There it is indicated the ability to insert your own dynamic library, formed in a special way according to api. The sources api says that I can write my own function for processing in cuda.
The first question is, can I use the nvidia npp library in ‘customer-lib-name = “libnvsample_cudaprocess.so”’, would that be good practice?
Second question - How at a low level can the preprocessing results obtained with “libnvsample_cudaprocess.so” be transferred to the buffer (input) of TensorRT?
You can use Cuda in the custom library. I.e. with the use of the following wrappers function https://github.com/dusty-nv/jetson-util
You can overlay text (see cuda/cudaFont) on images by Cuda
NPP is integrated into the CUDA library.
So you can use it as a general CUDA function.
You might also be interested in our Deepstream SDK.
The library uses the GStreamer interface and includes the preprocessing and TensorRT inference implementation.