTester() |
| 1140 | .batch_size(batch_size) |
| 1141 | .inplace(true) |
| 1142 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__sse_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1143 | } |
| 1144 | } |
| 1145 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1146 | |
| 1147 | |
| 1148 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1149 | TEST(F32_VSQRT__AVX_SQRT_X8, batch_eq_8) { |
| 1150 | TEST_REQUIRES_X86_AVX; |
| 1151 | VUnOpMicrokernelTester() |
| 1152 | .batch_size(8) |
| 1153 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1154 | } |
| 1155 | |
| 1156 | TEST(F32_VSQRT__AVX_SQRT_X8, batch_div_8) { |
| 1157 | TEST_REQUIRES_X86_AVX; |
| 1158 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 1159 | VUnOpMicrokernelTester() |
| 1160 | .batch_size(batch_size) |
| 1161 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1162 | } |
| 1163 | } |
| 1164 | |
| 1165 | TEST(F32_VSQRT__AVX_SQRT_X8, batch_lt_8) { |
| 1166 | TEST_REQUIRES_X86_AVX; |
| 1167 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 1168 | VUnOpMicrokernelTester() |
| 1169 | .batch_size(batch_size) |
| 1170 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1171 | } |
| 1172 | } |
| 1173 | |
| 1174 | TEST(F32_VSQRT__AVX_SQRT_X8, batch_gt_8) { |
| 1175 | TEST_REQUIRES_X86_AVX; |
| 1176 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 1177 | VUnOpMicrokernelTester() |
| 1178 | .batch_size(batch_size) |
| 1179 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1180 | } |
| 1181 | } |
| 1182 | |
| 1183 | TEST(F32_VSQRT__AVX_SQRT_X8, inplace) { |
| 1184 | TEST_REQUIRES_X86_AVX; |
| 1185 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 1186 | VUnOpMicrokernelTester() |
| 1187 | .batch_size(batch_size) |
| 1188 | .inplace(true) |
| 1189 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1190 | } |
| 1191 | } |
| 1192 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1193 | |
| 1194 | |
| 1195 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1196 | TEST(F32_VSQRT__AVX_SQRT_X16, batch_eq_16) { |
| 1197 | TEST_REQUIRES_X86_AVX; |
| 1198 | VUnOpMicrokernelTester() |
| 1199 | .batch_size(16) |
| 1200 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1201 | } |
| 1202 | |
| 1203 | TEST(F32_VSQRT__AVX_SQRT_X16, batch_div_16) { |
| 1204 | TEST_REQUIRES_X86_AVX; |
| 1205 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 1206 | VUnOpMicrokernelTester() |
| 1207 | .batch_size(batch_size) |
| 1208 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1209 | } |
| 1210 | } |
| 1211 | |
| 1212 | TEST(F32_VSQRT__AVX_SQRT_X16, batch_lt_16) { |
| 1213 | TEST_REQUIRES_X86_AVX; |
| 1214 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 1215 | VUnOpMicrokernelTester() |
| 1216 | .batch_size(batch_size) |
| 1217 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1218 | } |
| 1219 | } |
| 1220 | |
| 1221 | TEST(F32_VSQRT__AVX_SQRT_X16, batch_gt_16) { |
| 1222 | TEST_REQUIRES_X86_AVX; |
| 1223 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 1224 | VUnOpMicrokernelTester() |
| 1225 | .batch_size(batch_size) |
| 1226 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1227 | } |
| 1228 | } |
| 1229 | |
| 1230 | TEST(F32_VSQRT__AVX_SQRT_X16, inplace) { |
| 1231 | TEST_REQUIRES_X86_AVX; |
| 1232 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 1233 | VUnOpMicrokernelTester() |
| 1234 | .batch_size(batch_size) |
| 1235 | .inplace(true) |
| 1236 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx_sqrt_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1237 | } |
| 1238 | } |
| 1239 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1240 | |
| 1241 | |
| 1242 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1243 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X8, batch_eq_8) { |
| 1244 | TEST_REQUIRES_X86_FMA3; |
| 1245 | VUnOpMicrokernelTester() |
| 1246 | .batch_size(8) |
| 1247 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1248 | } |
| 1249 | |
| 1250 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X8, batch_div_8) { |
| 1251 | TEST_REQUIRES_X86_FMA3; |
| 1252 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 1253 | VUnOpMicrokernelTester() |
| 1254 | .batch_size(batch_size) |
| 1255 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1256 | } |
| 1257 | } |
| 1258 | |
| 1259 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X8, batch_lt_8) { |
| 1260 | TEST_REQUIRES_X86_FMA3; |
| 1261 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 1262 | VUnOpMicrokernelTester() |
| 1263 | .batch_size(batch_size) |
| 1264 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1265 | } |
| 1266 | } |
| 1267 | |
| 1268 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X8, batch_gt_8) { |
| 1269 | TEST_REQUIRES_X86_FMA3; |
| 1270 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 1271 | VUnOpMicrokernelTester() |
| 1272 | .batch_size(batch_size) |
| 1273 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1274 | } |
| 1275 | } |
| 1276 | |
| 1277 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X8, inplace) { |
| 1278 | TEST_REQUIRES_X86_FMA3; |
| 1279 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 1280 | VUnOpMicrokernelTester() |
| 1281 | .batch_size(batch_size) |
| 1282 | .inplace(true) |
| 1283 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1284 | } |
| 1285 | } |
| 1286 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1287 | |
| 1288 | |
| 1289 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1290 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X16, batch_eq_16) { |
| 1291 | TEST_REQUIRES_X86_FMA3; |
| 1292 | VUnOpMicrokernelTester() |
| 1293 | .batch_size(16) |
| 1294 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1295 | } |
| 1296 | |
| 1297 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X16, batch_div_16) { |
| 1298 | TEST_REQUIRES_X86_FMA3; |
| 1299 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 1300 | VUnOpMicrokernelTester() |
| 1301 | .batch_size(batch_size) |
| 1302 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1303 | } |
| 1304 | } |
| 1305 | |
| 1306 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X16, batch_lt_16) { |
| 1307 | TEST_REQUIRES_X86_FMA3; |
| 1308 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 1309 | VUnOpMicrokernelTester() |
| 1310 | .batch_size(batch_size) |
| 1311 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1312 | } |
| 1313 | } |
| 1314 | |
| 1315 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X16, batch_gt_16) { |
| 1316 | TEST_REQUIRES_X86_FMA3; |
| 1317 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 1318 | VUnOpMicrokernelTester() |
| 1319 | .batch_size(batch_size) |
| 1320 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1321 | } |
| 1322 | } |
| 1323 | |
| 1324 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X16, inplace) { |
| 1325 | TEST_REQUIRES_X86_FMA3; |
| 1326 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 1327 | VUnOpMicrokernelTester() |
| 1328 | .batch_size(batch_size) |
| 1329 | .inplace(true) |
| 1330 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1331 | } |
| 1332 | } |
| 1333 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1334 | |
| 1335 | |
| 1336 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1337 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X24, batch_eq_24) { |
| 1338 | TEST_REQUIRES_X86_FMA3; |
| 1339 | VUnOpMicrokernelTester() |
| 1340 | .batch_size(24) |
| 1341 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1342 | } |
| 1343 | |
| 1344 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X24, batch_div_24) { |
| 1345 | TEST_REQUIRES_X86_FMA3; |
| 1346 | for (size_t batch_size = 48; batch_size < 240; batch_size += 24) { |
| 1347 | VUnOpMicrokernelTester() |
| 1348 | .batch_size(batch_size) |
| 1349 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1350 | } |
| 1351 | } |
| 1352 | |
| 1353 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X24, batch_lt_24) { |
| 1354 | TEST_REQUIRES_X86_FMA3; |
| 1355 | for (size_t batch_size = 1; batch_size < 24; batch_size++) { |
| 1356 | VUnOpMicrokernelTester() |
| 1357 | .batch_size(batch_size) |
| 1358 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1359 | } |
| 1360 | } |
| 1361 | |
| 1362 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X24, batch_gt_24) { |
| 1363 | TEST_REQUIRES_X86_FMA3; |
| 1364 | for (size_t batch_size = 25; batch_size < 48; batch_size++) { |
| 1365 | VUnOpMicrokernelTester() |
| 1366 | .batch_size(batch_size) |
| 1367 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1368 | } |
| 1369 | } |
| 1370 | |
| 1371 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X24, inplace) { |
| 1372 | TEST_REQUIRES_X86_FMA3; |
| 1373 | for (size_t batch_size = 1; batch_size <= 120; batch_size += 23) { |
| 1374 | VUnOpMicrokernelTester() |
| 1375 | .batch_size(batch_size) |
| 1376 | .inplace(true) |
| 1377 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1378 | } |
| 1379 | } |
| 1380 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1381 | |
| 1382 | |
| 1383 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1384 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X32, batch_eq_32) { |
| 1385 | TEST_REQUIRES_X86_FMA3; |
| 1386 | VUnOpMicrokernelTester() |
| 1387 | .batch_size(32) |
| 1388 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1389 | } |
| 1390 | |
| 1391 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X32, batch_div_32) { |
| 1392 | TEST_REQUIRES_X86_FMA3; |
| 1393 | for (size_t batch_size = 64; batch_size < 320; batch_size += 32) { |
| 1394 | VUnOpMicrokernelTester() |
| 1395 | .batch_size(batch_size) |
| 1396 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1397 | } |
| 1398 | } |
| 1399 | |
| 1400 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X32, batch_lt_32) { |
| 1401 | TEST_REQUIRES_X86_FMA3; |
| 1402 | for (size_t batch_size = 1; batch_size < 32; batch_size++) { |
| 1403 | VUnOpMicrokernelTester() |
| 1404 | .batch_size(batch_size) |
| 1405 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1406 | } |
| 1407 | } |
| 1408 | |
| 1409 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X32, batch_gt_32) { |
| 1410 | TEST_REQUIRES_X86_FMA3; |
| 1411 | for (size_t batch_size = 33; batch_size < 64; batch_size++) { |
| 1412 | VUnOpMicrokernelTester() |
| 1413 | .batch_size(batch_size) |
| 1414 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1415 | } |
| 1416 | } |
| 1417 | |
| 1418 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X32, inplace) { |
| 1419 | TEST_REQUIRES_X86_FMA3; |
| 1420 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
| 1421 | VUnOpMicrokernelTester() |
| 1422 | .batch_size(batch_size) |
| 1423 | .inplace(true) |
| 1424 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1425 | } |
| 1426 | } |
| 1427 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1428 | |
| 1429 | |
| 1430 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1431 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X40, batch_eq_40) { |
| 1432 | TEST_REQUIRES_X86_FMA3; |
| 1433 | VUnOpMicrokernelTester() |
| 1434 | .batch_size(40) |
| 1435 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1436 | } |
| 1437 | |
| 1438 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X40, batch_div_40) { |
| 1439 | TEST_REQUIRES_X86_FMA3; |
| 1440 | for (size_t batch_size = 80; batch_size < 400; batch_size += 40) { |
| 1441 | VUnOpMicrokernelTester() |
| 1442 | .batch_size(batch_size) |
| 1443 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1444 | } |
| 1445 | } |
| 1446 | |
| 1447 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X40, batch_lt_40) { |
| 1448 | TEST_REQUIRES_X86_FMA3; |
| 1449 | for (size_t batch_size = 1; batch_size < 40; batch_size++) { |
| 1450 | VUnOpMicrokernelTester() |
| 1451 | .batch_size(batch_size) |
| 1452 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1453 | } |
| 1454 | } |
| 1455 | |
| 1456 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X40, batch_gt_40) { |
| 1457 | TEST_REQUIRES_X86_FMA3; |
| 1458 | for (size_t batch_size = 41; batch_size < 80; batch_size++) { |
| 1459 | VUnOpMicrokernelTester() |
| 1460 | .batch_size(batch_size) |
| 1461 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1462 | } |
| 1463 | } |
| 1464 | |
| 1465 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X40, inplace) { |
| 1466 | TEST_REQUIRES_X86_FMA3; |
| 1467 | for (size_t batch_size = 1; batch_size <= 200; batch_size += 39) { |
| 1468 | VUnOpMicrokernelTester() |
| 1469 | .batch_size(batch_size) |
| 1470 | .inplace(true) |
| 1471 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1472 | } |
| 1473 | } |
| 1474 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1475 | |
| 1476 | |
| 1477 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1478 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X48, batch_eq_48) { |
| 1479 | TEST_REQUIRES_X86_FMA3; |
| 1480 | VUnOpMicrokernelTester() |
| 1481 | .batch_size(48) |
| 1482 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1483 | } |
| 1484 | |
| 1485 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X48, batch_div_48) { |
| 1486 | TEST_REQUIRES_X86_FMA3; |
| 1487 | for (size_t batch_size = 96; batch_size < 480; batch_size += 48) { |
| 1488 | VUnOpMicrokernelTester() |
| 1489 | .batch_size(batch_size) |
| 1490 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1491 | } |
| 1492 | } |
| 1493 | |
| 1494 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X48, batch_lt_48) { |
| 1495 | TEST_REQUIRES_X86_FMA3; |
| 1496 | for (size_t batch_size = 1; batch_size < 48; batch_size++) { |
| 1497 | VUnOpMicrokernelTester() |
| 1498 | .batch_size(batch_size) |
| 1499 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1500 | } |
| 1501 | } |
| 1502 | |
| 1503 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X48, batch_gt_48) { |
| 1504 | TEST_REQUIRES_X86_FMA3; |
| 1505 | for (size_t batch_size = 49; batch_size < 96; batch_size++) { |
| 1506 | VUnOpMicrokernelTester() |
| 1507 | .batch_size(batch_size) |
| 1508 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1509 | } |
| 1510 | } |
| 1511 | |
| 1512 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X48, inplace) { |
| 1513 | TEST_REQUIRES_X86_FMA3; |
| 1514 | for (size_t batch_size = 1; batch_size <= 240; batch_size += 47) { |
| 1515 | VUnOpMicrokernelTester() |
| 1516 | .batch_size(batch_size) |
| 1517 | .inplace(true) |
| 1518 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1519 | } |
| 1520 | } |
| 1521 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1522 | |
| 1523 | |
| 1524 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1525 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X56, batch_eq_56) { |
| 1526 | TEST_REQUIRES_X86_FMA3; |
| 1527 | VUnOpMicrokernelTester() |
| 1528 | .batch_size(56) |
| 1529 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1530 | } |
| 1531 | |
| 1532 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X56, batch_div_56) { |
| 1533 | TEST_REQUIRES_X86_FMA3; |
| 1534 | for (size_t batch_size = 112; batch_size < 560; batch_size += 56) { |
| 1535 | VUnOpMicrokernelTester() |
| 1536 | .batch_size(batch_size) |
| 1537 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1538 | } |
| 1539 | } |
| 1540 | |
| 1541 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X56, batch_lt_56) { |
| 1542 | TEST_REQUIRES_X86_FMA3; |
| 1543 | for (size_t batch_size = 1; batch_size < 56; batch_size++) { |
| 1544 | VUnOpMicrokernelTester() |
| 1545 | .batch_size(batch_size) |
| 1546 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1547 | } |
| 1548 | } |
| 1549 | |
| 1550 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X56, batch_gt_56) { |
| 1551 | TEST_REQUIRES_X86_FMA3; |
| 1552 | for (size_t batch_size = 57; batch_size < 112; batch_size++) { |
| 1553 | VUnOpMicrokernelTester() |
| 1554 | .batch_size(batch_size) |
| 1555 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1556 | } |
| 1557 | } |
| 1558 | |
| 1559 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X56, inplace) { |
| 1560 | TEST_REQUIRES_X86_FMA3; |
| 1561 | for (size_t batch_size = 1; batch_size <= 280; batch_size += 55) { |
| 1562 | VUnOpMicrokernelTester() |
| 1563 | .batch_size(batch_size) |
| 1564 | .inplace(true) |
| 1565 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1566 | } |
| 1567 | } |
| 1568 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1569 | |
| 1570 | |
| 1571 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1572 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X64, batch_eq_64) { |
| 1573 | TEST_REQUIRES_X86_FMA3; |
| 1574 | VUnOpMicrokernelTester() |
| 1575 | .batch_size(64) |
| 1576 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1577 | } |
| 1578 | |
| 1579 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X64, batch_div_64) { |
| 1580 | TEST_REQUIRES_X86_FMA3; |
| 1581 | for (size_t batch_size = 128; batch_size < 640; batch_size += 64) { |
| 1582 | VUnOpMicrokernelTester() |
| 1583 | .batch_size(batch_size) |
| 1584 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1585 | } |
| 1586 | } |
| 1587 | |
| 1588 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X64, batch_lt_64) { |
| 1589 | TEST_REQUIRES_X86_FMA3; |
| 1590 | for (size_t batch_size = 1; batch_size < 64; batch_size++) { |
| 1591 | VUnOpMicrokernelTester() |
| 1592 | .batch_size(batch_size) |
| 1593 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1594 | } |
| 1595 | } |
| 1596 | |
| 1597 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X64, batch_gt_64) { |
| 1598 | TEST_REQUIRES_X86_FMA3; |
| 1599 | for (size_t batch_size = 65; batch_size < 128; batch_size++) { |
| 1600 | VUnOpMicrokernelTester() |
| 1601 | .batch_size(batch_size) |
| 1602 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1603 | } |
| 1604 | } |
| 1605 | |
| 1606 | TEST(F32_VSQRT__FMA3_NR1FMA1ADJ_X64, inplace) { |
| 1607 | TEST_REQUIRES_X86_FMA3; |
| 1608 | for (size_t batch_size = 1; batch_size <= 320; batch_size += 63) { |
| 1609 | VUnOpMicrokernelTester() |
| 1610 | .batch_size(batch_size) |
| 1611 | .inplace(true) |
| 1612 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1613 | } |
| 1614 | } |
| 1615 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1616 | |
| 1617 | |
| 1618 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1619 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X16, batch_eq_16) { |
| 1620 | TEST_REQUIRES_X86_AVX512F; |
| 1621 | VUnOpMicrokernelTester() |
| 1622 | .batch_size(16) |
| 1623 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1624 | } |
| 1625 | |
| 1626 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X16, batch_div_16) { |
| 1627 | TEST_REQUIRES_X86_AVX512F; |
| 1628 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 1629 | VUnOpMicrokernelTester() |
| 1630 | .batch_size(batch_size) |
| 1631 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1632 | } |
| 1633 | } |
| 1634 | |
| 1635 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X16, batch_lt_16) { |
| 1636 | TEST_REQUIRES_X86_AVX512F; |
| 1637 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 1638 | VUnOpMicrokernelTester() |
| 1639 | .batch_size(batch_size) |
| 1640 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1641 | } |
| 1642 | } |
| 1643 | |
| 1644 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X16, batch_gt_16) { |
| 1645 | TEST_REQUIRES_X86_AVX512F; |
| 1646 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 1647 | VUnOpMicrokernelTester() |
| 1648 | .batch_size(batch_size) |
| 1649 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1650 | } |
| 1651 | } |
| 1652 | |
| 1653 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X16, inplace) { |
| 1654 | TEST_REQUIRES_X86_AVX512F; |
| 1655 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 1656 | VUnOpMicrokernelTester() |
| 1657 | .batch_size(batch_size) |
| 1658 | .inplace(true) |
| 1659 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1660 | } |
| 1661 | } |
| 1662 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1663 | |
| 1664 | |
| 1665 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1666 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X32, batch_eq_32) { |
| 1667 | TEST_REQUIRES_X86_AVX512F; |
| 1668 | VUnOpMicrokernelTester() |
| 1669 | .batch_size(32) |
| 1670 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1671 | } |
| 1672 | |
| 1673 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X32, batch_div_32) { |
| 1674 | TEST_REQUIRES_X86_AVX512F; |
| 1675 | for (size_t batch_size = 64; batch_size < 320; batch_size += 32) { |
| 1676 | VUnOpMicrokernelTester() |
| 1677 | .batch_size(batch_size) |
| 1678 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1679 | } |
| 1680 | } |
| 1681 | |
| 1682 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X32, batch_lt_32) { |
| 1683 | TEST_REQUIRES_X86_AVX512F; |
| 1684 | for (size_t batch_size = 1; batch_size < 32; batch_size++) { |
| 1685 | VUnOpMicrokernelTester() |
| 1686 | .batch_size(batch_size) |
| 1687 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1688 | } |
| 1689 | } |
| 1690 | |
| 1691 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X32, batch_gt_32) { |
| 1692 | TEST_REQUIRES_X86_AVX512F; |
| 1693 | for (size_t batch_size = 33; batch_size < 64; batch_size++) { |
| 1694 | VUnOpMicrokernelTester() |
| 1695 | .batch_size(batch_size) |
| 1696 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1697 | } |
| 1698 | } |
| 1699 | |
| 1700 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X32, inplace) { |
| 1701 | TEST_REQUIRES_X86_AVX512F; |
| 1702 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
| 1703 | VUnOpMicrokernelTester() |
| 1704 | .batch_size(batch_size) |
| 1705 | .inplace(true) |
| 1706 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1707 | } |
| 1708 | } |
| 1709 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1710 | |
| 1711 | |
| 1712 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1713 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X48, batch_eq_48) { |
| 1714 | TEST_REQUIRES_X86_AVX512F; |
| 1715 | VUnOpMicrokernelTester() |
| 1716 | .batch_size(48) |
| 1717 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1718 | } |
| 1719 | |
| 1720 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X48, batch_div_48) { |
| 1721 | TEST_REQUIRES_X86_AVX512F; |
| 1722 | for (size_t batch_size = 96; batch_size < 480; batch_size += 48) { |
| 1723 | VUnOpMicrokernelTester() |
| 1724 | .batch_size(batch_size) |
| 1725 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1726 | } |
| 1727 | } |
| 1728 | |
| 1729 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X48, batch_lt_48) { |
| 1730 | TEST_REQUIRES_X86_AVX512F; |
| 1731 | for (size_t batch_size = 1; batch_size < 48; batch_size++) { |
| 1732 | VUnOpMicrokernelTester() |
| 1733 | .batch_size(batch_size) |
| 1734 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1735 | } |
| 1736 | } |
| 1737 | |
| 1738 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X48, batch_gt_48) { |
| 1739 | TEST_REQUIRES_X86_AVX512F; |
| 1740 | for (size_t batch_size = 49; batch_size < 96; batch_size++) { |
| 1741 | VUnOpMicrokernelTester() |
| 1742 | .batch_size(batch_size) |
| 1743 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1744 | } |
| 1745 | } |
| 1746 | |
| 1747 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X48, inplace) { |
| 1748 | TEST_REQUIRES_X86_AVX512F; |
| 1749 | for (size_t batch_size = 1; batch_size <= 240; batch_size += 47) { |
| 1750 | VUnOpMicrokernelTester() |
| 1751 | .batch_size(batch_size) |
| 1752 | .inplace(true) |
| 1753 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1754 | } |
| 1755 | } |
| 1756 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1757 | |
| 1758 | |
| 1759 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1760 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X64, batch_eq_64) { |
| 1761 | TEST_REQUIRES_X86_AVX512F; |
| 1762 | VUnOpMicrokernelTester() |
| 1763 | .batch_size(64) |
| 1764 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1765 | } |
| 1766 | |
| 1767 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X64, batch_div_64) { |
| 1768 | TEST_REQUIRES_X86_AVX512F; |
| 1769 | for (size_t batch_size = 128; batch_size < 640; batch_size += 64) { |
| 1770 | VUnOpMicrokernelTester() |
| 1771 | .batch_size(batch_size) |
| 1772 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1773 | } |
| 1774 | } |
| 1775 | |
| 1776 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X64, batch_lt_64) { |
| 1777 | TEST_REQUIRES_X86_AVX512F; |
| 1778 | for (size_t batch_size = 1; batch_size < 64; batch_size++) { |
| 1779 | VUnOpMicrokernelTester() |
| 1780 | .batch_size(batch_size) |
| 1781 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1782 | } |
| 1783 | } |
| 1784 | |
| 1785 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X64, batch_gt_64) { |
| 1786 | TEST_REQUIRES_X86_AVX512F; |
| 1787 | for (size_t batch_size = 65; batch_size < 128; batch_size++) { |
| 1788 | VUnOpMicrokernelTester() |
| 1789 | .batch_size(batch_size) |
| 1790 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1791 | } |
| 1792 | } |
| 1793 | |
| 1794 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X64, inplace) { |
| 1795 | TEST_REQUIRES_X86_AVX512F; |
| 1796 | for (size_t batch_size = 1; batch_size <= 320; batch_size += 63) { |
| 1797 | VUnOpMicrokernelTester() |
| 1798 | .batch_size(batch_size) |
| 1799 | .inplace(true) |
| 1800 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1801 | } |
| 1802 | } |
| 1803 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1804 | |
| 1805 | |
| 1806 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1807 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X80, batch_eq_80) { |
| 1808 | TEST_REQUIRES_X86_AVX512F; |
| 1809 | VUnOpMicrokernelTester() |
| 1810 | .batch_size(80) |
| 1811 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1812 | } |
| 1813 | |
| 1814 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X80, batch_div_80) { |
| 1815 | TEST_REQUIRES_X86_AVX512F; |
| 1816 | for (size_t batch_size = 160; batch_size < 800; batch_size += 80) { |
| 1817 | VUnOpMicrokernelTester() |
| 1818 | .batch_size(batch_size) |
| 1819 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1820 | } |
| 1821 | } |
| 1822 | |
| 1823 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X80, batch_lt_80) { |
| 1824 | TEST_REQUIRES_X86_AVX512F; |
| 1825 | for (size_t batch_size = 1; batch_size < 80; batch_size++) { |
| 1826 | VUnOpMicrokernelTester() |
| 1827 | .batch_size(batch_size) |
| 1828 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1829 | } |
| 1830 | } |
| 1831 | |
| 1832 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X80, batch_gt_80) { |
| 1833 | TEST_REQUIRES_X86_AVX512F; |
| 1834 | for (size_t batch_size = 81; batch_size < 160; batch_size++) { |
| 1835 | VUnOpMicrokernelTester() |
| 1836 | .batch_size(batch_size) |
| 1837 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1838 | } |
| 1839 | } |
| 1840 | |
| 1841 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X80, inplace) { |
| 1842 | TEST_REQUIRES_X86_AVX512F; |
| 1843 | for (size_t batch_size = 1; batch_size <= 400; batch_size += 79) { |
| 1844 | VUnOpMicrokernelTester() |
| 1845 | .batch_size(batch_size) |
| 1846 | .inplace(true) |
| 1847 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1848 | } |
| 1849 | } |
| 1850 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1851 | |
| 1852 | |
| 1853 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1854 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X96, batch_eq_96) { |
| 1855 | TEST_REQUIRES_X86_AVX512F; |
| 1856 | VUnOpMicrokernelTester() |
| 1857 | .batch_size(96) |
| 1858 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1859 | } |
| 1860 | |
| 1861 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X96, batch_div_96) { |
| 1862 | TEST_REQUIRES_X86_AVX512F; |
| 1863 | for (size_t batch_size = 192; batch_size < 960; batch_size += 96) { |
| 1864 | VUnOpMicrokernelTester() |
| 1865 | .batch_size(batch_size) |
| 1866 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1867 | } |
| 1868 | } |
| 1869 | |
| 1870 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X96, batch_lt_96) { |
| 1871 | TEST_REQUIRES_X86_AVX512F; |
| 1872 | for (size_t batch_size = 1; batch_size < 96; batch_size++) { |
| 1873 | VUnOpMicrokernelTester() |
| 1874 | .batch_size(batch_size) |
| 1875 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1876 | } |
| 1877 | } |
| 1878 | |
| 1879 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X96, batch_gt_96) { |
| 1880 | TEST_REQUIRES_X86_AVX512F; |
| 1881 | for (size_t batch_size = 97; batch_size < 192; batch_size++) { |
| 1882 | VUnOpMicrokernelTester() |
| 1883 | .batch_size(batch_size) |
| 1884 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1885 | } |
| 1886 | } |
| 1887 | |
| 1888 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X96, inplace) { |
| 1889 | TEST_REQUIRES_X86_AVX512F; |
| 1890 | for (size_t batch_size = 1; batch_size <= 480; batch_size += 95) { |
| 1891 | VUnOpMicrokernelTester() |
| 1892 | .batch_size(batch_size) |
| 1893 | .inplace(true) |
| 1894 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1895 | } |
| 1896 | } |
| 1897 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1898 | |
| 1899 | |
| 1900 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1901 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X112, batch_eq_112) { |
| 1902 | TEST_REQUIRES_X86_AVX512F; |
| 1903 | VUnOpMicrokernelTester() |
| 1904 | .batch_size(112) |
| 1905 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1906 | } |
| 1907 | |
| 1908 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X112, batch_div_112) { |
| 1909 | TEST_REQUIRES_X86_AVX512F; |
| 1910 | for (size_t batch_size = 224; batch_size < 1120; batch_size += 112) { |
| 1911 | VUnOpMicrokernelTester() |
| 1912 | .batch_size(batch_size) |
| 1913 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1914 | } |
| 1915 | } |
| 1916 | |
| 1917 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X112, batch_lt_112) { |
| 1918 | TEST_REQUIRES_X86_AVX512F; |
| 1919 | for (size_t batch_size = 1; batch_size < 112; batch_size++) { |
| 1920 | VUnOpMicrokernelTester() |
| 1921 | .batch_size(batch_size) |
| 1922 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1923 | } |
| 1924 | } |
| 1925 | |
| 1926 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X112, batch_gt_112) { |
| 1927 | TEST_REQUIRES_X86_AVX512F; |
| 1928 | for (size_t batch_size = 113; batch_size < 224; batch_size++) { |
| 1929 | VUnOpMicrokernelTester() |
| 1930 | .batch_size(batch_size) |
| 1931 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1932 | } |
| 1933 | } |
| 1934 | |
| 1935 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X112, inplace) { |
| 1936 | TEST_REQUIRES_X86_AVX512F; |
| 1937 | for (size_t batch_size = 1; batch_size <= 560; batch_size += 111) { |
| 1938 | VUnOpMicrokernelTester() |
| 1939 | .batch_size(batch_size) |
| 1940 | .inplace(true) |
| 1941 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1942 | } |
| 1943 | } |
| 1944 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1945 | |
| 1946 | |
| 1947 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1948 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X128, batch_eq_128) { |
| 1949 | TEST_REQUIRES_X86_AVX512F; |
| 1950 | VUnOpMicrokernelTester() |
| 1951 | .batch_size(128) |
| 1952 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1953 | } |
| 1954 | |
| 1955 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X128, batch_div_128) { |
| 1956 | TEST_REQUIRES_X86_AVX512F; |
| 1957 | for (size_t batch_size = 256; batch_size < 1280; batch_size += 128) { |
| 1958 | VUnOpMicrokernelTester() |
| 1959 | .batch_size(batch_size) |
| 1960 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1961 | } |
| 1962 | } |
| 1963 | |
| 1964 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X128, batch_lt_128) { |
| 1965 | TEST_REQUIRES_X86_AVX512F; |
| 1966 | for (size_t batch_size = 1; batch_size < 128; batch_size++) { |
| 1967 | VUnOpMicrokernelTester() |
| 1968 | .batch_size(batch_size) |
| 1969 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1970 | } |
| 1971 | } |
| 1972 | |
| 1973 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X128, batch_gt_128) { |
| 1974 | TEST_REQUIRES_X86_AVX512F; |
| 1975 | for (size_t batch_size = 129; batch_size < 256; batch_size++) { |
| 1976 | VUnOpMicrokernelTester() |
| 1977 | .batch_size(batch_size) |
| 1978 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1979 | } |
| 1980 | } |
| 1981 | |
| 1982 | TEST(F32_VSQRT__AVX512F_NR1FMA1ADJ_X128, inplace) { |
| 1983 | TEST_REQUIRES_X86_AVX512F; |
| 1984 | for (size_t batch_size = 1; batch_size <= 640; batch_size += 127) { |
| 1985 | VUnOpMicrokernelTester() |
| 1986 | .batch_size(batch_size) |
| 1987 | .inplace(true) |
| 1988 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1989 | } |
| 1990 | } |
| 1991 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1992 | |
| 1993 | |
| 1994 | #if XNN_ARCH_WASMSIMD |
| 1995 | TEST(F32_VSQRT__WASMSIMD_SQRT_X4, batch_eq_4) { |
| 1996 | VUnOpMicrokernelTester() |
| 1997 | .batch_size(4) |
| 1998 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 1999 | } |
| 2000 | |
| 2001 | TEST(F32_VSQRT__WASMSIMD_SQRT_X4, batch_div_4) { |
| 2002 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 2003 | VUnOpMicrokernelTester() |
| 2004 | .batch_size(batch_size) |
| 2005 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 2006 | } |
| 2007 | } |
| 2008 | |
| 2009 | TEST(F32_VSQRT__WASMSIMD_SQRT_X4, batch_lt_4) { |
| 2010 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 2011 | VUnOpMicrokernelTester() |
| 2012 | .batch_size(batch_size) |
| 2013 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 2014 | } |
| 2015 | } |
| 2016 | |
| 2017 | TEST(F32_VSQRT__WASMSIMD_SQRT_X4, batch_gt_4) { |
| 2018 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 2019 | VUnOpMicrokernelTester() |
| 2020 | .batch_size(batch_size) |
| 2021 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 2022 | } |
| 2023 | } |
| 2024 | |
| 2025 | TEST(F32_VSQRT__WASMSIMD_SQRT_X4, inplace) { |
| 2026 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 2027 | VUnOpMicrokernelTester() |
| 2028 | .batch_size(batch_size) |
| 2029 | .inplace(true) |
| 2030 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 2031 | } |
| 2032 | } |
| 2033 | #endif // XNN_ARCH_WASMSIMD |
| 2034 | |
| 2035 | |
| 2036 | #if XNN_ARCH_WASMSIMD |
| 2037 | TEST(F32_VSQRT__WASMSIMD_SQRT_X8, batch_eq_8) { |
| 2038 | VUnOpMicrokernelTester() |
| 2039 | .batch_size(8) |
| 2040 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 2041 | } |
| 2042 | |
| 2043 | TEST(F32_VSQRT__WASMSIMD_SQRT_X8, batch_div_8) { |
| 2044 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 2045 | VUnOpMicrokernelTester() |
| 2046 | .batch_size(batch_size) |
| 2047 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 2048 | } |
| 2049 | } |
| 2050 | |
| 2051 | TEST(F32_VSQRT__WASMSIMD_SQRT_X8, batch_lt_8) { |
| 2052 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 2053 | VUnOpMicrokernelTester() |
| 2054 | .batch_size(batch_size) |
| 2055 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 2056 | } |
| 2057 | } |
| 2058 | |
| 2059 | TEST(F32_VSQRT__WASMSIMD_SQRT_X8, batch_gt_8) { |
| 2060 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 2061 | VUnOpMicrokernelTester() |
| 2062 | .batch_size(batch_size) |
| 2063 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 2064 | } |
| 2065 | } |
| 2066 | |
| 2067 | TEST(F32_VSQRT__WASMSIMD_SQRT_X8, inplace) { |
| 2068 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 2069 | VUnOpMicrokernelTester() |
| 2070 | .batch_size(batch_size) |
| 2071 | .inplace(true) |
| 2072 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8), VUnOpMicrokernelTester::OpType::SquareRoot); |
| 2073 | } |
| 2074 | } |
| 2075 | #endif // XNN_ARCH_WASMSIMD |
| 2076 | |
| 2077 | |
| 2078 | TEST(F32_VSQRT__SCALAR_SQRT_X1, batch_eq_1) { |
| 2079 | VUnOpMicrokernelTester() |
| 2080 | .batch_size(1) |
| 2081 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x1), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2082 | } |
| 2083 | |
| 2084 | TEST(F32_VSQRT__SCALAR_SQRT_X1, batch_gt_1) { |
| 2085 | for (size_t batch_size = 2; batch_size < 10; batch_size++) { |
| 2086 | VUnOpMicrokernelTester() |
| 2087 | .batch_size(batch_size) |
| 2088 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x1), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2089 | } |
| 2090 | } |
| 2091 | |
| 2092 | TEST(F32_VSQRT__SCALAR_SQRT_X1, inplace) { |
| 2093 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
| 2094 | VUnOpMicrokernelTester() |
| 2095 | .batch_size(batch_size) |
| 2096 | .inplace(true) |
| 2097 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x1), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2098 | } |
| 2099 | } |
| 2100 | |
| 2101 | |
| 2102 | TEST(F32_VSQRT__SCALAR_SQRT_X2, batch_eq_2) { |
| 2103 | VUnOpMicrokernelTester() |
| 2104 | .batch_size(2) |
| 2105 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x2), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2106 | } |
| 2107 | |
| 2108 | TEST(F32_VSQRT__SCALAR_SQRT_X2, batch_div_2) { |
| 2109 | for (size_t batch_size = 4; batch_size < 20; batch_size += 2) { |
| 2110 | VUnOpMicrokernelTester() |
| 2111 | .batch_size(batch_size) |
| 2112 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x2), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2113 | } |
| 2114 | } |
| 2115 | |
| 2116 | TEST(F32_VSQRT__SCALAR_SQRT_X2, batch_lt_2) { |
| 2117 | for (size_t batch_size = 1; batch_size < 2; batch_size++) { |
| 2118 | VUnOpMicrokernelTester() |
| 2119 | .batch_size(batch_size) |
| 2120 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x2), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2121 | } |
| 2122 | } |
| 2123 | |
| 2124 | TEST(F32_VSQRT__SCALAR_SQRT_X2, batch_gt_2) { |
| 2125 | for (size_t batch_size = 3; batch_size < 4; batch_size++) { |
| 2126 | VUnOpMicrokernelTester() |
| 2127 | .batch_size(batch_size) |
| 2128 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x2), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2129 | } |
| 2130 | } |
| 2131 | |
| 2132 | TEST(F32_VSQRT__SCALAR_SQRT_X2, inplace) { |
| 2133 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
| 2134 | VUnOpMicrokernelTester() |
| 2135 | .batch_size(batch_size) |
| 2136 | .inplace(true) |
| 2137 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x2), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2138 | } |
| 2139 | } |
| 2140 | |
| 2141 | |
| 2142 | TEST(F32_VSQRT__SCALAR_SQRT_X4, batch_eq_4) { |
| 2143 | VUnOpMicrokernelTester() |
| 2144 | .batch_size(4) |
| 2145 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2146 | } |
| 2147 | |
| 2148 | TEST(F32_VSQRT__SCALAR_SQRT_X4, batch_div_4) { |
| 2149 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
| 2150 | VUnOpMicrokernelTester() |
| 2151 | .batch_size(batch_size) |
| 2152 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2153 | } |
| 2154 | } |
| 2155 | |
| 2156 | TEST(F32_VSQRT__SCALAR_SQRT_X4, batch_lt_4) { |
| 2157 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
| 2158 | VUnOpMicrokernelTester() |
| 2159 | .batch_size(batch_size) |
| 2160 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2161 | } |
| 2162 | } |
| 2163 | |
| 2164 | TEST(F32_VSQRT__SCALAR_SQRT_X4, batch_gt_4) { |
| 2165 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
| 2166 | VUnOpMicrokernelTester() |
| 2167 | .batch_size(batch_size) |
| 2168 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2169 | } |
| 2170 | } |
| 2171 | |
| 2172 | TEST(F32_VSQRT__SCALAR_SQRT_X4, inplace) { |
| 2173 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
| 2174 | VUnOpMicrokernelTester() |
| 2175 | .batch_size(batch_size) |
| 2176 | .inplace(true) |
| 2177 | .Test(xnn_f32_vunary_ukernel_function(xnn_f32_vsqrt_ukernel__scalar_sqrt_x4), VUnOpMicrokernelTester::OpType::SquareRoot, VUnOpMicrokernelTester::Variant::Scalar); |
| 2178 | } |
| 2179 | } |