[Mlir-commits] [mlir] [mlir][tensor] Update `GeneralizeOuterUnitDimsPackOpPattern` (PR #115312)

[Han-Chung Wang]

hanhanW wrote:

> > This PR also adds a check to verify that only the last N (for some value
> > of N) trailing dims that are being tiled. From what I can tell, that's
> > always the case in practice.
> 
> I have seen pack ops show up in practice that do not follow this restriction, so it may be worth supporting this case, but I'm not sure we rely much on this pattern anymore anyway. I'll approve, but pinging @hanhanW, who might have a better idea of whether or not support for this case would be desired as a follow up.

It is not a restriction in our use cases because we can data-tile any ops that implements ContractionOpInterface. E.g., a contraction generic op which has batch dimension being the innermost dimension. In this case, we don't data-tile batch dimension.

These patterns were built for the first take of pack/unpack vectorization, and they are no longer used on IREE CPU x86 codegen path. Because today we have direct vectorization and masking supports. 

https://github.com/llvm/llvm-project/pull/115312