[Mlir-commits] [mlir] [mlir][vector] Set InBound for vector read write after peeling (PR #89108)

[Diego Caballero]

dcaballe wrote:

> The issue is that it seems the affine dialect didn't have upstream support for peeling. Is it worth to introduce peeling in affine dialect?

There was someone asking about peeling as well on Discourse: https://discourse.llvm.org/t/rfc-mlir-vector-affine-supervectortize-optimization-for-misaligned-data/77860/5 Perhaps they have something working already :)

> Could I ask how to determine which loop to apply vectorization by vectorizer?

I guess it depends on what you want to achieve. You could vectorized both loops, one without masking an done with masking. I don't remember the logic of the Affine SuperVectorizer but if there is a way to disable masking, I guess you could also vectorize only the main vector loop and the vectorizer should bail out when trying to vectorize the remainder loop. Perhaps @sergei-grechanik can provide more context.

https://github.com/llvm/llvm-project/pull/89108