[Mlir-commits] [mlir] [NFC] Simplify the tiling implementation using cloning. (PR #72178)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Tue Nov 14 10:58:42 PST 2023


MaheshRavishankar wrote:

> Quick offhand question before I can devote more time to a review later this week: is this a step towards reducing the duplicated implementation problem ?

Yes it is (at least in my mind)... The reasons I did this was after looking at the implementation of the tiling using `scf.forall` under the hood of the transform dialect op, I realized that I could adapt the current implementation of tiling using `scf.for` much simpler by following the same path. So before I push forward on moving more of the implementation of tiling from under the transform dialect op in Linalg into here, took a slight detour to make the current implementation of `scf.for` consistent with the `scf.forall` implementation. I think there is a lot of logic here we could reuse for both.

In terms of taking time, yeah, I still need to test it with IREE and fix any issues. So its not urgent, but quicker review is appreciated. 

PS: irrespective of the above, I think this is just in general a good cleanup. It handles better managing the destination passing style through tiling.

https://github.com/llvm/llvm-project/pull/72178


More information about the Mlir-commits mailing list