Jetpack 4.4 DP issues with building with cudnn from docker

Hi!
I have an issue with building programs using the pytorch c++ frontend through docker using the latest jetpack.

I get the following error,
make[2]: *** No rule to make target ‘/usr/lib/aarch64-linux-gnu/libcudnn.so’, needed by ‘example-app’. Stop.

torch lib for c++ is built from source, and works as intended outside the docker image. The docker image in use is l4t-ml:r32.4.2-py3

It think that the root of my issue is the change in cudNN that was implemented in jetpack 4.4 by splitting it up.
In Jetpack 4.3 i got it working by manually collecting the libcudnn.so and cudnn.h from the host and inserting them at the same location in the docker image. But now, using jetpack 4.4 the header is split up in multiple header files. These i have also manually placed at the same location.

Is the libcudnn.so file also dependent on something i have forgotten to move? I can see that the filesize is much smaller than earlier. ~ 500MB vs 150KB in the current version.

Hi,

Just check nvcr.io/nvidia/l4t-ml:r32.4.3-py3 on JetPack4.4 DP.
The file libcudnn.so is located at /usr/lib/aarch64-linux-gnu/

root@xaiver8G:/# find / -iname libcudnn.so
/usr/lib/aarch64-linux-gnu/libcudnn.so

A possible cause is that cuDNN version information is moved from /usr/include/cudnn.h into /usr/include/cudnn_version.h.
So some corresponding change is required in the CMakeList.txt.

Thanks.