135         const float32x4_t vk20c0x0123 = vld1q_f32(w + 24);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
138         vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
139         vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
143         vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
144         vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
579         const float32x4_t vk20c0x0123 = vld1q_f32(w + 24);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
582         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
583         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
865         const float32x4_t vk20c0x0123 = vld1q_f32(w + 24);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()  local
868         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
869         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()