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 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 6 | #include <gtest/gtest.h> |
| 7 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 8 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 9 | #include <xnnpack/isa-checks.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 10 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 11 | #include <xnnpack/rmax.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 12 | #include "rmax-microkernel-tester.h" |
| 13 | |
| 14 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 15 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 16 | TEST(F32_RMAX__NEON, n_lt_16) { |
| 17 | TEST_REQUIRES_ARM_NEON; |
| 18 | for (size_t n = 1; n < 16; n++) { |
| 19 | RMaxMicrokernelTester() |
| 20 | .n(n) |
| 21 | .Test(xnn_f32_rmax_ukernel__neon); |
| 22 | } |
| 23 | } |
| 24 | |
| 25 | TEST(F32_RMAX__NEON, n_eq_16) { |
| 26 | TEST_REQUIRES_ARM_NEON; |
| 27 | RMaxMicrokernelTester() |
| 28 | .n(16) |
| 29 | .Test(xnn_f32_rmax_ukernel__neon); |
| 30 | } |
| 31 | |
| 32 | TEST(F32_RMAX__NEON, n_div_16) { |
| 33 | TEST_REQUIRES_ARM_NEON; |
| 34 | for (size_t n = 16; n < 128; n += 16) { |
| 35 | RMaxMicrokernelTester() |
| 36 | .n(n) |
| 37 | .Test(xnn_f32_rmax_ukernel__neon); |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | TEST(F32_RMAX__NEON, n_gt_16) { |
| 42 | TEST_REQUIRES_ARM_NEON; |
| 43 | for (size_t n = 16; n < 32; n++) { |
| 44 | RMaxMicrokernelTester() |
| 45 | .n(n) |
| 46 | .Test(xnn_f32_rmax_ukernel__neon); |
| 47 | } |
| 48 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 49 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 50 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 51 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 52 | TEST(F32_RMAX__SSE, n_lt_16) { |
| 53 | TEST_REQUIRES_X86_SSE2; |
| 54 | for (size_t n = 1; n < 16; n++) { |
| 55 | RMaxMicrokernelTester() |
| 56 | .n(n) |
| 57 | .Test(xnn_f32_rmax_ukernel__sse); |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | TEST(F32_RMAX__SSE, n_eq_16) { |
| 62 | TEST_REQUIRES_X86_SSE2; |
| 63 | RMaxMicrokernelTester() |
| 64 | .n(16) |
| 65 | .Test(xnn_f32_rmax_ukernel__sse); |
| 66 | } |
| 67 | |
| 68 | TEST(F32_RMAX__SSE, n_div_16) { |
| 69 | TEST_REQUIRES_X86_SSE2; |
| 70 | for (size_t n = 16; n < 128; n += 16) { |
| 71 | RMaxMicrokernelTester() |
| 72 | .n(n) |
| 73 | .Test(xnn_f32_rmax_ukernel__sse); |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | TEST(F32_RMAX__SSE, n_gt_16) { |
| 78 | TEST_REQUIRES_X86_SSE2; |
| 79 | for (size_t n = 16; n < 32; n++) { |
| 80 | RMaxMicrokernelTester() |
| 81 | .n(n) |
| 82 | .Test(xnn_f32_rmax_ukernel__sse); |
| 83 | } |
| 84 | } |
Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 85 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 86 | |
Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 87 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 88 | TEST(F32_RMAX__AVX, n_lt_32) { |
| 89 | TEST_REQUIRES_X86_AVX; |
| 90 | for (size_t n = 1; n < 32; n++) { |
| 91 | RMaxMicrokernelTester() |
| 92 | .n(n) |
| 93 | .Test(xnn_f32_rmax_ukernel__avx); |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | TEST(F32_RMAX__AVX, n_eq_32) { |
| 98 | TEST_REQUIRES_X86_AVX; |
| 99 | RMaxMicrokernelTester() |
| 100 | .n(32) |
| 101 | .Test(xnn_f32_rmax_ukernel__avx); |
| 102 | } |
| 103 | |
| 104 | TEST(F32_RMAX__AVX, n_div_32) { |
| 105 | TEST_REQUIRES_X86_AVX; |
| 106 | for (size_t n = 32; n < 256; n += 32) { |
| 107 | RMaxMicrokernelTester() |
| 108 | .n(n) |
| 109 | .Test(xnn_f32_rmax_ukernel__avx); |
| 110 | } |
| 111 | } |
| 112 | |
| 113 | TEST(F32_RMAX__AVX, n_gt_32) { |
| 114 | TEST_REQUIRES_X86_AVX; |
| 115 | for (size_t n = 32; n < 64; n++) { |
| 116 | RMaxMicrokernelTester() |
| 117 | .n(n) |
| 118 | .Test(xnn_f32_rmax_ukernel__avx); |
| 119 | } |
| 120 | } |
Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 121 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 122 | |
Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 123 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 124 | TEST(F32_RMAX__AVX512F, n_lt_64) { |
| 125 | TEST_REQUIRES_X86_AVX512F; |
| 126 | for (size_t n = 1; n < 64; n++) { |
| 127 | RMaxMicrokernelTester() |
| 128 | .n(n) |
| 129 | .Test(xnn_f32_rmax_ukernel__avx512f); |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | TEST(F32_RMAX__AVX512F, n_eq_64) { |
| 134 | TEST_REQUIRES_X86_AVX512F; |
| 135 | RMaxMicrokernelTester() |
| 136 | .n(64) |
| 137 | .Test(xnn_f32_rmax_ukernel__avx512f); |
| 138 | } |
| 139 | |
| 140 | TEST(F32_RMAX__AVX512F, n_div_64) { |
| 141 | TEST_REQUIRES_X86_AVX512F; |
| 142 | for (size_t n = 64; n < 512; n += 64) { |
| 143 | RMaxMicrokernelTester() |
| 144 | .n(n) |
| 145 | .Test(xnn_f32_rmax_ukernel__avx512f); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | TEST(F32_RMAX__AVX512F, n_gt_64) { |
| 150 | TEST_REQUIRES_X86_AVX512F; |
| 151 | for (size_t n = 64; n < 128; n++) { |
| 152 | RMaxMicrokernelTester() |
| 153 | .n(n) |
| 154 | .Test(xnn_f32_rmax_ukernel__avx512f); |
| 155 | } |
| 156 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 157 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 158 | |
Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 159 | #if XNN_ARCH_WASMSIMD |
| 160 | TEST(F32_RMAX__WASMSIMD_ARM, n_lt_16) { |
| 161 | for (size_t n = 1; n < 16; n++) { |
| 162 | RMaxMicrokernelTester() |
| 163 | .n(n) |
| 164 | .Test(xnn_f32_rmax_ukernel__wasmsimd_arm); |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | TEST(F32_RMAX__WASMSIMD_ARM, n_eq_16) { |
| 169 | RMaxMicrokernelTester() |
| 170 | .n(16) |
| 171 | .Test(xnn_f32_rmax_ukernel__wasmsimd_arm); |
| 172 | } |
| 173 | |
| 174 | TEST(F32_RMAX__WASMSIMD_ARM, n_div_16) { |
| 175 | for (size_t n = 16; n < 128; n += 16) { |
| 176 | RMaxMicrokernelTester() |
| 177 | .n(n) |
| 178 | .Test(xnn_f32_rmax_ukernel__wasmsimd_arm); |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | TEST(F32_RMAX__WASMSIMD_ARM, n_gt_16) { |
| 183 | for (size_t n = 16; n < 32; n++) { |
| 184 | RMaxMicrokernelTester() |
| 185 | .n(n) |
| 186 | .Test(xnn_f32_rmax_ukernel__wasmsimd_arm); |
| 187 | } |
| 188 | } |
| 189 | #endif // XNN_ARCH_WASMSIMD |
| 190 | |
| 191 | #if XNN_ARCH_WASMSIMD |
| 192 | TEST(F32_RMAX__WASMSIMD_X86, n_lt_16) { |
| 193 | for (size_t n = 1; n < 16; n++) { |
| 194 | RMaxMicrokernelTester() |
| 195 | .n(n) |
| 196 | .Test(xnn_f32_rmax_ukernel__wasmsimd_x86); |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | TEST(F32_RMAX__WASMSIMD_X86, n_eq_16) { |
| 201 | RMaxMicrokernelTester() |
| 202 | .n(16) |
| 203 | .Test(xnn_f32_rmax_ukernel__wasmsimd_x86); |
| 204 | } |
| 205 | |
| 206 | TEST(F32_RMAX__WASMSIMD_X86, n_div_16) { |
| 207 | for (size_t n = 16; n < 128; n += 16) { |
| 208 | RMaxMicrokernelTester() |
| 209 | .n(n) |
| 210 | .Test(xnn_f32_rmax_ukernel__wasmsimd_x86); |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | TEST(F32_RMAX__WASMSIMD_X86, n_gt_16) { |
| 215 | for (size_t n = 16; n < 32; n++) { |
| 216 | RMaxMicrokernelTester() |
| 217 | .n(n) |
| 218 | .Test(xnn_f32_rmax_ukernel__wasmsimd_x86); |
| 219 | } |
| 220 | } |
| 221 | #endif // XNN_ARCH_WASMSIMD |
| 222 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 223 | TEST(F32_RMAX__SCALAR, n_lt_4) { |
| 224 | for (size_t n = 1; n < 4; n++) { |
| 225 | RMaxMicrokernelTester() |
| 226 | .n(n) |
| 227 | .Test(xnn_f32_rmax_ukernel__scalar); |
| 228 | } |
| 229 | } |
| 230 | |
| 231 | TEST(F32_RMAX__SCALAR, n_eq_4) { |
| 232 | RMaxMicrokernelTester() |
| 233 | .n(4) |
| 234 | .Test(xnn_f32_rmax_ukernel__scalar); |
| 235 | } |
| 236 | |
| 237 | TEST(F32_RMAX__SCALAR, n_div_4) { |
| 238 | for (size_t n = 4; n < 32; n += 4) { |
| 239 | RMaxMicrokernelTester() |
| 240 | .n(n) |
| 241 | .Test(xnn_f32_rmax_ukernel__scalar); |
| 242 | } |
| 243 | } |
| 244 | |
| 245 | TEST(F32_RMAX__SCALAR, n_gt_4) { |
| 246 | for (size_t n = 4; n < 8; n++) { |
| 247 | RMaxMicrokernelTester() |
| 248 | .n(n) |
| 249 | .Test(xnn_f32_rmax_ukernel__scalar); |
| 250 | } |
| 251 | } |