What's the relationship between CUDA and CCCL?

[jiapei100]

I got CUDA 12.5 installed, as well as CCCL 2.4.0

But, it looks there are some incompatility?

I got a couple of folders included

• /usr/local/cuda/include
• /usr/local/cuda/include/cuda/std/detail/libcxx/include`

But, when I tried to build OpenCV 4.10, I obtained the following ERROR messages:

/usr/local/cuda/include/cuda/std/__type_traits/enable_if.h(27): error: incomplete type "cuda::std::__4::_CCCL_VISIBILITY_DEFAULT" is not allowed
  template <bool, class _Tp = void> struct _CCCL_VISIBILITY_DEFAULT enable_if {};
                                                                    ^

/usr/local/cuda/include/cuda/std/__type_traits/enable_if.h(28): error: incomplete type "cuda::std::__4::_CCCL_VISIBILITY_DEFAULT" is not allowed
  template <class _Tp> struct _CCCL_VISIBILITY_DEFAULT enable_if<true, _Tp> {typedef _Tp type;};
                                                       ^

/usr/local/cuda/include/cuda/std/__type_traits/enable_if.h(28): error: expected an expression
  template <class _Tp> struct _CCCL_VISIBILITY_DEFAULT enable_if<true, _Tp> {typedef _Tp type;};
                                                                             ^

/usr/local/cuda/include/cuda/std/__type_traits/enable_if.h(28): error: expected a "}"
  template <class _Tp> struct _CCCL_VISIBILITY_DEFAULT enable_if<true, _Tp> {typedef _Tp type;};
                                                                                             ^

/usr/local/cuda/include/cuda/std/__type_traits/enable_if.h(30): error: name followed by "::" must be a class or namespace name
  template <bool _Bp, class _Tp = void> using __enable_if_t = typename enable_if<_Bp, _Tp>::type;
                                                                       ^

/usr/local/cuda/include/cuda/std/__type_traits/enable_if.h(33): error: name followed by "::" must be a class or namespace name
  template <bool _Bp, class _Tp = void> using enable_if_t = typename enable_if<_Bp, _Tp>::type;

Well, I tried to solve my compilation issue from [https://github.com/opencv/opencv_contrib/issues/3690].(opencv/opencv_contrib#3690). However, it turns out CCCL is the solution .

Anyway, can anybody give me some suggestions?

Cheers
Pei

[miscco]

The error points you to an incomplete type:

error: incomplete type "cuda::std::__4::_CCCL_VISIBILITY_DEFAULT"

However, _CCCL_VISIBILITY_DEFAULT is not a type but a macro that conditionally defines an attribute. The fact that the compiler thinks this is a type tells us that the macro is not defined in your build.

The only way this can happen is when you mix different versions of CCCL. Could it be that you have a separate Thrust dependency or pulling in CUB?

[pauleonix]

Couldn't the cause also be directly adding /usr/local/cuda/include/cuda/std/detail/libcxx/include to the included folders? Having /usr/local/cuda/include should suffice as headers in /usr/local/cuda/include/cuda/std/detail/libcxx/include are included from headers in /usr/local/cuda/include/cuda/std.

[miscco]

So _CCCL_VISIBILITY_DEFAULT was only introduced after 2.4 was out, so you definitely have some newer cccl repo somewhere

[jiapei100]

@miscco
Yes, I believe I have a mixed version of CUDA or CCCL.

I have both CUDA 12.5 and CCCL 2.4.0 installed under the same folder /usr/local/cuda.

Cheers

[jrhemstad]

@jiapei100 can you clarify if you are just simply downloading the 12.5 CUDA Toolkit and trying to use #include<cuda/std/tuple> , or are you also downloading CCCL from GitHub and adding it to your include path?

It'd be helpful if you could show your exact nvcc compiler invocation.

[jrhemstad]

@jiapei100 to answer the larger question of "What is the relationship between CUDA and CCCL", CCCL is included by default with the CUDA Toolkit. You can download the CUDA toolkit and include a CCCL header like #include <cuda/std/tuple> and it "just works".

We also support using a newer version of CCCL in case you want to upgrade CCCL without upgrading the version of the CTK you're using.

For more details, you can see https://github.com/NVIDIA/cccl?tab=readme-ov-file#platform-support