106         float32x4_t vo0x1 = vo0x0;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()  local
120         vo0x1 = vmlaq_lane_f32(vo0x1, vk00c0, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
127         vo0x1 = vmlaq_lane_f32(vo0x1, vk10c0, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
134         vo0x1 = vmlaq_lane_f32(vo0x1, vk20c0, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
148         vo0x1 = vmlaq_lane_f32(vo0x1, vk00c1, vget_low_f32(vi0x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
155         vo0x1 = vmlaq_lane_f32(vo0x1, vk10c1, vget_low_f32(vi1x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
162         vo0x1 = vmlaq_lane_f32(vo0x1, vk20c1, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
169         vo0x1 = vmlaq_lane_f32(vo0x1, vk00c2, vget_low_f32(vi0x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
176         vo0x1 = vmlaq_lane_f32(vo0x1, vk10c2, vget_low_f32(vi1x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
183         vo0x1 = vmlaq_lane_f32(vo0x1, vk20c2, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
190         vo0x1 = vmlaq_lane_f32(vo0x1, vk01c0, vget_high_f32(vi0x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
197         vo0x1 = vmlaq_lane_f32(vo0x1, vk11c0, vget_high_f32(vi1x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
204         vo0x1 = vmlaq_lane_f32(vo0x1, vk21c0, vget_high_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
211         vo0x1 = vmlaq_lane_f32(vo0x1, vk01c1, vget_high_f32(vi0x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
218         vo0x1 = vmlaq_lane_f32(vo0x1, vk11c1, vget_high_f32(vi1x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
225         vo0x1 = vmlaq_lane_f32(vo0x1, vk21c1, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
239         vo0x1 = vmlaq_lane_f32(vo0x1, vk01c2, vget_low_f32(vi0x3), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
246         vo0x1 = vmlaq_lane_f32(vo0x1, vk11c2, vget_low_f32(vi1x3), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
253         vo0x1 = vmlaq_lane_f32(vo0x1, vk21c2, vget_low_f32(vi2x3), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
260         vo0x1 = vmlaq_lane_f32(vo0x1, vk02c0, vget_low_f32(vi0x3), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
267         vo0x1 = vmlaq_lane_f32(vo0x1, vk12c0, vget_low_f32(vi1x3), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
274         vo0x1 = vmlaq_lane_f32(vo0x1, vk22c0, vget_low_f32(vi2x3), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
281         vo0x1 = vmlaq_lane_f32(vo0x1, vk02c1, vget_high_f32(vi0x3), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
288         vo0x1 = vmlaq_lane_f32(vo0x1, vk12c1, vget_high_f32(vi1x3), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
295         vo0x1 = vmlaq_lane_f32(vo0x1, vk22c1, vget_high_f32(vi2x3), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
302         vo0x1 = vmlaq_lane_f32(vo0x1, vk02c2, vget_high_f32(vi0x3), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
309         vo0x1 = vmlaq_lane_f32(vo0x1, vk12c2, vget_high_f32(vi1x3), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
316         vo0x1 = vmlaq_lane_f32(vo0x1, vk22c2, vget_high_f32(vi2x3), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
327         vo0x1 = vmaxq_f32(vo0x1, vmin);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
332         vo0x1 = vminq_f32(vo0x1, vmax);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
335         const float32x4x2_t vo0c0123 = vzipq_f32(vo0x0, vo0x1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
353         float32x4_t vo0x1 = vo0x0;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()  local
368           vo0x1 = vmlaq_lane_f32(vo0x1, vk00c0, vget_high_f32(vi0x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
377           vo0x1 = vmlaq_lane_f32(vo0x1, vk10c0, vget_high_f32(vi1x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
386           vo0x1 = vmlaq_lane_f32(vo0x1, vk20c0, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
408         vo0x1 = vmlaq_lane_f32(vo0x1, vk00c1, vget_low_f32(vi0x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
415         vo0x1 = vmlaq_lane_f32(vo0x1, vk10c1, vget_low_f32(vi1x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
422         vo0x1 = vmlaq_lane_f32(vo0x1, vk20c1, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
429         vo0x1 = vmlaq_lane_f32(vo0x1, vk00c2, vget_low_f32(vi0x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
436         vo0x1 = vmlaq_lane_f32(vo0x1, vk10c2, vget_low_f32(vi1x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
443         vo0x1 = vmlaq_lane_f32(vo0x1, vk20c2, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
451           vo0x1 = vmlaq_lane_f32(vo0x1, vk01c0, vget_high_f32(vi0x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
460           vo0x1 = vmlaq_lane_f32(vo0x1, vk11c0, vget_high_f32(vi1x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
469           vo0x1 = vmlaq_lane_f32(vo0x1, vk21c0, vget_high_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
478           vo0x1 = vmlaq_lane_f32(vo0x1, vk01c1, vget_high_f32(vi0x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
487           vo0x1 = vmlaq_lane_f32(vo0x1, vk11c1, vget_high_f32(vi1x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
496           vo0x1 = vmlaq_lane_f32(vo0x1, vk21c1, vget_high_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
518         vo0x1 = vmlaq_lane_f32(vo0x1, vk01c2, vget_low_f32(vi0x3), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
525         vo0x1 = vmlaq_lane_f32(vo0x1, vk11c2, vget_low_f32(vi1x3), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
532         vo0x1 = vmlaq_lane_f32(vo0x1, vk21c2, vget_low_f32(vi2x3), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
584         vo0x1 = vmaxq_f32(vo0x1, vmin);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
589         vo0x1 = vminq_f32(vo0x1, vmax);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
594           const float32x4x2_t vo0c0123 = vzipq_f32(vo0x0, vo0x1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()