85       float32x4_t vi2x0 = vmovq_n_f32(0.0f);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()  local
111         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
113         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk00c0x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
136         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
138         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
157         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
159         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk00c1x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
182         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
184         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
196         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
198         vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk00c2x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
221         vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
223         vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
474         vi2x0 = vi2x3;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
567         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
569         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c0x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
582         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
584         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
591         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
593         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c1x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
606         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
608         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
615         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
617         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c2x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
630         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
632         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
795         vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
855         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
857         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c0x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
870         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
872         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c0x4567, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
879         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
881         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c1x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
894         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
896         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c1x4567, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
903         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
905         vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c2x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
918         vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
920         vo0c4567 = vmlaq_lane_f32(vo0c4567, vk20c2x4567, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()