[Mlir-commits] [mlir] [mlir][x86] Add pattern to tile and unroll vector.contract (PR #229076)

[Julian Oppermann]

jopperm wrote:

>     * Is this pass needs the `c type` to be in `f32`? and fails for the `bf16`.

Yes, `f32 += bf16 @ bf16` and `i32 += i8 @ i8` are the only supported dtype configuration in the initial version.

>     * The fusion of element-wise downconvert is to taken care separately?

Yes, that's the plan, but out-of-scope for this PR. My idea is to extend the pattern to fuse operations forming a single-user-chain from the loop result to the write-back, including type changes, so that would cover a down conversion.

https://github.com/llvm/llvm-project/pull/229076