[Mlir-commits] [mlir] [MLIR][Linalg] Add aggregate ops decomposition pass and softmax decom… (PR #97582)

[Petr Kurapov]

kurapov-peter wrote:

> Just to map back to what I said above, we can "recognize" that its a fill, but that seems like an unnecessary burden added to downstream users because it has been generalized too early without any control. I can go into details about why I think "fill" is special but thats a separate issue IMO.

I think you missed the point. The proposed decomposition only converts an aggregate operation into a sequence of non-aggregate ones. This has nothing to do with generalization. Downstreams don't need to recognize a `fill` from its generic form. The solution for you would be to do partial generalization, leaving `fill`s intact.

> Actually that brings me to maybe a potential solution. You can take the existing lowering for softmax and then add a pass to explicitly split out the broadcast and then generalize. That will get you to the state you want here?

Same here. Generalized IR with broadcasts is not the target state. The target is a sequence of named ops.

https://github.com/llvm/llvm-project/pull/97582