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>

$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() {}

$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() {}

	DR5:  "DR5",
	DR6:  "DR6",
	DR7:  "DR7",
	DR8:  "DR8",
	DR9:  "DR9",
	DR10: "DR10",
	DR11: "DR11",
	DR12: "DR12",
	DR13: "DR13",
	DR14: "DR14",
	DR15: "DR15",
	TR0:  "TR0",
	TR1:  "TR1",
	TR2:  "TR2",
	TR3:  "TR3",
	TR4:  "TR4",
	TR5:  "TR5",
	TR6:  "TR6",
	TR7:  "TR7",

	"CR9",
	"CR10",
	"CR11",
	"CR12",
	"CR13",
	"CR14",
	"CR15",
	"DR0", // [D_DR]
	"DR1",
	"DR2",
	"DR3",
	"DR4",
	"DR5",
	"DR6",
	"DR7",
	"TR0", // [D_TR]
	"TR1",
	"TR2",
	"TR3",
	"TR4",
	"TR5",
	"TR6",
	"TR7",
	"TLS",    // [D_TLS]
	"MAXREG", // [MAXREG]
}

func init() {
	obj.RegisterRegister(REG_AL, REG_AL+len(Register), rconv)

  // These are needed as the Nokia Symbian Compiler cannot decide between
  // const T& and const T* in a function template. The Nokia compiler _can_
  // decide between class template specializations for T and T*, so a
  // tr1::type_traits-like is_pointer works, and we can overload on that.
  template <typename T>
  inline void StreamHelper(internal::true_type /*is_pointer*/, T* pointer) {
    if (pointer == NULL) {
      *ss_ << "(null)";

  // These are needed as the Nokia Symbian Compiler cannot decide between
  // const T& and const T* in a function template. The Nokia compiler _can_
  // decide between class template specializations for T and T*, so a
  // tr1::type_traits-like is_pointer works, and we can overload on that.
  template <typename T>
  inline void StreamHelper(internal::true_type /*is_pointer*/, T* pointer) {
    if (pointer == NULL) {
      *ss_ << "(null)";