[Mlir-commits] [mlir] [mlir][linalg] Vectorize directly to a named contraction (PR #147296)

[Andrzej Warzyński]

https://github.com/banach-space commented:

Very nice, thanks!

I've made some small suggestions inline, but nothing major. The approach that you are taking makes a lot of sense to me.

> The new path is optional and disabled by default to avoid changing the default vectorizer behavior.

This is good time to discuss the future direction for this. 

@dcaballe has already hinted in [Vector Dialect: Refactoring + Re-design ideas](https://docs.google.com/document/d/10lfONA_T0SfnucXICUFAAWuvp4OICOw3XK3bid0eoV0/edit?tab=t.0) that we should look into removing `vector.multi_reduction`. I am not sure about that myself _just yet_, but avoiding `vector.multi_reduction` in the vectorization path makes a lot of sense to me (i.e. specifically for "Linalg contractions"). So:  
* Why shouldn't we make it the default? 
* What are the next step for you? 

I am happy to work with you to make this the default if we/you discover that's something is missing. Otherwise, I would prepare people to flip the switch soonish.

https://github.com/llvm/llvm-project/pull/147296