[Mlir-commits] [mlir] [MLIR][XeGPU] Add 2D `vector.multi_reduction` optimization (PR #171154)

[Artem Kroviakov]

================
@@ -77,6 +77,7 @@ void buildGPUPassPipeline(OpPassManager &pm,
     layoutOptions.layoutKind = "lane";
     pm.addNestedPass<gpu::GPUModuleOp>(
         xegpu::createXeGPUPropagateLayout(layoutOptions));
+    pm.addNestedPass<gpu::GPUModuleOp>(xegpu::createXeGPUOptimizeBlockLoads());
----------------
akroviakov wrote:

The new pattern expects layouts to be present and respects the layouts when creating new ops, similar to the load nd pattern in the same pass. Another propagation would not add new info, or do you have a specific case in mind?

https://github.com/llvm/llvm-project/pull/171154