[Intel-gfx] Modesetting bug on 945GSE (Atom N270) with Chris Wilson drm-intel-next kernel

[Chris Wilson]

On Mon, 22 Nov 2010 10:01:25 +0000, Simon Farnsworth <simon.farnsworth at onelan.co.uk> wrote:
> On Friday 19 November 2010, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> > On Fri, 19 Nov 2010 17:44:49 +0000, Simon Farnsworth <simon.farnsworth at onelan.co.uk> wrote:
> > > Hello,
> > > 
> > > I'm seeing an interesting bug on 945GSE, and I have absolutely no idea
> > > where to start looking. I'm trying to deal with performance issues on
> > > this platform, so I'm updating to master of as many bits as possible, to
> > > see if they're better.
> > 
> > Do you use a lot of fences? If so the pipelined fencing may be of interest
> > and also the relaxed fencing rules for i945. It would be good to get
> > some feedback on those.
> > 
> I'm a dumb applications programmer :)
> 
> I'm seeing an OpenGL compositor (using TFP to get window contents for circa 10
> windows) struggle to meet frame deadlines at 60Hz. I would therefore guess
> that I do use quite a few fences, as if I've understood properly, each of the
> various window contents is accessed via a fence.

A blit will consume a fence and libdrm overestimates the number of fence
it consumes in a batch so will emit far shorter batches than ideal.

The pipelined fencing is still under testing, but you can have a
sneak-peak at
http://cgit.freedesktop.org/~ickle/linux-2.6/log/?h=direct-gtt

Just using that would be interesting wrt to the deadlines, except for the
issue below...

The usual perf cpu profiling and latency charts would be interesting.

> > > drm-intel-next as of a17577c9 (not tried older versions yet) is
> > > consistently rendering to the middle third or so of the screen only;
> > > this is happening in X, as well as in inteldrmfb during boot.
> > 
> > Vertical or horizontal offset? Which outputs? Anything suspicious in
> > intel_reg_dumper?
> > -Chris
> It's horizontal that's broken. I'm using VGA (but can test DVI if you think
> it's output-specific). I didn't spot anything weird in intel_reg_dumper, but
> there are changes:

Hmm, looks like the change is both pipes are being enabled. xrandr output?
The effect of cloning would be that only the central 1280 pixels of your
1680 display is being used. (xrandr --output VGA1 --off --output DVI1
--preferred)

The question is why we now decide to use both outputs. drm.debug=0xe
dmesg?
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre