[Mlir-commits] [mlir] [mlir][gpu] reverse parallel loop to gpu dimension mapping order. (PR #79592)

[Jungwook Park]

jungpark-mlir wrote:

Thanks for the comment @grypp ! glad to know that.
I've found this [comment](https://discourse.llvm.org/t/understanding-parallel-loops-to-gpu-lowering/727/4?u=jungpark), it comes from several years back but I still agree and want to keep this mapping as simple as possible.
e.g., We might be able to merge multiple loops to gain enough trip count for the thread mapping. Also a loop can be split into two part so that we can map each part to blocks and threads.
I'd try to make further optimizations as loop transforms first and expand the mapping itself only when necessary.

Surely we need a place to keep those analysis and transforms together, any idea?

https://github.com/llvm/llvm-project/pull/79592