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

[Renato Golin]

https://github.com/rengolin commented:

Thanks Petr, this is looks good to me (but I'll allow time for others to review, so won't approve).

The main differences I see after generalizing the named ops:
* `linalg.fill`s become a genteric, but that's a minor thing and I think it's the right thing to have.
* `linalg.subf` and `linalg.exp` are not fused as before, but again, this is the correct lowering. One can fuse element-wise ops later if so inclined.
* The `linalg.reduce { maxnum }`'s init constant is wrong, will yield wrong max if all numbers are `<=0`.

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