Update FusePackIntoLoop -> FuseProducerIntoLoop to include linalg.copy op

[yzhang93]

In cases that linalg.copy ops are used instead of tensor.pack ops as the producers of the computation op, they should be able to be fused into the loops in the same way. This PR updates the pass to handle such cases.

[newling]

This looks good to me, but an alternative option would be to factorize this into a new pass iree-amdaie-copy-to-pack, run that after after any pass that might introduce linalg.copy (ire-amdaie-pad etc), and then we can always assume there are no linalg.copy

[yzhang93]

This looks good to me, but an alternative option would be to factorize this into a new pass iree-amdaie-copy-to-pack, run that after after any pass that might introduce linalg.copy (ire-amdaie-pad etc), and then we can always assume there are no linalg.copy

I know this trick would work as we used it as an intermediate step in ConvertToDmaPass, but I'm not sure if it's a good idea to write it as a formal pass at tensor level. In the definition of PackOp, the original tensor of rank n is packed into a result tensor of rank n + k where k>0 as shown here https://github.com/llvm/llvm-project/blob/main/mlir/include/mlir/Dialect/Tensor/IR/TensorOps.td#L1887

[newling]

:-0 so if those docs are correct you can have an identity IREE::LinalgExt::PackOp but not an identity Linalg::PackOp...

[Abhishek-Varma]

LGTM.