Re: [fluid-dev] Floats and doubles, simd and interpolation

[David Henningsson]

On 2010-11-27 10:35, Pedro Lopez-Cabanillas wrote:
> On Saturday 27 November 2010, David Henningsson wrote:
> > > > Yesterday, I tried to chase down a bug I discovered - I rendered a song but
> > > > a few notes were slightly out of tune, but only when using floats, not when
> > > > using doubles!
> [...]
> > > I wonder what is causing the bug you described though.  If we could
> > > setup a reproducible file rendering case it would probably help track
> > > down the problem.
> >
> > The soundfont is the standard FluidR3_GM.sf2, the midi file is
> > http://www.mididatabase.com/20100702/pop/steviewonder/Dont_You_Worry_About_a_Thin.mid
> > , and the first offending notes are on channel 2, ~25 seconds into the
> > song. The notes are affected by pitch bend, this is probably related
> > somehow. That's how long I've come so far.
>
> I couldn't reproduce the problem in any of my development machines. Using 
> floats or doubles it sounds pretty much alike.
>
> But I've remembered another report about a similar problem, also related to 
> notes and pitch bend events:
> http://lists.nongnu.org/archive/html/fluid-dev/2010-10/msg00101.html
> http://sourceforge.net/apps/trac/fluidsynth/ticket/91
>
> I couldn't reproduce a wrong behavior playing this song, either.

After long search, I finally found it! \o/

It's committed as http://sourceforge.net/apps/trac/fluidsynth/changeset/395/

Seems like we[1] used fluid_real_t to represent sample offsets at one 
time, and then the sample offsets got rounded off, so that the loop went 
slightly longer or shorter.

// David

[1] Looking at svn history, it was my fault originally.