[FFmpeg-devel] FFmpeg 0.7.5 is very much slower than FFmpeg 0.7 for decoding [SOLVED]

[Unga]

--- On Mon, 10/3/11, Michael Niedermayer <michaelni at gmx.at> wrote:

> From: Michael Niedermayer <michaelni at gmx.at>
> Subject: Re: [FFmpeg-devel] FFmpeg 0.7.5 is very much slower than FFmpeg 0.7 for decoding
> To: "FFmpeg development discussions and patches" <ffmpeg-devel at ffmpeg.org>
> Date: Monday, October 3, 2011, 12:11 AM
> On Sun, Oct 02, 2011 at 08:38:40AM
> -0700, Unga wrote:
> [...]
> >
> > I'll try to get correct timing for decoding.
> 
> see START/STOP_TIMER
> they provide quite precisse timing
> also make sure all asm optims are used and theres no
> difference in
> thread settings
> also if you use swscale make sure it too uses the asm
> 
> and please report your findings even if it turns out to be
> a bug in
> your code so noone else wastes time investigating this
> 

The issue was asm options. 

The ffmpeg version 0.7-rc1 was installed with asm options. (manual compilation)

ffmpeg version 0.7.5 was installed with asm options disabled! (FreeBSD port system)

Now, ffmpeg version 0.7.5 is installed without FreeBSD port system, just manual compilation with all asm options.

The top shows almost same load percentages.

The START/STOP_TIMER shows following:

ffmpeg version 0.7.5
====================

Decode
------

2nd run:
61534428 dezicycles in decode_line, 254 runs, 2 skips
112586535 dezicycles in decode_line, 510 runs, 2 skips
147411732 dezicycles in decode_line, 1022 runs, 2 skips

3rd run:
62363599 dezicycles in decode_line, 254 runs, 2 skips
113073326 dezicycles in decode_line, 510 runs, 2 skips
147496619 dezicycles in decode_line, 1022 runs, 2 skips


ffmpeg version 0.7-rc1
======================

Decode
------

61863204 dezicycles in decode_line, 254 runs, 2 skips
113255074 dezicycles in decode_line, 510 runs, 2 skips
147448029 dezicycles in decode_line, 1022 runs, 2 skips

63262963 dezicycles in decode_line, 254 runs, 2 skips
113094623 dezicycles in decode_line, 510 runs, 2 skips
147633460 dezicycles in decode_line, 1022 runs, 2 skips


ffmpeg version 0.7.5
====================

sws_scale
---------

61211763 dezicycles in sws_scale_line, 256 runs, 0 skips
61225849 dezicycles in sws_scale_line, 512 runs, 0 skips
61334792 dezicycles in sws_scale_line, 1024 runs, 0 skips


ffmpeg version 0.7-rc1
======================

sws_scale
---------

64247934 dezicycles in sws_scale_line, 256 runs, 0 skips
63500319 dezicycles in sws_scale_line, 512 runs, 0 skips
63078084 dezicycles in sws_scale_line, 1024 runs, 0 skips

Conclusion:

For decoding, both versions show similar performance.

For sws_scale, ffmpeg version 0.7.5 is obviously faster.

Very sorry that I over looked the asm options and rushed to post.

Best regards
Unga