A Look at std::mdspan

New library feature coming to C++23

These views do not in any way represent those of NVIDIA or any other organization or institution that I am professionally associated with. These views are entirely my own.


A YouTuber by the name of Context Free posted a few videos about tensors in various programming languages, including C++ (first video, second video). I loved these videos, but I noticed ContextFree failed to mention std::mdspan, a new library feature coming to C++23.


std::mdspan (or std::experimental::mdspan until the proposal is accepted-I’ll be using stdex::mdspan, as stdex is a common alias for std::experimental) is an alias for a more complex type, but at it’s core it is a pointer plus some number of extents. Extents are either sizes of a given dimension on an mdspan, or the sentinal std::dynamic_extent, which indicates the extent is dynamic, and doesn’t have to be supplied at compile-time. These extents and the sentinal dynamic_extent can be mixed and matched. This powerful capability allows users to work with data as if it were a complex matrix structure while the underlying data remain linear.

For example, given raw data, an mdspan may be constructed and passed to some library that expects a matrix with rank 3:

// https://godbolt.org/z/eWKev9nas
using mat_3d = std::mdspan<
                   , std::dynamic_extent
                   , std::dynamic_extent

template<typename T> void work(mat_3d<T> mat);

int main() {
  int raw[500] = {};

  work(stdex::mdspan(raw, 100, 5, 1));
  work(stdex::mdspan(raw, 4, 5, 5));
  work(stdex::mdspan(raw, 10, 1, 50));

Note that we will have to be more careful if we mix-and-match compile-time and run-time extents.

// https://godbolt.org/z/Phr1vh9zs
template<typename T> using mat_3d = stdex::mdspan<

template<typename T> void work(mat_3d<T> mat) {}

int main() {
  int raw[27] = {};
  work(mat_3d<int>(raw, 3, 3, 3));
  work(mat_3d<int>(raw, 9, 3, 0));
  // work(mat_3d<int>(raw, 3, 0, 9)) will fail bc extents don't match!
  return 0;

After supplying a dummy implementation of work to print the shape, we get

mat has shape [100, 5, 1]
mat has shape [4, 5, 5]
mat has shape [10, 1, 50]

In either case, the underlying data is the same, though it’s viewed differently in each of the invocations of work. It’s no coincidence that view seems like a natural name for mdspan - mdspan was developed by the authors of the portable execution library Kokkos and inspired by the Kokkos::View type.


Just like std::span, mdspan has support for taking subspans of a given span. This is more complicated with mdspan however, due to mdspan’s variadic extents.

There are three ways to take a slice of an mdspan:

  1. An integer index into the respective dimension
  2. A std::tuple or std::pair begin/end pair of indices
  3. The special value std::full_extent indicating all elements of the dimension should be selected in the subspan

For example:

// https://godbolt.org/z/Wrr4dhEs8
int main() {
  int raw[27] = {};
  std::iota(raw, raw+27, 0);

  // full will have shape [3,3,3]
  auto full = stdex::mdspan(raw, 3, 3, 3);

  // sub will have shape [1,3] and values [18,19,20]
  auto sub = stdex::submdspan(full, 2, std::pair{0, 1}, stdex::full_extent);

Note that using a single value as an extent passed to submdspan squashes the dimension to 0, while passing a pair or tuple will keep the dimension around, even if that dimension is 1. pair/tuple do not effect rank, while passing an index as an extent does.

I hope this article was enough to get you interested in mdspan and the future of C++! Make sure to check out Daisy Hollman’s appearance on CppCast, Context Free’s YouTube channel, and the reference implementation of C++23’s std::mdspan.


You can reach me at any of the links below:

These views do not in any way represent those of NVIDIA or any other organization or institution that I am professionally associated with. These views are entirely my own.

Written on Dec 23rd, 2021