[Mlir-commits] [mlir] [mlir][linalg] Enable masked vectorisation for depthwise convolutions (PR #81625)

[Diego Caballero]

https://github.com/dcaballe approved this pull request.

LGTM, for enablement purposes. I understand the criticality of this and I don't want to block the development but, as discussed (also offline with Nicolas), we know that this is a short-term specialization and that all the convolution support in the vectorizer needs to be revisited. As such, it may happen that some of this functionality has to be disabled/revisited in the future if it hinders a more generic path forward. I would also be concerned if there are plans to continue extending this path so let's make sure we discuss those and plan for a better approach if that is the case.

https://github.com/llvm/llvm-project/pull/81625