[llvm] [Offload] Introduce offload-tblgen and initial new API implementation (PR #108413)

[Callum Fare via llvm-commits]

callumfare wrote:

@jhuber6 

Thanks for the quick review, I'll try to address your other comments asap.

> I'm guessing right now this is mostly a demo and doesn't expose a functioning kernel launch? One downside to using this TableGen stuff is I have no clue what the headers look like after being generated.

Indeed, this doesn't expose anything to do with kernels, queues, memory, etc. right now. Just basic device discovery stuff which is enough for a quick demo - see the information in the PR description. I think implementing something that exposes all that would be achievable quickly, but getting a polished interface will obviously take much longer. Either way I didn't want to stick too much in the initial PR.

I agree that not seeing the generated headers is unfortunate. In Unified Runtime we check in the generated files (and have a CI job that fails if the generated headers don't match the output of re-running the generation). All things aside I think I'd prefer that approach, but I understand that's not how any other LLVM project works so it might be difficult to get traction there.

https://github.com/llvm/llvm-project/pull/108413