|
Size: 2287
Comment:
|
Size: 3292
Comment: Update to use the new CUDA toolkit SEPP package
|
| Deletions are marked like this. | Additions are marked like this. |
| Line 7: | Line 7: |
| Line 11: | Line 10: |
| Line 14: | Line 12: |
| === Header Files === * The OpenGL headers are found under `/usr/include/GL`. For NVidia programming, a CUDA toolkit matching the installed NVidia graphics card is installed (CUDA headers in `/usr/include`). |
=== OpenGL header files === The OpenGL headers are found under `/usr/include/GL`. === CUDA header files === For NVidia programming, a CUDA toolkit matching the installed NVidia driver is selected when a toolkit program like `nvcc`is run.<<BR>> Any CUDA headers have to bee included by manually matching the toolkit version used. Check the content of the SEPP package [[https://www.sepp.ee.ethz.ch/sepp-debian/cuda_toolkit-1x.x-sr.html|cuda_toolkit-1x.x-sr]] for available versions by listing its contents with `ls /usr/pack/cuda_toolkit-1x.x-sr` Usually, there is a lower driver/toolkit version combo for desktop PCs and a higher version combo for Slurm GPU nodes: * For desktop PCs: driver version `418` works with toolkit version `10.1.243` * For Slurm GPU nodes: driver version `460` works with toolkit version `11.2.2` The following environment variables have to be set with the correct driver or toolkit version: {{{#!highlight bash numbers=disable export CPPFLAGS='-I/usr/pack/cuda_toolkit-1x.x-sr/<driver or toolkit version>/include'` export LDFLAGS='-L/usr/pack/cuda_toolkit-1x.x-sr/<driver or toolkit version>/lib64 -Wl,-rpath,/usr/pack/cuda_toolkit-1x.x-sr/<driver or toolkit version>/lib64' ` }}} |
| Line 18: | Line 30: |
| * Make sure no special `CPPFLAGS` and `LDFLAGS` are set. The required headers and libraries are all found under the standard system paths. If you use your own OpenGL/ CUDA toolkit, e.g. installed in your home, make sure it is compatible with the graphics driver/ hardware installed on the system. Note: if you don't have brand new hardware in your computer, the most current CUDA release most probably won't work with it - use an older relase or ask ISG.EE if you are in doubt about the right version. * Then set `CPPFLAGS` (for the C preprocessor used by most compilers) and `LDFLAGS` (linker flags) as follows: |
|
| Line 19: | Line 33: |
| * Make sure no special `CFLAGS`,`CXXFLAGS` and `LDFLAGS` are set. The required headers and libraries are all found under the standard system paths. If you use your own OpenGL/ CUDA toolkit, e.g. installed in your home, make sure it is compatible with the graphics driver/ hardware installed on the system. Note: if you don't have brand new hardware in your computer, the most current CUDA release most probably won't work with it - use an older relase or ask ISG.EE if you are in doubt about the right version. * Then set `CFLAGS` (for C-compiler), `CXXFLAGS` (for C++ compiler) and `LDFLAGS` (linker flags) as follows: Assuming you installed the toolkit under `/home/$USER/toolkit` you set the mentioned environment variables in your Debian shell (`bash`, `tcsh`, ...) as follows, before you start compiling/ linking: {{{ export CFLAGS="-I/home/$USER/toolkit/include" export CXXFLAGS="$CFLAGS" |
Assuming you installed the toolkit under `/home/$USER/toolkit` you set the mentioned environment variables in your Debian shell (`bash`, `tcsh`, ...) as follows, before you start compiling/ linking: {{{#!highlight bash numbers=disable export CPPFLAGS="-I/home/$USER/toolkit/include" |
| Line 28: | Line 38: |
| `$USER` must of course be replaced with your ETH (login-)username. | `$USER` must of course be replaced with your ETH (login-)username. |
| Line 31: | Line 41: |
OpenGL
Compiling OpenGL/ NVidia CUDA-based applications on Debian Linux
The Debian platform managed by ISG.EE uses the following directory layout for the OpenGL libraries and headers:
Libraries
The /usr/lib/x86_64-linux-gnu/libGL*.so OpenGL libraries are symbolic links to the hardware-specific libraries. There is a mechanism in the startup process of each client workstation which detects the accurate OpenGL library based on the information about the graphics card and the kernel driver. The CUDA runtime libraries are found in the same directory.
OpenGL header files
The OpenGL headers are found under /usr/include/GL.
CUDA header files
For NVidia programming, a CUDA toolkit matching the installed NVidia driver is selected when a toolkit program like nvccis run.
Any CUDA headers have to bee included by manually matching the toolkit version used. Check the content of the SEPP package cuda_toolkit-1x.x-sr for available versions by listing its contents with ls /usr/pack/cuda_toolkit-1x.x-sr
Usually, there is a lower driver/toolkit version combo for desktop PCs and a higher version combo for Slurm GPU nodes:
For desktop PCs: driver version 418 works with toolkit version 10.1.243
For Slurm GPU nodes: driver version 460 works with toolkit version 11.2.2
The following environment variables have to be set with the correct driver or toolkit version:
export CPPFLAGS='-I/usr/pack/cuda_toolkit-1x.x-sr/<driver or toolkit version>/include'`
export LDFLAGS='-L/usr/pack/cuda_toolkit-1x.x-sr/<driver or toolkit version>/lib64 -Wl,-rpath,/usr/pack/cuda_toolkit-1x.x-sr/<driver or toolkit version>/lib64' `
Compiling sources and linking against NVidia libraries
Make sure no special CPPFLAGS and LDFLAGS are set. The required headers and libraries are all found under the standard system paths. If you use your own OpenGL/ CUDA toolkit, e.g. installed in your home, make sure it is compatible with the graphics driver/ hardware installed on the system. Note: if you don't have brand new hardware in your computer, the most current CUDA release most probably won't work with it - use an older relase or ask ISG.EE if you are in doubt about the right version.
Then set CPPFLAGS (for the C preprocessor used by most compilers) and LDFLAGS (linker flags) as follows:
Assuming you installed the toolkit under /home/$USER/toolkit you set the mentioned environment variables in your Debian shell (bash, tcsh, ...) as follows, before you start compiling/ linking:
export CPPFLAGS="-I/home/$USER/toolkit/include" export LDFLAGS="-L/home/$USER/toolkit/lib -Wl,-rpath,/home/$USER/toolkit/lib"
$USER must of course be replaced with your ETH (login-)username.
Recomendation
That said we encourage the use of CUDA Toolkit within Conda environments. More details can be found in the Working with GPU or CPU in data sciences and Conda subpages of this wiki.