More about de-normals - some test results. - Linux-audio-dev

4 May 2010

I wrote:
...
 > I tired a quick mod in MusE to do what the author
of the caps ladspa suite
>  did to handle de-normals. He said "A -80dB signal at the Nyquist frequency
>  or lower". No luck.
> But yeah, obviously at some signal level and type, it should stop.
> So I'll keep trying. Noise sounds like the best way. -100dB white to start?
> OK...
>
> Ugh. A new MusE options panel: Advanced de-normalization options, he he...
> Paul wrote:
...
 You might want to check ardour. It has 3 denormal
protection options (2 that
are h/w based, setting processor flags, and 1 that is software based, adding
a very very very very tiny constant value ("DC Bias") to every signal. See
libs/pbd/fpu.cc to find the h/w stuff. 
I raised MusE's de-normal DC bias to a very high 1.0e-2. 
Same problem.

So I slapped some printf traces on various float data streams in MusE
 to see what was really happening.

Here are the first four float values of the buffers involved, into and out of
 the plugin:

Into the plugin:  1.000000e-02 1.000000e-02 1.000000e-02 1.000000e-02                    
Out of the plugin: -1.398807e-19 8.058466e-19 1.999431e-18 3.264671e-18             

What comes out is a meandering, wandering series of floats which 
 eventually wanders into de-normal territory.

This data eventually makes its way to Jack for audio output. 

I might conclude then, that the plugin AC-couples it's input to remove 
 any DC bias.

So DC input bias won't help, and likely wouldn't help either if applied 
 to the wandering output. Seems only noise or other AC signal cures it.

I gather this also means we must be careful what we feed into Jack.

Haven't tested Ardour yet, but what do you think about this, Paul?
I suppose Ardour's processor flags option *must* cure it?

Still testing... Thanks. Tim.