[XviD-devel] Decoder performance

[Christoph Lampert]

On Sat, 9 Aug 2003, Edouard Gomez wrote:
> Well i was wrong for MVs < 0. As usual, the "most usual" rounding bug
> got me... i was using shifts instead of divisions.
> 
> This patch replaces all shifts:
> foo = bar >> dec->quaterpel
> by
> foo = bar / (1+dec->quaterpel)

Hm, can't you handle this differently? From all I know, integer division 
isn't pipelined on any CPU architecture. For dividing by constants,
compilers replace this by shift+some rounding trick, but when dividing
by some variable, at least gcc doesn't manage to:

foo = bar/2 

becomes
        movl %edx,%eax
        shrl $31,%eax
        addl %eax,%edx
        sarl $1,%edx

So for our case we might use

foo = (bar+((bar>>31)&quarterpel))>>quarterpel; 

Ugly, but should work. Anyone got a simpler idea? 

gruel


> The code is still a lot more branchless. I used some LUTs for small sets
> of values like for bframe dquants or bframe mb type... even if these two
> functions  are  called  not  so  often, their  ranking  in  the  profile
> changed... it's a bit better even if it's not an impressive speedup (you
> won't notice it ;-). 
> 
> I also added cbp test to skip complete macroblocks when possible. 
> 
> Still wanting some feedback to see if the patch impacts the decoding.
> 
> -- 
> Edouard Gomez