suggestions about efficiently binding typedef function taking double * to take np.array

[mike-gimelfarb]

The problem is fairly simple: I have a typedef function declared as follows

typedef std::function<double(const double*)> multivariate;

which is used as a parameter in other functions. I would like this function, from the python side, to take np.array as input and return float. Currently, the only way I know how is to create a wrapper that will work from python, e.g.

typedef std::function<double(const pybind11::array_t<double>&)> multivariate_wrapper;

and then convert this to a multivariate object in the binding code, e.g.

multivariate fc = [&](const double *x) -> double { const auto &arr = py::array(n, x); return f(arr); };

Here, n is taken as an additional argument provided by the python call, and f is an instance of multivariate_wrapper.

This compiles fine, but the python code is much slower (at least 5x) than the corresponding c++ code. I am wondering if there is a better way to achieve this within the pybind11 framework without having to create multivariate explicitly and capture? Would replacing std::function with a struct or similar (maybe with a call function) and pass pointer be a better alternative? Thanks, if you have any suggestions.