[FFmpeg-devel] Parallelization of FFmpeg

[David Ison]

It was necessary to change the email subscription, because I had used to 
subscribe was an alias.


The vc1 decoder could need frame parallelization.
That being said, what normally happens is that
you decide on which part of FFmpeg you would
like to work on and start sending patches. In
theory, this can lead to duplicated work but in
reality this does not happen very often afaict.

(And I suspect parallelization of vc1 is not
an easy task.)

Thank you for any contribution!

Are you referring to the file

libavcodec/vc1.c   ?

I am somewhat new to the idea of sending patches.  Am more familiar with, for 
example, doing pull requests using github.

If I have questions about getting started, is it ok to put them here?



___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] Parallelization of FFmpeg

[David Ison]

To process multiple slices at the same time - would depend  on whether 
independent slices could be scaled without data dependency from other 
slices.


If there is data dependency, then it would be more difficult.


On 12/29/14 9:00 AM, wm4 wrote:

On Mon, 29 Dec 2014 11:31:36 + (UTC)
Carl Eugen Hoyos  wrote:


David Ison  dtison.net> writes:


(And I suspect parallelization of vc1 is not
an easy task.)

Thank you for any contribution!

Are you referring to the file

libavcodec/vc1.c   ?

To libavcodec/vc1*, yes

An easier choice may be parallelization of the
scale filter: It is possible to let libswscale
work on slices, working on several slices at
the same time should allow for a big speedup
for some conversions.

I don't think this will work. libswscale makes very specific
requirements on how you scale by slices, and it doesn't look like you
can actually scale different slices with a different context.

I'm not a libswscale expert, though.
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel


___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] Require Help for accelerating ffmpeg video conversion using CUDA

[David Ison]

Begin by identifying portions where the processing can be arranged to 
run in parallel, with as little data dependencies as possible between 
processing threads.


So far we have identified two possible candidates:

   The vc1 decoder
   libavcodec/vc1.c

and

   the scale filter
   scale filter (libswscale)

Maybe I can work with you a little to get started.


On 12/29/14 1:56 PM, Akash Talole wrote:

I am a BE computer student from Sandip Foundation Nasik, India and i am
working on BE project "CUDA for accelerating video conversion". I want to
know from you that how can i accelerate ffmpeg for video conversion using
CUDA. Please give me a help in this project. In ffmpeg where i can use CUDA
so that conversion will get speedup.


Your Regards,
Akash Talole
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel


___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

[FFmpeg-devel] Parallelization

[David Ison]

These are some general ideas I have developed about how to approach this.

Parallelization Strategy and Goals:

1.  Implement in such manner as can easily be disabled and the program 
runs in serial with the same results as the parallel version.


2.  The parallel version likewise produces the same results as the serial.

3.  I am planning initial implementations using OpenMP.  There is good 
support on Gnu, LLVM and other compilers 
.


4.  Incremental Implementation - i.e., portions which are easiest, 
parallelized first, and moving forward from there.


Other contributors - can you comment?  You know a lot more concerning 
the internal details of FFMpeg than I do.



___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel