[MPlayer-dev-eng] Re: Re: PATCH [0/12] CoreAVCDecoder support

[Alexander Strange]

On Feb 20, 2007, at 9:01 AM, Michael Niedermayer wrote:

> well just take a random page of code put STRAT/STOP_TIMER around it  
> and
> simplify it, as loren already said the majority of the code in  
> h264.c is
> executed once per mb in some videos

 From what I get out of some quick profiling - cpu time and L2 cache  
miss on an Intel Core Duo/OS X - two good functions to check out are  
hl_decode_mb and its backup_mb_border.

It looks like you could decrease the active code size by having  
hl_decode_mb check for the common case (progressive mb that isn't PCM  
or svq3) and punt to a smaller function. I'll try it out and get some  
real numbers later.

Also, -O3 is faster than -Os here, but -O3 -fstrict-aliasing is  
faster still. -fstrict-aliasing is only off by default on Apple gccs,  
though.

On here, ffh264 is already so fast that my only video on hand with  
any slowdowns is 1080p :)