[Mesa-dev] [PATCH 1/2] Revert "i965: Implement guardband clipping on Sandybridge."

[Oliver McFadden]

On Wed, Oct 03, 2012 at 11:37:33AM +0300, Oliver McFadden wrote:
> On Tue, Oct 02, 2012 at 06:47:49PM -0700, Kenneth Graunke wrote:
> > On 10/02/2012 04:00 AM, Oliver McFadden wrote:
> > > This reverts commit 85cd30406ff8414bc6ac6de1064d059a4c564160.
> > > 
> > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=55523
> > > Signed-off-by: Oliver McFadden <oliver.mcfadden at linux.intel.com>
> > > CC: Kenneth Graunke <kenneth at whitecape.org>
> > > ---
> > >  src/mesa/drivers/dri/i965/gen6_clip_state.c     |    1 -
> > >  src/mesa/drivers/dri/i965/gen6_viewport_state.c |   24 +++++++++-------------
> > >  2 files changed, 10 insertions(+), 15 deletions(-)
> > 
> > Oliver,
> > 
> > I'm fine with reverting these patches on the 9.0 branch...especially
> > given how close we are to a release.  I'd like to leave them on master
> > for a little while and do some investigation.  If we can't solve it, we
> > can revert it...but if it ends up being something easy, I'd rather just
> > fix it...
> 
> Okay.  I am unfamiliar with the process for applying patches onto the
> release branches...  May I simply push the revert patches there, with
> your Reviewed-by (which I assume you have given implicitly,) or must I
> involve Brian?
> 
> > 
> > Since I can't reproduce this yet, I left a comment on the bug with an
> > experiment you might want to try.
> 
> Yes I saw that this morning.  I will try to run the test now.  I can
> provide an APItrace dump however it will be a few hundred megabytes; I
> can host it on my server or on people.freedesktop.org.
> 
> I will reply to the bug once this is done, and after checking your
> suggestions.

Unfortunately reducing maximum_post_clamp_delta had little effect on the
bug; maybe it made some polygons less likely to be guard band clipped
incorrectly, but ultimately the results were the same: entire polygons
would be incorrectly clipped from the scene or partially clipped.

I've uploaded two APItrace files at
http://people.freedesktop.org/~omcfadde/

The names are fairly self explanatory:
dante_9.1-devel_git-336cc64.trace.bz2 is vanilla Mesa at Git 336cc64,
and dante_maximum_post_clamp_delta_8000.trace.bz2 is the same with the
suggested modification.

I compressed them with bzip2 since it reduces each trace from ~500 MB to
~190 MB.  You will need to uncompress them for use with the APItrace
tools.

> 
> -- 
> Oliver McFadden.
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev

-- 
Oliver McFadden.