For performance reasons, it can be useful to write functions in C/C++ which can then be called within Python. This will be an introductory post, in where we will call a simple C++ function (with a dependency) within Python using SWIG.
First we need to install SWIG:
sudo apt install swig
We’re going to use GNU MP in order to have arbitrary precision arithmetic for our factorial function.
sudo apt install libgmp-dev
Source Setup
Normally people use headers for larger C++ programs, though we’re going to create one just so we can see how to include it later in SWIG. Let’s called this file factorial.hpp
#ifndef FACTORIAL_H
#define FACTORIAL_H
std::string fact(unsigned int n);
#endif
In order to get it the large number from C++ to Python. We are going to use std::string
as the return of our fact
function.
Here is the source factorial.cpp
#include <gmpxx.h>
#include "factorial.hpp"
std::string fact(unsigned int n) {
if (n == 0) {
n = 1;
}
mpz_class result(n);
while (n > 1) {
n--;
result *= n;
}
return result.get_str(10); // Base 10
}
Now that we have our C++ code, we need to create a swig template file called factorial.i
%module factorial
%{
#include "factorial.hpp"
%}
%include <std_string.i>
%include "factorial.hpp"
Since we’re returning a std::string
we need to tell SWIG what that is. We do this through the <std_string.i>
include.
We can now ask SWIG to write the C++ code that will interface with Python. This will create the files factorial_wrap.cxx
and factorial.py
.
swig -c++ -python factorial.i
Compilation and Linkage
Let’s compile our C++ code.
g++ -O2 -fPIC -c factorial.cpp
Flag | Description |
---|---|
-O2 | Perform nearly all supported optimizations that don’t involve a space-speed tradeoff. |
-fPIC | Create Position-Independent Code |
-c | Don’t link at this time |
To compile factorial_wrap.cxx
we need to include the directory where Python.h
lives. You can find this by issuing the command locate Python.h
. Below is where it is located on my system.
g++ -O2 -fPIC -c factorial_wrap.cxx -I/home/user/.pyenv/versions/3.8.2/include/python3.8/
Finally let’s create the needed shared object file by linking factorial.o
, factorial_wrap.o
, and the GNU MP libraries.
g++ -O2 -fPIC -shared factorial.o factorial_wrap.o -lgmpxx -lgmp -o _factorial.so
It is important that our final output is called _
+ module_name.so
We should at this time be able to open up python
and import our function.
import factorial
factorial.fact(5)
If you run into any errors, the SWIG Documentation is quite helpful.
In order to not have to type out the compiling and linking commands every time, here is a Makefile
CC=g++
CFLAGS=-O2 -fPIC -Wall
PYTHON_PATH=/home/user/.pyenv/versions/3.8.2/include/python3.8/
all: _factorial.so
_factorial.so: factorial.o factorial_wrap.o
$(CC) $(CFLAGS) -shared factorial.o factorial_wrap.o -lgmpxx -lgmp -o _factorial.so
factorial_wrap.o: factorial_wrap.cxx
$(CC) $(CFLAGS) -c factorial_wrap.cxx -I$(PYTHON_PATH)
factorial.o: factorial.cpp
$(CC) $(CFLAGS) -c factorial.cpp
factorial_wrap.cxx: factorial.i
swig -c++ -python factorial.i
clean:
rm *.o *.so factorial_wrap.cxx factorial.py
Then you can call make clean
to clean up everything and make
to run all the individual compilation steps we did before.
Top comments (0)