I have been using Cuda 10.0 which is installed manually I’ve got Cuda 10.0 from the Jetpack’s json file which has a url of downloadable cuda-repo-l4t-10-0-local-10.0.326_1.0-1_arm64.deb file. Without using the Jetpack, installed the Cuda with systemd-nspawn container and then flashed the Jetson with this custom rootfs.) I need to install Cuda 10.2 in order to use Cuda encode (which is not present in Cuda 10.0) ‘nvjpeg’ API coming with CUDA 10.0 that we have to use on Jetson boards has only decompression utility. If we can manage to use CUDA 10.2 on Jetson boards then we’ll be able to use compression utility of nvjpeg API.
I found that Cuda 10.2 is available according to this link https://developer.nvidia.com/cuda-toolkit/arm . I have followed the instructions and while installing there were no problems. But when I run the deviceQuery sample, it gives the following error
cuda-samples/NVIDIA_CUDA-10.2_Samples/1_Utilities/deviceQuery/deviceQuery Starting... CUDA Device Query (Runtime API) version (CUDART static linking) cudaGetDeviceCount returned 999 -> unknown error Result = FAIL
Previous version of Cuda 10.0 were giving the correct results. Installation procedure was almost the same but the only difference is I got Cuda 10.0 from the Jetpack’s json file on the other hand Cuda 10.2 is cuda-repo-ubuntu1804-10-2-local-10.2.107-435.17.01_1.0-1_arm64.deb file which is downloaded from the link I have mentioned in the first paragraph.
I know that Jetpack 4.3 supports Cuda 10.0 but isn’t Cuda on arm supposed to work too? What am I doing wrong?
I am sharing the output of Cuda 10.2 installation if you want to see it cuda_10-2_output.log (120.9 KB)
Thanks in advance.