What's the difference bewteen std::pair and cuda::std::pair?

[w3ntao]

When should I use them differently?

[miscco]

Generally speaking cuda::std::pair is usable on both host and device, whereas std::pair is not

Besides that we try very hard to be standard conforming. The only differences that come to mind are:

• cuda::std::pair is trivially copyable if the template arguments are. This comes at the cost that technically pointer to member point to a base class instead of members of the class but that is fine for us
• We aggressively backport constexpr and other features to older standard modes

To make live easier for our user we also added conversions to and from std::pair so that you can interoperate more easily.

So you should use cuda::std::pair if you intend to use it on device, or you are stuck on an older C++ standard and want newer features

[w3ntao]

Oh I see, thanks for the answer.
The fact that std::pair can be used directly in pure device environment (without interacting with host code) really confuse me. I guess they were translated into cuda::std::pair by compiler in such cases.

[miscco]

The issue is that std::pair is mostly constexpr from C++17 onwards. So if you compile with --relaxed-constexpr then it might work

[w3ntao]

Sorry you lost me.
Did you mean "with C++17 and compilation flag --relaxed-constexpr, std::pair will work just like cuda::std::pair"?

[pauleonix]

Basically, but there are good reasons not to use --relaxed-constexpr. See rapidsai/cudf#7795