[Ffmpeg-devel] [RFC] smallcpy for h264

[Luca Barbato]

Siarhei Siamashka wrote:
> On Saturday 07 October 2006 05:11, Luca Barbato wrote:
> 
>> here a pretty dumb patch that happens to work nicely on my system,
>> probably a better solution should be avoding memcpy.
> 
> ff_smallcpy_n in the patch assumes that size is divisible by 16, is it ok?

there is a comment about it and is done on purpose (on ppc I'm using
altivec for it)

> 
> A small suggestion is to probably have no return type for these functions as
> it is not used in ffmpeg anyway.

I'll update it soon

> It will make implementation a bit simplier
> and faster (no need to save destination buffer address in a register or stack
> for returning it later).
> Also an option for having direct calls instead of
> function pointer calls would be good to have in order to reduce call overhead.

I was thinking about either adding it as macro instead of dsputil in
order to cut the overhead since those function should be few lines of asm.

> 
> Nevertheless aligning buffers at 16 byte boundary is a good idea and having
> some special redefineable functions to handle special buffer address
> alignment/buffer size cases can help to improve performance on some
> architectures (especially for those which require strict alignment).

That was the idea, still I'm unsure which is the better method.

-- 

Luca Barbato

Gentoo/linux Gentoo/PPC
http://dev.gentoo.org/~lu_zero