You are not logged in.
I see that CUDA-9.2 from AUR can be installed together with CUDA-10.1 from official repos. Both packages install profile scripts and configs for LD:
[viktor@desolve-lab ~]$ pacman -Ql cuda | grep " /etc"
cuda /etc/
cuda /etc/ld.so.conf.d/
cuda /etc/ld.so.conf.d/cuda.conf
cuda /etc/profile.d/
cuda /etc/profile.d/cuda.sh
[viktor@desolve-lab ~]$ pacman -Ql cuda-9.2 | grep " /etc"
cuda-9.2 /etc/
cuda-9.2 /etc/ld.so.conf.d/
cuda-9.2 /etc/ld.so.conf.d/cuda-9.2.conf
cuda-9.2 /etc/profile.d/
cuda-9.2 /etc/profile.d/cuda-9.2.sh
I see that these scripts and configs set the same variable and paths:
[viktor@desolve-lab ~]$ cat /etc/ld.so.conf.d/cuda.conf
/opt/cuda/lib64
/opt/cuda/nvvm/lib64
/opt/cuda/extras/CUPTI/lib64
[viktor@desolve-lab ~]$ cat /etc/ld.so.conf.d/cuda-9.2.conf
/opt/cuda-9.2/lib64
/opt/cuda-9.2/lib
/opt/cuda-9.2/nvvm/lib64
/opt/cuda-9.2/nvvm/lib
/opt/cuda-9.2/extras/CUPTI/lib64
[viktor@desolve-lab ~]$ cat /etc/profile.d/cuda.sh
export PATH=$PATH:/opt/cuda/bin
[viktor@desolve-lab ~]$ cat /etc/profile.d/cuda-9.2.sh
export PATH=$PATH:/opt/cuda-9.2/bin
I have some applications which uses CUDA-9.2 and now I want to compile some another package which requires CUDA-10.1. So it's obvious that build system will become confused because of two sets of CUDA binaries and libs. The most straightforward way is to stash one or another set of config files/scripts but I wonder if there are any real better practice to use two CUDAs in parallel.
Last edited by dviktor (2019-04-19 09:42:15)
Offline