Everybody is good, now I will yolov3 - darknet encapsulated into a DLL, calls for the upper machine, target detection for a pair of big size pictures, size is 8000 x8000. Must be split into multiple small area into darknet again for testing, but found that single thread calls to each area scan testing is no problem, but the use of multi-threaded synchronous calls multiple darknet - DLL scans, appear problem, in CHECK_CUDA (cudaSetDeviceFlags (cudaDeviceScheduleBlockingSync)); This code error, should be cuda synchronization problems, someone had a similar problem? Excuse me, how to solve, thank you, or have other solution? - (multi-process tried, can, but is not suitable for application, and complicated)