167       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()  local
177       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
195       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
213       vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
231       vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
249       vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
267       vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
285       vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
303       vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
321       vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
339       vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
357       vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
375       vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
393       vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
399       vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
402       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
412       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()  local
418       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
430       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
442       vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
454       vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
466       vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
478       vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
490       vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
502       vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
514       vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
526       vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
538       vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
550       vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
562       vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
567       vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
569       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
580       __m256 vacc01234567p0 = _mm256_load_ps(w);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()  local
584       vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
592       vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
600       vacc01234567p0 = _mm256_fmadd_ps(vi4x01234567, vk4x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
608       vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
616       vacc01234567p0 = _mm256_fmadd_ps(vi8x01234567, vk8x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
624       vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
632       vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
640       vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
648       vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
656       vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
664       vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
672       vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
680       vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
683       vacc01234567p0 = _mm256_add_ps(vacc01234567p0, vacc01234567p1);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
685       __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);  in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()