Julien Fiore wrote:
> Thanks for the psyco information, Serge.
>
> > 2) Rewrite the code to be vectorized (don't use psyco) Right now your
> > code *doesn't* get any speed benefit from numpy
>
> I do not understand this point. How to rewrite the code ?
vectorized code means one operation work on multiple items. Here is
three variants of the most easy part of your function (i'm sure this
part is not hot, it's just an example, the hottest part I'm pretty sure
is the *whole* openness loop)
setup = """
import numpy, array
R = 4
winSize= 2*R + 1
cellSizeX = 11
cellSizeY = 23
"""
code1 = """
dist=numpy.zeros((winSize,winSize),float)
for i in range(winSize):
for j in range(winSize):
dist[i][j]=numpy.sqrt((cellSizeX*(i-R))**2
+(cellSizeY*(j-R))**2)
"""
code2 = """
ind = numpy.indices((winSize,winSize))
dist = numpy.sqrt((cellSizeX*(ind[0]-R))**2 +(cellSizeY*(ind[1]-R))**2)
"""
code3 = """
dist=array.array('f',chr(0)*(4*winSize*winSize))
for i in range(winSize):
for j in range(winSize):
dist[i*winSize+j]=((cellSizeX*(i-R))**2
+(cellSizeY*(j-R))**2)**0.5
"""
The first one is original, the second is vectorized, the last one is
using array.array. Here is the timing in microseconds for R=1,2,5,6,50
68.2184467579
95.1533784322
17.6263407779
>>>
173.079790867
108.401514718
39.155870369
>>>
802.235479649
174.977043172
161.933094849
>>>
1126.00120965
206.326781756
221.849145632
>>>
67636.3336681
6736.00415695
12675.9099271
>>>
As you see if R >=6 vectorized version is the fastest, otherwise
array.array is faster. You should also try psyco + array.array
> Do you mean in C ?
I didn't mean it, but it's also an option if you need the best speed.
You can try pyrex first, before resorting to raw C.
Serge.
--
http://mail.python.org/mailman/listinfo/python-list
