[pulseaudio-discuss] Virtual audio cable - high cpu usage

[Pierre-Louis Bossart]

On 2/4/21 7:15 PM, Sean Greenslade wrote:
> On Thu, Feb 04, 2021 at 06:29:47PM +0100, Renaud GHIA wrote:
>>>> Thank you for the tip.
>>>> Now I am sure that resampling does not apply (see below).
>>>> But unfortunately pulseaudio always consumes 30% of one CPU core!
>>>>
>>>
>>> The reason is that you are using module-loopback. Due to a potential
>>> difference between the clocks on the null sink and on the real sound card
>>> (even though both report 44100 Hz), it always has to resample in order to
>>> correct for the clock skew. If you don't understand, here is an analogy:
>>> you have two mechanical watches. Even though both claim that their hour
>>> hand makes one full circle every 12 hours, in fact, if left unattended,
>>> they will diverge over time. There is no way to avoid that, except by
>>> moving to Jack or PipeWire which simply don't introduce a null sink with an
>>> independent clock.
>>>
>>> If you decide to stay on PulseAudio, you can tweak the resampling method.
>>> The default, speex-float-1, is light on the CPU resources and should
>>> produce no distortions detectable by human ear on typical speech and music.
>>> It does produce easily detectable distortions on specifically crafted
>>> signals. If you want to make sure that the resampler is transparent no
>>> matter what is thrown at it, use speex-float-5. There is no point in going
>>> higher than that.
>>
>> Thanks for your response.
>> I understand the problem of clock drift when we have several audio hardware
>> devices (with different qwartz).
>> But here, it's a pity that the null video driver has an independent clock...
>> Concerning pipewire, for now it doesn't work with our particular
>> application (no sound), but it seems to evolve quickly.
>> I haven't tried jack for now, but if it depends on pulseaudio, won't I have
>> the same problem?
> 
> Jack doesn't depend on pulseaudio, it's an independent sound server.
> Though there are ways to make the two interconnect.
> 
> As far as I know, jack solves this problem by only allowing one hardware
> sound device and locking all clocks to that. See:
> 
> https://jackaudio.org/faq/multiple_devices.html
> 
> One thing that just occured to me was to ask whether you acutally need
> to use the remap module at all. I just did a quick test on a very weak
> machine of mine: a null sink, a music player outputting to that sink,
> and ffmpeg recording the null sink's monitor-source, all with matching
> 44.1 kHz sample rates. This resulted in no resamplers in the chain, and
> only 9% CPU usage on a 2010 Atom CPU.

If you have an HDaudio device, it's quite likely that the HDaudio bus 
clock and your MONOTONIC_RAW time are locked already. If you use the 
MONOTONIC option for your timers, it'll be subject to NTP corrections 
but for short-enough rendering you shouldn't see a problem without 
resampling.
My 2 cents.