gstreamer/ext/ogg

This document describes some things to know about the Ogg format, as well
as implementation details in GStreamer.

INTRODUCTION
============

ogg and the granulepos
----------------------

An ogg stream contains pages with a serial number and a granulepos.
The granulepos is a 64 bit signed integer.  It is a value that in some way
represents a time since the start of the stream.
The interpretation as such is however both codec-specific and
stream-specific.

ogg has no notion of time: it only knows about bytes and granulepos values
on pages.

The granule position is just a number; the only guarantee for a valid ogg
stream is that within a logical stream, this number never decreases.

While logically a granulepos value can be constructed for every ogg packet,
the page is marked with only one granulepos value: the granulepos of the
last packet to end on that page.

theora and the granulepos
-------------------------

The granulepos in theora is an encoding of the frame number of the last
key frame ("i frame"), and the number of frames since the last key frame
("p frame").  The granulepos is constructed as the sum of the first number,
shifted to the left for granuleshift bits, and the second number:
granulepos = pframe << granuleshift + iframe

(This means that given a framenumber or a timestamp, one cannot generate
 the one and only granulepos for that page; several granulepos possibilities
 correspond to this frame number.  You also need the last keyframe, as well
 as the granuleshift.
 However, given a granulepos, the theora codec can still map that to a
 unique timestamp and frame number for that theora stream)

 Note: currently theora stores the "presentation time" as the granulepos;
       ie. a first data page with one packet contains one video frame and
       will be marked with 0/0.  Changing that to be 1/0 (so that it
       represents the number of decodable frames up to that point, like
       for Vorbis) is being discussed.

vorbis and granulepos
---------------------

In Vorbis, the granulepos represents the number of samples that can be
decoded from all packets up to that point.

In GStreamer, the vorbisenc elements produces a stream where:
- OFFSET is the byte offset this buffer is at; ie a running count of the
  number of bytes produced before
- OFFSET_END is the granulepos of the produced vorbis buffer
- TIMESTAMP is the timestamp matching the begin of the buffer
- DURATION is set such that TIMESTAMP + DURATION is the correct 
in a raw vorbis stream we use the granulepos as the offset field.

Ogg media mapping
-----------------

Ogg defines a mapping for each media type that it embeds.

For Vorbis:

  - 3 header pages, with granulepos 0.
     - 1 page with 1 packet header identification
     - N pages with 2 packets comments and codebooks
  - granulepos is samplenumber of next page
  - one packet can contain a variable number of samples but one frame
    that should be handed to the vorbis decoder.
  
For Theora
     
  - 3 header pages, with granulepos 0.
     - 1 page with 1 packet header identification
     - N pages with 2 packets comments and codebooks
  - granulepos is framenumber of last packet in page, where framenumber
    is a combination of keyframe number and p frames since keyframe.
  - one packet contains 1 frame
  



DEMUXING
========

ogg demuxer
-----------

This ogg demuxer has two modes of operation, which both share a significant
amount of code. The first mode is the streaming mode which is automatically 
selected when the demuxer is connected to a non-getrange based element. When 
connected to a getrange based element the ogg demuxer can do full seeking
with great efficiency.

1) the streaming mode.

In this mode, the ogg demuxer receives buffers in the _chain() function which
are then simply submited to the ogg sync layer. Pages are then processed when
the sync layer detects them, pads are created for new chains and packets are
sent to the peer elements of the pads.

In this mode, no seeking is possible. This is the typical case when the
stream is read from a network source.

In this mode, no setup is done at startup, the pages are just read and decoded.
A new logical chain is detected when one of the pages has the BOS flag set. At
this point the existing pads are removed and new pads are created for all the
logical streams in this new chain.
  

2) the random access mode.

  In this mode, the ogg file is first scanned to detect the position and length
of all chains. This scanning is performed using a recursive binary search
algorithm that is explained below.

    find_chains(start, end)
    {
      ret1 = read_next_pages (start);
      ret2 = read_prev_page (end);
      
      if (WAS_HEADER (ret1)) {
      }
      else {
      }

    }

  a) read first and last pages

   start                                                      end
    V                                                          V 
    +-----------------------+-------------+--------------------+
    |  111                  |  222        |  333               |
   BOS                     BOS           BOS                  EOS

   
   after reading start, serial 111, BOS, chain[0] = 111
   after reading end,   serial 333, EOS

   start serialno != end serialno, binary search start, (end-start)/2

   start                    bisect                            end
    V                         V                                V 
    +-----------------------+-------------+--------------------+
    |  111                  |  222        |  333               |

   
   after reading start, serial 111, BOS, chain[0] = 111
   after reading end,   serial 222, EOS

   while (



testcases
---------
    
 a) stream without BOS

    +----------------------------------------------------------+
       111                                                     |
                                                              EOS

 b) chained stream, first chain without BOS
  
    +-------------------+--------------------------------------+
       111              | 222                                  |
                       BOS                                    EOS


 c) chained stream
  
    +-------------------+--------------------------------------+
    |  111              | 222                                  |
   BOS                 BOS                                    EOS


 d) chained stream, second without BOS

    +-------------------+--------------------------------------+
    |  111              | 222                                  |
   BOS                                                        EOS

What can an ogg demuxer do?
---------------------------

An ogg demuxer can read pages and get the granulepos from them.
It can ask the decoder elements to convert a granulepos to time.

An ogg demuxer can also get the granulepos of the first and the last page of a
stream to get the start and end timestamp of that stream.
It can also get the length in bytes of the stream
(when the peer is seekable, that is).

An ogg demuxer is therefore basically able to seek to any byte position and
timestamp.

When asked to seek to a given granulepos, the ogg demuxer should always convert
the value to a timestamp using the peer decoder element conversion function. It
can then binary search the file to eventually end up on the page with the given
granule pos or a granulepos with the same timestamp.

Seeking in ogg currently
------------------------

When seeking in an ogg, the decoders can choose to forward the seek event as a
granulepos or a timestamp to the ogg demuxer.

In the case of a granulepos, the ogg demuxer will seek back to the beginning of
the stream and skip pages until it finds one with the requested timestamp.

In the case of a timestamp, the ogg demuxer also seeks back to the beginning of
the stream. For each page it reads, it asks the decoder element to convert the
granulepos back to a timestamp. The ogg demuxer keeps on skipping pages until
the page has a timestamp bigger or equal to the requested one.

It is therefore important that the decoder elements in vorbis can convert a
granulepos into a timestamp or never seek on timestamp on the oggdemuxer.

The default format on the oggdemuxer source pads is currently defined as a the
granulepos of the packets, it is also the value of the OFFSET field in the
GstBuffer.

MUXING
======

Oggmux
------

The ogg muxer's job is to output complete Ogg pages such that the absolute
time represented by the valid (ie, not -1) granulepos values on those pages
never decreases. This has to be true for all logical streams in the group at
the same time.

To achieve this, encoders are required to pass along the exact time that the
granulepos represents for each ogg packet that it pushes to the ogg muxer.
This is ESSENTIAL: without this exact time representation of the granulepos,
the muxer can not produce valid streams.

The ogg muxer has a packet queue per sink pad.  From this queue a page can
be flushed when:
  - total byte size of queued packets exceeds a given value
  - total time duration of queued packets exceeds a given value
  - total byte size of queued packets exceeds maximum Ogg page size
  - eos of the pad
  - encoder sent a command to flush out an ogg page after this new packet
    (in 0.8, through a flush event; in 0.10, with a GstOggBuffer)
  - muxer wants a flush to happen (so it can output pages)

The ogg muxer also has a page queue per sink pad.  This queue collects
Ogg pages from the corresponding packet queue.  Each page is also marked
with the timestamp that the granulepos in the header represents.

A page can be flushed from this collection of page queues when:
- ideally, every page queue has at least one page with a valid granulepos
  -> choose the page, from all queues, with the lowest timestamp value
- if not, muxer can wait if the following limits aren't reached:
  - total byte size of any page queue exceeds a limit
  - total time duration of any page queue exceeds a limit
- if this limit is reached, then:
  - request a page flush from packet queue to page queue for each queue
    that does not have pages
  - now take the page from all queues with the lowest timestamp value
  - make sure all later-coming data is marked as old, either to be still
    output (but producing an invalid stream, though it can be fixed later)
    or dropped (which means it's gone forever)

The oggmuxer uses the offset fields to fill in the granulepos in the pages.

GStreamer implementation details
--------------------------------
As said before, the basic rule is that the ogg muxer needs an exact time
representation for each granulepos.  This needs to be provided by the encoder.

Potential problems are:
 - initial offsets for a raw stream need to be preserved somehow.  Example:
   if the first audio sample has time 0.5, the granulepos in the vorbis encoder
   needs to be adjusted to take this into account.
 - initial offsets may need be on rate boundaries.  Example:
   if the framerate is 5 fps, and the first video frame has time 0.1 s, the
   granulepos cannot correctly represent this timestamp.
   This can be handled out-of-band (initial offset in another muxing format,
   skeleton track with initial offsets, ...)

Given that the basic rule for muxing is that the muxer needs an exact timestamp
matching the granulepos, we need some way of communicating this time value
from encoders to the Ogg muxer.  So we need a mechanism to communicate
a granulepos and its time representation for each GstBuffer.

(This is an instance of a more generic problem - having a way to attach
 more fields to a GstBuffer)

Possible ways:
- setting TIMESTAMP to this value: bad - this value represents the end time
  of the buffer, and thus conflicts with GStreamer's idea of what TIMESTAMP
  is.  This would cause problems muxing the encoded stream in other muxing
  formats, or for streaming.  Note that this is what was done in GStreamer 0.8
- setting DURATION to GP_TIME - TIMESTAMP: bad - this breaks the concept of
  duration for this frame.  Take the video example above; each buffer would
  have a correct timestamp, but always a 0.1 s duration as opposed to the
  correct 0.2 s duration
- subclassing GstBuffer: clean, but requires a common header used between
  ogg muxer and all encoders that can be muxed into ogg.  Also, what if
  a format can be muxed into more than one container, and they each have
  their own "extra" info to communicate ?
- adding key/value pairs to GstBuffer: clean, but requires changes to
  core.  Also, the overhead of allocating e.g. a GstStructure for *each* buffer
  may be expensive.
- "cheating":
  - abuse OFFSET to store the timestamp matching this granulepos
  - abuse OFFSET_END to store the granulepos value
  The drawback here is that before, it made sense to use OFFSET and OFFSET_END
  to store a byte count.  Given that this is not used for anything critical
  (you can't store a raw theora or vorbis stream in a file anyway),
  this is what's being done for now.

In practice
-----------
- all encoders of formats that can be muxed into Ogg produce a stream where:
  - OFFSET is abused to be the granulepos of the encoded theora buffer
  - OFFSET_END is abused to be the timestamp corresponding exactly to the
    granulepos
  - TIMESTAMP is the timestamp matching the begin of the buffer
  - DURATION is the length in time of the buffer

- initial delays should be handled in the GStreamer encoders by mangling
  the granulepos of the encoded packet to take the delay into account as
  best as possible and store that in OFFSET;
  this then brings TIMESTAMP + DURATION to within less
  than a frame period of the granulepos's time representation
  The ogg muxer will then create new ogg packets with this OFFSET as
  the granulepos.  So in effect, the granulepos produced by the encoders
  does not get used directly.

TODO
----
- decide on a proper mechanism for communicating extra per-buffer fields
- the ogg muxer sets timestamp and duration on outgoing ogg pages based on
  timestamp/duration of incoming ogg packets.
  Note that:
  - since the ogg muxer *has* to output pages sorted by gp time, representing
    end time of the page, this means that the buffer's timestamps are not
    necessarily monotonically increasing
  - timestamp + duration of buffers don't match up; the duration represents
    the length of the ogg page *for that stream*.  Hence, for a normal
    two-stream file, the sum of all durations is twice the length of the
    muxed file.

TESTING
-------
Proper muxing can be tested by generating test files with command lines like:
- video and audio start from 0:
gst-launch -v videotestsrc ! theoraenc ! oggmux audiotestsrc ! audioconvert ! vorbisenc ! identity ! oggmux0. oggmux0. ! filesink location=test.ogg

- video starts after audio:
gst-launch -v videotestsrc timestamp-offset=500000000 ! theoraenc ! oggmux audiotestsrc ! audioconvert ! vorbisenc ! identity ! oggmux0. oggmux0. ! filesink location=test.ogg

- audio starts after video:
gst-launch -v videotestsrc ! theoraenc ! oggmux audiotestsrc timestamp-offset=500000000 ! audioconvert ! vorbisenc ! identity ! oggmux0. oggmux0. ! filesink location=test.ogg

The resulting files can be verified with oggz-validate for correctness.