| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-scalar-5x1.c | 85 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
111 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
115 vo3p0 += vi4x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
119 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
127 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
197 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
201 vo3p0 += vi4x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
205 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
|
| D | 5x5p2-minmax-scalar-3x1.c | 94 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
141 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
143 vo1p0 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
145 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
153 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
268 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
270 vo1p0 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
272 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
280 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
373 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
[all …]
|
| D | 5x5p2-minmax-scalar-2x1.c | 88 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
129 vo1p0 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
130 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
137 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
224 vo1p0 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
225 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
232 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
303 vo1p0 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
304 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 94 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
141 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
143 vo1p1 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
145 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
153 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
271 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
273 vo1p1 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
275 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
283 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
379 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
[all …]
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 88 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
129 vo1p0 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
130 vo0p1 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
137 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
228 vo1p0 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
229 vo0p1 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
236 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
311 vo1p0 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
312 vo0p1 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 88 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
129 vo1p1 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
130 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
137 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
226 vo1p1 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
227 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
234 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
307 vo1p1 += vi4x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
308 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-3x1.c | 111 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local
170 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
175 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
180 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
188 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
304 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
309 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
314 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
381 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
386 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
[all …]
|
| D | 5x5s2p2-minmax-scalar-3x1-acc2.c | 111 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
170 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
175 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
180 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
188 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
307 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
312 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
317 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
387 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
392 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
[all …]
|
| D | 3x3s2p1-minmax-scalar-3x1.c | 88 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1() local
104 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
109 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
141 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
181 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
186 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc4.c | 82 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
115 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
121 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
181 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
187 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
237 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
|
| D | 5x5p2-minmax-scalar-1x1-acc2.c | 82 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
115 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
121 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
179 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
185 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
233 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
|
| D | 5x5p2-minmax-scalar-1x1-acc3.c | 82 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
115 vo0p1 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
121 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
180 vo0p1 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
186 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
235 vo0p1 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-scalar-1x1.c | 82 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1() local
115 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
121 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
178 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
184 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
231 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
|
| D | 3x3p1-minmax-scalar-6x1.c | 91 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
120 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
125 vo3p0 += vi4x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
130 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
139 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
220 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
225 vo3p0 += vi4x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
230 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc5.c | 82 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
115 float vo0p4 = vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
121 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
182 float vo0p4 = vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
188 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
239 float vo0p4 = vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
|
| D | 3x3p1-minmax-scalar-4x1.c | 79 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
105 vo3p0 += vi4x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
108 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
115 vi4x0 = vi4x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
177 vo3p0 += vi4x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
180 vo2p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
|
| D | 5x5s2p2-minmax-scalar-2x1.c | 101 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
151 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
154 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
161 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
249 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
252 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
303 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
306 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc2.c | 101 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
151 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
154 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
161 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
251 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
254 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
307 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
310 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 101 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
151 float vo1p2 = vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
154 vo0p1 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
161 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
253 float vo1p2 = vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
256 vo0p1 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
311 float vo1p2 = vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
314 vo0p1 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
|
| D | 3x3s2p1-minmax-scalar-4x1.c | 98 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1() local
118 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
125 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
165 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
215 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
222 vo1p0 += vi4x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc4.c | 90 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
127 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
133 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
192 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
226 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
|
| D | 5x5s2p2-minmax-scalar-1x1.c | 90 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1() local
127 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
133 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
189 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
220 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc2.c | 90 float vi4x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
127 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
133 vi4x0 = vi4x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
190 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
222 vo0p0 += vi4x0 * vk40; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
|
| /external/XNNPACK/src/f32-conv-hwc/gen/ |
| D | 3x3s2p0p1c3x8-neonfma-2x1.c | 89 float32x4_t vi4x0 = vld1q_f32(i4); i4 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
118 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
120 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
142 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
144 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
166 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
168 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
190 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c0x0123, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
192 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
332 vi4x0 = vcombine_f32(vget_high_f32(vi4x1), vi4x2); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
[all …]
|
| D | 3x3s2p0p1c3x8-neon-2x1.c | 87 float32x4_t vi4x0 = vld1q_f32(i4); i4 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
116 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
118 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
140 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
142 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
164 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
166 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
188 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c0x0123, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
190 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
330 vi4x0 = vcombine_f32(vget_high_f32(vi4x1), vi4x2); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
[all …]
|