Add missing API for parallel_for with kernel and properties

[mfrancepillois]

Adds a missing parallel_for signature when used with sycl::kernel and user-defined properties.

[mfrancepillois]

Without this added signature, this type of call goes through the generic API parallel_for(nd_range<Dims> Range, PropertiesT Properties, _KERNELFUNCPARAM(KernelFunc)) and fails on an assert in detail::CheckDeviceCopyable<KernelType>();.

[victor-eds]

Can we test this not using SYCL-RTC? Something like https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html#_kernel_introspection

[maarquitos14]

I think the signature you're adding is not exactly the same as the one the spec shows:

  template <int Dimensions>
  void parallel_for(nd_range<Dimensions> ndRange, const kernel& kernelObject);

@steffenlarsen what do you think?

[steffenlarsen]

@maarquitos14 is right, this does not match any of the SYCL 2020 variants of parallel_for nor any of the ones added in sycl_ext_oneapi_kernel_properties. This was raised at one point, but due to some of the awkwardness of only really being able to apply launch-based properties, we went with only adding new overloads to sycl_ext_oneapi_enqueue_functions. Would it be possible to use that extension for these sort of launches?

[mfrancepillois]

Thanks for your suggestion. The sycl_ext_oneapi_enqueue_functions extension does indeed addresses this issue of launching sycl::kernel with properties.
This PR can therefore be closed.