CUDA Applications

The following resources are available to help you build and run CUDA sample applications:

The installations of CUDA Host x86 and Linux AArch64 are in the /usr/local/cuda-11.4/ folder.
The source code files of all CUDA samples are in the /usr/local/cuda-11.4/samples folder.
For more information about CUDA toolkit, see CUDA Toolkit Documentation v11.4.1.

Building CUDA Sample Applications for the Linux Target

Host cross-compile is supported for DRIVE OS releases only. After you finish installing CUDA x86 and cross-compile packages, perform the following steps:

Install the CUDA sample sources to a directory where you do not need root privileges to write, such as the $HOME directory as shown in the following example:
```
$ cd ~/
```
- If you installed the driveos-oobe-desktop filesystem, use the following command to run the cuda-install-samples-11.4.sh script from the CUDA installation in the target file system:
```
$ <$NV_WORKSPACE>/drive-linux/filesystem/targetfs/usr/local/cuda-11.4/bin/cuda-install-samples-11.4.sh .
```
- If you installed the default Debian local repo, install from the host CUDA directory and remove the nvJPEG folders:
```
$ /usr/local/cuda-11.4/bin/cuda-install-samples-11.4.sh .
$ cd ~/NVIDIA_CUDA-11.4_Samples/7_CUDALibraries/
$ rm -r nvJPEG*
$ cd ~/NVIDIA_CUDA-11.4_Samples
```
Build the sample application.
```
$ make TARGET_ARCH=aarch64 TARGET_OS=linux SMS=87
```
Note: This command will build a sample application for the Orin GPU (SMS=87). To build CUDA applications with other GPUs, replace the SMS value with the corresponding target compute version.

Running CUDA Sample Applications

To run a CUDA sample application,

Copy the sample files of your choice to the target.
From the target, run the sample application.

For example,

$ cd ~/
$ rcp -r <username>@<host ip address>:/path/to/NVIDIA_CUDA-11.4_Samples/ .
$ cd NVIDIA_CUDA-11.4_Samples/1_Utilities/deviceQuery
$ ./deviceQueryDrv // you can launch any sample application