cv::cuda::NvidiaOpticalFlow_2_0::create 大约需要 30 秒

Question

To use the official example script for NvidiaOpticalFlow, I built OpenCV from source following the instructions from the Nvidia Optical Flow SDK (with slightly modified build flags to enable JPEG, OPENEXR, and Eigen). The OpenCV version is 4.5.2. I can post the CMake options on request.

This line from the example script takes about 30 seconds. Is this normal or could there be an issue with my setup?

Other details:

• I'm on Ubuntu 20.04
• nvidia-smi gives Driver Version: 470.141.03 CUDA Version: 11.4
• To compile the example script I use CMake with: cmake_minimum_required(VERSION 3.16) and set(CMAKE_CXX_STANDARD 20) .

Answer 1

According to user @RobertCrovella in the comments, the delay may have been to do with JIT compilation. Indeed, when calling create twice in one script, the second time does not have a delay.

Whether or not that's the real reason, the root cause was that I had incorrectly specified the arch for my GPU in the OpenCV build flags. They were set as -DCUDA_ARCH_BIN:STRING=7.5 -DCUDA_ARCH_PTX:STRING=7.5 whereas for my GPU (Nvidia RTX 3090) they should have been set as -DCUDA_ARCH_BIN:STRING=8.6 -DCUDA_ARCH_PTX:STRING=8.6 . I rebuilt OpenCV with the correct flags and the delay went away.