559         float32x4_t vo0c4567 = vld1q_f32(w + 4);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
561         float32x4_t vo1c4567 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
568         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
576         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
584         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
592         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
600         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
608         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
616         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
624         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
632         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
640         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
648         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
656         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
671         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vget_low_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
679         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c1x4567, vget_low_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
687         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
695         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c2x4567, vget_low_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
703         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vget_low_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
711         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
719         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
727         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
735         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
743         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c1x4567, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
751         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
759         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c1x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
774         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c2x4567, vi0x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
782         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
790         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
802         vo0c4567 = vmaxq_f32(vo0c4567, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
807         vo0c4567 = vminq_f32(vo0c4567, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
815           vst1q_f32(o0 + 4, vo0c4567);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
824             vo0c0123 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
845         float32x4_t vo0c4567 = vld1q_f32(w + 4);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
847         float32x4_t vo1c4567 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
854         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c0x4567, vget_low_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
862         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
870         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
878         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c1x4567, vget_low_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
886         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
894         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c1x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
902         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c2x4567, vget_high_f32(vi0x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
910         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
918         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
926         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c0x4567, vget_high_f32(vi0x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
934         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
942         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
957         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c1x4567, vi0x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
965         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c1x4567, vi1x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
973         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vi2x1, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
981         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk01c2x4567, vi0x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
989         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c2x4567, vi1x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
997         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vi2x1, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
1003         vo0c4567 = vmaxq_f32(vo0c4567, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
1008         vo0c4567 = vminq_f32(vo0c4567, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
1016           vst1q_f32(o0 + 4, vo0c4567);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
1025             vo0c0123 = vo0c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()