Dan Hoeflinger

Rebased to pick up bugfix for legacy range implementation.

After some investigation of the original error, it seems that this PR is more an "avoidance" of an error than a direct fix (#1445 is more direct). Also, I have...

After a discussion offline with @rarutyun , it was decided to not go forward with this patch unless we find a specific use case we are fixing. Technically, this changes...

@akukanov > This patch is interesting but needs some more work. In particular, I think the semantics of transform_if should follow that of the regular transform algorithm, just extending it...

> I believe that we dont need up to 90% of these changes. Because a `transform_if` semantic is trivially implementing by a call `transform` algorithm with composition a passed `transform...

> I see... But in any case you can re-use oneapi::dpl::__internal::__pattern_walk2 and oneapi::dpl::__internal::__pattern_walk2 by passing special functor with conditional assignment. So, it is enough to define a couple of functions...

> Just "write" means read as well. I clarified it some time ago... Dont remember which way... probably, SYCL spec + discussion with guys who are responsible for SYCL language....

> other idea = to try just transform with conditinal_transform functor and output_transform iterator with std::ignore We don't have an output_transform iterator in oneDPL, currently only in SYCLomatic headers. Also,...

The important takeaways are this: - `write` and `read_write` are very similar or the same for SYCL kernel optimization. - We currently don't match SYCL spec with regard to host_iterators...

After some offline discussion with @MikeDvorskiy and @rarutyun, it has been decided that this PR will be delayed until after a discussion of iterators from std::vector (host iterators). The use...