[Mlir-commits] [mlir] [mlir][linalg] Introduce new `linalg.conv` op (PR #117688)

[Andrzej Warzyński]

banach-space wrote:

I agree with the points raised by Renato.

> This PR is the direct result of the discourse thread

While I appreciate the initiative, there was a [Call for Action](https://discourse.llvm.org/t/rfc-op-explosion-in-linalg/82863/31?u=banach-space) in the thread that remained mostly unanswered. This PR feels a bit unexpected as a result. For core logic like this, it's crucial to coordinate work in the original thread to ensure alignment.

In particular, Mahesh [proposed a plan](https://discourse.llvm.org/t/rfc-op-explosion-in-linalg/82863/7), and it's natural to assume he would like to follow through on it. Have you had the chance to discuss this with him?

> there is a reason for why discussion in that threa died down.

There’s been excellent feedback, and the next step involves proposing a refined design to address the concerns raised. This is a complex problem, and developing a solution that incorporates all the feedback takes time and thoughtful consideration. Could you clarify how this PR aligns with the key points from the thread?

One critical aspect missing here is alignment with the recent refactoring of `linalg.matmul`:

* https://github.com/llvm/llvm-project/pull/104783

That refactoring involved a thorough and detailed discussion to ensure future-proofing. Given the complexity and variety of Convs, achieving a similar state will likely require even more time and deliberation. Some delay is natural and expected.

---

That said, this PR provides a valuable data point that will undoubtedly contribute to the ongoing discussion. Thank you for sharing this! I’m a bit constrained this week, so apologies in advance for any delays on my side.

https://github.com/llvm/llvm-project/pull/117688