UnsatisfiedLinkError with cudnn on Windows 10

DL4J
1

I’ve done the following

  1. built my Java program with beta7 and Cuda 10.2 dependencies on Windows 10.
  2. installed Cuda10.2 and cudnn10.2. (copied dlls from cudnn/bin into Cuda10.2 bin)
  3. Set “-Djava.library.path=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\bin”

and get the following…

[main] INFO org.nd4j.linalg.factory.Nd4jBackend - Loaded [JCublasBackend] backend
[main] INFO org.nd4j.nativeblas.NativeOpsHolder - Number of threads used for linear algebra: 32
[main] INFO org.nd4j.linalg.api.ops.executioner.DefaultOpExecutioner - Backend used: [CUDA]; OS: [Windows 10]
[main] INFO org.nd4j.linalg.api.ops.executioner.DefaultOpExecutioner - Cores: [8]; Memory: [7.1GB];
[main] INFO org.nd4j.linalg.api.ops.executioner.DefaultOpExecutioner - Blas vendor: [CUBLAS]
[main] INFO org.nd4j.linalg.jcublas.JCublasBackend - ND4J CUDA build version: 10.2.89
[main] INFO org.nd4j.linalg.jcublas.JCublasBackend - CUDA device 0: [Quadro M2200]; cc: [5.2]; Total memory: [4294967296]
[main] INFO org.deeplearning4j.nn.multilayer.MultiLayerNetwork - Starting MultiLayerNetwork with WorkspaceModes set to [training: ENABLED; inference: ENABLED], cacheMode set to [NONE]
[main] WARN org.deeplearning4j.nn.layers.convolution.ConvolutionLayer - Could not initialize CudnnConvolutionHelper
java.lang.reflect.InvocationTargetException
…
Caused by: java.lang.UnsatisfiedLinkError: no jnicudnn in java.library.path
at java.lang.ClassLoader.loadLibrary(ClassLoader.java:1867)
…
Caused by: java.lang.UnsatisfiedLinkError: C:\Users\DavidWood.javacpp\cache\cuda-10.2-7.6-1.5.3-windows-x86_64.jar\org\bytedeco\cuda\windows-x86_64\jnicudnn.dll: The specified procedure could not be found
…

If I run the following (recommended on other posts):

		try {
		    Loader.load(cudnn.class);
		} catch (UnsatisfiedLinkError e) {
		    String path = Loader.cacheResource(cudnn.class, "windows-x86_64/jnicudnn.dll").getPath();
		    new ProcessBuilder("c:/temp/dependencies/DependenciesGui.exe", path).start().waitFor();
		}

I see that it can’t find cudnn64_7.dll. The cudnn 10.2 I installed contains cudnn64_8.dll. I copied this file to cudnn64_7.dll, but no joy.

Thoughts?

2

Just include deeplearning4j-cudnn as a dependency. No need to do anything manually.
More here: https://deeplearning4j.konduit.ai/config/backends/config-cudnn

3

I already have this dependency

<dependency>
    <groupId>org.deeplearning4j</groupId>
    <artifactId>deeplearning4j-cuda-10.2</artifactId>
    <version>1.0.0-beta7</version>
</dependency>

The page you reference above says that dependency plus a manual installation of cudnn is what is required. Are you suggesting something like this

<dependency>
    <groupId>org.deeplearning4j</groupId>
    <artifactId>deeplearning4j-cudnn</artifactId>
    <version>1.0.0-beta7</version>
</dependency>

I see no such artifact on the maven repo, so probably not?

4

Ah sorry yeah. Ok I thought you only had nd4j cuda. Try to use the redist artifact (this is also covered in the page I linked.)

5

I’d prefer not to use this solution as it would require me to distribute Nvidia’s code per their license. I would like to have my end customers download their code and accept the license themselves.

Alternatively, if I could find the cudnn v7 to download that would probably work, but I can’t find it at the NVidia site. Might there be a way to gracefully, in the future, allow an upgrade of cudnn or other libraries (I’m surprised renaming the file didn’t work), otherwise dl4j w/gpu on windows may age out like this in the future.

6

@dawood just use it for now for development. We don’t have that risk in practice. We always upgrade to the latest version of cuda/cudnn before doing a release.
If you’re paranoid about cudnn and co here’s how we handle that (we maintain javacpp as well):

Anaconda also redistributes anaconda fine: Cudnn :: Anaconda.org

7

In fact, I was able to find the cudnn7 downloads on NVidia at cuDNN Archive | NVIDIA Developer. I’m up and running and running into other issues. I’ll make a separate post.

8

Thanks yes please do!

9

Actually its all good and I’m up and running! Thanks for your help.