error: "forceinline" redefined in simple program

Question

Compiling the 3-line program test-cuda.cpp

#include <thrust/execution_policy.h>

int main() { return 0; }

results in a compiler warning/error:

$ g++ -std=c++17 test-cuda.cpp -I/opt/cuda/targets/x86_64-linux/include -Werror
In file included from /opt/cuda/targets/x86_64-linux/include/cuda_runtime_api.h:147,
                 from /opt/cuda/targets/x86_64-linux/include/cub/detail/detect_cuda_runtime.cuh:38,
                 from /opt/cuda/targets/x86_64-linux/include/cub/util_arch.cuh:41,
                 from /opt/cuda/targets/x86_64-linux/include/cub/util_debug.cuh:40,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/system/cuda/config.h:43,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/execution_policy.h:35,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/iterator/detail/device_system_tag.h:23,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/iterator/iterator_traits.h:62,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/pointer_traits.h:23,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/detail/raw_pointer_cast.h:20,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/detail/execute_with_allocator.h:23,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/execution_policy.h:25,
                 from test-cuda.cpp:1:
/opt/cuda/targets/x86_64-linux/include/crt/host_defines.h:86: error: "__forceinline__" redefined [-Werror]
   86 | #define __forceinline__ \
      | 
In file included from /opt/cuda/targets/x86_64-linux/include/cuda/std/version:13,
                 from /opt/cuda/targets/x86_64-linux/include/cuda/std/cstddef:20,
                 from /opt/cuda/targets/x86_64-linux/include/cuda/std/type_traits:17,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/detail/type_traits.h:27,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/detail/execute_with_allocator_fwd.h:21,
                 from /opt/cuda/targets/x86_64-linux/include/thrust/detail/execute_with_allocator.h:21:
/opt/cuda/targets/x86_64-linux/include/cuda/std/detail/__config:38: note: this is the location of the previous definition
   38 |         #define __forceinline__
      | 
cc1plus: all warnings being treated as errors

This is a stripped down version of an issue we're facing in a larger project in which we compile w/ -Werror.

Is there something we're missing, or a workaround without having to change the CUDA/thrust source code?

Similar warning occurs compiling w/ nvcc -std=c++17 test-cuda.cpp.

OS: arch linux

Cuda: 12.1.0

GCC: 12.2.1

According to lines 38-40 in this link, that should not be happening. — catnip
– catnip, Commented Apr 6, 2023 at 22:14
I don't have any trouble with nvcc when compiling it as a .cu file. That is my usual suggestion with thrust. — Robert Crovella
– Robert Crovella, Commented Apr 6, 2023 at 23:05
@PaulSanders That's where it is first defined. The problem is that host_defines.h unconditionally #defines __forceinline__ which conflicts with the prior #define in your link. Thank you Robert that does work. Seems we will need to find a way to name our files in such a way that builds on both gcc and nvcc. — Matt
– Matt, Commented Apr 7, 2023 at 12:48
@Matt You can use the -x option with most compilers to explicitly set the language (e.g. nvcc -x cu test.cpp or gcc -x cpp test.cu). I believe this can also be controlled using target / source-file properties in CMake if you're using that. — Allison Piper
– Allison Piper, Commented Jun 7, 2023 at 15:46

jrhemstad · Accepted Answer · 2023-06-07 15:46:21Z

4

Anis' answer is a good option for a short-term workaround.

The other option would be to do #undef __forceinline__ between the includes:

    #include <thrust/execution_policy.h>
    #undef __forceinline__
    #include <cuda_runtime.h>
    
    int main() { return 0; }

Also note that the root cause of this issue has already been address in libcudacxx here: https://github.com/NVIDIA/libcudacxx/pull/476

This will ship with the next 2.2 release of Thrust/libcudacxx.

answered Jun 7, 2023 at 15:46

jrhemstad

1411 silver badge4 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

Matt Over a year ago

This doesn't seem to help w/ Cuda 12.1. The problem is that thrust/execution_policy.h alone includes conflicting #define __forceinline__ lines as shown in the question.

Anis Ladram · Accepted Answer · 2023-06-08 14:34:19Z

3

Edit: please prefer Jake's solution below. Thanks!

Add #include <cuda_runtime.h> before including thrust headers in order to address your issue. Example:

$ cat test.cc
#include <cuda_runtime.h>
#include <thrust/execution_policy.h>

int main() { return 0; }
$ g++ -std=c++17 test.cc -I/usr/local/cuda/include -Werror
$ echo $?
0

edited Jun 8, 2023 at 14:34

answered Apr 7, 2023 at 14:50

Anis Ladram

1,6559 silver badges10 bronze badges

1 Comment

Matt Over a year ago

Sorry to disagree but yours is still the better answer.

Collectives™ on Stack Overflow

error: "forceinline" redefined in simple program

2 Answers 2

1 Comment

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

1 Comment

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related