| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-scalar-3x1.c | 111 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
178 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
192 vo2p0 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
194 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
202 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
305 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
319 vo2p0 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
321 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
329 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
408 vo2p0 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
[all …]
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 111 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
178 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
192 vo2p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
194 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
202 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
308 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
322 vo2p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
324 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
332 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
414 vo2p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
[all …]
|
| D | 5x5p2-minmax-scalar-2x1.c | 103 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
156 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
167 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
174 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
251 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
262 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
269 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
327 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 103 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
156 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
167 vo1p2 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
174 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
255 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
266 vo1p2 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
273 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
335 vo1p2 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 103 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
156 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
167 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
174 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
253 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
264 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
271 vi5x2 = vi5x3; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
331 vo1p0 += vi5x2 * vk42; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-3x1.c | 132 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local
189 vi5x0 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
225 vo2p0 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
230 vo1p0 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
241 vi5x2 = vi5x4; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
339 vo2p0 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
344 vo1p0 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
416 vo2p0 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
421 vo1p0 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-3x1-acc2.c | 132 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
189 vi5x0 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
225 vo2p1 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
230 vo1p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
241 vi5x2 = vi5x4; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
342 vo2p1 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
347 vo1p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
422 vo2p1 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
427 vo1p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-2x1.c | 118 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
162 vi5x0 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
191 vo1p0 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
200 vi5x2 = vi5x4; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
273 vo1p0 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
327 vo1p0 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc2.c | 118 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
162 vi5x0 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
191 vo1p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
200 vi5x2 = vi5x4; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
275 vo1p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
331 vo1p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 118 float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
162 vi5x0 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
191 vo1p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
200 vi5x2 = vi5x4; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
277 vo1p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
335 vo1p1 += vi5x2 * vk32; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
|
| D | 3x3p1-minmax-scalar-6x1.c | 112 const float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
168 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
177 vo5p0 += vi5x2 * vk02; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
182 vo4p0 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
187 vo3p0 += vi5x2 * vk22; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
|
| D | 3x3p1-minmax-scalar-5x1.c | 104 const float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
152 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
164 vo4p0 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
168 vo3p0 += vi5x2 * vk22; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
|
| D | 3x3s2p1-minmax-scalar-3x1.c | 122 const float vi5x2 = i5[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1() local
142 vi5x0 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
150 vo2p0 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
|
| D | 3x3p1-minmax-scalar-4x1.c | 96 const float vi5x2 = *i5++; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
136 vi5x1 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
149 vo3p0 += vi5x2 * vk22; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
|
| D | 3x3s2p1-minmax-scalar-4x1.c | 139 const float vi5x2 = i5[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1() local
166 vi5x0 = vi5x2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
177 vo2p0 += vi5x2 * vk12; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
|