Given that CUDA (or some other language; I don't remember which) doesn't implement this, because CUDA is designed for very "data-parallel" work (kind of), I wonder what type of work actually really wants something like this