[SYCL] Postpone creation of HostKernel copy

[Alexandr-Konovalov]

Create possibility to delay creation of copy of HostKernel till it became used out of submit stack, i.e. by scheduler. Do type erasure for kernel lambda via vptr in HostKernelRefBase.

[aelovikov-intel]

I think this can be empty. HostKernelRef<...> instantiates HostKernel<...> on line 256, and its InstantiateKernelOnHost already does the right thing (outside preview). And for preview we need a mechanism that doesn't require copy-paste.

[Alexandr-Konovalov]

It seems it can be empty. I try to describe the reason in comment.

And for preview we need a mechanism that doesn't require copy-paste.

Are we talking about GetInstantiateKernelOnHostPtr() call, right? It required template parameter, so unclear what can we done other then adding the call to templated constructor.

[aelovikov-intel]

IMO, this should std::unique_ptr because it has no overhead and one can always easily create shared via unique_ptr::release.

[Alexandr-Konovalov]

I'm not totally agree. Overhead is the creation/destructor of unique_ptr, meanwhile caller needs shared_ptr. (And for shared_ptr we return 2 pointers vs 1 for unique_ptr, so it's hard to judge). Is that a chance that someday caller would need unique_ptr?

[aelovikov-intel]

Overhead is the creation/destructor of unique_ptr

is exactly zero with optimizations enabled: https://godbolt.org/z/fcaos1Wr7

That is not true for std::shared_ptr (which not only has extra memory alloc but also involves atomics).

[Alexandr-Konovalov]

Thanks, good point. Done.

[aelovikov-intel]

I wish we could do

const HostKernelRefBase &TypeErasedKernel = HostKernerlRef<...>{KernelFunc};

(https://godbolt.org/z/h9v9s3TrG), but getPtr() isn't marked as const 😞

[vinser52]

Should we delete copy ctor here as well?

[Alexandr-Konovalov]

We need to create HostKernelRef from constant reference, as in sycl/include/sycl/queue.hpp, so we can't.

  HostKernelRef<KernelType, KernelTypeUniversalRef, TransformedArgType, Dims>
      HostKernel(std::forward<KernelTypeUniversalRef>(KernelFunc));

[vinser52]

I mean add HostKernelRef(const HostKernelRef&) = delete;

[Alexandr-Konovalov]

Can we explicitly delete copy ctor from HostKernelRefBase?