(TF_MapAndBatchDatasetOp $input_dataset, $other_arguments, $batch_size,
     (TF_ConstOp (GetI64ScalarElementsAttr<1>)), $drop_remainder, $f,
        $batch_output_types, $batch_output_shapes, $preserve_cardinality,
        // TODO(kramm): Should we merge batch_dataset_metadata and
        //              map_dataset_metadata?

    // Which dimensions (in the input) the two offset "columns" map to.
    SmallVector<int64_t, 2> start_index_map({num_dims - 2, num_dims - 1});

    // Gather the diagonal entries.
    // TODO(kramm): For a single diagonal, this might be slower than the
    //              mask + sum approach. Special-case num_diags==1?
    auto dims_attr = GatherDimensionNumbersAttr::get(
        rewriter.getContext(),