Add complete set of resource ref aliases

[nvdbaranec]

Adds new aliases to resource_ref.hpp

using host_resource_ref = cuda::mr::resource_ref<cuda::mr::host_accessible>;

using host_async_resource_ref = cuda::mr::async_resource_ref<cuda::mr::host_accessible>;

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[harrism]

Only requesting changes so we can discuss question below.

[harrism]

I think it will be common for users of async_resource_ref for pinned host memory to also want device_accessible. (I expect you will want it in your cuDF usage eventually too...) My question is, do we want to just add cuda::mr::device_accessible to this alias, or add another one:

using host_device_async_resource_ref = 
  cuda::mr::async_resource_ref<cuda::mr::host_accessible, cuda::mr::device_accessible>;

Thoughts @nvdbaranec @Missco @jrhemstad ?

[miscco]

I would not add device_accessible to the host_async_resource_ref as that would all potential memory resource to satisfy device_accessible

On the other hand an async resource has no real value without device_accessible

[harrism]

Good point. However I think an async resource ref that is host-only has value because it specifies that it will only be used on the host, but it may still be backed by pinned memory. And then it benefits from stream-ordered allocation (e.g. can be used in a stream-ordered pool) and also can be used as the host memory for SOL cudaMemcpyAsync.

So really for completeness I think we probably need the full set:

using host_resource_ref = cuda::mr::resource_ref<cuda::mr::host_accessible>;
using device_resource_ref = cuda::mr::resource_ref<cuda::mr::device_accessible>;
using host_device_resource_ref = cuda::mr::resource_ref<cuda::mr::host_accessible, cuda::mr::device_accessible>;
using host_async_resource_ref = cuda::mr::async_resource_ref<cuda::mr::host_accessible>;
using device_async_resource_ref = cuda::mr::async_resource_ref<cuda::mr::device_accessible>;
using host_device_async_resource_ref = cuda::mr::async_resource_ref<cuda::mr::host_accessible, cuda::mr::device_accessible>;

[nvdbaranec]

The use case for this right now in cudf is strictly in the hostdevice_vector class, whose sole purpose is to force explicit host<->device copying onto the user. That is: entirely avoiding the device touching pinned memory directly. So in that sense, stream ordering at the allocator level isn't necessary.

https://github.com/rapidsai/cudf/blob/branch-24.04/cpp/src/io/utilities/hostdevice_vector.hpp

As for whether cudf at large will ever need this, I'm not sure. At the moment it definitely prefers to be explicit about all host<-> device handoffs.

[harrism]

Opened #1480 to cover this work.

[harrism]

Thank you @nvdbaranec

[bdice]

/merge