[Mlir-commits] [mlir] [MLIR] Enable scalable vectorization for linalg.batch_matmul (PR #172333)

[Andrzej Warzyński]

banach-space wrote:

> > Looks trivial to me. @arun-thmn any feedback?
> > What about `batch_reduce_matmul`? Would that be as trivial, too?
> 
> +1 for `batch.reduce_matmul`

Is that a blocker for you? I am not against it, but as a rule of thumb, we only enable Ops that we actually run. This way, we have 100% confidence that everything works (i.e. lowering all the way to LLVM).

Perhaps that's too conservative - ultimately, this is just unblocking `Linalg` -> `Vector` lowering - but this way we avoid "scalable" vectors being enabled quite high-up (Linalg) and then not working further down the lowering pipeline.

https://github.com/llvm/llvm-project/pull/172333