[Mlir-commits] [mlir] [mlir][gpu] Add innermost-first policy when mapping loops to GPU IDs (PR #160634)

[Fabian Mora]

fabianmcg wrote:

> HW mapping is rather cheap compared to loop reordering in a way; especially if we unsettle the IR and cause any issues with analysis that we might have in place.

What I'm claiming is that we should have a loop interchange transform and use that to take care of this. However, I recognize the amount of work on that side is significantly greater. Could you add a TODO, saying replace this once we have an interchange/reordering transform?

https://github.com/llvm/llvm-project/pull/160634