I am in the process of wrapping a C++ library using Boost.Python. Some of the functions therein return Eigen::MatrixXd
objects (a dynamically sized double-precision matrix class). On the Python side I just need to access the dimensions of the matrices, this is easy, and retrieve some matrix elements with Eigen's overloaded operator()()
method. Unfortunately there are 4 such overloaded methods, one has to pick the right one manually, i.e. giving Boost.Python a function pointer typedef with the correct signature, something along the lines
namespace bpy = boost::python;
bpy::class_<Eigen::MatrixXd>("MatrixXd",
"Variable-size double-precision matrix class",
bpy::init<const Eigen::MatrixXd&>()
)
.def("__call__", static_cast<parop_signature>(&Eigen::MatrixXd::operator()))
// ...
;
The problem is that I cannot figure out what the correct signature of the function is. "Operationally" it should take two integer indices and return a double value. However,
typedef const double& (Eigen::MatrixXd::*parop_signature)(int, int) const;
results in the following compilation error (Mac OS X, clang++ in C++11 mode, Boost.Python V1.61):
address of overloaded function 'operator()' cannot be static_cast to type
'const double &(MatrixXd::*)(int, int) const'
...static_cast<const double& (Eigen::MatrixXd::*)(int, int) const>(&Eigen::MatrixXd::operator())
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/local/eigen/current/Eigen/src/Core/DenseCoeffsBase.h:111:41: note:
candidate function
EIGEN_STRONG_INLINE CoeffReturnType operator()(Index row, Index col) const
^
/usr/local/eigen/current/Eigen/src/Core/DenseCoeffsBase.h:171:5: note:
candidate function
operator()(Index index) const
^
/usr/local/eigen/current/Eigen/src/Core/DenseCoeffsBase.h:334:5: note:
candidate function
operator()(Index row, Index col)
^
/usr/local/eigen/current/Eigen/src/Core/DenseCoeffsBase.h:392:5: note:
candidate function
operator()(Index index)
Fair enough, you'd say: but I cannot figure out how I can tell Boost.Python that CoeffReturnType
is actually a double
here (or maybe a const double&
, who knows?), and that IndexType
will resolve to a plain int
at the end of the day. I have tried all sorts of combinations of typedef
-s with or without const
qualifiers.
Even tried to declare a C++11 -style function pointer like
auto eigen_indexfn = std::mem_fn<double(int,int)>(&Eigen::MatrixXd::operator());
, no success, I get
candidate template ignored: couldn't infer template argument '_Tp'
mem_fn(_Rp _Tp::* __pm)
^
Is there someone who has gone through this already and can provide me with the correct signature of what for all intents and purposes should be as simple as "double Eigen::MatrixXd::operator(int, int)
" ? Any hints would be much appreciated.
It seems the error origins from the fact that Eigen::Index
is not int
but defaults to ptrdiff_t
. Just because int
can implicitly be cast to Eigen::Index
does not mean that you can cast a function pointer which requires Eigen::Index
to a function pointer which requires int
. If that was possible, you would end up passing integers of the wrong size over the stack.
Addendum: If you really prefer int
over ptrdiff_t
, you can define EIGEN_DEFAULT_DENSE_INDEX_TYPE
to int
before including Eigen, as documented here, be aware that this will break ABI compatibility.