89       float32x4_t vi4x0 = vld1q_f32(i4); i4 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()  local
118         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
120         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
142         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
144         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
166         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
168         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
190         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
192         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
332         vi4x0 = vcombine_f32(vget_high_f32(vi4x1), vi4x2);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
405         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
407         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c0x4567, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
429         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
431         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c1x4567, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
453         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c2x0123, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
455         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk20c2x4567, vget_high_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
477         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk21c0x0123, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
479         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk21c0x4567, vget_high_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()