[Mlir-commits] [mlir] [MLIR] Allowing unsupported conv2d op to fail gracefully from vectorization (PR #130181)

[Andrzej Warzyński]

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

Thanks @jerryyin , convs are in real need of TLC 🙏🏻 

Now, IMO, what's missing is a very high-level check in [vectorizeOpPrecondition](https://github.com/llvm/llvm-project/blob/5ce4045384d7c2544185b0dbcb6222d06beb47dc/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp#L2185-L2216). >From what I can tell, it would be totally fine to check the dims of conv ops very early on.

With your current approach, I just feel that the current fairly "custom" logic is replaced with something else that is also fairly custom. Instead, I am proposing to re-use the high-level logic that's already available.

Thanks for helping us improve this!

https://github.com/llvm/llvm-project/pull/130181