| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-scalar-4x1.c | 78 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
101 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
104 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
107 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
114 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
173 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
176 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
179 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
|
| D | 3x3p1-minmax-scalar-5x1.c | 84 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
110 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
114 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
118 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
126 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
196 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
200 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
204 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
|
| D | 5x5p2-minmax-scalar-3x1.c | 93 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
138 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
140 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
142 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
152 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
265 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
267 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
269 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
279 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
370 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
[all …]
|
| D | 5x5p2-minmax-scalar-2x1.c | 87 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
127 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
128 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
136 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
222 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
223 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
231 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
301 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
302 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 3x3p1-minmax-scalar-3x1.c | 72 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1() local
94 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
96 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
102 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
152 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
154 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 93 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
138 float vo2p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
140 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
142 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
152 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
268 float vo2p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
270 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
272 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
282 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
376 float vo2p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
[all …]
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 87 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
127 float vo1p2 = vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
128 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
136 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
226 float vo1p2 = vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
227 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
235 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
309 float vo1p2 = vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
310 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 87 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
127 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
128 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
136 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
224 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
225 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
233 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
305 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
306 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
|
| D | 5x5p2-minmax-scalar-1x1-acc4.c | 81 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
114 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
120 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
180 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
186 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
236 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
|
| D | 5x5p2-minmax-scalar-1x1-acc3.c | 81 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
114 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
120 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
179 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
185 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
234 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-scalar-1x1.c | 81 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1() local
114 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
120 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
177 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
183 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
230 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc2.c | 81 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
114 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
120 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
178 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
184 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
232 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
|
| D | 3x3p1-minmax-scalar-6x1.c | 90 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
119 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
124 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
129 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
138 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
219 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
224 vo2p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
229 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc5.c | 81 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
114 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
120 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
181 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
187 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
238 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 100 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
149 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
152 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
160 vi3x0 = vi3x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
251 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
254 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
309 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
312 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc2.c | 100 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
149 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
152 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
160 vi3x0 = vi3x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
249 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
252 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
305 float vo1p1 = vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
308 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-2x1.c | 100 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
149 vo1p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
152 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
160 vi3x0 = vi3x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
247 vo1p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
250 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
301 vo1p0 += vi3x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
304 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
|
| D | 3x3p1-minmax-scalar-2x1-acc2.c | 66 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
85 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
90 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
131 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
|
| D | 3x3p1-minmax-scalar-2x1.c | 66 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1() local
85 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
90 vi3x0 = vi3x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
129 vo1p0 += vi3x0 * vk20; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc4.c | 89 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
126 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
132 vi3x0 = vi3x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
191 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
225 float vo0p3 = vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
|
| D | 5x5s2p2-minmax-scalar-1x1.c | 89 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1() local
126 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
132 vi3x0 = vi3x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
188 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
219 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc2.c | 89 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
126 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
132 vi3x0 = vi3x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
189 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
221 vo0p1 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc3.c | 89 float vi3x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
126 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
132 vi3x0 = vi3x2; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
190 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
223 vo0p0 += vi3x0 * vk30; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
|
| /external/XNNPACK/src/f32-conv-hwc/gen/ |
| D | 3x3s2p0p1c3x8-neonfma-2x1.c | 88 float32x4_t vi3x0 = vld1q_f32(i3); i3 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
110 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
112 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c0x4567, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
134 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
136 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c1x4567, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
158 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
160 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
182 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
184 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
331 vi3x0 = vcombine_f32(vget_high_f32(vi3x1), vi3x2); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
[all …]
|
| D | 3x3s2p0p1c3x8-neon-2x1.c | 86 float32x4_t vi3x0 = vld1q_f32(i3); i3 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
108 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
110 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c0x4567, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
132 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
134 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c1x4567, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
156 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
158 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c2x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
180 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
182 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
329 vi3x0 = vcombine_f32(vget_high_f32(vi3x1), vi3x2); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
[all …]
|