Hey! Cool work, bringing GPU compute to Go at this level. I found this through your RWA work.

I have AMD cards that support HIP, which is ~ a subset of CUDA that supports CUDA itself, but also AMD devices. If your bindings happen to fall within this subset of CUDA, it would be pretty cool to port it to HIP and support both.

From what I can see, in many cases this amounts to changing the header file pointed towards, and changing the names of some function calls to hipXYZ rather than cuXYZ, and so forth.

Given Go's conditional compilation support, it might even be possible to fully support both behind a build flag.

Consider this a low-priority request from me, as nothing is riding on it. But it'd be cool to support a broader range of cards in the Go ML ecosystem. :)

Is there any plan to port this lib to cuda10.1?