[Mlir-commits] [mlir] [mlir][xegpu] Add initial skeleton implementation for lowering ConvertLayoutOp (PR #146176)

[Chao Chen]

================
@@ -78,6 +78,20 @@ resolveUnrealizedConversionCastOp(UnrealizedConversionCastOp castOp) {
   }
 }
 
+struct ConvertLayoutOpPattern
+    : public OpRewritePattern<xegpu::ConvertLayoutOp> {
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(xegpu::ConvertLayoutOp op,
+                                PatternRewriter &rewriter) const override {
+    xegpu::LayoutAttr input_layout = op.getInputLayoutAttr().dropInstData();
+    xegpu::LayoutAttr target_layout = op.getTargetLayoutAttr().dropInstData();
----------------
chencha3 wrote:

UnrollPatterns always use ExtractStridedSliceOp to unroll the values and use InsertStrideSliceOp at the end to reconstruct the vector with original shape. So, the input and output value of the covertLayoutOp is always with original vector type.

https://github.com/llvm/llvm-project/pull/146176