[Mlir-commits] [mlir] [mlir][test] Add TD example for peel+vectorize (depthwise conv) (PR #90200)

[Andrzej Warzyński]

banach-space wrote:

>  wondering if the canonicalization would work in dynamic sizes. 

It should, but I won't have the bandwidth to check. It's not hard to update examples like this, just a bit fiddly, and I need to prioritise sth else in the coming weeks 😅 

> Is it required by decompose ? ( I am not too familiar with the op)

Yes. The vectoriser can only vectorise 1D convs. We deal with 2D convs by decomposing them into 1D convs. That's like a "vectoriser pre-processing" stage.

https://github.com/llvm/llvm-project/pull/90200