On Mon, Apr 09, 2018 at 03:00:17AM +0200, Michael Niedermayer wrote:
> On Sun, Apr 08, 2018 at 03:33:17PM +0200, Martin Vignali wrote:
> > Hello,
> > 
> > Patchs in attach remove mmxext version of shufflebytes2103 and uyvyto422
> > 
> > Checkasm test show that mmxext version is slower than MMX version
> > 
> > Also disable MMXext for "internal" func : extract_even2
> > (only use by uyvyto422)
> are you testing these functions with real image data in a real image
> processing pipeline ?
> synthetic tests would likely not give correct results with non temporal
> stores.
> in a real processing pipeline the output image will generally not fit in
> the L1 cache and will generally not be reused before other things overwrote
> the cache. So it should be better not to waste cache for it in the first
> place.
> But a synthetic test which just repeatly filters a short line will
> behave very differently.
> Its thus expected that mmx2 performs poorly here in such a synthetic test
> but that doesnt say anything about how it performs in a real use case

also IIRC all the really old SIMD code has been benchmarked when it was
written in real use cases, so if a function that is favoured is slower thats
"strange". And the question should be asked "why" its slower now


Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

It is dangerous to be right in matters on which the established authorities
are wrong. -- Voltaire

Attachment: signature.asc
Description: PGP signature

ffmpeg-devel mailing list

Reply via email to