cuDNN for TX1

Trying to download <cuDNN v4 Library for L4T (ARM64)> for TX1, but the link in the cuDNN download page after logon doesn’t work:
https://developer.nvidia.com/rdp/cudnn-download

Wondering where should I report this. Thanks a lot!

Hmm you appear correct - the downloading fails “no file”. We’ll look into it.

Also currently I think you might want to use “cuDNN v4 Library for L4T (ARMv7)” because of TX1 armhf root filesystem, which I have confirmed downloads ok: [url]https://developer.nvidia.com/rdp/assets/cudnn-70-linux-aarch64-v40[/url]

OK, it’s back up — thanks for reporting it.

Unfortunately this does not work with Torch… :(
Can someone fix the link?

Updated the arm64 link above.
For using Torch + cuDNN v4 on the default armhf sample root filesystem, you probably want to use the v4 bindings from [url]https://github.com/soumith/cudnn.torch/tree/R4[/url] and armhf cuDNN from here: [url]https://developer.nvidia.com/rdp/assets/cudnn-70-linux-armv7-v40[/url]

Hi.

Where is the cuDNN link for the ARM64 version for the TX1? Will the armhf version also work? I flashed the TX1 using Jetpack 3.1 (cuDNN 6.0). The deb packages are the /home/nvidia/cudnn directory. The packages have already been installed.

[url]https://developer.nvidia.com/rdp/assets/cudnn-70-linux-armv7-v40[/url]

nvidia@tegra-ubuntu:~$ uname -a
Linux tegra-ubuntu 4.4.38-jetsonbot-v0.2 #4 SMP PREEMPT Thu Sep 21 17:08:48 EDT 2017 aarch64 aarch64 aarch64 GNU/Linux

nvidia@tegra-ubuntu:~$ echo $LD_LIBRARY_PATH
/home/nvidia/torch/install/lib:/usr/local/cuda-8.0/lib64::/usr/local/cuda/lib64

I need to config the TX1 system for Torch extensions:
cp cudnn.h /usr/local/cuda-8.0/include
cp libcudnn* /usr/local/cuda-8.0/lib64


luarocks install cudnn


Update:
I went ahead and used the ARM7 cudnn files to perform the luarocks cudnn install. I ran into the following install message:
– CuDNN 5.1 not found at install-time. Please make sure it’s in LD_LIBRARY_PATH at runtime

The installation succeeded, but there could be an issue with using cudnn Release 4.

Thanks.

Hi,

cuDNN’s libraries is located at ‘/usr/lib/aarch64-linux-gnu/’

Thanks.

Hi,

For this link, https://developer.nvidia.com/rdp/assets/cudnn-70-linux-armv7-v40.

Is it CUDNN version 7.0? I got version 4.0 from its cudnn.h,
nvidia@tegra-ubuntu:/usr/local/cuda-8.0/include$ grep CUDNN_MA cudnn.h
#define CUDNN_MAJOR 4

I have a YOLO2 application running with CUDA8.0/CUDNN 6.0 on x86 Ubuntu 16.04 fine. When I ported it to Jetson TX1 with CUDA 8.0/above CUDNN 7.0 (or 4.0?), I got build error,
./src/convolutional_layer.c: In function ‘cudnn_convolutional_setup’:
./src/convolutional_layer.c:127:5: error: too many arguments to function ‘cudnnSetFilter4dDescriptor’
cudnnSetFilter4dDescriptor(l->dweightDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
^
compilation terminated due to -Wfatal-errors.
Makefile:79: recipe for target ‘obj/convolutional_layer.o’ failed

This looks like library incompatible issue.

thanks,

Steve

Hi,

From the error log, you are in the different version of cuDNN.

Please install Jetson with JetPack3.1.
JetPack 3.1 contains CUDA 8 and cuDNN 6. Align with your requirement.

Thanks.

Hi AastaLLL,

Thanks for the info. Actually I already have JetPack3.1/CUDA 8 installed. I couldn’t find where is my cuDNN so I just downloaded/installed https://developer.nvidia.com/rdp/assets/cudnn-70-linux-armv7-v40.

Is there a link for JetPack3.1’s cuDNN6? I would like to install it manually to avoid time consuming re-installation of everything.

thanks a lot!

Steve

Hi,

We don’t release individual cuDNN package for Jetson anymore.
All the package should be installed via JetPack.

The link you shared is cuDNNv4 and for 32-bit L4T. Can’t be used on rel-28.1, which is already 64-bit.

cuDNN can be download and installed via JetPack directly.
Package is located at /usr/lib/aarch64-linux-gnu/

Thanks.

thanks for the clarification, AastaLLL,

Steve

Hi Experts

I have installed tensorflow (python3.5) from jetsonhack. When I import
tensorflow in python I get:

ImportError: libcudnn.so.6: cannot open shared object file: No such file or directory

Any hints?

In my folder usr/lib/aarch64-linux/gnu

I have a file libcidnn.so.5

(In addition I have folders cuda and cuda-8.)

Thanks
Sojohans