Re: conv2 performance

[Jaroslav Hajek]

On Sun, Feb 28, 2010 at 4:53 PM, John W. Eaton <address@hidden> wrote:
> I recieved a complaint that Octave's conv2 function is about 5 times
> slower than Matlab's.  By making a few simple changes, I was able to
> improve the perfromance some, but not by a factor of 5.  My changes
> are here:
>
>  http://hg.savannah.gnu.org/hgweb/octave/rev/dc8637fd7a76
>
> On my system, I see the following improvement after these changes:
>
>  > n = 1024; m = 64; x = rand(n,n); k = rand(m,m); tic; y = conv2(x,k); toc
>
>  old: Elapsed time is 26.2576 seconds.
>  new: Elapsed time is 13.6104 seconds.
>
> and
>
>  > n = 1024; m = 512; x = rand(n,1); k = rand(m,1); m = rand (m,n);
>  > tic; y = conv2(x,k,m); toc
>
>  old: Elapsed time is 8.53273 seconds.
>  new: Elapsed time is 3.27103 seconds.
>
> Maybe there is still room for improvement here.  I would happily use a
> free software library with a GPL-compatible license to implement this
> function, but I don't know whether one is available.
>
> jwe
>
>


I have some ideas for speeding up the conv2 in stock. Using an
existing library would of course be better, but I don't know of a
suitable one. Unless anyone can come up with an existing solution,
I'll try to convert the ideas to code eventually.


-- 
RNDr. Jaroslav Hajek, PhD
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz