[Mlir-commits] [mlir] [mlir][linalg] Convert linalg.named to linalg.elementwise op. (PR #148424)

[Renato Golin]

rengolin wrote:

I'd like to have a longer term plan for this. Given the [operation tree](https://discourse.llvm.org/t/rfc-mlir-linalg-operation-tree/83586) discussion, we probably should start thinking about _"partial (de-)generalization"_ instead of lose passes.

```
generic <--> elementwise <--> add/mul/etc
          ^--> contract <--> matmul/etc
```

And also look at fitting the non-perfectly-nested ops like `contract` and composite ops like `softmax` to a DAG of operations:

```
DAG(generic ops) <--> DAG(named ops) <--> softmax
```

So that users can pick-and-choose separately which stage they stop for each step of the way.

https://github.com/llvm/llvm-project/pull/148424