397         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
402         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
407         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
412         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
417         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
422         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
427         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
432         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
437         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
442         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
454         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
459         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
464         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
469         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
474         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
479         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
484         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
489         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
494         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
499         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
504         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c0x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
509         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
521         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c1x0123, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
526         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
531         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
536         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
541         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
546         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c2x0123, vi4x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
556         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
559         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
562           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
570           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
573             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
588         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
593         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
598         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
603         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
608         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
613         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
618         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
623         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
628         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
633         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
645         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
650         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
655         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
660         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
665         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
670         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
675         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
680         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
685         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
689         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
692         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
695           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
703           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
706             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()