Thank you for contribution!

It would be nice if you could add simple test here: https://github.com/opencv/opencv/blob/4.5.2/modules/python/test/test_cuda.py

Hello @alalek

what kind of test do you need ?

1. that the patch does not break mat_device.upload(mat_host, stream) ?
2. that the patch correct the issue ? to test that the issue is corrected, we need to use nvprof from the CUDA Toolkit for launching the test.

I have a second question. My fix looks like a hack. Here is the problem creating the bug.
The generated Python code for GpuMat::upload is declaring Stream stream variables to parse the stream parameter. But the default ctor creates a new stream by calling cudaStreamCreate. The fix is to have Stream stream=Stream::Null() in the Python bindings code.

Other functions are using a default value in the C++ function prototype:
CV_EXPORTS_W void resize(InputArray src, OutputArray dst, Size dsize, double fx=0, double fy=0, int interpolation = INTER_LINEAR, Stream& stream = Stream::Null());

But this is not possible for GpuMat::upload because there is a circular dependency between GpuMat and Stream classes.

So patching gen2.py is the only solution I found to generate the correct stream parameter parsing.

jenkins cn please retry a build

Please remove "WIP: " when patch is ready (BTW, you can use GitHub's "drafts" instead)

Hello @alalek

I added a test_cuda_upload_download_stream to test_cuda.py as the current upload/download test were not using stream.

Running nvprof python3.6 test_cuda.py allows to see if cudaStreamCreate is called only 2 times.