[FFmpeg-devel] [PATCH 09/13] avcodec/svq1dec: clear MMX state after MB decode loop

[u-9iep at aetey.se]

On Tue, Oct 25, 2016 at 05:48:50PM +0200, Henrik Gramner wrote:
> On Tue, Oct 25, 2016 at 2:28 PM,  <u-9iep at aetey.se> wrote:
> > It would be nice to look at a benchmarking comparison, to be able to
> > see the actual practical performance gain of the decision not to follow
> > the ABI.
> 
> Just a quick comparison from adding EMMS to a random MMX function
> (from x264, because I happened to have the source for that but not
> ffmpeg on this machine and I was too lazy to download it).
> 
> sad_4x4_c: 359
> sad_4x4_mmx: 94
> sad_4x4_mmx (emms): 186

This is unfortunatly not exactly the numbers a user or a packager would
need to know. The relevant ones would be what time it takes to actually
compress or decompress a given file.

(if most of the time is spent calling this very function, then we can
extrapolate the numbers to the net result, otherwise we shouldn't)

If we nevertheless assume that the net performance is proportional
to the numbers above, we learn that omission of MMX acceleration
compatible with the standard ABI would lead to 50% performance loss
for the ABI-compliant users (pure C instead of compliant MMX).

This would be a pity, even if the circle of the users who depend
on strict ABI-compliance is limited, for the moment.

Regards,
Rune