CODEWITHSUNDEEP
c++templatestensoreigen3
Lukas Engelke
Lukas Engelke

Reputation: 174

Function taking Eigen::Tensor - template argument deduction fails

I'm trying to write a templated function taking an Eigen::Tensor as an argument. The same approach that works for Eigen::Matrix etc. does not work here.

Eigen recommends writing functions using a common base class. https://eigen.tuxfamily.org/dox/TopicFunctionTakingEigenTypes.html

A minimal example for Eigen::Matrix that compiles:

#include <Eigen/Dense>

template <typename Derived>
void func(Eigen::MatrixBase<Derived>& a) 
{
    a *= 2;
}

int main()
{
    Eigen::Matrix<int, 2, 2> matrix;
    func(matrix);
}

And the minimal example for Eigen::Tensor that does not compile:

#include <unsupported/Eigen/CXX11/Tensor>

template <typename Derived>
void func(Eigen::TensorBase<Derived>& a)
{
    a *= 2;
}

int main()
{
    Eigen::Tensor<int, 1> tensor;
    func(tensor);
}

$ g++ -std=c++11 -I /usr/include/eigen3 eigen_tensor_func.cpp
eigen_tensor_func.cpp: In function ‘int main()’:
eigen_tensor_func.cpp:12:16: error: no matching function for call to ‘func(Eigen::Tensor<int, 1>&)’
     func(tensor);
                ^
eigen_tensor_func.cpp:4:6: note: candidate: ‘template<class Derived> void func(Eigen::TensorBase<Derived>&)’
 void func(Eigen::TensorBase<Derived>& a)
      ^~~~
eigen_tensor_func.cpp:4:6: note:   template argument deduction/substitution failed:
eigen_tensor_func.cpp:12:16: note:   ‘Eigen::TensorBase<Derived>’ is an ambiguous base class of ‘Eigen::Tensor<int, 1>’
     func(tensor);

Upvotes: 4

Views: 317

Answers (1)

chtz
chtz

Reputation: 18807

The Tensor-Module is still far away from being fully compatible with the Eigen/Core functionality (this also implies the documentation of the core functionality does not necessarily apply to the Tensor-Module, of course).

First major difference is that TensorBase takes two template arguments instead of one, i.e., you need to write TensorBase<Derived, Eigen::WriteAccessors>. Also some functionality is either not implemented at all, or TensorBase does not properly forward it. The following works with current trunk (2019-04-03):

template <typename Derived>
void func(Eigen::TensorBase<Derived, Eigen::WriteAccessors>& a)
{
    // a *= 2;  // operator*=(Scalar) not implemented
    // a = 2*a; // operator=(...) not implemented/forwarded
    a *= a;     // ok
    a *= 2*a;   // ok
    a *= 0*a+2; // ok

    // a.derived() = 2*a; // derived() is not public
    static_cast<Derived&>(a) = a*2; // ok
}

Upvotes: 3

Related Questions