397         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()  local
402         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
407         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
412         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
417         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
422         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
427         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
432         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
437         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
442         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
447         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
452         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
457         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
469         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
474         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
479         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
484         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
489         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
494         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c2x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
499         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
504         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c0x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
509         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
514         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
519         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
524         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
536         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
541         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
546         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c2x0123, vi4x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
558         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
561         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
564           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
572           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
575             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
590         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()  local
595         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
600         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
605         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
610         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
615         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
620         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
625         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
630         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
635         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
640         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
645         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
650         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
662         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vi2x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
667         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vi3x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
672         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vi4x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
677         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
682         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vi3x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
687         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c2x0123, vi4x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
693         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
696         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
699           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
707           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
710             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()