Add TODO and failing test case

Author: nbpatel

mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp

@@ -68,6 +68,11 @@ namespace {
 ///
 /// Since the 24x24 matrix is divided into 8x8 distribution units, there will be
 /// 9 distribution units (3x3) in total. Hence the 9 subgroup level operations.
+
+/// The pass currently has entire distribution logic in the WgToSgCreateNdOp
+/// pattern and all the other ops just follow.
+/// TODO: Decouple the distribution logic from WgToSgCreateNdOp for all the
+/// ops in the pass.
 struct WgToSgCreateNdOp : public OpConversionPattern<xegpu::CreateNdDescOp> {
   using OpConversionPattern<xegpu::CreateNdDescOp>::OpConversionPattern;
 

mlir/test/Dialect/XeGPU/xegpu-wg-to-sg.mlir

@@ -121,4 +121,13 @@ gpu.func @test_dpas(%a: memref<24x32xf32>, %b: memref<32x24xf32>) {
       : !xegpu.tensor_desc<24x32xf32, #xegpu.layout<sg_layout = [2, 4], sg_data = [12, 8], lane_layout = [2, 8], lane_data = [1, 1]>>
     gpu.return
   }
-}
+
+  // CHECK-LABEL: test_dpas_with_no_create_nd_desc
+  gpu.func @test_dpas_with_no_create_nd_desc(%a: vector<24x32xf32>, %b: vector<32x24xf32>) {
+    // CHECK-NOT: vector<12x12xf32>
+    %dpas = xegpu.dpas %a, %b
+      {layout =  #xegpu.layout<sg_layout = [2, 2], sg_data = [12, 12], lane_layout = [2, 2], lane_data = [1, 1]>}
+      : vector<24x32xf32>, vector<32x24xf32> -> vector<24x24xf32>
+    gpu.return
+  }
+}