This commit is contained in:
Brandon Rozek 2020-10-28 00:32:15 -04:00
parent 0e645935c5
commit 65c04e204f

138
content/blog/pythonswig.md Normal file
View file

@ -0,0 +1,138 @@
---
title: "C++ within Python with SWIG"
date: 2020-10-27T23:49:54-04:00
draft: false
tags: []
---
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](http://swig.org/).
First we need to install SWIG:
```bash
sudo apt install swig
```
We're going to use [GNU MP](https://gmplib.org/) in order to have arbitrary precision arithmetic for our factorial function.
```bash
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`
```c++
#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`
```c++
#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`.
```bash
swig -c++ -python factorial.i
```
## Compilation and Linkage
Let's compile our C++ code.
```bash
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.
```bash
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.
```bash
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.
```python
import factorial
factorial.fact(5)
```
If you run into any errors, the [SWIG Documentation](http://www.swig.org/Doc3.0/Python.html#Python_nn3) is quite helpful.
In order to not have to type out the compiling and linking commands every time, here is a Makefile
```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.