NVIDIA Jetson Development Environment Setup

In previous posts on the NVIDIA Jetson posts I’ve talked about getting the device setup and some additional accessories that you may want to have. The OS image for the NVIDIA Jetson already contains a compiler and other development software. Technically someone can start developing with the OS image as it is when it ships.  But it is not desirable to develop this way.

There may be some things that you prefer to do on your primary computer and you’d like to be able to control the Jetson from your primary machine. The OS image for the Jetson already has SSH enabled. If you are using a Windows machine and net an SSH client I suggest using PuTTY for Windows. It’s a great SSH client and also works as a telnet or serial console when needed. It’s available from https://www.putty.org/.

When Putty is opened by default it is ready to connect to a device over SSH. You only need to enter the IP address to start the connection. Once connected enter your account name and password and you’ll have an active terminal available. For copying files over SSHFTP I use WinSCP (available from https://winscp.net/).

For development on the device I’ve chose Visual Studio Code as my IDE. Yes, it runs on ARMs too.  There are a number of guides available on how to get Visual Studio Code recompiled and installed for an ARMS system. The one that I used is available from code.headmelted.com. In a nutshell I followed two steps; I entered a super user session with the command

su -s

Then I ran the following (which downloads a script from the head melted site and runs it).

sudo apt-key adv --keyserver keyserver.ubuntu.com --recv-keys 0CC3FD642696BFC8^C
. <( wget -O - https://code.headmelted.com/installers/apt.sh )

The script only takes a few moments to run. After it’s done you are ready to start development either directly on the board or from another machine.

To make sure that everything works let’s make our first program with the Jetson Nano. This is a “Hello World” program; it is not going to do anything substantial. Let’s also use make to compile the program. make will take care of seeing what needs to be built and issuing the necessary commands. Here it’s use is going to be trivial. But I think starting with simple use of it will give an opportunity for those that are new to it to get oriented with it. Type the following code and save it as helloworld.cu

#include 

__global__ void cuda_hello()
{
    printf("Hello World from GPU!\n");
}

using namespace std;

int main()
{
	cout << "Hello world!" << endl;
	cuda_hello<<>>();
	return 0;
}

We also need to make a new file named makefile. The following couple of lines will say that if there is no file named helloworld (or if the file is out of date based on the date stamp on helloworld.cu) the to compile it using the command /usr/local/cuda/bin/nvcc helloworld.cu -o helloworld

helloworld: helloworld.cu
   usr/local/cuda/bin/nvcc helloworld.cu -o helloworld

Note that there should be a tab on the second line, not a space.
Save this in the same folder as helloworld.cs.

Type make and press enter to build the program. If you type it again nothing will happen. That’s because make sees that the source file hasn’t changed since the executable was build.

Now type ./helloworld and see the program run.

Congratulations, you have a working build environment. Now that we can compile code it’s time to move to something less trivial. In an upcoming post I’ll talk about what CUDA is and how you can use it for calculations with high parallelism.

Advertisements

Unboxing and Setting Up the NVIDIA Jetson Nano

I pre-ordered the NVIDIA Jetson Nano and had the opportunity to have my first experiences with it this week. For those that are considering the Nano I give you the gift of my hindsight so that you can have a smoother experience when you get started. My experience wasn’t bad by any measure. But there were some accessories that I would have ordered at the same time as the Jetson so that I would have everything that I needed at the start. I’ve also made a YouTube video covering this same information. You can view it here.

How does the Nano Compare to Other Jetson Devices?

The Jetson line of devices from NVIDIA can be compared across several dimensions. But where the Jetson Nano stands out is the price. It is priced at about 100 USD making it affordable to hobbiest. Compare this to the Jetson TK1 which is available for about 500 USD or the Jetson Xaviar available for about 1,200 USD. Another dimension of interest is the number of CUDA cores that the units have. CUDA cores are hardware units used for parallel execution.

  • TK1 – 192 CUDA Cores
  • TK2 – 256 CUDA cores
  • Nano – 128 CUDA Cores
  • Xavier – 512 CUDA Cores

In addition to the cores the other Jetson kits have support for other interfaces, such as SATA for adding hard drives or a CAN bus for interfacing with automotive systems. For someone getting started with experimentation the Jetson Nano is a good start.

What is In the Box?

Not much. You’ll find the unit, a small paper with the URL of the getting started page, and a cardboard cutout used for supporting the card on the case.

What Else Do I Need

  • SD Card
  • Power Supply
  • Keyboard
  • Mouse
  • Monitor (HDMI or Display Port)
  • 40mmx40mm Cooling Fan (optional)
  • WebCam (optional)
  • WiFi adapter or Ethernet cable to router

Most of the things on that list you might already have. For an SD card get one at least 8 gigs or larger.

Power Supply

A power supply! It uses a 5 volt power supply like what is used in a phone. Well, kind of. Don’t expect for any of your 5V power supplies to work. I found the hard way that many power supplies don’t deliver the amount of current that is needed. Even if the power supply is capable a USB cable might not allow the needed amount of current to pass. If this happens the device will just cut off. There’s no warning, no error message, nothing. It just cuts off. I only came to realize what was going on after I used a USB power meter on the device. I used a power meter for USB-A, but the board already has contacts for using a USB-C port. Depending on when you get your board it may have a USB-C port on it (possibly, speculatively).

Web Cam

A Raspberry Pi camera will work. But I used a Microsoft LifeCam. There are a number of off-the-shelf webcams that work. You’ll only need a camera if you plan on performing visual processing. If your going to be processing something non-visual or if your visual data is coming from a stream (file, network location) then of course this won’t be necessary.

WiFi

You have two options for WiFi. One option is a USB WiFi dongle. There are a number of them that are compatible with Linux that will also work here. I am using  the Edimax EW-7811UN. After being connected to one of the USB ports it just works. Another solution is to install a WiFi card into the M.2 adapter. It might not be apparent at first, but there is a M.2 slot on the case. I chose to use this solution. Like the USB solution there’s not much to be done here; inserting the WiFi adapter into the slot, securing it is most of the work. Note that you’ll also need to connect antennas to the wireless card.

Operating System Image

The instructions for writing a new operating system image are almost identical to that of a Raspberry Pi. The difference is the URL from which the OS image is downloaded. Otherwise you download an image, write it to an SD card, and insert it into the Nano. Everything else will be done on first boot. You’ll want to have a keyboard connected to the device so that you can respond to prompts. When everything is done you’ll have an ARMs build of Ubuntu installed.

For writing the OS image I used balenaEtcher. It is available for OS X and Linux. The usage is simple; select an OS image, select a target drive/memory device, and then let it start writing to the card.  The process takes a few minutes. But once it is done put the SD card in the Jetson Nano’s memory card slot.

Case Options

A case may be one of the last things that you need. But if you seriously have interest in having the Jetson Nano I suggest ordering the case at the start. There are no off-the-shelf cases available for purchase for the Nano. But there are a few 3D printable plans for the Jetson Nano. I’ve come across three and have settled on one.

First Place: Nano Mesh

NanoMesh case Image
NanoMesh 3D Printable Case

The case is a bit thick, but it isn’t lacking for ventilation. The case height accommodates a fan. While the design doesn’t include any holes for mounting antennas for WiFi drilling them is easy enough.

Second Place: Nano Box

NanoBox
Nanobox Case for nVidia Jetson Nano

The NanoBox will envelope the Jetson leaving the heat sink almost flush with the case. I’d suggest this one if you plan don’t plan to use a fan on the Jetson.  If you ever change your mind and decide that you want to have a fan it can be added. But it will be on the outside of the case.

Third Place: Nano-Pac

Nano-Pac 3D printable case
Nano-Pac case

There’s not much to say about this case. It fully envelopes the Jetson Nano. But I’ve got questions about the cooling effectiveness of this case.

It’s Assembled and Boots Up. Now What?

Once the Jetson is up and running the next thing to do is to setup a development environment. There is a lot of overlap between targeting the Jetson series and targeting a PC that has an NVIDIA GPU. What I write on this will be applicable to either except for when I state otherwise.

 

TIP: Creating a New CUDA Project with Visual Studio 2019

If you’ve installed Visual Studio 2019 and are trying to work with CUDA there are a couple of problems that you’ll encounter. The first is going to be an error that you receiving when trying to opan any CUDA project about missing properties. This is from the CUDA installer placing some of the files in the wrong place. It places the files based on what was in Visual Studio 2019 Preview. It was only recently that the full release of 2019 was made available (and for the full release theses files need to go into a different place). To work around that see this post for where to move the missing files to.

Once that is resolved the next problem is that the CUDA project templates are missing. An NVidia representative in the NVidia developer forums acknowledged the problem and says a fix will come in an upcoming release. Until then the current solution is to grab an existing CUDA project and rename it. If you need an existing CUDA project you can find them in the folder for the NVidia CUDA samples or download one from here:

https://devtalk.nvidia.com/cmd/default/download-comment-attachment/78613/

 

 

Tip: Installing CUDA SDK on Visual Studio 2019

If you try to install the nVidia CUDA SDK and plan to use Visual Studio 2019 there’s an additional manual step that you’ll need to take. The installer available for the current version of CUDA (10.1) doesn’t specifically target the recently released Visual Studio 2019, but it will mostly work with it. I say “mostly” because after installing it you’ll find that the CUDA related project templates are missing and you can’t open the sample projects.

Fixing this is as simple as copying a few files.  Copying everything from the following folder

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\extras\visual_studio_integration\MSBuildExtensions

Place it into this folder

C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\MSBuild\Microsoft\VC\v160\BuildCustomizations

You may have to reply to administrative prompts. But once those files are copied you should have access to the project templates and the samples.