It seems the author believes checking the return values of CUDA API calls is not "tiny" enough :-(