Add default nd_range constructor

[KornevNikita]

Extension Type

Improvement

Feature Description

The range and id classes have default constructor. It's possible that user is not aware of the required globalSize and `localSize, so it's not possible to construct the range using the existing constructor:

  // The offset is deprecated in SYCL 2020.
  nd_range(range<Dimensions> globalSize, range<Dimensions> localSize,
           id<Dimensions> offset = id<Dimensions>()) noexcept;

Therefore I suggest to add:

nd_range() noexcept;

Related Functionality in C++

No response

Related Functionality in Other Languages

No response

Related SYCL Extensions

No response

[KornevNikita]

@gmlueck we had this in intel/llvm, but I removed it recently to align with SYCL 2020. Now some customers complain about this, as they use this constructor in their code. I believe this should be a part of the spec.

[gmlueck]

The default range constructor creates a range with 0 in each dimension. Would the default nd_range constructor create an nd-range with 0 in each dimension for both the global and local ranges? (I.e. two default-constructed ranges?)

[KornevNikita]

Yes, if we'd proceed with adding it we could use something like this as the description:

Construct an nd_range with the value 0 for each dimension of the global and local ranges.

[gmlueck]

I added the "Agenda" label to this because we have customers asking for this functionality. @rolandschulz, I think the request is coming from your team (CMPLRLLVM-77706).

It seems reasonable to me that we would add this as a clarification to SYCL 2020. We made a similar change already for range in #453. I'm not sure why we didn't add a default constructor for nd_range at the same time -- it seems like it has all the same issues. I think the motivation for adding a default constructor is that user code may wrap an nd_range in some other struct like:

template <int dimensions>
struct my_nd_range_holder {  sycl::nd_range<dimensions> ndr; };

It's awkward for the user to write a default constructor for my_nd_range_holder unless nd_range also has a default constructor. (See #90 for more details.)

[TApplencourt]

I will put that on the Thursday agenda then :)

[KornevNikita]

@gmlueck @TApplencourt any decision so far?

[gmlueck]

Unfortunately, we did not have time to discuss this yesterday. @TApplencourt, it would be good to discuss this next week. I think Nikita is quite keen to get an answer from us.

In the meantime, maybe we can make progress on this offline. Does anyone object to adding a default nd_range constructor? I can't think of any reason why we added a default range constructor but not a default nd_range constructor.

[TApplencourt]

Talking about Thomas:

#90 doesn't really give me warm feeling about re-opening this can of worm if too much details : )

If we bite the bullet forrange, it make sense to do it for nd_range too IMO.

[gmlueck]

If we bite the bullet forrange, it make sense to do it for nd_range too IMO.

We already bit the bullet for range. This was done in #453.

I think there are no complicated issues to discuss. An nd_range just contains two range components and an id (the optional offset). If we add a default nd_range constructor, I assume it would just default-construct each of those things. Isn't that the only behavior that makes sense? We have already defined default constructors for both range and id.

[TApplencourt]

Sorry, yes typos. I wanted to said as we bitten.

[TApplencourt]

• PR welcome to add nd_range default constructor.
• Need to add CTS tests : )
• Clarification in SYCL 2020

---

Thomas will do PR for adaptive cpp.

[KornevNikita]

I'll create a PR for the spec and tests.

#1046
KhronosGroup/SYCL-CTS#1231