[Mlir-commits] [mlir] [mlir][linalg] Add patterns to convert matmul to transposed variants (PR #89075)

[Oleksandr Alex Zinenko]

================
@@ -73,6 +73,21 @@ def ApplyTilingCanonicalizationPatternsOp : Op<Transform_Dialect,
   let assemblyFormat = "attr-dict";
 }
 
+def ApplyTransposeMatmulPatternsOp : Op<Transform_Dialect,
+    "apply_patterns.linalg.transpose_matmul",
+    [DeclareOpInterfaceMethods<PatternDescriptorOpInterface>]> {
+  let description = [{
+    Collects patterns to convert Linalg matmul ops to transposed variants.
+
+    By default the A matrix is transposed, set `transpose_a=false` to transpose
+    B matrix.
+  }];
+
+  let arguments = (ins DefaultValuedAttr<BoolAttr, "true">:$transpose_a);
----------------
ftynse wrote:

I find this interface/API highly confusing. `transpose_a = false` reads like we shouldn't be transposing anything at all. How about we have two unit attributes, `transpose_lhs` and `transpose_rhs` that indicate which of the operands we have to transpose? Alternatively, an enum or string attribute specifying which of the operands is transposed.

https://github.com/llvm/llvm-project/pull/89075