[Cycles] CUDA compilation failed for GTX760

Hey,

today made a linux build and tried to compile the CUDA kernels for my graphics card (one PC): GTX550Ti and GTX760
The GTX550Ti compiles fine, but for the GTX760 I always get the typical CUDA compile error with a “Segmentation fault”.

florian@florian-desktop ~/Entwicklung/blender-git/build_linux/bin $ ./blenderfound bundled python: /home/florian/Entwicklung/blender-git/build_linux/bin/2.70/python
Compiling CUDA kernel ...
"nvcc" -arch=sm_30 -m64 --cubin "/home/florian/Entwicklung/blender-git/build_linux/bin/2.70/scripts/addons/cycles/kernel/kernel.cu" -o "/home/florian/.config/blender/2.70/cache/cycles_kernel_sm30_54E3C696A61A2FF05BA769C862D83BAC.cubin" --ptxas-options="-v" --maxrregcount=32 --use_fast_math -I"/home/florian/Entwicklung/blender-git/build_linux/bin/2.70/scripts/addons/cycles/kernel" -DNVCC -D__KERNEL_CUDA_VERSION__=50
Segmentation fault
CUDA kernel compilation failed, see console for details.


Refer to the Cycles GPU rendering documentation for possible solutions:
http://wiki.blender.org/index.php/Doc:2.6/Manual/Render/Cycles/GPU_Rendering



The GTX760 is also my display graphics card, while the GTX550Ti is just for rendering. In older builds both worked together for rendering. The problem is that my last linux build was quite old so I can’t say since which commit this is broken.

Blender version: df5cbfb, latest GIT-version
CUDA toolkit:5.0.35
NVidia driver: 319.60
OS: Linux Mint 16 (based on Ubuntu 13.10)

Thanks for your help,
Florian

EDIT: Rendering with the buidbot-version with the GTX760 works…

Hi floric, I can´t build kernel for my EVGA GTX 760 with latest master 6e9ee13 too but got different error.

Compiling CUDA kernel ...
"/usr/local/cuda-5.0/bin/nvcc" -arch=sm_30 -m64 --cubin "/daten/blender-git/build/bin/2.70/scripts/addons/cycles/kernel/kernel.cu" -o "/home/pepo/.config/blender/2.70/cache/cycles_kernel_sm30_3C862F6180324926C4EC34BDE0F825E4.cubin" --ptxas-options="-v" --maxrregcount=32 --use_fast_math -I"/daten/blender-git/build/bin/2.70/scripts/addons/cycles/kernel" -DNVCC -D__KERNEL_CUDA_VERSION__=50
nvcc error   : 'ptxas' died due to signal 11 (Invalid memory reference)
CUDA kernel compilation failed, see console for details.

Cuda 5.0.35, Driver 331.49, specs in my signature.
Will ask for help in #blendercoders before I make a bug report.

Cheers, mib.

Hi, info from DingTo on IRC:

[18:52] <DingTo> Disable Branched Path for now
[18:52] <DingTo> Multi Light Sampling breaks sm_30

Comment out in /intern/cycles/kernel/kernel_types.h

Line 69 //#define BRANCHED_PATH

Will post here if it is fixed in master., mib.

Ok, then thanks for the fast fix. Thats what I like about Blender. :slight_smile: