// only when height = width.
    if (params_type.getShape().size() != 4 ||
        indices_type.getShape().size() != 2)
      return failure();
    if (params_type.getShape()[1] != 1) return failure();
    if (params_type.getShape()[0] != params_type.getShape()[2])
      return failure();
    if (result_type.getShape()[0] != params_type.getShape()[0] * 2)
      return failure();

    if (!shaped_type || !shaped_type.hasStaticShape()) {
      return false;
    }
    if (operand_num == 0) {
      shape = shaped_type.getShape();
    } else {
      if (shape != shaped_type.getShape()) {
        return false;
      }
    }
    ++operand_num;
  }
  return true;
}

// Utility function to map final permutation to initial permutation

  dot_dimension_numbers.ParseFromString(dot_dimension_numbers_str.str());
  SmallVector<Value> input_arguments = {lhs, rhs};
  const int lhs_rank = mlir::cast<ShapedType>(lhs.getType()).getShape().size();
  const int rhs_rank = mlir::cast<ShapedType>(rhs.getType()).getShape().size();

  const std::string einsum_equation =
      CreateEinsumEquation(dot_dimension_numbers, lhs_rank, rhs_rank);

            mlir::cast<ShapedType>(lhs.getType()).getShape(),
            mlir::cast<ShapedType>(rhs.getType()).getShape(),
            symbolic_broadcast_shape)) {
      return failure();
    }

    // Calculates the broadcast shape using BroadcastArgs op.
    Value lhs_shape = GetShape(lhs, op->getLoc(), rewriter);
    Value rhs_shape = GetShape(rhs, op->getLoc(), rewriter);
    auto broadcast_shape =

                           Location loc) {
  auto buffer_type = mlir::cast<RankedTensorType>(buffer.getType());
  if (buffer_type.getShape().size() == 1) return index;
  // Create a concat of index and trailing zeros.
  llvm::SmallVector<int64_t, 8> zeros(buffer_type.getShape().size() - 1, 0);
  auto zeros_tensor = GetR1Const(zeros, builder, loc);
  return builder.create<TF::ConcatV2Op>(
      loc,

  if (auto tensor_type = dyn_cast<RankedTensorType>(input_type))
    return RankedTensorType::get(tensor_type.getShape(), elemental_type);
  if (auto tensor_type = dyn_cast<UnrankedTensorType>(input_type))
    return UnrankedTensorType::get(elemental_type);
  if (auto vector_type = dyn_cast<VectorType>(input_type))
    return VectorType::get(vector_type.getShape(), elemental_type);

  // If the expressed types match, just use the new elemental type.

  // Extract Z from %dot.
  ArrayRef<int64_t> shape_z =
      dot.getType().getShape().drop_front(shape_b.size() + shape_y2.size());

  // Check %after shape.
  if (reshape_after.getType().getShape() !=
      ArrayRef<int64_t>(llvm::to_vector(
          llvm::concat<const int64_t>(shape_b, shape_y1, shape_z)))) {
    return failure();
  }

  rewriter.replaceOpWithNewOp<mhlo::DotGeneralOp>(

      }
      for (auto idx = 0; idx < output_type.getShape().size(); idx++) {
        if (input_type.getShape()[idx] != output_type.getShape()[idx]) {
          op.emitOpError()
              << "the " << idx << "th dim of output tensor is "
              << output_type.getShape()[idx]
              << ". It is not equal to the one in input tensor, which is "
              << input_type.getShape()[idx];
          return failure();

  RankedTensorType tensorType = mlir::cast<RankedTensorType>(value.getType());
  Type element_type = tensorType.getElementType();

  int rank = tensorType.getShape().size();
  int num_rows = tensorType.getShape()[rank - 2];
  int num_cols = tensorType.getShape()[rank - 1];

  std::vector<Value> sliced;

  if (batch_size == 1) {
    // Batch size is 1, no splitting is required

      if (input_ty_i.hasStaticShape() &&
          input_ty_i.getShape() != input_ty_0.getShape()) {
        return op.emitOpError()
               << "inputs[" << i << "] has shape [" << input_ty_i.getShape()
               << "] different than the shape of inputs[0]: "
               << input_ty_0.getShape();
      }
    }

    if (op.getDimensionsToReduce().size() > input_ty_0.getRank()) {