142         const float32x4_t vi2x2 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()  local
149         vo1x1 = vmlaq_lane_f32(vo1x1, vk00c1, vget_low_f32(vi2x2), 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()
170         vo1x1 = vmlaq_lane_f32(vo1x1, vk00c2, vget_low_f32(vi2x2), 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()
191         vo1x1 = vmlaq_lane_f32(vo1x1, vk01c0, vget_high_f32(vi2x2), 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()
212         vo1x1 = vmlaq_lane_f32(vo1x1, vk01c1, vget_high_f32(vi2x2), 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()
280         vo1x0 = vmlaq_lane_f32(vo1x0, vk02c1, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
293         vo0x0 = vmlaq_lane_f32(vo0x0, vk22c1, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
301         vo1x0 = vmlaq_lane_f32(vo1x0, vk02c2, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
314         vo0x0 = vmlaq_lane_f32(vo0x0, vk22c2, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
394         float32x4_t vi2x2 = vmovq_n_f32(0.0f);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()  local
401           vi2x2 = vld1q_f32(i2 + 4);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
409         vo1x1 = vmlaq_lane_f32(vo1x1, vk00c1, vget_low_f32(vi2x2), 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()
430         vo1x1 = vmlaq_lane_f32(vo1x1, vk00c2, vget_low_f32(vi2x2), 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()
452           vo1x1 = vmlaq_lane_f32(vo1x1, vk01c0, vget_high_f32(vi2x2), 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()
479           vo1x1 = vmlaq_lane_f32(vo1x1, vk01c1, vget_high_f32(vi2x2), 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()
554           vo1x0 = vmlaq_lane_f32(vo1x0, vk02c1, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
563           vo0x0 = vmlaq_lane_f32(vo0x0, vk22c1, vget_low_f32(vi2x2), 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
569           vo1x0 = vmlaq_lane_f32(vo1x0, vk02c2, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
578           vo0x0 = vmlaq_lane_f32(vo0x0, vk22c2, vget_low_f32(vi2x2), 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()