[SYCL][NVPTX] Do not decompose SYCL functor unless necessary

[Naghasan]

CUDA backend can support passing pointer in the generic address space. The patch prevent the decomposition of the SYCL functor if there is no special types in it.

[Naghasan]

@elizabethandrews @tahonermann That's the patch I mentioned the other day, still early and missing tests obviously. But if you can spare a bit of time to give a high level feedback, that would be great. Thanks.

Note: the patch is aimed to be used with #14332

[Naghasan]

Note 2: I'm aware of the test failures (and why) and I'm also restricting to nvptx at the moment to avoid fixing a large number of tests.

[mdtoguchi]

OK for driver

[uditagarwal97]

SYCL Changes LGTM.

[victor-eds]

Kernel fusion test change LGTM

[elizabethandrews]

One thing to note here is that pointers in decomposed structs i.e. those in a struct with special types will have different address space handing than those which are in structs not decomposed. Is that discrepancy ok?

[Naghasan]

One thing to note here is that pointers in decomposed structs i.e. those in a struct with special types will have different address space handing than those which are in structs not decomposed. Is that discrepancy ok?

Yes, and it can even be important to have it. When not decomposed, pointers are basically treated as CUDA would but that only applies to global pointers. Where it is important to decompose and keep the address space is for the local one as for CUDA we need to pass an offset instead.

[Naghasan]

@intel/llvm-gatekeepers The PR is ready to merge, the windows failure is unrelated.

Fails during AOT compilation for the following reason:

Could not determine device target: cfl.
Cannot get HW Info for device cfl.

[Naghasan]

FYI @intel/dpcpp-devops-reviewers