P2299R1
mdspan and CTAD

Published Proposal,

Author:
(NVIDIA)
Source:
GitHub
Issue Tracking:
GitHub
Project:
ISO/IEC JTC1/SC22/WG21 14882: Programming Language — C++
Audience:
WG21

1. Introduction

[P0009R10]'s mdspan is a convenience template alias for simpler use cases of basic_mdspan:

template <class ElementType, size_t... Extents>
using mdspan = basic_mdspan<ElementType, extents<Extents...>>;

In the basic_mdspan/span interface, extents can be either static, e.g. expressed at compile time:

mdspan<double, 64, 64> a(data);

or dynamic, e.g. expressed at run time:

mdspan<double, dynamic_extent, dynamic_extent> a(data, 64, 64);

You can also use a mix of the two styles:

mdspan<double, 64, dynamic_extent> a(data, 64);

2. Problem

The [P0009R10] interface style for expressing extents currently interacts poorly with Class Template Argument Deduction (CTAD). As of C++20, CTAD now works with template aliases, so you can write this:

mdspan a(data, 64, 64);

This syntax would be very nice. It would remove the need to explicitly spell out the verbose dynamic_extent when dealing with run time extents, which I believe is the common case for most users.

However, the above code does not appear to do what you might expect. This appears to instantiate mdspan<double>, and then basic_mdspan<double, extents<>>; a multi-dimensional array of rank 0. This will lead to a static assertion as basic_mdspan's dynamic extent constructor. You can see the code on Godbolt here.

3. Solutions

3.1. Make mdspan A Template Class

If mdspan was a template class, not a template alias, we could simply add a deduction guide to handle this:

template <class ElementType, class... IndexType>
explicit mdspan(ElementType*, IndexType...)
  -> mdspan<ElementType, [] (auto) constexpr
                         { return dynamic_extent; }
                         (identity<IndexType>{})...>;

However, it seems you cannot add a deduction guide for a template alias. Perhaps there is a way of formulating a deduction guide for basic_mdspan that would help here, however I could not find a way of doing this.

One way we could solve this problem would be to make mdspan a template class, not a template alias, and then give that template class such a deduction guide. You can see a sketch of this on Godbolt here.

However, making mdspan a template class instead of a template alias would introduce all sorts of other challenges, as it would become a distinct type from basic_mdspan and functions taking a basic_mdspan would not take an mdspan. Perhaps we could come up with a solution, possibly involving conversions, that would be acceptable.

3.2. Add dynamic_extents And dynamic_mdspan

We could add some helper aliases or classes to simplify the use case of all dynamic extents:

template <size_t N, size_t... Extents>
struct __generate_dynamic_extents {
    using type = typename __generate_dynamic_extents<
        N - 1, dynamic_extent, Extents...
    >::type;
};

template <size_t... Extents>
struct __generate_dynamic_extents<0, Extents...> {
    using type = extents<Extents...>;
};

template <size_t N>
using dynamic_extents = typename __generate_dynamic_extents<N>::type;

template <class ElementType, size_t Rank>
using dynamic_mdspan = basic_mdspan<ElementType, dynamic_extents<Rank>>;

You can see a sketch of this on Godbolt here.

3.3. Rename (mdspan|extents) to static_(extents|mdspan)

If we do add something like dynamic_extents and dynamic_mdspan, why not make it the default and give them the good names of extents and mdspan? All dynamic extents seems like the far more common use case.

We could rename the current extents to static_extents and the current mdspan to static_mdspan.

4. P.S. No ptrdiff_t

A final, partially related note: [P0009R10] must be updated to use size_t instead of ptrdiff_t. As much as I love signed types, dynamic_extent, which shipped in <span> in C++20, is defined as a size_t. If extents, mdspan, and basic_mdspan use ptrdiff_t, users will experience all sorts of signedness and narrowing warnings when trying to use dynamic_extent with these facilities.

References

Informative References

[P0009R10]
Christian Trott, Bryce Adelstein Lelbach, Daniel Sunderland, D. S. Hollman, H. Carter Edwards, Mauro Bianco, Ben Sander, Athanasios Iliopoulos, John Michopoulos, Mark Hoemmen. mdspan. 28 February 2020. URL: https://wg21.link/p0009r10