CUDA < 12.4.1 bug affects use of openPMD-api in C++20 projects

[franzpoeschel]

Describe the bug
This bug basically makes std::variant unusable in the affected CUDA versions.

To Reproduce
I've seen this issue when trying to compile the latest dev branch of PIConGPU (2981b07a7160e64d057b5653f3d26410d54a8135) against the openPMD-api (version does not really matter, I used a4d66cc).

The error affects multiple usages of std::variant, e.g.:

/nix/store/6vclfnmcqpm1k5bbh5syf89f21gip2v3-gcc-12.3.0/include/c++/12.3.0/variant(391): error: function "std::__detail::__variant::_Variadic_union<_First, _Rest...>::~_Variadic_union() [with
 _First=unsigned int, _Rest=<unsigned long, unsigned long long, float, double, long double, std::complex<float>, std::complex<double>, std::complex<long double>, std::string, std::vector<cha
r, std::allocator<char>>, std::vector<short, std::allocator<short>>, std::vector<int, std::allocator<int>>, std::vector<long, std::allocator<long>>, std::vector<long long, std::allocator<lon
g long>>, std::vector<unsigned char, std::allocator<unsigned char>>, std::vector<unsigned short, std::allocator<unsigned short>>, std::vector<std::seed_seq::result_type, std::allocator<std::
seed_seq::result_type>>, std::vector<std::size_t, std::allocator<std::size_t>>, std::vector<unsigned long long, std::allocator<unsigned long long>>, std::vector<float, std::allocator<float>>
, std::vector<double, std::allocator<double>>, std::vector<long double, std::allocator<long double>>, std::vector<std::complex<float>, std::allocator<std::complex<float>>>, std::vector<std::
complex<double>, std::allocator<std::complex<double>>>, std::vector<std::complex<long double>, std::allocator<std::complex<long double>>>, std::vector<signed char, std::allocator<signed char
>>, std::vector<std::string, std::allocator<std::string>>, openpmd_array_double_7, bool>]" (declared at line 400) cannot be referenced -- it is a deleted function                            
   : _M_rest(in_place_index<_Np-1>, std::forward<_Args>(__args)...)                                                                                                                           
            ^                                                                                                                                                                                 
          detected during:
…
…
…
            instantiation of "std::__detail::__variant::_Move_ctor_base<<unnamed>, _Types...>::_Move_ctor_base(std::__detail::__variant::_Move_ctor_base<<unnamed>, _Types...> &&) [with <unna
med>=false, _Types=<char, unsigned char, signed char, short, int, long, long long, unsigned short, unsigned int, unsigned long, unsigned long long, float, double, long double, std::complex<f
loat>, std::complex<double>, std::complex<long double>, std::string, std::vector<char, std::allocator<char>>, std::vector<short, std::allocator<short>>, std::vector<int, std::allocator<int>>
, std::vector<long, std::allocator<long>>, std::vector<long long, std::allocator<long long>>, std::vector<unsigned char, std::allocator<unsigned char>>, std::vector<unsigned short, std::allo
cator<unsigned short>>, std::vector<std::seed_seq::result_type, std::allocator<std::seed_seq::result_type>>, std::vector<std::size_t, std::allocator<std::size_t>>, std::vector<unsigned long 
long, std::allocator<unsigned long long>>, std::vector<float, std::allocator<float>>, std::vector<double, std::allocator<double>>, std::vector<long double, std::allocator<long double>>, std:
:vector<std::complex<float>, std::allocator<std::complex<float>>>, std::vector<std::complex<double>, std::allocator<std::complex<double>>>, std::vector<std::complex<long double>, std::alloca
tor<std::complex<long double>>>, std::vector<signed char, std::allocator<signed char>>, std::vector<std::string, std::allocator<std::string>>, openpmd_array_double_7, bool>]" at line 59 of /
home/franzpoeschel/picongpu_env/env/include/openPMD/backend/Attribute.hpp

Line:

openPMD-api/include/openPMD/backend/Attribute.hpp

Line 55 in a4d66cc

class Attribute : public auxiliary::Variant<Datatype, attribute_types>

or

/nix/store/6vclfnmcqpm1k5bbh5syf89f21gip2v3-gcc-12.3.0/include/c++/12.3.0/variant(391): error: function "std::__detail::__variant::_Variadic_union<_First, _Rest...>::~_Variadic_union() [with
 _First=openPMD::UniquePtrWithLambda<void>, _Rest=<>]" (declared at line 400) cannot be referenced -- it is a deleted function                                                                
   : _M_rest(in_place_index<_Np-1>, std::forward<_Args>(__args)...)                                                                                                                           
            ^                                                                                                                                                                                 
          detected during:  
…
…
…
            instantiation of "std::variant<_Types...>::variant(_Tp &&) [with _Types=<std::shared_ptr<const void>, openPMD::UniquePtrWithLambda<void>>, _Tp=openPMD::UniquePtrWithLambda<void>,
 <unnamed>=void, <unnamed>=void, _Tj=openPMD::UniquePtrWithLambda<void>, <unnamed>=void]" at line 181 of /home/franzpoeschel/picongpu_env/env/include/openPMD/auxiliary/Memory.hpp

Line:

openPMD-api/include/openPMD/auxiliary/Memory.hpp

Line 178 in a4d66cc

EligibleTypes m_buffer;

What should we do

I don't see an easy way out of this. Looking at the linked bug report, even trivial use of std::variant seems to be impossible. In openPMD::Attribute, std::variant is part of our public API, so we can't easily replace it with something else.

• Option 1: Check the affected CUDA versions, add some include guard to cancel compilation with the note that the CUDA version is broken and something else should be used.
• Option 2: Reintroduce mpark::variant? It seems to be unmaintained though

[ax3l]

Thanks for documenting this C++20 issue with CUDA compilers! I encounter no problems with C++17 and std::variant in my projects.

It looks to me like using CUDA 12.4.1+ fixes this for C++20:
https://forums.developer.nvidia.com/t/nvcc-c-20-std-variant-complie-failed/270162/5

So I think there are two solutions for people:

• If you want to use C++20 with Nvidia compilers, use the latest compilers only for now. They already fixed this issue and we should keep reporting other bugs we find to https://developer.nvidia.com This is what we usually do when new standards are supported in vendor compilers and bugs are being shaken out still.
• If you want to use std::variant and older Nvidia compilers, only require C++17 for now in downstream projects. (openPMD-api only needs C++17 currently)

[franzpoeschel]

This seems to particularly affect gcc 12.3.0, other versions (even within this minor release range) seem to work. Need to investigate again and maybe narrow down the error check.