I think we should keep the baresip related config examples in the baresip repo and reference to ajb docs if necessary.

This is not meant as documentation. It is the baresip config for generate_plots.sh. Maybe it is displaced.

I wonder if auframe_level performance is really a problem? In my quick measurements its nearly the same (and we can reuse the level for vumeter etc.). Silence tracks are more expensive since comparison costs much more than addition.

We do not need the vumeter on our products and the silence detection works already. We don't want to have 640 sqrt and logarithm operations for each frame.

The sqrt is only done once per frame sum. not per sample.

Okay, silence detection is only needed for AJB_LOW and AJB_HIGH situations. But it is not the same like computing a volume in dB. This might justify a separate function auframe_silence(). The note can be removed, it is meant only for the review.

Ah, yes, you're right! Logarithm also. Thus there are only the O(n)

• double multiplications instead of int
• double additions instead of int
• and the early exit with return false which is not possible for auframe_level().

On one product we have a very old hardware (balckfin CPU). I will check how much the impact would be to replace the silence detection with the auframe_level() computation.

Yes the double multiplications has much more impact, i will prepare a patch (i dont think we need this), here are some results:

auframe_silence 3468 usecs (2000)
auframe_silence2 665 usecs (2000)
auframe_silence 3572 usecs (2000)
auframe_silence2 645 usecs (2000)

https://gist.github.com/sreimers/67440f237cae34537c9c1667900a7c57

On one product we have a very old hardware (balckfin CPU).

We can use maybe a square root alternative:

static float square_root(float x)
{
	unsigned int i = *(unsigned int *)&x;

	// adjust bias
	i += 127 << 23;
	// approximation of square root
	i >>= 1;

	return *(float *)&i;
}

On modern hardware I see no difference with sqrt.

Thanks, improvement here would be welcome! If the performance loss is only little, than I would remove the extra silence detection.

Edit:

• arm would be interesting
• and maybe somebody of my colleagues tries this on our blackfin platform

@sreimers Do you have a sample program for this benchmarks?

https://gist.github.com/sreimers/67440f237cae34537c9c1667900a7c57

And this patch: #38

Compiled with RELEASE=1

thanks!