How do I make the l4t-Cuda container access to cudnn?

Robotics & Edge Computing Jetson Systems Jetson Nano
,
1

Hi sir,
I use a x86 host computer to compile the c++ CUDA program for the jetson nano. I use l4t-CUDA docker image to compile the code and copy the program into the l4t-base image, then put the docker image into the jetson machine.

Now I’m trying to use torch. I use the library in the pre-built image l4t-pytorch. However, The compiler ask me for the cudnn library like:

//home/gino/libtorch/lib//libtorch_cuda.so: undefined reference to cudnnDropoutGetStatesSize@libcudnn.so.8' //home/gino/libtorch/lib//libtorch_cuda.so: undefined reference to cudnnSetFilterNdDescriptor@libcudnn.so.8’
/home/gino/tesseract_install/lib//libtesseract.so: undefined reference to log2@GLIBC_2.29' //home/gino/libtorch/lib//libtorch_cuda.so: undefined reference to cudnnDestroySpatialTransformerDescriptor@libcudnn.so.8’
/home/gino/tesseract_install/lib//libtesseract.so: undefined reference to pow@GLIBC_2.29' //home/gino/libtorch/lib//libtorch_cuda.so: undefined reference to cudnnCreateLRNDescriptor@libcudnn.so.8’
//home/gino/libtorch/lib//libtorch_cuda.so: undefined reference to cudnnRestoreDropoutDescriptor@libcudnn.so.8' //home/gino/libtorch/lib//libtorch_cuda.so: undefined reference to cudnnDestroyCTCLossDescriptor@libcudnn.so.8’
//home/gino/libtorch/lib//libtorch_cuda.so: undefined reference to cudnnBatchNormalizationBackwardEx@libcudnn.so.8' //home/gino/libtorch/lib//libtorch_cuda.so: undefined reference to cudnnFindConvolutionBackwardDataAlgorithm@libcudnn.so.8’
//home/gino/libtorch/lib//libtorch_cuda.so: undefined reference to `cudnnFindConvolutionBackwardFilterAlgorithmEx@libcudnn.so.8’
…

The cudnn library seems not available in l4t-CUDA container. I also take a look other container but none of them are like l4t-CUDA image which has whole cuda libraries so I can compile the code. How do I get the l4t version of cudnn library for compiling in the container?

3

Hi,

The base image only has CUDA library pre-installed.

You can manually install it with the aarch64 cuDNN package (available from SDK manager).
Or you can use TensorRT container (CUDA, cuDNN and TensorRT) instead.

Thanks.