Repository navigation
UnsatisfiedLinkError with Fedora 40 (cuda 12.6) #45
Description
The latest version of jcuda on Maven is 12.0.0, but when I use that I get
Stack trace from the attempt to load the library as a file:
java.lang.UnsatisfiedLinkError: no JCudaDriver-12.0.0-linux-x86_64 in java.library.path: /usr/local/cuda/lib:/usr/local/cuda/lib::/usr/java/packages/lib:/usr/lib64:/lib64:/lib:/usr/lib
at java.base/java.lang.ClassLoader.loadLibrary(ClassLoader.java:2434)
at java.base/java.lang.Runtime.loadLibrary0(Runtime.java:818)
at java.base/java.lang.System.loadLibrary(System.java:2006)
at jcuda.LibUtils.loadLibrary(LibUtils.java:168)
at jcuda.LibUtilsCuda.loadLibrary(LibUtilsCuda.java:68)
at jcuda.driver.JCudaDriver.(JCudaDriver.java:350)
at ...
...
Stack trace from the attempt to load the library as a resource:
java.lang.UnsatisfiedLinkError: /tmp/libJCudaDriver-12.0.0-linux-x86_64.so: libcuda.so.1: cannot open shared object file: No such file or directory
at java.base/jdk.internal.loader.NativeLibraries.load(Native Method)
at java.base/jdk.internal.loader.NativeLibraries$NativeLibraryImpl.open(NativeLibraries.java:388)
at java.base/jdk.internal.loader.NativeLibraries.loadLibrary(NativeLibraries.java:232)
at java.base/jdk.internal.loader.NativeLibraries.loadLibrary(NativeLibraries.java:174)
at java.base/java.lang.ClassLoader.loadLibrary(ClassLoader.java:2394)
at java.base/java.lang.Runtime.load0(Runtime.java:755)
at java.base/java.lang.System.load(System.java:1970)
at jcuda.LibUtils.loadLibraryResource(LibUtils.java:285)
at jcuda.LibUtils.loadLibrary(LibUtils.java:183)
at jcuda.LibUtilsCuda.loadLibrary(LibUtilsCuda.java:68)
at jcuda.driver.JCudaDriver.(JCudaDriver.java:350)
which does make somem sense I guess, as the current libraries are from cuda-libraries-devel-12-6-12.6.2-1.x86_64
I'm not sure if I can make a simple work around, need to simply rebuild against this version, or better roll back to the formally obsolete Fedora 39 :(
Bram
Activity
jcuda commentedon Nov 14, 2024
The CUDA version and the JCuda version must match, insofar that it will not be possible to use JCuda 12.0.0 to access CUDA 12.6.
I know that an update is overdue, but ... I'm not sooo actively maintaining JCuda (for many reasons).
However, an update for the latest CUDA version is already in the pipeline (as a follow-up of #44 (comment) ). I already downloaded and installed the latest CUDA version, and need to do an update pass for possible changes in cuDNN, and then will hopefully be able to create a new release (for CUDA 12.6) soon.
jcuda commentedon Nov 14, 2024
An aside:
This is theoretically possible, but it's not clear (yet) whether there have been breaking changes in the API (meaning that you would have to update the code). If you can compile the
.sofiles, it should be possible to load them (with caveats).btbouwens commentedon Nov 15, 2024
OK, trying that. Cmake complains about the
cmake_minimum_required(VERSION 3.1)being deprecated, and it doesn't knowCUDA_CUDA_LIBRARYandCUDA_cuDNN_LIBRARY, but that should be easy to fix I think.btbouwens commentedon Nov 15, 2024
There are some other differences: the newer cmake has a different way of finding the CUDA toolkit, and/or it's in a different place.
Apart from that, nvcc doesn't like gcc-14, a work around could be to use clang instead, there's a trick for that I think.
btbouwens commentedon Nov 15, 2024
Actually, the problem I had originally is much simpler: I simply didn't have
nvidia-driver-cuda-libs-3:565.57.01-1.fc40.x86_64installed, which contains the required libraries.To fix, it's just
sudo dnf -y module install nvidia-driver:open-dkms.This doesn't fix the issue with gcc-14 though, for that the
"-ccbin /usr/bin/clang++ "helps in thenvcccall.jcuda commentedon Nov 15, 2024
I cannot say much about the specifics of the Linux setup and compilation. But
You may have to manually set these to the
.libfiles that correspond toC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.6\lib\x64\cuda.liband
C:\Program Files\NVIDIA\CUDNN\v9.5\lib\12.6\x64\cudnn.libon Linux.
(Similarly, it may not find the proper
_INCLUDE...directories, and these may also have to be set manually).There's that
FindCUDA.cmakescript that usually does most of the work, but ... it may not always work and may not always be up to date.