Video Hackfest conclusions

[Benjamin Otte]

Hi,

As you may be aware, we held a video hackfest last week in Barcelona.
Developers met to discuss how best to improve GPU support for video
applications. See http://gstreamer.freedesktop.org/wiki/VideoHackfest
for more details. In particular, you might be interested in the notes
some people took while the hackfest was going on. These can be found
at http://gstreamer.freedesktop.org/wiki/VideoHackfest/Notes

What follows is the results we arrived at after the hackfest. They are
taken from http://gstreamer.freedesktop.org/wiki/VideoHackfest/Conclusions
but certainly deserve wide circulation. We're interested in any
comments (or questions) you might have about them, so please don't
heistate to reply.

If you do reply, please make sure to reduce the recipient list appropriately.

Cheers,
Benjamin,

on behalf of all hackfest paticipants.


== YUV in pixman ==
It has been a goal for a while to make video a first-class citizen in
the image stack. A concrete API proposal on how to integrate video
into pixman was reached and reviewed.
Links:
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest/YCbCr_Formats
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest/pixman
Actions:
 * GStreamer developers to provide a specification of important YUV formats
 * Write implementation

== YUV in Cairo ==
To make it easy for applications to use these different formats, the
Cairo API needs to be extended to allow them. A proposal does exist
and has been reviewed, details need to be finalized.
Links:
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest/Cairo
Actions:
 * Finalize proposal
 * Write implementation

== locking in Cairo ==
GStreamer takes a lot of freedom in deciding which threads to schedule
elements in. Most hardware backends however require proper
serialization of commands. The current gst-plugins-cairo and
gst-plugins-gl code provide different, but ugly ways to achieve this.
Cairo's internal thread safety guarantees match these requirements
very well. But the different Cairo backends don't always keep these
guarantees. Interaction with these locking mechanisms from outside
applications is not possible yet either.
Links:
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest/Cairo
Actions:
 * Improve Cairo backend implementations, in particular X11 and GL
 * Expose Cairo's locking API to allow interaction with it
 * Write testcases to squash bugs

== locking in Mesa ==
The GLX specification does not allow binding the same GLX context in
multiple threads at the same time. This is a requirement for both
Cairo's and GStreamer's threading model. An extension was proposed and
initial code developed to support this requirement in the same way as
Apple's GL does by default. Windows does not support this and would
require potentially expensive fallback code.
Links:
 * http://people.freedesktop.org/~anholt/MESA_multithread_makecurrent.spec
Actions:
 * Get review for suggested extension and include it in future Mesa releases
 * Make Windows developers investigate the situation

== switch GStreamer to Cairo as default video transport model ==
The current approach to handling video in GStreamer is very outdated.
It does not allow hw-accelerated buffers nor does it provide a unified
API to modify video buffers, which leads to fragmentation and
duplication. It was agreed that using Cairo to provide an abstraction
as a drawing API and abstraction over different backends was a good
idea.
Links:
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest/Notes
 * http://cgit.freedesktop.org/~company/gst-plugins-cairo
Actions:
 * Rework gst-plugins-cairo to match improvements listed in previous points
 * Get more review on API and to avoid regressions
 * Merge into gst-plugins-base
 * Switch plugins gradually to use new Cairo code

== gst-plugins-gl ==
The general consensus was that gst-plugins-gl is a hack. It was
necessary in the past to get things to work, but is not a good way
forward. However, the functionality or performance provided by the
current elements needs to continue working. Developers shared the
opinion that gst-plugins-cairo with cairo-gl surfaces is the best way
to achieve this.
Links:
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest/Notes
Actions:
 * Get review of cairo-gl/gst-plugins-cairo code from gst-plugins-gl developers
 * Port gst-plugins-gl elements to use Cairo early to ensure required
features are available

== XRenderPutImage ==
Currently there is no way to upload video data to the X server for
later use. The XV extension is not sufficient for anything more
complicated than a simple video player.
Links:
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest/Cairo
Actions:
 * Figure out the best way to move YUV data into hw-accelerated Cairo
surfaces (GL vs X)
 * Possibly draft an extension to XRender to handle this in the no-GL case

== JIT in Pixman ==
ORC was investigated as a potential JIT for pixman. A lot of talk has
happened, but no clear conclusions were reached. The idea to share a
JIT with Mesa's GLSL stack was also brought up.
Links:
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest/Notes
Actions:
 * Continue investigating JIT options
 * Prototype ORC integration into Pixman

== Video decoding Acceleration ==
Hardware video decoding is a feature that everybody is very excited
about and wants to integrate into GStreamer as soon as possible. But
developers feel that the current APIs available - the focus was on
VAAPI and VDPAU - to do hardware acceleration fail to integrate well.
No developer managed to create even demo-quality integration with
gst-plugins-cairo, even though quite a few tried. In particular, both
VDPAU and VAAPI...

... do not integrate with existing frameworks
Both libraries fail to use (or allow easy conversion to) existing
objects (GEM handles, GL textures, ...) but only provide access to an
encapsulation. This makes it hard for a flexible framework like
GStreamer to make use of its features without limiting itself.

... do not have a clear scope
In particular VDPAU advertises itself as a video playback framework
using hardware acceleration. Unfortunately, they both limit themselves
to only the formats and functionalities supported by the hardware and
see fullscreen video playback as the single usecase. This needlessly
complicates integration with existing frameworks

... provide a very complicated API
For decoding video, an API like provided by libtheora, libmpeg2 or
even ffmpeg is very simple and allows easy integration with both the
demuxing  and the postprocessing/display side. VAAPI and VDPAU require
complex setups and preprocessing of the data to work properly.

... overlap with GL functionality
It was felt that a lot of processing features provided by these
libraries are equally well available using shaders in GL or by
providing GL extensions. It was unclear why it is necessary to use a
separate way to do this.

Links:
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest/Notes
 * http://gstreamer.freedesktop.org/wiki/VideoHackfest?action=AttachFile&do=view&target=VAinGST.pdf
Actions:
 * Get in contact with developers of the APIs
 * Encourage revision of public
 * Work closer together in integrating the provided functionality into GStreamer

= Timeline =
The end goal of all of this is to get the code into users' hands
quickly, but allow a smooth and non-disrupting upgrade. Desired dates
have been put forth to achieve this:
 * End of January: release new versions of Cairo and Pixman that
contain the new APIs reuired by GStreamer
 * End of March: X server and Mesa releases are due - make sure the
required Mesa extension is part of this. Ideally XrenderPutImage would
be included, too.
 * April/May, after next Fedora/Ubuntu releases: Merge Cairo support
into gst-plugins-base and start porting elements to it. Encourage
application developers (browser, Flash players) to make use of the new
APIs
 * October: another Fedora/Ubuntu release that switches all users to
the new APIs