Importing the C API module of numpy requires more code than just including a header. xtensor-python simplifies a lot this import, however some actions are still required in the user code.

When writing an extension module that is self-contained in a single file, its author should pay attention to the following points:

• FORCE_IMPORT_ARRAY must be defined before including any header of xtensor-python.
• xt::import_numpy() must be called in the function initializing the module.

Thus the basic skeleton of the module looks like:

#define FORCE_IMPORT_ARRAY
#include "xtensor-python/pyarray.hpp"

PYBIND11_MODULE(plugin_name, m)
{
    xt::import_numpy();
    //...
}

If the extension module contains many source files that include xtensor-python header files, the previous points are still required. However, the symbol FORCE_IMPORT_ARRAY must be defined only once. The simplest is to define it int the file that contains the initializing code of the module, you can then directly include xtensor-python headers in other files. Let's illustrate this with an extension modules containing the following files:

• main.cpp: initializing code of the module
• image.hpp: declaration of the image class embedding an xt::pyarray object
• image.cpp: implementation of the image class

The basic skeleton of the module looks like:

// image.hpp
// Do NOT define FORCE_IMPORT_ARRAY here
#include "xtensor-python/pyarray.hpp"

class image
{
// ....
private:
    xt::pyarray<double> m_data;
};

// image.cpp
// Do NOT define FORCE_IMPORT_ARRAY here
#include "image.hpp"
// definition of the image class

// main.cpp
// FORCE_IMPORT_ARRAY must be define ONCE, BEFORE including
// any header from xtensor-python (even indirectly)
#define FORCE_IMPORT_ARRAY
#include "image.hpp"
PYBIND11_MODULE(plugin_name, m)
{
    xt::import_numpy();
    //...
}

Including an header of xtensor-python actually defines PY_ARRAY_UNIQUE_SYMBOL to xtensor_python_ARRAY_API. This might be problematic if you import another library that defines its own PY_ARRAY_UNIQUE_SYMBOL, or if you define yours. If so, you can override the behavior of xtensor-python by explicitly defining PY_ARRAY_UNIQUE_SYMBOL prior to including any stenxor-python header:

// in every source file
#define PY_ARRAY_UNIQUE_SYMBOL my_uniqe_array_api
#include "xtensor-python/pyarray.hpp"