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

[Yichun Wei]

Hi Travis,

Travis Oliphant wrote:

> So, something like the following untested code should get you started:
> 
> def convolvefft(arr1,arr2):
>        s1 = array(arr1.shape())
>        s2 = array(arr2.shape())
>        fftsize = s1 + s2 - 1
>        # finds the closest power of 2 in each dimension (you may comment 
> this out and compare speeds)
>        fftsize= pow(2,ceil(log2(fftsize))) 

I tried for larger kernels. This does matters!

> 
>        ARR1 = fftn(arr1,fftsize)
>        ARR2 = fftn(arr2,fftsize)
>       
>        RES = ifftn(ARR1*ARR2) 
>        #RES=RES[validpart]  # I'm not sure how to get the correct part 
> --- first try would be to just truncate to the shape you wanted
>        return RES  

Using this code a convolution of (16,16,40) kernel with (16,16,1800) 
signal takes 5s to solve on my 1.8G P-IV CPU.

Because I am only interested in the "valid" convolution. What could be 
used to speed this up a bit more? I am really in need of speed for I 
have to do this convolution lots of times.

best,

- yichun






-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: convolvefft.py
URL: <http://mail.scipy.org/pipermail/scipy-user/attachments/20050118/af36fa32/attachment.ksh>