I think you are misunderstanding something, you don’t need a rocm kernel. What you need is the rocm-opencl-runtime.

This video is a year old, but should be enough to get you started: https://www.youtube.com/watch?v=d_CgaHyA_n4

You can get it to work on arch, rocm is in the repos.

I suggest you use a container if you proceed though.

While i didn’t benchmark previously, I remember quite vividly the speeds were much slower 7-8 months ago.