[Mlir-commits] [mlir] [MLIR][XeGPU] Xegpu distribution patterns for load_nd, store_nd, and create_nd_tdesc. (PR #112945)

[llvmlistbot at llvm.org]

MaheshRavishankar wrote:

> This PR introduces distribution patterns for a portion of xegpu dialect similarly to the vector dialect, as well as moving some of the common functionality to the vector utilities.
> 
> Xegpu ops rewrite patterns distribute the vector and xegpu tensor descriptor types when sinked through yield op of a `vector.warp_execute_on_lane_0` according to the `xegpu.sg_map` attribute. The validation of distributed types in the `warp_execute_on_lane_0` was hence relaxed to allow `ShapedType` return values to have distributed shapes.

I dont think relaxing a `vector` verification to allow `ShapedType` makes sense. If you are really relying on vector distribution like thing, you probably need to implement an interface so that you can implement a different op with that interface. Thats just a suggestion. There are other suggestions here as well in terms of smaller PRs etc. that might help, but I think we need more information/better design to achieve what you are looking for. 

https://github.com/llvm/llvm-project/pull/112945