[pulseaudio-discuss] Latency problem with long latencies

[Georg Chini]

On 20.03.2016 19:13, Alexander E. Patrakov wrote:
> 20.03.2016 22:41, Georg Chini пишет:
>> Hello,
>>
>> I am still working on module-loopback and hit a problem that I cannot
>> explain.
>> When running a HDA card with long latency (333ms) I see that the 
>> resulting
>> latency is not stable but varies around 300 usec. What is worse, the
>> changes
>> in latency are not even reported, module-loopback sees a completely 
>> stable
>> value ( +/-10 usec).
>> My setup: I have a bitscope dual channel USB oscilloscope attached to
>> another
>> machine. This device also has a function generator that I set to 1kHz
>> rectangle.
>> The signal is input to the HDA card and fed to the output of the same
>> card via
>> module-loopback. Input and output signal are measured using the
>> oscilloscope.
>> The movement of the edge of the output signal now shows the "latency
>> jitter"
>> which can be made visible by using the overlay mode of the oscilloscope.
>>
>> The result for HDA (module-loopback running with 500ms adjust time 
>> and 1s
>> latency) is shown at http://georg.chini.tk/hda_333ms_3.png. Some of the
>> original
>> input signal is somehow fed into the output signal, don't know where
>> this happens,
>> so please ignore it as it has no impact on the result.
>
> No idea either, just some blind pokes...
>
> 1. Are you using the trivial resampler? Does anything change if you 
> change the resampler type?
> 2. Do you get the same trace with a "weird" latency (not something 
> that can be derived from one second by multiplication or division by 
> small factors)? E.g. something like 318 ms.
>
> To me, it looks like something is sometimes creating an 
> off-by-one-sample error when copying the data. Which is exactly the 
> job of the trivial resampler, that's why the question.
>
Hi Alexander,

no, I am working with speex-float-5. I just re-tested zitaj-bridge, and 
I was wrong,
it shows the same behavior. I am also seeing double lines, so it is 
probably deeper
down in the alsa driver.
I also tested your 318 ms - same result.
To rule out artifacts produced by the oscilloscope I fed the input 
signal directly to the
second channel - this looks fine.
The choice of the resampler does no longer have any impact on the 
performance
of module-loopback. For small latencies (70ms - 100ms) I can keep the 
latency jitter
now well below 100 usec for HDA -> USB on my system, this corresponds to a
difference of 35 mm in hearing distance. When the USB device is run in 
batch mode
the result is even better but less robust to changes in system load.
zitaj-bridge is still somewhat better for HDA -> USB and USB -> HDA (~50 
usec)
but always fails after a few minutes complaining about excessive timing 
errors
even for HDA -> HDA on the same card.