[FFmpeg-devel] [PATCH] G.729A (floating-point) decoder and ACT demuxer

[Reimar Döffinger]

On Thu, Feb 28, 2008 at 01:10:46AM +0600, Vladimir Voroshilov wrote:
> > > +/**
> >  > + * L1 codebook (10-dimensional, with 128 entries (3.2.4)
> >  > + */
> >  > +static const float cb_L1[128][10] = {
> > >
> > > [...]
> > >
> >
> >  *10000 and it fits in uint16_t
> 
> This and tables below replaced with fixed-point tables from AnnexA
> reference code but this decreased PSNR significantly.
> Multiplying by 10000 did not decrease  PSNR, though

Hm, as you said the fixed-point tables differ in rounding from the float
ones, probably to make up for different rounding in float operations.
Unless someone volunteers to do a fixed-point implementation it probably
makes more sense to use the float tables scaled to fit in uint16_t.