[Mlir-commits] [mlir] [mlir][linalg] Relax tensor.extract vectorization (PR #99299)

[Andrzej Warzyński]

banach-space wrote:

> Hey! I was looking at the changes and thinking that we may end up introducing too much complexity if we add support for all these unit dimension special cases. 

Thanks for taking a look! Note that this is actual simplifying the current logic 😅 (and reducing the number of special cases).

> I think we discussed this in the past and decided to leverage the drop unit dimension utilities for that. For this particular case, could we generate a gather and then have a canonicalization pattern that turns it into a load when the unit dimension is removed?

This is still not obvious to me, tbh. My concern is that generating gather loads introduces a chain of not-so-trivial address calculation Ops. But this is definitely worth exploring - I agree that we should reduce the complexity in the vectoriser. 

And yes, we have discussed this in the past - I finally have some cycles to revisit :)

https://github.com/llvm/llvm-project/pull/99299