build failure with GCC14

[prusnak]

line here should set C++ standard to 14 instead of 11

oidn/cmake/oidn_platform.cmake

Line 31 in 24baabd

set(CMAKE_CXX_STANDARD 11)

otherwise the CUDA build fails with

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/type_traits(1610): error: "__is_nothrow_new_constructible" is not a function or static data member
      constexpr bool __is_nothrow_new_constructible
                     ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/type_traits(1610): error: "constexpr" is not valid here
      constexpr bool __is_nothrow_new_constructible
      ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/type_traits(1610): error: "__is_nothrow_new_constructible" is not a function or static data member
      constexpr bool __is_nothrow_new_constructible
                     ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/type_traits(1610): error: "constexpr" is not valid here
      constexpr bool __is_nothrow_new_constructible
      ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/bits/new_allocator.h(190): error: identifier "__is_nothrow_new_constructible" is undefined
   noexcept(__is_nothrow_new_constructible<_Up, _Args...>)
            ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/bits/new_allocator.h(190): error: type name is not allowed
   noexcept(__is_nothrow_new_constructible<_Up, _Args...>)
                                           ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/bits/new_allocator.h(190): error: expected a ")"
   noexcept(__is_nothrow_new_constructible<_Up, _Args...>)
                                              ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/bits/alloc_traits.h(107): error: variable template "std::__allocator_traits_base::__has_construct_impl" is not a valid class member template
        static constexpr bool __has_construct_impl = false;
                              ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/bits/alloc_traits.h(110): error: __has_construct_impl is not a template
        __has_construct_impl<_Alloc, _Tp,
        ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/bits/alloc_traits.h(110): error: variable template "std::__allocator_traits_base::_Args" is not a valid class member template
        __has_construct_impl<_Alloc, _Tp,
        ^

/nix/store/il0r2x1l03qjda9nxva18c1mjsqggcn9-gcc-14-20241116/include/c++/14-20241116/bits/alloc_traits.h(115): error: variable template "std::__allocator_traits_base::__has_construct" is not a valid class member template
        static constexpr bool __has_construct
                              ^

[atafra]

I cannot reproduce this issue with GCC 14.2.0 on Ubuntu 24.10.

I also don't think that the global C++ standard should be changed from C++11 to C++14 (and we can't do that anyway since we're currently committed to require only C++11 for CPU-only builds). The CUDA backend indeed requires more than C++11 but that target already has CXX_STANDARD 17 set, and the non-device specific code should be C++11 compliant.

[prusnak]

Same issue with Gentoo: https://bugs.gentoo.org/950261

It seems they fix by commenting the whole line completely (and the line below!) - via matching all lines that contain set(CMAKE_CXX_STANDARD .

https://gitweb.gentoo.org/repo/gentoo.git/tree/media-libs/oidn/oidn-2.3.2.ebuild#n77

[atafra]

I'm not familiar with Gentoo so I can't really comment much on this. I don't understand why removing CMAKE_CXX_STANDARD fixed it in Gentoo but in their build log -std=c++11 is passed to NVCC, which is obviously wrong and should be -std=c++17 because C++17 is set for that CMake target. Since Gentoo applies some patches to OIDN, this could be a side effect of some of their changes, which could be similar to your case. Maybe they removed CXX_STANDARD 17 from the CUDA target, I don't know. If they didn't, it's very odd that CMake doesn't respect that option, for which I have no explanation. My only guess is that it might be a bug in the CMake version both you and Gentoo are using, which was later fixed. However, I don't know which CMake version is being used in the failure cases.

What I can say for sure is that we're testing builds for the original, unmodified CMake files on Rocky Linux and Ubuntu, and in both cases -std=c++17 is correctly passed to NVCC, so there is no such issue. Unless it would be possible to create a reproducer in Docker for an unmodified OIDN build, unfortunately there's not much I could do about this. At this point it's unclear whether this is a real issue in OIDN or the issue happens because of third-party modifications to the CMake files. So far I don't see that OIDN is doing anything wrong.

[AngryLoki]

Dropping -std= flags for GCC effectively acts as setting -std=c++17 for GCC>=11 - see https://godbolt.org/z/cfboErMo3 (and Gentoo does not support GCC 10 anymore), that's why it works. It is still better to set a specific standard for future compilers, yes.

Regarding on what happens with nvcc -std=c++11 (and who supposed to override it with -std=c++17) - no idea / environment / time to check. OIDN uses standard CMake enable_language(CUDA) - maybe CMake was instructed to override specified standard only since specific version (when CUDA_STANDARD is NOT set).