Output tr1 = ops::Transpose(on_gpu.WithOpName("tr1"), in1, perm);
  }

  std::unique_ptr<Graph> result;
  TF_ASSERT_OK(CloneConstantsForBetterClustering(root, &result));

  OutputTensor tr0_perm;
  TF_ASSERT_OK(FindNodeByName(result.get(), "tr0")->input_tensor(1, &tr0_perm));

  OutputTensor tr1_perm;
  TF_ASSERT_OK(FindNodeByName(result.get(), "tr1")->input_tensor(1, &tr1_perm));

template <typename T>
struct ByRef<T&> { typedef T& type; };  // NOLINT

// A handy wrapper for ByRef.
#define GTEST_BY_REF_(T) typename ::std::tr1::gtest_internal::ByRef<T>::type

// AddRef<T>::type is T if T is a reference; otherwise it's T&.  This
// is the same as tr1::add_reference<T>::type.
template <typename T>
struct AddRef { typedef T& type; };  // NOLINT
template <typename T>

template <typename T>
struct ByRef<T&> { typedef T& type; };  // NOLINT

// A handy wrapper for ByRef.
#define GTEST_BY_REF_(T) typename ::std::tr1::gtest_internal::ByRef<T>::type

// AddRef<T>::type is T if T is a reference; otherwise it's T&.  This
// is the same as tr1::add_reference<T>::type.
template <typename T>
struct AddRef { typedef T& type; };  // NOLINT
template <typename T>

// tuples of up-to 10 fields.  The following implementation works
// regardless of whether tr1::tuple is implemented using the
// non-standard variadic template feature or not.

inline void PrintTo(const ::std::tr1::tuple<>& t, ::std::ostream* os) {
  PrintTupleTo(t, os);
}

template <typename T1>
void PrintTo(const ::std::tr1::tuple<T1>& t, ::std::ostream* os) {
  PrintTupleTo(t, os);
}

// tuples of up-to 10 fields.  The following implementation works
// regardless of whether tr1::tuple is implemented using the
// non-standard variadic template feature or not.

inline void PrintTo(const ::std::tr1::tuple<>& t, ::std::ostream* os) {
  PrintTupleTo(t, os);
}

template <typename T1>
void PrintTo(const ::std::tr1::tuple<T1>& t, ::std::ostream* os) {
  PrintTupleTo(t, os);
}

// Until version 4.3.2, gcc has a bug that causes <tr1/functional>,
// which is #included by <tr1/tuple>, to not compile when RTTI is
// disabled.  _TR1_FUNCTIONAL is the header guard for
// <tr1/functional>.  Hence the following #define is a hack to prevent
// <tr1/functional> from being included.
#   define _TR1_FUNCTIONAL 1
#   include <tr1/tuple>
#   undef _TR1_FUNCTIONAL  // Allows the user to #include

$range i 2..maxtuple
$for i [[
$range j 1..i
$range k 2..i

template <$for j, [[typename T$j]]>
class CartesianProductGenerator$i
    : public ParamGeneratorInterface< ::std::tr1::tuple<$for j, [[T$j]]> > {
 public:
  typedef ::std::tr1::tuple<$for j, [[T$j]]> ParamType;

  CartesianProductGenerator$i($for j, [[const ParamGenerator<T$j>& g$j]])
      : $for j, [[g$(j)_(g$j)]] {}
  virtual ~CartesianProductGenerator$i() {}

// Until version 4.3.2, gcc has a bug that causes <tr1/functional>,
// which is #included by <tr1/tuple>, to not compile when RTTI is
// disabled.  _TR1_FUNCTIONAL is the header guard for
// <tr1/functional>.  Hence the following #define is a hack to prevent
// <tr1/functional> from being included.
#   define _TR1_FUNCTIONAL 1
#   include <tr1/tuple>
#   undef _TR1_FUNCTIONAL  // Allows the user to #include

$range i 2..maxtuple
$for i [[
$range j 1..i
$range k 2..i

template <$for j, [[typename T$j]]>
class CartesianProductGenerator$i
    : public ParamGeneratorInterface< ::std::tr1::tuple<$for j, [[T$j]]> > {
 public:
  typedef ::std::tr1::tuple<$for j, [[T$j]]> ParamType;

  CartesianProductGenerator$i($for j, [[const ParamGenerator<T$j>& g$j]])
      : $for j, [[g$(j)_(g$j)]] {}
  virtual ~CartesianProductGenerator$i() {}

template <typename T>
struct ByRef<T&> { typedef T& type; };  // NOLINT

// A handy wrapper for ByRef.
#define GTEST_BY_REF_(T) typename ::std::tr1::gtest_internal::ByRef<T>::type

// AddRef<T>::type is T if T is a reference; otherwise it's T&.  This
// is the same as tr1::add_reference<T>::type.
template <typename T>
struct AddRef { typedef T& type; };  // NOLINT
template <typename T>