Re: [PATCH 17/29] staging: bcm2835-audio: Add 10ms period constraint [Resend in plain text...]

[Mike Brady <mikebrady@eircom.net>]

Hi Takashi.

> On 9 Oct 2018, at 14:44, Takashi Iwai <tiwai@suse.de> wrote:
> 
> On Tue, 09 Oct 2018 15:18:15 +0200,
> Mike Brady wrote:
>> 
>>>> @Mike: Do you want to write a patch series which upstream "interpolate
>>>> audio delay" and addresses Takashi's comments?
>>>> 
>>>> I would help you, in case you have questions about setup a Raspberry Pi
>>>> with Mainline kernel or patch submission.
>>> 
>>> Well, the question is who really wants this.  The value given by that
>>> patch is nothing but some estimation and might be even incorrect.
>>> 
>>> PulseAudio won't need it any longer when you set the BATCH flag.
>>> Then it'll switch from tsched mode to the old mode, and the delay
>>> value would be almost irrelevant.
>> 
>> Well, two answers. First, Shairport Sync
>> (https://github.com/mikebrady/shairport-sync) needs it — whenever a
>> packet of audio frames is about to be added to the output queue (at
>> approximately 7.9 millisecond intervals), the delay is checked to
>> try to maintain sync to within a few milliseconds. The BCM2835 audio
>> device is the only one I have yet come across with so much
>> jitter. Whatever other drivers do, the delay they report doesn’t
>> suffer from anything like this level of jitter.
> 
> OK, if there is another application using that delay value, it's worth
> to consider providing a fine-grained value.
> 
>> The second answer is that the veracity of the ALSA documentation
>> depends on it — any application using the ALSA system for
>> synchronisation will rely on this being an accurate reflection of
>> the situation. AFAIK there is really no workaround it if the
>> application is confined to “safe” ALSA
>> (http://0pointer.de/blog/projects/guide-to-sound-apis).
> 
>> On LMKL.org, Takashi wrote:
>> 
>>> Date	Wed, 19 Sep 2018 11:52:33 +0200
>>> From	Takashi Iwai <>
>>> Subject	Re: [PATCH 17/29] staging: bcm2835-audio: Add 10ms period constraint
>> 
>>> [snip]
>> 
>>> That's OK, as long as the computation is accurate enough (at least not
>>> exceed the actual position) and is light-weight.
>> 
>>> [snip]
>> 
>> The overhead is small -- an extra ktime_get() every time a GPU message
>> is sent -- and another call and a few calculations whenever the delay
>> is sought from userland.
>> 
>> At 48,000 frames per second, i.e. approximately 20 microseconds per
>> frame, it would take a clock inaccuracy of roughly
>> 20 microseconds in 10 milliseconds -- 2,000 parts per million — to
>> result in an inaccurate estimate. 
>> Crystal or resonator-based clocks typically have an inaccuracy of
>> 10s to 100s of parts per million.
>> 
>> Finally, to see the effect of the absence and presence of this
>> interpolation, please have a look at this:
>> https://github.com/raspberrypi/firmware/issues/1026#issuecomment-415746016,
>> where a downstream version of this fix was being discussed.
> 
> I'm not opposing to the usage of delay value.  The attribute is
> provided exactly for such a purpose.  It's a good thing (tm).
> 
> The potential problem is, however, rather the implementation: it's
> using a system timer for interpolation, which is known to drift from
> the actual clocks.  Though, one may say that in such a use case, we
> may ignore the drift since the interpolation is so narrow.

Yes, that was my thought. I guess another thing in its favour is that this audio device will always
be in partnership with a processor as part of an SoC, so it will always be likely to have a reasonably
accurate clock.

> But another question is whether it should be implemented in each
> driver level.  The time-stamping is basically a PCM core
> functionality, and nothing specific to the hardware, especially when
> it's referring to the system timer.

That’s a fair point. I don’t know what is done in other drivers, but can only report that with one possible exception,
the DACs used with Shairport Sync by many end users report well-behaved delay figures, certainly to within two microseconds. I’m afraid I don’t know how they do it.

> e.g. you can think in a different way, too: we may put a timestamp at
> each hwptr update, and pass it as-is, instead of updating the
> timestamp at each position query.  This will effectively gives the
> accurate position-timestamp pair, and user-space may interpolate as it
> likes, too.

That’s not a bad idea, and I might take it up on the alsa-devel mailing list, as you suggest.

> In anyway, if *this* kind of feature needs to be merged, it's
> definitely to be discussed with the upstream.  So, if you're going to
> merge that sort of path, please keep Cc to alsa-devel ML.

In the meantime, would you think that the balance of convenience lies with this interpolation scheme? (Finally, I have a patch ready….)
Regards
Mike

> 
> thanks,
> 
> Takashi