System information (version)

• OpenCV => 4.0
• Operating System / Platform => Windows 64 Bit
• Compiler => Visual Studio 2015
• Cuda => 10.1

Detailed description

I will duplicate the problem described here:
https://stackoverflow.com/questions/59075944/why-do-i-get-corrupted-results-if-i-run-gpu-dense-optical-flow-algorithm-concurr

While using OpenCV CUDA dense optical flow in parallel I noticed that sometimes I get corrupted optical flow results, though I run it on different cuda::GpuMats and in separate cuda::Streams with separate Algorithm instances.

After some experiments with the code I found out that if I protect DenseOpticalFlow::calc() call with mutex or run only one thread, I always get correct (not corrupted) results.

I reduced my code to minimum reproducible example where I run optical flow algorithm on same input images multiple times in multiple threads

Furthermore, I checked that calling cuda::Stream::waitForCompletion() before calling DenseOpticalFlow::calc() asynchronously still causes the corruption.

However, if i use default stream cuda::Stream::Null() or use only DenseOpticalFlow::calc() synchronously, no corruption occurs.

Seems like using DenseOpticalFlow::calc() asynchronously causes data races.

Steps to reproduce

To reproduce the problem, you can use the example provided by the link above.