311         const float32x4_t vi4x3 = vld1q_f32(i4); i4 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
345         vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk21c2x0123, vi4x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
347         vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk21c2x4567, vi4x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
384         vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk22c0x0123, vi4x3, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
386         vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk22c0x4567, vi4x3, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
423         vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk22c1x0123, vi4x3, 2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
425         vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk22c1x4567, vi4x3, 2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
462         vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk22c2x0123, vi4x3, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
464         vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk22c2x4567, vi4x3, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
470         vi4x0 = vi4x3;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
798         float32x4_t vi4x3 = vmovq_n_f32(0.0f);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
805           vi4x3 = vld1q_lane_f32(i4 + 8, vi4x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
840         vo1x1c0123 = vfmaq_laneq_f32(vo1x1c0123, vk21c2x0123, vi4x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
842         vo1x1c4567 = vfmaq_laneq_f32(vo1x1c4567, vk21c2x4567, vi4x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()