Creating a Python module from C++ functions using pybind11

Setting up the build environment

Required dependencies for GNU/Linux

Install required packages

$ python -m pip install pybind11
$ python -m pip install numpy

Required dependencies for Windows

Install MSYS2

Install required packages using pacman from within MSYS2 command window

$ pacman -S mingw-w64-x86_64-pybind11
$ pacman -S mingw-w64-x86_64-python
$ pacman -S mingw-w64-x86_64-gcc
$ pacman -S mingw-w64-x86_64-python-numpy

Creating the pybind11 bindings source file

Create the file my_module.cpp which is mostly copied from the pybind11 documentation but with a few extras added.

#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>

namespace py = pybind11;

py::array_t<double> add_arrays(py::array_t<double> input1, py::array_t<double> input2) {
    py::buffer_info buf1 = input1.request(), buf2 = input2.request();

    if (buf1.ndim != 1 || buf2.ndim != 1)
        throw std::runtime_error("Number of dimensions must be one");

    if (buf1.size != buf2.size)
        throw std::runtime_error("Input shapes must match");

    /* No pointer is passed, so NumPy will allocate the buffer */
    auto result = py::array_t<double>(buf1.size);

    py::buffer_info buf3 = result.request();

    double *ptr1 = static_cast<double *>(buf1.ptr);
    double *ptr2 = static_cast<double *>(buf2.ptr);
    double *ptr3 = static_cast<double *>(buf3.ptr);

    size_t len1 = static_cast<size_t>(buf1.shape[0]);

    for (size_t idx = 0; idx < len1; idx++)
        ptr3[idx] = ptr1[idx] + ptr2[idx];

    return result;

float add(float a, float b)
    return a+b;

py::array_t<double> half_length(py::array_t<double> input)
    py::buffer_info buf = input.request();
    auto result = py::array_t<double>(buf.size);
    py::buffer_info ret = result.request();
    double *ptr1 = static_cast<double *>(buf.ptr);
    double *ptr2 = static_cast<double *>(ret.ptr);

    size_t len = static_cast<size_t>(buf.shape[0]);

    for (size_t i = 0; i < len; i++)
        ptr2[i] = ptr1[i];
        ptr1[i] = 0.0;


    return result; 

PYBIND11_MODULE(my_module, m)
    m.def("add", &add, "Add two numbers");
    m.def("add_arrays", &add_arrays, "Add two NumPy arrays");
    m.def("half_length", &half_length, "Return first half of array and set input array to zeros");


To build the python extension module

$ c++ -O3 -Wall -shared -std=c++11 -fPIC `python3 -m pybind11 --includes` my_module.cpp -o my_module`python3-config --extension-suffix` `python3-config --libs`


#!/usr/bin/env python

import my_module as mm
import numpy as np


# The input array will not be zeroed as it is first converted to double.

# The input array type is double and a conversion is not performed. The C++ function will zero the input array.