But it depends on some very sketchy code in cuda_ndarray, so it will need
further work.