Transfer array from Singe

[BrendanKKrueger]

PR Summary

An implementation of std::array that works on both the CPU and the GPU.

PR Checklist

• Any changes to code are appropriately documented.
• Code is formatted.
• Install test passes.
• Docs build.
• If preparing for a new release, update the version in cmake.

[Yurlungur]

Big picture I think this is a great addition and would like to see this included. Given discussion with @mauneyc-LANL and @rbberger I think we cna merge without CI actually running, so long as you ran tests locally and we can wait for the great test rewrite from @rbberger .

A few comments/questions/requested changes below. Feel free to push back on anything.

[Yurlungur]

Oh also there should be at least a mention this exists in the docs.

[mauneyc-LANL]

I was digging around a bit and found this

NVIDIA/libcudacxx#51 (comment)

TLDR: 1) --expt-relaxed-constexpr & -std=c++17 seems to provide sufficient functionality for std::array on device, 2.) libcu++ rolls a std::array in cuda/std/array (as of 2023).

Would this make a hand-rolled std::array necessary?

[Yurlungur]

I was digging around a bit and found this

NVIDIA/libcudacxx#51 (comment)

TLDR: 1) --expt-relaxed-constexpr & -std=c++17 seems to provide sufficient functionality for std::array on device, 2.) libcu++ rolls a std::array in cuda/std/array (as of 2023).

Would this make a hand-rolled std::array necessary?

👍 This is what we do in parthenon. But is that fully portable?

[BrendanKKrueger]

I was digging around a bit and found this
NVIDIA/libcudacxx#51 (comment)
TLDR: 1) --expt-relaxed-constexpr & -std=c++17 seems to provide sufficient functionality for std::array on device, 2.) libcu++ rolls a std::array in cuda/std/array (as of 2023).
Would this make a hand-rolled std::array necessary?

👍 This is what we do in parthenon. But is that fully portable?

Most of std::array works on GPUs with relaxed constexpr, but not all. In particular, std::array::fill and std::array::swap aren't constexpr until C++20 and I believe Singe code uses std::array:fill. Once we move to C++20 (hopefully much faster than the 14-to-17 transition), then this code becomes redundant if we're comfortable relying on relaxed constexpr to get the STL onto the GPU.

[Yurlungur]

I was digging around a bit and found this
NVIDIA/libcudacxx#51 (comment)
TLDR: 1) --expt-relaxed-constexpr & -std=c++17 seems to provide sufficient functionality for std::array on device, 2.) libcu++ rolls a std::array in cuda/std/array (as of 2023).
Would this make a hand-rolled std::array necessary?

👍 This is what we do in parthenon. But is that fully portable?

Most of std::array works on GPUs with relaxed constexpr, but not all. In particular, std::array::fill and std::array::swap aren't constexpr until C++20 and I believe Singe code uses std::array:fill. Once we move to C++20 (hopefully much faster than the 14-to-17 transition), then this code becomes redundant if we're comfortable relying on relaxed constexpr to get the STL onto the GPU.

that's a satisfying explanation for me. 👍

[BrendanKKrueger]

This is ready for another review

[Yurlungur]

Looks great! Only one concern in the tests otherwise I think this is good to go.

[BrendanKKrueger]

@Yurlungur, is this namespacing correct?

[Yurlungur]

Yep 👍