Are you aware of AMD’s efforts to make it possible to compile CUDA sources for AMD cards with minimal source modification? It’s already possible to compile PyTorch for AMD this way. I’m not sure if you’re using any frameworks or are working with CUDA directly, but doing it this way should be less work than writing an entirely new backend for OpenCL.
I tried searching the forums for any official information regarding AMD GPU support, but couldn’t really find anything. Apologies if this has already been discussed.