Re: [fluid-dev] Automatic polyphony reduction?

[Josh Green]

On Mon, 2005-10-17 at 16:22 +0300, Kimmo Sundqvist wrote:
> I made an ebuild for this purpose.
> 
> Configure says:
> **************************************************************
> Summary:
> ALSA:                  yes
> OSS:                   yes
> MidiShare:             no
> JACK:                  no
> CoreAudio:             no
> LADSPA support:        no
> LADCCA support:        no
> Readline:              yes
> Debug:                 no
> Profiling:             no
> use long long:         no
> Pentium 3+ SSE:        no
> **************************************************************
> 
> > Ohh, it just occurred to me that the PIII has problems with denormal
> > floating point numbers (very inefficient when processing very tiny
> > floating point numbers).  There were some fixes a long while back in
> > regards to this issue, but perhaps some have surfaced again.
> 
> Playing music didn't bring the load down, and now we also know the problem is 
> not LADSPA.  Any switches for gcc I could try out?  Or should I try disabling 
> readline and oss support?
> 
> -Kimmo S.
> 

Ahh, another thought occurred to me.  Which audio driver are you using?
Are you certain that the sample rate that FluidSynth runs at (defaults
to 44100) is natively supported by your sound card?  I think the newer
version of ALSA will do sample rate conversion if a program tries to use
a sample rate that isn't natively supported by the card.  In practice
this often is extremely inefficient.

One test you could try is to use the file output driver (run "fluidsynth
-a file").  This will dump a fluidsynth.raw file, which should be
deleted afterwards.  If you don't see max CPU usage, then its probably
some issue with the audio driver.  You could try different sample rates
with the -r switch ("fluidsynth -r 48000" for example), different audio
drivers (OSS, JACK, etc).

I was going to recommend enabling profiling (--enable-profiling), but I
just noticed that this is not working currently (results don't seem
accurate).  Best regards,
        Josh Green

signature.asc
Description: This is a digitally signed message part