parallel_for_ flagNestedParallelFor cause multi task in difference thread may not parallel

[DaydreamCoding]

System Information

opencv 4.9

Detailed description

    static std::atomic<bool> flagNestedParallelFor(false);
    if (isNotNestedRegion)
            isNotNestedRegion = !flagNestedParallelFor.exchange(true); // multi threads in this line, may cause isNotNestedRegion to be false
    if (isNotNestedRegion)
    {
        try
        {
            parallel_for_impl(range, body, nstripes); // before done, flagNestedParallelFor always true, and another job could not fall into parallel_for_impl
            flagNestedParallelFor = false;
        }
        catch (...)
        {
            flagNestedParallelFor = false;
            throw;
        }
    }
    else // nested parallel_for_() calls are not parallelized
    {
        CV_UNUSED(nstripes);
        body(range);
    }

When multiple threads are using cv::resize, the cv::resize of some threads cannot be fall into parallel for

Steps to reproduce

auto t1 = std::thread([]() {
    cv::resize(); // this fall into parallel_for_impl
});

auto t2 = std::thread([]() {
    cv::resize(); // can not fall into parallel_for_impl
});

t1.join();
t2.join();

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
• I updated to the latest OpenCV version and the issue is still there
• There is reproducer code and related data files (videos, images, onnx, etc)

[asmorkalov]

@opencv-alalek could you take a look?

[fengyuentau]

It is not a bug. Parallel for is designed to work in this way. Nested parallel for does not lead to better performance. See #25260 (comment).

[DaydreamCoding]

It is not a bug. Parallel for is designed to work in this way. Nested parallel for does not lead to better performance. See #25260 (comment).

This design is flawed, resulting in external non-nested calls that probabilistically cannot be concurrent. And for multi-threaded calls to a function that uses parral_for, this scheme may be invalidated

auto t1 = std::thread([]() {
    cv::resize(); // this fall into parallel_for_impl
});

auto t2 = std::thread([]() {
    cv::resize(); // can not fall into parallel_for_impl
});

t1.join();
t2.join();

@opencv-alalek