[cairo] potential bug in cairo relative to intel drivers

[Chris Wilson]

On Thu, 2009-07-30 at 14:02 -0700, Carl Worth wrote:
> On Thu, 2009-07-30 at 16:57 -0400, Mike Gualtieri wrote:
> > Setting NoAccel to true fixes the issue.  When it is set to false (the
> > default) the problem occurs.
> 
> Yes. :-)
> 
> If "the problem" means "blending the scaled image with the destination
> in the case of EXTEND_NONE and cairo_paint()" then that's not a bug at
> all. It's the expected behavior that EXTEND_NONE and cairo_paint() will
> give you this blending when you scale up.
> 
> So if setting NoAccel to true makes that blending go away, then yes,
> that does sound like a bug in the NoAccel case.
> 
> Does that make sense? Or am I still misunderstanding something?

Although from the description, thank you Mike for being clear and
concise, we have a good idea of exactly what bug we are looking for, we
have been developing a tool that should making this kind of bug report
easier in future. The idea is that the user runs cairo-trace on the
application and captures a trace which will contain sufficient
information to replicate the drawing instructions that exhibit the bug.
And then we can feed this trace through a tool called cairo-test-trace
which can be run against two X servers in parallel and compare the
output after every context. Once it has found a discrepancy, it will
save the images of the differing surfaces and a minimal trace to
reproduce the error. (At the moment this is likely to find a difference
in rendering long before we hit the scaling bug! :(

One of the goals for the future, is to then run a technique called
delta-debugging, which essentially performs a binary bisection on the
trace, to reduce the trace even further down to the precise minimum
required to reproduce a bug. (The only difficulty here is that to do the
bisection efficiently we need to be able to determine illegal cuts,
faster than simply executing the trace and hitting an error -- that is
to bisect the state rather than the script.)

Mike, I would appreciate if if you could find the time to download
cairo-trace either from the 1.9.2 snapshot, or from git and capture a
trace from your application:
  # cairo-trace --profile ./how-to-exhibit-bug [args]
Note that the trace will capture everything drawn, so be careful no to
include any confidential information, or anything else you might not
want to distribute.

That will produce a file called ./how-to-exhibit-bug.$pid.lzma which
will probably be too large to attach here, so upload it or send it to
me, and I can see how well our tools fare. Already I can see that we're
encouraging people to use --profile more often than not, so it might be
worth making that the default.

Thanks,
-ickle