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

[Nishant Patel]

================
@@ -273,26 +273,76 @@ LayoutAttr::verify(llvm::function_ref<mlir::InFlightDiagnostic()> emitError,
 FailureOr<SmallVector<Value>>
 LayoutAttr::delinearizeSubgroupId(OpBuilder &builder, Location loc,
                                   Value linearId) {
-  // delinearizeSubgroupId is only available for
-  // workgroup-level layout attribute
   if (!isForWorkgroup())
     return failure();
 
-  // TODO: handle order attribute
-  auto hasDefaultOrder = [&]() {
-    DenseI32ArrayAttr order = getOrder();
-    return !order || isIdentityPermutation(llvm::to_vector_of<int64_t>(
-                         llvm::reverse(order.asArrayRef())));
-  };
-  if (!hasDefaultOrder())
-    return mlir::emitError(loc, "order attribute is currently not supported.");
+  SmallVector<int64_t> sgLayoutInt = getEffectiveSgLayoutAsInt();
+  DenseI32ArrayAttr orderAttr = getOrder();
 
-  auto dims =
-      llvm::map_to_vector(getEffectiveSgLayoutAsInt(), [&](int64_t d) -> Value {
-        return builder.createOrFold<arith::ConstantIndexOp>(loc, d);
-      });
+  // Handle order attribute
+  SmallVector<int64_t> order;
+  if (orderAttr && !orderAttr.empty()) {
+    order = llvm::to_vector(
+        llvm::map_range(orderAttr.asArrayRef(),
+                        [](int32_t idx) { return static_cast<int64_t>(idx); }));
+  } else {
+    // Default order: [1, 0] for 2D (row-major), [2, 1, 0] for 3D, etc.
+    order = llvm::to_vector(
+        llvm::reverse(llvm::seq<int64_t>(0, sgLayoutInt.size())));
+  }
 
-  return affine::delinearizeIndex(builder, loc, linearId, dims);
+  if (order.size() != sgLayoutInt.size()) {
+    return failure();
+  }
+
+  SmallVector<Value> result(sgLayoutInt.size());
+  Value remaining = linearId;
+
+  /// Process dimensions in the order they appear in the order array
+  /// The first dimension in order is the fastest-changing
+  ///
+  /// Example walkthrough for linearId=22, sgLayout=[2,4,4], order=[2,1,0]:
+  ///
+  /// Initial: remaining=22, result=[?,?,?]
+  ///
+  /// i=0 (process columns, dimIdx=2, dimSize=4):
+  ///   result[2] = 22 % 4 = 2  (column coordinate)
+  ///   remaining = 22 / 4 = 5  (5 complete groups of 4 columns processed)
+  ///
+  /// i=1 (process rows, dimIdx=1, dimSize=4):
+  ///   result[1] = 5 % 4 = 1   (row coordinate)
+  ///   remaining = 5 / 4 = 1   (1 complete group of 4 rows processed)
+  ///
+  /// i=2 (process layers, dimIdx=0, dimSize=2):
+  ///   result[0] = 1 % 2 = 1   (layer coordinate)
+  ///   (no remaining update - last iteration)
+  ///
+  /// Final result: [1,1,2] = Layer 1, Row 1, Column 2
+  for (size_t i = 0; i < order.size(); ++i) {
+    int64_t dimIdx = order[i];
+    int64_t dimSize = sgLayoutInt[dimIdx];
+
+    Value dimSizeVal =
+        builder.createOrFold<arith::ConstantIndexOp>(loc, dimSize);
----------------
nbpatel wrote:

createOrFold is supported, just the create is deprecated

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