Latency compensation in scsynth & supernova

[smiarx]

There is currently a heated debate on the Ableton Link PR on how to implement accurate sample-level synchronization between instances. While doing some tests I noticed some inconsistencies regarding latency compensation in scsynth and supercollider for timestamped events:

• scsynth with jack audio backend has a sign error when computing true output time, which leads into a doubling of the latency ! (I can't believed that I never noticed this).
• supernova has no mechanism for latency compensation. I implemented it similarly to how it's done in scsynth.

[telephon]

There is currently a heated debate

this is not what I'd call a heated debate! It is really a friendly conversation on what might be the best solution.

accurate sample-level synchronization

There is only (very) accurate UTP-level time-stamp synchronzation in supercollider. We (unfortunately) have no means for sample level sync.

Where exactly is the difference between scsynth and supernova? Both do support OSC bundles with time stamps, I had thought. What is it you corrected?

[smiarx]

Sorry I didn't mean that in a negative way at all ! I read with interest everyone's ideas. My apologies.

Most audio applications use estimated audio latency to compute signals ahead of time so that the audio is produced just at the right moment (see PortAudio doc for example).

For example you have a latency of 32ms and you want to produce a sound at a time T. If you start writing in the buffer at T you will hear the sound at T + 32ms. So by advancing the internal clock by 32ms, the buffer will be written ahead of time, and the audio will produced at T.

That's what scsynth is doing, except that in the Jack backend the latency was subtracted instead of added, so the audio was computed late.
Supernova wasn't implementing this mechanism, this can be highlighted by running one instance of scsynth and one instance of supernova and sending them exactly the same commands. The instances produce delayed signals.

This is also done in the Link demo

[telephon]

Ah I see. I had supposed that not only for scsynth but also for supernova the scheduled time means the nanosecond (according to network time) when the signal hits the wire. Thanks for the fix!

[nhthn]

i feel like i would have a better grip on what this is doing if i had some test code. i'm not asking for a unit test, just a quick experiment to demonstrate the effect of this change.

also, congrats on finding a )/( :)

[telephon]

@smiarx could you post a quick example of how you verified that it works?

[smiarx]

thanks @snappizz :).

So I used two examples to test it. The first one checks that scsynth & supernova are in sync.

(
~scsynth = Server(\scsynth, NetAddr("localhost", 57110));
~scsynth.waitForBoot{
	~supernova = Server(\supernova, NetAddr("localhost", 57111));
	Server.supernova;
	~supernova.boot;
	~scsynth.latency = 0.2; ~supernova.latency = 0.2;
}
)

(
SynthDef(\click, {
	OffsetOut.ar(0, Impulse.ar(0, mul:0.6));
	Line.kr(dur:0.01, doneAction:Done.freeSelf)
}).send(~scsynth).send(~supernova);
)

(
~pattern = Pbind(\instrument, \click);
Ppar([
	Pbindf(~pattern, \server, ~scsynth),
	Pbindf(~pattern, \server, ~supernova)
]).play();
)

One can then use a recording tool to check the waveform of the output of the two servers. Before the fix supernova should be late by a time corresponding to the estimated latency (or the opposite for Jack). After the fix the clicks should be in sync.

I also tested it against an external tool using the LinkClock branch. In my case I used linkhut from the Link demos, but I believe that any other Link metronome should do it. Note that the Jack version of linkhut use the wrong latency estimation, use this patch to make it use the Jack API estimated latency (same as scsynth).

s.boot;
s.latency = 0.2;
l = LinkClock.new;
l.tempo = 1; // be sure that the tempo is 60 bpm on both sides

(
SynthDef(\click, {
	OffsetOut.ar(0, Impulse.ar(0, mul:0.6));
	Line.kr(dur:0.01, doneAction:Done.freeSelf)
}).add;
)

// I use the phase argument of Quant to compute ahead of time
// so that clicks happen right on time
Pbind(\instrument, \click).play(l, quant:[1, -0.2]);

I get <1ms of latency between the two programs with the fix.

[nhthn]

this works as expected and the code looks fine. thanks!

[sonoro1234]

SC_PortAudio was left out!!
https://github.com/supercollider/supercollider/blob/develop/server/scsynth/SC_PortAudio.cpp#L209