[MPlayer-dev-eng] [PATCH] more swscale altivec fixes

Alan Curry pacman at theworld.com
Tue Apr 25 04:00:29 CEST 2006


Michael Niedermayer writes the following:
>> 
>> At least 3 more. There are 2 cases of assuming stride==width and 1 case of
>> expecting the Y plane to be aligned at 16 bytes. They're independent bugs,
>> but they all appear in the DEFCSP420_CVT macro which defines the functions
>> altivec_yuv2_{argb,bgra,rgb24,...}.
>
>do we have a swscale-altivec maintainer? if no, then apply it, but with
>special case(s) for the correctly aligned case so that the code isnt slowed
>down

Special case as in a whole new set of scaler functions? Or adding an extra
copy of the main loop with a single test being done every time the function
is called?

The current version is neither fully optimized for aligned input nor fully
correct for unaligned input, but halfway in between. Only the Y plane is
assumed to be aligned; the U and V planes are loaded and shifted. There are 3
independent assumptions that could be made there, so I could make 8 versions
of the same loop. I assume nobody would want to see that.




More information about the MPlayer-dev-eng mailing list