Fri Dec 28 05:06:56 CET 2007
On Thu, Dec 27, 2007 at 11:37:01PM +0300, Anatoliy wrote:
> Michael Niedermayer wrote:
> > Noone tried AFAIK, but if you want to try patches are certainly welcome
> > though likely just (I)DCT will not lead to much speedup.
> > [...]
> As I have seen after code reseaching there is one big structure
> DSPContext containing all available methods for DSP optimization. There
> are too much methods. Which of them mostly important for codec time
I do not think that adding cuda support via DSPContext will work out very
well. The reason is that these functions are all working on small amounts
of data at a time ...
It would make more sense if you would look at the code below HAVE_XVMC.
And implement CUDA support similarely. The XVMC code does the bitstream
decoding in libavcodec while passing motion vectors and idct coeffs to
XVMC, that way its easy to collect data for many blocks and then work on
them in parallel.
XVMC IIRC supports just mpeg1/mpeg2 though the design can easily be extended
to mpeg4 ASP (divx4/5) and msmpeg4v3 (divx3). H.264 though is a different
beast and will require a seperate (and more complex) solution.
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
I have often repented speaking, but never of holding my tongue.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 189 bytes
Desc: Digital signature
More information about the ffmpeg-devel