560         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()  local
567         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
575         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
583         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 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()
599         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
607         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 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()
623         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
631         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
646         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
654         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
662         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
670         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
678         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
686         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
694         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
702         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
710         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
718         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
726         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c0x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
734         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
749         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c1x0123, vi2x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
757         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
765         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
773         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
781         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
789         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c2x0123, vi4x2, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
803         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
808         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
813           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
823             vst1q_f32(o1_tmp, vo1c0123); o1_tmp += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
824             vo1c0123 = vo1c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
829           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
832             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
848         float32x4_t vo1c0123 = vo0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()  local
855         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
863         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
871         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 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()
887         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
895         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 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()
911         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c2x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
919         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
934         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
942         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c0x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
950         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c0x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
958         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
966         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c1x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
974         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c1x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
982         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
990         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk11c2x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
998         vo1c0123 = vmlaq_lane_f32(vo1c0123, vk21c2x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1006         vo1c0123 = vmaxq_f32(vo1c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1011         vo1c0123 = vminq_f32(vo1c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1016           vst1q_f32(o1, vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1026             vst1q_f32(o1_tmp, vo1c0123); o1_tmp += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1027             vo1c0123 = vo1c4567;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1032           float32x2_t vo1c01 = vget_low_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
1035             vo1c01 = vget_high_f32(vo1c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()