121         const float32x4_t vk20c0x0123 = vld1q_f32(w + 12);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
123         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
124         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
126         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
127         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
409         const float32x4_t vk20c0x0123 = vld1q_f32(w + 12);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
411         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
412         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
600         const float32x4_t vk20c0x0123 = vld1q_f32(w + 12);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()  local
602         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
603         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()