[Mlir-commits] [mlir] [mlir][vector] Fold `in_bounds` for transfers with loop-derived indices (PR #215340)

[Federico Bruzzone]

https://github.com/FedericoBruzzone commented:

Thank you a lot @dhairyashilRG!

Quite impressive [your study](https://gist.github.com/dhairyashilRG/d461b45231f19b49db8d3614a5891401). Since I'm also working on studying MLIR performance for other aspects, I really appreciate seeing these studies backed by actual contributions (I believe the community benefits from this as well).
As you explain, NEON doesn't support predicated/masked memory operations, whereas SVE does.
I have a quick question: _Genuinely, why haven't you measured the performance on SVE? Is there a particular issue that makes it difficult? Just out of curiosity._ 

I'm limiting this comment to pure curiosity, given the comment of @matthias-springer :D







https://github.com/llvm/llvm-project/pull/215340