std::mdspan views one flat buffer as a matrix

C++ programs have always stored matrices, images, and grids in a flat one-dimensional buffer and done the index arithmetic by hand: data[r * cols + c]. It works, it is fast, and it is a steady source of off-by-one and row-versus-column-major bugs. std::mdspan (C++23, P0009) is the standard fix: a non-owning view that wraps the flat buffer you already have and gives it a shape.

The design is three separable pieces. Storage is your contiguous array, owned by whatever already owns it. Extents are the shape, any mix of compile-time and runtime dimensions. A layout maps multidimensional indices to offsets (row-major by default, column-major and strided are built in). mdspan owns none of the data and adds no allocation; it is span with more than one dimension.

The demo takes a std::vector<int> of twelve elements and views it as three rows of four. The indexing uses C++23’s real multidimensional subscript, m[r, c], not a workaround:

shape: 3 x 4
  0  1  2  3
  4  5  6  7
  8  9 10 11

No copy happened, and the vector still owns its storage. Change the extents to view the same buffer as 4x3, or hand the mdspan to a function that neither knows nor cares how the memory was allocated.

That last point is why it matters beyond convenience. mdspan is the lingua franca the numerical parts of the standard library are being built around: it is how you hand a slice of a tensor to a BLAS-style routine without committing to a container. Mark Hoemmen, one of its authors, gave the “Multidimensional Parallel Standard C++” keynote at C++Now 2026 (videos now public) on exactly that direction, mdspan plus the parallel algorithms as the standard’s answer for array-heavy and GPU-adjacent code. GCC 16.1’s libstdc++ ships mdspan today, so the everyday half of that story runs in the browser above.

Sources: cppreference: std::mdspan · P0009 “mdspan” · Mark Hoemmen, “Multidimensional Parallel Standard C++” (C++Now 2026).