[FFmpeg-devel] VP8 profiling status

[David Conrad]

On Jul 20, 2010, at 8:06 PM, Ronald S. Bultje wrote:

> Hi,
> 
> On Mon, Jul 19, 2010 at 4:06 PM, Ronald S. Bultje <rsbultje at gmail.com> wrote:
>> time ./ffmpeg_g -i Elephants_Dream-360p-Stereo.webm -f md5 -an -vcodec
>> rawvideo -

You should use -f null for benchmarking. And add -vsync 0 for webm since it's likely that there will be double the number of output frames (this only really matters outside of -f null though.)

With higher cache-unfriendly resolutions it should be more, e.g. with sintel_trailer_1080p_vp8_vorbis.webm on my Penryn (x86-64):

Us:
real	0m18.670s
user	0m18.359s
sys	0m0.116s

libvpx:
real	0m23.498s
user	0m23.348s
sys	0m0.130s

> That was for a Core1, now here's one for x86-64:
> 
> C:
> real	1m47.649s
> user	1m47.411s
> sys	0m0.218s
> 
> SIMD:
> real	0m52.142s
> user	0m51.943s
> sys	0m0.187s
> 
> libvpx:
> real	0m50.346s
> user	0m50.041s
> sys	0m0.295s
> 
> Jason has suggested a variety of ways to take more advantage of
> SSSE3/SSE4 instructions (which I basically didn't really focus on
> because my CPU doesn't support it), and more generally to be a little
> more efficient with registers in non-looping functions (e.g. all
> loopfilter functions in >=SSE2), which should allow us to speed it up
> quite a bit more. David also showed some progress on block coeff
> optimizations, so we'll probably be able to shave off a few seconds
> more.
> 
> Ronald
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at mplayerhq.hu
> https://lists.mplayerhq.hu/mailman/listinfo/ffmpeg-devel