167       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()  local
175       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
182       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
189       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
196       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
203       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
210       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi5x01234567, vk5x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
217       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
224       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi7x01234567, vk7x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
231       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
238       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
245       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
252       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
259       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
266       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
273       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
280       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
287       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
294       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
301       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
308       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
315       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
322       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
329       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
336       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
343       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
347       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
356       float16x8_t vacc01234567p0 = vld1q_f16(w); w += 8;  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()  local
361       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
365       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
369       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
373       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
377       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
381       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi5x01234567, vk5x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
385       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
389       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi7x01234567, vk7x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
393       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
397       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
401       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
405       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
409       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
413       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
417       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
421       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
425       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
429       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
433       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
437       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
441       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
445       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
449       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
453       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
457       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
460       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
466       float16x8_t vacc01234567p0 = vld1q_f16(w);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()  local
471       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi0x01234567, vk0x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
475       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi1x01234567, vk1x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
479       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi2x01234567, vk2x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
483       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi3x01234567, vk3x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
487       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi4x01234567, vk4x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
491       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi5x01234567, vk5x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
495       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi6x01234567, vk6x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
499       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi7x01234567, vk7x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
503       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi8x01234567, vk8x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
507       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
511       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
515       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
519       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
523       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
527       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
531       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
535       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
539       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
543       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
547       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
551       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
555       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
559       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
563       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
567       vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
570       float16x8_t vacc01234567 = vmaxq_f16(vacc01234567p0, vmin);  in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()