[SYCL] piextQueueCreateWithNativeHandle should not be followed by a piQueueRetain

[smaslov-intel]

Signed-off-by: Sergey V Maslov [email protected]

[romanovvlad]

Could you please add a unit test to sycl/unittests/?

[smaslov-intel]

The fix for the failing test in precommit comes in intel/llvm-test-suite#577

[smaslov-intel]

Could you please add a unit test to sycl/unittests/?

I don't know how to write a good unit test for this. The issue was appearing as a "leak" of a queue essentially, since it's reference counter never got to 0 in this case. We have special testing (in process of being enabled by default) design to test imbalance between create/release calls with ZE_DEBUG=4

[romanovvlad]

Could you please add a unit test to sycl/unittests/?

I don't know how to write a good unit test for this. The issue was appearing as a "leak" of a queue essentially, since it's reference counter never got to 0 in this case. We have special testing (in process of being enabled by default) design to test imbalance between create/release calls with ZE_DEBUG=4

We can check that for a specific code(make_queue call) there are expected number of calls to piQueueRetain by mocking this API like in this test https://github.com/intel/llvm/blob/sycl/sycl/unittests/queue/Wait.cpp .

[smaslov-intel]

Could you please add a unit test to sycl/unittests/?

I don't know how to write a good unit test for this. The issue was appearing as a "leak" of a queue essentially, since it's reference counter never got to 0 in this case. We have special testing (in process of being enabled by default) design to test imbalance between create/release calls with ZE_DEBUG=4

We can check that for a specific code(make_queue call) there are expected number of calls to piQueueRetain by mocking this API like in this test https://github.com/intel/llvm/blob/sycl/sycl/unittests/queue/Wait.cpp .

@romanovvlad : please check

[smaslov-intel]

Just in case, the "Plugin/interop-level-zero.cpp" fail is expected with a fix coming in intel/llvm-test-suite#577

[romanovvlad]

Thanks, a couple of minor comments.

[romanovvlad]

Just in case, the "Plugin/interop-level-zero.cpp" fail is expected with a fix coming in intel/llvm-test-suite#577

Should this patches be committed simultaneously? Do they depend on each other?

[smaslov-intel]

Just in case, the "Plugin/interop-level-zero.cpp" fail is expected with a fix coming in intel/llvm-test-suite#577

Should this patches be committed simultaneously? Do they depend on each other?

The llvm-test-suite change can be committed first, the test still passes with that change. Now CI reports unrelated CUDA failures there.

[smaslov-intel]

@romanovvlad : please merge this