CODEWITHSUNDEEP
c++cmakecudac++17
Dominic
Dominic

Reputation: 491

CMake cannot set CUDA standard c++17

I'm using CMake to configure a CUDA/C++ project. Some of the files compiled with NVCC require C++ 17 features. To enable those, I would use:

cmake_minimum_required(VERSION 3.19)
project(RISA LANGUAGES CXX CUDA)
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CUDA_STANDARD 17)

which results in an error message:

Target "RISA" requires the language dialect "CUDA17" (with compiler extensions), but CMake does not know the compile flags to use to enable it.

Double checking the CMAKE_CUDA_COMPILE_FEATURES variable reveals cuda_std_03 cuda_std_11 cuda_std_14. For some reason, CMake doesn't seem to know about anything past C++14 regarding CUDA.

I'm running ubuntu 20.04 with

What can I do to get C++17 configured and compiled with this setup?

I know about this question - however the solution to just compile cuda files with C++14 does not suffice for me.

Upvotes: 4

Views: 7481

Answers (1)

Dominic
Dominic

Reputation: 491

As is turns out, CMake was not using the correct nvcc binary.

In my case, I've had two versions of nvcc on the system:

  • nvcc v10.1.243 in /usr/bin and
  • nvcc v11.2.152 in /usr/local/cuda-11.2/bin

In my CMake configuration,

  • CUDA_NVCC_EXECUTABLE was set to /usr/local/cuda-11.2/bin, but
  • CMAKE_CUDA_COMPILER was set to the older version in /usr/bin.

After setting CMAKE_CUDA_COMPILER to the correct path, CMake was able to detect nvcc 11.2.152 and could apply the C++17 standard.

Upvotes: 7

Related Questions