[Mlir-commits] [mlir] [mlir][linalg] Add mixed precision folding pattern in vectorize_children_and_apply_patterns TD Op (PR #148684)

[Andrzej Warzyński]

banach-space wrote:

’m OK with extending `transform.structured.vectorize_children_and_apply_patterns` for now. However, if we plan to extend it further, we should step back and discuss the design goals of this Op more explicitly. Thanks!

My only outstanding request for this PR is to move the tests to:
* https://github.com/llvm/llvm-project/tree/main/mlir/test/Dialect/Linalg/vectorization
For the rationale and naming scheme, please see:
* https://github.com/llvm/llvm-project/issues/141025

I feel the current TD-specific test files can be removed - we already test TD Ops extensively elsewhere.

---

@rolfmorel Thanks for sharing all the ideas! Looks like we have quite a few directions to consider. We should probably move this discussion to a more suitable place - maybe a dedicated GitHub issue?

Two high-level concerns I have about combining the existing **vectorize** Ops into e.g. `transform.apply_patterns.vector.vectorize`:

* Op granularity in [Transform dialect](https://mlir.llvm.org/docs/Dialects/Transform/) is intentionally low - even single-pattern Ops exist. Introducing large meta-Ops may go against that design.

* The tests in mlir/test/Dialect/Linalg/vectorization currently follow the split between `vectorize_children_and_apply_patterns` and `vectorize`. If we unify the Ops, we’ll need to rethink that split and redesign the test organisation - a fair bit of churn. And to be honest, test organisation and discoverability are already weak points.

---

Thanks again to everyone for the great discussion so far! 🙏🏻

https://github.com/llvm/llvm-project/pull/148684