[BDW] dEQP-VK.synchronization.*read*_image*fence_fd are flaky
Submitted by Samuel Iglesias Gonsálvez
Assigned to Intel 3D Bugs Mailing List
Device: Mesa DRI Intel(R) HD Graphics 5500 (Broadwell GT2) (0x1616) Mesa HEAD: 767ca5bd (master) VK-GL-CTS HEAD: ab6b0ac1824ce0b091830042110cf3b3ea17e895 (master)
$ ./deqp-vk -n dEQP-VK.synchronization.read_image*fence_fd
How to reproduce it:
I run these in my updated Fedora GNOME 26 installation on my BDW laptop. The easiest way to reproduce it is to run them in a terminal and lock/unlock my GNOME session. Any time I do that, some tests fail. If I don't, they succeed.
Another easy way is to just run one of them in an infinite loop and repeat the aforementioned steps:
$ while(true); do ./deqp-vk -n dEQP-VK.synchronization.read_image* | grep "Passed" | grep -v "1/1" ; done
Also it is possible to reproduce it when executing a full cts run with piglit with concurrency mode enabled using dylan's wip/deqp-group-at-a-time branch. In that case, usually one or two tests fail.
- There are no GPU hangs or any other kind of error in dmesg.
- These patches from Jason https://patchwork.freedesktop.org/series/31771/ (already landed in mesa master) improved the situation a lot, however it is still possible to reproduce this bug.
- I was not able to reproduce it either on SKL or KBL.