[Tutor] Cython vs Python-C API

[Steven D'Aprano]

Jaidev Deshpande wrote:

> 1. I can either compile the C code into a module using the Python-C/C++
> API, through which I can simple 'import' the required function.
> 2. I can use the Python code and extend it using Cython.
> 
> Which will give me a better performance?

The only way to be sure is to do both, measure the speed of each, and 
compare. But in *principle* hand-written C may be faster, but in 
*practice* Cython is likely to be as fast and possibly faster. It's 
certainly easier: you can afford to try Cython and only bother with the 
pure C code if you actually need it.

A more interesting question is "which will give you sufficiently fast 
performance for the least development time?"

Some further options:

3. Just write it in pure Python, and don't worry about speeding it up 
unless you actually need to. Unless you have actually *measured* the 
time taken, how do you know which (if any) bits are too slow? Perhaps it 
is already fast enough, and making it faster just throws away developer 
time for no benefit.

If the cubic spline interpolation is too slow, you still have additional 
options:

4. If you only need to support Python 2 on 32-bit platforms, use the 
Psyco JIT compiler to speed it up. This only works on the standard 
CPython version (not to be confused with Cython), not IronPython or 
Jython compilers.

5. Use PyPy, which is an optimizing compiler for Python. This is likely 
to be ten times as fast as CPython. PyPy is fast enough to do some video 
processing tasks in real-time, so it likely will be fast enough for 
cubic splines.

http://morepypy.blogspot.com/2011/07/realtime-image-processing-in-python.html

6. Extend the Python code with Cython, and see if that is sufficiently 
fast enough for your needs.

7. And finally, if the Cython code is too slow, re-write in C by hand.



-- 
Steven