[Mlir-commits] [mlir] [MLIR][XeGPU] Support order attribute and add pattern for vector.transpose in WgToSg Pass (PR #165307)

[Nishant Patel]

================
@@ -411,29 +408,32 @@ gpu.module @test_distribution {
   // CHECK-LABEL: vector_step_op
   gpu.func @vector_step_op_slice_attr() {
     //CHECK: [[sgId:%.+]] = gpu.subgroup_id : index
-    //CHECK-DAG: [[IDY:%.+]] = affine.apply #map2()[[[sgId]]]
-    //CHECK-DAG: [[c32:%.+]] = arith.constant 32 : index
-    //CHECK-DAG: [[LY:%.+]] = index.mul [[IDY]], [[c32]]
-    //CHECK-DAG: [[c0:%.+]] = arith.constant 0 : index
-    //CHECK-DAG: [[c128:%.+]] = arith.constant 128 : index
-    //CHECK-DAG: [[MODY:%.+]] = index.remu [[LY]], [[c128]]
-    //CHECK-DAG: [[BASE:%.+]] = vector.step : vector<32xindex>
-    //CHECK-DAG: [[CAST:%.+]] = vector.broadcast [[MODY]] : index to vector<32xindex>
+    //CHECK: [[c8:%.+]] = arith.constant 8 : index
+    //CHECK: [[sgidx:%.+]] = index.remu [[sgId]], [[c8]]
+    //CHECK: [[sgidy_tmp:%.+]] = index.divu [[sgId]], [[c8]]
+    //CHECK: [[c4:%.+]] = arith.constant 4 : index
+    //CHECK: [[sgidy:%.+]] = index.remu [[sgidy_tmp]], [[c4]]
+    //CHECK: [[c32:%.+]] = arith.constant 32 : index
+    //CHECK: [[LY:%.+]] = index.mul [[sgidy]], [[c32]]
+    //CHECK: [[c128:%.+]] = arith.constant 128 : index
+    //CHECK: [[MODY:%.+]] = index.remu [[LY]], [[c128]]
+    //CHECK: [[BASE:%.+]] = vector.step : vector<32xindex>
+    //CHECK: [[CAST:%.+]] = vector.broadcast [[MODY]] : index to vector<32xindex>
----------------
nbpatel wrote:

yes, not using affine maps to delinearize any more since that helper doesn't support order

https://github.com/llvm/llvm-project/pull/165307