With CMake 3.18 we have native support for clang in cuda mode. We should investigate if we can profit from this. (Maybe only once hip is also there?)