Cuda 8.0, Fedora 25, gcc stddef.h nullptr, GTX1060

Question

I have a MSI GTX 1060, running Fedora 25, Blender 2.78a, Nvidia drivers 375.26, and Cuda 8.0.44.

I got the Cuda toolkit from Nvidia's website, installed the repo, and installed the cuda packages but it insists that I need to install xorg-x11-drv-nvidia from RPMFusion*. Which I have done and it all seems to work fine. Once those drivers are installed, if I run glxspheres it indicates I'm using the nvidia kernel module.

When I first try to render something using Cycles Render, I get an error about an unsupported version of GCC. After looking up the fix, I edit the appropriate .h file to remove the offending line. After that, the next time I try to render, I get a long output starting with undefined "nullptr" then multiple lines about expecting semi-colons.

The full error can be found here: http://pastebin.com/zpw6xFYU

Any ideas on what I'm doing wrong? I've tried to search for an answer for nothing comes up.

*I recognize this is not the preferred method to install proprietary drivers, but I was unable to get around this dependency. If the solution involves removing these drivers and cuda, installing the binary manually from Nvidia (which I can do), then installing Cuda in another manner, I'm open to that.

Answer 1

I found the root problem, it was an incompatibility with C++ 11. So passing in a different standard along with compiling the CUDA kernel manually did the trick. Here is the command I ran (I got it from an output linked above):

/usr/local/cuda/bin/nvcc \
-arch=sm_61 -Xcompiler -std=c++98 --cubin \
"/usr/share/blender/2.78/scripts/addons/cycles/kernel/kernels/cuda/kernel.cu" \
-o "/home/user/.config/blender/2.78/cache/cycles_kernel_sm61_D8F831DA74C0363DC018432CF25F3AD8.cubin" \
-m64 --ptxas-options="-v" --use_fast_math -DNVCC -D__KERNEL_CUDA_VERSION__=80 \
-I"/usr/share/blender/2.78/scripts/addons/cycles/kernel"

Which did give me this warning:

cc1: warning: command line option ‘-std=c++98’ is valid for C++/ObjC++ but not for C

However, this doesn't seem to affect Blender's ability to use my video card. I saw a near 50% improvement in render times on the file I have been working on.

For reference, I found these pages that helped me finally hunt down a solution:

Nvidia forum post, CUDA using wrong GCC

Arch Linux bug report

Cplusplus.com forum post