cmake: /usr/include/c++/10.1.0/type_traits(1396): “error: type name is not allowed” or “error: identifier “__is_same_as” is undefined”

If the gcc version being used is incompatible with cuda version, there will be errors such as:

[ 68%] Building NVCC (Device) object src/nvtt/CMakeFiles/cuda_compile_1.dir/cuda/
/usr/include/c++/10.1.0/type_traits(1396): error: type name is not allowed

/usr/include/c++/10.1.0/type_traits(1396): error: type name is not allowed

/usr/include/c++/10.1.0/type_traits(1396): error: identifier "__is_same_as" is undefined

3 errors detected in the compilation of "/home/nick/Code/Git/".
CMake Error at (message):
  Error generating file

make[2]: *** [src/nvtt/CMakeFiles/nvtt.dir/build.make:85: src/nvtt/CMakeFiles/cuda_compile_1.dir/cuda/] Error 1
make[1]: *** [CMakeFiles/Makefile2:721: src/nvtt/CMakeFiles/nvtt.dir/all] Error 2
make: *** [Makefile:183: all] Error 2

See discussion on GitHub: CUDA Build error on Arch Linux and CUDA Compilers – nvcc compatibility chart.

This can be resolved by using an earlier version of GCC which is compatible with the CUDA version. In my example setting the following CMake variable resolved the issue:


Leave a Reply

Your email address will not be published. Required fields are marked *