How to achieve the official cloudson nano mirror with opencv with cuda acceleration

Autonomous Machines Jetson & Embedded Systems Jetson Nano
1

I now need to use jetson nano to deploy opencv to use cuda accelerated development environment, but after downloading and compiling the source code of opencv4.1.0, run the test program, cv::gpu::GetCudaEnabledDeviceCount() return value is always 0

2

Hi,

Do you compile the OpenCV with the correct GPU architecture.
Nano GPU capacity is sm=5.3 so please add the following option as the the cmake option:

cmake -D WITH_CUDA=ON -D CUDA_ARCH_BIN="5.3" -D CUDA_ARCH_PTX="" .....

Here is a sample script for v4.0.0:
https://github.com/AastaNV/JEP/blob/master/script/install_opencv4.0.0_Nano.sh
You can modify the script into v4.1.0 and recompile the OpenCV again.

Thanks.

3

How necessary are those second two? And at which versions of OpenCV? I’m asking because in my script I didn’t add them explicitly and those flags ended up getting set to the same things. I didn’t run tests for every version, however.

Since my script is supposed to support multiple versions, I suppose I better set those explicitly.

4

Hi,

-D WITH_CUDA=ON

Please enable this configuration or OpenCV won’t build with GPU support.

-D CUDA_ARCH_BIN="5.3"

5.3 is Nano GPU architecture. OpenCV only includes dGPU architecture by default.
You will need to add the Tegra capacity manually:
https://github.com/opencv/opencv/blob/ae279966c2c2e4d2db4a295b6397cc9aa832c794/cmake/OpenCVDetectCUDA.cmake#L120

If the package doesn’t build with correct GPU architecture, it’s expected to have some runtime failure.

Thanks.