87       float32x4_t vi2x0 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()  local
99         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
108         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
114         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
123         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
129         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
138         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
144         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
153         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
247         vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
291         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
300         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
306         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
315         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c1x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
321         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
330         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c2x0123, vget_high_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
336         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
345         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_high_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()