Quoting Vittorio Giovara (2015-05-01 03:44:30)
> On Thu, Apr 30, 2015 at 2:03 PM, Anton Khirnov <[email protected]> wrote:
> > Quoting Vittorio Giovara (2015-04-27 20:13:25)
> >> They are just duplicates of AVCodecContext members so use them directly.
> >
> > Have you benchmarked this?
> >
> > The reason there are copies of those fields is probably more efficient
> > cache usage. I am not sure it actually matters anywhere, but it should
> > be tested.
> 
> Well I did a rough check with make -j fate on the currently published
> set (6 patches) under a few configurations, but didn't find a huge
> impact on performance (as expected).
> 
> * clean master
>   (+debug -opt +asm)
>     real    0m59.735s
>     user    6m26.844s
>     sys    0m9.752s
>   (+debug -opt -asm)
>     real    1m41.182s
>     user    9m54.168s
>     sys    0m10.552s
>   (-debug +opt +asm)
>     real    0m24.618s
>     user    2m29.692s
>     sys    0m9.124s
>   (-debug +opt -asm)
>     real    0m37.459s
>     user    3m18.604s
>     sys    0m9.376s
> 
> * my set
>   (+debug -opt +asm)
>     real    0m59.714s
>     user    6m24.404s
>     sys    0m9.648s
>   (+debug -opt -asm)
>     real    1m40.503s
>     user    9m57.104s
>     sys    0m9.880s
>   (-debug +opt +asm)
>     real    0m24.618s
>     user    2m31.316s
>     sys    0m9.080s
>   (-debug +opt -asm)
>     real    0m33.260s
>     user    3m15.884s
>     sys    0m9.304s
> 
> Thanks to Luca for lending me the test machine.

FATE tests so many things that it's not really a good benchmark of any
specific codec.

-- 
Anton Khirnov
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel

Reply via email to