You are not logged in.

#1 2017-09-09 15:29:35

bartus
Member
Registered: 2013-05-13
Posts: 48

glibc 2.26 breaks blender cycles cuda kernel compilation.

Lately I'm having a problem when building blender cycles CUDA kernel. Got this error:

[ 19%] Generating kernel_sm_52.cubin
/usr/include/bits/floatn.h(61): error: invalid argument to attribute "__mode__"

/usr/include/bits/floatn.h(73): error: identifier "__float128" is undefined

/home/bartus/_src/blender-git/intern/cycles/kernel/../util/util_math.h(227): warning: integer conversion resulted in a change of sign

2 errors detected in the compilation of "/tmp/tmpxft_00004cdd_00000000-7_kernel.cpp1.ii".

Gcc5 (CUDA 8.0 host compiler) throws this errors when getting to this new files introduced with glibc 2.26
Previous version of glibc 2.25, which lack this file, works fine.

Last edited by bartus (2017-09-09 15:30:44)

Offline

#2 2017-09-10 22:09:54

loqs
Member
Registered: 2014-03-06
Posts: 17,197

Re: glibc 2.26 breaks blender cycles cuda kernel compilation.

https://bugs.archlinux.org/task/55580?project=5
Have you reported this upstream against CUDA? nvcc is taking preprocessed source from a compiler that does support __float128 gcc-5 so passes the preprocessor checks but nvcc does not support it.
Edit:
replace MACRO with preprocessor

Last edited by loqs (2017-09-10 23:01:23)

Offline

#3 2017-09-13 13:35:20

Svenstaro
Administrator
From: Germany
Registered: 2008-11-19
Posts: 388

Re: glibc 2.26 breaks blender cycles cuda kernel compilation.

Should be fixed now, please test with current CUDA.

Offline

Board footer

Powered by FluxBB