307         const float32x4_t vi0x3 = vld1q_f32(i0); i0 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
318         vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk01c2x0123, vi0x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
320         vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk01c2x4567, vi0x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
357         vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk02c0x0123, vi0x3, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
359         vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk02c0x4567, vi0x3, 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
396         vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk02c1x0123, vi0x3, 2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
398         vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk02c1x4567, vi0x3, 2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
435         vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk02c2x0123, vi0x3, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
437         vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk02c2x4567, vi0x3, 3);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
466         vi0x0 = vi0x3;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
794         float32x4_t vi0x3 = vmovq_n_f32(0.0f);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()  local
801           vi0x3 = vld1q_lane_f32(i0 + 8, vi0x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
813         vo0x1c0123 = vfmaq_laneq_f32(vo0x1c0123, vk01c2x0123, vi0x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
815         vo0x1c4567 = vfmaq_laneq_f32(vo0x1c4567, vk01c2x4567, vi0x3, 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()