[Mlir-commits] [mlir] [mlir][xegpu] Refine layout assignment in XeGPU SIMT distribution. (PR #142687)
Chao Chen
llvmlistbot at llvm.org
Tue Jun 10 09:48:12 PDT 2025
================
@@ -27,10 +27,18 @@ def XeGPUSubgroupDistribute : Pass<"xegpu-subgroup-distribute"> {
}];
let dependentDialects = ["memref::MemRefDialect", "xegpu::XeGPUDialect",
"vector::VectorDialect"];
- let options = [Option<
- "printOnly", "print-analysis-only", "bool",
- /*default=*/"false",
- "Print the result of the subgroup map propagation analysis and exit.">];
+}
+
+def XeGPULayoutPropagate : Pass<"xegpu-layout-propagate"> {
----------------
chencha3 wrote:
I partially agree. We could have another case, e.g., after WgToSg lowering, the LayoutAttr for some values are missing. It happens when there is 1:N conversion. In such case, the LayoutAttr for some VectorType values (e.g., results of scf ops) are missing, in this case, we potentially need a forward pass to propagate the results. In comparison, I feel the logic of this pass is closer to determine a proper layout for each OpResult.
https://github.com/llvm/llvm-project/pull/142687
More information about the Mlir-commits
mailing list