XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -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 | #include <cpuinfo.h> |
| 7 | #include <gtest/gtest.h> |
| 8 | |
| 9 | #include <xnnpack/isa-checks.h> |
| 10 | #include <xnnpack/clamp.h> |
| 11 | |
| 12 | #include "clamp-microkernel-tester.h" |
| 13 | |
| 14 | |
| 15 | #if !CPUINFO_ARCH_WASM && !CPUINFO_ARCH_ASMJS |
| 16 | TEST(F32_CLAMP__PSIMD, n_eq_4) { |
| 17 | TEST_REQUIRES_PSIMD; |
| 18 | ClampMicrokernelTester() |
| 19 | .n(4) |
| 20 | .Test(xnn_f32_clamp_ukernel__psimd, ClampMicrokernelTester::Variant::Scalar); |
| 21 | } |
| 22 | |
| 23 | TEST(F32_CLAMP__PSIMD, n_div_4) { |
| 24 | TEST_REQUIRES_PSIMD; |
| 25 | for (size_t n = 4; n < 256; n += 4) { |
| 26 | ClampMicrokernelTester() |
| 27 | .n(n) |
| 28 | .Test(xnn_f32_clamp_ukernel__psimd, ClampMicrokernelTester::Variant::Scalar); |
| 29 | } |
| 30 | } |
| 31 | |
| 32 | TEST(F32_CLAMP__PSIMD, n_gt_4) { |
| 33 | TEST_REQUIRES_PSIMD; |
| 34 | for (size_t n = 5; n < 8; n++) { |
| 35 | ClampMicrokernelTester() |
| 36 | .n(n) |
| 37 | .Test(xnn_f32_clamp_ukernel__psimd, ClampMicrokernelTester::Variant::Scalar); |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | TEST(F32_CLAMP__PSIMD, n_lt_4) { |
| 42 | TEST_REQUIRES_PSIMD; |
| 43 | for (size_t n = 1; n < 4; n++) { |
| 44 | ClampMicrokernelTester() |
| 45 | .n(n) |
| 46 | .Test(xnn_f32_clamp_ukernel__psimd, ClampMicrokernelTester::Variant::Scalar); |
| 47 | } |
| 48 | } |
| 49 | |
| 50 | TEST(F32_CLAMP__PSIMD, inplace) { |
| 51 | TEST_REQUIRES_PSIMD; |
| 52 | for (size_t n = 1; n < 64; n += 3) { |
| 53 | ClampMicrokernelTester() |
| 54 | .iterations(1) |
| 55 | .n(n) |
| 56 | .inplace(true) |
| 57 | .Test(xnn_f32_clamp_ukernel__psimd, ClampMicrokernelTester::Variant::Scalar); |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | TEST(F32_CLAMP__PSIMD, qmin) { |
| 62 | TEST_REQUIRES_PSIMD; |
| 63 | for (size_t n = 1; n < 64; n += 5) { |
| 64 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 65 | ClampMicrokernelTester() |
| 66 | .iterations(1) |
| 67 | .n(n) |
| 68 | .qmin(qmin) |
| 69 | .qmax(255) |
| 70 | .Test(xnn_f32_clamp_ukernel__psimd, ClampMicrokernelTester::Variant::Scalar); |
| 71 | } |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | TEST(F32_CLAMP__PSIMD, qmax) { |
| 76 | TEST_REQUIRES_PSIMD; |
| 77 | for (size_t n = 1; n < 64; n += 5) { |
| 78 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 79 | ClampMicrokernelTester() |
| 80 | .iterations(1) |
| 81 | .n(n) |
| 82 | .qmin(0) |
| 83 | .qmax(qmax) |
| 84 | .Test(xnn_f32_clamp_ukernel__psimd, ClampMicrokernelTester::Variant::Scalar); |
| 85 | } |
| 86 | } |
| 87 | } |
| 88 | #endif // !CPUINFO_ARCH_WASM && !CPUINFO_ARCH_ASMJS |
| 89 | |
| 90 | |
| 91 | TEST(F32_CLAMP__SCALAR, n_eq_2) { |
| 92 | ClampMicrokernelTester() |
| 93 | .n(2) |
| 94 | .Test(xnn_f32_clamp_ukernel__scalar, ClampMicrokernelTester::Variant::Scalar); |
| 95 | } |
| 96 | |
| 97 | TEST(F32_CLAMP__SCALAR, n_div_2) { |
| 98 | for (size_t n = 4; n < 128; n += 2) { |
| 99 | ClampMicrokernelTester() |
| 100 | .n(n) |
| 101 | .Test(xnn_f32_clamp_ukernel__scalar, ClampMicrokernelTester::Variant::Scalar); |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | TEST(F32_CLAMP__SCALAR, n_gt_2) { |
| 106 | for (size_t n = 3; n < 4; n++) { |
| 107 | ClampMicrokernelTester() |
| 108 | .n(n) |
| 109 | .Test(xnn_f32_clamp_ukernel__scalar, ClampMicrokernelTester::Variant::Scalar); |
| 110 | } |
| 111 | } |
| 112 | |
| 113 | TEST(F32_CLAMP__SCALAR, n_lt_2) { |
| 114 | for (size_t n = 1; n < 2; n++) { |
| 115 | ClampMicrokernelTester() |
| 116 | .n(n) |
| 117 | .Test(xnn_f32_clamp_ukernel__scalar, ClampMicrokernelTester::Variant::Scalar); |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | TEST(F32_CLAMP__SCALAR, inplace) { |
| 122 | for (size_t n = 1; n < 32; n += 3) { |
| 123 | ClampMicrokernelTester() |
| 124 | .iterations(1) |
| 125 | .n(n) |
| 126 | .inplace(true) |
| 127 | .Test(xnn_f32_clamp_ukernel__scalar, ClampMicrokernelTester::Variant::Scalar); |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | TEST(F32_CLAMP__SCALAR, qmin) { |
| 132 | for (size_t n = 1; n < 32; n += 3) { |
| 133 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 134 | ClampMicrokernelTester() |
| 135 | .iterations(1) |
| 136 | .n(n) |
| 137 | .qmin(qmin) |
| 138 | .qmax(255) |
| 139 | .Test(xnn_f32_clamp_ukernel__scalar, ClampMicrokernelTester::Variant::Scalar); |
| 140 | } |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | TEST(F32_CLAMP__SCALAR, qmax) { |
| 145 | for (size_t n = 1; n < 32; n += 3) { |
| 146 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 147 | ClampMicrokernelTester() |
| 148 | .iterations(1) |
| 149 | .n(n) |
| 150 | .qmin(0) |
| 151 | .qmax(qmax) |
| 152 | .Test(xnn_f32_clamp_ukernel__scalar, ClampMicrokernelTester::Variant::Scalar); |
| 153 | } |
| 154 | } |
| 155 | } |
| 156 | |
| 157 | #if CPUINFO_ARCH_ARM || CPUINFO_ARCH_ARM64 |
| 158 | TEST(F32_CLAMP__NEON, n_eq_4) { |
| 159 | TEST_REQUIRES_ARM_NEON; |
| 160 | ClampMicrokernelTester() |
| 161 | .n(4) |
| 162 | .Test(xnn_f32_clamp_ukernel__neon); |
| 163 | } |
| 164 | |
| 165 | TEST(F32_CLAMP__NEON, n_div_4) { |
| 166 | TEST_REQUIRES_ARM_NEON; |
| 167 | for (size_t n = 4; n < 256; n += 4) { |
| 168 | ClampMicrokernelTester() |
| 169 | .n(n) |
| 170 | .Test(xnn_f32_clamp_ukernel__neon); |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | TEST(F32_CLAMP__NEON, n_gt_4) { |
| 175 | TEST_REQUIRES_ARM_NEON; |
| 176 | for (size_t n = 5; n < 8; n++) { |
| 177 | ClampMicrokernelTester() |
| 178 | .n(n) |
| 179 | .Test(xnn_f32_clamp_ukernel__neon); |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | TEST(F32_CLAMP__NEON, n_lt_4) { |
| 184 | TEST_REQUIRES_ARM_NEON; |
| 185 | for (size_t n = 1; n < 4; n++) { |
| 186 | ClampMicrokernelTester() |
| 187 | .n(n) |
| 188 | .Test(xnn_f32_clamp_ukernel__neon); |
| 189 | } |
| 190 | } |
| 191 | |
| 192 | TEST(F32_CLAMP__NEON, inplace) { |
| 193 | TEST_REQUIRES_ARM_NEON; |
| 194 | for (size_t n = 1; n < 64; n += 3) { |
| 195 | ClampMicrokernelTester() |
| 196 | .iterations(1) |
| 197 | .n(n) |
| 198 | .inplace(true) |
| 199 | .Test(xnn_f32_clamp_ukernel__neon); |
| 200 | } |
| 201 | } |
| 202 | |
| 203 | TEST(F32_CLAMP__NEON, qmin) { |
| 204 | TEST_REQUIRES_ARM_NEON; |
| 205 | for (size_t n = 1; n < 64; n += 5) { |
| 206 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 207 | ClampMicrokernelTester() |
| 208 | .iterations(1) |
| 209 | .n(n) |
| 210 | .qmin(qmin) |
| 211 | .qmax(255) |
| 212 | .Test(xnn_f32_clamp_ukernel__neon); |
| 213 | } |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | TEST(F32_CLAMP__NEON, qmax) { |
| 218 | TEST_REQUIRES_ARM_NEON; |
| 219 | for (size_t n = 1; n < 64; n += 5) { |
| 220 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 221 | ClampMicrokernelTester() |
| 222 | .iterations(1) |
| 223 | .n(n) |
| 224 | .qmin(0) |
| 225 | .qmax(qmax) |
| 226 | .Test(xnn_f32_clamp_ukernel__neon); |
| 227 | } |
| 228 | } |
| 229 | } |
| 230 | #endif // CPUINFO_ARCH_ARM || CPUINFO_ARCH_ARM64 |
| 231 | |
| 232 | #if CPUINFO_ARCH_X86 || CPUINFO_ARCH_X86_64 |
| 233 | TEST(F32_CLAMP__SSE, n_eq_4) { |
| 234 | TEST_REQUIRES_X86_SSE2; |
| 235 | ClampMicrokernelTester() |
| 236 | .n(4) |
| 237 | .Test(xnn_f32_clamp_ukernel__sse); |
| 238 | } |
| 239 | |
| 240 | TEST(F32_CLAMP__SSE, n_div_4) { |
| 241 | TEST_REQUIRES_X86_SSE2; |
| 242 | for (size_t n = 4; n < 256; n += 4) { |
| 243 | ClampMicrokernelTester() |
| 244 | .n(n) |
| 245 | .Test(xnn_f32_clamp_ukernel__sse); |
| 246 | } |
| 247 | } |
| 248 | |
| 249 | TEST(F32_CLAMP__SSE, n_gt_4) { |
| 250 | TEST_REQUIRES_X86_SSE2; |
| 251 | for (size_t n = 5; n < 8; n++) { |
| 252 | ClampMicrokernelTester() |
| 253 | .n(n) |
| 254 | .Test(xnn_f32_clamp_ukernel__sse); |
| 255 | } |
| 256 | } |
| 257 | |
| 258 | TEST(F32_CLAMP__SSE, n_lt_4) { |
| 259 | TEST_REQUIRES_X86_SSE2; |
| 260 | for (size_t n = 1; n < 4; n++) { |
| 261 | ClampMicrokernelTester() |
| 262 | .n(n) |
| 263 | .Test(xnn_f32_clamp_ukernel__sse); |
| 264 | } |
| 265 | } |
| 266 | |
| 267 | TEST(F32_CLAMP__SSE, inplace) { |
| 268 | TEST_REQUIRES_X86_SSE2; |
| 269 | for (size_t n = 1; n < 64; n += 3) { |
| 270 | ClampMicrokernelTester() |
| 271 | .iterations(1) |
| 272 | .n(n) |
| 273 | .inplace(true) |
| 274 | .Test(xnn_f32_clamp_ukernel__sse); |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | TEST(F32_CLAMP__SSE, qmin) { |
| 279 | TEST_REQUIRES_X86_SSE2; |
| 280 | for (size_t n = 1; n < 64; n += 5) { |
| 281 | for (uint8_t qmin = 1; qmin < 255; qmin++) { |
| 282 | ClampMicrokernelTester() |
| 283 | .iterations(1) |
| 284 | .n(n) |
| 285 | .qmin(qmin) |
| 286 | .qmax(255) |
| 287 | .Test(xnn_f32_clamp_ukernel__sse); |
| 288 | } |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | TEST(F32_CLAMP__SSE, qmax) { |
| 293 | TEST_REQUIRES_X86_SSE2; |
| 294 | for (size_t n = 1; n < 64; n += 5) { |
| 295 | for (uint8_t qmax = 1; qmax < 255; qmax++) { |
| 296 | ClampMicrokernelTester() |
| 297 | .iterations(1) |
| 298 | .n(n) |
| 299 | .qmin(0) |
| 300 | .qmax(qmax) |
| 301 | .Test(xnn_f32_clamp_ukernel__sse); |
| 302 | } |
| 303 | } |
| 304 | } |
| 305 | #endif // CPUINFO_ARCH_X86 || CPUINFO_ARCH_X86_64 |