[SciPy-user] Help on performance of signal.convolve

[Yichun Wei]

Dear Experts,

Sorry if I was not concrete or even not correct last time I posted this
for help.

I'd like to convolve a (64,64,41) kernel with a (64,64,1800)
array with mode='valid' . What would be the fastest method in scipy?

Here I tried with signal.convolve and it takes >400 s to solve.
a.shape is (64,64,41), b.shape is (64,64,1800)

res = signal.convolve (a, b, mode='valid')

it took around 420 s CPU time to solve on my P-IV 1.8G CPU. I have the
file dumped from profile, if you want to have a look I can attach it.
'same' and 'full' never ends when I ran them. I am using the Enthought
Python with scipy 0.3. Is this performance normal on a P-IV 1.8G CPU?

>>> p.sort_stats('cumulative').print_stats(10)
Wed Jan 12 11:31:08 2005    Profile_k_GetRespons_same

    1631 function calls (1623 primitive calls) in 420.407 CPU seconds

    Ordered by: cumulative time
    List reduced from 175 to 10 due to restriction <10>

    ncalls  tottime  percall  cumtime  percall filename:lineno(function)
         1    0.001    0.001  420.407  420.407 profile:0(res =
k.GetResponse())
         1    0.000    0.000  420.406  420.406 <string>:1(?)
         1    0.000    0.000  420.406  420.406 
F:\tmp\py\cte\kernel.py:173(GetResponse)
         1  419.705  419.705  419.705  419.705
C:\Python23\Lib\site-packages\scipy\signal\signaltools.py:79(convolve)
       5/1    0.000    0.000    0.701    0.701
C:\Python23\Lib\site-packages\scipy_base\ppimport.py:299(__getattr__)
       5/1    0.033    0.007    0.701    0.701
C:\Python23\Lib\site-packages\scipy_base\ppimport.py:252(_ppimport_importer)
         1    0.091    0.091    0.699    0.699
C:\Python23\Lib\site-packages\scipy\signal\__init__.py:5(?)
         1    0.000    0.000    0.395    0.395
C:\Python23\Lib\site-packages\scipy\signal\signaltools.py:4(?)
         1    0.091    0.091    0.313    0.313
C:\Python23\Lib\site-packages\scipy\stats\__init__.py:5(?)
         1    0.019    0.019    0.196    0.196
C:\Python23\Lib\site-packages\scipy\signal\bsplines.py:1(?)


<pstats.Stats instance at 0x007E9DA0>


I read some performance guide like the one by Prabhu at
http://www.scipy.org/documentation/weave/weaveperformance.html. But
since this is only a function call to sigtools._correlateND, I think it
is already implemented in C++. If it is the case, I think it is not
profitable to use blitz, swig or f2py.

Also, I find there is a fftpack.convolve, however I am not sure if it 
works only on 1-d array, or if it is appropriate to use fft in this 
convolution I will do. (I also find in numarray the convolution object 
got an option to decide whether or not to use fft.)

Could you be kind enought to point out where the effort should be put to 
improve the performance of such a convolution? Any hint will be greatly 
appreciated!!

- yichun