[mlt] / docs / framework.txt

Framework Documentation

Copyright (C) 2004 Ushodaya Enterprises Limited
Author: Charles Yates <charles.yates@pandora.be>
Last Revision: 2004-03-20


MLT FRAMEWORK
-------------

Preamble:

        MLT is a multimedia framework designed for television broadcasting. As such, 
        it provides a pluggable architecture for the inclusion of new audio/video 
        sources, filters, transitions and playback devices.

        The framework provides the structure and utility functionality on which
        all of the MLT applications and services are defined. 

        On its own, the framework provides little more than 'abstract classes' and
        utilities for managing resources, such as memory, properties, dynamic object
        loading and service instantiation. 

        This document is split roughly into 3 sections. The first section provides a
        basic overview of MLT, the second section shows how it's used and the final
        section shows structure and design, with an emphasis on how the system is 
        extended.


Target Audience:

        This document is provided as a 'road map' for the framework and should be
        considered mandatory reading for anyone wishing to develop code at the MLT
        level. 

        This includes:

        1. framework maintainers;
        2. module developers;
        3. application developers;
        4. anyone interested in MLT.

        The emphasis of the document is in explaining the public interfaces, as
        opposed to the implementation details.

        It is not required reading for the MLT client/server integration - please
        refer to valerie.txt and dvcp.txt for more details on this area.


SECTION 1 - BASIC OVERVIEW
--------------------------

Basic Design Information:

        MLT is written in C. 
        
        The framework has no dependencies other than the standard C99 and POSIX 
        libraries. 

        It follows a basic Object Oriented design paradigm, and as such, much of the
        design is loosely based on the Producer/Consumer design pattern. 
        
        It employs Reverse Polish Notation for the application of audio and video FX.

        The framework is designed to be colour space neutral - the currently
        implemented modules, however, are very much 8bit YUV422 oriented. In theory,
        the modules could be entirely replaced.

        A vague understanding of these terms is assumed throughout the remainder of
        this document.


Structure and Flow:

        The general structure of an MLT 'network' is simply the connection of a
        'producer' to a 'consumer':

        +--------+   +--------+
        |Producer|-->|Consumer|
        +--------+   +--------+

        A typical consumer requests MLT Frame objects from the producer, does 
        something with them and when finished with a frame, closes it. 
        
         /\  A common confusion with the producer/consumer terminology used here is 
        /!!\ that a consumer may 'produce' something. For example, the libdv consumer
        \!!/ produces DV and the libdv producer seems to consume DV. However, the
         \/  naming conventions refer only to producers and consumers of MLT Frames.

        To put it another way - a producer produces MLT Frame objects and a consumer 
        consumes MLT Frame objects.

        An MLT Frame essentially provides an uncompressed image and its associated
        audio samples.

        Filters may also be placed between the producer and the consumer:

        +--------+   +------+   +--------+
        |Producer|-->|Filter|-->|Consumer|
        +--------+   +------+   +--------+

        A service is the collective name for producers, filters, transitions and
        consumers. 

        The communications between a connected consumer and producer or service are 
        carried out in 3 phases:

        * get the frame
        * get the image
        * get the audio

        MLT employs 'lazy evaluation' - the image and audio need not be extracted
        from the source until the get image and audio methods are invoked. 

        In essence, the consumer pulls from what it's connected to - this means that
        threading is typically in the domain of the consumer implementation and some
        basic functionality is provided on the consumer class to ensure realtime
        throughput.


SECTION 2 - USAGE
-----------------

Hello World:

        Before we go in to the specifics of the framework architecture, a working
        example of usage is provided. 
        
        The following simply provides a media player:

        #include <stdio.h>
        #include <unistd.h>
        #include <framework/mlt.h>

        int main( int argc, char *argv[] )
        {
            // Initialise the factory
            if ( mlt_factory_init( NULL ) == 0 )
            {
                // Create the default consumer
                mlt_consumer hello = mlt_factory_consumer( NULL, NULL );

                // Create via the default producer
                mlt_producer world = mlt_factory_producer( NULL, argv[ 1 ] );

                // Connect the producer to the consumer
                mlt_consumer_connect( hello, mlt_producer_service( world ) );

                // Start the consumer
                mlt_consumer_start( hello );

                // Wait for the consumer to terminate
                while( !mlt_consumer_is_stopped( hello ) )
                    sleep( 1 );

                // Close the consumer
                mlt_consumer_close( hello );

                // Close the producer
                mlt_producer_close( world );

                // Close the factory
                mlt_factory_close( );
            }
            else
            {
                // Report an error during initialisation
                fprintf( stderr, "Unable to locate factory modules\n" );
            }

            // End of program
            return 0;
        }

        This is a simple example - it doesn't provide any seeking capabilities or
        runtime configuration options. 

        The first step of any MLT application is the factory initialisation - this
        ensures that the environment is configured and MLT can function. The factory
        is covered in more detail below.

        All services are instantiated via the factories, as shown by the
        mlt_factory_consumer and mlt_factory_producer calls above. There are similar
        factories for filters and transitions. There are details on all the standard
        services in services.txt. 

        The defaults requested here are a special case - the NULL usage requests
        that we use the default producers and consumers. 
        
        The default producer is "fezzik". This producer matches file names to 
        locate a service to use and attaches 'normalising filters' (such as scalers,
        deinterlacers, resamplers and field normalisers) to the loaded content -
        these filters ensure that the consumer gets what it asks for.

        The default consumer is "sdl". The combination of fezzik and sdl will
        provide a media player.

        In this example, we connect the producer and then start the consumer. We
        then wait until the consumer is stopped (in this case, by the action of the
        user closing the SDL window) and finally close the consumer, producer and
        factory before exiting the application.

        Note that the consumer is threaded - waiting for an event of some sort is 
        always required after starting and before stopping or closing the consumer.

        Also note, you can override the defaults as follows:

        $ MLT_CONSUMER=westley ./hello file.avi

        This will create a westley xml document on stdout.

        $ MLT_CONSUMER=westley MLT_PRODUCER=avformat ./hello file.avi

        This will play the video using the avformat producer directly, thus it will
        bypass the normalising functions.

        $ MLT_CONSUMER=libdv ./hello file.avi > /dev/dv1394

        This might, if you're lucky, do on the fly, realtime conversions of file.avi
        to DV and broadcast it to your DV device.


Factories:

        As shown in the 'Hello World' example, factories create service objects.

        The framework itself provides no services - they are provided in the form of
        a plugin structure. A plugin is organised in the form of a 'module' and a
        module can provide many services of different types. 

        Once the factory is initialised, all the configured services are available
        for use.

        The complete set of methods associated to the factory are as follows:

        int mlt_factory_init( char *prefix );
        const char *mlt_factory_prefix( );
        char *mlt_environment( char *name );
        mlt_producer mlt_factory_producer( char *name, void *input );
        mlt_filter mlt_factory_filter( char *name, void *input );
        mlt_transition mlt_factory_transition( char *name, void *input );
        mlt_consumer mlt_factory_consumer( char *name, void *input );
        void mlt_factory_close( );

        The mlt_factory_prefix returns the path to the location of the installed
        modules directory. This can be specified in the mlt_factory_init call
        itself, or it can be specified via the MLT_REPOSITORY environment variable,
        or in the absence of either of those, it will default to the install
        prefix/shared/mlt/modules. 

        The mlt_environment provides read only access to a collection of name=value
        pairs as shown in the following table:

        +------------------+------------------------------------+------------------+
        |Name              |Description                         |Values            |
        +------------------+------------------------------------+------------------+
        |MLT_NORMALISATION |The normalisation of the system     |PAL or NTSC       |
        +------------------+------------------------------------+------------------+
        |MLT_PRODUCER      |The default producer                |"fezzik" or other |
        +------------------+------------------------------------+------------------+
        |MLT_CONSUMER      |The default consumer                |"sdl" or other    |
        +------------------+------------------------------------+------------------+
        |MLT_TEST_CARD     |The default test card producer      |any producer      |
        +------------------+------------------------------------+------------------+

        These values are initialised from the environment variables of the same
        name.

        As shown above, a producer can be created using the 'default normalising'
        producer, and they can also be requested by name. Filters and transitions 
        are always requested by name - there is no concept of a 'default' for these.


Service Properties:

        As shown in the services.txt document, all services have their own set of
        properties than can be manipulated to affect their behaviour.

        In order to set properties on a service, we need to retrieve the properties
        object associated to it. For producers, this is done by invoking:

            mlt_properties properties = mlt_producer_properties( producer );

        All services have a similar method associated to them.

        Once retrieved, setting and getting properties can be done directly on this
        object, for example:

            mlt_properties_set( properties, "name", "value" );
        
        A more complete description of the properties object is found below.


Playlists:

        So far, we've shown a simple producer/consumer configuration - the next
        phase is to organise producers in playlists.

        Let's assume that we're adapting the Hello World example, and wish to queue
        a number of files for playout, ie:

            hello *.avi

        Instead of invoking mlt_factory_producer directly, we'll create a new
        function called create_playlist. This function is responsible for creating
        the playlist, creating each producer and appending to the playlist.

        mlt_producer create_playlist( int argc, char **argv )
        {
            // We're creating a playlist here
            mlt_playlist playlist = mlt_playlist_init( );

            // We need the playlist properties to ensure clean up
            mlt_properties properties = mlt_playlist_properties( playlist );

            // Loop through each of the arguments
            int i = 0;
            for ( i = 1; i < argc; i ++ )
            {
                // Create the producer
                mlt_producer producer = mlt_factory_producer( NULL, argv[ i ] );

                // Add it to the playlist
                mlt_playlist_append( playlist, producer );

                        // Close the producer (see below)
                        mlt_producer_close( producer );
            }

            // Return the playlist as a producer
            return mlt_playlist_producer( playlist );
        }

        Notice that we close the producer after the append. Actually, what we're 
        doing is closing our reference to it - the playlist creates its own reference
        to the producer on append and insert, and it will close its reference 
        when the playlist is destroyed[*].

        Note also that if you append multiple instances of the same producer, it 
        will create multiple references to it.

        Now all we need do is to replace these lines in the main function:

            // Create a normalised producer
            mlt_producer world = mlt_factory_producer( NULL, argv[ 1 ] );

        with:

            // Create a playlist
            mlt_producer world = create_playlist( argc, argv );

        and we have a means to play multiple clips.

        [*] This reference functionality was introduced in mlt 0.1.2 - it is 100%
        compatable with the early mechanism of registering the reference and 
        destructor with the properties of the playlist object.


Filters:

        Inserting filters between the producer and consumer is just a case of
        instantiating the filters, connecting the first to the producer, the next
        to the previous filter and the last filter to the consumer.

        For example:

            // Create a producer from something
            mlt_producer producer = mlt_factory_producer( ... );

            // Create a consumer from something
            mlt_consumer consumer = mlt_factory_consumer( ... );

            // Create a greyscale filter
            mlt_filter filter = mlt_factory_filter( "greyscale", NULL );

            // Connect the filter to the producer
            mlt_filter_connect( filter, mlt_producer_service( producer ), 0 );

            // Connect the consumer to filter
            mlt_consumer_connect( consumer, mlt_filter_service( filter ) );

        As with producers and consumers, filters can be manipulated via their
        properties object - the mlt_filter_properties method can be invoked and
        properties can be set as needed.

        The additional argument in the filter connection is an important one as it
        dictates the 'track' on which the filter operates. For basic producers and
        playlists, there's only one track (0), and as you will see in the next
        section, even multiple tracks have a single track output.


Multiple Tracks and Transitions:

        MLT's approach to multiple tracks is governed by two requirements:

        1) The need for a consumer and producer to communicate with one another via
        a single frame;
        2) The desire to be able to serialise and manipulate a 'network' (or filter
        graph if you prefer).

        We can visualise a multitrack in the way that an NLE presents it:

           +-----------------+                          +-----------------------+
        0: |a1               |                          |a2                     |
           +---------------+-+--------------------------+-+---------------------+
        1:                 |b1                            |
                           +------------------------------+

        The overlapping areas of track 0 and 1 would (presumably) have some kind of
        transition - without a transition, the frames from b1 and b2 would be shown 
        during the areas of overlap (ie: by default, the higher numbered track takes 
        precedence over the lower numbered track). 

        MLT has a multitrack object, but it is not a producer in the sense that it
        can be connected directly to a consumer and everything will work correctly.
        A consumer would treat it precisely as it would a normal producer, and, in
        the case of the multitrack above, you would never see anything from track 1
        other than the transitions between the clips - the gap between a1 and a2 
        would show test frames.

        This happens because a consumer pulls one frame from the producer it's 
        connected to while a multitrack will provide one frame per track.
        Something, somewhere, must ensure that all frames are pulled from the
        multitrack and elect the correct frame to pass on.

        Hence, MLT provides a wrapper for the multitrack, which is called a
        'tractor', and its the tractors task to ensure that all tracks are pulled
        evenly, the correct frame is output and that we have 'producer like'
        behaviour.

        Thus, a multitrack is conceptually 'pulled' by a tractor as shown here:

        +----------+
        |multitrack|
        | +------+ |    +-------+
        | |track0|-|--->|tractor|
        | +------+ |    |\      |
        |          |    | \     |
        | +------+ |    |  \    |
        | |track1|-|--->|---o---|--->
        | +------+ |    |  /    |
        |          |    | /     |
        | +------+ |    |/      |
        | |track2|-|--->|       |
        | +------+ |    +-------+
        +----------+

        With a combination of the two, we can now connect multitracks to consumers.
        The last non-test card will be retrieved and passed on. 

        The tracks can be producers, playlists, or even other tractors. 

        Now we wish to insert filters and transitions between the multitrack and the
        tractor. We can do this directly by inserting filters directly between the
        tractor and the multitrack, but this involves a lot of connecting and
        reconnecting left and right producers and consumers, and it seemed only fair
        that we should be able to automate that process. 

        So in keeping with our agricultural theme, the concept of the 'field' was 
        born. We 'plant' filters and transitions in the field and the tractor pulls 
        the multitrack (think of a combine harvester :-)) over the field and 
        produces a 'bail' (sorry - kidding - frame :-)).

        Conceptually, we can see it like this:

        +----------+
        |multitrack|
        | +------+ |    +-------------+    +-------+
        | |track0|-|--->|field        |--->|tractor|
        | +------+ |    |             |    |\      |
        |          |    |   filters   |    | \     |
        | +------+ |    |     and     |    |  \    |
        | |track1|-|--->| transitions |--->|---o---|--->
        | +------+ |    |             |    |  /    |
        |          |    |             |    | /     |
        | +------+ |    |             |    |/      |
        | |track2|-|--->|             |--->|       |
        | +------+ |    +-------------+    +-------+
        +----------+

        In reality, we create a field first, and from that we obtain a multitrack
        and a tractor. We can then populate the multitrack, field and finally,
        connect the tractor to the consumer. 
        
        The reasoning behind this is possibly flawed - it might have made more 
        sense to produce the tractor and have it encapsulate the field and the
        multitrack as that is how it looks to a connected consumer:

        +-----------------------------------------------+
        |tractor          +--------------------------+  |
        | +----------+    | +-+    +-+    +-+    +-+ |  |
        | |multitrack|    | |f|    |f|    |t|    |t| |  |
        | | +------+ |    | |i|    |i|    |r|    |r| |  |
        | | |track0|-|--->| |l|- ->|l|- ->|a|--->|a|\|  |
        | | +------+ |    | |t|    |t|    |n|    |n| |  |
        | |          |    | |e|    |e|    |s|    |s| |\ |
        | | +------+ |    | |r|    |r|    |i|    |i| | \|
        | | |track1|-|- ->| |0|--->|1|--->|t|--->|t|-|--o--->
        | | +------+ |    | | |    | |    |i|    |i| | /|
        | |          |    | | |    | |    |o|    |o| |/ |
        | | +------+ |    | | |    | |    |n|    |n| |  |
        | | |track2|-|- ->| | |- ->| |--->|0|- ->|1|/|  |
        | | +------+ |    | | |    | |    | |    | | |  |
        | +----------+    | +-+    +-+    +-+    +-+ |  |
        |                 +--------------------------+  |
        +-----------------------------------------------+

        An example will hopefully clarify this. 
        
        Let's assume that we want to provide a 'watermark' to our hello world 
        example. We have already extended the example to play multiple clips,
        and now we will place a text based watermark, reading 'Hello World' in 
        the top left hand corner:

        mlt_producer create_tracks( int argc, char **argv )
        {
            // Create the field
            mlt_field field = mlt_field_init( );
        
            // Obtain the multitrack
            mlt_multitrack multitrack = mlt_field_multitrack( field );
        
            // Obtain the tractor
            mlt_tractor tractor = mlt_field_tractor( field );
        
            // Obtain a composite transition
            mlt_transition transition = mlt_factory_transition( "composite", "10%,10%:15%x15%" );
        
            // Create track 0
            mlt_producer track0 = create_playlist( argc, argv );
        
            // Create the watermark track - note we NEED fezzik for scaling here
            mlt_producer track1 = mlt_factory_producer( "fezzik", "pango" );
        
            // Get the length of track0
            mlt_position length = mlt_producer_get_playtime( track0 );
        
            // Set the properties of track1
            mlt_properties properties = mlt_producer_properties( track1 );
            mlt_properties_set( properties, "text", "Hello\nWorld" );
            mlt_properties_set_position( properties, "in", 0 );
            mlt_properties_set_position( properties, "out", length - 1 );
            mlt_properties_set_position( properties, "length", length );
            mlt_properties_set_int( properties, "a_track", 0 );
            mlt_properties_set_int( properties, "b_track", 1 );

            // Now set the properties on the transition
            properties = mlt_transition_properties( transition );
            mlt_properties_set_position( properties, "in", 0 );
            mlt_properties_set_position( properties, "out", length - 1 );
        
            // Add our tracks to the multitrack
            mlt_multitrack_connect( multitrack, track0, 0 );
            mlt_multitrack_connect( multitrack, track1, 1 );
        
            // Now plant the transition
            mlt_field_plant_transition( field, transition, 0, 1 );
        
            // Now set the properties on the tractor
            properties = mlt_tractor_properties( tractor );
            mlt_properties_set_data( properties, "multitrack", multitrack, 0, ( mlt_destructor )mlt_multitrack_close, NULL );
            mlt_properties_set_data( properties, "field", field, 0, ( mlt_destructor )mlt_field_close, NULL );
            mlt_properties_set_data( properties, "track0", track0, 0, ( mlt_destructor )mlt_producer_close, NULL );
            mlt_properties_set_data( properties, "track1", track1, 0, ( mlt_destructor )mlt_producer_close, NULL );
            mlt_properties_set_data( properties, "transition", transition, 0, ( mlt_destructor )mlt_transition_close, NULL );
        
            // Return the tractor
            return mlt_tractor_producer( tractor );
        }

        Now all we need do is to replace these lines in the main function:

            // Create a playlist
            mlt_producer world = create_playlist( argc, argv );

        with:

            // Create a watermarked playlist
            mlt_producer world = create_tracks( argc, argv );

        and we have a means to play multiple clips with a horribly obtrusive
        watermark - just what the world needed, right? ;-)

        Incidentally, the same thing could be achieved with the more trivial
        watermark filter inserted between the producer and the consumer.


SECTION 3 - STRUCTURE AND DESIGN
--------------------------------

Class Hierarchy:

        The mlt framework consists of an OO class hierarchy which consists of the
        following public classes and abstractions:

        mlt_properties
          mlt_frame
          mlt_service
            mlt_producer
              mlt_playlist
              mlt_tractor
            mlt_filter
            mlt_transition
            mlt_consumer
        mlt_deque
        mlt_pool
        mlt_factory

        Each class defined above can be read as extending the classes above and to
        the left.

        The following sections describe the properties, stacking/queuing and memory 
        pooling functionality provided by the framework - these are key components 
        and a basic understanding of these is required for the remainder of the
        documentation.


mlt_properties:

        The properties class is the base class for the frame and service classes.

        It is designed to provide an efficient lookup table for various types of
        information, such as strings, integers, floating points values and pointers
        to data and data structures. 

        All properties are indexed by a unique string. 
        
        The most basic use of properties is as follows:

            // 1. Create a new, empty properties set;
            mlt_properties properties = mlt_properties_new( );

            // 2. Assign the value "world" to the property "hello";
            mlt_properties_set( properties, "hello", "world" );

            // 3. Retrieve and print the value of "hello";
            printf( "%s\n", mlt_properties_get( properties, "hello" ) );

            // 4. Reassign "hello" to "world!";
            mlt_properties_set( properties, "hello", "world!" );

            // 5. Retrieve and print the value of "hello";
            printf( "%s\n", mlt_properties_get( properties, "hello" ) );

            // 6. Assign the value "0" to "int";
            mlt_properties_set( properties, "int", "0" );

            // 7. Retrieve and print the integer value of "int";
            printf( "%d\n", mlt_properties_get_int( properties, "int" ) );

            // 8. Assign the integer value 50 to "int2";
            mlt_properties_set_int( properties, "int2", 50 );

            // 9. Retrieve and print the double value of "int2";
            printf( "%s\n", mlt_properties_get( properties, "int2" ) );

        Steps 2 through 5 demonstrate that the "name" is unique - set operations on 
        an existing "name" change the value. They also free up memory associated to 
        the previous value. Note that it also possible to change type in this way
        too.

        Steps 6 and 7 demonstrate that the properties object handles deserialisation
        from strings. The string value of "0" is set, the integer value of 0 is
        retrieved.

        Steps 8 and 9 demonstrate that the properties object handles serialisation
        to strings.

        To show all the name/value pairs in a properties, it is possible to iterate
        through them:

            for ( i = 0; i < mlt_properties_count( properties ); i ++ )
                printf( "%s = %s\n", mlt_properties_get_name( properties, i ),
                                     mlt_properties_get_value( properties, i ) );

        Note that properties are retrieved in the order in which they are set. 

        Properties are also used to hold pointers to memory. This is done via the
        set_data call:

        uint8_t *image = malloc( size );
        mlt_properties_set_data( properties, "image", image, size, NULL, NULL );

        In this example, we specify that the pointer can be retrieved from
        properties by a subsequent request to get_data:

            image = mlt_properties_get_data( properties, "image", &size );

        or:

            image = mlt_properties_get_data( properties, "image", NULL );

        if we don't wish to retrieve the size.

        Two points here:

        1) The allocated memory remains after the properties object is closed unless
           you specify a destructor. In the case above, this can be done with:

           mlt_properties_set_data( properties, "image", image, size, free, NULL );

           When the properties are closed, or the value of "image" is changed, the 
           destructor is invoked.
        
        2) The string value returned by mlt_properties_get is NULL. Typically, you
           wouldn't wish to serialise an image as a string, but other structures
           might need such functionality - you can specify a serialiser as the last
           argument if required (declaration is char *serialise( void * )).

        Properties also provides some more advanced usage capabilities. 
        
        It has the ability to inherit all serialisable values from another properties 
        object:

            mlt_properties_inherit( this, that );

        It has the ability to mirror properties set on this on another set of
        properties:

            mlt_properties_mirror( this, that );

        After this call, all serialisable values set on this are passed on to that.


mlt_deque:

        Stacks and queues are essential components in the MLT framework. Being of a
        lazy disposition, we elected to implement a 'Double Ended Queue' (deque) -
        this encapsulates the functionality of both.

        The API of the deque is defined as follows:

        mlt_deque mlt_deque_init( );
        int mlt_deque_count( mlt_deque this );
        int mlt_deque_push_back( mlt_deque this, void *item );
        void *mlt_deque_pop_back( mlt_deque this );
        int mlt_deque_push_front( mlt_deque this, void *item );
        void *mlt_deque_pop_front( mlt_deque this );
        void *mlt_deque_peek_back( mlt_deque this );
        void *mlt_deque_peek_front( mlt_deque this );
        void mlt_deque_close( mlt_deque this );

        The stacking operations are used in a number of places:

        * Reverse Polish Notation (RPN) image and audio operations
        * memory pooling

        The queuing operations are used in:
        
        * the consumer base class;
        * consumer implementations may require further queues.


mlt_pool:

        The MLT framework provides memory pooling capabilities through the mlt_pool
        API. Once initilialised, these can be seen as a straightforward drop in
        replacement for malloc/realloc/free functionality.

        The background behind this API is that malloc/free operations are
        notoriously inefficient, especially when dealing with large blocks of memory
        (such as an image). On linux, malloc is optimised for memory allocations
        less than 128k - memory blocks allocated of these sizes or less are retained
        in the process heap for subsequent reuse, thus bypassing the kernel calls
        for repeated allocation/frees for small blocks of memory. However, blocks of
        memory larger than that require kernel calls and this has a detrimental
        impact on performance.

        The mlt_pool design is simply to hold a list of stacks - there is one stack
        per 2^n bytes (where n is between 8 and 31). When an alloc is called, the
        requested size is rounded to the next 2^n, the stack is retrieved for that
        size, and an item is popped or created if the stack is empty. 

        Each item has a 'header', situated immediately before the returned address - 
        this holds the 'stack' to which the item belongs.

        When an item is released, we retrieve the header, obtain the stack and push
        it back.

        Thus, from the programmers point of view, the API is the same as the
        traditional malloc/realloc/free calls:

        void *mlt_pool_alloc( int size );
        void *mlt_pool_realloc( void *ptr, int size );
        void mlt_pool_release( void *release );


mlt_frame:

        A frame object is essentially defined as:

        +------------+
        |frame       |
        +------------+
        | properties |
        | image stack|
        | audio stack|
        +------------+

        The life cycle of a frame can be represented as follows:

        +-----+----------------------+-----------------------+---------------------+
        |Stage|Producer              |Filter                 |Consumer             |
        +-----+----------------------+-----------------------+---------------------+
        | 0.0 |                      |                       |Request frame        |
        +-----+----------------------+-----------------------+---------------------+
        | 0.1 |                      |Receives request       |                     |
        |     |                      |Request frame          |                     |
        +-----+----------------------+-----------------------+---------------------+
        | 0.2 |Receives request      |                       |                     |
        |     |Generates frame for   |                       |                     |
        |     |current position      |                       |                     |
        |     |Increments position   |                       |                     |
        +-----+----------------------+-----------------------+---------------------+
        | 0.3 |                      |Receives frame         |                     |
        |     |                      |Updates frame          |                     |
        +-----+----------------------+-----------------------+---------------------+
        | 0.4 |                      |                       |Receives frame       |
        +-----+----------------------+-----------------------+---------------------+

        Note that neither the filter nor the consumer have any conception of
        'position' until they receive a frame. Speed and position are properties of
        the producer, and they are assigned to the frame object when the producer
        creates it.

        Step 0.3 is a critical one here - if the filter determines that the frame is
        of interest to it, then it should manipulate the image and/or audio stacks
        and properties as required.

        Assuming that the filter deals with both image and audio, then it should
        push data and methods on to the stacks which will deal with the processing. 
        This can be done with the mlt_frame_push_image and audio methods. In order for 
        the filter to register interest in the frame, the stacks should hold:

        image stack:
        [ producer_get_image ] [ data1 ] [ data2 ] [ filter_get_image ]

        audio stack:
        [ producer_get_audio ] [ data ] [ filter_get_audio ]

        The filter_get methods are invoked automatically when the consumer invokes a
        get_image on the frame. 

        +-----+----------------------+-----------------------+---------------------+
        |Stage|Producer              |Filter                 |Consumer             |
        +-----+----------------------+-----------------------+---------------------+
        | 1.0 |                      |                       |frame_get_image      |
        +-----+----------------------+-----------------------+---------------------+
        | 1.1 |                      |filter_get_image:      |                     |
        |     |                      | pop data2 and data1   |                     |
        |     |                      | frame_get_image       |                     |
        +-----+----------------------+-----------------------+---------------------+
        | 1.2 |producer_get_image    |                       |                     |
        |     |  Generates image     |                       |                     |
        +-----+----------------------+-----------------------+---------------------+
        | 1.3 |                      |Receives image         |                     |
        |     |                      |Updates image          |                     |
        +-----+----------------------+-----------------------+---------------------+
        | 1.4 |                      |                       |Receives image       |
        +-----+----------------------+-----------------------+---------------------+

        Obviously, if the filter isn't interested in the image, then it should leave
        the stack alone, and then the consumer will retrieve its image directly from
        the producer.

        Similarly, audio is handled as follows:

        +-----+----------------------+-----------------------+---------------------+
        |Stage|Producer              |Filter                 |Consumer             |
        +-----+----------------------+-----------------------+---------------------+
        | 2.0 |                      |                       |frame_get_audio      |
        +-----+----------------------+-----------------------+---------------------+
        | 2.1 |                      |filter_get_audio:      |                     |
        |     |                      | pop data              |                     |
        |     |                      | frame_get_audio       |                     |
        +-----+----------------------+-----------------------+---------------------+
        | 2.2 |producer_get_audio    |                       |                     |
        |     |  Generates audio     |                       |                     |
        +-----+----------------------+-----------------------+---------------------+
        | 2.3 |                      |Receives audio         |                     |
        |     |                      |Updates audio          |                     |
        +-----+----------------------+-----------------------+---------------------+
        | 2.4 |                      |                       |Receives audio       |
        +-----+----------------------+-----------------------+---------------------+

        And finally, when the consumer is done with the frame, it should close it.

        Note that a consumer may not evaluate both image and audio for any given
        frame, especially in a realtime environment. See 'Realtime Considerations'
        below.

        By default, a frame has the following properties:

        +------------------+------------------------------------+------------------+
        |Name              |Description                         |Values            |
        +------------------+------------------------------------+------------------+
        |_position         |The producers frame position        |0 to n            |
        +------------------+------------------------------------+------------------+
        |_speed            |The producers speed                 |double            |
        +------------------+------------------------------------+------------------+
        |image             |The generated image                 |NULL or pointer   |
        +------------------+------------------------------------+------------------+
        |alpha             |The generated alpha mask            |NULL or pointer   |
        +------------------+------------------------------------+------------------+
        |width             |The width of the image              |                  |
        +------------------+------------------------------------+------------------+
        |height            |The height of the image             |                  |
        +------------------+------------------------------------+------------------+
        |normalised_width  |The normalised width of the image   |720               |
        +------------------+------------------------------------+------------------+
        |normalised_height |The normalised height of the image  |576 or 480        |
        +------------------+------------------------------------+------------------+
        |progressive       |Indicates progressive/interlaced    |0 or 1            |
        +------------------+------------------------------------+------------------+
        |top_field_first   |Indicates top field first           |0 or 1            |
        +------------------+------------------------------------+------------------+
        |audio             |The generated audio                 |NULL or pointer   |
        +------------------+------------------------------------+------------------+
        |frequency         |The frequency of the audio          |                  |
        +------------------+------------------------------------+------------------+
        |channels          |The channels of the audio           |                  |
        +------------------+------------------------------------+------------------+
        |samples           |The samples of the audio            |                  |
        +------------------+------------------------------------+------------------+
        |aspect_ratio      |The sample aspect ratio of the image|double            |
        +------------------+------------------------------------+------------------+
        |test_image        |Used to indicate no image available |0 or 1            |
        +------------------+------------------------------------+------------------+
        |test_audio        |Used to indicate no audio available |0 or 1            |
        +------------------+------------------------------------+------------------+

        The consumer can attach the following properties which affect the default
        behaviour of a frame:

        +------------------+------------------------------------+------------------+
        |test_card_producer|Synthesise test images from here    |NULL or pointer   |
        +------------------+------------------------------------+------------------+
        |consumer_aspect_  |Apply this aspect ratio to the test |double            |
        |ratio             |card producer                       |                  |
        +------------------+------------------------------------+------------------+
        |rescale.interp    |Use this scale method for test image|"string"          |
        +------------------+------------------------------------+------------------+

        While most of these are mainly self explanatory, the normalised_width and
        normalised_height values require a little explanation. These are required
        to ensure that effects are consistently handled as PAL or NTSC, regardless 
        of the consumers or producers width/height image request. 

        The test_image and audio flags are used to determine when images and audio
        should be synthesised.

        Additional properties may be provided by the producer implementation, and
        filters, transitions and consumers may add additional properties to
        communicate specific requests. These are documented in modules.txt.

        The complete API for the mlt frame is as follows:

        mlt_frame mlt_frame_init( );
        mlt_properties mlt_frame_properties( mlt_frame this );
        int mlt_frame_is_test_card( mlt_frame this );
        int mlt_frame_is_test_audio( mlt_frame this );
        double mlt_frame_get_aspect_ratio( mlt_frame this );
        int mlt_frame_set_aspect_ratio( mlt_frame this, double value );
        mlt_position mlt_frame_get_position( mlt_frame this );
        int mlt_frame_set_position( mlt_frame this, mlt_position value );
        int mlt_frame_get_image( mlt_frame this, uint8_t **buffer, mlt_image_format *format, int *width, int *height, int writable );
        uint8_t *mlt_frame_get_alpha_mask( mlt_frame this );
        int mlt_frame_get_audio( mlt_frame this, int16_t **buffer, mlt_audio_format *format, int *frequency, int *channels, int *samples );
        int mlt_frame_push_get_image( mlt_frame this, mlt_get_image get_image );
        mlt_get_image mlt_frame_pop_get_image( mlt_frame this );
        int mlt_frame_push_frame( mlt_frame this, mlt_frame that );
        mlt_frame mlt_frame_pop_frame( mlt_frame this );
        int mlt_frame_push_service( mlt_frame this, void *that );
        void *mlt_frame_pop_service( mlt_frame this );
        int mlt_frame_push_audio( mlt_frame this, void *that );
        void *mlt_frame_pop_audio( mlt_frame this );
        void mlt_frame_close( mlt_frame this );


mlt_service:

        The service base class extends properties and allows 0 to m inputs and 0 to
        n outputs and is represented as follows:

            +-----------+
        - ->|           |- ->
        - ->|  Service  |- ->
        - ->|           |
            +-----------+
            | properties|
            +-----------+

        Descendents of service impose restrictions on how inputs and outputs can be 
        connected and will provide a basic set of properties. Typically, the service 
        instance is encapsulated by the descendent in order for it to ensure that
        its connection rules are followed.

        A service does not define any properties when constructed. It should be
        noted that producers, filters and transitions my be serialised (say, via the
        westley consumer), and care should be taken to distinguish between
        serialisable and transient properties. The convention used is to prefix
        transient properties with an underscore.

        The public interface is defined by the following functions:

        int mlt_service_init( mlt_service this, void *child );
        mlt_properties mlt_service_properties( mlt_service this );
        int mlt_service_connect_producer( mlt_service this, mlt_service producer, int index );
        int mlt_service_get_frame( mlt_service this, mlt_frame_ptr frame, int index );
        void mlt_service_close( mlt_service this );

        Typically, only direct descendents of services need invoke these methods and
        developers are encouraged to use those extensions when defining new services. 


mlt_producer:

        A producer has 0 inputs and 1 output:

            +-----------+
            |           |
            | Producer  |--->
            |           |
            +-----------+
            | service   |
            +-----------+

        A producer provides an abstraction for file readers, pipes, streams or any
        other image or audio input. 

        When instantiated, a producer has the following properties:

        +------------------+------------------------------------+------------------+
        |Name              |Description                         |Values            |
        +------------------+------------------------------------+------------------+
        |mlt_type          |The producers type                  |mlt_producer      |
        +------------------+------------------------------------+------------------+
        |_position         |The producers frame position        |0 to n            |
        +------------------+------------------------------------+------------------+
        |_speed            |The producers speed                 |double            |
        +------------------+------------------------------------+------------------+
        |fps               |The output frames per second        |25 or 29.97       |
        +------------------+------------------------------------+------------------+
        |in                |The in point in frames              |0 to length - 1   |
        +------------------+------------------------------------+------------------+
        |out               |The out point in frames             |in to length - 1  |
        +------------------+------------------------------------+------------------+
        |length            |The length of the input in frames   |0 to n            |
        +------------------+------------------------------------+------------------+
        |aspect_ratio      |aspect_ratio of the source          |0 to n            |
        +------------------+------------------------------------+------------------+
        |eof               |end of clip behaviour               |"pause" or "loop" |
        +------------------+------------------------------------+------------------+
        |resource          |Constructor argument (ie: file name)|"<resource>"      |
        +------------------+------------------------------------+------------------+

        Additional properties may be provided by the producer implementation.

        The public interface is defined by the following functions:

        mlt_producer mlt_producer_new( );
        int mlt_producer_init( mlt_producer this, void *child );
        mlt_service mlt_producer_service( mlt_producer this );
        mlt_properties mlt_producer_properties( mlt_producer this );
        int mlt_producer_seek( mlt_producer this, mlt_position position );
        mlt_position mlt_producer_position( mlt_producer this );
        mlt_position mlt_producer_frame( mlt_producer this );
        int mlt_producer_set_speed( mlt_producer this, double speed );
        double mlt_producer_get_speed( mlt_producer this );
        double mlt_producer_get_fps( mlt_producer this );
        int mlt_producer_set_in_and_out( mlt_producer this, mlt_position in, mlt_position out );
        mlt_position mlt_producer_get_in( mlt_producer this );
        mlt_position mlt_producer_get_out( mlt_producer this );
        mlt_position mlt_producer_get_playtime( mlt_producer this );
        mlt_position mlt_producer_get_length( mlt_producer this );
        void mlt_producer_prepare_next( mlt_producer this );
        void mlt_producer_close( mlt_producer this );


mlt_filter:

        The public interface is defined by the following functions:

        int mlt_filter_init( mlt_filter this, void *child );
        mlt_filter mlt_filter_new( );
        mlt_service mlt_filter_service( mlt_filter this );
        mlt_properties mlt_filter_properties( mlt_filter this );
        mlt_frame mlt_filter_process( mlt_filter this, mlt_frame that );
        int mlt_filter_connect( mlt_filter this, mlt_service producer, int index );
        void mlt_filter_set_in_and_out( mlt_filter this, mlt_position in, mlt_position out );
        int mlt_filter_get_track( mlt_filter this );
        mlt_position mlt_filter_get_in( mlt_filter this );
        mlt_position mlt_filter_get_out( mlt_filter this );
        void mlt_filter_close( mlt_filter );


mlt_transition:

        The public interface is defined by the following functions:

        int mlt_transition_init( mlt_transition this, void *child );
        mlt_transition mlt_transition_new( );
        mlt_service mlt_transition_service( mlt_transition this );
        mlt_properties mlt_transition_properties( mlt_transition this );
        int mlt_transition_connect( mlt_transition this, mlt_service producer, int a_track, int b_track );
        void mlt_transition_set_in_and_out( mlt_transition this, mlt_position in, mlt_position out );
        int mlt_transition_get_a_track( mlt_transition this );
        int mlt_transition_get_b_track( mlt_transition this );
        mlt_position mlt_transition_get_in( mlt_transition this );
        mlt_position mlt_transition_get_out( mlt_transition this );
        mlt_frame mlt_transition_process( mlt_transition this, mlt_frame a_frame, mlt_frame b_frame );
        void mlt_transition_close( mlt_transition this );


mlt_consumer:

        The public interface is defined by the following functions:

        int mlt_consumer_init( mlt_consumer this, void *child );
        mlt_service mlt_consumer_service( mlt_consumer this );
        mlt_properties mlt_consumer_properties( mlt_consumer this );
        int mlt_consumer_connect( mlt_consumer this, mlt_service producer );
        int mlt_consumer_start( mlt_consumer this );
        mlt_frame mlt_consumer_get_frame( mlt_consumer this );
        mlt_frame mlt_consumer_rt_frame( mlt_consumer this );
        int mlt_consumer_stop( mlt_consumer this );
        int mlt_consumer_is_stopped( mlt_consumer this );
        void mlt_consumer_close( mlt_consumer );


Specialised Producers:

        There are two major types of specialised producers - playlists and tractors.

        The following sections describe these.


mlt_playlist:

        mlt_playlist mlt_playlist_init( );
        mlt_producer mlt_playlist_producer( mlt_playlist this );
        mlt_service mlt_playlist_service( mlt_playlist this );
        mlt_properties mlt_playlist_properties( mlt_playlist this );
        int mlt_playlist_count( mlt_playlist this );
        int mlt_playlist_clear( mlt_playlist this );
        int mlt_playlist_append( mlt_playlist this, mlt_producer producer );
        int mlt_playlist_append_io( mlt_playlist this, mlt_producer producer, mlt_position in, mlt_position out );
        int mlt_playlist_blank( mlt_playlist this, mlt_position length );
        mlt_position mlt_playlist_clip( mlt_playlist this, mlt_whence whence, int index );
        int mlt_playlist_current_clip( mlt_playlist this );
        mlt_producer mlt_playlist_current( mlt_playlist this );
        int mlt_playlist_get_clip_info( mlt_playlist this, mlt_playlist_clip_info *info, int index );
        int mlt_playlist_insert( mlt_playlist this, mlt_producer producer, int where, mlt_position in, mlt_position out );
        int mlt_playlist_remove( mlt_playlist this, int where );
        int mlt_playlist_move( mlt_playlist this, int from, int to );
        int mlt_playlist_resize_clip( mlt_playlist this, int clip, mlt_position in, mlt_position out );
        void mlt_playlist_close( mlt_playlist this );

mlt_tractor: