Remove deprecated target::host_buffer

[BenBrock]

Replacing a use of deprecated target::host_buffer with target::host_task. There is one other use of host_buffer here, but it is using the internal namespace, so not sure if it should be replaced as well.

[akukanov]

I think we should instead use host_accessor here. @MikeDvorskiy please take a look.

[danhoeflinger]

I agree that sycl::host_accessor should be used as the return type.

Also, the sycl buffer member function used to get the return value, get_access<mode>(), is also deprecated in SYCL 2020 and get_host_access() should be used instead. These should change together.

[akukanov]

Requested a few more developers to look at this.

[BenBrock]

I've updated the PR to use sycl::host_accessor instead. I quickly searched through the codebase for any usage of get_mode<..., target...>, but didn't find any.

[danhoeflinger]

I've updated the PR to use sycl::host_accessor instead. I quickly searched through the codebase for any usage of get_mode<..., target...>, but didn't find any.

To be more specific because there are many get_access overloads...
From the SYCL spec for the buffer class:

template <access_mode Mode>
accessor<T, Dimensions, Mode, target::host_buffer> get_access()

"Deprecated in SYCL 2020. Use get_host_access() instead.
Returns a valid host accessor to the buffer with the specified access mode and target."

---

The instance I was referring to is right after your change, include/oneapi/dpl/internal/iterator_impl.h: line 53.

The function which makes use of this accessor_type should be changed.

    static accessor_type
    get(oneapi::dpl::__internal::sycl_iterator<Mode, T, Allocator>& iter)
    {
        return iter.get_buffer().template get_access<mode>();  // <-- here
    }

I think there are some instances elsewhere as well in include/oneapi/dpl/internal/function.h (a helper function which may not be used anymore) and in a test test/xpu_api/functional/test_functional.cpp.

[akukanov]

The possible problem with trying to change get_access to get_host_access now is that the latter cannot be parameterized by the mode, instead it uses deduction tags as function arguments. So more changes will likely be needed in the class, and etc. And I am not even sure that this class is needed in the long term, as we now re-think the approach for buffer access modes in our pseudo-iterator class.

So I suggest to either accept the current one-line change in this PR, or postpone it completely until we figure out how we deal with access modes in general.

[danhoeflinger]

So I suggest to either accept the current one-line change in this PR, or postpone it completely until we figure out how we deal with access modes in general.

Its a fair point, I hadn't fully considered the implications of how this will work in context. However, I believe these changes must be done together. To return a sycl::host_accessor<T,dim,mode> from this function, we need to use get_host_access(), as there is no conversion from the result of get_access<mode>() (the old return type) to a sycl::host_accessor<T,dim,mode>. Testing this briefly as a one-line change results in an error.

Therefore, I think its best to postpone this. As you mention, this struct may not be needed in general. As far as I can find, I don't believe it is used anywhere in the repository (and hasn't been in some time).

[BenBrock]

Closing this PR.