Re: XF4 on Mach64 and 2.2 kernel?
On Tue, Dec 05, 2000 at 05:31:54PM +0100, Michel Dänzer wrote:
> Sven LUTHER wrote:
> >
> > On Mon, Dec 04, 2000 at 04:29:24PM +0100, Michel Dänzer wrote:
> > > Sven LUTHER wrote:
> > >
> > > > > > Adam C Powell IV wrote:
> > > > >
> > > > > Hmm, given your warning, maybe I'll stick to fbdev for now. It's fast
> > > > > enough for everything- except hashing the screen during a GNOME
> > > > > logout, but I don't do that too often. :-)
> > > >
> > > > Not sure if here is the right place for this, but i noticed that
> > > > supplying the clipping pseudo accel will improve this a lot. Is there
> > > > any chance that this accel could be supported by the fbdev driver ?
> > > > It would need support from the fbdev, isn't it ?
> > >
> > > I'm not convinced at all. For now, fbdev doesn't use XAA at all, so I
> > > can't imagine how just enabling it for nothing (?) could help any.
> > > However, this is one of the rare cases where I'd really like you to go
> > > ahead and prove me wrong. :)
> >
> > Did you see the other mail talking about the gnome logout being slow when
> > shadowfb was activated, are you able to confirm this ? if yes, then it is a
> > problem with shadowfb. Now i remember that when i tried the just clipping
> > accel driver, i had to use accel, which disables shadowfb. Maybe this was
> > the source of the problem.
>
> Sounds very plausible to me.
>
> > I don't know shadowfb enough to know how it works, but i guess you write to
> > the shadowfb, which in turn get copied to the true screen every now and
> > then.
>
> Yes. The ShadowFB we are talking about does the copy after each operation.
> Keith Packard wrote another shadow framebuffer implementation for Tiny X which
> only does it in regular intervals which should further improve performance at
> the cost of latency, don't know if it will (or even can) be integrated into
> the 'normal' XFree86 server.
mmm, ...
don't know if this would really help here, especially if you want to retain
the visual aspect of it.
> > I guess what happens is that the gnome logout screen does some very small
> > writing to the shadowfb, and it has to be copied (all of the shadow buffer
> > ?) to the screen after each line draw or so.
> >
> > Am i wrong in this ?
>
> Yes and no. :)
>
> GNOME draws stippled rectangles, which leads to the whole area of the
> rectangles being copied from the shadow framebuffer to the real one. If it
> drew the rectangles using four lines instead, it should be much faster.
Or if we have some magic in shadowfb that know whatdid change and what not,
but i think this would be expensive, well not in bus cycle, but in memory
accesses.
So no real way out of this, maybe we should fill a bug against gnome ?
Friendly,
sven Luther
Reply to: