Created attachment 28414 [details]
Bug detailed description:
Some rendercheck cases(fill,blend,repeat,triangles) failed because of adding support of render BGRA formats now, which used to be not supported. It also fails on G45.
2. run rendercheck -t fill
With commit ce1fe8ddb4a4dbe6cfd909e5b1b73b459d742bec in xserver master branch, it disables b8r8g8x8 in depth 24. rendercheck works well with (fill, repeat, triangles) now. But blend still fails.
Author: Michel D?nzer <email@example.com>
Date: Wed Sep 23 13:10:05 2009 +0200
render: Don't add b8g8r8x8 format for depth 24.
The components are required to be packed in the bottom of the pixel, so this
format can't fit in depth 24.
Also fix up a comment for the addition of BGRA formats.
(In reply to comment #1)
> With commit ce1fe8ddb4a4dbe6cfd909e5b1b73b459d742bec in xserver master branch,
> it disables b8r8g8x8 in depth 24. rendercheck works well with (fill, repeat,
> triangles) now. But blend still fails.
Do you have rendercheck logs for the blend failures? Are the failures produced by hardware or software rendering? If the latter, you may also want to try a newer pixman version.
The issue lies with pixman. I've ported rendercheck into the pixman test suite and Soeren has begun fixing the failing cases, so the bug is in progress. ;-)
Eric, why did you reassign this to me? Reassigning per comment #3.
With pixman 0.16.2 or newer, do you still see these problems?
(In reply to comment #5)
> With pixman 0.16.2 or newer, do you still see these problems?
Yes, we tested with pixman always updated, now it is 0.17.1 and blend still fails.
On the nouveau driver, which uses pixman to render into a shadow framebuffer, "rendercheck -t blend" passes without issue, so any further bugs here must be in the Intel driver.
Chris might have some good ideas on this one.
(And, Chris! It's not a bug with strange corruption!)
Had a good long battle with the driver and rendercheck, and I think I have corrected all the bugs I've gradually introduced since the reduction in signal-to-noise in rendercheck from the additional formats.
In short rendercheck is now much happier on i915, and I believe this should equally extend to i830 and i965. Now to check that I have not just caused everything to fallback to software...
rendercheck passed here.
Yes, with the newest code, it now passed on both Pinetrail, G45 and Piketon. So verified.