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

[Federico Bruzzone]

FedericoBruzzone wrote:

Thank you for this work, it's a great example to follow! These visuals are impressive because AI allows us to quickly create prototypes of these charts.

Before diving into the review, the following sentence caught my attention, as I believe it to be intuitively true:

> For smaller shapes, a single fill over a larger area could be more efficient than 4 fills (using the 2D case as an example) for the different boundaries. So, I’ve made this an optional flag.

However, looking at your graphs right now, we only see speedups with your approach, so one might ask: _But then why not always have it as the default?_ (without a cost model, I mean).

May I ask you, genuinely, if you still have that conversation with Claude to see if he can find any sizes where this approach performs worse? (Don't hesitate to say no to me :'D)

This is purely out of curiosity, and non-blocking on my part. Usually, these kinds of efforts aren't requested in a PR, so please forgive me.

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