add transparent amd gpu support

[tristanitschner]

Hi there,

this is early RFC. I added transparent AMD GPU support. I would be very happy about feedback :)

To build for amd, simply run make TARGET=AMD. By default, it will still build for NVIDIA of course without any functional changes.

Currently, gpu temp support is not available. While I have though about implementing the same pattern as taken with nvidia-smi, I don't think this is a clean solution. My take here would be to compile amdsmi as a static library and directly link against it. The licenses should be compatible, but correct me if I'm wrong. Also I haven't checked yet whether and how amdsmi handles global state, so there might be a technical issue.

Also I had to comment out some cublas functions. I do not know whether this leads to functional differences, so that should definitely be looked at.

Please let me know what you think :)

[wilicc]

Hi @tristanitschner, thanks for the patience on this. Looping in #111 where @kerwenwwer did a separate AMD HIP port and shared hipblas perf observations from MI210 — worth coordinating between threads.

A few things would need to happen for me to merge here:

• Revert the /usr/bin → /opt/cuda change in the Makefile (would break existing Linux users; unrelated to AMD support)
• Fix the segfault on AMD when nvidia-smi isn't present
• Either basic AMD temp polling via rocm-smi or a clear README note that AMD temps aren't reported
• Some plan for ongoing validation — I don't have AMD hardware myself, so any AMD path landing here will need community ownership

Are you still actively on this? If yes, no rush. If not, I'll close as stale in a month or so and link to your branch in case someone picks it up.

[tristanitschner]

Hello, thanks for the reply.
No, not currently working on this, just did it to test my new GPU and decided to open pull request so the work is not lost.
I'm fine if @kerwenwwer finishes his port.
Otherwise I will finish this port according to your comments.