[Mlir-commits] [mlir] [mlir][xegpu] Fix shape_cast source layout inference for split dims (PR #226178)

[llvmlistbot at llvm.org]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir

Author: Naaz30

<details>
<summary>Changes</summary>

inferShapeCastSourceLayout collapsed sg_layout and sg_data independently when a 1-D source was split into several result dims. That is only valid when each subgroup's elements stay a single round-robin pattern after row-major flattening. For sg_layout=[1,2], sg_data=[2,2] on 2x4 it inferred sg_data=[4] instead of [2], so the source and result disagreed on which elements each subgroup owns.

Fixes #<!-- -->224655

---
Full diff: https://github.com/llvm/llvm-project/pull/226178.diff


2 Files Affected:

- (modified) mlir/lib/Dialect/XeGPU/Transforms/XeGPULayoutImpl.cpp (+77-2) 
- (modified) mlir/test/Dialect/XeGPU/propagate-layout-subgroup.mlir (+33) 


``````````diff
The server is unavailable at this time. Please wait a few minutes before you try again.
``````````

</details>


https://github.com/llvm/llvm-project/pull/226178