[Libav-user] Hardware accelerated decoding

[Kirill Gavrilov]

Hi,

On Wed, Sep 4, 2013 at 5:33 PM, Attila Sukosd <attila.sukosd at gmail.com>wrote:

> I've looked at OS X's VDA acceleration, but it seems as though you need to
> do some extra work in the application to support the different hwaccels,
> but I haven't found any nice examples on how to do it.
>
Most accelerated decoders decode picture into API-specific surface in GPU
memory, which can be drawn using OpenGL (VDPAU) or Direct3D (DXVA2) without
extra copying it back to CPU memory.
For this reason you need to do a lot of extra work to configure FFmpeg
using specific hardware decoder (or detect when it can not be used), render
result on the screen using more complicated scenarios.
Because this stuff is really overcomplicated, painful and contradicts to
implemented decoding+rendering pipeline - I haven't tried it in my
application yet.

VDA is somewhat simpler than most others because the original Apple API
doesn't provide the way to render result directly and you should copy frame
back to CPU memory anyway.
Technically you should just try to open another decoder and use it instead
of auto-detected one in avcodec_open2 (with extra checks and + probably
overridden get_format2 if you like planar YUV420P):

AVCodec* aCodecVda = avcodec_find_decoder_by_name ("h264_vda");
avcodec_open2 (theCodecCtx, aCodecVda, NULL);

I have tried this decoder on my old Macbook and it is significantly slower
than software decoder.
There is also patch in mail list which introduces similar decoder
(automatic GPU->CPU memory copying) with DXVA2 acceleration.
-----------------------------------------------
Kirill Gavrilov,
Software designer.
<kirill at sview.ru>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://ffmpeg.org/pipermail/libav-user/attachments/20130904/7cc3bd7a/attachment.html>