[cairo] pixman 0.12.8 and later fail image transform tests

[Chris Wilson]

On Tue, 2008-11-25 at 21:50 -0800, Keith Packard wrote:
> I'm sitting here releasing pixman 0.13.2 and tested it against current
> cairo bits. I'm getting three unexpected failures:
> 
> image (argb32): 3 failed - rotate-image-surface-paint scale-offset-image
> scale-offset-similar
> 
> (same on rgb24)
> 
> It looks like the nearest neighbor point sampling code is not working
> right (the old results look much nicer than the new ones), but of
> course, I'm not sure which is correct.
> 
> I can't find any pixman version which builds with current cairo and
> which gives correct results.

The rotate-image-surface-paint is a regression as the current reference
image depends upon the pixel sampling at x.5, but after tweaking the
matrix to improve the results for scale-offset-*, the sample coordinate
has drifted slightly generating a different result. The scale-offset-*
tests are two new known failure cases where we have insufficient
precision in the matrices to handle the transformation. According to
Carl, I should have made them XFAIL to prevent exactly this
misinterpretation of the results - but that currently hides the failures
making it harder to track how changes to the code fare.

If those are your only failures, it sounds like (as far as the limited
testing Cairo provides) pixman is good to go.
-- 
Chris