Add P3242: Copy and fill for mdspan

[nmm0]

No description provided.

[mhoemmen]

mdspans are views; would you therefore consider passing them by value, as is done in P1673?

Suggested change

	void copy(const mdspan<SrcElementType, SrcExtents, SrcLayoutPolicy, SrcAccessorPolicy> &src, const mdspan<DstElementType, DstExtents, DstLayoutPolicy, DstAccessorPolicy> &dst)
	void copy(mdspan<SrcElementType, SrcExtents, SrcLayoutPolicy, SrcAccessorPolicy> src, mdspan<DstElementType, DstExtents, DstLayoutPolicy, DstAccessorPolicy> dst);

[nmm0]

revisiting this, because even though they are views, the size of the mdspan is proportional to the number of dynamic extents; so it's very possible to have mdspans that are not passible in registers. I'm happy to keep it as we have in other papers, but this particular aspect might get challenged

[mhoemmen]

@nmm0 I don't really mind either way. For std::linalg, all the mdspan are no more than rank 2, but here, they might have arbitrary rank.

I suspect LEWG would want to hear that argument. They might also want some performance measurements to justify a difference from the status quo.

[dalg24]

Naive implementation is not tedious

for i0 in [0, extents(0)
  for i1 in [0, extents(1)
    ...
      for iK in [0, extents(K)
  dst[i0, i1, ..., iK] = src[i0, i1, ..., iK];

just (sometimes) really inefficient.

[nmm0]

Haha well, I think writing an 8-nested for loop is pretty tedious :P. I guess it's a matter of opinion

[mhoemmen]

I already submitted a PR to kokkos/mdspan that does generic iteration over extents of any rank. It's not really tedious, but would be really inefficient without at least the specializations for exhaustive layouts.