[SYCL] Don't return last event in ext_oneapi_submit_barrier

[uditagarwal97]

Problem

As an optimization in ext_oneapi_submit_barrier, for in-order queue, when waitlist is empty, we return the last event, if it is available. However, after #18277, we no longer store last event and instead submit a marker to get the last event. And this caused performance regression in some of our internal benchmarks.

Solution
Don't return last event and instead submit a barrier with empty waitlist.

[uditagarwal97]

@jbrodman @vinser52 fyi.

[steffenlarsen]

I am a little surprised that a marker event would be slower than actually submitting a barrier. Of course, the adapters/backends could be optimizing the barriers on in-order queues themselves, i.e. by storing the latest event themselves or by submitting markers internally. As such, I am all for dropping the complexity from the runtime.

[againull]

I am a little surprised that a marker event would be slower than actually submitting a barrier. Of course, the adapters/backends could be optimizing the barriers on in-order queues themselves, i.e. by storing the latest event themselves or by submitting markers internally. As such, I am all for dropping the complexity from the runtime.

Yes, L0 adapter has the same optimization, that's why removing optimization from sycl level restores performance, because markers are not inserted, but optimization is still in place at adapter level.
Optimization was introduced at sycl level to have same benefit for all backends.