EXA and MigrationHeuristic

[Steven J Newbury]

On Thu, 2008-05-29 at 12:12 +0200, Michel Dänzer wrote:
> On Thu, 2008-05-29 at 12:10 +0200, Lukas Hejtmanek wrote:
> > On Thu, May 29, 2008 at 12:06:29PM +0200, Michel Dänzer wrote:
> > > "greedy" is faster than "always" in some cases because it's quite
> > > conservative about migrating pixmaps offscreen. Any single low level
> > > operation can only be accelerated if all pixmaps involved are offscreen,
> > > so this results in fewer operations being accelerated, thereby avoiding
> > > migration overhead for operations that can't be accelerated for other
> > > reasons.
> > 
> > hmm what about shared memory graphics (such as Intel) and migration from video
> > memory to system memory. I guess this kind of migration could be avoided at
> > all as CPU should be equally fast with both the "video" memory and the system
> > memory. Or is there any reason why to memcpy pixmaps from vram to ram in
> > shared graphics case?
> 
> No, this is eliminated on the intel batchbuffer branch. The whole EXA
> migration code is basically inactive there.
> 
I never managed to get the intel-batchbuffer branch to be stable on my
GM965 but it was very fast until it crashed.  Anychance of seeing some
of the improvements hitting master (or drm-gem?) anytime soon?