[llvm-branch-commits] [llvm] [offload] Write small unordered host-to-device transfers directly on AMDGPU (PR #231569)

[Alex Duran via llvm-branch-commits]

adurang wrote:

I'm not trying to argue that is not useful (it seems like other backends could make a similar optimization based on the explanation you gave) but if we keep extending the Plugin interface without extending the liboffload interface makes the effort to migrate liboffload kind of impossible...

In fact, I think given that AsyncQueue already has an ol_queue_t, I think it should already be possible to use olMemCpy here.

I don't have a strong opinion on how we could pass this information:  a flag to the queue, a flag to the MemCpy operation (this would be similar to what you're already proposing I guess), or something else. But I feel we should have that defined  before making this change.

https://github.com/llvm/llvm-project/pull/231569