[Mlir-commits] [mlir] [MLIR][Linalg] Scalable Vectorization of Reduction (PR #97788)

[Andrzej Warzyński]

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

I've finally had a bit more time for this. I think that a lot of complexities in this PR stem from the fact that `vectorizeScalableVectorPrecondition` is a bit messy and hard to extend (my fault!). Also:

> Simply removing it will break some useful cases like matmul. Making sure we allow all correct combinations of vector sizes and op types and prevent unsupported cases is beyond the scope of this PR. I'm happy to work on it later.

I think that we should refactor things a bit first and then build this PR on top of that. To quickly unblock you, here's what I'm proposing:
* https://github.com/llvm/llvm-project/pull/98639

I think that you should be able to enable "reductions" quite easily. Sadly re-basing won't be straightforward :(

Also:
> [MLIR][Linalg] Scalable Vectorization of Reduction

To me this says that you are adding scalable vectorisation of e.g. `linalg.reduce`, but in practice you are doing something more generic - allowing reduction dimensions to be scalable. It's worth updating the summary.

https://github.com/llvm/llvm-project/pull/97788