Feature Request: HIP Support about cuda HOT 6 OPEN

Took me a while to find this, it's HIP's porting guide: https://github.com/ROCm-Developer-Tools/HIP/blob/master/docs/markdown/hip_porting_guide.md

from cuda.

Interesting proposal. I'm curious why one might use HIP instead of something like OpenCL? Is it just for easier porting?

To me, the proper approach for supporting AMD cards would be to create a separate opencl or hip package. The package would come with its own plugin for anyvec, which is the abstraction layer under anynet as it is.

from cuda.

from cuda.

I see what you did here: https://github.com/unixpickle/anyplugin

This is a promising route, I think? Is the idea that, instead of directly referencing/building-upon "github.com/unixpickle/cuda" you'd build on "github.com/unixpickle/anyplugin" and reference the build tag that best suits the back-end?

So, similar to having build-tags on the vector-back-end repository, but decoupling it to keep architecture cleaner overall?

from cuda.

@cathalgarvey I am surprised you noticed this so fast! Yes, that is the plan. I want to avoid the Torch7 problem you mentioned by having a universal plugin package that everyone should include. That way, anyvec plugins for HIP, OpenCL, or cuDNN can be dropped in after the fact.

from cuda.

from cuda.