[FFmpeg-devel] [HACK] 50% faster H.264 decoding

Alex Converse alex.converse
Sun Aug 22 21:50:48 CEST 2010


On Sun, Aug 22, 2010 at 3:42 PM, Luca Barbato <lu_zero at gentoo.org> wrote:
> On 08/22/2010 04:58 AM, Jason Garrett-Glaser wrote:
>>> Keep in mind that some compilers (like open64) might support inline asm
>>> and cannot do link time optimizations on yasm generated binaries...
>>
>> Why would I want the compiler doing link-time "optimization" on my
>> asm? ?I would consider that inability a benefit.
>
> It cannot perform link time optimization at all if you are trying to
> link something foreign apparently, that is an annoying limitation but
> that is. By your reasoning you would consider hand writing everything in
> asm since you would get it right better than a compiler.
>
> Is still the whole asm vs intrinsics again. I do agree that using asm
> _now_ on platforms with depressing implementations (hi sse and neon
> intrinsics) is the only safe way.
>
> Still you end up with code that wishfully would work fine across
> different cpus only because the scheduling of those instructions might
> not had changed a lot yet.
>
> ASM optimizations are single target and could be outright wrong when
> recycled (e.g x86 vs amd64).
>

Meanwhile in the real world, converting the imdct to from inline to
yasm didn't break a single platform the inline code ran on and also
works properly on win64 and suncc which were previously broken.



More information about the ffmpeg-devel mailing list