[SYCL] Add scheduler-bypass for handler-less kernel submission path

[slawekptak]

The handler-less kernel submission path has been extended to support the fast, scheduler-bypass submission.

[vinser52]

@slawekptak In your first PR #19294 the main goal was to introduce interfaces with some minimal implementation. This PR is only about implementation. Should we start unifying the handler and no-handler implementations?

[slawekptak]

@slawekptak In your first PR #19294 the main goal was to introduce interfaces with some minimal implementation. This PR is only about implementation. Should we start unifying the handler and no-handler implementations?

Sure - my plan is to introduce a series of new PRs right after this one, where the no-handler APIs are extended. This one seemed urgent to me, since we can start optimizing the HostKernel allocation, and we can run the performance benchmarks.

[vinser52]

@slawekptak In your first PR #19294 the main goal was to introduce interfaces with some minimal implementation. This PR is only about implementation. Should we start unifying the handler and no-handler implementations?

Sure - my plan is to introduce a series of new PRs right after this one, where the no-handler APIs are extended. This one seemed urgent to me, since we can start optimizing the HostKernel allocation, and we can run the performance benchmarks.

To test no-handler path in perf CI we must enable it by default (I mean, remove the __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT macros). Did you say that you are going to do that in the next PR?

Also, I did not understand which new APIs you meant. The Submitkernel benchmark uses sycl::ext::oneapi::experimental::nd_launch and this API would be enabled with no-handler path after we remove the __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT macros, right? Do we need any other APIs now?

I suggest focusing on the implementation and removing the __DPCPP_ENABLE_UNFINISHED_NO_CGH_SUBMIT macros (it means that the no-handler path should be mature enough).

[uditagarwal97]

LGTM overall. Just had a few NITs

[github-actions]

@intel/llvm-gatekeepers please consider merging

[github-actions]

@intel/llvm-gatekeepers please consider merging