[Mlir-commits] [mlir] [mlir] [linalg] Check for dim shape to decide unit dim for each operand in dropUnitDims pass. (PR #91673)

[Sayan Saha]

================
@@ -1087,3 +1087,47 @@ func.func @drop_known_unit_constant_low_high(%arg0: tensor<1x383x128xf32>) -> te
 //       CHECK:   } : tensor<383x128xf32> to tensor<384x128xf32>
 //       CHECK:   tensor.expand_shape %[[PADDED]]
 //  CHECK-SAME:     {{\[}}[0, 1], [2]] output_shape [1, 384, 128] : tensor<384x128xf32> into tensor<1x384x128xf32>
+
+
+// -----
+
+// CHECK: #[[$MAP0:.+]] = affine_map<()[s0, s1] -> (s0 * s1)>
+// CHECK: #[[$MAP1:.+]] = affine_map<(d0) -> (0, d0)>
+// CHECK: #[[$MAP2:.+]] = affine_map<(d0) -> ()>
+
+// CHECK-LABEL: func @drop_unit_dim_corresponding_to_dynamic_dim
+// CHECK-SAME:                    %[[ARG0:.*]]: tensor<1x?x?x1xf32>,
+// CHECK-SAME:                    %[[ARG1:.*]]: index) -> tensor<?x1x61x1xf32> {
+// CHECK:           %[[VAL_0:.*]] = arith.constant 0 : index
+// CHECK:           %[[VAL_1:.*]] = arith.constant 1 : index
+// CHECK:           %[[VAL_2:.*]] = arith.constant dense<1.000000e+00> : tensor<f32>
+// CHECK:           %[[VAL_3:.*]] = tensor.collapse_shape %[[ARG0]] {{\[\[}}0, 1], [2, 3]] : tensor<1x?x?x1xf32> into tensor<?x?xf32>
+// CHECK:           %[[VAL_4:.*]] = tensor.empty(%[[ARG1]]) : tensor<?x61xf32>
+// CHECK:           %[[VAL_5:.*]] = affine.apply #[[$MAP0]](){{\[}}%[[ARG1]], %[[VAL_1]]]
+// CHECK:           %[[VAL_6:.*]] = tensor.empty(%[[VAL_5]]) : tensor<?x61xf32>
+// CHECK:           %[[VAL_7:.*]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP2]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel"]} ins(%[[VAL_3]], %[[VAL_2]], %[[VAL_4]] : tensor<?x?xf32>, tensor<f32>, tensor<?x61xf32>) outs(%[[VAL_6]] : tensor<?x61xf32>) {
----------------
sahas3 wrote:

Seems like adding the verifier to `verifyStructuredOpInterface` is an option to enable it for all `Linalg` ops. But that caused some failures for existing passes:

`mlir-opt ---pass-pipeline="builtin.module(func.func(tosa-to-linalg-named))"`
```
func.func @matmul_dyn_output(%arg0: tensor<1x1x8xf32>, %arg1: tensor<1x8x1xf32>) -> tensor<?x1x1xf32> {
  %0 = tosa.matmul %arg0, %arg1 : (tensor<1x1x8xf32>, tensor<1x8x1xf32>) -> tensor<?x1x1xf32>
  return %0 : tensor<?x1x1xf32>
}
```

will lead to an error because the `output` dim `0` is dynamic though it should be `1` based on input dims. Running `tosa-infer-shapes` before `tosa-to-linalg-named` solves the problem here but seems like adding the verifier can cause spurious errors like this?


https://github.com/llvm/llvm-project/pull/91673