[Feature Request] Kernel loop per-iteration data reading/fetching size/type settings

As i understand correctly, you can tell attribute to read more data per iteration/loop-cycle. This way we could tell to pick 9 numbers as example, it's a 3x3 matrix make calculations on them and return 9 calculated numbers back.

It would give more freedom to many non-stateless calculations. For example a Gauss Elimination could be accomplished due to pivoting.

Can this be accomplished?
What do you think?