[llvm] [Offload][AMDGPU] Add env var to control device-to-device memory access (PR #215385)

[via llvm-commits]

hidekisaito wrote:

> Do we really need more environment variables? I would assume that we should do the optimal thing in every case, according to what the system and device supports. Is there ever a case where we wouldn't be able to determine this?

I can add code to detect single GPU device case, but that doesn't help single GPU device code running on a multi-GPU system. I don't think we can check the need-for-d2d at every memory access. Are you suggesting to try adding a handler for GPU memory access fault?

https://github.com/llvm/llvm-project/pull/215385