gstreamer/docs/design/draft-latency.txt

Latency
-------

The latency is the time it takes for a sample captured at timestamp 0 to reach the
sink. This time is measured against the clock in the pipeline. For pipelines
where the only elements that synchronize against the clock are the sinks, the
latency is always 0 since no other element is delaying the buffer.

For pipelines with live sources, a latency is introduced, mostly because of the
way a live source works. Consider an audio source, it will start capturing the
first sample at time 0. If the source pushes buffers with 44100 samples at a
time at 44100Hz it will have collected the buffer at second 1.
Since the timestamp of the buffer is 0 and the time of the clock is now >= 1
second, the sink will drop this buffer because it is too late.
Without an latency compensation in the sink, all buffers will be dropped.

The situation becomes more complex in the presence of:

 - 2 live sources connected to 2 live sinks with different latencies
    * audio/video capture with synchronized live preview.
    * added latencies due to effects (delays, resamplers...)
 - 1 live source connected to 2 live sinks
    * firewire DV
    * RTP, with added latencies because of jitter buffers.
 - mixed live source and non-live source scenarios.
    * synchronized audio capture with non-live playback. (overdubs,..)
 - clock slaving in the sinks due to the live sources providing their own
   clocks.

To perform the needed latency corrections in the above scenarios, we must
develop an algorithm to calculate a global latency for the pipeline. The
algorithm must be extensible so that it can optimize the latency at runtime. 
It must also be possible to disable or tune the algorithm based on specific
application needs (required minimal latency).


Pipelines without latency compensation
--------------------------------------

We show some examples to demonstrate the problem of latency in typical
capture pipelines.

- Example 1

  An audio capture/playback pipeline.

   asrc: audio source, provides a clock
   asink audio sink, provides a clock

   .--------------------------.
   | pipeline                 |
   | .------.      .-------.  |
   | | asrc |      | asink |  |
   | |     src -> sink     |  |
   | '------'      '-------'  |
   '--------------------------'

 NULL->READY:
  asink: NULL->READY: probes device, returns SUCCESS 
  asrc: NULL->READY:  probes device, returns SUCCESS

 READY->PAUSED:
  asink: READY:->PAUSED open device, returns ASYNC 
  asrc: READY->PAUSED:  open device, returns NO_PREROLL

  * Since the source is a live source, it will only produce data in the
    PLAYING state. To note this fact, it returns NO_PREROLL from the state change
    function.
  * This sink returns ASYNC because it can only complete the state change to
    PAUSED when it receives the first buffer.

  At this point the pipeline is not processing data and the clock is not
  running. Unless a new action is performed on the pipeline, this situation will
  never change.

 PAUSED->PLAYING:
  asrc clock selected because it is the most upstream clock provider.
  asink: PAUSED:->PLAYING, returns ASYNC because it is not prerolled 
  asrc: PAUSED->PLAYING: sets pending state to PLAYING, returns ASYNC. The sink
                         will commit state to PLAYING when it prerolls.

  * since the sink is still performing a state change from READY -> PAUSED, it
    remains ASYNC. The pending state will be set to PLAYING.
  * The clock starts running as soon as all the elements have been set to
    PLAYING.
  * the source is a live source with a latency. Since it is synchronized with
    the clock, it will produce a buffer with timestamp 0 and duration D after
    time D, ie. it will only be able to produce the last sample of the buffer
    (with timestamp D) at time D. This latency depends on the size of the
    buffer.
  * the sink will receive the buffer with timestamp 0 at time >= D. At this
    point the buffer is too late already and might be dropped. This state of
    constantly dropping data will not change unless a constant latency
    correction is added to the incomming buffer timestamps.

  The problem is due to the fact that the sink is set to (pending) PLAYING
  without being prerolled, which only happens in live pipelines. 

- Example 2

  An audio/video capture/playback pipeline. We capture both audio and video and
  have them played back synchronized again.

   asrc: audio source, provides a clock
   asink audio sink, provides a clock
   vsrc: video source
   vsink video sink

   .--------------------------.
   | pipeline                 |
   | .------.      .-------.  |
   | | asrc |      | asink |  |
   | |     src -> sink     |  |
   | '------'      '-------'  |
   | .------.      .-------.  |
   | | vsrc |      | vsink |  |
   | |     src -> sink     |  |
   | '------'      '-------'  |
   '--------------------------'

  The state changes happen in the same way as example 1. Both sinks end up with
  pending state of PLAYING and a return value of ASYNC until they receive the
  first buffer.

  For audio and video to be played in sync, both sinks must compensate for the
  latency of its source but must also use exactly the same latency correction.  

  Suppose asrc has a latency of 20ms and vsrc a latency of 33ms, the total
  latency in the pipeline has to be at least 33ms. This also means that the
  pipeline must have at least a 33 - 20 = 13ms buffering on the audio stream or
  else the audio src will underrun while the audiosink waits for the previous
  sample to play.

- Example 3

  An example of the combination of a non-live (file) and a live source (vsrc) 
  connected to live sinks (vsink, sink). 

   .--------------------------.
   | pipeline                 |
   | .------.      .-------.  |
   | | file |      | sink  |  |
   | |     src -> sink     |  |
   | '------'      '-------'  |
   | .------.      .-------.  |
   | | vsrc |      | vsink |  |
   | |     src -> sink     |  |
   | '------'      '-------'  |
   '--------------------------'

  The state changes happen in the same way as example 1. Except sink will be
  able to preroll (commit its state to PAUSED).

  In this case sink will have no latency but vsink will. The total latency
  should be that of vsink.

  Note that because of the presence of a live source (vsrc), the pipeline can be
  set to playing before sink is able to preroll. Without compensation for the
  live source, this might lead to synchronisation problems because the latency
  should be configured in the element before it can go to PLAYING.


- Example 4

  An example of the combination of a non-live and a live source. The non-live
  source is connected to a live sink and the live source to a non-live sink.

   .--------------------------.
   | pipeline                 |
   | .------.      .-------.  |
   | | file |      | sink  |  |
   | |     src -> sink     |  |
   | '------'      '-------'  |
   | .------.      .-------.  |
   | | vsrc |      | files |  |
   | |     src -> sink     |  |
   | '------'      '-------'  |
   '--------------------------'

  The state changes happen in the same way as example 3. Sink will be
  able to preroll (commit its state to PAUSED). files will not be able to
  preroll.

  sink will have no latency since it is not connected to a live source. files
  does not do synchronisation so it does not care about latency.

  The total latency in the pipeline is 0. The vsrc captures in sync with the
  playback in sink.

  As in example 3, sink can only be set to PLAYING after it successfully
  prerolled.


State Changes revised
---------------------

As a first step in a generic solution we propose to modify the state changes so
that no sink is set to PLAYING before it is prerolled.

In order to do this, the pipeline (at the GstBin level) keeps track of all
elements that require preroll (the ones that return ASYNC from the state
change). It keeps a custom GstBinState GST_MESSAGE_ELEMENT internally for
those elements.

When the pipeline did not receive a NO_PREROLL state change return from any
element, it can forget about the GstBinState messages because the state change
to PLAYING will proceed when all elements commited their state when they are
prerolled automatically.

When the pipeline received a NO_PREROLL state change return from an element, it
keeps the ELEMENT messages. The pipeline will not change the state of the
elements that are still doing an ASYNC state change.

When an ASYNC element prerolls, it commits its state to PAUSED and posts a
PREROLLED message. The pipeline notices this PREROLLED message and matches it
with the GstBinState message it cached for the corresponding element. 

When all GstBinState messages are matched with a PREROLLED message, the
pipeline proceeds with setting the PREROLLED sinks to their pending state.
The base time of the element was already set by the pipeline when it changed the
nopreroll element to PLAYING. This operation has to be performed in the
separate async state change thread (like the one currently used for going from
PAUSED->PLAYING in a non-live pipeline).

implications:

 - the current async_play vmethod in basesink can be deprecated since we now
   always call the state change function when going from PAUSED->PLAYING


Latency compensation
--------------------

As an extension to the revised state changes we can perform latency calculation
and compensation before we proceed to the PLAYING state. 

To the PREROLLED message posted by the sinks when then go to PAUSED we add the
following fields:

  - (boolean) live 
  - (boolean) upstream-live 
  - (int_range) latency  (min and max latency in microseconds, could also be 
                          expressed as int_list or min/max fields)

When the pipeline collected all PREROLLED messages it can calculate the global
latency as follows:

  - if no message has live, latency = 0 (no sink syncs against the clock)
  - if no message has upstream-live, latency = 0 (no live source)

  - latency = MAX (MIN (all latencies)) 
  - if MIN (MAX (all latencies) < latency we have an impossible situation.

The sinks gather this information with a LATENCY query upstream. Intermediate
elements pass the query upstream and add the amount of latency they add to the
result.


ex1:
  sink1: [20 - 20]
  sink2: [33 - 40]

  MAX (20, 33) = 33
  MIN (20, 40) = 20 < 33 -> impossible

ex2:
  sink1: [20 - 50]
  sink2: [33 - 40]

  MAX (20, 33) = 33
  MIN (50, 40) = 40 >= 33 -> latency = 33

The latency is set on the pipeline by sending a SET_LATENCY event to the sinks
that posted the PREROLLED message. This event configures the total latency on
the sinks. The sink forwards this SET_LATENCY event upstream so that
intermediate elements can configure themselves as well.

After this step, the pipeline continues setting the pending state on the sinks.

A sink adds the latency value, received in the SET_LATENCY event, to
the times used for synchronizing against the clock. This will effectively
delay the rendering of the buffer with the required latency.


Flushing a playing pipeline
---------------------------

Using the new state change mechanism we can implement resynchronisation after an
uncontrolled FLUSH in (part of) a pipeline. Indeed, when a flush is performed on
a PLAYING live element, a new base time must be distributed to this element.

A flush in a pipeline can happen in the following cases:

 - flushing seek in the pipeline
    - performed by the application on the pipeline
    - performed by the application on an element
 - flush preformed by an element
    - after receiving a navigation event (DVD, ...)

When a playing sink is flushed by a FLUSH_START event, a LOST_PREROLL message is
posted by the element. As part of the message, the state of the element is
included. The element also goes to a pending PAUSED state.

The message is kept in a GstBinState message by the parent bin. When the element
prerolls, it posts a PREROLLED message. 

When all LOST_PREROLL messages are matched with a PREROLLED message, the bin
will capture a new base time from the clock and will bring all the prerolled
sinks back to playing (their pending state) after setting the new base time on
them. It's also possible to perform additional latency calculations and
adjustments before doing this. 

The difference with the NEED_PREROLL/PREROLLED and LOST_PREROLL/PREROLLED
message pair is that the latter makes the pipeline acquire a new base time for
the PREROLLED elements.