[SYCL][NATIVECPU] Materialize floating point atomics in LLVM pass

[PietroGhg]

This PR sets SYCL_USE_NATIVE_FP_ATOMICS when compiling for Native CPU, and provides an implementation for said atomics via an LLVM pass that defines them through atomicrmw instructions.

[maarquitos14]

What about fetch_sub? I can see the following in SYCL spec for atomic_ref:

Floating fetch_sub(Floating operand,
                     memory_order order = default_read_modify_write_order,
                     memory_scope scope = default_scope) const noexcept;

[PietroGhg]

What about fetch_sub? I can see the following in SYCL spec for atomic_ref:

Floating fetch_sub(Floating operand,
                     memory_order order = default_read_modify_write_order,
                     memory_scope scope = default_scope) const noexcept;

Thanks @maarquitos14 , it appears to be implemented as atomicfadd(-operand) see atomic_ref.hpp

[maarquitos14]

LGTM.

[maarquitos14]

Would it be possible to add e2e test to also check that we get the expected result?

[elizabethandrews]

Can you add a FE test? I assume there is already a FE test checking the existence of this macro for other targets. You can just add a RUN there.

[PietroGhg]

Added it, thank you

[PietroGhg]

Would it be possible to add e2e test to also check that we get the expected result?

There are already e2e tests that cover fp atomics, currently we track the pass rate of those internally for Native CPU

[elizabethandrews]

FE LGTM. Thanks!

[PietroGhg]

@intel/llvm-gatekeepers, this looks ready to merge, thank you