We have been able to use the NVIDIAOpticalFlowSDK and generate good flow data but we are running into issues when attempting to run optical flow on frames that have already been allocated on the GPU. Are there any examples or details for using optical flow w/ existing on-device frames versus uploading new frames? We tried various ways such as mimicking the NV_OF_EXECUTE_INPUT_PARAMS, making modified OF buffers that use the pointers for the existing frames, etc. but we are not seeing valid output that’s even remotely close to what works when re-uploading and running through the SDK code.