I’m not using JetPack, because my host is upgraded to 17.10 and JetPack refuses to unpack either on this OS, or inside a docker image based on 16.04. I’ve downloaded the separate sources, and flashed Linux4Tegra 28.1 to my TX2, using the same root file system. So far, it works fine.
(Well, the sample root file system includes stuff like libreoffice which isn’t really all that useful for most Jetson installs, but that can be uninstalled)
I also have the “factory” image available and un-tarred. Unfortunately, this doesn’t come with a “root file system,” but instead just with a completed “system.img” image file.
When looking for “apt search cuda” on the TX2, I find nothing, so apparently the NVIDIA tools and libraries aren’t in the repository configured by the sample root file system.
How can I install the jetson-side packages, such as cuda, cudnn, tensorrt, and so forth?
Is there a repository I can add to apt that will let me install this?
Should I loop-mount the system.img and extract a new root file system that way? It seems to be an Android Sparse Image, so I don’t know how to actually extract files from it.
Is there some download on the downloads section that has these packages ready, that I’ve missed?
We don’t have too much experience on Ubuntu17.04.
This topic may give you some information:
Thanks for the answer, but it doesn’t answer my question.
Ignore what version of Ubuntu I’m using. The question is: What repository/download can I use to get the packages that support TensorRT, OpenCV4Tegra, MultimediaAPI, libargus, CUDA, cuDNN, and the rest of the Jetson-accelerated software, directly onto the Jetson?
Or, as a separate question: Where can I find a “rootfs” that contains all of those packages? The example R28 rootfs does NOT contain those packages.
You have to use wget on URLs you dig out of “_installer/repository.json”. The first one to get is the CUDA repo…this is what adds a reference to a local repository from which apt/apt-get can then install things. Many packages are unavailable in any other way.
As linuxdev mentioned, you can get the download link of each package in repository.json.
AastaLLL: You (nvidia) really should document the appropriate paths/repositories.
For what it’s worth, here are the debs it wants to install:
To access this list, I did:
- mkdir -p /opt/jetson/jetpack/_installer
- cd /opt/jetson/jetpack/_installer
- …/wherever/JetPack-L4T-3.1-linux-x64.run --tar xvf
- vim start_up.sh # remove all the things testing for whether I’m on ubuntu / 16.04 but keep the rest
- Find a copy of libpng12.so (I found mine in the Arduino distribution; you can find it elsewhere) and point LD_LIBRARY_PATH at the directory where it is.
- select custom -> select “no action” for the host things and flash-disk things
Now, there is a file called “local.ini” in the _installer directory.
grep for “url”: in that file, and then further grep for “arm” in those paths.
Now you have the list of packages to download!
This repository link may change.
To keep the link up to date, we recommended user getting the latest repository.json from JetPack each time.
There is no “repository.json” in JetPack 3.1. There is, however, a local.ini file, which contains these URLs.
Btw: For some of these packages, there are multiple versions; I assume that the TX2 wants the latest version (r21 vs r24 vs r26, for example.)
The repository.json is generated when you first run JetPack…until then it doesn’t exist. Even on a non-Ubuntu system, where JetPack is rejected, this file gets generated (you don’t have to actually use JetPack to install anything, this is a side-effect of initial run).