Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -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/f16-vsqrdiff.yaml |
| 8 | // Generator: tools/generate-vbinary-test.py |
| 9 | |
| 10 | |
| 11 | #include <gtest/gtest.h> |
| 12 | |
| 13 | #include <xnnpack/common.h> |
| 14 | #include <xnnpack/isa-checks.h> |
| 15 | |
Marat Dukhan | 87bd511 | 2021-08-02 11:43:53 -0700 | [diff] [blame] | 16 | #include <xnnpack/params-init.h> |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 17 | #include <xnnpack/vbinary.h> |
| 18 | #include "vbinary-microkernel-tester.h" |
| 19 | |
| 20 | |
| 21 | #if XNN_ARCH_ARM64 |
| 22 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X8, batch_eq_8) { |
| 23 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 24 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 25 | .batch_size(8) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 26 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 27 | } |
| 28 | |
| 29 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X8, batch_div_8) { |
| 30 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 31 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 32 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 33 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 34 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 35 | } |
| 36 | } |
| 37 | |
| 38 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X8, batch_lt_8) { |
| 39 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 40 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 41 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 42 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 43 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 44 | } |
| 45 | } |
| 46 | |
| 47 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X8, batch_gt_8) { |
| 48 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 49 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 50 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 51 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 52 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 53 | } |
| 54 | } |
| 55 | |
| 56 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X8, inplace_a) { |
| 57 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 58 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 59 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 60 | .batch_size(batch_size) |
| 61 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 62 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 63 | } |
| 64 | } |
| 65 | |
| 66 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X8, inplace_b) { |
| 67 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 68 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 69 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 70 | .batch_size(batch_size) |
| 71 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 72 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 73 | } |
| 74 | } |
| 75 | |
| 76 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X8, inplace_a_and_b) { |
| 77 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 78 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 79 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 80 | .batch_size(batch_size) |
| 81 | .inplace_a(true) |
| 82 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 83 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 84 | } |
| 85 | } |
| 86 | #endif // XNN_ARCH_ARM64 |
| 87 | |
| 88 | |
| 89 | #if XNN_ARCH_ARM64 |
| 90 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X16, batch_eq_16) { |
| 91 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 92 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 93 | .batch_size(16) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 94 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 95 | } |
| 96 | |
| 97 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X16, batch_div_16) { |
| 98 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 99 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 100 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 101 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 102 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 103 | } |
| 104 | } |
| 105 | |
| 106 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X16, batch_lt_16) { |
| 107 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 108 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 109 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 110 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 111 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 112 | } |
| 113 | } |
| 114 | |
| 115 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X16, batch_gt_16) { |
| 116 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 117 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 118 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 119 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 120 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 121 | } |
| 122 | } |
| 123 | |
| 124 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X16, inplace_a) { |
| 125 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 126 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 127 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 128 | .batch_size(batch_size) |
| 129 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 130 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 131 | } |
| 132 | } |
| 133 | |
| 134 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X16, inplace_b) { |
| 135 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 136 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 137 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 138 | .batch_size(batch_size) |
| 139 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 140 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 141 | } |
| 142 | } |
| 143 | |
| 144 | TEST(F16_VSQRDIFF__NEONFP16ARITH_X16, inplace_a_and_b) { |
| 145 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 146 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 147 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 148 | .batch_size(batch_size) |
| 149 | .inplace_a(true) |
| 150 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 151 | .Test(xnn_f16_vsqrdiff_ukernel__neonfp16arith_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 152 | } |
| 153 | } |
| 154 | #endif // XNN_ARCH_ARM64 |
Marat Dukhan | d454545 | 2022-01-10 16:13:11 -0800 | [diff] [blame] | 155 | |
| 156 | |
| 157 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 158 | TEST(F16_VSQRDIFF__F16C_X8, batch_eq_8) { |
| 159 | TEST_REQUIRES_X86_F16C; |
| 160 | VBinaryMicrokernelTester() |
| 161 | .batch_size(8) |
| 162 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 163 | } |
| 164 | |
| 165 | TEST(F16_VSQRDIFF__F16C_X8, batch_div_8) { |
| 166 | TEST_REQUIRES_X86_F16C; |
| 167 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
| 168 | VBinaryMicrokernelTester() |
| 169 | .batch_size(batch_size) |
| 170 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | TEST(F16_VSQRDIFF__F16C_X8, batch_lt_8) { |
| 175 | TEST_REQUIRES_X86_F16C; |
| 176 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
| 177 | VBinaryMicrokernelTester() |
| 178 | .batch_size(batch_size) |
| 179 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | TEST(F16_VSQRDIFF__F16C_X8, batch_gt_8) { |
| 184 | TEST_REQUIRES_X86_F16C; |
| 185 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
| 186 | VBinaryMicrokernelTester() |
| 187 | .batch_size(batch_size) |
| 188 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 189 | } |
| 190 | } |
| 191 | |
| 192 | TEST(F16_VSQRDIFF__F16C_X8, inplace_a) { |
| 193 | TEST_REQUIRES_X86_F16C; |
| 194 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 195 | VBinaryMicrokernelTester() |
| 196 | .batch_size(batch_size) |
| 197 | .inplace_a(true) |
| 198 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 199 | } |
| 200 | } |
| 201 | |
| 202 | TEST(F16_VSQRDIFF__F16C_X8, inplace_b) { |
| 203 | TEST_REQUIRES_X86_F16C; |
| 204 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 205 | VBinaryMicrokernelTester() |
| 206 | .batch_size(batch_size) |
| 207 | .inplace_b(true) |
| 208 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | TEST(F16_VSQRDIFF__F16C_X8, inplace_a_and_b) { |
| 213 | TEST_REQUIRES_X86_F16C; |
| 214 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
| 215 | VBinaryMicrokernelTester() |
| 216 | .batch_size(batch_size) |
| 217 | .inplace_a(true) |
| 218 | .inplace_b(true) |
| 219 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 220 | } |
| 221 | } |
| 222 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 223 | |
| 224 | |
| 225 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 226 | TEST(F16_VSQRDIFF__F16C_X16, batch_eq_16) { |
| 227 | TEST_REQUIRES_X86_F16C; |
| 228 | VBinaryMicrokernelTester() |
| 229 | .batch_size(16) |
| 230 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 231 | } |
| 232 | |
| 233 | TEST(F16_VSQRDIFF__F16C_X16, batch_div_16) { |
| 234 | TEST_REQUIRES_X86_F16C; |
| 235 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
| 236 | VBinaryMicrokernelTester() |
| 237 | .batch_size(batch_size) |
| 238 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | TEST(F16_VSQRDIFF__F16C_X16, batch_lt_16) { |
| 243 | TEST_REQUIRES_X86_F16C; |
| 244 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
| 245 | VBinaryMicrokernelTester() |
| 246 | .batch_size(batch_size) |
| 247 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | TEST(F16_VSQRDIFF__F16C_X16, batch_gt_16) { |
| 252 | TEST_REQUIRES_X86_F16C; |
| 253 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
| 254 | VBinaryMicrokernelTester() |
| 255 | .batch_size(batch_size) |
| 256 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 257 | } |
| 258 | } |
| 259 | |
| 260 | TEST(F16_VSQRDIFF__F16C_X16, inplace_a) { |
| 261 | TEST_REQUIRES_X86_F16C; |
| 262 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 263 | VBinaryMicrokernelTester() |
| 264 | .batch_size(batch_size) |
| 265 | .inplace_a(true) |
| 266 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 267 | } |
| 268 | } |
| 269 | |
| 270 | TEST(F16_VSQRDIFF__F16C_X16, inplace_b) { |
| 271 | TEST_REQUIRES_X86_F16C; |
| 272 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 273 | VBinaryMicrokernelTester() |
| 274 | .batch_size(batch_size) |
| 275 | .inplace_b(true) |
| 276 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | TEST(F16_VSQRDIFF__F16C_X16, inplace_a_and_b) { |
| 281 | TEST_REQUIRES_X86_F16C; |
| 282 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
| 283 | VBinaryMicrokernelTester() |
| 284 | .batch_size(batch_size) |
| 285 | .inplace_a(true) |
| 286 | .inplace_b(true) |
| 287 | .Test(xnn_f16_vsqrdiff_ukernel__f16c_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
| 288 | } |
| 289 | } |
| 290 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |