[Mlir-commits] [mlir] [MLIR][Linalg] Add pass to convert linalg.generic back to named ops (PR #95656)

[Javed Absar]

================
@@ -58,6 +68,195 @@ static bool areBinOpsSwapped(GenericOp genericOp) {
   return swapped;
 }
 
+//===----------------------------------------------------------------------===//
+// Specialize linalg generic to matmul variants.
+//===----------------------------------------------------------------------===//
+/// Identifies linalg.generic that is essentially named op of the form:
+//    ` linalg.{batch_}?matmul{_transpose_a | _transpose_b}? `
+//
+// It is possible that a linalg.generic may be implementing a matmul but not
+// in a straight-forward way e.g. below is matrix multiply over some slice
+// ```
+//  %0 = linalg.generic {
+//          indexing_maps = [affine_map<(d0, d1, d2) -> (3, d1, d0)>,
+//                           affine_map<(d0, d1, d2) -> (d0, 5, d2)>,
+//                           affine_map<(d0, d1, d2) -> (d2, d1, 13)>],
+//          iterator_types = ["parallel", "parallel", "parallel"]}
+//          ins(%A, %B : tensor<20x20x20xf32>,  tensor<20x20x20xf32>)
+//          outs(%C : tensor<20x20x20xf32>) {
+//             ^bb0(%a: f32, %b: f32, %c : f32):
+//                %mul = arith.mulf %a, %b : f32
+//                %add = arith.addf %mul, %c : f32
+//                linalg.yield %add : f32
+//       } -> tensor<20x20x20xf32>
+// ```
+// It is not possible to represent above as named op.
+// e.g. linalg.batch_matmul(%A, %B :  tensor<20x20x20xf32>, ...) is
+// not  the same as linalg.generic above.
+namespace {
+enum class IndexMatchResult {
+  Match = 0,  // identity map.
----------------
javedabsar1 wrote:

Yeah. its a taste issue. the class is '*Match*'. there is another field mis-match..

https://github.com/llvm/llvm-project/pull/95656