Cython

Cython

Tutorial: cython

Hello world

First, write the code in the file helloworld.pyx

print("Hello World")

Then, write a file setup.py with codes below:

from setuptools import setup
from Cython.Build import cythonize

setup(
    ext_modules = cythonize("helloworld.pyx")
)

Now, let’s setup the Cython in the terminal/cmd

python setup.py build_ext --inplace

Compiling helloworld.pyx because it changed.
[1/1] Cythonizing helloworld.pyx
/home/ken/.local/lib/python3.8/site-packages/Cython/Compiler/Main.py:369: FutureWarning: Cython directive 'language_level' not set, using 2 for now (Py2). This will change in a later release! File: /home/ken/test/helloworld.pyx
  tree = Parsing.p_module(s, pxd, full_module_name)
running build_ext
building 'helloworld' extension
gcc -pthread -B /mnt/8A26661926660713/Conda/miniconda/compiler_compat -Wno-unused-result -Wsign-compare -DNDEBUG -fwrapv -O2 -Wall -fPIC -O2 -isystem /mnt/8A26661926660713/Conda/miniconda/include -fPIC -O2 -isystem /mnt/8A26661926660713/Conda/miniconda/include -fPIC -I/mnt/8A26661926660713/Conda/miniconda/include/python3.8 -c helloworld.c -o build/temp.linux-x86_64-cpython-38/helloworld.o
gcc -pthread -B /mnt/8A26661926660713/Conda/miniconda/compiler_compat -shared -Wl,--allow-shlib-undefined -Wl,-rpath,/mnt/8A26661926660713/Conda/miniconda/lib -Wl,-rpath-link,/mnt/8A26661926660713/Conda/miniconda/lib -L/mnt/8A26661926660713/Conda/miniconda/lib -Wl,--allow-shlib-undefined -Wl,-rpath,/mnt/8A26661926660713/Conda/miniconda/lib -Wl,-rpath-link,/mnt/8A26661926660713/Conda/miniconda/lib -L/mnt/8A26661926660713/Conda/miniconda/lib build/temp.linux-x86_64-cpython-38/helloworld.o -o build/lib.linux-x86_64-cpython-38/helloworld.cpython-38-x86_64-linux-gnu.so
copying build/lib.linux-x86_64-cpython-38/helloworld.cpython-38-x86_64-linux-gnu.so -> 

Finally, import it in any python interpreter:

import helloworld

Hello World

Speed up the for loop

First, let’s create a file named example.pyx with the following Cython code:

# example.pyx
def sum_elements(double[:] array):
    cdef int i
    cdef double result = 0

    for i in range(array.shape[0]):
        result += array[i]

    return result

To compile and use the Cython code, you’ll need a setup.py file. Create a setup.py file with the following content:

# setup.py
from setuptools import setup
from Cython.Build import cythonize
import numpy as np

setup(
    ext_modules=cythonize("example.pyx"),
    include_dirs=[np.get_include()]
)

Next, compile the example.pyx file:

python setup.py build_ext --inplace

Now, you can test it in the python

import time
import numpy as np
from example import sum_elements

# define the same function in python to compare the time consuming
def sum_elements_py(array):
  result = 0
  for i in range(array.shape[0]):
    result += array[i]
  return(result)


array = np.array([1.0, 2.0, 3.0, 4.0, 5.0]*200000, dtype=np.float64)

A = time.time()
sum_elements(array)
print(time.time() - A)

A = time.time()
sum_elements_py(array)
print(time.time() - A)

from numba import njit

@njit
def sum_elements_numba(array):
    result = 0.0
    for i in range(array.shape[0]):
        result += array[i]
    return result

0.001369476318359375
0.11698126792907715

When there are 1M loops, Cython is 85 times faster than python.

array = np.array([1.0, 2.0, 3.0, 4.0, 5.0]*20000000, dtype=np.float64)

A = time.time()
sum_elements_py(array)
print(time.time() - A)


A = time.time()
sum_elements(array)
print(time.time() - A)

A = time.time()
sum_elements_numba(array)
print(time.time() - A)

13.218583345413208
0.11868596076965332
0.19266653060913086

When there are 10000M loops, Cython is 85 times faster than python.