[Mlir-commits] [mlir] [MLIR][Linalg] Add boundary fill option to pad fusion pass (PR #223792)

[Andrzej Warzyński]

banach-space wrote:

> Options and documentation all come down to maintenance cost and also a cost to MLIR user who need to read them and understand the nuances.
> 

I agree that special-casing like this is problematic from maintenance point of view. I'm not sure whether a dedicated pattern would be better though, the code duplication would be quite bad. 

> > I wonder how your micro-test looks like, is it a loop doing a element-wise padding inside and running on a single core? @bruteforceboy

I would also like to make sure that the benchmark that's used to justify this is not "hermetic" (i.e. not related to any real use case). 

https://github.com/llvm/llvm-project/pull/223792