Two ways:
As of version 0.93, PyCUDA supports threading. There is an example of how this can be done in examples/multiple_threads.py in the PyCUDA distribution. When you use threading in PyCUDA, you should be aware of one peculiarity, though. Contexts in CUDA are a per-thread affair, and as such all contexts associated with a thread as well as GPU memory, arrays and other resources in that context will be automatically freed when the thread exits. PyCUDA will notice this and will not try to free the corresponding resource–it’s already gone after all.
There is another, less intended consequence, though: If Python’s garbage collector finds a PyCUDA object it wishes to dispose of, and PyCUDA, upon trying to free it, determines that the object was allocated outside of the current thread of execution, then that object is quietly leaked. This properly handles the above situation, but it mishandles a situation where:
To entirely avoid the problem, do one of the following:
In the Visual Profiler, create a new session with File set to your Python executable and Arguments set to your main Python script.
There is one more important change you have to make before the timeline will show any events. Before your application quits, make the call
pycuda.driver.stop_profiler() so that the profiling data buffers get flushed to file.