Weave is a python module that includes a method of including C and C++ code inline with python. Has a slightly bigger overhead than the swig approach but is a lot easier to implement. And the code can be any dynamically created string. For example this block of python code calculates the fixed point multiplication of var1 and var2. The code is modified from a matlab algorithm worked on during my summer internship at Tait Electronics.

from scipy.weave import inline as inlineC

r= rounding and 1 or 0# Blatent hack to convert from pythons True/False to 1/0 for C/C++exp =

"""

signed int K,result;

signed long int temp;

temp = (long int)var1 * (long int)var2;

signed int sign = temp/abs(temp);

if(r) {

K = 1 << (numBits-1); temp = temp + K; } result = sign*(abs(temp) >> numBits);

if(result > fp_Max){

result = fp_Max;

}

if(result < result =" fp_Min;" return_val =" result;">

result = inlineC(exp,['var1','var2','numBits','fp_Min','fp_Max','r'])

It must be noted the C code is just a string to python, the string is then passed to scipy.weave.inline along with any variables required by the code. The only non standard piece of c is assigning return_val, this is set up by weave and can be assigned with any type to return into python. Also note the commented out printf statement in the C code – was very useful for debugging as it still outputs to standard out.

Weave is a quick and powerful way to include c code or optimize a bottleneck in an algorithm. The above example is probably to simple to get any speed up as its a very trivial problem and the overhead of calling C must be factored in. Weave is pretty awesome for when you have a loop in python you cannot get rid of that MUST go faster.

As with all optimization tho, ask yourself if you really need it?

Here is another example...

The original code:

from numpy import *

defwastedLoop(n):

"""wastedLoop does alot of looping adding up the numbers 1 + 2 + ... N

wastedLoop(n) loops n times"""

p = 0

for k in range(n+1):

p = p + k

print p

Then at an ipython prompt do a quick timing test:

>>> %timeit wastedLoop(1000000)

10 loops, best of 3: 464 ms per loop

Ok so now lets use numpy functions instead of looping:

def

numpySumMethod(n):

return sum(arange(n+1))

>>> %timeit numpySumMethod(1000000)So we already have achieved a 35 times faster run time than the original. And have clearer code!

100 loops, best of 3: 12.9 ms per loop

*We see an improvement here but its not amazing... Just one more example, doing the same thing but with the knowledge about the problem. Using the algorithm:*

from scipy import weave

defwastedLoop(n):"""

wastedLoop does alot of looping adding up the numbers 1 + 2 + ... N

wastedLoop(n) loops n times

"""p = int(0)

exp =

"""

int i;

for(i=0;iweave.inline(exp,[

'p','n'])

return p

%timeit wastedLoop(1000000)

10 loops, best of 3: 2.8 ms per loop

And there we have it: 1.73µs There is a lesson in that! A bit of math goes a long way! For the cases where you don't have insider knowledge however, it is clear numpy alone can suffice in most situations.def

efficientMethod(n):

"""Return sum of 1+2+3...+N"""

return n*(n+1)/2.0

>>> timeit efficientMethod(1000000)

1000000 loops, best of 3: 0.00000173 s per loop

Stay tuned for a look at python extensions with swig.

P.S: What is with this editor for blogger? It is crap!!!

## Comments

## Post a Comment