Do not execute tests using EventHostManualTrigger if not supported

[BenjaminW3]

Adds a trait and a test method isEventHostManualTriggerSupported to not use the EventHostManualTrigger when it is not supported. It was always supported with CUDA 8.0 but is only conditionally enabled in CUDA 9.
This fixes the failing tests due to this disabled feature in #537 by checking the trait before executing the test.

[psychocoderHPC]

Why is it only conditional enabled since CUDA 9. What is the reason for it?

[BenjaminW3]

I do not know. I have not yet found any reasoning.

[BenjaminW3]

When you are looking for other usages of CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS in github you will find two occurences where they do the same. I have had a look at those places before opening the PR.

[tdd11235813]

There is still the error. I try to investigate this, why the test is executed.

$ test/unit/event/event

Running 20 test cases...
unknown location(0): fatal error: in "event/eventTestShouldBeFalseWhileInQueueAndTrueAfterBeingProcessed<std__tuple<alpaka__dev__DevCudaRt,_alpaka__queue__QueueCudaRtAsync>>": std::runtime_error: /home/matwerne/cuda-workspace/alpaka/test/common/include/alpaka/test/event/EventHostManualTrigger.hpp(649) cuStreamWaitValue32( static_cast<CUstream>(queue.m_spQueueImpl->m_CudaQueue), reinterpret_cast<CUdeviceptr>(event.m_spEventImpl->m_devMem), 0x01010101u, CU_STREAM_WAIT_VALUE_GEQ) : 'unrecognized error code': 'unrecognized error code'!
/home/matwerne/cuda-workspace/alpaka/test/unit/event/src//EventTest.cpp(70): last checkpoint: "eventTestShouldBeFalseWhileInQueueAndTrueAfterBeingProcessed" entry.
unknown location(0): fatal error: in "event/eventReEnqueueShouldBePossibleIfNobodyWaitsFor<std__tuple<alpaka__dev__DevCudaRt,_alpaka__queue__QueueCudaRtAsync>>": std::runtime_error: /home/matwerne/cuda-workspace/alpaka/test/common/include/alpaka/test/event/EventHostManualTrigger.hpp(649) cuStreamWaitValue32( static_cast<CUstream>(queue.m_spQueueImpl->m_CudaQueue), reinterpret_cast<CUdeviceptr>(event.m_spEventImpl->m_devMem), 0x01010101u, CU_STREAM_WAIT_VALUE_GEQ) : 'unrecognized error code': 'unrecognized error code'!
/home/matwerne/cuda-workspace/alpaka/test/unit/event/src//EventTest.cpp(115): last checkpoint: "eventReEnqueueShouldBePossibleIfNobodyWaitsFor" entry.
unknown location(0): fatal error: in "event/eventReEnqueueShouldBePossibleIfSomeoneWaitsFor<std__tuple<alpaka__dev__DevCudaRt,_alpaka__queue__QueueCudaRtAsync>>": std::runtime_error: /home/matwerne/cuda-workspace/alpaka/test/common/include/alpaka/test/event/EventHostManualTrigger.hpp(649) cuStreamWaitValue32( static_cast<CUstream>(queue.m_spQueueImpl->m_CudaQueue), reinterpret_cast<CUdeviceptr>(event.m_spEventImpl->m_devMem), 0x01010101u, CU_STREAM_WAIT_VALUE_GEQ) : 'unrecognized error code': 'unrecognized error code'!
/home/matwerne/cuda-workspace/alpaka/test/unit/event/src//EventTest.cpp(176): last checkpoint: "eventReEnqueueShouldBePossibleIfSomeoneWaitsFor" entry.
unknown location(0): fatal error: in "event/waitForEventThatAlreadyFinishedShouldBeSkipped<std__tuple<alpaka__dev__DevCudaRt,_alpaka__queue__QueueCudaRtAsync>>": std::runtime_error: /home/matwerne/cuda-workspace/alpaka/test/common/include/alpaka/test/event/EventHostManualTrigger.hpp(649) cuStreamWaitValue32( static_cast<CUstream>(queue.m_spQueueImpl->m_CudaQueue), reinterpret_cast<CUdeviceptr>(event.m_spEventImpl->m_devMem), 0x01010101u, CU_STREAM_WAIT_VALUE_GEQ) : 'unrecognized error code': 'unrecognized error code'!
/home/matwerne/cuda-workspace/alpaka/test/unit/event/src//EventTest.cpp(254): last checkpoint: "waitForEventThatAlreadyFinishedShouldBeSkipped" entry.

*** 4 failures are detected in the test module "event"

[tdd11235813]

it is only an issue on CUDA 8. CUDA 9 does not run the test, as stream memory operations are not supported (result is 0 by querying CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS).
However, CUDA 8 does not have this attribute and if stream mem ops are supported on CUDA 8 by default, the issue is somewhere else ...

[BenjaminW3]

Maybe they are not supported by default on CUDA 8? Maybe the flag was simply missing and it was expected to work everywhere but it did not so they introduced the new flag?

[tdd11235813]

I agree with you. It is not really supported in CUDA 8, so that test can be disabled in this case!?

/alpaka/test/common/include/alpaka/test/event/EventHostManualTrigger.hpp(588) cuStreamWaitValue32( static_cast(queue.m_spQueueImpl->m_CudaQueue), reinterpret_cast(event.m_spEventImpl->m_devMem), 0x01010101u, CU_STREAM_WAIT_VALUE_GEQ) : 'CUDA_ERROR_NOT_SUPPORTED': 'operation not supported'!
/alpaka/test/unit/event/src//EventTest.cpp(232): last checkpoint: "waitForEventThatAlreadyFinishedShouldBeSkipped" entry.

(see PR #548, which fixes error reporting for driver API errors)

[BenjaminW3]

@psychocoderHPC As far as I remember, this test once worked on your system using CUDA 8. Can you confirm or reevaluate this?
If we can not be sure that this works with CUDA 8 we may have to disable the test for CUDA 8.

[BenjaminW3]

@psychocoderHPC Could you please retest this on your CUDA 8 system?
I will nevertheless deactivate the EventHostManualTrigger on CUDA 8 because there does not seem to be a way to tell if those operations are supported.

[BenjaminW3]

@tdd11235813 Now the tests should not fail anymore on your system and #537 should be finished. Please approve if it is ok for you.
I am integrating make test / ctest into the CI in another PR.

[tdd11235813]

works for me, thanks a bunch for your work!