gstreamer/docs/slides/outline

145 lines
6.5 KiB
Text
Raw Normal View History

Introduction (1)
(sorry, no cool logo/graphic yet, ideas?)
GStreamer is a library and set of tools to build arbitrary,
reconfigurable filter graphs. It derives from the OGI Pipeline
and DirectShow (docs, no experience), and is in its second
generation (first was completed/abandonded *in* Atlanta on the
way to the Linux Expo).
0.1.0 release is scheduled for Oct 31, 1999.
Will cover Background, Goals, Design, and Futures
Why do we need this?
launch reads the command line to create the graph, from .so's
Connections (queues) are made by launcher, lots of switchout code
Argument handling is messy, at start-time only
...thus there is basically only one state: running
There is no real master process capable of seeing the whole
pipeline as a graph, so wiring A->B->C with some backchannel
(parameter, not data stream) from C to A is hard
Channels try to do IPC, file, and network I/O, excess abstraction
Goals (1)
Provide a clean way to both build graphs and write new elements
Make things easier by providing auto-connect, stock sub-graphs
Include tools sorely lacking in OGI pipeline, like editor, saves
Enable Linux to catch up with M$'s world, allowing commercial
plugins to a stable API (as of 1.0) so we don't end up with N
wheels from N-M different people (multiple projects)
Overview (1)
Object hierarchy capable of run-time discovery, based on GtkObject
Deeply nested parent-child relationships enable threads, blackboxes
Buffers can point to anything, are typed, and can carry metadata
Plugins can be loaded at any point, and registry reduces loads
Symbiotic editor lets you design/run/save graphs visually
What are filter graphs? (1)
Filters take data in and spit data out, doing something to it
Filters have N>=0 inputs and M>=0 outputs
Filter graphs are many filters connected together, like a circuit
The goal is typically to move data from 'left' to 'right', towards
some kind of user-visible conclusion, i.e. audio or video
Architecture (3?)
- Graphs of Elements
(here lies screen-grab from editor)
Element is core Object, Bins hold (and are) Elements
Pads are fundamental to an Element, are cross-wired with pointers
Since Bins hold Elements, and Bins are Elements, Bins hold Bins
'Ghostpads' provide interfaces for Bins without native interfaces
# Threads are type of Bin that actually run in separate threads
- States
(table of states, invariants, and descriptions)
COMPLETE Element has all needed information
RUNNING Element has acquired resources, ready to go
DISCOVERY ... (unimplemented)
PREROLL ... (unimplemented)
PLAYING Element is actively trading data
PAUSED Special state where things just don't run (?..)
States are used to keep elements in check
- Buffers
Buffers designed to be versatile, with arbitrary typing/metadata
Has pointer to data, length, so can point to someone else's data
Type system (crude so far) ensures buffers don't go stray
Metadata can be attached, such as the audio parameters
Ref-counting and copy-on-write avoids most copies, not complete
Sub-buffers can be created from bigger buffer, limitting copies
Gtk+ Object System (2)
- Pros
C-language object system, well tested (Gtk+, Gnome...)
Arguments of any fundamental type, read/write, built-in hooks
Signals used for hooks into object events, overridable
Run-time discovery of args, signals (quarks)
- Cons
No multiple-inheritance, though I haven't *needed* it
There are some holes (can't attach quarks to *eveything*)
- Design
Classes, instances are structs; 1st item is parent class
Type system allows clean casting, ^^^^^^
Arguments are set/get by string, use functions to do the work,
thus setting an arg might trigger a redraw of the GUI
Signals are strings, use marshallers, various firing methods
Basic GStreamer objects (1)
- Elements
(show class,instance structs)
Very simple, just provides a means to handle pads, state
- Bins
(show class,instance structs)
Supports children, handles group state transitions
Pads (1)
Pad list type, direction, and chaining function ptr
When creating a sink pad (!src) you set the chaining function
gst_pad_connect() sets the peers, and copies chain function to src
Passing buffer to a src pad transparently calls the chain function
(graph goes here...)
Sources (1)
Source provides functions to push data
Regular push() function just takes next N bytes and sends them
Async push_region() grabs N bytes at offset O and sends them
EOF signal [will] reset the state from PLAYING down to !RUNNING
"location" argument is global by convention, for filenames...URIs
Connections (1)
Special type of Filter that
Threads (1)
Special case of Bin that actually creates a thread transparently
When RUNNING, thread exists, mutex/cond used to go [!]PLAYING
Automatically determines how to start sub-graph
Looks for both Sources and Elements wired to Connection
Will cooperate with Pipelines when threading is not available
Typing and Metadata (1)
- Types
Based on MIME types, set up as quarks, and dealt with as int
Usable entirely at run-time, since they're registerable by plugins
- Metadata
Also registered as an int, but must be compile time due to structs
Have refcounts and CoW semantics, since they travel with buffers
Plugins (1)
Plugin architecture designed around class system
Arguments and signals provide interface over standard base class
Each Element defined by ElementFactory, which is queried by name
At plugin load, any number of ElementFactories and Types registered
Element registers against Type as source or sink
Editor (2+?)
(show filter graph snapshot, a different one, more complex)
Built as a parallel object hierarchy on top of GNOME Canvas
Every object in filter graph has equivalent in editor, plus some
Canvas is designed with groups and signal-propagation, so...
Why not build the whole thing as subclasses of CanvasGroup?
...because updates get messy/recursive (the way I *was* doing it)
Solution is to modify objects so they own Group rather than being
Relatively trivial modification, but requires lots of repointering
Still a genealogical mess of parents and children...
XML
The goal is to use XML heavily, with an eye towards DOM
Used for both saving and providing pre-build components
Both graph and editor will have namespace, they'll interleave
A generic save function will exist for Elements, with hooks
Saving an EditorElement will also save Element
Also used for a plugin registry, to avoid loading all plugins
leaky bucket is trivial
applications - generic conferencing tool (repluggable codecs), mixing
environment (data flow graphs)