setting up CUDA manually on TX2

For me the Jetpack3 didnt really succeed.
It did, however, download the cuda-repo-l4t???.deb

I was able to set CUDA up following the instructions here that still work:
http://elinux.org/Jetson/Installing_CUDA

I did neither find the cuda-l4t-deb package online in the download area, nor any specific
information on setting it up manually, which is why I created this thread.

Maybe it helps someone.
Cheers,
Martin

Hi enthusi,

Do you see cuda-repo-l4t-8-0-local_8.0.64-1_arm64.deb under your jetpack_download folder?

Hi Wayne,

yes, maybe I was not clear about this. Jetpack3 failed at some late point but did download all files and I took it from there.
But I saw no other independent link/url anywhere. It is also missing from the embedded download center.

Hi enthusi,

There would have some log when jetpack install. Could you paste it?

Do you only install cuda?

Sorry for the delay.
The problem is, that I started Jetpack in the same directory again/several times.
Next time I will keep a better record of the log files.

This, however, is the complete Launcher.log

03-28 23:35:38.6 I: 0x63 0x68 0x6d 0x6f 0x64 0x3a 0x20 0x63 0x68 0x61 0x6e 0x67 0x69 0x6e 0x67 0x20 0x70 0x65 0x72 0x6d 0x69 0x73 0x73 0x69 0x6f 0x6e 0x73 0x20 0x6f 0x66 0x20 0x27 0x2f 0x74 0x6d 0x70 0x2f 0x6a 0x65 0x74 0x70 0x61 0x63 0x6b 0x5f 0x64 0x65 0x62 0x75 0x67 0x2e 0x6c 0x6f 0x67 0x27 0x3a 0x20 0x4f 0x70 0x65 0x72 0x61 0x74 0x69 0x6f 0x6e 0x20 0x6e 0x6f 0x74 0x20 0x70 0x65 0x72 0x6d 0x69 0x74 0x74 0x65 0x64 0x0a 0x6a 0x65 0x74 0x70 0x61 0x63 0x6b 0x5f 0x6e 0x65 0x74 0x77 0x6f 0x72 0x6b 0x5f 0x6c 0x61 0x79 0x6f 0x75 0x74 0x0a 
03-28 23:35:51.1 E: nextpage error occurs

and this is the head/tail of JetPackLog:

03-28 23:23:26.2 E: fopen failed, errno = 2
03-28 23:23:28.8 W: One invalid entry
03-28 23:23:28.8 W: One invalid entry
03-28 23:23:28.8 W: One invalid entry
03-28 23:23:28.8 W: One invalid entry
03-28 23:24:58.4 N: Install com.nvidia.nvtools.battle 2.4.17023;
(...)
03-28 23:35:38.0 N: Installing CUDA Toolkit 8.0.64 succeeded.
03-28 23:35:38.0 N: Installing Compile CUDA Samples 8.0.64 succeeded.
03-28 23:35:38.1 N: Installing OpenCV for Tegra 2.4.13-17 succeeded.
03-29 00:49:33.8 E: Couldn't resolve host name

I will try all this again carefully with the next L4T version.
Right now I have 27.1 set up.

enthusi,

Just confirmed that we don’t provide a link to download cuda for L4T now. The only way to install cuda toolkit on tegra is through jetpack.

Hi Wayne,
thanks for that info. This is most unfortunate, however.
May I ask, is there any particular reason that host should be Ubuntu 14.04 and not 16.xx?

You may find this as a workaround…I use a Fedora host so I also can’t use JetPack. Someone else figured this out a while back, and the trick is a life saver.
https://devtalk.nvidia.com/default/topic/982848/jetson-tx1/tx1-specific-arm64-deb-repo-for-cuda-8/post/5063053/#5063053

To get JetPack to unpack on a host which isn’t the correct Ubuntu:

bash ./JetPack-L4T-2.3.1-linux-x64.run --noexec

Running “./Chooser” produces produces human-readable “repository.json”…URLs are in this. For JetPack 3 this implies:

wget http://developer.download.nvidia.com/devzone/devcenter/mobile/jetpack_l4t/006/linux-x64/nv-gie-repo-ubuntu1604-6-rc-cuda8.0_1.0.2-1_arm64.deb

If you copy that “.deb” file to the Jetson and use dpkg to install it, then you will be able to search for, find, and install anything provided in that local repo via apt.

Thank you!
I wish there was a less ‘cryptic’ channel for such actions at some point.
I am well aware of the dev-board status, however :)

Edit:
and in fact I am quite happy with your prompt supports here in the forum!

Resolved. The message removed.

Thanks guys - good to know!

"May I ask, is there any particular reason that host should be Ubuntu 14.04 and not 16.xx?
enthusi,
The r28.1 release was mainly based and tested on Ubuntu 14.04. The upcoming r28.2 will have official support for 16.04. Sorry for the inconvenience that should end very soon :)

for me the JetPack 3.3 installer just hangs after copy cuda-repo…deb file to TX2 (over ssh)
it asks for ‘nvidia’ ssh pass in new terminal window. then does file copy.
then nothing happens.

if i kill it after an hour i see nothing in /usr/local – no ‘cuda’ directory in there.

no ‘nvcc’ anywhere on TX2 either.

my Ubuntu host where I run jetpack 3.3 is 16.04.

so… is there manual install of CUDA 9.0 on TX2?

Check first if command line ssh works from your host. Check both of these from the regular user doing the JetPack run:

ssh nvidia@<whatever_the_address_is>
sudo ssh nvidia@<whatever_the_address_is>

See if ssh is offering clues. If ssh fails in JetPack using the same address which works on command line, then you’ll at least be able to rule out ssh being picky as a cause.

yes, ssh works fine.
i think something was wrong with JetPack 3.3
went to JetPack 3.2, done full reinstall without any issues.

I am wondering if it is possible now to install Cuda without jetpack?

No other method is supported (and the supported method is much easier), and only JetPack can be used to download CUDA for a Jetson…at least normally. Unofficially there are essentially two options (both require extra work, but one requires far more work than the other). Is this JetPack4.2? Or something earlier? The answer can change depending on which one.

The basic theme is that both can be told to not delete temporary files, and you tell it you are going to add packages (deselect every other operation), and then upon actual install don’t install packages. Depending on release the packages can then be found on the host PC for manual copy to the Jetson and manual install.

Furthermore, there is also a manifest downloaded. The manifest contains a base web server URL, and then notation exists describing each download URL for specific files…you can use wget to download files that way as well. Then you must still copy to the Jetson and manually install.

For SDK Manager 4.2 here is some information (earlier JetPack will have close similarities):
https://devtalk.nvidia.com/default/topic/1050477/jetson-tx2/jetpack4-2-flashing-issues-and-how-to-resolve/post/5340560/#5340560