[FFmpeg-devel] Fwd: Fixpoint FFT optimization, with MDCT and IMDCT wrappers for audio optimization

[Marc Hoffman]

Hi,

On 8/17/07, Michael Niedermayer <michaelni at gmx.at> wrote:
> Hi
>
> On Fri, Aug 17, 2007 at 03:58:16PM +0200, Benjamin Larsson wrote:
> > Michael Niedermayer wrote:
> > > Hi
> > > [...]
> > >
> > > dont fear ill revert any cooley tukey fixed point FFTs which get commited
> > >
> >
> > Ok, then you can start by reverting revision 5.
> > http://svn.mplayerhq.hu/ffmpeg/trunk/libavcodec/ac3enc.c?view=markup&pathrev=5
> >
> > I don't see it as obvious that a split radix fft would be faster on
> > every fixed point platform out there.
>
> well on which is it not?
>
>
> > Would it really hurt to have
> > several fft algos available in the source?
>
> depends on their speed, accuracy, ...
> theres no sense in keeping things which are always inferrior
> keeping them until they have been optimized and benchmarked is all fine
> but we still are at step 0 (that is we dont have an implementation we rather
> disscuss what we could do if we did had one which we dont ...)
>
>
> > Without a radix-2 algo we
> > wouldn't have anything to compare against to back that claim anyway. I'm
> > not really doubting that a split radix fft would be faster in most cases
> > but I'd rather let the benchmark numbers prove the point.
>
> who did suggest that we shouldnt benchmark the code?
> iam just saying that id like to see a split radix fft in lavc, if someone
> wants to write a radix2 one thats fine i wont stop him and ill surely look
> at any benchmarks

Can someone tell me all the data set sizes of the input signal used by
the FFT in the system (ffmpeg)?

Thanks
Marc