Fixed point (was: RE: [iiwusynth-devel] Fw: Re: [linux-audio-dev] more about iiwusynth)

[M. Nentwig]

Hei,

> Not much luck over here, the fixed point mixer itself is  fast and it
> works, but the integer->float conversions to save back to the buffers
> decompensate the speed. I'll try later converting the values to float
> manually to see if i can gain some speed.

That's odd. The samples are fixed point anyway, and the output buffer is
float.
The existing iiwusynth converts the samples to float and then continues
to process in float. And you probably take the samples as they are,
process fixed point and then convert to float. In both cases, one (1)
fixed-to-float conversion is needed. Now I don't understand why this
isn't much faster...
Well, maybe long => float takes much more time than short => float, that
could be an explanation?

> I think that i'll also try
> rewriting the filter to use a two poles FIR with 2 coheficients only.
> Any suggestions?

The current filter is two-pole IIR (you probably meant IIR, too?). It
actually has two poles and two zeroes. I just wonder what happens, if we
get rid of the zeroes, that could save two coefficients.

> I think the fastest would be to use 32bits integer
> buffers instead, and convert to float only for effects..

Agree.

>
> Regads!
>
> Juan Linietsky
>

Regards

Markus