Annotated i965_prepare_composite ready for optimization

[Wang Zhenyu]

On 2007.06.22 12:35:28 +0000, Michel Dänzer wrote:
> On Thu, 2007-06-21 at 15:47 -0700, Carl Worth wrote:
> > As feedback to a recent blog post[*] I made showing
> > i965_prepare_composite to be a bottleneck in a Mozilla rendering
> > benchmark with an i965 and with EXA, I received a kind message cluing
> > me into opannotate, (thanks Roland!).
> > 
> > Here's a new entry explaining what I learned:
> > 
> > 	http://www.cworth.org/exa/opannotate_i965_prepare_composite/
> > 
> > and here are the annotated source and assembly versions of the
> > functions of interest:
> > 
> > 	http://www.cworth.org/exa/i965_prepare_composite.source_annotate
> > 
> > 	http://www.cworth.org/exa/i965_prepare_composite.assembly_annotate
> > 
> > I'd greatly appreciate any ideas for improving the performance of this
> > function.

Great thanks Carl! I did i965 exa code for correctly rendering first, so
no optimization has been done.

> 
> The state memory is probably in the AGP aperture, so uncacheable. It
> should also be write-combined though, so an easy improvement could be to
> initialize a malloced copy and then memcpy it to the final destination
> in one go.
> 

I think we can assign multiple buffers for all types of render composite,
and each time buffer change will be small just includes buffer infos, blend
op, etc. And we need consistent shader programs.