[Mlir-commits] [mlir] Revert "[mlir][tensor] Refine the semantics of `createPadHighOp`" (PR #110153)

[Han-Chung Wang]

hanhanW wrote:

> Interestingly, in your example you are padding the static dim and for dynamic dims the transformation fails elsewhere. 

I don't remember all the details, but the order to make fusion work is `tile parallel loops -> pad parallel dimensions-> tile reduction loops -> pad reduction dimensions`. The first padding avoids the padding along reduction dimensions because they are not tiled yet. Then the reduction dimension is tiled, so we can pad the operands with a smallest bounding box.

> Btw, this made me wonder - why pack_paddings and not nofold 🤔 ?

I have no ideas either. :)

https://github.com/llvm/llvm-project/pull/110153