Speex dev - Sep 2005 - Precomputing the remaining floating pointoperations.

Firstly, running for more channels will not break my hack. All that's needed
 is to call RECOPLAY_MARK with different identifiers (say nb, wb or uwb)
 before doing the appropriate initialization.

 Secondly, my attempts to do the Gaussian in fixed point went like this :
 Define a new constant lag_factor_gauss that is manually set equal to
 exp(sqr(2*M_PI*lag_factor)/-2) by whoever changes the lag_factor. From bash
 you can say
echo 'e((8*a(1)*.012)^2/-2)' |bc -l
Then we can replace the code in the init functions with :
   gauss = 16384;
   for (i=0;i<st->lpcSize+1;i++) {
      st->lagWindow[i]=gauss;
      for (j=i*i; j < (i+1)*(i+1); j++) {
         gauss = MULT16_16_Q15 (gauss, st->lag_factor_gauss);
      }
   }
A Taylor expansion will be more accurate.

Le mardi 27 septembre 2005 ? 23:24 +0200, Nic Roets a ?crit
:
> Firstly, running for more channels will not break my hack. All that's
needed
>  is to call RECOPLAY_MARK with different identifiers (say nb, wb or uwb)
>  before doing the appropriate initialization.
Still, too many ways it can fail. You can maintain it as a separate
patch it you like, but this is not going in the main tree.
>  Secondly, my attempts to do the Gaussian in fixed point went like this :
...
>    for (i=0;i<st->lpcSize+1;i++) {
>       st->lagWindow[i]=gauss;
>       for (j=i*i; j < (i+1)*(i+1); j++) {
>          gauss = MULT16_16_Q15 (gauss, st->lag_factor_gauss);
>       }
>    }
Way more complicated that it needs to be
> A Taylor expansion will be more accurate.
Not only more accurate, but a lot simpler because I'm only using a very
small portion of the gaussian. I think a third order approximation would
work.