[FFmpeg-cvslog] r9307 - in trunk/libavcodec: dsputil.c mpegvideo.c

[Jan Knutar]

On Thursday 14 June 2007 11:44, diego wrote:

> -#ifdef HAVE_MLIB
> +#elif defined(HAVE_MLIB)
>      dsputil_init_mlib(c, avctx);
> -#endif
> -#ifdef ARCH_SPARC
> +#elif defined(ARCH_SPARC)
>     dsputil_init_vis(c,avctx);
> -#endif

Ok, I benchmarked in mplayer with a svn snapshot having a lavc before r9307.
4 variants, no accel, vis accel, mlib, vis+mlib. I compiled with --enable-mlib
and without, and then modified libavcodec/sparc/dsputil_vis.c to enable/disable
vis.

I used -ac dummy -ao null -vo x11 -vf format=bgr24 -benchmark -quiet file.avi
3 runs each, the best result of each below. The difference between runs was <0.25s.
The test file's resolution was 320x240, the video codec used in it is lavc's mpeg4.
2212 frames. CPU UltraSparc IIe 650Mhz.

The VC value is most meaningful, I guess.

no accel
BENCHMARKs: VC:  16.545s VO:  22.799s A:   0.000s Sys:   0.710s =   40.053s

vis-only:
BENCHMARKs: VC:  15.364s VO:  22.584s A:   0.000s Sys:   0.716s =   38.664s

mlib-only:
BENCHMARKs: VC:  13.273s VO:  22.740s A:   0.000s Sys:   0.905s =   36.917s

mlib+vis:
BENCHMARKs: VC:  12.308s VO:  22.789s A:   0.000s Sys:   0.689s =   35.786s


It would be interesting to compare the areas where vis and mlib overlap, too. Or see
if there's any significant difference if you change the init order.