Re: [fluid-dev] FluidSynth 1.1.2 release planning

[David Henningsson]

2010-07-17 15:26, Bernd Casper skrev:
> Hi David,
> 
> many thanks for your efforts in preparing a new release.
> I'm testing the FS 1.1.2 beta extension Sven Meier has compiled for jOrgan, 
> on WinXPSP3, 4GB RAM, 2x3 GHz DualCore AMD processor.
> 
> With 1.1.2 beta I experience strange cracklings at the attack of every note. 
> No problems with 1.1.1.
> First scale in the example: FS 1.1.1, second scale: FS 1.1.2 beta. The preset 
> uses more than one sample. Altering interpolation did not result an 
> improvement.
> http://bca.free-artists.net/data/FS112_trial_crackles.wav  

Strange. Could you tell me a little more about this test, are you
1) Fast-rendering to file
2) Output to file
3) Analog loopback (recording "stereo mix", or a loopback cable),

and does it matter which strategy you use?

> Regarding the voice stealing algorithm:
> 
> Organ instruments live from a steady mixture of sounds.
> The term "sound" equals single instruments in a GM soundfonts. At the organ 
> they're called "ranks". So the organ can be the same as whether I'd use all 
> 127 sounds of a GM soundfonts at the same time, potentially, in certain 
> situations.
> 
> The principle of creating sound colours, resulting from mixing several organ 
> ranks steady together, seems not to allow to steal sounds. An organ expects 
> as much samples remaining as possible. Best would be if the voice stealing 
> mechanism could be disabled completely by a command.
> 
> For the purpose of not to suffer from voice stealing, I use up to 20 
> instances of FS in one organ model. By testing I came to the result, I can 
> run 7 sounds per instance successfully. After that, I've to end one sound to 
> free resources for starting another. Also, the size of the samples seems to 
> influence FS performance. 

I've played some organ myself (real ones only, not MIDI versions). And
it's logical to think that the polyphony should be equal to the number
of pipes in the organ. That would be true for FluidSynth, if it weren't
for the release time, i e if you quickly unpress and then press the key
again, both voices will be active at the same point in time (since the
first one is in its release stage and the second one in attack stage).

Is that what you're trying to explain?

> The bigger the sample size, the bigger the latency.

Could you elaborate on this?

> On my PC FS 1.1.2 beta seems to use much more CPU resources.

There shouldn't be that much difference; if anything, if you enable
multi-core (and possibly have high latencies), it should be faster. It's
possible that MIDI operations might be a little slower though.

Could it be that Sven accidentally compiled a debug build? They're 3-4
times slower than a release build (at least on my machine), since all
compiler optimizations are turned off to ease debugging.

> Did you already alter or re-animate something in the voice stealing 
> mechanism, so I could test rejecting some of the many instances?

Voice stealing is currently not functional. I'm working on it. That
said, I'll have to go away for a week now, so don't expect much the
coming week.

> Helpful wound be, if you'd provide information, which computer system 
> parametres influence the FS performance in fact - like amount of RAM memory, 
> CPU speed etc.

I would say it would be nice with some testing here. More RAM might help
to keep big samples in memory, whereas CPU speed (and probably CPU cache
size/speed) would speed up the rendering and enable more voices and/or
lower latency.
(Latency depends on many other factors as well.)

Thanks for all testing so far, I really appreciate it!

// David