Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -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. |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 5 | // |
| 6 | // Auto-generated file. Do not edit! |
| 7 | // Specification: test/f32-vscaleexpminusmax.yaml |
| 8 | // Generator: tools/generate-vscaleexpminusmax-test.py |
| 9 | |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 10 | |
| 11 | #include <gtest/gtest.h> |
| 12 | |
| 13 | #include <xnnpack/common.h> |
| 14 | #include <xnnpack/isa-checks.h> |
| 15 | |
| 16 | #include <xnnpack/vscaleexpminusmax.h> |
| 17 | #include "vscaleexpminusmax-microkernel-tester.h" |
| 18 | |
| 19 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 20 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 21 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X8, elements_eq_8) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 22 | TEST_REQUIRES_X86_AVX2; |
| 23 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 24 | .elements(8) |
| 25 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x8); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 26 | } |
| 27 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 28 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X8, elements_div_8) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 29 | TEST_REQUIRES_X86_AVX2; |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 30 | for (size_t elements = 16; elements < 80; elements += 8) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 31 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 32 | .elements(elements) |
| 33 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x8); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 34 | } |
| 35 | } |
| 36 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 37 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X8, elements_lt_8) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 38 | TEST_REQUIRES_X86_AVX2; |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 39 | for (size_t elements = 1; elements < 8; elements++) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 40 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 41 | .elements(elements) |
| 42 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x8); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 43 | } |
| 44 | } |
| 45 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 46 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X8, elements_gt_8) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 47 | TEST_REQUIRES_X86_AVX2; |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 48 | for (size_t elements = 9; elements < 16; elements++) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 49 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 50 | .elements(elements) |
| 51 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x8); |
| 52 | } |
| 53 | } |
| 54 | |
| 55 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X8, scale) { |
| 56 | TEST_REQUIRES_X86_AVX2; |
| 57 | for (size_t elements = 1; elements <= 40; elements += 7) { |
| 58 | VScaleExpMinusMaxMicrokernelTester() |
| 59 | .elements(elements) |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 60 | .scale(0.01f) |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 61 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x8); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 62 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 63 | .elements(elements) |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 64 | .scale(100.0f) |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 65 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x8); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 66 | } |
| 67 | } |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 68 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 69 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 70 | |
| 71 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 72 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X16, elements_eq_16) { |
| 73 | TEST_REQUIRES_X86_AVX2; |
| 74 | VScaleExpMinusMaxMicrokernelTester() |
| 75 | .elements(16) |
| 76 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x16); |
| 77 | } |
| 78 | |
| 79 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X16, elements_div_16) { |
| 80 | TEST_REQUIRES_X86_AVX2; |
| 81 | for (size_t elements = 32; elements < 160; elements += 16) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 82 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 83 | .elements(elements) |
| 84 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x16); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 85 | } |
| 86 | } |
| 87 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 88 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X16, elements_lt_16) { |
| 89 | TEST_REQUIRES_X86_AVX2; |
| 90 | for (size_t elements = 1; elements < 16; elements++) { |
| 91 | VScaleExpMinusMaxMicrokernelTester() |
| 92 | .elements(elements) |
| 93 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x16); |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X16, elements_gt_16) { |
| 98 | TEST_REQUIRES_X86_AVX2; |
| 99 | for (size_t elements = 17; elements < 32; elements++) { |
| 100 | VScaleExpMinusMaxMicrokernelTester() |
| 101 | .elements(elements) |
| 102 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x16); |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X16, scale) { |
| 107 | TEST_REQUIRES_X86_AVX2; |
| 108 | for (size_t elements = 1; elements <= 80; elements += 15) { |
| 109 | VScaleExpMinusMaxMicrokernelTester() |
| 110 | .elements(elements) |
| 111 | .scale(0.01f) |
| 112 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x16); |
| 113 | VScaleExpMinusMaxMicrokernelTester() |
| 114 | .elements(elements) |
| 115 | .scale(100.0f) |
| 116 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x16); |
| 117 | } |
| 118 | } |
| 119 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 120 | |
| 121 | |
| 122 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 123 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X24, elements_eq_24) { |
| 124 | TEST_REQUIRES_X86_AVX2; |
| 125 | VScaleExpMinusMaxMicrokernelTester() |
| 126 | .elements(24) |
| 127 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x24); |
| 128 | } |
| 129 | |
| 130 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X24, elements_div_24) { |
| 131 | TEST_REQUIRES_X86_AVX2; |
| 132 | for (size_t elements = 48; elements < 240; elements += 24) { |
| 133 | VScaleExpMinusMaxMicrokernelTester() |
| 134 | .elements(elements) |
| 135 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x24); |
| 136 | } |
| 137 | } |
| 138 | |
| 139 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X24, elements_lt_24) { |
| 140 | TEST_REQUIRES_X86_AVX2; |
| 141 | for (size_t elements = 1; elements < 24; elements++) { |
| 142 | VScaleExpMinusMaxMicrokernelTester() |
| 143 | .elements(elements) |
| 144 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x24); |
| 145 | } |
| 146 | } |
| 147 | |
| 148 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X24, elements_gt_24) { |
| 149 | TEST_REQUIRES_X86_AVX2; |
| 150 | for (size_t elements = 25; elements < 48; elements++) { |
| 151 | VScaleExpMinusMaxMicrokernelTester() |
| 152 | .elements(elements) |
| 153 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x24); |
| 154 | } |
| 155 | } |
| 156 | |
| 157 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X24, scale) { |
| 158 | TEST_REQUIRES_X86_AVX2; |
| 159 | for (size_t elements = 1; elements <= 120; elements += 23) { |
| 160 | VScaleExpMinusMaxMicrokernelTester() |
| 161 | .elements(elements) |
| 162 | .scale(0.01f) |
| 163 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x24); |
| 164 | VScaleExpMinusMaxMicrokernelTester() |
| 165 | .elements(elements) |
| 166 | .scale(100.0f) |
| 167 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x24); |
| 168 | } |
| 169 | } |
| 170 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 171 | |
| 172 | |
| 173 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 174 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X32, elements_eq_32) { |
| 175 | TEST_REQUIRES_X86_AVX2; |
| 176 | VScaleExpMinusMaxMicrokernelTester() |
| 177 | .elements(32) |
| 178 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x32); |
| 179 | } |
| 180 | |
| 181 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X32, elements_div_32) { |
| 182 | TEST_REQUIRES_X86_AVX2; |
| 183 | for (size_t elements = 64; elements < 320; elements += 32) { |
| 184 | VScaleExpMinusMaxMicrokernelTester() |
| 185 | .elements(elements) |
| 186 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x32); |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X32, elements_lt_32) { |
| 191 | TEST_REQUIRES_X86_AVX2; |
| 192 | for (size_t elements = 1; elements < 32; elements++) { |
| 193 | VScaleExpMinusMaxMicrokernelTester() |
| 194 | .elements(elements) |
| 195 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x32); |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X32, elements_gt_32) { |
| 200 | TEST_REQUIRES_X86_AVX2; |
| 201 | for (size_t elements = 33; elements < 64; elements++) { |
| 202 | VScaleExpMinusMaxMicrokernelTester() |
| 203 | .elements(elements) |
| 204 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x32); |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X32, scale) { |
| 209 | TEST_REQUIRES_X86_AVX2; |
| 210 | for (size_t elements = 1; elements <= 160; elements += 31) { |
| 211 | VScaleExpMinusMaxMicrokernelTester() |
| 212 | .elements(elements) |
| 213 | .scale(0.01f) |
| 214 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x32); |
| 215 | VScaleExpMinusMaxMicrokernelTester() |
| 216 | .elements(elements) |
| 217 | .scale(100.0f) |
| 218 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x32); |
| 219 | } |
| 220 | } |
| 221 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 222 | |
| 223 | |
| 224 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 225 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X40, elements_eq_40) { |
| 226 | TEST_REQUIRES_X86_AVX2; |
| 227 | VScaleExpMinusMaxMicrokernelTester() |
| 228 | .elements(40) |
| 229 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x40); |
| 230 | } |
| 231 | |
| 232 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X40, elements_div_40) { |
| 233 | TEST_REQUIRES_X86_AVX2; |
| 234 | for (size_t elements = 80; elements < 400; elements += 40) { |
| 235 | VScaleExpMinusMaxMicrokernelTester() |
| 236 | .elements(elements) |
| 237 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x40); |
| 238 | } |
| 239 | } |
| 240 | |
| 241 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X40, elements_lt_40) { |
| 242 | TEST_REQUIRES_X86_AVX2; |
| 243 | for (size_t elements = 1; elements < 40; elements++) { |
| 244 | VScaleExpMinusMaxMicrokernelTester() |
| 245 | .elements(elements) |
| 246 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x40); |
| 247 | } |
| 248 | } |
| 249 | |
| 250 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X40, elements_gt_40) { |
| 251 | TEST_REQUIRES_X86_AVX2; |
| 252 | for (size_t elements = 41; elements < 80; elements++) { |
| 253 | VScaleExpMinusMaxMicrokernelTester() |
| 254 | .elements(elements) |
| 255 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x40); |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X40, scale) { |
| 260 | TEST_REQUIRES_X86_AVX2; |
| 261 | for (size_t elements = 1; elements <= 200; elements += 39) { |
| 262 | VScaleExpMinusMaxMicrokernelTester() |
| 263 | .elements(elements) |
| 264 | .scale(0.01f) |
| 265 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x40); |
| 266 | VScaleExpMinusMaxMicrokernelTester() |
| 267 | .elements(elements) |
| 268 | .scale(100.0f) |
| 269 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x40); |
| 270 | } |
| 271 | } |
| 272 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 273 | |
| 274 | |
| 275 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 276 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X48, elements_eq_48) { |
| 277 | TEST_REQUIRES_X86_AVX2; |
| 278 | VScaleExpMinusMaxMicrokernelTester() |
| 279 | .elements(48) |
| 280 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48); |
| 281 | } |
| 282 | |
| 283 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X48, elements_div_48) { |
| 284 | TEST_REQUIRES_X86_AVX2; |
| 285 | for (size_t elements = 96; elements < 480; elements += 48) { |
| 286 | VScaleExpMinusMaxMicrokernelTester() |
| 287 | .elements(elements) |
| 288 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48); |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X48, elements_lt_48) { |
| 293 | TEST_REQUIRES_X86_AVX2; |
| 294 | for (size_t elements = 1; elements < 48; elements++) { |
| 295 | VScaleExpMinusMaxMicrokernelTester() |
| 296 | .elements(elements) |
| 297 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48); |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X48, elements_gt_48) { |
| 302 | TEST_REQUIRES_X86_AVX2; |
| 303 | for (size_t elements = 49; elements < 96; elements++) { |
| 304 | VScaleExpMinusMaxMicrokernelTester() |
| 305 | .elements(elements) |
| 306 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48); |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X48, scale) { |
| 311 | TEST_REQUIRES_X86_AVX2; |
| 312 | for (size_t elements = 1; elements <= 240; elements += 47) { |
| 313 | VScaleExpMinusMaxMicrokernelTester() |
| 314 | .elements(elements) |
| 315 | .scale(0.01f) |
| 316 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48); |
| 317 | VScaleExpMinusMaxMicrokernelTester() |
| 318 | .elements(elements) |
| 319 | .scale(100.0f) |
| 320 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48); |
| 321 | } |
| 322 | } |
| 323 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 324 | |
| 325 | |
| 326 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 327 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X56, elements_eq_56) { |
| 328 | TEST_REQUIRES_X86_AVX2; |
| 329 | VScaleExpMinusMaxMicrokernelTester() |
| 330 | .elements(56) |
| 331 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56); |
| 332 | } |
| 333 | |
| 334 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X56, elements_div_56) { |
| 335 | TEST_REQUIRES_X86_AVX2; |
| 336 | for (size_t elements = 112; elements < 560; elements += 56) { |
| 337 | VScaleExpMinusMaxMicrokernelTester() |
| 338 | .elements(elements) |
| 339 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56); |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X56, elements_lt_56) { |
| 344 | TEST_REQUIRES_X86_AVX2; |
| 345 | for (size_t elements = 1; elements < 56; elements++) { |
| 346 | VScaleExpMinusMaxMicrokernelTester() |
| 347 | .elements(elements) |
| 348 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56); |
| 349 | } |
| 350 | } |
| 351 | |
| 352 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X56, elements_gt_56) { |
| 353 | TEST_REQUIRES_X86_AVX2; |
| 354 | for (size_t elements = 57; elements < 112; elements++) { |
| 355 | VScaleExpMinusMaxMicrokernelTester() |
| 356 | .elements(elements) |
| 357 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56); |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X56, scale) { |
| 362 | TEST_REQUIRES_X86_AVX2; |
| 363 | for (size_t elements = 1; elements <= 280; elements += 55) { |
| 364 | VScaleExpMinusMaxMicrokernelTester() |
| 365 | .elements(elements) |
| 366 | .scale(0.01f) |
| 367 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56); |
| 368 | VScaleExpMinusMaxMicrokernelTester() |
| 369 | .elements(elements) |
| 370 | .scale(100.0f) |
| 371 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56); |
| 372 | } |
| 373 | } |
| 374 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 375 | |
| 376 | |
| 377 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 378 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X64, elements_eq_64) { |
| 379 | TEST_REQUIRES_X86_AVX2; |
| 380 | VScaleExpMinusMaxMicrokernelTester() |
| 381 | .elements(64) |
| 382 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64); |
| 383 | } |
| 384 | |
| 385 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X64, elements_div_64) { |
| 386 | TEST_REQUIRES_X86_AVX2; |
| 387 | for (size_t elements = 128; elements < 640; elements += 64) { |
| 388 | VScaleExpMinusMaxMicrokernelTester() |
| 389 | .elements(elements) |
| 390 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64); |
| 391 | } |
| 392 | } |
| 393 | |
| 394 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X64, elements_lt_64) { |
| 395 | TEST_REQUIRES_X86_AVX2; |
| 396 | for (size_t elements = 1; elements < 64; elements++) { |
| 397 | VScaleExpMinusMaxMicrokernelTester() |
| 398 | .elements(elements) |
| 399 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64); |
| 400 | } |
| 401 | } |
| 402 | |
| 403 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X64, elements_gt_64) { |
| 404 | TEST_REQUIRES_X86_AVX2; |
| 405 | for (size_t elements = 65; elements < 128; elements++) { |
| 406 | VScaleExpMinusMaxMicrokernelTester() |
| 407 | .elements(elements) |
| 408 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64); |
| 409 | } |
| 410 | } |
| 411 | |
| 412 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X64, scale) { |
| 413 | TEST_REQUIRES_X86_AVX2; |
| 414 | for (size_t elements = 1; elements <= 320; elements += 63) { |
| 415 | VScaleExpMinusMaxMicrokernelTester() |
| 416 | .elements(elements) |
| 417 | .scale(0.01f) |
| 418 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64); |
| 419 | VScaleExpMinusMaxMicrokernelTester() |
| 420 | .elements(elements) |
| 421 | .scale(100.0f) |
| 422 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x64); |
| 423 | } |
| 424 | } |
| 425 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 426 | |
| 427 | |
| 428 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 429 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X72, elements_eq_72) { |
| 430 | TEST_REQUIRES_X86_AVX2; |
| 431 | VScaleExpMinusMaxMicrokernelTester() |
| 432 | .elements(72) |
| 433 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72); |
| 434 | } |
| 435 | |
| 436 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X72, elements_div_72) { |
| 437 | TEST_REQUIRES_X86_AVX2; |
| 438 | for (size_t elements = 144; elements < 720; elements += 72) { |
| 439 | VScaleExpMinusMaxMicrokernelTester() |
| 440 | .elements(elements) |
| 441 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72); |
| 442 | } |
| 443 | } |
| 444 | |
| 445 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X72, elements_lt_72) { |
| 446 | TEST_REQUIRES_X86_AVX2; |
| 447 | for (size_t elements = 1; elements < 72; elements++) { |
| 448 | VScaleExpMinusMaxMicrokernelTester() |
| 449 | .elements(elements) |
| 450 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72); |
| 451 | } |
| 452 | } |
| 453 | |
| 454 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X72, elements_gt_72) { |
| 455 | TEST_REQUIRES_X86_AVX2; |
| 456 | for (size_t elements = 73; elements < 144; elements++) { |
| 457 | VScaleExpMinusMaxMicrokernelTester() |
| 458 | .elements(elements) |
| 459 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72); |
| 460 | } |
| 461 | } |
| 462 | |
| 463 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X72, scale) { |
| 464 | TEST_REQUIRES_X86_AVX2; |
| 465 | for (size_t elements = 1; elements <= 360; elements += 71) { |
| 466 | VScaleExpMinusMaxMicrokernelTester() |
| 467 | .elements(elements) |
| 468 | .scale(0.01f) |
| 469 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72); |
| 470 | VScaleExpMinusMaxMicrokernelTester() |
| 471 | .elements(elements) |
| 472 | .scale(100.0f) |
| 473 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72); |
| 474 | } |
| 475 | } |
| 476 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 477 | |
| 478 | |
| 479 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 480 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X80, elements_eq_80) { |
| 481 | TEST_REQUIRES_X86_AVX2; |
| 482 | VScaleExpMinusMaxMicrokernelTester() |
| 483 | .elements(80) |
| 484 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80); |
| 485 | } |
| 486 | |
| 487 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X80, elements_div_80) { |
| 488 | TEST_REQUIRES_X86_AVX2; |
| 489 | for (size_t elements = 160; elements < 800; elements += 80) { |
| 490 | VScaleExpMinusMaxMicrokernelTester() |
| 491 | .elements(elements) |
| 492 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80); |
| 493 | } |
| 494 | } |
| 495 | |
| 496 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X80, elements_lt_80) { |
| 497 | TEST_REQUIRES_X86_AVX2; |
| 498 | for (size_t elements = 1; elements < 80; elements++) { |
| 499 | VScaleExpMinusMaxMicrokernelTester() |
| 500 | .elements(elements) |
| 501 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80); |
| 502 | } |
| 503 | } |
| 504 | |
| 505 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X80, elements_gt_80) { |
| 506 | TEST_REQUIRES_X86_AVX2; |
| 507 | for (size_t elements = 81; elements < 160; elements++) { |
| 508 | VScaleExpMinusMaxMicrokernelTester() |
| 509 | .elements(elements) |
| 510 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80); |
| 511 | } |
| 512 | } |
| 513 | |
| 514 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X80, scale) { |
| 515 | TEST_REQUIRES_X86_AVX2; |
| 516 | for (size_t elements = 1; elements <= 400; elements += 79) { |
| 517 | VScaleExpMinusMaxMicrokernelTester() |
| 518 | .elements(elements) |
| 519 | .scale(0.01f) |
| 520 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80); |
| 521 | VScaleExpMinusMaxMicrokernelTester() |
| 522 | .elements(elements) |
| 523 | .scale(100.0f) |
| 524 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x80); |
| 525 | } |
| 526 | } |
| 527 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 528 | |
| 529 | |
| 530 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 531 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X88, elements_eq_88) { |
| 532 | TEST_REQUIRES_X86_AVX2; |
| 533 | VScaleExpMinusMaxMicrokernelTester() |
| 534 | .elements(88) |
| 535 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88); |
| 536 | } |
| 537 | |
| 538 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X88, elements_div_88) { |
| 539 | TEST_REQUIRES_X86_AVX2; |
| 540 | for (size_t elements = 176; elements < 880; elements += 88) { |
| 541 | VScaleExpMinusMaxMicrokernelTester() |
| 542 | .elements(elements) |
| 543 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88); |
| 544 | } |
| 545 | } |
| 546 | |
| 547 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X88, elements_lt_88) { |
| 548 | TEST_REQUIRES_X86_AVX2; |
| 549 | for (size_t elements = 1; elements < 88; elements++) { |
| 550 | VScaleExpMinusMaxMicrokernelTester() |
| 551 | .elements(elements) |
| 552 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88); |
| 553 | } |
| 554 | } |
| 555 | |
| 556 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X88, elements_gt_88) { |
| 557 | TEST_REQUIRES_X86_AVX2; |
| 558 | for (size_t elements = 89; elements < 176; elements++) { |
| 559 | VScaleExpMinusMaxMicrokernelTester() |
| 560 | .elements(elements) |
| 561 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88); |
| 562 | } |
| 563 | } |
| 564 | |
| 565 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X88, scale) { |
| 566 | TEST_REQUIRES_X86_AVX2; |
| 567 | for (size_t elements = 1; elements <= 440; elements += 87) { |
| 568 | VScaleExpMinusMaxMicrokernelTester() |
| 569 | .elements(elements) |
| 570 | .scale(0.01f) |
| 571 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88); |
| 572 | VScaleExpMinusMaxMicrokernelTester() |
| 573 | .elements(elements) |
| 574 | .scale(100.0f) |
| 575 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x88); |
| 576 | } |
| 577 | } |
| 578 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 579 | |
| 580 | |
| 581 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 582 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X96, elements_eq_96) { |
| 583 | TEST_REQUIRES_X86_AVX2; |
| 584 | VScaleExpMinusMaxMicrokernelTester() |
| 585 | .elements(96) |
| 586 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96); |
| 587 | } |
| 588 | |
| 589 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X96, elements_div_96) { |
| 590 | TEST_REQUIRES_X86_AVX2; |
| 591 | for (size_t elements = 192; elements < 960; elements += 96) { |
| 592 | VScaleExpMinusMaxMicrokernelTester() |
| 593 | .elements(elements) |
| 594 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96); |
| 595 | } |
| 596 | } |
| 597 | |
| 598 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X96, elements_lt_96) { |
| 599 | TEST_REQUIRES_X86_AVX2; |
| 600 | for (size_t elements = 1; elements < 96; elements++) { |
| 601 | VScaleExpMinusMaxMicrokernelTester() |
| 602 | .elements(elements) |
| 603 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96); |
| 604 | } |
| 605 | } |
| 606 | |
| 607 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X96, elements_gt_96) { |
| 608 | TEST_REQUIRES_X86_AVX2; |
| 609 | for (size_t elements = 97; elements < 192; elements++) { |
| 610 | VScaleExpMinusMaxMicrokernelTester() |
| 611 | .elements(elements) |
| 612 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96); |
| 613 | } |
| 614 | } |
| 615 | |
| 616 | TEST(F32_VSCALEEXPMINUSMAX__AVX2_P5_X96, scale) { |
| 617 | TEST_REQUIRES_X86_AVX2; |
| 618 | for (size_t elements = 1; elements <= 480; elements += 95) { |
| 619 | VScaleExpMinusMaxMicrokernelTester() |
| 620 | .elements(elements) |
| 621 | .scale(0.01f) |
| 622 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96); |
| 623 | VScaleExpMinusMaxMicrokernelTester() |
| 624 | .elements(elements) |
| 625 | .scale(100.0f) |
| 626 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x96); |
| 627 | } |
| 628 | } |
| 629 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 630 | |
| 631 | |
| 632 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 633 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X16, elements_eq_16) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 634 | TEST_REQUIRES_X86_AVX512F; |
| 635 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 636 | .elements(16) |
| 637 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 638 | } |
| 639 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 640 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X16, elements_div_16) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 641 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 642 | for (size_t elements = 32; elements < 160; elements += 16) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 643 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 644 | .elements(elements) |
| 645 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 646 | } |
| 647 | } |
| 648 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 649 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X16, elements_lt_16) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 650 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 651 | for (size_t elements = 1; elements < 16; elements++) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 652 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 653 | .elements(elements) |
| 654 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 655 | } |
| 656 | } |
| 657 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 658 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X16, elements_gt_16) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 659 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 660 | for (size_t elements = 17; elements < 32; elements++) { |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 661 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 662 | .elements(elements) |
| 663 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16); |
| 664 | } |
| 665 | } |
| 666 | |
| 667 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X16, scale) { |
| 668 | TEST_REQUIRES_X86_AVX512F; |
| 669 | for (size_t elements = 1; elements <= 80; elements += 15) { |
| 670 | VScaleExpMinusMaxMicrokernelTester() |
| 671 | .elements(elements) |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 672 | .scale(0.01f) |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 673 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 674 | VScaleExpMinusMaxMicrokernelTester() |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 675 | .elements(elements) |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 676 | .scale(100.0f) |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 677 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x16); |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 678 | } |
| 679 | } |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 680 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 681 | |
| 682 | |
| 683 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 684 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X32, elements_eq_32) { |
| 685 | TEST_REQUIRES_X86_AVX512F; |
| 686 | VScaleExpMinusMaxMicrokernelTester() |
| 687 | .elements(32) |
| 688 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32); |
| 689 | } |
| 690 | |
| 691 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X32, elements_div_32) { |
| 692 | TEST_REQUIRES_X86_AVX512F; |
| 693 | for (size_t elements = 64; elements < 320; elements += 32) { |
| 694 | VScaleExpMinusMaxMicrokernelTester() |
| 695 | .elements(elements) |
| 696 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32); |
| 697 | } |
| 698 | } |
| 699 | |
| 700 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X32, elements_lt_32) { |
| 701 | TEST_REQUIRES_X86_AVX512F; |
| 702 | for (size_t elements = 1; elements < 32; elements++) { |
| 703 | VScaleExpMinusMaxMicrokernelTester() |
| 704 | .elements(elements) |
| 705 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32); |
| 706 | } |
| 707 | } |
| 708 | |
| 709 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X32, elements_gt_32) { |
| 710 | TEST_REQUIRES_X86_AVX512F; |
| 711 | for (size_t elements = 33; elements < 64; elements++) { |
| 712 | VScaleExpMinusMaxMicrokernelTester() |
| 713 | .elements(elements) |
| 714 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32); |
| 715 | } |
| 716 | } |
| 717 | |
| 718 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X32, scale) { |
| 719 | TEST_REQUIRES_X86_AVX512F; |
| 720 | for (size_t elements = 1; elements <= 160; elements += 31) { |
| 721 | VScaleExpMinusMaxMicrokernelTester() |
| 722 | .elements(elements) |
| 723 | .scale(0.01f) |
| 724 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32); |
| 725 | VScaleExpMinusMaxMicrokernelTester() |
| 726 | .elements(elements) |
| 727 | .scale(100.0f) |
| 728 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32); |
| 729 | } |
| 730 | } |
| 731 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 732 | |
| 733 | |
| 734 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 735 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X48, elements_eq_48) { |
| 736 | TEST_REQUIRES_X86_AVX512F; |
| 737 | VScaleExpMinusMaxMicrokernelTester() |
| 738 | .elements(48) |
| 739 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x48); |
| 740 | } |
| 741 | |
| 742 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X48, elements_div_48) { |
| 743 | TEST_REQUIRES_X86_AVX512F; |
| 744 | for (size_t elements = 96; elements < 480; elements += 48) { |
| 745 | VScaleExpMinusMaxMicrokernelTester() |
| 746 | .elements(elements) |
| 747 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x48); |
| 748 | } |
| 749 | } |
| 750 | |
| 751 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X48, elements_lt_48) { |
| 752 | TEST_REQUIRES_X86_AVX512F; |
| 753 | for (size_t elements = 1; elements < 48; elements++) { |
| 754 | VScaleExpMinusMaxMicrokernelTester() |
| 755 | .elements(elements) |
| 756 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x48); |
| 757 | } |
| 758 | } |
| 759 | |
| 760 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X48, elements_gt_48) { |
| 761 | TEST_REQUIRES_X86_AVX512F; |
| 762 | for (size_t elements = 49; elements < 96; elements++) { |
| 763 | VScaleExpMinusMaxMicrokernelTester() |
| 764 | .elements(elements) |
| 765 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x48); |
| 766 | } |
| 767 | } |
| 768 | |
| 769 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X48, scale) { |
| 770 | TEST_REQUIRES_X86_AVX512F; |
| 771 | for (size_t elements = 1; elements <= 240; elements += 47) { |
| 772 | VScaleExpMinusMaxMicrokernelTester() |
| 773 | .elements(elements) |
| 774 | .scale(0.01f) |
| 775 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x48); |
| 776 | VScaleExpMinusMaxMicrokernelTester() |
| 777 | .elements(elements) |
| 778 | .scale(100.0f) |
| 779 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x48); |
| 780 | } |
| 781 | } |
| 782 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 783 | |
| 784 | |
| 785 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 786 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X64, elements_eq_64) { |
| 787 | TEST_REQUIRES_X86_AVX512F; |
| 788 | VScaleExpMinusMaxMicrokernelTester() |
| 789 | .elements(64) |
| 790 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x64); |
| 791 | } |
| 792 | |
| 793 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X64, elements_div_64) { |
| 794 | TEST_REQUIRES_X86_AVX512F; |
| 795 | for (size_t elements = 128; elements < 640; elements += 64) { |
| 796 | VScaleExpMinusMaxMicrokernelTester() |
| 797 | .elements(elements) |
| 798 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x64); |
| 799 | } |
| 800 | } |
| 801 | |
| 802 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X64, elements_lt_64) { |
| 803 | TEST_REQUIRES_X86_AVX512F; |
| 804 | for (size_t elements = 1; elements < 64; elements++) { |
| 805 | VScaleExpMinusMaxMicrokernelTester() |
| 806 | .elements(elements) |
| 807 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x64); |
| 808 | } |
| 809 | } |
| 810 | |
| 811 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X64, elements_gt_64) { |
| 812 | TEST_REQUIRES_X86_AVX512F; |
| 813 | for (size_t elements = 65; elements < 128; elements++) { |
| 814 | VScaleExpMinusMaxMicrokernelTester() |
| 815 | .elements(elements) |
| 816 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x64); |
| 817 | } |
| 818 | } |
| 819 | |
| 820 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X64, scale) { |
| 821 | TEST_REQUIRES_X86_AVX512F; |
| 822 | for (size_t elements = 1; elements <= 320; elements += 63) { |
| 823 | VScaleExpMinusMaxMicrokernelTester() |
| 824 | .elements(elements) |
| 825 | .scale(0.01f) |
| 826 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x64); |
| 827 | VScaleExpMinusMaxMicrokernelTester() |
| 828 | .elements(elements) |
| 829 | .scale(100.0f) |
| 830 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x64); |
| 831 | } |
| 832 | } |
| 833 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 834 | |
| 835 | |
| 836 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 837 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X80, elements_eq_80) { |
| 838 | TEST_REQUIRES_X86_AVX512F; |
| 839 | VScaleExpMinusMaxMicrokernelTester() |
| 840 | .elements(80) |
| 841 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x80); |
| 842 | } |
| 843 | |
| 844 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X80, elements_div_80) { |
| 845 | TEST_REQUIRES_X86_AVX512F; |
| 846 | for (size_t elements = 160; elements < 800; elements += 80) { |
| 847 | VScaleExpMinusMaxMicrokernelTester() |
| 848 | .elements(elements) |
| 849 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x80); |
| 850 | } |
| 851 | } |
| 852 | |
| 853 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X80, elements_lt_80) { |
| 854 | TEST_REQUIRES_X86_AVX512F; |
| 855 | for (size_t elements = 1; elements < 80; elements++) { |
| 856 | VScaleExpMinusMaxMicrokernelTester() |
| 857 | .elements(elements) |
| 858 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x80); |
| 859 | } |
| 860 | } |
| 861 | |
| 862 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X80, elements_gt_80) { |
| 863 | TEST_REQUIRES_X86_AVX512F; |
| 864 | for (size_t elements = 81; elements < 160; elements++) { |
| 865 | VScaleExpMinusMaxMicrokernelTester() |
| 866 | .elements(elements) |
| 867 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x80); |
| 868 | } |
| 869 | } |
| 870 | |
| 871 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X80, scale) { |
| 872 | TEST_REQUIRES_X86_AVX512F; |
| 873 | for (size_t elements = 1; elements <= 400; elements += 79) { |
| 874 | VScaleExpMinusMaxMicrokernelTester() |
| 875 | .elements(elements) |
| 876 | .scale(0.01f) |
| 877 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x80); |
| 878 | VScaleExpMinusMaxMicrokernelTester() |
| 879 | .elements(elements) |
| 880 | .scale(100.0f) |
| 881 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x80); |
| 882 | } |
| 883 | } |
| 884 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 885 | |
| 886 | |
| 887 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 888 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X96, elements_eq_96) { |
| 889 | TEST_REQUIRES_X86_AVX512F; |
| 890 | VScaleExpMinusMaxMicrokernelTester() |
| 891 | .elements(96) |
| 892 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96); |
| 893 | } |
| 894 | |
| 895 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X96, elements_div_96) { |
| 896 | TEST_REQUIRES_X86_AVX512F; |
| 897 | for (size_t elements = 192; elements < 960; elements += 96) { |
| 898 | VScaleExpMinusMaxMicrokernelTester() |
| 899 | .elements(elements) |
| 900 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96); |
| 901 | } |
| 902 | } |
| 903 | |
| 904 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X96, elements_lt_96) { |
| 905 | TEST_REQUIRES_X86_AVX512F; |
| 906 | for (size_t elements = 1; elements < 96; elements++) { |
| 907 | VScaleExpMinusMaxMicrokernelTester() |
| 908 | .elements(elements) |
| 909 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96); |
| 910 | } |
| 911 | } |
| 912 | |
| 913 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X96, elements_gt_96) { |
| 914 | TEST_REQUIRES_X86_AVX512F; |
| 915 | for (size_t elements = 97; elements < 192; elements++) { |
| 916 | VScaleExpMinusMaxMicrokernelTester() |
| 917 | .elements(elements) |
| 918 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96); |
| 919 | } |
| 920 | } |
| 921 | |
| 922 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X96, scale) { |
| 923 | TEST_REQUIRES_X86_AVX512F; |
| 924 | for (size_t elements = 1; elements <= 480; elements += 95) { |
| 925 | VScaleExpMinusMaxMicrokernelTester() |
| 926 | .elements(elements) |
| 927 | .scale(0.01f) |
| 928 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96); |
| 929 | VScaleExpMinusMaxMicrokernelTester() |
| 930 | .elements(elements) |
| 931 | .scale(100.0f) |
| 932 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96); |
| 933 | } |
| 934 | } |
| 935 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 936 | |
| 937 | |
| 938 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 939 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X112, elements_eq_112) { |
| 940 | TEST_REQUIRES_X86_AVX512F; |
| 941 | VScaleExpMinusMaxMicrokernelTester() |
| 942 | .elements(112) |
| 943 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112); |
| 944 | } |
| 945 | |
| 946 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X112, elements_div_112) { |
| 947 | TEST_REQUIRES_X86_AVX512F; |
| 948 | for (size_t elements = 224; elements < 1120; elements += 112) { |
| 949 | VScaleExpMinusMaxMicrokernelTester() |
| 950 | .elements(elements) |
| 951 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112); |
| 952 | } |
| 953 | } |
| 954 | |
| 955 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X112, elements_lt_112) { |
| 956 | TEST_REQUIRES_X86_AVX512F; |
| 957 | for (size_t elements = 1; elements < 112; elements++) { |
| 958 | VScaleExpMinusMaxMicrokernelTester() |
| 959 | .elements(elements) |
| 960 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112); |
| 961 | } |
| 962 | } |
| 963 | |
| 964 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X112, elements_gt_112) { |
| 965 | TEST_REQUIRES_X86_AVX512F; |
| 966 | for (size_t elements = 113; elements < 224; elements++) { |
| 967 | VScaleExpMinusMaxMicrokernelTester() |
| 968 | .elements(elements) |
| 969 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112); |
| 970 | } |
| 971 | } |
| 972 | |
| 973 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X112, scale) { |
| 974 | TEST_REQUIRES_X86_AVX512F; |
| 975 | for (size_t elements = 1; elements <= 560; elements += 111) { |
| 976 | VScaleExpMinusMaxMicrokernelTester() |
| 977 | .elements(elements) |
| 978 | .scale(0.01f) |
| 979 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112); |
| 980 | VScaleExpMinusMaxMicrokernelTester() |
| 981 | .elements(elements) |
| 982 | .scale(100.0f) |
| 983 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x112); |
| 984 | } |
| 985 | } |
| 986 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 987 | |
| 988 | |
| 989 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 990 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X128, elements_eq_128) { |
| 991 | TEST_REQUIRES_X86_AVX512F; |
| 992 | VScaleExpMinusMaxMicrokernelTester() |
| 993 | .elements(128) |
| 994 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128); |
| 995 | } |
| 996 | |
| 997 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X128, elements_div_128) { |
| 998 | TEST_REQUIRES_X86_AVX512F; |
| 999 | for (size_t elements = 256; elements < 1280; elements += 128) { |
| 1000 | VScaleExpMinusMaxMicrokernelTester() |
| 1001 | .elements(elements) |
| 1002 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128); |
| 1003 | } |
| 1004 | } |
| 1005 | |
| 1006 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X128, elements_lt_128) { |
| 1007 | TEST_REQUIRES_X86_AVX512F; |
| 1008 | for (size_t elements = 1; elements < 128; elements++) { |
| 1009 | VScaleExpMinusMaxMicrokernelTester() |
| 1010 | .elements(elements) |
| 1011 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128); |
| 1012 | } |
| 1013 | } |
| 1014 | |
| 1015 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X128, elements_gt_128) { |
| 1016 | TEST_REQUIRES_X86_AVX512F; |
| 1017 | for (size_t elements = 129; elements < 256; elements++) { |
| 1018 | VScaleExpMinusMaxMicrokernelTester() |
| 1019 | .elements(elements) |
| 1020 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128); |
| 1021 | } |
| 1022 | } |
| 1023 | |
| 1024 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X128, scale) { |
| 1025 | TEST_REQUIRES_X86_AVX512F; |
| 1026 | for (size_t elements = 1; elements <= 640; elements += 127) { |
| 1027 | VScaleExpMinusMaxMicrokernelTester() |
| 1028 | .elements(elements) |
| 1029 | .scale(0.01f) |
| 1030 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128); |
| 1031 | VScaleExpMinusMaxMicrokernelTester() |
| 1032 | .elements(elements) |
| 1033 | .scale(100.0f) |
| 1034 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128); |
| 1035 | } |
| 1036 | } |
| 1037 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1038 | |
| 1039 | |
| 1040 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1041 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X144, elements_eq_144) { |
| 1042 | TEST_REQUIRES_X86_AVX512F; |
| 1043 | VScaleExpMinusMaxMicrokernelTester() |
| 1044 | .elements(144) |
| 1045 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144); |
| 1046 | } |
| 1047 | |
| 1048 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X144, elements_div_144) { |
| 1049 | TEST_REQUIRES_X86_AVX512F; |
| 1050 | for (size_t elements = 288; elements < 1440; elements += 144) { |
| 1051 | VScaleExpMinusMaxMicrokernelTester() |
| 1052 | .elements(elements) |
| 1053 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144); |
| 1054 | } |
| 1055 | } |
| 1056 | |
| 1057 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X144, elements_lt_144) { |
| 1058 | TEST_REQUIRES_X86_AVX512F; |
| 1059 | for (size_t elements = 1; elements < 144; elements++) { |
| 1060 | VScaleExpMinusMaxMicrokernelTester() |
| 1061 | .elements(elements) |
| 1062 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144); |
| 1063 | } |
| 1064 | } |
| 1065 | |
| 1066 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X144, elements_gt_144) { |
| 1067 | TEST_REQUIRES_X86_AVX512F; |
| 1068 | for (size_t elements = 145; elements < 288; elements++) { |
| 1069 | VScaleExpMinusMaxMicrokernelTester() |
| 1070 | .elements(elements) |
| 1071 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144); |
| 1072 | } |
| 1073 | } |
| 1074 | |
| 1075 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X144, scale) { |
| 1076 | TEST_REQUIRES_X86_AVX512F; |
| 1077 | for (size_t elements = 1; elements <= 720; elements += 143) { |
| 1078 | VScaleExpMinusMaxMicrokernelTester() |
| 1079 | .elements(elements) |
| 1080 | .scale(0.01f) |
| 1081 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144); |
| 1082 | VScaleExpMinusMaxMicrokernelTester() |
| 1083 | .elements(elements) |
| 1084 | .scale(100.0f) |
| 1085 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x144); |
| 1086 | } |
| 1087 | } |
| 1088 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1089 | |
| 1090 | |
| 1091 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1092 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X160, elements_eq_160) { |
| 1093 | TEST_REQUIRES_X86_AVX512F; |
| 1094 | VScaleExpMinusMaxMicrokernelTester() |
| 1095 | .elements(160) |
| 1096 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160); |
| 1097 | } |
| 1098 | |
| 1099 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X160, elements_div_160) { |
| 1100 | TEST_REQUIRES_X86_AVX512F; |
| 1101 | for (size_t elements = 320; elements < 1600; elements += 160) { |
| 1102 | VScaleExpMinusMaxMicrokernelTester() |
| 1103 | .elements(elements) |
| 1104 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160); |
| 1105 | } |
| 1106 | } |
| 1107 | |
| 1108 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X160, elements_lt_160) { |
| 1109 | TEST_REQUIRES_X86_AVX512F; |
| 1110 | for (size_t elements = 1; elements < 160; elements++) { |
| 1111 | VScaleExpMinusMaxMicrokernelTester() |
| 1112 | .elements(elements) |
| 1113 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160); |
| 1114 | } |
| 1115 | } |
| 1116 | |
| 1117 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X160, elements_gt_160) { |
| 1118 | TEST_REQUIRES_X86_AVX512F; |
| 1119 | for (size_t elements = 161; elements < 320; elements++) { |
| 1120 | VScaleExpMinusMaxMicrokernelTester() |
| 1121 | .elements(elements) |
| 1122 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160); |
| 1123 | } |
| 1124 | } |
| 1125 | |
| 1126 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X160, scale) { |
| 1127 | TEST_REQUIRES_X86_AVX512F; |
| 1128 | for (size_t elements = 1; elements <= 800; elements += 159) { |
| 1129 | VScaleExpMinusMaxMicrokernelTester() |
| 1130 | .elements(elements) |
| 1131 | .scale(0.01f) |
| 1132 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160); |
| 1133 | VScaleExpMinusMaxMicrokernelTester() |
| 1134 | .elements(elements) |
| 1135 | .scale(100.0f) |
| 1136 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160); |
| 1137 | } |
| 1138 | } |
| 1139 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1140 | |
| 1141 | |
| 1142 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1143 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X176, elements_eq_176) { |
| 1144 | TEST_REQUIRES_X86_AVX512F; |
| 1145 | VScaleExpMinusMaxMicrokernelTester() |
| 1146 | .elements(176) |
| 1147 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176); |
| 1148 | } |
| 1149 | |
| 1150 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X176, elements_div_176) { |
| 1151 | TEST_REQUIRES_X86_AVX512F; |
| 1152 | for (size_t elements = 352; elements < 1760; elements += 176) { |
| 1153 | VScaleExpMinusMaxMicrokernelTester() |
| 1154 | .elements(elements) |
| 1155 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176); |
| 1156 | } |
| 1157 | } |
| 1158 | |
| 1159 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X176, elements_lt_176) { |
| 1160 | TEST_REQUIRES_X86_AVX512F; |
| 1161 | for (size_t elements = 1; elements < 176; elements++) { |
| 1162 | VScaleExpMinusMaxMicrokernelTester() |
| 1163 | .elements(elements) |
| 1164 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176); |
| 1165 | } |
| 1166 | } |
| 1167 | |
| 1168 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X176, elements_gt_176) { |
| 1169 | TEST_REQUIRES_X86_AVX512F; |
| 1170 | for (size_t elements = 177; elements < 352; elements++) { |
| 1171 | VScaleExpMinusMaxMicrokernelTester() |
| 1172 | .elements(elements) |
| 1173 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176); |
| 1174 | } |
| 1175 | } |
| 1176 | |
| 1177 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X176, scale) { |
| 1178 | TEST_REQUIRES_X86_AVX512F; |
| 1179 | for (size_t elements = 1; elements <= 880; elements += 175) { |
| 1180 | VScaleExpMinusMaxMicrokernelTester() |
| 1181 | .elements(elements) |
| 1182 | .scale(0.01f) |
| 1183 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176); |
| 1184 | VScaleExpMinusMaxMicrokernelTester() |
| 1185 | .elements(elements) |
| 1186 | .scale(100.0f) |
| 1187 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x176); |
| 1188 | } |
| 1189 | } |
| 1190 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1191 | |
| 1192 | |
| 1193 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 1194 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X192, elements_eq_192) { |
| 1195 | TEST_REQUIRES_X86_AVX512F; |
| 1196 | VScaleExpMinusMaxMicrokernelTester() |
| 1197 | .elements(192) |
| 1198 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192); |
| 1199 | } |
| 1200 | |
| 1201 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X192, elements_div_192) { |
| 1202 | TEST_REQUIRES_X86_AVX512F; |
| 1203 | for (size_t elements = 384; elements < 1920; elements += 192) { |
| 1204 | VScaleExpMinusMaxMicrokernelTester() |
| 1205 | .elements(elements) |
| 1206 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192); |
| 1207 | } |
| 1208 | } |
| 1209 | |
| 1210 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X192, elements_lt_192) { |
| 1211 | TEST_REQUIRES_X86_AVX512F; |
| 1212 | for (size_t elements = 1; elements < 192; elements++) { |
| 1213 | VScaleExpMinusMaxMicrokernelTester() |
| 1214 | .elements(elements) |
| 1215 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192); |
| 1216 | } |
| 1217 | } |
| 1218 | |
| 1219 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X192, elements_gt_192) { |
| 1220 | TEST_REQUIRES_X86_AVX512F; |
| 1221 | for (size_t elements = 193; elements < 384; elements++) { |
| 1222 | VScaleExpMinusMaxMicrokernelTester() |
| 1223 | .elements(elements) |
| 1224 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192); |
| 1225 | } |
| 1226 | } |
| 1227 | |
| 1228 | TEST(F32_VSCALEEXPMINUSMAX__AVX512F_P5_SCALEF_X192, scale) { |
| 1229 | TEST_REQUIRES_X86_AVX512F; |
| 1230 | for (size_t elements = 1; elements <= 960; elements += 191) { |
| 1231 | VScaleExpMinusMaxMicrokernelTester() |
| 1232 | .elements(elements) |
| 1233 | .scale(0.01f) |
| 1234 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192); |
| 1235 | VScaleExpMinusMaxMicrokernelTester() |
| 1236 | .elements(elements) |
| 1237 | .scale(100.0f) |
| 1238 | .Test(xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192); |
| 1239 | } |
| 1240 | } |
| 1241 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |