[Ffmpeg-devel] Re: about mmx instructions

[Chengji Zhao]

Hi, all
Firstly, I did rewrite that part of the code in SSE2 instructions and  I 
achieved about 20% performance gain on my Pentium 4 pc
for that part of the code.
Secondly, if you take a closer look at  the original implementation. you 
will find that the code can still be improved  in terms of  CPU resource
 utilization . For example , if you changing some of the movq 
instructions into pshufw xxx, xxx, 0xe4 ( which means use  shift  unit 
instead of
load unit), you can gain some improvement.

By the way, I had a hard time to understand that magic too. I finally 
gave up( but I did write a small code to verify that formula :)
I will try again this weekend.


-- 

Regards

Chengji Zhao

Mediaware International
Phone: 61-2-6163 8776