144         const float32x4_t vi4x2 = vld1q_f32(i4); i4 += 4;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()  local
163         vo1x1 = vmlaq_lane_f32(vo1x1, vk20c1, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
184         vo1x1 = vmlaq_lane_f32(vo1x1, vk20c2, vget_low_f32(vi4x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
205         vo1x1 = vmlaq_lane_f32(vo1x1, vk21c0, vget_high_f32(vi4x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
226         vo1x1 = vmlaq_lane_f32(vo1x1, vk21c1, vget_high_f32(vi4x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
294         vo1x0 = vmlaq_lane_f32(vo1x0, vk22c1, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
315         vo1x0 = vmlaq_lane_f32(vo1x0, vk22c2, vget_low_f32(vi4x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
396         float32x4_t vi4x2 = vmovq_n_f32(0.0f);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()  local
403           vi4x2 = vld1q_f32(i4 + 4);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
423         vo1x1 = vmlaq_lane_f32(vo1x1, vk20c1, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
444         vo1x1 = vmlaq_lane_f32(vo1x1, vk20c2, vget_low_f32(vi4x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
470           vo1x1 = vmlaq_lane_f32(vo1x1, vk21c0, vget_high_f32(vi4x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
497           vo1x1 = vmlaq_lane_f32(vo1x1, vk21c1, vget_high_f32(vi4x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
564           vo1x0 = vmlaq_lane_f32(vo1x0, vk22c1, vget_low_f32(vi4x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
579           vo1x0 = vmlaq_lane_f32(vo1x0, vk22c2, vget_low_f32(vi4x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()