Possible bug with arrays and out-of-order queues
As evidenced by this CI job failure:
Job #88646 failed for 1dd0183d.
Trying to reproduce on koelsch:
i=0; while PYOPENCL_TEST=int python test_array.py 'test_outoforderqueue_indexing(cl._csc)'; do echo $i; i=$((i+1)); done
seems to show that this seems to be a bug that's not hit with a high probability. (i=185 in one trial)
One principled way to help debug stuff like this would be to build instrumentation that can draw a dependency DAG based on CL functions called. That probably wouldn't be hard to monkey-patch into place.
Edited by Andreas Klöckner