UMat memory leak on Intel GPUs when using multiple threads

[leugenea]

System information (version)

• OpenCV => 3.3.1
• Operating System / Platform => Windows 10 64 Bit
• Compiler => Visual Studio 2013

Detailed description

Using UMat with OpenCL on Intel videocards in multiple host threads causes GPU memory leak.

Steps to reproduce

1. Create function that: loads image into cv::Mat, copies it to cv::UMat using copyTo method, runs some OpenCL-related function like CascadeClassifier's detectMultiScale or cv::equalizeHist, and returns.
2. Create loop: create thread, run function from 1 in this thread, join thread.
3. Watch GPU memory increasing indefinitely.

Additional notes

1. If not using Intel GPU (tested on various NVidia GPUs), there's no memory leak.
2. If using OpenCL from fixed number of threads (for example, in fixed-size threadpool), there's no memory leak.
3. Even if I release UMat's memory using release() method, memory will leak.

[alalek]

Consider to use ThreadPools instead of spawning new threads.

[leugenea]

Yes, of course. But isn't it a bit strange that OpenCL memory isn't deallocated until program terminates?