[MPlayer-dev-eng] New Codec Interface

[D Richard Felker III]

On Wed, Feb 27, 2002 at 01:12:41PM +0100, Michael Niedermayer wrote:
> Hi
> 
> On Wednesday 27 February 2002 12:48, Arpi wrote:
> [...]
> > > I have a question about direct rendering (see A'rpi mail in DECODING
> > > AHEAD thread): if a decoder uses method 2 (copy to video mem after each
> > > macroblock/stride) it has to take care of color conversion: will
> > > mplayer provide a library to do it, at least for most commonly used
> > > block sizes (4x4, 8x8, 16x16)? We have rgb2rgb, but it works on a
> > > linear buffer, not on blocks.
> >
> > I hope so. Michael? :)
> imho doing a row/slice of blocks is easyer and faster, as the current cpus 
> dont have enough non temporal write buffers afaik for these blocks and for a 
> slice the data comes from L2 cache and is written to agp memory, the limiting 
> factor will be agp mem, a per block variant would read from L1 but it still 
> writes to agp mem so i wouldnt expect much speedup especially as the accesses 
> are not sequential anymore

Don't underestimate AGP memory; it's insanely fast. On a texture
mapper I once wrote, I got a 2x performance boost just passing it a
pointer to video memory rather than system memory, and that doesn't
even count the copies. Also keep in mind this was just AGP 2x, not
4x...

I'm no expert on this stuff, but I would expect per-MB writes to be
much faster than per-slice...

Rich