92         float32x4_t vo1x0c0123 = vo0x0c0123;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()  local
106         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
114         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
122         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
130         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
138         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_low_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
146         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
161         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
169         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c2x0123, vget_high_f32(vi3x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
177         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
185         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
193         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c0x0123, vget_high_f32(vi3x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
201         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
209         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c1x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
217         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c1x0123, vget_low_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
225         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c1x0123, vget_low_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
233         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk01c2x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
241         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk11c2x0123, vget_low_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
249         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk21c2x0123, vget_low_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
264         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
272         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c0x0123, vget_high_f32(vi3x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
280         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
288         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c1x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
296         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_high_f32(vi3x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
304         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
312         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
320         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
328         vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c2x0123, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
343         vo1x0c0123 = vmaxq_f32(vo1x0c0123, vmin);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
349         vo1x0c0123 = vminq_f32(vo1x0c0123, vmax);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
355           vst1q_f32(o1, vo1x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
368           float32x2_t vo1x0c01 = vget_low_f32(vo1x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
378             vo1x0c01 = vget_high_f32(vo1x0c0123);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()