[Mlir-commits] [mlir] [MLIR][XeGPU] XeGPU Layout adds support for fractional-subgroup-size vector (PR #183434)

[Artem Kroviakov]

================
@@ -58,7 +57,7 @@ func.func @dpas_f16(%arg0: memref<8x16xf16>, %arg1: memref<16x16xf16>, %arg2: me
 
 // -----
 gpu.module @test_kernel {
-  gpu.func @elementwise_with_inst_data_only(%A: memref<1024x1024xf16>, %B: memref<1024x1024xf16>, %C: memref<1024x1024xf16>) {
+  func.func @elementwise_with_inst_data_only(%A: memref<1024x1024xf16>, %B: memref<1024x1024xf16>, %C: memref<1024x1024xf16>) {
----------------
akroviakov wrote:

Certain passes tend to retrieve op's parent function/module, which is not necessarily `module` or `func.func`, so while lits pass, real use cases crash.

https://github.com/llvm/llvm-project/pull/183434