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

[Adam Siemieniuk]

adam-smnk wrote:

> Why are you framing this as a CPU-only transformation? Are there any implicit limitations that we should be aware of?

AFAIK, it's not necessarily beneficial to pack for GPU to give a counter example. We've also used packing just for CPUs so, it's the main motivation or use case I can confidently bring up.
It might be equally useful for other targets which might use the same or a completely different packing layout/scheme.

> How configurable the layout would be? Could we do plain transposes, decide which operands to transpose and how many blocking levels to use? (Sorry if I missed something while skimming through the code)

This one worked best based on our use cases. I'll try to generalize the pass and see what can be done.

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