if os.IsNotExist(err) {
		return nil, err
	}

	return &AtomicWriter{targetDir: targetDir, logContext: logContext}, nil
}

const (
	dataDirName    = "..data"
	newDataDirName = "..data_tmp"
)

// Write does an atomic projection of the given payload into the writer's target
// directory.  Input paths must not begin with '..'.
// setPerms is an optional pointer to a function that caller can provide to set the

  for (auto data_type : llvm::enumerate(input_dtypes_vector)) {
    if (!DataType_Parse(data_type.value(), &data_types[data_type.index()]))
      return errors::InvalidArgument("Invalid dtype at index ",
                                     data_type.index(), ": ",
                                     data_type.value());
    const auto& resolved_dtype = data_types[data_type.index()];

  owner->reference.Unref();
  delete owner;
}

// Converts TF_DATAType to DLPack data type.
DLDataType GetDlDataType(TF_DataType data_type, TF_Status* status) {
  DLDataType dtype;
  dtype.lanes = 1;
  dtype.bits = TF_DataTypeSize(data_type) * 8;
  switch (data_type) {
    case TF_DataType::TF_BOOL:
      dtype.code = DLDataTypeCode::kDLBool;
      break;
    case TF_DataType::TF_HALF:

  enum OpType { RESIZE_NEAREST, RESIZE_BILINEAR, RESIZE_BICUBIC };

  template <typename T>
  Tensor MakeData(const TensorShape& data_shape) {
    DataType data_type = DataTypeToEnum<T>::v();
    Tensor data(data_type, data_shape);
    auto data_flat = data.flat<T>();
    for (int i = 0; i < data_flat.size(); ++i) {
      data_flat(i) = T(i);
    }
    return data;
  }

  template <typename T>

  DataType data_type = DT_INT32;
  std::initializer_list<int64_t> dims = {2, 3, 4, 5};
  Tensor tensor(data_type, TensorShape(dims));
  for (int i = 0; i < 2 * 3 * 4 * 5; ++i) {
    tensor.flat<int32>()(i) = i;
  }

  NodeDef node;
  auto builder = NodeDefBuilder("some_node", "Const")
                     .Attr("dtype", data_type)
                     .Attr("value", tensor);

          hoisted_read(nullptr),
          data_type(nullptr),
          result_index(-1) {}

    bool IsResultIndexAssigned() { return result_index != -1; }

    // Refine the resource type using the given type `type`.
    void RefineType(Type type) {
      if (!data_type) {
        data_type = type;
      } else {
        data_type = TF::GetCastCompatibleType(data_type, type,

Type TensorFlowExecutorDialect::parseType(DialectAsmParser &parser) const {
  StringRef data_type;
  if (parser.parseKeyword(&data_type)) return Type();

  if (data_type == "control") return ControlType::get(getContext());
  if (data_type == "token") return TokenType::get(getContext());
  parser.emitError(parser.getNameLoc())
      << "unknown tf_executor type: " << data_type;
  return nullptr;
}

  }
  if (inp.node()) {
    return NodeBuilder::NodeOut(inp.node(), inp.index());
  }
  if (!inp.node_name().empty()) {
    return NodeBuilder::NodeOut(inp.node_name(), inp.index(), inp.data_type());
  }
  auto transformed = Input{
      Const(scope.NewSubScope("Const"), Input::Initializer(inp.tensor()))};
  return NodeBuilder::NodeOut{transformed.node(), transformed.index()};
}

                             const int arg_idx, const DataType& data_type,
                             Node* arg_node) {
  NodeDefBuilder ret_builder(absl::StrCat("ret_", arg_idx), "_Retval");
  ret_builder.Attr("T", data_type);
  ret_builder.Attr("index", arg_idx);
  ret_builder.Input(arg_node->name(), 0, data_type);
  NodeDef ret_def;
  TF_RETURN_IF_ERROR(ret_builder.Finalize(&ret_def));

  static ArgType CreateInput(const OpDef::ArgDef& arg_def);
  static ArgType CreateInputRef(const OpDef::ArgDef& arg_def);
  static ArgType CreateOutput(const OpDef::ArgDef& arg_def);

  const tensorflow::DataType data_type() const { return data_type_; }
  const string type_attr_name() const { return type_attr_name_; }
  const bool is_read_only() const { return kind_ == kInput; }
  const bool is_list() const { return is_list_; }

 private: