grad_outputs->push_back(Gather(scope, grad_inputs[0], index));
  }

  return scope.status();
}
REGISTER_GRADIENT_OP("DynamicStitch", DynamicStitchGrad);
REGISTER_GRADIENT_OP("ParallelDynamicStitch", DynamicStitchGrad);

}  // anonymous namespace
}  // namespace ops

  func.return %0 : tensor<1x2xf32>
}

// CHECK-LABEL: func @ParallelDynamicStitch
func.func @ParallelDynamicStitch(%arg0: tensor<2x2xf32>) -> tensor<2x2xf32> {
  %indices = "tf.Const"() {value = dense<[1, 0]> : tensor<2xi32>} : () -> tensor<2xi32>
  // CHECK-NOT: tf.ParallelDynamicStitch
  %0 = "tf.ParallelDynamicStitch"(%indices, %arg0) : (tensor<2xi32>, tensor<2x2xf32>) -> tensor<2x2xf32>

      "MaxPoolGradGradV2",
      "MaxPoolGradV2",
      "MaxPoolV2",
      "Multinomial",
      "NextAfter",
      "NonMaxSuppressionV3",
      "NonMaxSuppressionV4",
      "ParallelDynamicStitch",
      "ParameterizedTruncatedNormal",
      "PartitionedCall",
      "Polygamma",
      "PopulationCount",
      "Qr",
      "QuantizeAndDequantizeV2",
      "QuantizeAndDequantizeV3",

Values are merged in order, so if an index appears in both `indices[m][i]` and
`indices[n][j]` for `(m,i) < (n,j)` the slice `data[n][j]` will appear in the
merged result. If you do not need this guarantee, ParallelDynamicStitch might
perform better on some devices.

For example:

```python
    indices[0] = 6
    indices[1] = [4, 1]
    indices[2] = [[5, 2], [0, 3]]
    data[0] = [61, 62]