[HIP] Fix host/device synchronization

[veselypeta]

This PR fixes the entry-point urDeviceGetGlobalTimestamp:

• Initialize the global EvBase event when getting platforms
• Re-order fetching host/device times
  • Fetching the device time takes longer so fetching this first results in a more synchronized time between host/device time.

llvm-testing: intel/llvm#11669

[fabiomestre]

For sanity check, do you mind running the SuccessSynchronizedTime test in a loop 1000 times or so?

On CUDA (at least on my machine), this test was unstable (sometimes passes and sometimes fails with a large variance of the result). I struggle to understand how moving the hostTimestamp code solves the issue. Maybe it is the fact that cuEventSynchronize is called right away after cuEventRecord?

[veselypeta]

For sanity check, do you mind running the SuccessSynchronizedTime test in a loop 1000 times or so?

On CUDA (at least on my machine), this test was unstable (sometimes passes and sometimes fails with a large variance of the result). I struggle to understand how moving the hostTimestamp code solves the issue. Maybe it is the fact that cuEventSynchronize is called right away after cuEventRecord?

Yep you're right this test does fail sometimes after multiple repeated calls, so I have made it optional again. However, the reason why we take the host time last, is that to fetch the device time takes for time that for host time. If we get the host-time first, then we have to wait for hip to sync the streams and calculate the time, by which time the host clock is already old.

[fabiomestre]

For sanity check, do you mind running the SuccessSynchronizedTime test in a loop 1000 times or so?
On CUDA (at least on my machine), this test was unstable (sometimes passes and sometimes fails with a large variance of the result). I struggle to understand how moving the hostTimestamp code solves the issue. Maybe it is the fact that cuEventSynchronize is called right away after cuEventRecord?

Yep you're right this test does fail sometimes after multiple repeated calls, so I have made it optional again. However, the reason why we take the host time last, is that to fetch the device time takes for time that for host time. If we get the host-time first, then we have to wait for hip to sync the streams and calculate the time, by which time the host clock is already old.

Yeah, I think it makes sense for the time to be a bit more synchronized this way. It just didn't completely justify the variance that I saw on CUDA. I think there is something else going on that makes this entrypoint unreliable. My guess is that it's related to the operating system.

Anyway, this looks good to me now 👍

[npmiller]

Note that for CUDA we moved the EvBase to the device instead, I'm not 100% sure if HIP behaves the same but it could make sense to mirror that change for HIP as well, and it might also address this issue:

• [SYCL][CUDA] Move base event into the device intel/llvm#8312

[veselypeta]

Note that for CUDA we moved the EvBase to the device instead, I'm not 100% sure if HIP behaves the same but it could make sense to mirror that change for HIP as well, and it might also address this issue:

* [[SYCL][CUDA] Move base event into the device intel/llvm#8312](https://github.com/intel/llvm/pull/8312)

Thanks @npmiller - I won't make this change as part of this PR, but I've created #1007 to make the change

[veselypeta]

@oneapi-src/unified-runtime-hip-write can you please review?

[veselypeta]

Covered in #1030