Re: [Help-gsl] gsl_vector stride

[G1ul10]

On Mon, 7 Feb 2005 09:50:35 +0000 (GMT)
Henry Braun <address@hidden> wrote:

> I have some convolutions $C_j = \sum_{i = 0}^n x_i y_{j-i}$
> to evaluate and the obviously slick technique would be to create
> views of the $y$ vector with negative stride to pass to gsl_blas_ddot.

Well, I don't know if this technique could be considered slick, but it
is tremendously inefficient. If you have very short arrays, then I would
suggest you to just write down explicitly the for loops. But if you have
large arrays, probably it is better to switch to DFT (discrete fourier
transform). See Numerical recipes for an intorduction (chapter 13). If
you have very large arrays and speed is an issue, then I suggest you to
use FFTW as a FFT package instead of the GSL implementation.

> 
> Not to be able to do this looks like a serious gap in functionality.
> Is there a workaround?
> 

If you follow my advice, there's no need to work around anything. But
for the size_t vs. signed int issue see the examples in the GSL design
document.


        G.

pgp1NEXjKUvbJ.pgp
Description: PGP signature