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> |
| 10 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 11 | #include <xnnpack/dwconv.h> |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 12 | #include "dwconv-chw-microkernel-tester.h" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 13 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 14 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 15 | TEST(F32_DWCONV_CHW_3X3P1__SSE, input_width_eq_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 16 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 17 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 18 | .input_tuple_size(4) |
| 19 | .output_tuple_size(4) |
| 20 | .input_width(4) |
| 21 | .padding_left(1) |
| 22 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 23 | .padding_top(1) |
| 24 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 25 | .kernel_height(3) |
| 26 | .kernel_width(3) |
| 27 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 28 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 29 | } |
| 30 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 31 | TEST(F32_DWCONV_CHW_3X3P1__SSE, input_width_lt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 32 | TEST_REQUIRES_X86_SSE; |
| 33 | for (size_t input_width = 1; input_width < 4; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 34 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 35 | .input_tuple_size(4) |
| 36 | .output_tuple_size(4) |
| 37 | .input_width(input_width) |
| 38 | .padding_left(1) |
| 39 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 40 | .padding_top(1) |
| 41 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 42 | .kernel_height(3) |
| 43 | .kernel_width(3) |
| 44 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 45 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 46 | } |
| 47 | } |
| 48 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 49 | TEST(F32_DWCONV_CHW_3X3P1__SSE, input_width_gt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 50 | TEST_REQUIRES_X86_SSE; |
| 51 | for (size_t input_width = 5; input_width < 8; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 52 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 53 | .input_tuple_size(4) |
| 54 | .output_tuple_size(4) |
| 55 | .input_width(input_width) |
| 56 | .padding_left(1) |
| 57 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 58 | .padding_top(1) |
| 59 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 60 | .kernel_height(3) |
| 61 | .kernel_width(3) |
| 62 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 63 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 64 | } |
| 65 | } |
| 66 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 67 | TEST(F32_DWCONV_CHW_3X3P1__SSE, input_width_div_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 68 | TEST_REQUIRES_X86_SSE; |
| 69 | for (size_t input_width = 8; input_width < 32; input_width += 4) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 70 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 71 | .input_tuple_size(4) |
| 72 | .output_tuple_size(4) |
| 73 | .input_width(input_width) |
| 74 | .padding_left(1) |
| 75 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 76 | .padding_top(1) |
| 77 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 78 | .kernel_height(3) |
| 79 | .kernel_width(3) |
| 80 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 81 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 82 | } |
| 83 | } |
| 84 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 85 | TEST(F32_DWCONV_CHW_3X3P1__SSE, input_width_stride) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 86 | TEST_REQUIRES_X86_SSE; |
| 87 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 88 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 89 | .input_tuple_size(4) |
| 90 | .output_tuple_size(4) |
| 91 | .input_width(input_width) |
| 92 | .input_width_stride(36) |
| 93 | .padding_left(1) |
| 94 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 95 | .padding_top(1) |
| 96 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 97 | .kernel_height(3) |
| 98 | .kernel_width(3) |
| 99 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 100 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 101 | } |
| 102 | } |
| 103 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 104 | TEST(F32_DWCONV_CHW_3X3P1__SSE, input_tuple_stride) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 105 | TEST_REQUIRES_X86_SSE; |
| 106 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 107 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 108 | .input_tuple_size(4) |
| 109 | .output_tuple_size(4) |
| 110 | .input_width(input_width) |
| 111 | .input_width_stride(4) |
| 112 | .input_tuple_stride(3 * 4) |
| 113 | .padding_left(1) |
| 114 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 115 | .padding_top(1) |
| 116 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 117 | .kernel_height(3) |
| 118 | .kernel_width(3) |
| 119 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 120 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 121 | } |
| 122 | } |
| 123 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 124 | TEST(F32_DWCONV_CHW_3X3P1__SSE, output_height_gt_1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 125 | TEST_REQUIRES_X86_SSE; |
| 126 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 127 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 128 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 129 | .input_tuple_size(4) |
| 130 | .output_tuple_size(4) |
| 131 | .input_width(input_width) |
| 132 | .padding_left(1) |
| 133 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 134 | .padding_top(1) |
| 135 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 136 | .kernel_height(3) |
| 137 | .kernel_width(3) |
| 138 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 139 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 140 | } |
| 141 | } |
| 142 | } |
| 143 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 144 | TEST(F32_DWCONV_CHW_3X3P1__SSE, output_width_stride) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 145 | TEST_REQUIRES_X86_SSE; |
| 146 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 147 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 148 | .input_tuple_size(4) |
| 149 | .output_tuple_size(4) |
| 150 | .input_width(input_width) |
| 151 | .padding_left(1) |
| 152 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 153 | .padding_top(1) |
| 154 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 155 | .kernel_height(3) |
| 156 | .kernel_width(3) |
| 157 | .output_height(5) |
| 158 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 159 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 160 | } |
| 161 | } |
| 162 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 163 | TEST(F32_DWCONV_CHW_3X3P1__SSE, output_tuple_stride) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 164 | TEST_REQUIRES_X86_SSE; |
| 165 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 166 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 167 | .input_tuple_size(4) |
| 168 | .output_tuple_size(4) |
| 169 | .input_width(input_width) |
| 170 | .padding_left(1) |
| 171 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 172 | .padding_top(1) |
| 173 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 174 | .kernel_height(3) |
| 175 | .kernel_width(3) |
| 176 | .output_height(5) |
| 177 | .output_width_stride(4) |
| 178 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 179 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 180 | } |
| 181 | } |
| 182 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 183 | TEST(F32_DWCONV_CHW_3X3P1__SSE, chw_layout) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 184 | TEST_REQUIRES_X86_SSE; |
| 185 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 186 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 187 | .input_tuple_size(4) |
| 188 | .output_tuple_size(4) |
| 189 | .input_width(input_width) |
| 190 | .input_width_stride(input_width) |
| 191 | .padding_left(1) |
| 192 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 193 | .padding_top(1) |
| 194 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 195 | .kernel_height(3) |
| 196 | .kernel_width(3) |
| 197 | .output_height(5) |
| 198 | .output_width_stride(input_width) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 199 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 200 | } |
| 201 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 202 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 203 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 204 | #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 205 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_eq_4_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 206 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 207 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 208 | .input_tuple_size(4) |
| 209 | .output_tuple_size(4) |
| 210 | .input_width(4) |
| 211 | .padding_left(1) |
| 212 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 213 | .padding_top(0) |
| 214 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 215 | .kernel_height(3) |
| 216 | .kernel_width(3) |
| 217 | .subsampling(2) |
| 218 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 219 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 220 | } |
| 221 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 222 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_lt_4_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 223 | TEST_REQUIRES_X86_SSE; |
| 224 | for (size_t input_width = 1; input_width < 4; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 225 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 226 | .input_tuple_size(4) |
| 227 | .output_tuple_size(4) |
| 228 | .input_width(input_width) |
| 229 | .padding_left(1) |
| 230 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 231 | .padding_top(0) |
| 232 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 233 | .kernel_height(3) |
| 234 | .kernel_width(3) |
| 235 | .subsampling(2) |
| 236 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 237 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 238 | } |
| 239 | } |
| 240 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 241 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_gt_4_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 242 | TEST_REQUIRES_X86_SSE; |
| 243 | for (size_t input_width = 5; input_width < 8; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 244 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 245 | .input_tuple_size(4) |
| 246 | .output_tuple_size(4) |
| 247 | .input_width(input_width) |
| 248 | .padding_left(1) |
| 249 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 250 | .padding_top(0) |
| 251 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 252 | .kernel_height(3) |
| 253 | .kernel_width(3) |
| 254 | .subsampling(2) |
| 255 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 256 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 257 | } |
| 258 | } |
| 259 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 260 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_div_4_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 261 | TEST_REQUIRES_X86_SSE; |
| 262 | for (size_t input_width = 8; input_width < 32; input_width += 4) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 263 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 264 | .input_tuple_size(4) |
| 265 | .output_tuple_size(4) |
| 266 | .input_width(input_width) |
| 267 | .padding_left(1) |
| 268 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 269 | .padding_top(0) |
| 270 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 271 | .kernel_height(3) |
| 272 | .kernel_width(3) |
| 273 | .subsampling(2) |
| 274 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 275 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 276 | } |
| 277 | } |
| 278 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 279 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_stride_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 280 | TEST_REQUIRES_X86_SSE; |
| 281 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 282 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 283 | .input_tuple_size(4) |
| 284 | .output_tuple_size(4) |
| 285 | .input_width(input_width) |
| 286 | .input_width_stride(36) |
| 287 | .padding_left(1) |
| 288 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 289 | .padding_top(0) |
| 290 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 291 | .kernel_height(3) |
| 292 | .kernel_width(3) |
| 293 | .subsampling(2) |
| 294 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 295 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 296 | } |
| 297 | } |
| 298 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 299 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_tuple_stride_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 300 | TEST_REQUIRES_X86_SSE; |
| 301 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 302 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 303 | .input_tuple_size(4) |
| 304 | .output_tuple_size(4) |
| 305 | .input_width(input_width) |
| 306 | .input_width_stride(4) |
| 307 | .input_tuple_stride(3 * 4) |
| 308 | .padding_left(1) |
| 309 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 310 | .padding_top(0) |
| 311 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 312 | .kernel_height(3) |
| 313 | .kernel_width(3) |
| 314 | .subsampling(2) |
| 315 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 316 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 317 | } |
| 318 | } |
| 319 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 320 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, output_height_gt_1_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 321 | TEST_REQUIRES_X86_SSE; |
| 322 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 323 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 324 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 325 | .input_tuple_size(4) |
| 326 | .output_tuple_size(4) |
| 327 | .input_width(input_width) |
| 328 | .padding_left(1) |
| 329 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 330 | .padding_top(0) |
| 331 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 332 | .kernel_height(3) |
| 333 | .kernel_width(3) |
| 334 | .subsampling(2) |
| 335 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 336 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 337 | } |
| 338 | } |
| 339 | } |
| 340 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 341 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, output_width_stride_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 342 | TEST_REQUIRES_X86_SSE; |
| 343 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 344 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 345 | .input_tuple_size(4) |
| 346 | .output_tuple_size(4) |
| 347 | .input_width(input_width) |
| 348 | .padding_left(1) |
| 349 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 350 | .padding_top(0) |
| 351 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 352 | .kernel_height(3) |
| 353 | .kernel_width(3) |
| 354 | .subsampling(2) |
| 355 | .output_height(5) |
| 356 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 357 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 358 | } |
| 359 | } |
| 360 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 361 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, output_tuple_stride_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 362 | TEST_REQUIRES_X86_SSE; |
| 363 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 364 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 365 | .input_tuple_size(4) |
| 366 | .output_tuple_size(4) |
| 367 | .input_width(input_width) |
| 368 | .padding_left(1) |
| 369 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 370 | .padding_top(0) |
| 371 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 372 | .kernel_height(3) |
| 373 | .kernel_width(3) |
| 374 | .subsampling(2) |
| 375 | .output_height(5) |
| 376 | .output_width_stride(4) |
| 377 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 378 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 379 | } |
| 380 | } |
| 381 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 382 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, chw_layout_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 383 | TEST_REQUIRES_X86_SSE; |
| 384 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 385 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 386 | .input_tuple_size(4) |
| 387 | .output_tuple_size(4) |
| 388 | .input_width(input_width) |
| 389 | .input_width_stride(input_width) |
| 390 | .padding_left(1) |
| 391 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 392 | .padding_top(0) |
| 393 | .padding_bottom(1) |
| 394 | .kernel_height(3) |
| 395 | .kernel_width(3) |
| 396 | .subsampling(2) |
| 397 | .output_height(5) |
| 398 | .output_width_stride((input_width - 1) / 2 + 1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 399 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 400 | } |
| 401 | } |
| 402 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 403 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_eq_4_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 404 | TEST_REQUIRES_X86_SSE; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 405 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 406 | .input_tuple_size(4) |
| 407 | .output_tuple_size(4) |
| 408 | .input_width(4) |
| 409 | .padding_left(1) |
| 410 | .padding_right(1) |
| 411 | .padding_top(1) |
| 412 | .padding_bottom(1) |
| 413 | .kernel_height(3) |
| 414 | .kernel_width(3) |
| 415 | .subsampling(2) |
| 416 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 417 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 418 | } |
| 419 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 420 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_lt_4_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 421 | TEST_REQUIRES_X86_SSE; |
| 422 | for (size_t input_width = 1; input_width < 4; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 423 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 424 | .input_tuple_size(4) |
| 425 | .output_tuple_size(4) |
| 426 | .input_width(input_width) |
| 427 | .padding_left(1) |
| 428 | .padding_right(1) |
| 429 | .padding_top(1) |
| 430 | .padding_bottom(1) |
| 431 | .kernel_height(3) |
| 432 | .kernel_width(3) |
| 433 | .subsampling(2) |
| 434 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 435 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 436 | } |
| 437 | } |
| 438 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 439 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_gt_4_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 440 | TEST_REQUIRES_X86_SSE; |
| 441 | for (size_t input_width = 5; input_width < 8; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 442 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 443 | .input_tuple_size(4) |
| 444 | .output_tuple_size(4) |
| 445 | .input_width(input_width) |
| 446 | .padding_left(1) |
| 447 | .padding_right(1) |
| 448 | .padding_top(1) |
| 449 | .padding_bottom(1) |
| 450 | .kernel_height(3) |
| 451 | .kernel_width(3) |
| 452 | .subsampling(2) |
| 453 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 454 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 455 | } |
| 456 | } |
| 457 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 458 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_div_4_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 459 | TEST_REQUIRES_X86_SSE; |
| 460 | for (size_t input_width = 8; input_width < 32; input_width += 4) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 461 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 462 | .input_tuple_size(4) |
| 463 | .output_tuple_size(4) |
| 464 | .input_width(input_width) |
| 465 | .padding_left(1) |
| 466 | .padding_right(1) |
| 467 | .padding_top(1) |
| 468 | .padding_bottom(1) |
| 469 | .kernel_height(3) |
| 470 | .kernel_width(3) |
| 471 | .subsampling(2) |
| 472 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 473 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 474 | } |
| 475 | } |
| 476 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 477 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_width_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 478 | TEST_REQUIRES_X86_SSE; |
| 479 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 480 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 481 | .input_tuple_size(4) |
| 482 | .output_tuple_size(4) |
| 483 | .input_width(input_width) |
| 484 | .input_width_stride(36) |
| 485 | .padding_left(1) |
| 486 | .padding_right(1) |
| 487 | .padding_top(1) |
| 488 | .padding_bottom(1) |
| 489 | .kernel_height(3) |
| 490 | .kernel_width(3) |
| 491 | .subsampling(2) |
| 492 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 493 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 494 | } |
| 495 | } |
| 496 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 497 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, input_tuple_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 498 | TEST_REQUIRES_X86_SSE; |
| 499 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 500 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 501 | .input_tuple_size(4) |
| 502 | .output_tuple_size(4) |
| 503 | .input_width(input_width) |
| 504 | .input_width_stride(4) |
| 505 | .input_tuple_stride(3 * 4) |
| 506 | .padding_left(1) |
| 507 | .padding_right(1) |
| 508 | .padding_top(1) |
| 509 | .padding_bottom(1) |
| 510 | .kernel_height(3) |
| 511 | .kernel_width(3) |
| 512 | .subsampling(2) |
| 513 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 514 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 515 | } |
| 516 | } |
| 517 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 518 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, output_height_gt_1_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 519 | TEST_REQUIRES_X86_SSE; |
| 520 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 521 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 522 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 523 | .input_tuple_size(4) |
| 524 | .output_tuple_size(4) |
| 525 | .input_width(input_width) |
| 526 | .padding_left(1) |
| 527 | .padding_right(1) |
| 528 | .padding_top(1) |
| 529 | .padding_bottom(1) |
| 530 | .kernel_height(3) |
| 531 | .kernel_width(3) |
| 532 | .subsampling(2) |
| 533 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 534 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 535 | } |
| 536 | } |
| 537 | } |
| 538 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 539 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, output_width_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 540 | TEST_REQUIRES_X86_SSE; |
| 541 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 542 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 543 | .input_tuple_size(4) |
| 544 | .output_tuple_size(4) |
| 545 | .input_width(input_width) |
| 546 | .padding_left(1) |
| 547 | .padding_right(1) |
| 548 | .padding_top(1) |
| 549 | .padding_bottom(1) |
| 550 | .kernel_height(3) |
| 551 | .kernel_width(3) |
| 552 | .subsampling(2) |
| 553 | .output_height(5) |
| 554 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 555 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 556 | } |
| 557 | } |
| 558 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 559 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, output_tuple_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 560 | TEST_REQUIRES_X86_SSE; |
| 561 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 562 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 563 | .input_tuple_size(4) |
| 564 | .output_tuple_size(4) |
| 565 | .input_width(input_width) |
| 566 | .padding_left(1) |
| 567 | .padding_right(1) |
| 568 | .padding_top(1) |
| 569 | .padding_bottom(1) |
| 570 | .kernel_height(3) |
| 571 | .kernel_width(3) |
| 572 | .subsampling(2) |
| 573 | .output_height(5) |
| 574 | .output_width_stride(4) |
| 575 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 576 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 577 | } |
| 578 | } |
| 579 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 580 | TEST(F32_DWCONV_CHW_3X3S2P1__SSE, chw_layout_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 581 | TEST_REQUIRES_X86_SSE; |
| 582 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 583 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 584 | .input_tuple_size(4) |
| 585 | .output_tuple_size(4) |
| 586 | .input_width(input_width) |
| 587 | .input_width_stride(input_width) |
| 588 | .padding_left(1) |
| 589 | .padding_right(1) |
| 590 | .padding_top(1) |
| 591 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 592 | .kernel_height(3) |
| 593 | .kernel_width(3) |
| 594 | .subsampling(2) |
| 595 | .output_height(5) |
| 596 | .output_width_stride((input_width - 1) / 2 + 1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 597 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 598 | } |
| 599 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 600 | #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 601 | |
| 602 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 603 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 604 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, input_width_eq_4_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 605 | TEST_REQUIRES_ARM_NEON_FMA; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 606 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 607 | .input_tuple_size(4) |
| 608 | .output_tuple_size(4) |
| 609 | .input_width(4) |
| 610 | .padding_left(1) |
| 611 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 612 | .padding_top(1) |
| 613 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 614 | .kernel_height(3) |
| 615 | .kernel_width(3) |
| 616 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 617 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 618 | } |
| 619 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 620 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, input_width_lt_4_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 621 | TEST_REQUIRES_ARM_NEON_FMA; |
| 622 | for (size_t input_width = 1; input_width < 4; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 623 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 624 | .input_tuple_size(4) |
| 625 | .output_tuple_size(4) |
| 626 | .input_width(input_width) |
| 627 | .padding_left(1) |
| 628 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 629 | .padding_top(1) |
| 630 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 631 | .kernel_height(3) |
| 632 | .kernel_width(3) |
| 633 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 634 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 635 | } |
| 636 | } |
| 637 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 638 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, input_width_gt_4_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 639 | TEST_REQUIRES_ARM_NEON_FMA; |
| 640 | for (size_t input_width = 5; input_width < 8; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 641 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 642 | .input_tuple_size(4) |
| 643 | .output_tuple_size(4) |
| 644 | .input_width(input_width) |
| 645 | .padding_left(1) |
| 646 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 647 | .padding_top(1) |
| 648 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 649 | .kernel_height(3) |
| 650 | .kernel_width(3) |
| 651 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 652 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 653 | } |
| 654 | } |
| 655 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 656 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, input_width_div_4_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 657 | TEST_REQUIRES_ARM_NEON_FMA; |
| 658 | for (size_t input_width = 8; input_width < 32; input_width += 4) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 659 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 660 | .input_tuple_size(4) |
| 661 | .output_tuple_size(4) |
| 662 | .input_width(input_width) |
| 663 | .padding_left(1) |
| 664 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 665 | .padding_top(1) |
| 666 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 667 | .kernel_height(3) |
| 668 | .kernel_width(3) |
| 669 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 670 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 671 | } |
| 672 | } |
| 673 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 674 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, input_width_stride_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 675 | TEST_REQUIRES_ARM_NEON_FMA; |
| 676 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 677 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 678 | .input_tuple_size(4) |
| 679 | .output_tuple_size(4) |
| 680 | .input_width(input_width) |
| 681 | .input_width_stride(36) |
| 682 | .padding_left(1) |
| 683 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 684 | .padding_top(1) |
| 685 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 686 | .kernel_height(3) |
| 687 | .kernel_width(3) |
| 688 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 689 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 690 | } |
| 691 | } |
| 692 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 693 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, input_tuple_stride_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 694 | TEST_REQUIRES_ARM_NEON_FMA; |
| 695 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 696 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 697 | .input_tuple_size(4) |
| 698 | .output_tuple_size(4) |
| 699 | .input_width(input_width) |
| 700 | .input_width_stride(4) |
| 701 | .input_tuple_stride(3 * 4) |
| 702 | .padding_left(1) |
| 703 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 704 | .padding_top(1) |
| 705 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 706 | .kernel_height(3) |
| 707 | .kernel_width(3) |
| 708 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 709 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 710 | } |
| 711 | } |
| 712 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 713 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, output_height_gt_1_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 714 | TEST_REQUIRES_ARM_NEON_FMA; |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 715 | for (size_t output_height = 2; output_height <= 5; output_height++) { |
| 716 | for (size_t input_width = 8; input_width < 9; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 717 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 718 | .input_tuple_size(4) |
| 719 | .output_tuple_size(4) |
| 720 | .input_width(input_width) |
| 721 | .padding_left(1) |
| 722 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 723 | .padding_top(1) |
| 724 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 725 | .kernel_height(3) |
| 726 | .kernel_width(3) |
| 727 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 728 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 729 | } |
| 730 | } |
| 731 | } |
| 732 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 733 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, output_width_stride_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 734 | TEST_REQUIRES_ARM_NEON_FMA; |
| 735 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 736 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 737 | .input_tuple_size(4) |
| 738 | .output_tuple_size(4) |
| 739 | .input_width(input_width) |
| 740 | .padding_left(1) |
| 741 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 742 | .padding_top(1) |
| 743 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 744 | .kernel_height(3) |
| 745 | .kernel_width(3) |
| 746 | .output_height(5) |
| 747 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 748 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 749 | } |
| 750 | } |
| 751 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 752 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, output_tuple_stride_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 753 | TEST_REQUIRES_ARM_NEON_FMA; |
| 754 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 755 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 756 | .input_tuple_size(4) |
| 757 | .output_tuple_size(4) |
| 758 | .input_width(input_width) |
| 759 | .padding_left(1) |
| 760 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 761 | .padding_top(1) |
| 762 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 763 | .kernel_height(3) |
| 764 | .kernel_width(3) |
| 765 | .output_height(5) |
| 766 | .output_width_stride(4) |
| 767 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 768 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 769 | } |
| 770 | } |
| 771 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 772 | TEST(F32_DWCONV_CHW_3X3P1__NEONFMA, chw_layout_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 773 | TEST_REQUIRES_ARM_NEON_FMA; |
| 774 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 775 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 776 | .input_tuple_size(4) |
| 777 | .output_tuple_size(4) |
| 778 | .input_width(input_width) |
| 779 | .input_width_stride(input_width) |
| 780 | .padding_left(1) |
| 781 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 782 | .padding_top(1) |
| 783 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 784 | .kernel_height(3) |
| 785 | .kernel_width(3) |
| 786 | .output_height(5) |
| 787 | .output_width_stride(input_width) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 788 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 789 | } |
| 790 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 791 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 792 | |
| 793 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 794 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 795 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_eq_4_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 796 | TEST_REQUIRES_ARM_NEON_FMA; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 797 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 798 | .input_tuple_size(4) |
| 799 | .output_tuple_size(4) |
| 800 | .input_width(4) |
| 801 | .padding_left(1) |
| 802 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 803 | .padding_top(0) |
| 804 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 805 | .kernel_height(3) |
| 806 | .kernel_width(3) |
| 807 | .subsampling(2) |
| 808 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 809 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 810 | } |
| 811 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 812 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_eq_4_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 813 | TEST_REQUIRES_ARM_NEON_FMA; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 814 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 815 | .input_tuple_size(4) |
| 816 | .output_tuple_size(4) |
| 817 | .input_width(4) |
| 818 | .padding_left(1) |
| 819 | .padding_right(1) |
| 820 | .padding_top(1) |
| 821 | .padding_bottom(1) |
| 822 | .kernel_height(3) |
| 823 | .kernel_width(3) |
| 824 | .subsampling(2) |
| 825 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 826 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 827 | } |
| 828 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 829 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_lt_4_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 830 | TEST_REQUIRES_ARM_NEON_FMA; |
| 831 | for (size_t input_width = 1; input_width < 4; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 832 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 833 | .input_tuple_size(4) |
| 834 | .output_tuple_size(4) |
| 835 | .input_width(input_width) |
| 836 | .padding_left(1) |
| 837 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 838 | .padding_top(0) |
| 839 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 840 | .kernel_height(3) |
| 841 | .kernel_width(3) |
| 842 | .subsampling(2) |
| 843 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 844 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 845 | } |
| 846 | } |
| 847 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 848 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_lt_4_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 849 | TEST_REQUIRES_ARM_NEON_FMA; |
| 850 | for (size_t input_width = 1; input_width < 4; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 851 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 852 | .input_tuple_size(4) |
| 853 | .output_tuple_size(4) |
| 854 | .input_width(input_width) |
| 855 | .padding_left(1) |
| 856 | .padding_right(1) |
| 857 | .padding_top(1) |
| 858 | .padding_bottom(1) |
| 859 | .kernel_height(3) |
| 860 | .kernel_width(3) |
| 861 | .subsampling(2) |
| 862 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 863 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 864 | } |
| 865 | } |
| 866 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 867 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_gt_4_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 868 | TEST_REQUIRES_ARM_NEON_FMA; |
| 869 | for (size_t input_width = 5; input_width < 8; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 870 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 871 | .input_tuple_size(4) |
| 872 | .output_tuple_size(4) |
| 873 | .input_width(input_width) |
| 874 | .padding_left(1) |
| 875 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 876 | .padding_top(0) |
| 877 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 878 | .kernel_height(3) |
| 879 | .kernel_width(3) |
| 880 | .subsampling(2) |
| 881 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 882 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 883 | } |
| 884 | } |
| 885 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 886 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_gt_4_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 887 | TEST_REQUIRES_ARM_NEON_FMA; |
| 888 | for (size_t input_width = 5; input_width < 8; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 889 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 890 | .input_tuple_size(4) |
| 891 | .output_tuple_size(4) |
| 892 | .input_width(input_width) |
| 893 | .padding_left(1) |
| 894 | .padding_right(1) |
| 895 | .padding_top(1) |
| 896 | .padding_bottom(1) |
| 897 | .kernel_height(3) |
| 898 | .kernel_width(3) |
| 899 | .subsampling(2) |
| 900 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 901 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 902 | } |
| 903 | } |
| 904 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 905 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_div_4_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 906 | TEST_REQUIRES_ARM_NEON_FMA; |
| 907 | for (size_t input_width = 8; input_width < 32; input_width += 4) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 908 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 909 | .input_tuple_size(4) |
| 910 | .output_tuple_size(4) |
| 911 | .input_width(input_width) |
| 912 | .padding_left(1) |
| 913 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 914 | .padding_top(0) |
| 915 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 916 | .kernel_height(3) |
| 917 | .kernel_width(3) |
| 918 | .subsampling(2) |
| 919 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 920 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 921 | } |
| 922 | } |
| 923 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 924 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_div_4_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 925 | TEST_REQUIRES_ARM_NEON_FMA; |
| 926 | for (size_t input_width = 8; input_width < 32; input_width += 4) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 927 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 928 | .input_tuple_size(4) |
| 929 | .output_tuple_size(4) |
| 930 | .input_width(input_width) |
| 931 | .padding_left(1) |
| 932 | .padding_right(1) |
| 933 | .padding_top(1) |
| 934 | .padding_bottom(1) |
| 935 | .kernel_height(3) |
| 936 | .kernel_width(3) |
| 937 | .subsampling(2) |
| 938 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 939 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 940 | } |
| 941 | } |
| 942 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 943 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_stride_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 944 | TEST_REQUIRES_ARM_NEON_FMA; |
| 945 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 946 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 947 | .input_tuple_size(4) |
| 948 | .output_tuple_size(4) |
| 949 | .input_width(input_width) |
| 950 | .input_width_stride(36) |
| 951 | .padding_left(1) |
| 952 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 953 | .padding_top(0) |
| 954 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 955 | .kernel_height(3) |
| 956 | .kernel_width(3) |
| 957 | .subsampling(2) |
| 958 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 959 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 960 | } |
| 961 | } |
| 962 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 963 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_width_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 964 | TEST_REQUIRES_ARM_NEON_FMA; |
| 965 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 966 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 967 | .input_tuple_size(4) |
| 968 | .output_tuple_size(4) |
| 969 | .input_width(input_width) |
| 970 | .input_width_stride(36) |
| 971 | .padding_left(1) |
| 972 | .padding_right(1) |
| 973 | .padding_top(1) |
| 974 | .padding_bottom(1) |
| 975 | .kernel_height(3) |
| 976 | .kernel_width(3) |
| 977 | .subsampling(2) |
| 978 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 979 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 980 | } |
| 981 | } |
| 982 | |
| 983 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 984 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_tuple_stride_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 985 | TEST_REQUIRES_ARM_NEON_FMA; |
| 986 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 987 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 988 | .input_tuple_size(4) |
| 989 | .output_tuple_size(4) |
| 990 | .input_width(input_width) |
| 991 | .input_width_stride(4) |
| 992 | .input_tuple_stride(3 * 4) |
| 993 | .padding_left(1) |
| 994 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 995 | .padding_top(0) |
| 996 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 997 | .kernel_height(3) |
| 998 | .kernel_width(3) |
| 999 | .subsampling(2) |
| 1000 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1001 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1002 | } |
| 1003 | } |
| 1004 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1005 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, input_tuple_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1006 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1007 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1008 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1009 | .input_tuple_size(4) |
| 1010 | .output_tuple_size(4) |
| 1011 | .input_width(input_width) |
| 1012 | .input_width_stride(4) |
| 1013 | .input_tuple_stride(3 * 4) |
| 1014 | .padding_left(1) |
| 1015 | .padding_right(1) |
| 1016 | .padding_top(1) |
| 1017 | .padding_bottom(1) |
| 1018 | .kernel_height(3) |
| 1019 | .kernel_width(3) |
| 1020 | .subsampling(2) |
| 1021 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1022 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1023 | } |
| 1024 | } |
| 1025 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1026 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, output_height_gt_1_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1027 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1028 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 1029 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1030 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1031 | .input_tuple_size(4) |
| 1032 | .output_tuple_size(4) |
| 1033 | .input_width(input_width) |
| 1034 | .padding_left(1) |
| 1035 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1036 | .padding_top(0) |
| 1037 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1038 | .kernel_height(3) |
| 1039 | .kernel_width(3) |
| 1040 | .subsampling(2) |
| 1041 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1042 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1043 | } |
| 1044 | } |
| 1045 | } |
| 1046 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1047 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, output_height_gt_1_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1048 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1049 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 1050 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1051 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1052 | .input_tuple_size(4) |
| 1053 | .output_tuple_size(4) |
| 1054 | .input_width(input_width) |
| 1055 | .padding_left(1) |
| 1056 | .padding_right(1) |
| 1057 | .padding_top(1) |
| 1058 | .padding_bottom(1) |
| 1059 | .kernel_height(3) |
| 1060 | .kernel_width(3) |
| 1061 | .subsampling(2) |
| 1062 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1063 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1064 | } |
| 1065 | } |
| 1066 | } |
| 1067 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1068 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, output_width_stride_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1069 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1070 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1071 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1072 | .input_tuple_size(4) |
| 1073 | .output_tuple_size(4) |
| 1074 | .input_width(input_width) |
| 1075 | .padding_left(1) |
| 1076 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1077 | .padding_top(0) |
| 1078 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1079 | .kernel_height(3) |
| 1080 | .kernel_width(3) |
| 1081 | .subsampling(2) |
| 1082 | .output_height(5) |
| 1083 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1084 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1085 | } |
| 1086 | } |
| 1087 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1088 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, output_width_stride_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1089 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1090 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1091 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1092 | .input_tuple_size(4) |
| 1093 | .output_tuple_size(4) |
| 1094 | .input_width(input_width) |
| 1095 | .padding_left(1) |
| 1096 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1097 | .padding_top(1) |
| 1098 | .padding_bottom(1) |
| 1099 | .kernel_height(3) |
| 1100 | .kernel_width(3) |
| 1101 | .subsampling(2) |
| 1102 | .output_height(5) |
| 1103 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1104 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1105 | } |
| 1106 | } |
| 1107 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1108 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, output_tuple_stride_pad0) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1109 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1110 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1111 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1112 | .input_tuple_size(4) |
| 1113 | .output_tuple_size(4) |
| 1114 | .input_width(input_width) |
| 1115 | .padding_left(1) |
| 1116 | .padding_right(1) |
| 1117 | .padding_top(0) |
| 1118 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1119 | .kernel_height(3) |
| 1120 | .kernel_width(3) |
| 1121 | .subsampling(2) |
| 1122 | .output_height(5) |
| 1123 | .output_width_stride(4) |
| 1124 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1125 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1126 | } |
| 1127 | } |
| 1128 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1129 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, output_tuple_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1130 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1131 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1132 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1133 | .input_tuple_size(4) |
| 1134 | .output_tuple_size(4) |
| 1135 | .input_width(input_width) |
| 1136 | .padding_left(1) |
| 1137 | .padding_right(1) |
| 1138 | .padding_top(1) |
| 1139 | .padding_bottom(1) |
| 1140 | .kernel_height(3) |
| 1141 | .kernel_width(3) |
| 1142 | .subsampling(2) |
| 1143 | .output_height(5) |
| 1144 | .output_width_stride(4) |
| 1145 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1146 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1147 | } |
| 1148 | } |
| 1149 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1150 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, chw_layout_pad0) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1151 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1152 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1153 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1154 | .input_tuple_size(4) |
| 1155 | .output_tuple_size(4) |
| 1156 | .input_width(input_width) |
| 1157 | .input_width_stride(input_width) |
| 1158 | .padding_left(1) |
| 1159 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1160 | .padding_top(0) |
| 1161 | .padding_bottom(1) |
| 1162 | .kernel_height(3) |
| 1163 | .kernel_width(3) |
| 1164 | .subsampling(2) |
| 1165 | .output_height(5) |
| 1166 | .output_width_stride((input_width - 1) / 2 + 1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1167 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1168 | } |
| 1169 | } |
| 1170 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1171 | TEST(F32_DWCONV_CHW_3X3S2P1__NEONFMA, chw_layout_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1172 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1173 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1174 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1175 | .input_tuple_size(4) |
| 1176 | .output_tuple_size(4) |
| 1177 | .input_width(input_width) |
| 1178 | .input_width_stride(input_width) |
| 1179 | .padding_left(1) |
| 1180 | .padding_right(1) |
| 1181 | .padding_top(1) |
| 1182 | .padding_bottom(1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1183 | .kernel_height(3) |
| 1184 | .kernel_width(3) |
| 1185 | .subsampling(2) |
| 1186 | .output_height(5) |
| 1187 | .output_width_stride((input_width - 1) / 2 + 1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1188 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1189 | } |
| 1190 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1191 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1192 | |
| 1193 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1194 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1195 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, input_width_eq_4_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1196 | TEST_REQUIRES_ARM_NEON_FMA; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1197 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1198 | .input_tuple_size(4) |
| 1199 | .output_tuple_size(4) |
| 1200 | .input_width(4) |
| 1201 | .padding_left(2) |
| 1202 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1203 | .padding_top(2) |
| 1204 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1205 | .kernel_height(5) |
| 1206 | .kernel_width(5) |
| 1207 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1208 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1209 | } |
| 1210 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1211 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, input_width_lt_4_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1212 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1213 | for (size_t input_width = 1; input_width < 4; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1214 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1215 | .input_tuple_size(4) |
| 1216 | .output_tuple_size(4) |
| 1217 | .input_width(input_width) |
| 1218 | .padding_left(2) |
| 1219 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1220 | .padding_top(2) |
| 1221 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1222 | .kernel_height(5) |
| 1223 | .kernel_width(5) |
| 1224 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1225 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1226 | } |
| 1227 | } |
| 1228 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1229 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, input_width_gt_4_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1230 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1231 | for (size_t input_width = 5; input_width < 8; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1232 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1233 | .input_tuple_size(4) |
| 1234 | .output_tuple_size(4) |
| 1235 | .input_width(input_width) |
| 1236 | .padding_left(2) |
| 1237 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1238 | .padding_top(2) |
| 1239 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1240 | .kernel_height(5) |
| 1241 | .kernel_width(5) |
| 1242 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1243 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1244 | } |
| 1245 | } |
| 1246 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1247 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, input_width_div_4_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1248 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1249 | for (size_t input_width = 8; input_width < 32; input_width += 4) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1250 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1251 | .input_tuple_size(4) |
| 1252 | .output_tuple_size(4) |
| 1253 | .input_width(input_width) |
| 1254 | .padding_left(2) |
| 1255 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1256 | .padding_top(2) |
| 1257 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1258 | .kernel_height(5) |
| 1259 | .kernel_width(5) |
| 1260 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1261 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1262 | } |
| 1263 | } |
| 1264 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1265 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, input_width_stride_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1266 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1267 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1268 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1269 | .input_tuple_size(4) |
| 1270 | .output_tuple_size(4) |
| 1271 | .input_width(input_width) |
| 1272 | .input_width_stride(36) |
| 1273 | .padding_left(2) |
| 1274 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1275 | .padding_top(2) |
| 1276 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1277 | .kernel_height(5) |
| 1278 | .kernel_width(5) |
| 1279 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1280 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1281 | } |
| 1282 | } |
| 1283 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1284 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, input_tuple_stride_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1285 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1286 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1287 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1288 | .input_tuple_size(4) |
| 1289 | .output_tuple_size(4) |
| 1290 | .input_width(input_width) |
| 1291 | .input_width_stride(4) |
| 1292 | .input_tuple_stride(3 * 4) |
| 1293 | .padding_left(2) |
| 1294 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1295 | .padding_top(2) |
| 1296 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1297 | .kernel_height(5) |
| 1298 | .kernel_width(5) |
| 1299 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1300 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1301 | } |
| 1302 | } |
| 1303 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1304 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, output_height_eq_2_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1305 | TEST_REQUIRES_ARM_NEON_FMA; |
Erich Elsen | 4ad5115 | 2019-11-19 13:11:53 -0800 | [diff] [blame] | 1306 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1307 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4ad5115 | 2019-11-19 13:11:53 -0800 | [diff] [blame] | 1308 | .input_tuple_size(4) |
| 1309 | .output_tuple_size(4) |
| 1310 | .input_width(input_width) |
| 1311 | .padding_left(2) |
| 1312 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1313 | .padding_top(2) |
| 1314 | .padding_bottom(2) |
Erich Elsen | 4ad5115 | 2019-11-19 13:11:53 -0800 | [diff] [blame] | 1315 | .kernel_height(5) |
| 1316 | .kernel_width(5) |
| 1317 | .output_height(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1318 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
Erich Elsen | 4ad5115 | 2019-11-19 13:11:53 -0800 | [diff] [blame] | 1319 | } |
| 1320 | } |
| 1321 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1322 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, output_height_gt_2_pad2) { |
Erich Elsen | 4ad5115 | 2019-11-19 13:11:53 -0800 | [diff] [blame] | 1323 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1324 | for (size_t output_height = 3; output_height < 5; output_height++) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1325 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1326 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1327 | .input_tuple_size(4) |
| 1328 | .output_tuple_size(4) |
| 1329 | .input_width(input_width) |
| 1330 | .padding_left(2) |
| 1331 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1332 | .padding_top(2) |
| 1333 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1334 | .kernel_height(5) |
| 1335 | .kernel_width(5) |
| 1336 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1337 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1338 | } |
| 1339 | } |
| 1340 | } |
| 1341 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1342 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, output_width_stride_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1343 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1344 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1345 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1346 | .input_tuple_size(4) |
| 1347 | .output_tuple_size(4) |
| 1348 | .input_width(input_width) |
| 1349 | .padding_left(2) |
| 1350 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1351 | .padding_top(2) |
| 1352 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1353 | .kernel_height(5) |
| 1354 | .kernel_width(5) |
| 1355 | .output_height(5) |
| 1356 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1357 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1358 | } |
| 1359 | } |
| 1360 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1361 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, output_tuple_stride_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1362 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1363 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1364 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1365 | .input_tuple_size(4) |
| 1366 | .output_tuple_size(4) |
| 1367 | .input_width(input_width) |
| 1368 | .padding_left(2) |
| 1369 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1370 | .padding_top(2) |
| 1371 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1372 | .kernel_height(5) |
| 1373 | .kernel_width(5) |
| 1374 | .output_height(5) |
| 1375 | .output_width_stride(4) |
| 1376 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1377 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1378 | } |
| 1379 | } |
| 1380 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1381 | TEST(F32_DWCONV_CHW_5X5P2__NEONFMA, chw_layout_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1382 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1383 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
| 1384 | for (size_t output_height = 1; output_height < 32; output_height += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1385 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1386 | .input_tuple_size(4) |
| 1387 | .output_tuple_size(4) |
| 1388 | .input_width(input_width) |
| 1389 | .input_width_stride(input_width) |
| 1390 | .padding_left(2) |
| 1391 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1392 | .padding_top(2) |
| 1393 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1394 | .kernel_height(5) |
| 1395 | .kernel_width(5) |
| 1396 | .output_height(5) |
| 1397 | .output_width_stride(input_width) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1398 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1399 | } |
| 1400 | } |
| 1401 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1402 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1403 | |
| 1404 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1405 | #if XNN_ARCH_ARM64 |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1406 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_eq_8_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1407 | TEST_REQUIRES_ARM_NEON_FMA; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1408 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1409 | .input_tuple_size(4) |
| 1410 | .output_tuple_size(4) |
| 1411 | .input_width(8) |
| 1412 | .padding_left(2) |
| 1413 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1414 | .padding_top(2) |
| 1415 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1416 | .kernel_height(5) |
| 1417 | .kernel_width(5) |
| 1418 | .subsampling(2) |
| 1419 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1420 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1421 | } |
| 1422 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1423 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_eq_8_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1424 | TEST_REQUIRES_ARM_NEON_FMA; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1425 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1426 | .input_tuple_size(4) |
| 1427 | .output_tuple_size(4) |
| 1428 | .input_width(8) |
| 1429 | .padding_left(2) |
| 1430 | .padding_right(2) |
| 1431 | .padding_top(1) |
| 1432 | .padding_bottom(2) |
| 1433 | .kernel_height(5) |
| 1434 | .kernel_width(5) |
| 1435 | .subsampling(2) |
| 1436 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1437 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1438 | } |
| 1439 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1440 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_lt_8_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1441 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1442 | for (size_t input_width = 1; input_width < 8; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1443 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1444 | .input_tuple_size(4) |
| 1445 | .output_tuple_size(4) |
| 1446 | .input_width(input_width) |
| 1447 | .padding_left(2) |
| 1448 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1449 | .padding_top(1) |
| 1450 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1451 | .kernel_height(5) |
| 1452 | .kernel_width(5) |
| 1453 | .subsampling(2) |
| 1454 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1455 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1456 | } |
| 1457 | } |
| 1458 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1459 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_lt_8_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1460 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1461 | for (size_t input_width = 1; input_width < 8; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1462 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1463 | .input_tuple_size(4) |
| 1464 | .output_tuple_size(4) |
| 1465 | .input_width(input_width) |
| 1466 | .padding_left(2) |
| 1467 | .padding_right(2) |
| 1468 | .padding_top(2) |
| 1469 | .padding_bottom(2) |
| 1470 | .kernel_height(5) |
| 1471 | .kernel_width(5) |
| 1472 | .subsampling(2) |
| 1473 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1474 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1475 | } |
| 1476 | } |
| 1477 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1478 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_gt_8_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1479 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1480 | for (size_t input_width = 8; input_width < 16; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1481 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1482 | .input_tuple_size(4) |
| 1483 | .output_tuple_size(4) |
| 1484 | .input_width(input_width) |
| 1485 | .padding_left(2) |
| 1486 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1487 | .padding_top(1) |
| 1488 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1489 | .kernel_height(5) |
| 1490 | .kernel_width(5) |
| 1491 | .subsampling(2) |
| 1492 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1493 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1494 | } |
| 1495 | } |
| 1496 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1497 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_gt_8_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1498 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1499 | for (size_t input_width = 8; input_width < 16; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1500 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1501 | .input_tuple_size(4) |
| 1502 | .output_tuple_size(4) |
| 1503 | .input_width(input_width) |
| 1504 | .padding_left(2) |
| 1505 | .padding_right(2) |
| 1506 | .padding_top(2) |
| 1507 | .padding_bottom(2) |
| 1508 | .kernel_height(5) |
| 1509 | .kernel_width(5) |
| 1510 | .subsampling(2) |
| 1511 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1512 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1513 | } |
| 1514 | } |
| 1515 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1516 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_div_4_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1517 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1518 | for (size_t input_width = 16; input_width < 32; input_width += 4) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1519 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1520 | .input_tuple_size(4) |
| 1521 | .output_tuple_size(4) |
| 1522 | .input_width(input_width) |
| 1523 | .padding_left(2) |
| 1524 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1525 | .padding_top(1) |
| 1526 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1527 | .kernel_height(5) |
| 1528 | .kernel_width(5) |
| 1529 | .subsampling(2) |
| 1530 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1531 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1532 | } |
| 1533 | } |
| 1534 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1535 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_div_4_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1536 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1537 | for (size_t input_width = 16; input_width < 32; input_width += 4) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1538 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1539 | .input_tuple_size(4) |
| 1540 | .output_tuple_size(4) |
| 1541 | .input_width(input_width) |
| 1542 | .padding_left(2) |
| 1543 | .padding_right(2) |
| 1544 | .padding_top(2) |
| 1545 | .padding_bottom(2) |
| 1546 | .kernel_height(5) |
| 1547 | .kernel_width(5) |
| 1548 | .subsampling(2) |
| 1549 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1550 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1551 | } |
| 1552 | } |
| 1553 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1554 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_stride_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1555 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1556 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1557 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1558 | .input_tuple_size(4) |
| 1559 | .output_tuple_size(4) |
| 1560 | .input_width(input_width) |
| 1561 | .input_width_stride(36) |
| 1562 | .padding_left(2) |
| 1563 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1564 | .padding_top(1) |
| 1565 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1566 | .kernel_height(5) |
| 1567 | .kernel_width(5) |
| 1568 | .subsampling(2) |
| 1569 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1570 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1571 | } |
| 1572 | } |
| 1573 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1574 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_width_stride_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1575 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1576 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1577 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1578 | .input_tuple_size(4) |
| 1579 | .output_tuple_size(4) |
| 1580 | .input_width(input_width) |
| 1581 | .input_width_stride(36) |
| 1582 | .padding_left(2) |
| 1583 | .padding_right(2) |
| 1584 | .padding_top(2) |
| 1585 | .padding_bottom(2) |
| 1586 | .kernel_height(5) |
| 1587 | .kernel_width(5) |
| 1588 | .subsampling(2) |
| 1589 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1590 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1591 | } |
| 1592 | } |
| 1593 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1594 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_tuple_stride_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1595 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1596 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1597 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1598 | .input_tuple_size(4) |
| 1599 | .output_tuple_size(4) |
| 1600 | .input_width(input_width) |
| 1601 | .input_width_stride(4) |
| 1602 | .input_tuple_stride(3 * 4) |
| 1603 | .padding_left(2) |
| 1604 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1605 | .padding_top(1) |
| 1606 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1607 | .kernel_height(5) |
| 1608 | .kernel_width(5) |
| 1609 | .subsampling(2) |
| 1610 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1611 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1612 | } |
| 1613 | } |
| 1614 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1615 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, input_tuple_stride_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1616 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1617 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1618 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1619 | .input_tuple_size(4) |
| 1620 | .output_tuple_size(4) |
| 1621 | .input_width(input_width) |
| 1622 | .input_width_stride(4) |
| 1623 | .input_tuple_stride(3 * 4) |
| 1624 | .padding_left(2) |
| 1625 | .padding_right(2) |
| 1626 | .padding_top(2) |
| 1627 | .padding_bottom(2) |
| 1628 | .kernel_height(5) |
| 1629 | .kernel_width(5) |
| 1630 | .subsampling(2) |
| 1631 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1632 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1633 | } |
| 1634 | } |
| 1635 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1636 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, output_height_gt_1_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1637 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1638 | for (size_t output_height = 3; output_height < 4; output_height++) { |
| 1639 | for (size_t input_width = 4; input_width < 5; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1640 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1641 | .input_tuple_size(4) |
| 1642 | .output_tuple_size(4) |
| 1643 | .input_width(input_width) |
| 1644 | .padding_left(2) |
| 1645 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1646 | .padding_top(1) |
| 1647 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1648 | .kernel_height(5) |
| 1649 | .kernel_width(5) |
| 1650 | .subsampling(2) |
| 1651 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1652 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1653 | } |
| 1654 | } |
| 1655 | } |
| 1656 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1657 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, output_height_gt_1_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1658 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1659 | for (size_t output_height = 3; output_height < 4; output_height++) { |
| 1660 | for (size_t input_width = 4; input_width < 5; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1661 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1662 | .input_tuple_size(4) |
| 1663 | .output_tuple_size(4) |
| 1664 | .input_width(input_width) |
| 1665 | .padding_left(2) |
| 1666 | .padding_right(2) |
| 1667 | .padding_top(2) |
| 1668 | .padding_bottom(2) |
| 1669 | .kernel_height(5) |
| 1670 | .kernel_width(5) |
| 1671 | .subsampling(2) |
| 1672 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1673 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1674 | } |
| 1675 | } |
| 1676 | } |
| 1677 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1678 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, output_width_stride_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1679 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1680 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1681 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1682 | .input_tuple_size(4) |
| 1683 | .output_tuple_size(4) |
| 1684 | .input_width(input_width) |
| 1685 | .padding_left(2) |
| 1686 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1687 | .padding_top(1) |
| 1688 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1689 | .kernel_height(5) |
| 1690 | .kernel_width(5) |
| 1691 | .subsampling(2) |
| 1692 | .output_height(5) |
| 1693 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1694 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1695 | } |
| 1696 | } |
| 1697 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1698 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, output_width_stride_pad2) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1699 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1700 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1701 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1702 | .input_tuple_size(4) |
| 1703 | .output_tuple_size(4) |
| 1704 | .input_width(input_width) |
| 1705 | .padding_left(2) |
| 1706 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1707 | .padding_top(2) |
| 1708 | .padding_bottom(2) |
| 1709 | .kernel_height(5) |
| 1710 | .kernel_width(5) |
| 1711 | .subsampling(2) |
| 1712 | .output_height(5) |
| 1713 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1714 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1715 | } |
| 1716 | } |
| 1717 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1718 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, output_tuple_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1719 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1720 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1721 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1722 | .input_tuple_size(4) |
| 1723 | .output_tuple_size(4) |
| 1724 | .input_width(input_width) |
| 1725 | .padding_left(2) |
| 1726 | .padding_right(2) |
| 1727 | .padding_top(1) |
| 1728 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1729 | .kernel_height(5) |
| 1730 | .kernel_width(5) |
| 1731 | .subsampling(2) |
| 1732 | .output_height(5) |
| 1733 | .output_width_stride(4) |
| 1734 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1735 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1736 | } |
| 1737 | } |
| 1738 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1739 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, output_tuple_stride_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1740 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1741 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1742 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1743 | .input_tuple_size(4) |
| 1744 | .output_tuple_size(4) |
| 1745 | .input_width(input_width) |
| 1746 | .padding_left(2) |
| 1747 | .padding_right(2) |
| 1748 | .padding_top(2) |
| 1749 | .padding_bottom(2) |
| 1750 | .kernel_height(5) |
| 1751 | .kernel_width(5) |
| 1752 | .subsampling(2) |
| 1753 | .output_height(5) |
| 1754 | .output_width_stride(4) |
| 1755 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1756 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1757 | } |
| 1758 | } |
| 1759 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1760 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, chw_layout_pad1) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1761 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1762 | for (size_t input_width = 1; input_width < 32; input_width += 1) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1763 | DWConvCHWMicrokernelTester() |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1764 | .input_tuple_size(4) |
| 1765 | .output_tuple_size(4) |
| 1766 | .input_width(input_width) |
| 1767 | .input_width_stride(input_width) |
| 1768 | .padding_left(2) |
| 1769 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1770 | .padding_top(1) |
| 1771 | .padding_bottom(2) |
| 1772 | .kernel_height(5) |
| 1773 | .kernel_width(5) |
| 1774 | .subsampling(2) |
| 1775 | .output_height(5) |
| 1776 | .output_width_stride((input_width - 1) / 2 + 1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1777 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1778 | } |
| 1779 | } |
| 1780 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1781 | TEST(F32_DWCONV_CHW_5X5S2P2__NEONFMA, chw_layout_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1782 | TEST_REQUIRES_ARM_NEON_FMA; |
| 1783 | for (size_t input_width = 1; input_width < 32; input_width += 1) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1784 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1785 | .input_tuple_size(4) |
| 1786 | .output_tuple_size(4) |
| 1787 | .input_width(input_width) |
| 1788 | .input_width_stride(input_width) |
| 1789 | .padding_left(2) |
| 1790 | .padding_right(2) |
| 1791 | .padding_top(2) |
| 1792 | .padding_bottom(2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1793 | .kernel_height(5) |
| 1794 | .kernel_width(5) |
| 1795 | .subsampling(2) |
| 1796 | .output_height(5) |
| 1797 | .output_width_stride((input_width - 1) / 2 + 1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1798 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1799 | } |
| 1800 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1801 | #endif // XNN_ARCH_ARM64 |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1802 | |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1803 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1804 | TEST(F32_DWCONV_CHW_3X3P1__SCALAR, input_width_eq_1) { |
| 1805 | DWConvCHWMicrokernelTester() |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1806 | .input_tuple_size(1) |
| 1807 | .output_tuple_size(1) |
| 1808 | .input_width(1) |
| 1809 | .padding_left(1) |
| 1810 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1811 | .padding_top(1) |
| 1812 | .padding_bottom(1) |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1813 | .kernel_height(3) |
| 1814 | .kernel_width(3) |
| 1815 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1816 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1817 | } |
| 1818 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1819 | TEST(F32_DWCONV_CHW_3X3P1__SCALAR, input_width_gt_1) { |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1820 | for (size_t input_width = 2; input_width < 32; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1821 | DWConvCHWMicrokernelTester() |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1822 | .input_tuple_size(1) |
| 1823 | .output_tuple_size(1) |
| 1824 | .input_width(input_width) |
| 1825 | .padding_left(1) |
| 1826 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1827 | .padding_top(1) |
| 1828 | .padding_bottom(1) |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1829 | .kernel_height(3) |
| 1830 | .kernel_width(3) |
| 1831 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1832 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1833 | } |
| 1834 | } |
| 1835 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1836 | TEST(F32_DWCONV_CHW_3X3P1__SCALAR, input_width_stride) { |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1837 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1838 | DWConvCHWMicrokernelTester() |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1839 | .input_tuple_size(1) |
| 1840 | .output_tuple_size(1) |
| 1841 | .input_width(input_width) |
| 1842 | .input_width_stride(36) |
| 1843 | .padding_left(1) |
| 1844 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1845 | .padding_top(1) |
| 1846 | .padding_bottom(1) |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1847 | .kernel_height(3) |
| 1848 | .kernel_width(3) |
| 1849 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1850 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1851 | } |
| 1852 | } |
| 1853 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1854 | TEST(F32_DWCONV_CHW_3X3P1__SCALAR, input_tuple_stride) { |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1855 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1856 | DWConvCHWMicrokernelTester() |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1857 | .input_tuple_size(1) |
| 1858 | .output_tuple_size(1) |
| 1859 | .input_width(input_width) |
| 1860 | .input_width_stride(4) |
| 1861 | .input_tuple_stride(3 * 4) |
| 1862 | .padding_left(1) |
| 1863 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1864 | .padding_top(1) |
| 1865 | .padding_bottom(1) |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1866 | .kernel_height(3) |
| 1867 | .kernel_width(3) |
| 1868 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1869 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1870 | } |
| 1871 | } |
| 1872 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1873 | TEST(F32_DWCONV_CHW_3X3P1__SCALAR, output_height_gt_1) { |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1874 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 1875 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1876 | DWConvCHWMicrokernelTester() |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1877 | .input_tuple_size(1) |
| 1878 | .output_tuple_size(1) |
| 1879 | .input_width(input_width) |
| 1880 | .padding_left(1) |
| 1881 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1882 | .padding_top(1) |
| 1883 | .padding_bottom(1) |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1884 | .kernel_height(3) |
| 1885 | .kernel_width(3) |
| 1886 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1887 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1888 | } |
| 1889 | } |
| 1890 | } |
| 1891 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1892 | TEST(F32_DWCONV_CHW_3X3P1__SCALAR, output_width_stride) { |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1893 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1894 | DWConvCHWMicrokernelTester() |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1895 | .input_tuple_size(1) |
| 1896 | .output_tuple_size(1) |
| 1897 | .input_width(input_width) |
| 1898 | .padding_left(1) |
| 1899 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1900 | .padding_top(1) |
| 1901 | .padding_bottom(1) |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1902 | .kernel_height(3) |
| 1903 | .kernel_width(3) |
| 1904 | .output_height(5) |
| 1905 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1906 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1907 | } |
| 1908 | } |
| 1909 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1910 | TEST(F32_DWCONV_CHW_3X3P1__SCALAR, output_tuple_stride) { |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1911 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1912 | DWConvCHWMicrokernelTester() |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1913 | .input_tuple_size(1) |
| 1914 | .output_tuple_size(1) |
| 1915 | .input_width(input_width) |
| 1916 | .padding_left(1) |
| 1917 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1918 | .padding_top(1) |
| 1919 | .padding_bottom(1) |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1920 | .kernel_height(3) |
| 1921 | .kernel_width(3) |
| 1922 | .output_height(5) |
| 1923 | .output_width_stride(4) |
| 1924 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1925 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1926 | } |
| 1927 | } |
| 1928 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1929 | TEST(F32_DWCONV_CHW_3X3P1__SCALAR, chw_layout) { |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1930 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1931 | DWConvCHWMicrokernelTester() |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1932 | .input_tuple_size(1) |
| 1933 | .output_tuple_size(1) |
| 1934 | .input_width(input_width) |
| 1935 | .input_width_stride(input_width) |
| 1936 | .padding_left(1) |
| 1937 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1938 | .padding_top(1) |
| 1939 | .padding_bottom(1) |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1940 | .kernel_height(3) |
| 1941 | .kernel_width(3) |
| 1942 | .output_height(5) |
| 1943 | .output_width_stride(input_width) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1944 | .Test(xnn_f32_dwconv_chw_ukernel_3x3p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 0cc2c53 | 2019-10-15 04:44:18 -0700 | [diff] [blame] | 1945 | } |
| 1946 | } |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1947 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1948 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, input_width_eq_1_pad0) { |
| 1949 | DWConvCHWMicrokernelTester() |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1950 | .input_tuple_size(1) |
| 1951 | .output_tuple_size(1) |
| 1952 | .input_width(1) |
| 1953 | .padding_left(1) |
| 1954 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1955 | .padding_top(0) |
| 1956 | .padding_bottom(1) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1957 | .kernel_height(3) |
| 1958 | .kernel_width(3) |
| 1959 | .subsampling(2) |
| 1960 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1961 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1962 | } |
| 1963 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1964 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, input_width_gt_1_pad0) { |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1965 | for (size_t input_width = 2; input_width < 32; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1966 | DWConvCHWMicrokernelTester() |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1967 | .input_tuple_size(1) |
| 1968 | .output_tuple_size(1) |
| 1969 | .input_width(input_width) |
| 1970 | .padding_left(1) |
| 1971 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1972 | .padding_top(0) |
| 1973 | .padding_bottom(1) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1974 | .kernel_height(3) |
| 1975 | .kernel_width(3) |
| 1976 | .subsampling(2) |
| 1977 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1978 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1979 | } |
| 1980 | } |
| 1981 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1982 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, input_width_stride_pad0) { |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1983 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1984 | DWConvCHWMicrokernelTester() |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1985 | .input_tuple_size(1) |
| 1986 | .output_tuple_size(1) |
| 1987 | .input_width(input_width) |
| 1988 | .input_width_stride(36) |
| 1989 | .padding_left(1) |
| 1990 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 1991 | .padding_top(0) |
| 1992 | .padding_bottom(1) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1993 | .kernel_height(3) |
| 1994 | .kernel_width(3) |
| 1995 | .subsampling(2) |
| 1996 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 1997 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 1998 | } |
| 1999 | } |
| 2000 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2001 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, input_tuple_stride_pad0) { |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2002 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2003 | DWConvCHWMicrokernelTester() |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2004 | .input_tuple_size(1) |
| 2005 | .output_tuple_size(1) |
| 2006 | .input_width(input_width) |
| 2007 | .input_width_stride(4) |
| 2008 | .input_tuple_stride(3 * 4) |
| 2009 | .padding_left(1) |
| 2010 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2011 | .padding_top(0) |
| 2012 | .padding_bottom(1) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2013 | .kernel_height(3) |
| 2014 | .kernel_width(3) |
| 2015 | .subsampling(2) |
| 2016 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2017 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2018 | } |
| 2019 | } |
| 2020 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2021 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, output_height_gt_1_pad0) { |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2022 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 2023 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2024 | DWConvCHWMicrokernelTester() |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2025 | .input_tuple_size(1) |
| 2026 | .output_tuple_size(1) |
| 2027 | .input_width(input_width) |
| 2028 | .padding_left(1) |
| 2029 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2030 | .padding_top(0) |
| 2031 | .padding_bottom(1) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2032 | .kernel_height(3) |
| 2033 | .kernel_width(3) |
| 2034 | .subsampling(2) |
| 2035 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2036 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2037 | } |
| 2038 | } |
| 2039 | } |
| 2040 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2041 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, output_width_stride_pad0) { |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2042 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2043 | DWConvCHWMicrokernelTester() |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2044 | .input_tuple_size(1) |
| 2045 | .output_tuple_size(1) |
| 2046 | .input_width(input_width) |
| 2047 | .padding_left(1) |
| 2048 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2049 | .padding_top(0) |
| 2050 | .padding_bottom(1) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2051 | .kernel_height(3) |
| 2052 | .kernel_width(3) |
| 2053 | .subsampling(2) |
| 2054 | .output_height(5) |
| 2055 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2056 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2057 | } |
| 2058 | } |
| 2059 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2060 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, output_tuple_stride_pad0) { |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2061 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2062 | DWConvCHWMicrokernelTester() |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2063 | .input_tuple_size(1) |
| 2064 | .output_tuple_size(1) |
| 2065 | .input_width(input_width) |
| 2066 | .padding_left(1) |
| 2067 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2068 | .padding_top(0) |
| 2069 | .padding_bottom(1) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2070 | .kernel_height(3) |
| 2071 | .kernel_width(3) |
| 2072 | .subsampling(2) |
| 2073 | .output_height(5) |
| 2074 | .output_width_stride(4) |
| 2075 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2076 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2077 | } |
| 2078 | } |
| 2079 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2080 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, chw_layout_pad0) { |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2081 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2082 | DWConvCHWMicrokernelTester() |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2083 | .input_tuple_size(1) |
| 2084 | .output_tuple_size(1) |
| 2085 | .input_width(input_width) |
| 2086 | .input_width_stride(input_width) |
| 2087 | .padding_left(1) |
| 2088 | .padding_right(1) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2089 | .padding_top(0) |
| 2090 | .padding_bottom(1) |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2091 | .kernel_height(3) |
| 2092 | .kernel_width(3) |
| 2093 | .subsampling(2) |
| 2094 | .output_height(5) |
| 2095 | .output_width_stride(input_width) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2096 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | ac4de80 | 2019-10-16 04:35:30 -0700 | [diff] [blame] | 2097 | } |
| 2098 | } |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2099 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2100 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, input_width_eq_1_pad1) { |
| 2101 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2102 | .input_tuple_size(1) |
| 2103 | .output_tuple_size(1) |
| 2104 | .input_width(1) |
| 2105 | .padding_left(1) |
| 2106 | .padding_right(1) |
| 2107 | .padding_top(1) |
| 2108 | .padding_bottom(1) |
| 2109 | .kernel_height(3) |
| 2110 | .kernel_width(3) |
| 2111 | .subsampling(2) |
| 2112 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2113 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2114 | } |
| 2115 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2116 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, input_width_gt_1_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2117 | for (size_t input_width = 2; input_width < 32; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2118 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2119 | .input_tuple_size(1) |
| 2120 | .output_tuple_size(1) |
| 2121 | .input_width(input_width) |
| 2122 | .padding_left(1) |
| 2123 | .padding_right(1) |
| 2124 | .padding_top(1) |
| 2125 | .padding_bottom(1) |
| 2126 | .kernel_height(3) |
| 2127 | .kernel_width(3) |
| 2128 | .subsampling(2) |
| 2129 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2130 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2131 | } |
| 2132 | } |
| 2133 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2134 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, input_width_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2135 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2136 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2137 | .input_tuple_size(1) |
| 2138 | .output_tuple_size(1) |
| 2139 | .input_width(input_width) |
| 2140 | .input_width_stride(36) |
| 2141 | .padding_left(1) |
| 2142 | .padding_right(1) |
| 2143 | .padding_top(1) |
| 2144 | .padding_bottom(1) |
| 2145 | .kernel_height(3) |
| 2146 | .kernel_width(3) |
| 2147 | .subsampling(2) |
| 2148 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2149 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2150 | } |
| 2151 | } |
| 2152 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2153 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, input_tuple_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2154 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2155 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2156 | .input_tuple_size(1) |
| 2157 | .output_tuple_size(1) |
| 2158 | .input_width(input_width) |
| 2159 | .input_width_stride(4) |
| 2160 | .input_tuple_stride(3 * 4) |
| 2161 | .padding_left(1) |
| 2162 | .padding_right(1) |
| 2163 | .padding_top(1) |
| 2164 | .padding_bottom(1) |
| 2165 | .kernel_height(3) |
| 2166 | .kernel_width(3) |
| 2167 | .subsampling(2) |
| 2168 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2169 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2170 | } |
| 2171 | } |
| 2172 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2173 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, output_height_gt_1_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2174 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 2175 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2176 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2177 | .input_tuple_size(1) |
| 2178 | .output_tuple_size(1) |
| 2179 | .input_width(input_width) |
| 2180 | .padding_left(1) |
| 2181 | .padding_right(1) |
| 2182 | .padding_top(1) |
| 2183 | .padding_bottom(1) |
| 2184 | .kernel_height(3) |
| 2185 | .kernel_width(3) |
| 2186 | .subsampling(2) |
| 2187 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2188 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2189 | } |
| 2190 | } |
| 2191 | } |
| 2192 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2193 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, output_width_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2194 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2195 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2196 | .input_tuple_size(1) |
| 2197 | .output_tuple_size(1) |
| 2198 | .input_width(input_width) |
| 2199 | .padding_left(1) |
| 2200 | .padding_right(1) |
| 2201 | .padding_top(1) |
| 2202 | .padding_bottom(1) |
| 2203 | .kernel_height(3) |
| 2204 | .kernel_width(3) |
| 2205 | .subsampling(2) |
| 2206 | .output_height(5) |
| 2207 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2208 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2209 | } |
| 2210 | } |
| 2211 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2212 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, output_tuple_stride_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2213 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2214 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2215 | .input_tuple_size(1) |
| 2216 | .output_tuple_size(1) |
| 2217 | .input_width(input_width) |
| 2218 | .padding_left(1) |
| 2219 | .padding_right(1) |
| 2220 | .padding_top(1) |
| 2221 | .padding_bottom(1) |
| 2222 | .kernel_height(3) |
| 2223 | .kernel_width(3) |
| 2224 | .subsampling(2) |
| 2225 | .output_height(5) |
| 2226 | .output_width_stride(4) |
| 2227 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2228 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2229 | } |
| 2230 | } |
| 2231 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2232 | TEST(F32_DWCONV_CHW_3X3S2P1__SCALAR, chw_layout_pad1) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2233 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2234 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2235 | .input_tuple_size(1) |
| 2236 | .output_tuple_size(1) |
| 2237 | .input_width(input_width) |
| 2238 | .input_width_stride(input_width) |
| 2239 | .padding_left(1) |
| 2240 | .padding_right(1) |
| 2241 | .padding_top(1) |
| 2242 | .padding_bottom(1) |
| 2243 | .kernel_height(3) |
| 2244 | .kernel_width(3) |
| 2245 | .subsampling(2) |
| 2246 | .output_height(5) |
| 2247 | .output_width_stride(input_width) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2248 | .Test(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2249 | } |
| 2250 | } |
| 2251 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2252 | TEST(F32_DWCONV_CHW_5X5P2__SCALAR, input_width_eq_1_pad2) { |
| 2253 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2254 | .input_tuple_size(1) |
| 2255 | .output_tuple_size(1) |
| 2256 | .input_width(1) |
| 2257 | .padding_left(2) |
| 2258 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2259 | .padding_top(2) |
| 2260 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2261 | .kernel_height(5) |
| 2262 | .kernel_width(5) |
| 2263 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2264 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2265 | } |
| 2266 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2267 | TEST(F32_DWCONV_CHW_5X5P2__SCALAR, input_width_gt_1_pad2) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2268 | for (size_t input_width = 2; input_width < 32; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2269 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2270 | .input_tuple_size(1) |
| 2271 | .output_tuple_size(1) |
| 2272 | .input_width(input_width) |
| 2273 | .padding_left(2) |
| 2274 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2275 | .padding_top(2) |
| 2276 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2277 | .kernel_height(5) |
| 2278 | .kernel_width(5) |
| 2279 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2280 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2281 | } |
| 2282 | } |
| 2283 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2284 | TEST(F32_DWCONV_CHW_5X5P2__SCALAR, input_width_stride_pad2) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2285 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2286 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2287 | .input_tuple_size(1) |
| 2288 | .output_tuple_size(1) |
| 2289 | .input_width(input_width) |
| 2290 | .input_width_stride(36) |
| 2291 | .padding_left(2) |
| 2292 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2293 | .padding_top(2) |
| 2294 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2295 | .kernel_height(5) |
| 2296 | .kernel_width(5) |
| 2297 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2298 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2299 | } |
| 2300 | } |
| 2301 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2302 | TEST(F32_DWCONV_CHW_5X5P2__SCALAR, input_tuple_stride_pad2) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2303 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2304 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2305 | .input_tuple_size(1) |
| 2306 | .output_tuple_size(1) |
| 2307 | .input_width(input_width) |
| 2308 | .input_width_stride(4) |
| 2309 | .input_tuple_stride(3 * 4) |
| 2310 | .padding_left(2) |
| 2311 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2312 | .padding_top(2) |
| 2313 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2314 | .kernel_height(5) |
| 2315 | .kernel_width(5) |
| 2316 | .output_height(1) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2317 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2318 | } |
| 2319 | } |
| 2320 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2321 | TEST(F32_DWCONV_CHW_5X5P2__SCALAR, output_height_gt_1_pad2) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2322 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 2323 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2324 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2325 | .input_tuple_size(1) |
| 2326 | .output_tuple_size(1) |
| 2327 | .input_width(input_width) |
| 2328 | .padding_left(2) |
| 2329 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2330 | .padding_top(2) |
| 2331 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2332 | .kernel_height(5) |
| 2333 | .kernel_width(5) |
| 2334 | .output_height(output_height) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2335 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2336 | } |
| 2337 | } |
| 2338 | } |
| 2339 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2340 | TEST(F32_DWCONV_CHW_5X5P2__SCALAR, output_width_stride_pad2) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2341 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2342 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2343 | .input_tuple_size(1) |
| 2344 | .output_tuple_size(1) |
| 2345 | .input_width(input_width) |
| 2346 | .padding_left(2) |
| 2347 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2348 | .padding_top(2) |
| 2349 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2350 | .kernel_height(5) |
| 2351 | .kernel_width(5) |
| 2352 | .output_height(5) |
| 2353 | .output_width_stride(36) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2354 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2355 | } |
| 2356 | } |
| 2357 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2358 | TEST(F32_DWCONV_CHW_5X5P2__SCALAR, output_tuple_stride_pad2) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2359 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2360 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2361 | .input_tuple_size(1) |
| 2362 | .output_tuple_size(1) |
| 2363 | .input_width(input_width) |
| 2364 | .padding_left(2) |
| 2365 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2366 | .padding_top(2) |
| 2367 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2368 | .kernel_height(5) |
| 2369 | .kernel_width(5) |
| 2370 | .output_height(5) |
| 2371 | .output_width_stride(4) |
| 2372 | .output_tuple_stride(5 * 4) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2373 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2374 | } |
| 2375 | } |
| 2376 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2377 | TEST(F32_DWCONV_CHW_5X5P2__SCALAR, chw_layout_pad2) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2378 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2379 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2380 | .input_tuple_size(1) |
| 2381 | .output_tuple_size(1) |
| 2382 | .input_width(input_width) |
| 2383 | .input_width_stride(input_width) |
| 2384 | .padding_left(2) |
| 2385 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2386 | .padding_top(2) |
| 2387 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2388 | .kernel_height(5) |
| 2389 | .kernel_width(5) |
| 2390 | .output_height(5) |
| 2391 | .output_width_stride(input_width) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2392 | .Test(xnn_f32_dwconv_chw_ukernel_5x5p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2393 | } |
| 2394 | } |
| 2395 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2396 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, input_width_eq_1_pad1) { |
| 2397 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2398 | .input_tuple_size(1) |
| 2399 | .output_tuple_size(1) |
| 2400 | .input_width(1) |
| 2401 | .padding_left(2) |
| 2402 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2403 | .padding_top(2) |
| 2404 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2405 | .kernel_height(5) |
| 2406 | .kernel_width(5) |
| 2407 | .output_height(1) |
| 2408 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2409 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2410 | } |
| 2411 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2412 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, input_width_gt_1_pad1) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2413 | for (size_t input_width = 2; input_width < 32; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2414 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2415 | .input_tuple_size(1) |
| 2416 | .output_tuple_size(1) |
| 2417 | .input_width(input_width) |
| 2418 | .padding_left(2) |
| 2419 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2420 | .padding_top(2) |
| 2421 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2422 | .kernel_height(5) |
| 2423 | .kernel_width(5) |
| 2424 | .output_height(1) |
| 2425 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2426 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2427 | } |
| 2428 | } |
| 2429 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2430 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, input_width_stride_pad1) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2431 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2432 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2433 | .input_tuple_size(1) |
| 2434 | .output_tuple_size(1) |
| 2435 | .input_width(input_width) |
| 2436 | .input_width_stride(36) |
| 2437 | .padding_left(2) |
| 2438 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2439 | .padding_top(2) |
| 2440 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2441 | .kernel_height(5) |
| 2442 | .kernel_width(5) |
| 2443 | .output_height(1) |
| 2444 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2445 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2446 | } |
| 2447 | } |
| 2448 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2449 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, input_tuple_stride_pad1) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2450 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2451 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2452 | .input_tuple_size(1) |
| 2453 | .output_tuple_size(1) |
| 2454 | .input_width(input_width) |
| 2455 | .input_width_stride(4) |
| 2456 | .input_tuple_stride(3 * 4) |
| 2457 | .padding_left(2) |
| 2458 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2459 | .padding_top(2) |
| 2460 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2461 | .kernel_height(5) |
| 2462 | .kernel_width(5) |
| 2463 | .output_height(1) |
| 2464 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2465 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2466 | } |
| 2467 | } |
| 2468 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2469 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, output_height_gt_1_pad1) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2470 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 2471 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2472 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2473 | .input_tuple_size(1) |
| 2474 | .output_tuple_size(1) |
| 2475 | .input_width(input_width) |
| 2476 | .padding_left(2) |
| 2477 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2478 | .padding_top(2) |
| 2479 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2480 | .kernel_height(5) |
| 2481 | .kernel_width(5) |
| 2482 | .output_height(output_height) |
| 2483 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2484 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2485 | } |
| 2486 | } |
| 2487 | } |
| 2488 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2489 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, output_width_stride_pad1) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2490 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2491 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2492 | .input_tuple_size(1) |
| 2493 | .output_tuple_size(1) |
| 2494 | .input_width(input_width) |
| 2495 | .padding_left(2) |
| 2496 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2497 | .padding_top(2) |
| 2498 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2499 | .kernel_height(5) |
| 2500 | .kernel_width(5) |
| 2501 | .output_height(5) |
| 2502 | .output_width_stride(36) |
| 2503 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2504 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2505 | } |
| 2506 | } |
| 2507 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2508 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, output_tuple_stride_pad1) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2509 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2510 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2511 | .input_tuple_size(1) |
| 2512 | .output_tuple_size(1) |
| 2513 | .input_width(input_width) |
| 2514 | .padding_left(2) |
| 2515 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2516 | .padding_top(2) |
| 2517 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2518 | .kernel_height(5) |
| 2519 | .kernel_width(5) |
| 2520 | .output_height(5) |
| 2521 | .output_width_stride(4) |
| 2522 | .output_tuple_stride(5 * 4) |
| 2523 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2524 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2525 | } |
| 2526 | } |
| 2527 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2528 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, chw_layout_pad1) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2529 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2530 | DWConvCHWMicrokernelTester() |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2531 | .input_tuple_size(1) |
| 2532 | .output_tuple_size(1) |
| 2533 | .input_width(input_width) |
| 2534 | .input_width_stride(input_width) |
| 2535 | .padding_left(2) |
| 2536 | .padding_right(2) |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2537 | .padding_top(2) |
| 2538 | .padding_bottom(2) |
| 2539 | .kernel_height(5) |
| 2540 | .kernel_width(5) |
| 2541 | .output_height(5) |
| 2542 | .output_width_stride(input_width) |
| 2543 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2544 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2545 | } |
| 2546 | } |
| 2547 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2548 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, input_width_eq_1_pad2) { |
| 2549 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2550 | .input_tuple_size(1) |
| 2551 | .output_tuple_size(1) |
| 2552 | .input_width(1) |
| 2553 | .padding_left(2) |
| 2554 | .padding_right(2) |
| 2555 | .padding_top(2) |
| 2556 | .padding_bottom(2) |
| 2557 | .kernel_height(5) |
| 2558 | .kernel_width(5) |
| 2559 | .output_height(1) |
| 2560 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2561 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2562 | } |
| 2563 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2564 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, input_width_gt_1_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2565 | for (size_t input_width = 2; input_width < 32; input_width++) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2566 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2567 | .input_tuple_size(1) |
| 2568 | .output_tuple_size(1) |
| 2569 | .input_width(input_width) |
| 2570 | .padding_left(2) |
| 2571 | .padding_right(2) |
| 2572 | .padding_top(2) |
| 2573 | .padding_bottom(2) |
| 2574 | .kernel_height(5) |
| 2575 | .kernel_width(5) |
| 2576 | .output_height(1) |
| 2577 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2578 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2579 | } |
| 2580 | } |
| 2581 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2582 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, input_width_stride_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2583 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2584 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2585 | .input_tuple_size(1) |
| 2586 | .output_tuple_size(1) |
| 2587 | .input_width(input_width) |
| 2588 | .input_width_stride(36) |
| 2589 | .padding_left(2) |
| 2590 | .padding_right(2) |
| 2591 | .padding_top(2) |
| 2592 | .padding_bottom(2) |
| 2593 | .kernel_height(5) |
| 2594 | .kernel_width(5) |
| 2595 | .output_height(1) |
| 2596 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2597 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2598 | } |
| 2599 | } |
| 2600 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2601 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, input_tuple_stride_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2602 | for (size_t input_width = 1; input_width < 32; input_width += 5) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2603 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2604 | .input_tuple_size(1) |
| 2605 | .output_tuple_size(1) |
| 2606 | .input_width(input_width) |
| 2607 | .input_width_stride(4) |
| 2608 | .input_tuple_stride(3 * 4) |
| 2609 | .padding_left(2) |
| 2610 | .padding_right(2) |
| 2611 | .padding_top(2) |
| 2612 | .padding_bottom(2) |
| 2613 | .kernel_height(5) |
| 2614 | .kernel_width(5) |
| 2615 | .output_height(1) |
| 2616 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2617 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2618 | } |
| 2619 | } |
| 2620 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2621 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, output_height_gt_1_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2622 | for (size_t output_height = 2; output_height < 5; output_height++) { |
| 2623 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2624 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2625 | .input_tuple_size(1) |
| 2626 | .output_tuple_size(1) |
| 2627 | .input_width(input_width) |
| 2628 | .padding_left(2) |
| 2629 | .padding_right(2) |
| 2630 | .padding_top(2) |
| 2631 | .padding_bottom(2) |
| 2632 | .kernel_height(5) |
| 2633 | .kernel_width(5) |
| 2634 | .output_height(output_height) |
| 2635 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2636 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2637 | } |
| 2638 | } |
| 2639 | } |
| 2640 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2641 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, output_width_stride_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2642 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2643 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2644 | .input_tuple_size(1) |
| 2645 | .output_tuple_size(1) |
| 2646 | .input_width(input_width) |
| 2647 | .padding_left(2) |
| 2648 | .padding_right(2) |
| 2649 | .padding_top(2) |
| 2650 | .padding_bottom(2) |
| 2651 | .kernel_height(5) |
| 2652 | .kernel_width(5) |
| 2653 | .output_height(5) |
| 2654 | .output_width_stride(36) |
| 2655 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2656 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2657 | } |
| 2658 | } |
| 2659 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2660 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, output_tuple_stride_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2661 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2662 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2663 | .input_tuple_size(1) |
| 2664 | .output_tuple_size(1) |
| 2665 | .input_width(input_width) |
| 2666 | .padding_left(2) |
| 2667 | .padding_right(2) |
| 2668 | .padding_top(2) |
| 2669 | .padding_bottom(2) |
| 2670 | .kernel_height(5) |
| 2671 | .kernel_width(5) |
| 2672 | .output_height(5) |
| 2673 | .output_width_stride(4) |
| 2674 | .output_tuple_stride(5 * 4) |
| 2675 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2676 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2677 | } |
| 2678 | } |
| 2679 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2680 | TEST(F32_DWCONV_CHW_5X5S2P2__SCALAR, chw_layout_pad2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2681 | for (size_t input_width = 1; input_width < 32; input_width += 3) { |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2682 | DWConvCHWMicrokernelTester() |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 2683 | .input_tuple_size(1) |
| 2684 | .output_tuple_size(1) |
| 2685 | .input_width(input_width) |
| 2686 | .input_width_stride(input_width) |
| 2687 | .padding_left(2) |
| 2688 | .padding_right(2) |
| 2689 | .padding_top(2) |
| 2690 | .padding_bottom(2) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2691 | .kernel_height(5) |
| 2692 | .kernel_width(5) |
| 2693 | .output_height(5) |
| 2694 | .output_width_stride(input_width) |
| 2695 | .subsampling(2) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame^] | 2696 | .Test(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar, DWConvCHWMicrokernelTester::Variant::Scalar); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 2697 | } |
| 2698 | } |