cuda::span redefinition and ambiguous constructor errors in CUDA 12

[StonerLing]

Description

When compiling a .cu file with CUDA 12, using cuda-api-wrappers together with Thrust leads to cuda::span redefinition and ambiguous constructor errors. The same code compiles fine under CUDA 11.

---

Minimal Reproducer

#include <cuda/api.hpp> // Must come first to trigger the error

#include <thrust/device_vector.h>
#include <thrust/execution_policy.h>

void trigger() { cuda::span<float> s{nullptr, 0}; }

int main() {
    trigger();
    return 0;
}

Compile with:

nvcc -x cu -std=c++20 -arch=sm_75 -cudart shared \
     -I <path-to-cuda-api-wrappers> \
     -I <CUDA-include-path> \
     -c main.cpp -o main.obj

Observed Behavior

Errors like:

error C2977: 'cuda::span': too many template arguments
error C2641: cannot deduce template arguments for 'cuda::span'
error C2780: expects 0 arguments - 2 provided

Notes

• The issue only appears when compiling the file as a CUDA source (-x cu) in CUDA 12.
• Treated as a normal C++ file, the code compiles fine.
• Likely related to changes in CUDA 12 handling of cuda::span and C++20 aggregate initialization.

[eyalroz]

First, a nitpick: In C++, it is not necessary to return from main. There's an implicit return 0;.

More importantly, though - this is not even a minimal reproducer, and the problem isn't specific to spans. Look:

#include <thrust/device_vector.h>
#include <cuda/event.hpp>

int main() { }

This already triggers an error:

$ /usr/local/cuda/bin/nvcc -I"/home/eyalroz/src/mine/cuda-api-wrappers/src" -isystem "/usr/local/cuda/targets/x86_64-linux/include" -isystem "/usr/local/cuda/targets/x86_64-linux/include/cccl" -std=c++20 -x cu -c /home/eyalroz/src/mine/cuda-api-wrappers/examples/other/delme.cu
/home/eyalroz/src/mine/cuda-api-wrappers/src/cuda/api/graph/../event.hpp(374): error: "cuda::event" is ambiguous
   ::cuda::event::handle_t new_event_handle;
           ^

/home/eyalroz/src/mine/cuda-api-wrappers/src/cuda/api/graph/../event.hpp(374): error: expected a ";"
   ::cuda::event::handle_t new_event_handle;
                           ^

/home/eyalroz/src/mine/cuda-api-wrappers/src/cuda/api/graph/../event.hpp(375): error: identifier "new_event_handle" is undefined
   auto status = cuEventCreate(&new_event_handle, flags);
                                ^

/home/eyalroz/src/mine/cuda-api-wrappers/src/cuda/api/multi_wrapper_impls/context.hpp(324): error: "cuda::event" is ambiguous
   return cuda::event::detail_::create(
                ^

this is with CUDA 13.1 (nvcc V13.1.80).

[StonerLing]

Interesting, I haven't test with CUDA 13.

This code could be compiled with CUDA 12:

#include <thrust/detail/config.h> // This is included at the beginning of most user-facing thrust headers
#include <cuda/api.hpp>

int main() { }

However, this couldn't:

#include <cuda/api.hpp> // Must come first to trigger the error
#include <thrust/device_vector.h>

int main() { 
    float f = 0;
    cuda::span s{&f, 1};
}

The errors are:

cuda\std\detail\libcxx\include\__iterator/wrap_iter.h(243): error C2977: 'cuda::span': too many template arguments
cuda-api-wrappers\cuda\api\detail/span.hpp(38): note: see declaration of 'cuda::span'
cuda\std\detail\libcxx\include\__iterator/wrap_iter.h(243): note: the template instantiation context (the oldest one first) is
cuda\std\detail\libcxx\include\__iterator/wrap_iter.h(112): note: while compiling class template 'cuda::std::__4::__wrap_iter'
main.cpp(6): error C2641: cannot deduce template arguments for 'cuda::span'
main.cpp(6): error C2893: Failed to specialize function template 'cuda::span<T> cuda::span(T *,cuda::span<T>::size_type)'
cuda-api-wrappers\cuda\api\detail/span.hpp(38): note: see declaration of 'cuda::span'
main.cpp(6): note: With the following template arguments:
main.cpp(6): note: 'T=float'

If we only consider span-related errors, I think not every Thrust include will involve cuda::span.

[eyalroz]

I wonder, though, if the errors are triggered by something within thrust, or by something within the cuda::std . Perhaps we could get the thrust people to play nicer?

[eyalroz]

Had to tweak my earlier reproducer; just the config.h is not enough. device_vector.h is enough, and I couldn't quite put my finger on what it is that breaks the compilation, since there's a mess of includes in there.

Also, in your last comment, what was your compilation command-line? Specifically, what was the C++ standard version? And which NVCC version were you using, exactly?

[StonerLing]

I compiled it with :

nvcc -x cu -std=c++20 -arch=sm_75 -cudart shared \
     -I <path-to-cuda-api-wrappers> \
     -I <CUDA-include-path> \
     -c main.cpp -o main.obj

NVCC version:

nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2024 NVIDIA Corporation
Built on Tue_Feb_27_16:28:36_Pacific_Standard_Time_2024
Cuda compilation tools, release 12.4, V12.4.99
Build cuda_12.4.r12.4/compiler.33961263_0

[eyalroz]

So, this is due to CCCL expanding into the CUDA namespace. See:

• Bug against CCCL: [BUG]: Implicit inclusion of libcu++ headers + 'expansionism' into in-use names breaks the CUDA API wrappers & user code NVIDIA/cccl#7569
• Bug against NVIDIA: https://developer.nvidia.com/bugs/5867511