SuReLI / tuto_pybind

Make your code fast again

Make your code fast again

👷 Repo in constuction 👷

Repository Purpose

How to integrate C++ code into your Python code in order to make faster programs.

Tools used: Pybind11,Cmake

For more details: Pybind.

For more details Cmake.

Minimum Requierements (For Linux)

If GCC is not installed :

sudo apt install build-essential

If Cmake is not installed :

sudo apt-get -y install cmake

If the repo is not cloned :

git clone https://github.com/SuReLI/tuto_pybind.git

If the module pybind11 and nanobind is not installed :

python3 -m venv venv
source venv/bin/activate
pip install pybind11 nanobind

Pybind11: Function Harmonic sum

Our toy test is the harmonic sum.

$$H_m = \sum_{k=1}^{m} \frac{1}{k}$$

Methodology

A program coded in C++ and a program coded in python compute the harmonic sum. In our implementation, $m=10^n$ where n equals to 9 is tested.

Test 1

cd pybind_examples/harmonic_sum
mkdir build
cd build
cmake ..
make
cd ..
python3 harmonic_sum.py

Results :
Python
Time Process 22.109s
Harmonic sum 22.300481502349225 for n equals to 9

Python powered by C++
Time Process 7.352s
Harmonic sum 22.300481502349225 for n equals to 9

Explanations

It is necessary to create a file named 'build'. Initially, the command 'cmake ..' refers to the execution of the CMakeLists.txt. CMakeLists.txt is a file containing all the code needed for CMake to build, generate, and orchestrate the project. Finally, the command line 'make' executes the Makefile outputted by CMake. A .so file is created, and now it is possible to import your C++ function into Python. harmonic_sum.py compares the Python code to the bound Python code.

Pybind11: OOP Matrix Multiplication,Trace

Our toy test is to compute the trace of a dot product between two matrices.

Methodology

The two matrixes equals to $10*I_{850,850}$. A program is coded in C++ and a program is coded in Python. We bind the C++ object with Python and compare it to the same object coded in Python.

Test 2

cd pybind_examples/mat_mul_trace
mkdir build
cd build
cmake ..
make
cd ..
python3 class_matrix.py

Python powered by C++ Time Process 0.416s
Multiplication, trace result : 72250000.0 for l,m,n : 850,850,850

Python powered by Numpy
Time Process 0.253s
Multiplication, trace result : 72250000 for l,m,n : 850,850,850

Python
Time Process 52.638s
Multiplication, trace result : 72250000 for l,m,n : 850,850,850

Nanobind

You can test the bind between C++ and Python thanks to Nanobind however i find than it is slower compared to Pybind11.

Update

I added #pragma omp parallel for to my code and my C++ binding was quicker than Numpy.

Python powered by C++ Time Process 0.098s Multiplication, trace result : 72250000.0 for l,m,n : 850,850,850

Python powered by Numpy Time Process 0.271s Multiplication, trace result : 72250000 for l,m,n : 850,850,850

Python Time Process 50.278s Multiplication, trace result : 72250000 for l,m,n : 850,850,850

Python powered by C++ Time Process 15.16s Multiplication, trace result : 1600000000.0 for l,m,n : 4000,4000,4000

Python powered by Numpy Time Process 75.008s Multiplication, trace result : 1600000000 for l,m,n : 4000,4000,4000