Using pybind11

The package pybind11 is provides an elegant way to wrap C++ code for Python, including automatic conversions for numpy arrays and the C++ Eigen linear algebra library. Used with the cppimport package, this provides a very nice work flow for integrating C++ and Python:

  • Edit C++ code
  • Run Python code
pip install pybind11
pip install cppimport

Clone the Eigen library - no installation is required as Eigen is a header only library.

hg clone https://bitbucket.org/eigen/eigen/

A first example of using pybind11

Create a new subdirectory - e.g. example1 and create the following 5 files in it:

  • funcs.hpp
  • funcs.cpp
  • wrap.cpp
  • setup.py
  • test_funcs.py

First write the C++ header and implementation files

In funcs.hpp

int add(int i, int j);

In funcs.cpp

int add(int i, int j) {
    return i + j;
};

Next write the C++ wrapper code using pybind11 in wrap.cpp. The arguments "i"_a=1, "j"_a=2 in the exported function definition tells pybind11 to generate variables named i with default value 1 and j with default value 2 for the add function.

#include <pybind11/pybind11.h>
#include "funcs.hpp"

namespace py = pybind11;

using namespace pybind11::literals;

PYBIND11_PLUGIN(wrap) {
    py::module m("wrap", "pybind11 example plugin");
    m.def("add", &add, "A function which adds two numbers",
          "i"_a=1, "j"_a=2);
    return m.ptr();
}

Finally, write the setup.py file to compile the extension module. This is mostly boilerplate.

import os, sys

from distutils.core import setup, Extension
from distutils import sysconfig

cpp_args = ['-std=c++11', '-stdlib=libc++', '-mmacosx-version-min=10.7']

ext_modules = [
    Extension(
    'wrap',
        ['funcs.cpp', 'wrap.cpp'],
        include_dirs=['pybind11/include'],
    language='c++',
    extra_compile_args = cpp_args,
    ),
]

setup(
    name='wrap',
    version='0.0.1',
    author='Cliburn Chan',
    author_email='cliburn.chan@duke.edu',
    description='Example',
    ext_modules=ext_modules,
)

Now build the extension module in the subdirectory with these files

python setup.py build_ext -i

And if you are successful, you should now see a new funcs.so extension module. We can write a test_funcs.py file test the extension module:

import wrap

def test_add():
    assert(wrap.add(3, 4) == 7)

if __name__ == '__main__':
    test_add()

And finally, running the test should not generate any error messages:

$ python test_funcs.py

Using cppimport

In the development stage, it can be distracting to have to repeatedly rebuild the extension module by running

python setup.py clean
python setup.py build_ext -i

every single time you modify the C++ code. The cppimport package does this for you.

Create a new sub-directory exaample2 and copy the files func.hpp, funcs.cpp and wrap.cpp from example1 over. For the previous example, we just need to add some annotation (between <% and %> delimiters) to the top of the wrap.cpp file

<%
cfg['compiler_args'] = ['-std=c++11', '-stdlib=libc++', '-mmacosx-version-min=10.7']
cfg['sources'] = ['funcs.cpp']
setup_pybind11(cfg)
%>

#include "funcs.hpp"
#include <pybind11/pybind11.h>

namespace py = pybind11;

PYBIND11_PLUGIN(wrap) {
    py::module m("wrap", "pybind11 example plugin");
    m.def("add", &add, "A function which adds two numbers");
    return m.ptr();
}

and use cpppimport in the test_funcs.py file.

import cppimport
funcs = cppimport.imp("wrap")

def test_add():
    assert(funcs.add(3, 4) == 7)

if __name__ == '__main__':
    test_add()

You can now run

$ python test_funcs.py

without any need to manually build the extension module. Any updates will be detected by cppimport and it will automatically trigger a re-build.

Vectorizing functions for use with numpy arrays

Example showing how to vectorize a square function. Note that from here on, we don’t bother to use separate header and implementation files for these code snippets, and just write them together with the wrapping code in a code.cpp file. This means that with cppimport, there are only two files that we actually code for, a C++ code.cpp file and a python test file.

<%
cfg['compiler_args'] = ['-std=c++11', '-stdlib=libc++', '-mmacosx-version-min=10.7']
setup_pybind11(cfg)
%>

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

namespace py = pybind11;

double square(double x) {
    return x * x;
}

PYBIND11_PLUGIN(code) {
    pybind11::module m("code", "auto-compiled c++ extension");
    m.def("square", py::vectorize(square));
    return m.ptr();
}

and the test file

import cppimport
import numpy as np

code = cppimport.imp("code")

if __name__ == '__main__':
    xs = np.random.rand(10)
    print(xs)
    print(code.square(xs))

    ys = range(10)
    print(code.square(ys))

Using numpy arrays as function arguments and return values

Example showing how to pass numpy arrays in and out of functions. These numpy array arguments can either be generic py:array or typed py:array_t<double>. The properties of the numpy array can be obtained by calling its request method. This returns a struct of the following form:

struct buffer_info {
    void *ptr;
    size_t itemsize;
    std::string format;
    int ndim;
    std::vector<size_t> shape;
    std::vector<size_t> strides;
};

Here is C++ code for two functions - the function twice shows how to change a passed in numpy array in-place using pointers; the function sum shows how to sum the elements of a numpy array. By taking advantage of the information in buffer_info, the code will work for arbitrary n-d arrays.

<%
cfg['compiler_args'] = ['-std=c++11', '-stdlib=libc++', '-mmacosx-version-min=10.7']
setup_pybind11(cfg)
%>

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

namespace py = pybind11;

// Passing in an array of doubles
void twice(py::array_t<double> xs) {
    py::buffer_info info = xs.request();
    auto ptr = static_cast<double *>(info.ptr);

    int n = 1;
    for (auto r: info.shape) {
      n *= r;
    }

    for (int i = 0; i <n; i++) {
        *ptr++ *= 2;
    }
}

// Passing in a generic array
double sum(py::array xs) {
    py::buffer_info info = xs.request();
    auto ptr = static_cast<double *>(info.ptr);

    int n = 1;
    for (auto r: info.shape) {
      n *= r;
    }

    double s = 0.0;
    for (int i = 0; i <n; i++) {
        s += *ptr++;
    }

    return s;
}

PYBIND11_PLUGIN(code) {
    pybind11::module m("code", "auto-compiled c++ extension");
    m.def("sum", &sum);
    m.def("twice", &twice);
    return m.ptr();
}

and the test code

import cppimport
import numpy as np

code = cppimport.imp("code")

if __name__ == '__main__':
    xs = np.arange(12).reshape(3,4).astype('float')
    print(xs)
    print("np :", xs.sum())
    print("cpp:", code.sum(xs))

    print()
    code.twice(xs)
    print(xs)

More on working with numpy arrays

This example shows how to use array access for numpy arrays within the C++ function. It is taken from the pybind11 documentation, but fixes a small bug in the official version. As noted in the documentation, the function would be more easily coded using py::vectorize.

<%
cfg['compiler_args'] = ['-std=c++11', '-stdlib=libc++', '-mmacosx-version-min=10.7']
setup_pybind11(cfg)
%>

#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) {
    auto 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.shape[0] != buf2.shape[0])
        throw std::runtime_error("Input shapes must match");

    auto result = py::array(py::buffer_info(
        nullptr,            /* Pointer to data (nullptr -> ask NumPy to allocate!) */
        sizeof(double),     /* Size of one item */
        py::format_descriptor<double>::value, /* Buffer format */
        buf1.ndim,          /* How many dimensions? */
        { buf1.shape[0] },  /* Number of elements for each dimension */
        { sizeof(double) }  /* Strides for each dimension */
    ));

    auto buf3 = result.request();

    double *ptr1 = (double *) buf1.ptr,
           *ptr2 = (double *) buf2.ptr,
           *ptr3 = (double *) buf3.ptr;

    for (size_t idx = 0; idx < buf1.shape[0]; idx++)
        ptr3[idx] = ptr1[idx] + ptr2[idx];

    return result;
}

PYBIND11_PLUGIN(code) {
    py::module m("code");
    m.def("add_arrays", &add_arrays, "Add two NumPy arrays");
    return m.ptr();
}

with test code

import cppimport
import numpy as np

code = cppimport.imp("code")

if __name__ == '__main__':
    xs = np.arange(12)
    print(xs)

    print(code.add_arrays(xs, xs))

Using the C++ eigen library to calculate matrix inverse and determinant

Example showing how Eigen vectors and matrices can be passed in and out of C++ functions. Note that Eigen arrays are automatically converted to/from numpy arrays simply by including the pybind/eigen.h header. Because of this, it is probably simplest in most cases to work with Eigen vectors and matrices rather than py::buffer or py::array where py::vectorize is insufficient.

<%
cfg['compiler_args'] = ['-std=c++11', '-stdlib=libc++', '-mmacosx-version-min=10.7']
cfg['include_dirs'] = ['/Users/cliburn/hg/eigen']
setup_pybind11(cfg)
%>

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

#include <Eigen/LU>

namespace py = pybind11;

// convenient matrix indexing comes for free
double get(Eigen::MatrixXd xs, int i, int j) {
    return xs(i, j);
}

// takes numpy array as input and returns double
double det(Eigen::MatrixXd xs) {
    return xs.determinant();
}

// takes numpy array as input and returns another numpy array
Eigen::MatrixXd inv(Eigen::MatrixXd xs) {
    return xs.inverse();
}

PYBIND11_PLUGIN(code) {
    pybind11::module m("code", "auto-compiled c++ extension");
    m.def("inv", &inv);
    m.def("det", &det);
    return m.ptr();
}

and test code

import cppimport
import numpy as np

code = cppimport.imp("code")

if __name__ == '__main__':
    A = np.array([[1,2,1],
                  [2,1,0],
                  [-1,1,2]])

    print(A)
    print(code.det(A))
    print(code.inv(A))

Using pybind11 with openmp

Here is a standard example of using OpenMP to integrate the value of \(\pi\) written using pybind11.

<%
cfg['compiler_args'] = ['-std=c++11', '-fopenmp']
cfg['linker_args'] = ['-fopenmp']
setup_pybind11(cfg)
%>

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

namespace py = pybind11;

double calc_pi(int n) {
  /* Acquire GIL before calling Python code */
  py::gil_scoped_acquire acquire;

  int i;
  double step = 1.0/n;
  double s = 0;

  #pragma omp parallel
  {
    double x;
    #pragma omp for reduction(+:s)
    for (i=0; i<n; i++) {
      x = (i+0.5) * step;
      s += 4.0/(1 + x*x);
    }
  }
  return step * s;
};

PYBIND11_PLUGIN(code) {
  pybind11::module m("code", "auto-compiled c++ extension");
  m.def("calc_pi", [](int n) {
      /* Release GIL before calling into C++ code */
      py::gil_scoped_release release;
      return calc_pi(n);
    });

  return m.ptr();
}

And here is the test code.

import cppimport
import numpy as np

code = cppimport.imp("code")

if __name__ == '__main__':
    n = int(1e9)

    print(code.calc_pi(n))

In [ ]: