The current minimum target latency is partially based upon the default qemu
behaviour whereby audio packets are delivered in a sawtooth pattern, with
packet timestamps drifting between 5ms above and below the measured clock.
This 5ms error is baked into the minimum target latency to avoid
underrunning.
This sawtooth pattern can be reduced by specifying a lower timer period in
the qemu configuration, so remove it from the hardcoded minimum latency and
add it to the default configurable buffer latency instead. This allows
users that have configured their VM appropriately to reduce the overall
latency.
The best quality resampler has an intrinsic latency of about 3ms, and the
processing itself takes another 1-2ms per 10ms block. The faster setting
has an intrinsic latency of about 0.4ms, with about 0.04ms processing time.
This makes for an overall saving of about 4ms, with negligible loss in
quality.
This adds a new `audio:bufferLatency` option which allows the user to
adjust the amount of buffering LG does over the absolute bare minimum. By
default, this is set large enough to absorb typical timing jitter from
Spice. Users may reduce this if they care more about latency than audio
quality.
This adds a new `audio:periodSize` option which defaults to 2048 frames.
For PipeWire, this controls the `PIPEWIRE_LATENCY` value. For PulseAudio,
the controls the target buffer length (`tlength`) value.
If the audio device starts earlier than required, we slew the read pointer
backwards to avoid underrunning. We need to apply this same offset to the
recorded device position, otherwise the Spice thread will think playback is
further ahead than it really is and inject unnecessary latency to
compensate.
The actual time between opening the device and the device starting to pull
data can range anywhere between nearly instant and hundreds of
milliseconds. To minimise startup latency, open the device as soon as the
first playback data is received from Spice. If the device starts earlier
than required, insert a period of silence at the beginning of playback to
avoid underrunning. If it starts later, just accept the higher latency and
let the adaptive resampling deal with it.
We can set the startup latency for the next playback far more precisely if
we have the device open already.
Only keep the device open with no playback for 30 seconds to avoid keeping
the device open unnecessarily forever.
Underruns can still happen quite easily at the beginning of playback,
particularly at very low latency settings. Further increase the startup
latency to avoid this.
The target latency is now based upon the device maximum period size
(which may be configured by setting the `PIPEWIRE_LATENCY` environment
variable if using PipeWire), with some allowance for timing jitter from
Spice and the audio device.
PipeWire can change the period size dynamically at any time which must be
taken into account when selecting the target latency to avoid underruns
when the period size is increased. This is explained in detail within the
commit body.
This removes the need for locking while also giving a better result in
the graph output. Also when the graph is disabled via the overlay
options it will no longer cause redraws.
This change allows the audiodevs to return the minimum period frames
needed to start playback instead of having to rely on a pull to obtain
these details.
Additionally we are using this information to select an initial start
latency as well as to train the desired latency in order to keep it as
low as possible.
This change is based on the techniques described in [1] and [2].
The input audio stream from Spice is not synchronised to the audio playback
device. While the input and output may be both nominally running at 48 kHz,
when compared against each other, they will differ by a tiny fraction of a
percent. Given enough time (typically on the order of a few hours), this
will result in the ring buffer becoming completely full or completely
empty. It will stay in this state permanently, periodically resulting in
glitches as the buffer repeatedly underruns or overruns.
To address this, adjust the speed of the received data to match the rate at
which it is being consumed by the audio device. This will result in a
slight pitch shift, but the changes should be small and smooth enough that
this is unnoticeable to the user.
The process works roughly as follows:
1. Every time audio data is received from Spice, or consumed by the audio
device, sample the current time. These are fed into a pair of delay
locked loops to produce smoothed approximations of the two clocks.
2. Compute the difference between the two clocks and compare this against
the target latency to produce an error value. This error value will be
quite stable during normal operation, but can change quite rapidly due
to external factors, particularly at the start of playback. To smooth
out any sudden changes in playback speed, which would be noticeable to
the user, this value is also filtered through another delay locked loop.
3. Feed this error value into a PI controller to produce a ratio value.
This is the target playback speed in order to bring the error value
towards zero.
4. Resample the input audio using the computed ratio to apply the speed
change. The output of the resampler is what is ultimately inserted into
the ring buffer for consumption by the audio device.
Since this process targets a specific latency value, rather than simply
trying to rate match the input and output, it also has the effect of
'correcting' latency issues. If a high latency application (such as a media
player) is already running, the time between requesting the start of
playback and the audio device actually starting to consume samples can be
very high, easily in the hundreds of milliseconds. The changes here will
automatically adjust the playback speed over the course of a few minutes to
bring the latency back down to the target value.
[1] https://kokkinizita.linuxaudio.org/papers/adapt-resamp.pdf
[2] https://kokkinizita.linuxaudio.org/papers/usingdll.pdf
In unbounded mode, the read and write pointers are free to move
independently of one another. This is useful where the input and output
streams are progressing at the same rate on average, and we want to keep
the latency stable in the event than an underrun or overrun occurs.
If an underrun occurs (i.e., there is not enough data in the buffer to
satisfy a read request), the missing values with be filled with zeros. When
the writer catches up, the same number of values will be skipped from the
input.
If an overrun occurs (i.e., there is not enough free space in the buffer to
satisfy a write request), excess values will be discarded. When the reader
catches up, the same number of values will be zeroed in the output.
Unbounded mode is currently unused since our audio input and output
streams are not synchronised. This will be implemented in a later commit.
Also reimplemented as a lock-free queue which is safer for use in audio
device callbacks.
If a new playback is started while the previous playback is still flushing,
we simply allow the stream to continue playing and effectively cancel the
flush. In general this is not safe because there may not be enough data in
the buffer to avoid underrunning. We could handle this better later by
trying to insert the right number of silent samples into the buffer, but
for now just completely stop the previous stream before starting the new
one.
