[Mlir-commits] [mlir] [mlir] Expose linalg vectorization without replacement (PR #144158)

[llvmlistbot at llvm.org]

Max191 wrote:

> Thanks for sending this!
> 
> What is the rationale and the end-goal here?
> 
> > This allows for more control over vectorization transformations.
> 
> Sure, but without tests upstream, this change feels very arbitrary. _What_ is meant to decide whether to use or not the vectorization output? This could be a job for a cost-model?

I have a use case downstream in IREE, although I haven't sent the code out for it yet. I want to reuse the linalg.generic vectorization implementation for a custom op in IREE, since the vectorization for them is similar. My transformation creates a new generic op with a similar body to the custom op, and then I vectorized the generic, but the results of the generic cannot immediately replace anything, since it is just a step in the vectorization transformation, so I want a way to vectorize and just get the results of the vectorized operation.

https://github.com/llvm/llvm-project/pull/144158