Performance change from X in Fedora Core 4 to Fedora Core 5

[Felix Bellaby]

Carsten Haitzler wrote: 
> > I based my gloomy forecast on my recent experience of trying to allocate
> > and deallocate GLXPixmaps on nVidia hardware. Even when the new
> > GLXPixmap references exactly the same Pixmap (which remains X allocated)
> > this process takes longer than copying the Pixmap. I can only presume
> > that this is a synchronisation problem, but I do not really know.
> 
> i suspect it is a sync and lock/unlokc/round trip performance issue. alloc and
> de-alloc of 2d pixmaps inside of x is fast.

I observed the problems in a indirect rendered compositor which held a
server grab during all its drawing operations regardless of whether the
GLXPixmaps were being realloced or not. The performance loss was
spectacular, slowing glxgears and cairogears by 66% under the
compositor.

I would need to prototype some simple tests to gauge the effect in less
unusual circumstances. It might be down to how well the nvidia GLX
implementation copes inside a server grab. Otherwise a GL specific sync
problem may be responsible.

Felix