NVRTC error failed to load builtins on Linux

Hi guys,

I’m having some problem with the NVRTC on Linux, my build for MacOS works fine though. The story is like that, I have compiled my code in one Linux machine, copy the libcuda, libnvrtc and libnvrtc-builtins to the build. When I bring my build to another Linux machine, which don’t have CUDA SDK installed but It have the latest CUDA driver (VIDIA UNIX x86_64 Kernel Module 352.39). When my program try to compile to the kernel code using NVRTC this error message pop up:
“compilation failed - nvrtc: error: failed to load builtin”

Does anyone have experience with Linux can help me with this. Thanks a lot :D.

What version of CUDA are you using? You might to also want to mention which exact version of Linux this is, they are not all exactly the sane. Are the two Linux systems you mention (the “build” and the “run” systems), running the same version of Linux? Is the version of Linux they are running supported by the version of CUDA you are using?

I’m using CUDA 7.5, 2 Linux machine running on same CentOS 6.6. My other CUDA kernels are running fine with these machines. I only have this nvrtc problem.

I haven’t use NVRTC. I assume by “builtins” you are referring to the files called libnvrtc-builtins* ?

These are dynamic link libraries, and you may have to include their location in LD_LIBRARY_PATH. Also, when you copy these files make sure any symbolic links are handled appropriately.