[FFmpeg-devel] [flamefest-start] A little something on MMX/SSE intrinsics
Thu Feb 28 23:39:30 CET 2008
Without taking sides in the debate...
Thursday 28 February 2008 22:15-kor Luca Barbato ezt ?rta:
> No, even a 3% can be important, BUT getting a large gain for every cpu
> in a family with minimal effort is nicer that get the theoretical
> maximum for a small subset and leave the rest in a painful situation.
I am not knowledgeable about ppc at all - but what is the big difference - in
scheduling, as you claim - on the different cores? I mean instructions can
have different latency and throughput - but at the end of the day these don't
really matter, as you always need to strive for a reasonable amount of
non-dependent instructions after each. And when you do that, I don't think
the code will really slow down by 5% on different cores.
So I don't really see the big deal with scheduling - you just interlave a
reasonable number of dependency chains, that's all you can do - and it's
enough - independent of the cores and latencies.
Now if these cores have differing number of registers is indeed another issue,
which in an ideal world could be handled automatically by the compiler for
parts using intrinsics.
More information about the ffmpeg-devel