[Mlir-commits] [mlir] [MLIR][XeGPU] Lower vector.multi_reduction before linearization in XeGPUVectorLinearize (PR #190272)

[Jianhui Li]

================
@@ -265,4 +265,21 @@ gpu.module @test_kernel {
   }
 }
 
+// -----
+// CHECK-LABEL: func.func @test_vector_multi_reduction_add
+// CHECK-SAME: (%[[ARG0:.*]]: vector<16x1xf16>, %[[ARG1:.*]]: vector<1xf16>) -> vector<1xf16>
----------------
Jianhui-Li wrote:

The code look very inefficient to me.  Why it is not just removing the training dimension and use a vector.reduction of 16 elements? 



https://github.com/llvm/llvm-project/pull/190272