Bug 24379

Assigning to the registered maintainer.

CC: (none) => marja11
Assignee: bugsquad => geiger.david68210

Latest git snapshot from branch blender2.7 is now available on Cauldron, please test it, thanks!

blender-2.79b-14.git20190219.1.mga7

Yep, that's fixed it - thank you very much

Resolution: (none) => FIXED
Status: NEW => RESOLVED

Ouch. Sorry. Blender started without complaint but fails on compiling the CUDA kernel for a render.

It now mentions a gcc problem;

CUDA version 9.1 detected, build may succeed but only CUDA 8.0 is officially supported.
Compiling CUDA kernel ...
"nvcc" -arch=sm_61 --cubin "/usr/share/blender/2.79/scripts/addons/cycles/source/kernel/kernels/cuda/kernel.cu" -o "/home/richard/.cache/cycles/kernels/cycles_kernel_sm61_7541DDBE6B1A613331389550DF3BCB6B.cubin" -m64 --ptxas-options="-v" --use_fast_math -DNVCC -I"/usr/share/blender/2.79/scripts/addons/cycles/source" 
In file included from /usr/include/host_config.h:50,
                 from /usr/include/cuda_runtime.h:78,
                 from <command-line>:
/usr/include/crt/host_config.h:121:2: error: #error -- unsupported GNU version! gcc versions later than 6 are not supported!
 #error -- unsupported GNU version! gcc versions later than 6 are not supported!
  ^~~~~
CUDA kernel compilation failed, see console for details.


Sorry for being so careless.

Resolution: FIXED => (none)
Status: RESOLVED => REOPENED

Can you report this bug upstream, please?

OK, I can report it upstream, but I am not sure what I would say. The 2.79b release from Blender.org works just fine and I don't know how the Mageia version was built.

How would you describe the problem?

Hmmm. From a very quick poke around it seems the gcc compiler complaint is coming from the nvidia side (/usr/include/crt/host_config.h:121:2). That still doesn't explain why all of the blender.org builds (2.79b, 2.79 and 2.80 nightly builds) all manage to render with CUDA on my GPU. 

It's late now. I'll take another look after work tomorrow.

Richard

That took longer than I expected, but I think I understand the problem now.
The most significant difference between the current Cauldron Blender "2.79b" rpm, built from git 2.79 branch on 19 Feb 2019, and the build from Blender.org dailies, timed at 15:34 on the same day, is that the Cauldron version starts rendering by compiling a CUDA kernel (which fails as noted above) whereas the Blender-built version goes straight in to rendering with no delay.

The reason for the difference is that the Blender.org dailies come with pre-compiled CUDA kernels in the 2.79/scripts/addons/cycles/lib directory so the CUDA kernel compile step is skipped.

Presumably the Blender devs have more recent CUDA headers than those shipped in our nvidia-cuda-toolkit-devel package (CUDA 9.1.85 dating from some time in 2017).

It may solve the problem if our CUDA packages were updated to CUDA 10.0.130 but that might prevent users, with nvidia hardware requiring pre-410.48 nvidia drivers, from getting CUDA accelerated rendering in Blender.

Perhaps we should consider driver-versioned rpms for the CUDA toolkit stuff too....

Meanwhile I can show that copying the Blender.org-supplied cycles addon from blender-2.79-667033e89e7f-linux-glibc224-x86_64/2.79/scripts/addons/cycles to ~/.config/blender/2.79/scripts/addons/cycles does indeed allow the Cauldron rpm version to start and render using my nvidia gpu.

Would you agree that this bug can be closed as "not valid" or whatever, and a new bug be raised against nvidia-cuda-toolkit-devel?

Yes for me as we don't provides any pre-compiled sources in ours packages!

So make more sense in a cuda driver version issue.

Opened a new bug for this: 24411

Thank you for your help, David

Status: REOPENED => RESOLVED
Resolution: (none) => INVALID