[RFC] Move `tensor.pack` and `tensor.unpack` into Linalg

[banach-space]

ATM, these Ops sit at the boundary between Tensor and Linalg, so some clarity here would be welcome.

However, constructs like:

• GeneralizeOuterUnitDimsPackOpPattern

decompose tensor.pack into a combination of tensor.pad (tensor) and linalg.transpose (linalg). It seems that even after a move, these Ops may remain somewhat at the boundary due to patterns like GeneralizeOuterUnitDimsPackOpPattern.

Overall, I don’t have a strong stance on where these Ops belong, but I’d appreciate avoiding a rushed decision, especially considering ongoing discussions around Linalg:

• RFC: Op Explosion in Linalg,
• RFC: MLIR Project Charter and Restructuring.

Additionally, I’m currently updating various patterns to support dynamic tile sizes. I’ve mostly completed tensor.pack and will next focus on tensor.unpack. I’m happy to coordinate with anyone handling this move or copy effort—please feel free to reach out.

Lastly,

@rengolin

This gives us a second, more important, reason to move (or copy?) pack and unpack to Linalg: if we restrict its usage to “tensor compiler” logic, then we can inbue it with HW information, because that’s what tensor compilers do.

I don’t quite follow this point.