[Wave] Reproducer for reported issue

[harsh-nod]

No description provided.

[harsh-nod]

@adedespirlet - does this PR look correct to you?

[adedespirlet]

Thanks for taking a look at it! Unfortunately I’m still seeing the same code being emitted and the same incorrect results as described in #981.
The code still results in multiple threads reading from the same location in the result vector of the MMA. I think fixing this requires changes in the codegen logic that propagates affine maps across ops with mismatched data layouts. The access pattern for the result of the MMA maybe needs to be recomputed rather than inherited. I am happy to dive more into that after completing RMSnorm kernel