[Ffmpeg-devel] [PATCH] another vorbis optimization
Fri Aug 18 19:23:09 CEST 2006
On Fri, 18 Aug 2006, Rich Felker wrote:
> On Fri, Aug 11, 2006 at 02:17:29AM +0200, Luca Barbato wrote:
>> Loren Merritt wrote:
>>> One branch (perfectly predictable) vs saving 224 integer additions when
>>> non-simd. Well, if you value simplicity of the code over speed in this
>>> less-common case (and I guess I do too, else I would've mmxed this
>>> loop), then I can remove it.
>> At least on ppc the branchless version is slightly faster (better
>> average with lower deviation)
> This should be true on any sane cpu. Only exception might be shit with
> slow bit arithmetic, P4 anyone??
Really? one branch should be slower than 224 adds?
Granted, there is a tradeoff only without the sse/3dnow/altivec version
of float2int. If an optimized version is used, then both choices use the
same amount of arithmatic, and the only difference is branch vs no branch.
More information about the ffmpeg-devel