144         const float32x4_t vi4x2 = vld1q_f32(i4); i4 += 4;  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
163         vo1x1 = vfmaq_laneq_f32(vo1x1, vk20c1, vi4x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
184         vo1x1 = vfmaq_laneq_f32(vo1x1, vk20c2, vi4x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
205         vo1x1 = vfmaq_laneq_f32(vo1x1, vk21c0, vi4x2, 2);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
226         vo1x1 = vfmaq_laneq_f32(vo1x1, vk21c1, vi4x2, 3);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
294         vo1x0 = vfmaq_laneq_f32(vo1x0, vk22c1, vi4x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
315         vo1x0 = vfmaq_laneq_f32(vo1x0, vk22c2, vi4x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
398         float32x4_t vi4x2 = vmovq_n_f32(0.0f);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
405           vi4x2 = vld1q_f32(i4 + 4);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
425         vo1x1 = vfmaq_laneq_f32(vo1x1, vk20c1, vi4x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
446         vo1x1 = vfmaq_laneq_f32(vo1x1, vk20c2, vi4x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
472           vo1x1 = vfmaq_laneq_f32(vo1x1, vk21c0, vi4x2, 2);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
499           vo1x1 = vfmaq_laneq_f32(vo1x1, vk21c1, vi4x2, 3);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
566           vo1x0 = vfmaq_laneq_f32(vo1x0, vk22c1, vi4x2, 0);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
581           vo1x0 = vfmaq_laneq_f32(vo1x0, vk22c2, vi4x2, 1);  in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()