Ctrl+K

Adding Cython to SciPy#

As written on the Cython website:

Cython is an optimising static compiler for both the Python programming language and the extended Cython programming language (based on Pyrex). It makes writing C extensions for Python as easy as Python itself.

If your code currently performs a lot of loops in Python, it might benefit from compilation with Cython. This document is intended to be a very brief introduction: just enough to see how to use Cython with SciPy. Once you have your code compiling, you can learn more about how to optimize it by reviewing the Cython documentation.

There are only two things you need to do in order for SciPy compile your code with Cython:

  1. Include your code in a file with a .pyx extension rather than a .py extension. All files with a .pyx extension are automatically converted by Cython to .c or .cpp files when SciPy is built.

  2. Add the new .pyx file to the meson.build build configuration of the subpackage in which your code lives. Typically, there are already other .pyx patterns present (if not, look in another submodule) so there’s an example to follow for what exact content to add to meson.build.

Example#

scipy.optimize._linprog_rs.py contains the implementation of the revised simplex method for scipy.optimize.linprog. The revised simplex method performs many elementary row operations on matrices, and so it was a natural candidate to be Cythonized.

Note that scipy/optimize/_linprog_rs.py imports the BGLU and LU classes from ._bglu_dense exactly as if they were regular Python classes. But they’re not. BGLU and LU are Cython classes defined in /scipy/optimize/_bglu_dense.pyx. There is nothing about the way they are imported or used that suggests that they are written in Cython; the only way so far that we can tell they are Cython classes is that they are defined in a file with a .pyx extension.

Even in /scipy/optimize/_bglu_dense.pyx, most of the code resembles Python. The most notable differences are the presence of cimport, cdef, and Cython decorators. None of these are strictly necessary. Without them, the pure Python code can still be compiled by Cython. The Cython language extensions are *just* tweaks to improve performance. This .pyx file is automatically converted to a .c file by Cython when SciPy is built.

The only thing left is to add the build configuration, which will look something like:

_bglu_dense_c = opt_gen.process('_bglu_dense.pyx')

py3.extension_module('_bglu_dense',
  _bglu_dense_c,
  c_args: cython_c_args,
  dependencies: np_dep,
  link_args: version_link_args,
  install: true,
  subdir: 'scipy/optimize'
)

When SciPy is built, _bglu_dense.pyx will be transpiled by cython to C code, and then that generated C file is treated by Meson like any other C code in SciPy - producing an extension modules that we will be able to import and use the LU and BGLU classes from.

Exercise#

See a video run-through of this exercise: Cythonizing SciPy Code

  1. Update Cython and create a new branch (e.g., git checkout -b cython_test) in which to make some experimental changes to SciPy

  2. Add some simple Python code in a .py file in the /scipy/optimize directory, say /scipy/optimize/mypython.py. For example:

    def myfun():
    i = 1
    while i < 10000000:
        i += 1
    return i

  3. Let’s see how long this pure-Python loop takes so we can compare the performance of Cython. For example, in an IPython console in Spyder:

    from scipy.optimize.mypython import myfun
%timeit myfun()

    I get something like:

    715 ms ± 10.7 ms per loop

  4. Save your .py file to a .pyx file, e.g. mycython.pyx.

  5. Add the .pyx to scipy/optimize/meson.build, in the way described in the previous section.

  6. Rebuild SciPy. Note that an extension module (a .so or .pyd file) has been added to the build/scipy/optimize/ directory.

  7. Time it, e.g. by dropping into IPython with python dev.py ipython and then:

    from scipy.optimize.mycython import myfun
%timeit myfun()

    I get something like:

    359 ms ± 6.98 ms per loop

    Cython sped up the pure Python code by a factor of ~2.

  8. That’s not much of an improvement in the scheme of things. To see why, it helps to have Cython create an “annotated” version of our code to show bottlenecks. In a terminal window, call Cython on your .pyx file with the -a flag:

    cython -a scipy/optimize/mycython.pyx

    Note that this creates a new .html file in the /scipy/optimize directory. Open the .html file in any browser.

  9. The yellow-highlighted lines in the file indicate potential interaction between the compiled code and Python, which slows things down considerably. The intensity of the highlighting indicates the estimated severity of the interaction. In this case, much of the interaction can be avoided if we define the variable i as an integer so that Cython doesn’t have to consider the possibility of it being a general Python object:

    def myfun():
    cdef int i = 1  # our first line of Cython code
    while i < 10000000:
        i += 1
    return i

    Recreating the annotated .html file shows that most of the Python interaction has disappeared.

  10. Rebuild SciPy, open an fresh IPython console, and %timeit:

from scipy.optimize.mycython import myfun
%timeit myfun()

I get something like: 68.6 ns ± 1.95 ns per loop. The Cython code ran about 10 million times faster than the original Python code.

In this case, the compiler probably optimized away the loop, simply returning the final result. This sort of speedup is not typical for real code, but this exercise certainly illustrates the power of Cython when the alternative is many low-level operations in Python.

On this page