lumiera_/doc/design/architecture/playRender.txt

Design: Playback and Render Control
===================================
:Author: Ichthyostega
:Date: 5/2011

//MENU: label Player subsystem

**************************************************************************************
This part of the architecture is concerned with how to initiate, coordinate and control
the calculation processes necessary to generate visible / audible data from our source
media. This design is _not so much_ concerned with the actual calculations; indeed it
takes a ``birds eye view'', treating the actual render engine mostly just like an
opaque entity providing some service.
**************************************************************************************

Collecting initial drafts
-------------------------

At the start of 2011, several design sketches emerged as an offspring of other
discussions and brainstorming sessions.

 * the link:/documentation/devel/rfc_pending/EngineInterfaceOverview.html[Engine Interface draft]
   contains some fundamental terms and definitions.
 * discussions regarding time values and timecode handling shed some light on the requirements
   to be fulfilled by a player subsystem
   
   - link:http://article.gmane.org/gmane.comp.video.lumiera.general/2116[Timecode Widget and Framerates]
   - link:http://article.gmane.org/gmane.comp.video.lumiera.general/2109[Time, Time-spans, Quatisation and Timecode]
   - link:{l}/wiki/renderengine.html#TimeQuant[Definition of Time entities]
     from the implementation notes in the TiddlyWiki


Reasoning Behind the Player Subsystem's Structure
------------------------------------------------

The name ``Player'' might seem surprising at first. 
The full, complete and precise name would be something along
the lines of ``_render- and playback-process coordination subsystem_''.
However we need to shorten that into a single word. The more obvious
abbreviation (assuming importance of words to be the selection criteria) would be
_Render_, but that would be strikingly misleading, because everyone would take
that to refer to the render engine. Thus, we are left with _Player_. A second
consideration highlights similarities to the structure of a
typical software player -- the term we selected turns out to be suitably
aligned to the actual nature of that subsystem.

Influences and Requirements
~~~~~~~~~~~~~~~~~~~~~~~~~~~

In accordance with the spirit of modern software development, the analysis
starts by determining the _Service_ that is provided by such a
player. Examination of the Use Cases determines the fundamental forces 
to be _multiplicity_, combined with _differentiation in detail_, all under
the government of _time-bound delivery_, combined with _live reconfiguration_.

The basic service turns out to be *performing of a preconfigured object model*.
This performance is *time based*. Multiple usage instances of this service can
be expected to coexist, each of which can be broken down into a set of *elementary
streams to be delivered in sync*. The parameters of delivery can be *reconfigured
on-the-fly*.

Modes of Operation
^^^^^^^^^^^^^^^^^^
Delivery can be either _free-wheeling_, so as to cover a predefined time interval
with fixed quality calculations (final render), and it can be _throttled_ for using
excess computation power (background rendering). Or -- alternatively -- delivery can
be _time-bound_ (for classical playback), following the projection of wall clock time
onto a predefined timeline interval (the so called ``playhead'' or ``playback cursor''
proceeding along a timeline).

Reconfiguration
^^^^^^^^^^^^^^^
Some of these operation modes need to be prepared to encounter an unpredictable live
reconfiguration driven by user interactions:

- any part of background rendering can be invalidated and restarted, while other parts
  should be re-integrated, possibly with a re-adjusted position
- playback, which evolves in a linear manner, can be in one of any of these
  states at any particular time :

  * paused
  * reversed in direction
  * playback speed re-adjusted

- playback can be looped, with _unlimited_ adjustments of the loop boundaries at
  any time.



Conclusions for the Design
~~~~~~~~~~~~~~~~~~~~~~~~~~

Based on these observations, the following design would appear to be pretty much
obvious: 

The overall player subsystem can be described as ``play/render-this''-service.
Given a suitable (high-level) object, the player has the ability to ``perform''
(play or render) the object.

- the standard operation is _playing a timeline_.
- it is conceivable that playing a selection of other objects would be
  permitted, for example, directly playing a clip or even a media asset. In
  these cases, it is the player's job to prepare the scaffolding that
  is required on the fly.

Yet each such performance of an object is a _stateful instance_, a player application:
on application of the player service, the client obtains a front-end handle, a *play-controller*,
which is a _state machine_. This provides states and transitions such as 'play', 'pause', 'ffwd',
'rew', 'goto', 'step', 'scrub' and similar.  Moreover, it maintains (or connects to) a distinct
playback location, and it can be hooked up to a play-control GUI widget
(or something simpler in the case of a render process, which is free wheeling).

Each play-controller, in turn, is then associated with several *play/render-processes*,
one for each independent media stream (channel) to be produced. Of course this
isn't an operating system process; rather, each such process is a compound of entries
in a registration table, which serves the purpose of tying several other services together,
which we initiate and use in order to make that render process happen.
Most notably, we'll use the services of the actual engine which provides us with
a kind of a *calculation stream service*: the ability to deliver a sequence of
calculated data frames in a timely fashion.

When a client requests such an instance of the player service, we build up these
parts providing that service, which cascades down to the individual elements.
At that point, we need to pull and combine two kinds of information:

- the ``what'' to render: this information stems from the session/model.
- the ``how'' to render: this information is guided by the derived output configuration.

Viewer and Output Connection
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Creating a player instance binds three partners together: a _timeline_, a _viewer_
and _the engine_. While the timeline provides the contents to play, the _viewer connection_
is crucial for working out the actual output sink(s) and, hence, the output format to use.
Thus, a viewer connection is a prerequisite in order to create a player instance.

Viewer connections exist as associations in the session/model -- which are entities separate
from the player. Usually, a timeline has (at least) one viewer connection. But in case
such a connection is (still) missing, building a player instance recurs to the session
to get a suitable viewer _allocated_. The viewer connection can't be broken during the
lifetime of that player instance (or putting it the other way: breaking that viewer
connection, e.g. by forcing a different connection or by shutting down the viewer,
immediately terminates the player.) This detaching works synchronously, i.e. it
blocks until all the allocated _output slots_ can be released.

Live switching
^^^^^^^^^^^^^^
While the viewer connection can be treated as fixed during the lifespan of a player
instance, several life switching and reconfiguration operations might happen at any time:
the _model port_ (the place where data is retrieved from calculation), the output characteristics
(framerate, direction) and the delivery goals (playback position, loop playing, scrubbing)
all may be changed during playback -- we need a way for the player to ``cancel'' and
reconfigure the vault layer services.

Frame quantisation
^^^^^^^^^^^^^^^^^^
link:../TimeQuant.html[Quantisation] is a kind of rounding; like any kind of rounding, quantisation is
a dangerous operation because it kills information content.

Thus, there are three fundamental guidelines when it comes to rounding:

. don't do it
. do it at most once
. do it as late as possible

These may guide the process of finding the right place for the Quantiser(s) to
apply. We need some information flows to be joined in order to be able to do
the quantisation, which leaves us with just a few possible junction points
where to place quantisation: the Vault, the GUI, the player and the session.

- putting it into the vault layer seems to be the most reasonable at first sight:
  we can ``do away'' with nasty things soon, especially if they are technicalities,
  ``get a clean state soon'' -- and hasn't frame quantisation something to do
  with media data, which is handled in the vault?
+  
Well, actually, all of those are pitfalls to trap the unwary. About
cleanliness, well, sigh! Doing rounding soon will leave us with a huge
amount of degraded information flows throughout the whole system; as a
consequence,  the general rule is to do it as late as possible. Uncrippled
information is an enablement. And last but not least: the frame quantisation is connected
to the _output_ format -- and the vault layer is likely within the whole
application the subsystem most remote and unaware of output requirements.

- rounding/quantising in the GUI is extremely common within media applications;
  unfortunately there does not appear to be a single rational argument in
  support of this habit. Most of all, it is a violation of the subsidiarity
  principle. 

Which leaves us with the player and the session. Both positions could
arguably be supported. Here, a more careful consideration shows, that
the ``act of frame rounding'' can be decomposed: into the _act of quantisation_
and the _frame grid_. Basically it is the session which has the ability
to form the *frame grid*, but it is lacking crucial information about
the output. Only when connecting both -- which is the essence of the
player -- frame quantisation can actually be performed. Thus, the
player is the natural location to perform that quantisation operation.