[SYCL] Add test for sorting over sub-group

[romanovvlad]

While extending the original test did some refactoring:

1. Reduced types it tests
2. Test only 1 and 2 dimensions

[andreyfe1]

Big thanks for making such changes! I very appreciate it!
Any plans to test 3 dimensions?

[andreyfe1]

LGTM. Make sure that CI is passed

[romanovvlad]

The fail KernelAndProgram/kernel-bundle-merge-options-env.cpp is unrelated. Merging.

[npmiller]

@romanovvlad this is failing on CUDA and HIP, they don't support intel::reqd_sub_group_size so they end up running the test with the wrong sub group size which leads to incorrect results, setting ReqSubGroupSize to 32 for CUDA I got the test to pass locally, maybe we should query the device for the sub-group size instead of imposing it?

[romanovvlad]

@romanovvlad this is failing on CUDA and HIP, they don't support intel::reqd_sub_group_size so they end up running the test with the wrong sub group size which leads to incorrect results, setting ReqSubGroupSize to 32 for CUDA I got the test to pass locally, maybe we should query the device for the sub-group size instead of imposing it?

Cannot find an API for querying sub-groups size for a given kernel with a given work-group size. Am I missing something?

[npmiller]

I don't think there's a way to query for that, I think either we need to not specify it, let the device pick for us and use get_sub_group(), in the kernel.

Or there should be a way to query what the device support, info::device::sub_group_sizes maybe, I'm not sure it's properly implemented.

[romanovvlad]

I don't think there's a way to query for that, I think either we need to not specify it, let the device pick for us and use get_sub_group(), in the kernel.

This won't help as I need to know size of the sub-group before running the kernel to understand how much additional memory required.

Should we disable the test for CUDA and HIP until we find a solution?

[npmiller]

Should we disable the test for CUDA and HIP until we find a solution?

Yeah we can do that