[Mlir-commits] [mlir] Add missing FillOp to winograd lowering (PR #108181)

[llvmlistbot at llvm.org]

================
@@ -44,9 +51,13 @@ module attributes {transform.with_named_sequence} {
 // CHECK:  %[[S1:.*]] = scf.for %[[ARG3:.*]] = %[[C0]] to %[[C2]] step %[[C1]] iter_args(%[[ARG4:.*]] = %[[S0]])
 // CHECK:    %[[S9:.*]] = scf.for %[[ARG5:.*]] = %[[C0]] to %[[C5]] step %[[C1]] iter_args(%[[ARG6:.*]] = %[[ARG4]])
 // CHECK:      %[[EXTRACTED_SLICE:.*]] = tensor.extract_slice %[[ARG1]][%[[ARG3]], 0, 0, %[[ARG5]]] [1, 3, 3, 1] [1, 1, 1, 1]
-// CHECK:      %[[S11:.*]] = linalg.matmul
-// CHECK:      %[[S13:.*]] = linalg.matmul
-// CHECK:      %[[INSERTED_SLICE:.*]] = tensor.insert_slice %[[S13]] into %[[ARG6]][0, 0, %[[ARG5]], %[[ARG3]]] [6, 6, 1, 1] [1, 1, 1, 1]
+// CHECK:      %[[S10:.*]] = tensor.empty() : tensor<6x3xf32>
+// CHECK:      %[[S11:.*]] = linalg.fill ins(%[[CST_6]] : f32) outs(%[[S10]] : tensor<6x3xf32>) -> tensor<6x3xf32>
+// CHECK:      %[[S12:.*]] = linalg.matmul ins(%[[CST_5]], %[[EXTRACTED_SLICE]] : tensor<6x3xf32>, tensor<3x3xf32>) outs(%[[S11]] : tensor<6x3xf32>) -> tensor<6x3xf32>
+// CHECK:      %[[S13:.*]] = tensor.empty() : tensor<6x6xf32>
+// CHECK:      %[[S14:.*]] = linalg.fill ins(%[[CST_6]] : f32) outs(%[[S13]] : tensor<6x6xf32>) -> tensor<6x6xf32>
+// CHECK:      %[[S15:.*]] = linalg.matmul ins(%[[S12]], %[[CST_4]] : tensor<6x3xf32>, tensor<3x6xf32>) outs(%14 : tensor<6x6xf32>) -> tensor<6x6xf32>
----------------
Max191 wrote:

```suggestion
// CHECK:      %[[S13:.*]] = tensor.empty() : tensor<6x6xf32>
// CHECK:      %[[S14:.*]] = linalg.fill ins(%[[CST_6]] : f32) outs(%[[S13]] : tensor<6x6xf32>) -> tensor<6x6xf32>
// CHECK:      %[[S15:.*]] = linalg.matmul ins(%[[S12]], %[[CST_4]] : tensor<6x3xf32>, tensor<3x6xf32>) outs(%[[S14]] : tensor<6x6xf32>) -> tensor<6x6xf32>
```

https://github.com/llvm/llvm-project/pull/108181