Python + cython: faster than C?

[Note added on January 15, 2013: I am amazed that this post, clearly written tongue-in-cheek 3 years ago, can still spur people into doing their own test, and feeling the urge to comment.]

Now that I have your attention... ;-)

I've been playing around with cython and will write more about it soon.  But I could not resist doing a quick test when I read this post, comparing C and Java on some toy micro benchmark.

First, the result:

andre$ gcc -o fib -O3 fib.c
andre$ time ./fib
433494437

real    0m3.765s
user    0m3.463s
sys     0m0.028s

andre$ time python fib_test.py
433494437

real    0m2.953s
user    0m2.452s
sys     0m0.380s

Yes, Python+cython is faster than C!  :-)

Now, the code.  First, the C program taken straight from this post, except for a different, more meaningful value for "num" ;-)

#include 

double fib(int num)
{
   if (num <= 1)
       return 1;
   else
       return fib(num-1) + fib(num-2);
 }
 int main(void)
{
     printf("%.0f\n", fib(42));
    return 0;
 }

Next, the cython code ...

# fib.pyx

cdef inline int fib(int num):
    if (num <= 1):
        return 1
    else:
        return fib(num-1) + fib(num-2)

def fib_import (int num):
    return fib(num)

... and the Python file that calls it

# fib_test.py

import pyximport
pyximport.install()

from fib import fib_import

print fib_import(42)

1. I know, I declared fib() to be of type "double" in the C code (like what was done in the original post) and "int" in the cython code; however, if I declare it to be of type "int" in the C code, I get -0 as the answer instead of 433494437.  I could declare it to be of type "unsigned int" ... but even when I did that, the Python code was marginally faster.
2. If I declared fib to be of type "double" in the cython code, it was slightly slower than the corresponding C code.  However, what Python user would ever think of an integer variable as a floating point number! ;-)
   
3. What is most impressive is that the cython code is NOT pre-compiled, unlike the C code.  However, to be fair ... I did run it a few times and took the best result, and it was always slower the first time.
   
4. Yes, it is silly to compare such micro-benchmarks ... but it can be fun, no? ;-)

More seriously: using cython is a relatively painless way to significantly increase the speed of Python applications ... without having to learn a totally different language.