• Home
  • Raw
  • Download

Lines Matching refs:vget_low_s16

76         vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk00), vget_low_s16(vxi00));  in xnn_q8_dwconv_ukernel_up8x9__neon()
78 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk00), vget_low_s16(vxi01)); in xnn_q8_dwconv_ukernel_up8x9__neon()
80 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk00), vget_low_s16(vxi02)); in xnn_q8_dwconv_ukernel_up8x9__neon()
91 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk10), vget_low_s16(vxi10)); in xnn_q8_dwconv_ukernel_up8x9__neon()
93 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk10), vget_low_s16(vxi11)); in xnn_q8_dwconv_ukernel_up8x9__neon()
95 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk10), vget_low_s16(vxi12)); in xnn_q8_dwconv_ukernel_up8x9__neon()
106 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk20), vget_low_s16(vxi20)); in xnn_q8_dwconv_ukernel_up8x9__neon()
108 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk20), vget_low_s16(vxi21)); in xnn_q8_dwconv_ukernel_up8x9__neon()
110 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk20), vget_low_s16(vxi22)); in xnn_q8_dwconv_ukernel_up8x9__neon()
117 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk01), vget_low_s16(vxi01)); in xnn_q8_dwconv_ukernel_up8x9__neon()
119 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk01), vget_low_s16(vxi02)); in xnn_q8_dwconv_ukernel_up8x9__neon()
121 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk01), vget_low_s16(vxi03)); in xnn_q8_dwconv_ukernel_up8x9__neon()
128 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk11), vget_low_s16(vxi11)); in xnn_q8_dwconv_ukernel_up8x9__neon()
130 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk11), vget_low_s16(vxi12)); in xnn_q8_dwconv_ukernel_up8x9__neon()
132 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk11), vget_low_s16(vxi13)); in xnn_q8_dwconv_ukernel_up8x9__neon()
139 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk21), vget_low_s16(vxi21)); in xnn_q8_dwconv_ukernel_up8x9__neon()
141 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk21), vget_low_s16(vxi22)); in xnn_q8_dwconv_ukernel_up8x9__neon()
143 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk21), vget_low_s16(vxi23)); in xnn_q8_dwconv_ukernel_up8x9__neon()
150 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk02), vget_low_s16(vxi02)); in xnn_q8_dwconv_ukernel_up8x9__neon()
152 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk02), vget_low_s16(vxi03)); in xnn_q8_dwconv_ukernel_up8x9__neon()
154 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk02), vget_low_s16(vxi04)); in xnn_q8_dwconv_ukernel_up8x9__neon()
161 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk12), vget_low_s16(vxi12)); in xnn_q8_dwconv_ukernel_up8x9__neon()
163 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk12), vget_low_s16(vxi13)); in xnn_q8_dwconv_ukernel_up8x9__neon()
165 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk12), vget_low_s16(vxi14)); in xnn_q8_dwconv_ukernel_up8x9__neon()
172 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk22), vget_low_s16(vxi22)); in xnn_q8_dwconv_ukernel_up8x9__neon()
174 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk22), vget_low_s16(vxi23)); in xnn_q8_dwconv_ukernel_up8x9__neon()
176 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk22), vget_low_s16(vxi24)); in xnn_q8_dwconv_ukernel_up8x9__neon()
234 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk00), vget_low_s16(vxi00)); in xnn_q8_dwconv_ukernel_up8x9__neon()
236 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk00), vget_low_s16(vxi01)); in xnn_q8_dwconv_ukernel_up8x9__neon()
238 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk00), vget_low_s16(vxi02)); in xnn_q8_dwconv_ukernel_up8x9__neon()
249 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk10), vget_low_s16(vxi10)); in xnn_q8_dwconv_ukernel_up8x9__neon()
251 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk10), vget_low_s16(vxi11)); in xnn_q8_dwconv_ukernel_up8x9__neon()
253 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk10), vget_low_s16(vxi12)); in xnn_q8_dwconv_ukernel_up8x9__neon()
264 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk20), vget_low_s16(vxi20)); in xnn_q8_dwconv_ukernel_up8x9__neon()
266 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk20), vget_low_s16(vxi21)); in xnn_q8_dwconv_ukernel_up8x9__neon()
268 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk20), vget_low_s16(vxi22)); in xnn_q8_dwconv_ukernel_up8x9__neon()
275 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk01), vget_low_s16(vxi01)); in xnn_q8_dwconv_ukernel_up8x9__neon()
277 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk01), vget_low_s16(vxi02)); in xnn_q8_dwconv_ukernel_up8x9__neon()
279 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk01), vget_low_s16(vxi03)); in xnn_q8_dwconv_ukernel_up8x9__neon()
286 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk11), vget_low_s16(vxi11)); in xnn_q8_dwconv_ukernel_up8x9__neon()
288 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk11), vget_low_s16(vxi12)); in xnn_q8_dwconv_ukernel_up8x9__neon()
290 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk11), vget_low_s16(vxi13)); in xnn_q8_dwconv_ukernel_up8x9__neon()
297 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk21), vget_low_s16(vxi21)); in xnn_q8_dwconv_ukernel_up8x9__neon()
299 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk21), vget_low_s16(vxi22)); in xnn_q8_dwconv_ukernel_up8x9__neon()
301 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk21), vget_low_s16(vxi23)); in xnn_q8_dwconv_ukernel_up8x9__neon()
308 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk02), vget_low_s16(vxi02)); in xnn_q8_dwconv_ukernel_up8x9__neon()
310 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk02), vget_low_s16(vxi03)); in xnn_q8_dwconv_ukernel_up8x9__neon()
312 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk02), vget_low_s16(vxi04)); in xnn_q8_dwconv_ukernel_up8x9__neon()
319 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk12), vget_low_s16(vxi12)); in xnn_q8_dwconv_ukernel_up8x9__neon()
321 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk12), vget_low_s16(vxi13)); in xnn_q8_dwconv_ukernel_up8x9__neon()
323 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk12), vget_low_s16(vxi14)); in xnn_q8_dwconv_ukernel_up8x9__neon()
330 vacc0_lo = vmlal_s16(vacc0_lo, vget_low_s16(vxk22), vget_low_s16(vxi22)); in xnn_q8_dwconv_ukernel_up8x9__neon()
332 vacc1_lo = vmlal_s16(vacc1_lo, vget_low_s16(vxk22), vget_low_s16(vxi23)); in xnn_q8_dwconv_ukernel_up8x9__neon()
334 vacc2_lo = vmlal_s16(vacc2_lo, vget_low_s16(vxk22), vget_low_s16(vxi24)); in xnn_q8_dwconv_ukernel_up8x9__neon()
426 int32x4_t vaccX0_lo = vmull_s16(vget_low_s16(vxk0), vget_low_s16(vxi0)); in xnn_q8_dwconv_ukernel_up8x9__neon()
433 vaccX1_lo = vmlal_s16(vaccX1_lo, vget_low_s16(vxk1), vget_low_s16(vxi1)); in xnn_q8_dwconv_ukernel_up8x9__neon()
440 vaccX0_lo = vmlal_s16(vaccX0_lo, vget_low_s16(vxk2), vget_low_s16(vxi2)); in xnn_q8_dwconv_ukernel_up8x9__neon()
447 vaccX1_lo = vmlal_s16(vaccX1_lo, vget_low_s16(vxk3), vget_low_s16(vxi3)); in xnn_q8_dwconv_ukernel_up8x9__neon()
454 vaccX0_lo = vmlal_s16(vaccX0_lo, vget_low_s16(vxk4), vget_low_s16(vxi4)); in xnn_q8_dwconv_ukernel_up8x9__neon()
461 vaccX1_lo = vmlal_s16(vaccX1_lo, vget_low_s16(vxk5), vget_low_s16(vxi5)); in xnn_q8_dwconv_ukernel_up8x9__neon()
468 vaccX0_lo = vmlal_s16(vaccX0_lo, vget_low_s16(vxk6), vget_low_s16(vxi6)); in xnn_q8_dwconv_ukernel_up8x9__neon()
475 vaccX1_lo = vmlal_s16(vaccX1_lo, vget_low_s16(vxk7), vget_low_s16(vxi7)); in xnn_q8_dwconv_ukernel_up8x9__neon()
482 vaccX0_lo = vmlal_s16(vaccX0_lo, vget_low_s16(vxk8), vget_low_s16(vxi8)); in xnn_q8_dwconv_ukernel_up8x9__neon()
517 int32x4_t vaccX0_lo = vmull_s16(vget_low_s16(vxk0), vget_low_s16(vxi0)); in xnn_q8_dwconv_ukernel_up8x9__neon()
524 vaccX1_lo = vmlal_s16(vaccX1_lo, vget_low_s16(vxk1), vget_low_s16(vxi1)); in xnn_q8_dwconv_ukernel_up8x9__neon()
531 vaccX0_lo = vmlal_s16(vaccX0_lo, vget_low_s16(vxk2), vget_low_s16(vxi2)); in xnn_q8_dwconv_ukernel_up8x9__neon()
538 vaccX1_lo = vmlal_s16(vaccX1_lo, vget_low_s16(vxk3), vget_low_s16(vxi3)); in xnn_q8_dwconv_ukernel_up8x9__neon()
545 vaccX0_lo = vmlal_s16(vaccX0_lo, vget_low_s16(vxk4), vget_low_s16(vxi4)); in xnn_q8_dwconv_ukernel_up8x9__neon()
552 vaccX1_lo = vmlal_s16(vaccX1_lo, vget_low_s16(vxk5), vget_low_s16(vxi5)); in xnn_q8_dwconv_ukernel_up8x9__neon()
559 vaccX0_lo = vmlal_s16(vaccX0_lo, vget_low_s16(vxk6), vget_low_s16(vxi6)); in xnn_q8_dwconv_ukernel_up8x9__neon()
566 vaccX1_lo = vmlal_s16(vaccX1_lo, vget_low_s16(vxk7), vget_low_s16(vxi7)); in xnn_q8_dwconv_ukernel_up8x9__neon()
573 vaccX0_lo = vmlal_s16(vaccX0_lo, vget_low_s16(vxk8), vget_low_s16(vxi8)); in xnn_q8_dwconv_ukernel_up8x9__neon()