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/f32-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 | |
| 16 | #include <xnnpack/vbinary.h> |
| 17 | #include "vbinary-microkernel-tester.h" |
| 18 | |
| 19 | |
| 20 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 21 | TEST(F32_VSQRDIFF__NEON_X4, batch_eq_4) { |
| 22 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 23 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 24 | .batch_size(4) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 25 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 26 | } |
| 27 | |
| 28 | TEST(F32_VSQRDIFF__NEON_X4, batch_div_4) { |
| 29 | TEST_REQUIRES_ARM_NEON; |
| 30 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 31 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 32 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 33 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 34 | } |
| 35 | } |
| 36 | |
| 37 | TEST(F32_VSQRDIFF__NEON_X4, batch_lt_4) { |
| 38 | TEST_REQUIRES_ARM_NEON; |
| 39 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 40 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 41 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 42 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 43 | } |
| 44 | } |
| 45 | |
| 46 | TEST(F32_VSQRDIFF__NEON_X4, batch_gt_4) { |
| 47 | TEST_REQUIRES_ARM_NEON; |
| 48 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 49 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 50 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 51 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 52 | } |
| 53 | } |
| 54 | |
| 55 | TEST(F32_VSQRDIFF__NEON_X4, inplace_a) { |
| 56 | TEST_REQUIRES_ARM_NEON; |
| 57 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 58 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 59 | .batch_size(batch_size) |
| 60 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 61 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 62 | } |
| 63 | } |
| 64 | |
| 65 | TEST(F32_VSQRDIFF__NEON_X4, inplace_b) { |
| 66 | TEST_REQUIRES_ARM_NEON; |
| 67 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 68 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 69 | .batch_size(batch_size) |
| 70 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 71 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 72 | } |
| 73 | } |
| 74 | |
| 75 | TEST(F32_VSQRDIFF__NEON_X4, inplace_a_and_b) { |
| 76 | TEST_REQUIRES_ARM_NEON; |
| 77 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 78 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 79 | .batch_size(batch_size) |
| 80 | .inplace_a(true) |
| 81 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 82 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 83 | } |
| 84 | } |
| 85 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 86 | |
| 87 | |
| 88 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 89 | TEST(F32_VSQRDIFF__NEON_X8, batch_eq_8) { |
| 90 | TEST_REQUIRES_ARM_NEON; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 91 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 92 | .batch_size(8) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 93 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 94 | } |
| 95 | |
| 96 | TEST(F32_VSQRDIFF__NEON_X8, batch_div_8) { |
| 97 | TEST_REQUIRES_ARM_NEON; |
| 98 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 99 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 100 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 101 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 102 | } |
| 103 | } |
| 104 | |
| 105 | TEST(F32_VSQRDIFF__NEON_X8, batch_lt_8) { |
| 106 | TEST_REQUIRES_ARM_NEON; |
| 107 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 108 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 109 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 110 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 111 | } |
| 112 | } |
| 113 | |
| 114 | TEST(F32_VSQRDIFF__NEON_X8, batch_gt_8) { |
| 115 | TEST_REQUIRES_ARM_NEON; |
| 116 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 117 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 118 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 119 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 120 | } |
| 121 | } |
| 122 | |
| 123 | TEST(F32_VSQRDIFF__NEON_X8, inplace_a) { |
| 124 | TEST_REQUIRES_ARM_NEON; |
| 125 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 126 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 127 | .batch_size(batch_size) |
| 128 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 129 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 130 | } |
| 131 | } |
| 132 | |
| 133 | TEST(F32_VSQRDIFF__NEON_X8, inplace_b) { |
| 134 | TEST_REQUIRES_ARM_NEON; |
| 135 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 136 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 137 | .batch_size(batch_size) |
| 138 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 139 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 140 | } |
| 141 | } |
| 142 | |
| 143 | TEST(F32_VSQRDIFF__NEON_X8, inplace_a_and_b) { |
| 144 | TEST_REQUIRES_ARM_NEON; |
| 145 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 146 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 147 | .batch_size(batch_size) |
| 148 | .inplace_a(true) |
| 149 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 150 | .Test(xnn_f32_vsqrdiff_ukernel__neon_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 151 | } |
| 152 | } |
| 153 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 154 | |
| 155 | |
| 156 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 157 | TEST(F32_VSQRDIFF__SSE_X4, batch_eq_4) { |
| 158 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 159 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 160 | .batch_size(4) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 161 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 162 | } |
| 163 | |
| 164 | TEST(F32_VSQRDIFF__SSE_X4, batch_div_4) { |
| 165 | TEST_REQUIRES_X86_SSE; |
| 166 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 167 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 168 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 169 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 170 | } |
| 171 | } |
| 172 | |
| 173 | TEST(F32_VSQRDIFF__SSE_X4, batch_lt_4) { |
| 174 | TEST_REQUIRES_X86_SSE; |
| 175 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 176 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 177 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 178 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 179 | } |
| 180 | } |
| 181 | |
| 182 | TEST(F32_VSQRDIFF__SSE_X4, batch_gt_4) { |
| 183 | TEST_REQUIRES_X86_SSE; |
| 184 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 185 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 186 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 187 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 188 | } |
| 189 | } |
| 190 | |
| 191 | TEST(F32_VSQRDIFF__SSE_X4, inplace_a) { |
| 192 | TEST_REQUIRES_X86_SSE; |
| 193 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 194 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 195 | .batch_size(batch_size) |
| 196 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 197 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 198 | } |
| 199 | } |
| 200 | |
| 201 | TEST(F32_VSQRDIFF__SSE_X4, inplace_b) { |
| 202 | TEST_REQUIRES_X86_SSE; |
| 203 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 204 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 205 | .batch_size(batch_size) |
| 206 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 207 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 208 | } |
| 209 | } |
| 210 | |
| 211 | TEST(F32_VSQRDIFF__SSE_X4, inplace_a_and_b) { |
| 212 | TEST_REQUIRES_X86_SSE; |
| 213 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 214 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 215 | .batch_size(batch_size) |
| 216 | .inplace_a(true) |
| 217 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 218 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 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_VSQRDIFF__SSE_X8, batch_eq_8) { |
| 226 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 227 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 228 | .batch_size(8) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 229 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 230 | } |
| 231 | |
| 232 | TEST(F32_VSQRDIFF__SSE_X8, batch_div_8) { |
| 233 | TEST_REQUIRES_X86_SSE; |
| 234 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 235 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 236 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 237 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 238 | } |
| 239 | } |
| 240 | |
| 241 | TEST(F32_VSQRDIFF__SSE_X8, batch_lt_8) { |
| 242 | TEST_REQUIRES_X86_SSE; |
| 243 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 244 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 245 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 246 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 247 | } |
| 248 | } |
| 249 | |
| 250 | TEST(F32_VSQRDIFF__SSE_X8, batch_gt_8) { |
| 251 | TEST_REQUIRES_X86_SSE; |
| 252 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 253 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 254 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 255 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 256 | } |
| 257 | } |
| 258 | |
| 259 | TEST(F32_VSQRDIFF__SSE_X8, inplace_a) { |
| 260 | TEST_REQUIRES_X86_SSE; |
| 261 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 262 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 263 | .batch_size(batch_size) |
| 264 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 265 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 266 | } |
| 267 | } |
| 268 | |
| 269 | TEST(F32_VSQRDIFF__SSE_X8, inplace_b) { |
| 270 | TEST_REQUIRES_X86_SSE; |
| 271 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 272 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 273 | .batch_size(batch_size) |
| 274 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 275 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 276 | } |
| 277 | } |
| 278 | |
| 279 | TEST(F32_VSQRDIFF__SSE_X8, inplace_a_and_b) { |
| 280 | TEST_REQUIRES_X86_SSE; |
| 281 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 282 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 283 | .batch_size(batch_size) |
| 284 | .inplace_a(true) |
| 285 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 286 | .Test(xnn_f32_vsqrdiff_ukernel__sse_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 287 | } |
| 288 | } |
| 289 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 290 | |
| 291 | |
| 292 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 293 | TEST(F32_VSQRDIFF__AVX_X8, batch_eq_8) { |
| 294 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 295 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 296 | .batch_size(8) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 297 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 298 | } |
| 299 | |
| 300 | TEST(F32_VSQRDIFF__AVX_X8, batch_div_8) { |
| 301 | TEST_REQUIRES_X86_AVX; |
| 302 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 303 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 304 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 305 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 306 | } |
| 307 | } |
| 308 | |
| 309 | TEST(F32_VSQRDIFF__AVX_X8, batch_lt_8) { |
| 310 | TEST_REQUIRES_X86_AVX; |
| 311 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 312 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 313 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 314 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 315 | } |
| 316 | } |
| 317 | |
| 318 | TEST(F32_VSQRDIFF__AVX_X8, batch_gt_8) { |
| 319 | TEST_REQUIRES_X86_AVX; |
| 320 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 321 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 322 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 323 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 324 | } |
| 325 | } |
| 326 | |
| 327 | TEST(F32_VSQRDIFF__AVX_X8, inplace_a) { |
| 328 | TEST_REQUIRES_X86_AVX; |
| 329 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 330 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 331 | .batch_size(batch_size) |
| 332 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 333 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 334 | } |
| 335 | } |
| 336 | |
| 337 | TEST(F32_VSQRDIFF__AVX_X8, inplace_b) { |
| 338 | TEST_REQUIRES_X86_AVX; |
| 339 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 340 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 341 | .batch_size(batch_size) |
| 342 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 343 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 344 | } |
| 345 | } |
| 346 | |
| 347 | TEST(F32_VSQRDIFF__AVX_X8, inplace_a_and_b) { |
| 348 | TEST_REQUIRES_X86_AVX; |
| 349 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 350 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 351 | .batch_size(batch_size) |
| 352 | .inplace_a(true) |
| 353 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 354 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 355 | } |
| 356 | } |
| 357 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 358 | |
| 359 | |
| 360 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 361 | TEST(F32_VSQRDIFF__AVX_X16, batch_eq_16) { |
| 362 | TEST_REQUIRES_X86_AVX; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 363 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 364 | .batch_size(16) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 365 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 366 | } |
| 367 | |
| 368 | TEST(F32_VSQRDIFF__AVX_X16, batch_div_16) { |
| 369 | TEST_REQUIRES_X86_AVX; |
| 370 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 371 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 372 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 373 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 374 | } |
| 375 | } |
| 376 | |
| 377 | TEST(F32_VSQRDIFF__AVX_X16, batch_lt_16) { |
| 378 | TEST_REQUIRES_X86_AVX; |
| 379 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 380 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 381 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 382 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 383 | } |
| 384 | } |
| 385 | |
| 386 | TEST(F32_VSQRDIFF__AVX_X16, batch_gt_16) { |
| 387 | TEST_REQUIRES_X86_AVX; |
| 388 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 389 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 390 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 391 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 392 | } |
| 393 | } |
| 394 | |
| 395 | TEST(F32_VSQRDIFF__AVX_X16, inplace_a) { |
| 396 | TEST_REQUIRES_X86_AVX; |
| 397 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 398 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 399 | .batch_size(batch_size) |
| 400 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 401 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 402 | } |
| 403 | } |
| 404 | |
| 405 | TEST(F32_VSQRDIFF__AVX_X16, inplace_b) { |
| 406 | TEST_REQUIRES_X86_AVX; |
| 407 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 408 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 409 | .batch_size(batch_size) |
| 410 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 411 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 412 | } |
| 413 | } |
| 414 | |
| 415 | TEST(F32_VSQRDIFF__AVX_X16, inplace_a_and_b) { |
| 416 | TEST_REQUIRES_X86_AVX; |
| 417 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 418 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 419 | .batch_size(batch_size) |
| 420 | .inplace_a(true) |
| 421 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 422 | .Test(xnn_f32_vsqrdiff_ukernel__avx_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 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_VSQRDIFF__AVX512F_X16, batch_eq_16) { |
| 430 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 431 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 432 | .batch_size(16) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 433 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 434 | } |
| 435 | |
| 436 | TEST(F32_VSQRDIFF__AVX512F_X16, batch_div_16) { |
| 437 | TEST_REQUIRES_X86_AVX512F; |
| 438 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 439 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 440 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 441 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 442 | } |
| 443 | } |
| 444 | |
| 445 | TEST(F32_VSQRDIFF__AVX512F_X16, batch_lt_16) { |
| 446 | TEST_REQUIRES_X86_AVX512F; |
| 447 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 448 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 449 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 450 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 451 | } |
| 452 | } |
| 453 | |
| 454 | TEST(F32_VSQRDIFF__AVX512F_X16, batch_gt_16) { |
| 455 | TEST_REQUIRES_X86_AVX512F; |
| 456 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 457 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 458 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 459 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 460 | } |
| 461 | } |
| 462 | |
| 463 | TEST(F32_VSQRDIFF__AVX512F_X16, inplace_a) { |
| 464 | TEST_REQUIRES_X86_AVX512F; |
| 465 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 466 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 467 | .batch_size(batch_size) |
| 468 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 469 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 470 | } |
| 471 | } |
| 472 | |
| 473 | TEST(F32_VSQRDIFF__AVX512F_X16, inplace_b) { |
| 474 | TEST_REQUIRES_X86_AVX512F; |
| 475 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 476 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 477 | .batch_size(batch_size) |
| 478 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 479 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 480 | } |
| 481 | } |
| 482 | |
| 483 | TEST(F32_VSQRDIFF__AVX512F_X16, inplace_a_and_b) { |
| 484 | TEST_REQUIRES_X86_AVX512F; |
| 485 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 486 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 487 | .batch_size(batch_size) |
| 488 | .inplace_a(true) |
| 489 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 490 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 491 | } |
| 492 | } |
| 493 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 494 | |
| 495 | |
| 496 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 497 | TEST(F32_VSQRDIFF__AVX512F_X32, batch_eq_32) { |
| 498 | TEST_REQUIRES_X86_AVX512F; |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 499 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 500 | .batch_size(32) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 501 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x32, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 502 | } |
| 503 | |
| 504 | TEST(F32_VSQRDIFF__AVX512F_X32, batch_div_32) { |
| 505 | TEST_REQUIRES_X86_AVX512F; |
| 506 | for (size_t batch_size = 64; batch_size < 320; batch_size += 32) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 507 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 508 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 509 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x32, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 510 | } |
| 511 | } |
| 512 | |
| 513 | TEST(F32_VSQRDIFF__AVX512F_X32, batch_lt_32) { |
| 514 | TEST_REQUIRES_X86_AVX512F; |
| 515 | for (size_t batch_size = 1; batch_size < 32; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 516 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 517 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 518 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x32, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 519 | } |
| 520 | } |
| 521 | |
| 522 | TEST(F32_VSQRDIFF__AVX512F_X32, batch_gt_32) { |
| 523 | TEST_REQUIRES_X86_AVX512F; |
| 524 | for (size_t batch_size = 33; batch_size < 64; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 525 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 526 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 527 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x32, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 528 | } |
| 529 | } |
| 530 | |
| 531 | TEST(F32_VSQRDIFF__AVX512F_X32, inplace_a) { |
| 532 | TEST_REQUIRES_X86_AVX512F; |
| 533 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 534 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 535 | .batch_size(batch_size) |
| 536 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 537 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x32, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 538 | } |
| 539 | } |
| 540 | |
| 541 | TEST(F32_VSQRDIFF__AVX512F_X32, inplace_b) { |
| 542 | TEST_REQUIRES_X86_AVX512F; |
| 543 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 544 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 545 | .batch_size(batch_size) |
| 546 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 547 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x32, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 548 | } |
| 549 | } |
| 550 | |
| 551 | TEST(F32_VSQRDIFF__AVX512F_X32, inplace_a_and_b) { |
| 552 | TEST_REQUIRES_X86_AVX512F; |
| 553 | for (size_t batch_size = 1; batch_size <= 160; batch_size += 31) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 554 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 555 | .batch_size(batch_size) |
| 556 | .inplace_a(true) |
| 557 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 558 | .Test(xnn_f32_vsqrdiff_ukernel__avx512f_x32, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 559 | } |
| 560 | } |
| 561 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| 562 | |
| 563 | |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 564 | #if XNN_ARCH_WASMSIMD |
| 565 | TEST(F32_VSQRDIFF__WASMSIMD_X4, batch_eq_4) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 566 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 567 | .batch_size(4) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 568 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 569 | } |
| 570 | |
| 571 | TEST(F32_VSQRDIFF__WASMSIMD_X4, batch_div_4) { |
| 572 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 573 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 574 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 575 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 576 | } |
| 577 | } |
| 578 | |
| 579 | TEST(F32_VSQRDIFF__WASMSIMD_X4, batch_lt_4) { |
| 580 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 581 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 582 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 583 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 584 | } |
| 585 | } |
| 586 | |
| 587 | TEST(F32_VSQRDIFF__WASMSIMD_X4, batch_gt_4) { |
| 588 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 589 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 590 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 591 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 592 | } |
| 593 | } |
| 594 | |
| 595 | TEST(F32_VSQRDIFF__WASMSIMD_X4, inplace_a) { |
| 596 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 597 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 598 | .batch_size(batch_size) |
| 599 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 600 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 601 | } |
| 602 | } |
| 603 | |
| 604 | TEST(F32_VSQRDIFF__WASMSIMD_X4, inplace_b) { |
| 605 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 606 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 607 | .batch_size(batch_size) |
| 608 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 609 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 610 | } |
| 611 | } |
| 612 | |
| 613 | TEST(F32_VSQRDIFF__WASMSIMD_X4, inplace_a_and_b) { |
| 614 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 615 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 616 | .batch_size(batch_size) |
| 617 | .inplace_a(true) |
| 618 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 619 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x4, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 620 | } |
| 621 | } |
| 622 | #endif // XNN_ARCH_WASMSIMD |
| 623 | |
| 624 | |
| 625 | #if XNN_ARCH_WASMSIMD |
| 626 | TEST(F32_VSQRDIFF__WASMSIMD_X8, batch_eq_8) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 627 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 628 | .batch_size(8) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 629 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 630 | } |
| 631 | |
| 632 | TEST(F32_VSQRDIFF__WASMSIMD_X8, batch_div_8) { |
| 633 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 634 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 635 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 636 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 637 | } |
| 638 | } |
| 639 | |
| 640 | TEST(F32_VSQRDIFF__WASMSIMD_X8, batch_lt_8) { |
| 641 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 642 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 643 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 644 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 645 | } |
| 646 | } |
| 647 | |
| 648 | TEST(F32_VSQRDIFF__WASMSIMD_X8, batch_gt_8) { |
| 649 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 650 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 651 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 652 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 653 | } |
| 654 | } |
| 655 | |
| 656 | TEST(F32_VSQRDIFF__WASMSIMD_X8, inplace_a) { |
| 657 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 658 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 659 | .batch_size(batch_size) |
| 660 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 661 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 662 | } |
| 663 | } |
| 664 | |
| 665 | TEST(F32_VSQRDIFF__WASMSIMD_X8, inplace_b) { |
| 666 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 667 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 668 | .batch_size(batch_size) |
| 669 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 670 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 671 | } |
| 672 | } |
| 673 | |
| 674 | TEST(F32_VSQRDIFF__WASMSIMD_X8, inplace_a_and_b) { |
| 675 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 676 | VBinaryMicrokernelTester() |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 677 | .batch_size(batch_size) |
| 678 | .inplace_a(true) |
| 679 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 680 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x8, VBinaryMicrokernelTester::OpType::SqrDiff); |
Marat Dukhan | 93d1ba1 | 2020-06-26 12:33:35 -0700 | [diff] [blame] | 681 | } |
| 682 | } |
| 683 | #endif // XNN_ARCH_WASMSIMD |
| 684 | |
| 685 | |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 686 | #if XNN_ARCH_WASMSIMD |
| 687 | TEST(F32_VSQRDIFF__WASMSIMD_X16, batch_eq_16) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 688 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 689 | .batch_size(16) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 690 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 691 | } |
| 692 | |
| 693 | TEST(F32_VSQRDIFF__WASMSIMD_X16, batch_div_16) { |
| 694 | for (size_t batch_size = 32; batch_size < 160; batch_size += 16) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 695 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 696 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 697 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 698 | } |
| 699 | } |
| 700 | |
| 701 | TEST(F32_VSQRDIFF__WASMSIMD_X16, batch_lt_16) { |
| 702 | for (size_t batch_size = 1; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 703 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 704 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 705 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 706 | } |
| 707 | } |
| 708 | |
| 709 | TEST(F32_VSQRDIFF__WASMSIMD_X16, batch_gt_16) { |
| 710 | for (size_t batch_size = 17; batch_size < 32; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 711 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 712 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 713 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 714 | } |
| 715 | } |
| 716 | |
| 717 | TEST(F32_VSQRDIFF__WASMSIMD_X16, inplace_a) { |
| 718 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 719 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 720 | .batch_size(batch_size) |
| 721 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 722 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 723 | } |
| 724 | } |
| 725 | |
| 726 | TEST(F32_VSQRDIFF__WASMSIMD_X16, inplace_b) { |
| 727 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 728 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 729 | .batch_size(batch_size) |
| 730 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 731 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 732 | } |
| 733 | } |
| 734 | |
| 735 | TEST(F32_VSQRDIFF__WASMSIMD_X16, inplace_a_and_b) { |
| 736 | for (size_t batch_size = 1; batch_size <= 80; batch_size += 15) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 737 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 738 | .batch_size(batch_size) |
| 739 | .inplace_a(true) |
| 740 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 741 | .Test(xnn_f32_vsqrdiff_ukernel__wasmsimd_x16, VBinaryMicrokernelTester::OpType::SqrDiff); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 742 | } |
| 743 | } |
| 744 | #endif // XNN_ARCH_WASMSIMD |
| 745 | |
| 746 | |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 747 | TEST(F32_VSQRDIFF__SCALAR_X1, batch_eq_1) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 748 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 749 | .batch_size(1) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 750 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x1, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 751 | } |
| 752 | |
| 753 | TEST(F32_VSQRDIFF__SCALAR_X1, batch_gt_1) { |
| 754 | for (size_t batch_size = 2; batch_size < 10; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 755 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 756 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 757 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x1, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 758 | } |
| 759 | } |
| 760 | |
| 761 | TEST(F32_VSQRDIFF__SCALAR_X1, inplace_a) { |
| 762 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 763 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 764 | .batch_size(batch_size) |
| 765 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 766 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x1, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 767 | } |
| 768 | } |
| 769 | |
| 770 | TEST(F32_VSQRDIFF__SCALAR_X1, inplace_b) { |
| 771 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 772 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 773 | .batch_size(batch_size) |
| 774 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 775 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x1, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 776 | } |
| 777 | } |
| 778 | |
| 779 | TEST(F32_VSQRDIFF__SCALAR_X1, inplace_a_and_b) { |
| 780 | for (size_t batch_size = 1; batch_size <= 5; batch_size += 1) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 781 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 782 | .batch_size(batch_size) |
| 783 | .inplace_a(true) |
| 784 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 785 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x1, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 786 | } |
| 787 | } |
| 788 | |
| 789 | |
| 790 | TEST(F32_VSQRDIFF__SCALAR_X2, batch_eq_2) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 791 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 792 | .batch_size(2) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 793 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x2, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 794 | } |
| 795 | |
| 796 | TEST(F32_VSQRDIFF__SCALAR_X2, batch_div_2) { |
| 797 | for (size_t batch_size = 4; batch_size < 20; batch_size += 2) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 798 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 799 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 800 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x2, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 801 | } |
| 802 | } |
| 803 | |
| 804 | TEST(F32_VSQRDIFF__SCALAR_X2, batch_lt_2) { |
| 805 | for (size_t batch_size = 1; batch_size < 2; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 806 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 807 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 808 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x2, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 809 | } |
| 810 | } |
| 811 | |
| 812 | TEST(F32_VSQRDIFF__SCALAR_X2, batch_gt_2) { |
| 813 | for (size_t batch_size = 3; batch_size < 4; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 814 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 815 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 816 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x2, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 817 | } |
| 818 | } |
| 819 | |
| 820 | TEST(F32_VSQRDIFF__SCALAR_X2, inplace_a) { |
| 821 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 822 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 823 | .batch_size(batch_size) |
| 824 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 825 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x2, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 826 | } |
| 827 | } |
| 828 | |
| 829 | TEST(F32_VSQRDIFF__SCALAR_X2, inplace_b) { |
| 830 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 831 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 832 | .batch_size(batch_size) |
| 833 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 834 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x2, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 835 | } |
| 836 | } |
| 837 | |
| 838 | TEST(F32_VSQRDIFF__SCALAR_X2, inplace_a_and_b) { |
| 839 | for (size_t batch_size = 1; batch_size <= 10; batch_size += 1) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 840 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 841 | .batch_size(batch_size) |
| 842 | .inplace_a(true) |
| 843 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 844 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x2, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 845 | } |
| 846 | } |
| 847 | |
| 848 | |
| 849 | TEST(F32_VSQRDIFF__SCALAR_X4, batch_eq_4) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 850 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 851 | .batch_size(4) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 852 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x4, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 853 | } |
| 854 | |
| 855 | TEST(F32_VSQRDIFF__SCALAR_X4, batch_div_4) { |
| 856 | for (size_t batch_size = 8; batch_size < 40; batch_size += 4) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 857 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 858 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 859 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x4, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 860 | } |
| 861 | } |
| 862 | |
| 863 | TEST(F32_VSQRDIFF__SCALAR_X4, batch_lt_4) { |
| 864 | for (size_t batch_size = 1; batch_size < 4; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 865 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 866 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 867 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x4, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 868 | } |
| 869 | } |
| 870 | |
| 871 | TEST(F32_VSQRDIFF__SCALAR_X4, batch_gt_4) { |
| 872 | for (size_t batch_size = 5; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 873 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 874 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 875 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x4, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 876 | } |
| 877 | } |
| 878 | |
| 879 | TEST(F32_VSQRDIFF__SCALAR_X4, inplace_a) { |
| 880 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 881 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 882 | .batch_size(batch_size) |
| 883 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 884 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x4, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 885 | } |
| 886 | } |
| 887 | |
| 888 | TEST(F32_VSQRDIFF__SCALAR_X4, inplace_b) { |
| 889 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 890 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 891 | .batch_size(batch_size) |
| 892 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 893 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x4, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 894 | } |
| 895 | } |
| 896 | |
| 897 | TEST(F32_VSQRDIFF__SCALAR_X4, inplace_a_and_b) { |
| 898 | for (size_t batch_size = 1; batch_size <= 20; batch_size += 3) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 899 | VBinaryMicrokernelTester() |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 900 | .batch_size(batch_size) |
| 901 | .inplace_a(true) |
| 902 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 903 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x4, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 904 | } |
| 905 | } |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 906 | |
| 907 | |
| 908 | TEST(F32_VSQRDIFF__SCALAR_X8, batch_eq_8) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 909 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 910 | .batch_size(8) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 911 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x8, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 912 | } |
| 913 | |
| 914 | TEST(F32_VSQRDIFF__SCALAR_X8, batch_div_8) { |
| 915 | for (size_t batch_size = 16; batch_size < 80; batch_size += 8) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 916 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 917 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 918 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x8, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 919 | } |
| 920 | } |
| 921 | |
| 922 | TEST(F32_VSQRDIFF__SCALAR_X8, batch_lt_8) { |
| 923 | for (size_t batch_size = 1; batch_size < 8; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 924 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 925 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 926 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x8, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 927 | } |
| 928 | } |
| 929 | |
| 930 | TEST(F32_VSQRDIFF__SCALAR_X8, batch_gt_8) { |
| 931 | for (size_t batch_size = 9; batch_size < 16; batch_size++) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 932 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 933 | .batch_size(batch_size) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 934 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x8, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 935 | } |
| 936 | } |
| 937 | |
| 938 | TEST(F32_VSQRDIFF__SCALAR_X8, inplace_a) { |
| 939 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 940 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 941 | .batch_size(batch_size) |
| 942 | .inplace_a(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 943 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x8, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 944 | } |
| 945 | } |
| 946 | |
| 947 | TEST(F32_VSQRDIFF__SCALAR_X8, inplace_b) { |
| 948 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 949 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 950 | .batch_size(batch_size) |
| 951 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 952 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x8, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 953 | } |
| 954 | } |
| 955 | |
| 956 | TEST(F32_VSQRDIFF__SCALAR_X8, inplace_a_and_b) { |
| 957 | for (size_t batch_size = 1; batch_size <= 40; batch_size += 7) { |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 958 | VBinaryMicrokernelTester() |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 959 | .batch_size(batch_size) |
| 960 | .inplace_a(true) |
| 961 | .inplace_b(true) |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 962 | .Test(xnn_f32_vsqrdiff_ukernel__scalar_x8, VBinaryMicrokernelTester::OpType::SqrDiff, VBinaryMicrokernelTester::Variant::Scalar); |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 963 | } |
| 964 | } |