[Mlir-commits] [mlir] [mlir][linalg] Pack matmul pass (PR #89782)

[llvmlistbot at llvm.org]

MaheshRavishankar wrote:

> > Thanks for the change. We have been doing something similar downstream as well.
> > @yzhang93 can you help review this and guide it based on the use case we have worked through.
> 
> I think this pass may not be general enough for a broader audience . It limits the usage to linalg.matmul/linalg.batch_matmul ops (without any linalg.generic variants) and fixes the innerPerm/outerPerm to {1, 0}.
> 
> In our downstream development, I found it very fixable to directly call `linalg::pack` and `packTranspose` to meet our particular requirements for dimension permutations and various generic versions of matmul. On the other hand, this pass looks like a wrapper for `packMatmulGreedily` and `packTranspose` methods but has a rather narrow scope.

What does `packMatmulGreedily` do? It is fine to start with something that has a narrow scope to begin with but can be generalized as needed.

https://github.com/llvm/llvm-project/pull/89782