Re: conv2 performance

[Michael D. Godfrey]

On 3/1/10 10:56 AM, Soren Hauberg wrote:
> man, 01 03 2010 kl. 10:28 -0800, skrev Robert T. Short:
>    
> > >  Create separate functions for
> > >  dft-based convolutions (fftconv?).
> >      
> We have 'fftconv' in the core already. The 'image' package has
> 'fftconv2'. It seems we don't have the N-dimensional version anywhere,
> though.
>
> Soren
>
>    
fftconv is a .m file and the implementation, while apparently sound, is
not very efficient.  A case can be made that the core fft should be a C++
routine which could be used to implement various fft-based functions.
A well thought out C++ implementation is described at:
http://www.drdobbs.com/cpp/199500857

It is not, as far as I could find, made clear if this is free software, 
but it
appears to be.  The author's email address is included in the article.
And, the references at the end are all good sources of up-to-date work.

Michael