[pulseaudio-discuss] alsa sink latency - how to account for startup delay

[Tanu Kaskinen]

On Wed, 2016-04-06 at 21:17 +0200, Georg Chini wrote:
> On 06.04.2016 15:55, Tanu Kaskinen wrote:
> > On Tue, 2016-04-05 at 21:54 +0200, Georg Chini wrote:
> > > On 05.04.2016 17:42, Tanu Kaskinen wrote:
> > > > I can't follow that line of reasoning. In the beginning the ring buffer
> > > > is filled to max, and once you call snd_pcm_start(), data starts to
> > > > move from the ring buffer to other buffers (I'll call the other buffers
> > > > the "not-ring-buffer"). Apparently the driver "sees" the not-ring-
> > > > buffer only partially, since it reports a larger latency than just the
> > > > ring buffer fill level, but it still doesn't report the full latency.
> > > > The time between snd_pcm_start() and the point where the reported delay
> > > > does not any more equal the written amount tells the size of the
> > > > visible part of the not-ring-buffer - it's the time it took for the
> > > > first sample to travel from the ring buffer to the invisible part of
> > > > the not-ring-buffer. I don't understand how the time could say anything
> > > > about the size of the invisible part of the not-ring-buffer. Your logic
> > > > "works" only if the visible and invisible parts happen to be of the
> > > > same size.
> > > > 
> > > > You should get the same results by calculating
> > > > 
> > > >     adjusted delay = ring buffer fill level + 2 * (reported delay - ring buffer fill level)
> > > > 
> > > > That formula doesn't make sense, but that's how I understand your logic
> > > > works, with the difference that your fix is based on one measurement
> > > > only, so it's constant over time, while my formula recalculates the
> > > > adjustment every time the delay is queried, so the adjustment size
> > > > varies somewhat depending on the granularity at which audio moves to
> > > > and from the visible part of the not-ring-buffer.
> > > > 
> > > > In any case, even if your logic actually makes sense and I'm just
> > > > misunderstanding something, I don't see why the correction should be
> > > > done in pulseaudio instead of the alsa driver.
> > > 
> > > Well, now I don't understand what you mean. The logic is very simple:
> > > If there is a not reported delay between the time snd_pcm_start() is
> > > called and the time when the first sample is delivered to the DAC, then
> > > this delay will persist and become part of the continuous latency.
> > > That's all, what causes the delay is completely irrelevant.
> > 
> > The code can't know when the first sample hits the DAC. The delay
> > reported by alsa is supposed to tell that, but if the reported delay is
> > wrong, I don't think you have any way to know the real delay.
> 
> Yes, the code can know when the first sample hits the DAC. I explained it
> already. Before the first sample hits the DAC, the delay is growing and
> larger or equal than the number of samples you have written to the
> buffer.
> At the moment the delay is smaller than the write count, you can be
> sure that at least some audio has been delivered. Since the delay is
> decreased by the amount of audio that has been delivered to the DAC,
> you can work back in time to the moment when the first sample has been
> played.

Yes, you explained that already, but you didn't give a convincing
explanation of why the point in time when the delay stops growing would
indicate the point when the first sample hit the DAC.

> > > Maybe what I said above was not complete. At the point in time when
> > > the first audio is played, there are two delays: First the one that is
> > > reported
> > > by alsa and the other is the difference between the time stamps minus
> > > the played audio. If these two delays don't match, then there is an
> > > "extra delay" that has to be taken into account.
> > 
> > The difference between the time stamps is not related to how big the
> > invisible part of the buffer is. I'll try to illustrate:
> > 
> > In the beginning, pulseaudio has written 10 ms of audio to the ring
> > buffer, and snd_pcm_start() hasn't been called:
> > 
> > DAC <- ssssssssss|sss|dddddddddd <- pulseaudio
> > 
> > Here "ssssssssss|sss|ddddddddd" is the whole buffer between the DAC and
> > pulseaudio. It's divided into three parts; the pipe characters separate
> > the different parts. Each letter represents 1 ms of data. "s" stands
> > for silence and "d" stands for data. The first part of the buffer is
> > the invisible part that is not included in the delay reports. I've put
> > 10 ms of data there, but it's unknown to the driver how big the
> > invisible part is. The middle part of the buffer is the "send buffer"
> > that the driver maintains, its size is 3 ms in this example. It's
> > filled with silence in the beginning. The third part is the ring
> > buffer, containing 10 ms of data from pulseaudio.
> > 
> > At this point the driver reports 10 ms latency. It knows it has 3 ms of
> > silence buffered too, which it should include in its latency report,
> > but it's stupid, so it only reports the data in the ring buffer. The
> > driver has no idea how big the invisible part is, so it doesn't include
> > it in the report.
> > 
> > Now pulseaudio calls snd_pcm_start(), which causes data to start moving
> > from the ring buffer to the send buffer. After 1 ms the situation looks
> > like this:
> > 
> > DAC <- ssssssssss|ssd|ddddddddd  <- pulseaudio
> > 
> > There's 2 ms of silence in the send buffer and 1 ms of data. The driver
> > again ignores the silence in the send buffer, and reports that the
> > delay is 10 ms, which consists of 1 ms of data in the send buffer and 9
> > ms of data in the ring buffer.
> > 
> > After 2 ms:
> > 
> > DAC <- ssssssssss|sdd|dddddddd   <- pulseaudio
> > 
> > Reported delay: 10 ms
> > 
> > After 3 ms:
> > 
> > DAC <- ssssssssss|ddd|ddddddd    <- pulseaudio
> > 
> > Reported delay: 10 ms
> > 
> > Let's say pulseaudio refills the ring buffer now.
> > 
> > DAC <- ssssssssss|ddd|dddddddddd <- pulseaudio
> > 
> > Reported delay: 13 ms
> > 
> > After 4 ms:
> > 
> > DAC <- sssssssssd|ddd|ddddddddd  <- pulseaudio
> > 
> > The first data chunk has now entered the invisible part of the buffer,
> > but it will still take 9 ms before it hits the DAC. At this point
> > pulseaudio has written 13 ms of audio, and the reported delay is 12 ms.
> > According to your logic, the adjusted delay is 12 + (4 - 1) = 15 ms,
> > while in reality the latency is 22 ms.
> 
> At this point, no audio has been played yet. You still have silence in the
> buffer, so alsa would not report back, that samples have been played.

But the reported delay stopped growing! That's the point where you
claim the first sample hits the DAC, but as my example illustrates,
that doesn't seem to be true.

> I choose the point where the first d hits the DAC and that is reported
> back by alsa. (see above) I've tried put it all together in a document.
> I hope I can finish the part that deals with the smoother code today.
> If so, I will send it to you privately because the part about 
> module-loopback
> is still missing.
> Anyway, even if you think it is wrong I am still measuring the correct
> end-to-end latency with my code, so something I am doing must be
> right ...

>From what I can tell, that's a coincidence.

> > I don't know how well this model reflects the reality of how the usb
> > audio driver works, but this model seems like a plausible explanation
> > for why the driver reports delays equalling the amount of written data
> > in the beginning, and why the real latency is higher than the reported
> > latency at later times.
> > 
> > I hope this also clarifies why I don't buy your argument that the time
> > stamp difference is somehow related to the unreported latency.
> 
> No, in fact it doesn't.
> 
> > > Trying to fix up that delay on every iteration does not make any sense
> > > at all, it is there from the start and it is constant.
> > 
> > Commenting on "it is constant": The playback latency is the sum of data
> > in various buffers. The DAC consumes one sample at a time from the very
> > last buffer, but I presume that all other places move data in bigger
> > chunks than one sample. The unreported delay can only be constant if
> > data moves to the invisible part of the buffering in one sample chunks.
> > Otherwise the latency goes down every time the DAC reads a sample, and
> > then when the buffer is refilled at the other end, the latency jumps up
> > by the refill amount.
> 
> I only said the "extra latency" is constant, not the latency as such.
> See your own example above that your argument is wrong. Even
> if the audio is moved in chunks through your invisible buffer part,
> that part still has the same length all the time. When one "d" is
> moved forward another one will replace it.

No, the invisible part is not constant, even though my presentation
didn't show the variance. The DAC consumes data from the invisible
buffer one sample at a time, and each time it does that, the extra
latency decreases by one sample. Data moves from the visible part of
the buffer to the invisible part in bigger chunks. I didn't specify the
chunk size, but if we assume 1 ms chunks, the extra latency grows by 1
ms every time a chunk is transferred from the visible part to the
invisible part.

> > > This is not a negative delay reported by alsa, but my "extra latency"
> > > is getting negative, which means playback must have started
> > > before snd_pcm_start().
> > > According to Raymond Yau playback seems in fact to be started
> > > before snd_pcm_start() for HDA devices, at least if I read his last
> > > mail on that topic right. Then the negative delays would even make
> > > sense, since data is written to the buffer before snd_pcm_start().
> > 
> > I had a look at the code to verify the claim that we configure alsa to
> > start playback already before we call snd_pcm_start(). If we really do
> > that intentionally, then it doesn't make sense to call snd_pcm_start()
> > explicitly.
> > 
> > This is what we do:
> > snd_pcm_sw_params_set_start_threshold(pcm, swparams, (snd_pcm_uframes_t) -1)
> > 
> > Note the casting of -1 to an unsigned integer. It seems that the
> > intention is to set as high threshold as possible to avoid automatic
> > starting. However, alsa-lib casts the threshold back to a signed value
> > when it's used, and I believe the end result is indeed that playback
> > starts immediately after the first write. I don't know if that matters,
> > since we do the manual snd_pcm_start() call immediately after the first
> > write anyway, but it seems like a bug in any case.

Not very important, but I'll clarify one thing: I had another look, and
I'm not any more sure that the code where I saw the casting back to a
signed integer is actually used by pulseaudio. The function
is snd_pcm_write_areas(), but pulseaudio doesn't call that at least
directly, and I did some searching in alsa-lib too, and I didn't find a
call path that would cause snd_pcm_write_areas() to be used by
pulseaudio. Even if snd_pcm_write_areas() isn't used, though, it's
entirely possible that there's some other code that does a similar
cast. I don't know the code is that triggers the snd_pcm_start() call
when the ring buffer fill level exceeds the configured threshold. It
might be in the kernel.

> OK, this it why I measure an "extra latency" of -60 to -20 usec.
> So again, if I can measure it and even detect a bug that way,
> don't you think there must be some truth in what I'm saying?

Do I understand correctly that your "extra latency" is affected by
whether snd_pcm_start() is called implicitly in mmap_write() or
explicitly after mmap_write()? The time when mmap_write() is called
doesn't affect the latency in the long term. The smoother will produce
wrong values if it's not started at the same time as snd_pcm_start() is
called, but I presume the smoother is able to fix such inaccuracies
over time, so it doesn't matter that much when the snd_pcm_start() is
called. So isn't it a bad thing if your "extra latency" permanently
includes something that doesn't have any real effect after some time?

> > Have you investigated why there are so many memblocks active that the
> > pool gets full? It seems fishy to me. The limit could be triggered e.g.
> > by thousand 1 ms blocks in a 1 second buffer, but why would you do
> > that?
> 
> I have no idea at all. It always happens when you run a sink at
> very low latencies and I have observed this also with the code
> from git. It is not something that has been introduced by my
> changes, so I did not investigate further and just accepted that
> there seems to be a lower limit to the sink latency at around 2.3ms
> for HDA. Now that USB devices also use timer based scheduling
> I am seeing it there as well at around 8ms configured latency.

The message might indicate some bug somewhere, but I think it's
sufficient if it's printed just once. If it's printed once, it's likely
to be printed many more times, and those additional messages don't
really add any value. So you could add some flag that allows us to
print the message only once.

-- 
Tanu