[Mlir-commits] [mlir] [MLIR][XeGPU] Distribute create_nd_desc op without offset from Wg to Sg (PR #152351)

[Jianhui Li]

================
@@ -161,6 +161,18 @@ struct WgToSgCreateNdOp : public OpConversionPattern<xegpu::CreateNdDescOp> {
   LogicalResult
   matchAndRewrite(xegpu::CreateNdDescOp op, OneToNOpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override {
+
+    // Ensure that the op has explicit offsets specified (either dynamic or
+    // constant).
+    int64_t offsetSize = static_cast<int64_t>(op.getOffsets().size());
+    if (offsetSize == 0) {
+      auto constOffsetsAttr = op.getConstOffsetsAttr();
+      if (!constOffsetsAttr || constOffsetsAttr.empty() ||
+          llvm::all_of(constOffsetsAttr.asArrayRef(),
----------------
Jianhui-Li wrote:

I think we'd better leave create_nd_tdesc as is. and only work on load_nd/store_nd/prefetch_nd but override the offset set by create_nd_tdesc distribution.  

https://github.com/llvm/llvm-project/pull/152351