lhs_batch_idx.insert(dot_dimension_numbers.lhs_batch_dimensions().begin(),
                       dot_dimension_numbers.lhs_batch_dimensions().end());
  lhs_contract_idx.insert(
      dot_dimension_numbers.lhs_contracting_dimensions().begin(),
      dot_dimension_numbers.lhs_contracting_dimensions().end());
  rhs_batch_idx.insert(dot_dimension_numbers.rhs_batch_dimensions().begin(),

                                const bool copy_left = true) {
  if (copy_left) {
    for (auto v : d1.lhs_batch_dimensions()) {
      d2.add_lhs_batch_dimensions(v);
    }
    for (auto v : d1.lhs_contracting_dimensions()) {
      d2.add_lhs_contracting_dimensions(v);
    }
  } else {
    for (auto v : d1.rhs_batch_dimensions()) {
      d2.add_rhs_batch_dimensions(v);
    }
    for (auto v : d1.rhs_contracting_dimensions()) {

    auto reshaped =
        rewriter.create<mhlo::ReshapeOp>(loc, result_type, matmul.getResult());
    return reshaped.getResult();
  }

  // Reshape for dynamic shaped operands. The result shape is
  // [lhs_batch_dimensions, lhs_out_dimensions, rhs_out_dimensions].
  auto lhs_shape = rewriter.create<TFL::ShapeOp>(
      loc,
      RankedTensorType::get(static_cast<int32_t>(lhs_type.getRank()),

        rewriter.create<mhlo::ReshapeOp>(loc, result_type, matmul.getResult());
    return reshaped.getResult();
  }

  // Reshape for dynamic shaped operands. The result shape is
  // [lhs_batch_dimensions, lhs_out_dimensions, rhs_out_dimensions].
  BoolAttr true_attr = rewriter.getBoolAttr(true);
  auto lhs_shape = rewriter.create<TF::ShapeOp>(loc, lhs, true_attr);