Barriers give inconsistent results on NVIDIA backend.
Job #1666 failed for 23274e4d:
While the pipelines on AMDGPU, and POCL are both successful for both GPU acceleration and CPU based operation it would seem NVIDIA GPU accelerated implementations fail due to inconsistencies in the results for the Varrier Tests. See the link above for an example of the failing job. Interesting enough when run on NVIDIA in CPU mode it does seem to pass the tests, or at the very least the pipeline isnt failing (it is possible the tests dont run on CPU as many of our tests dont).
Below are the specific tests that fail along with the results:
BarrierSupportTest.testBarrier1:83->validate:210 targetArray: arrays first differed at element [0]; expected:<1048575> but was:<1023> BarrierSupportTest.testBarrier1Explicit:100->validate:210 targetArray: arrays first differed at element [0]; expected:<1048575> but was:<1023> BarrierSupportTest.testBarrier2:135->validate:210 targetArray: arrays first differed at element [0]; expected:<1048575> but was:<1023> BarrierSupportTest.testBarrier2Explicit:152->validate:210 targetArray: arrays first differed at element [0]; expected:<1048575> but was:<1023>