[MPlayer-dev-eng] [PATCH]Add support for CoreAVC h264 codec

[Luca Barbato]

Guillaume POIRIER wrote:
> Hi,
> 
> On 10/5/06, Luca Barbato <lu_zero at gentoo.org> wrote:
>> Michael Niedermayer wrote:
>>
>> > could you post benchmarks with h264 videos with various parameters with
>> > coreavc and ffh264? that should be VERY usefull to find out which
>> parts are
>> > faster then ffh264 which could then help to improve ffh264
>> >
>> > things to test
>> > * low resolution where all reference frames+1 _easily_ fit in the L2
>> cache
>> > * CABAC / CAVLC
>> > * high bitrate / low bitrate
>> > * intra only
>> > * B frames vs. no B frames
>> > * loop filter / disabled loop filter
>> >
>>
>> Please also use oprofile (latest release/cvs) or equivalent tools,
>> recently I started digging a bit and looks like that:
>>
>> for G4 you spend 23% memcpying data (equally balanced smallcopys and
>> bigcopys, I implemented a naif smallcopy and I shaved 1/25 of the time)
>>
>> for G5 you spend nearly 10% of the time (the glibc in gentoo includes G5
>> specific improvements over standard memcpy)
> 
> Alternatively, can decoding be done without so much mem copies?
> 

Hadn't look at the code that much...

> 
>> I guess we could move the discussion on ffmpeg-devel (a patch with the
>> smallcopys is available but quite dirty)
> 
> Could you point me out the G4 and G5 memcopy routines? x264 could
> probably benefit from this, as currently on PPC the copy is really a
> trivial implementation.

http://penguinppc.org/dev/glibc/glibc-powerpc-cpu-addon.html

the smallcpy stuff is just adding 4 dsputil functions to copy the small
and fixed amount of data in trivial ways in C and it could be
implemented in not so trivial ways in asm.

lu