[XviD-devel] zero-biased dequant/idct

[daniel smith]

> I doubt it. Speeding up C code and speeding up 
MMX
> assembler are two different pairs of shoes... 
> One branch prediction might stall everything 
longer than 
> 50 extra MMX instruction. 
> 
> But please test, if you have the time.

tests concluded - zero bias is a no-go.  idct_zero 
was never faster than normal idct ,even when 
given a zeroed array.  each took 18000 +- 50 
milliseconds per 10,000,000 calls.

dan
-- 
_______________________________________________
Get your free email from http://www.astroboymail.com

Powered by Outblaze