In CUDA, Nsight doesn't show cuInit(0) call in timeline

Question

I have following code in my program that works with pure driver API,

cuInit(0)

and Nvidia Visual Profiler 8.0 is not showing it on timeline graph.

Can I call it once in a while, or as frequently as pc wake-ups or should it be called only maximum once per process(which would need a singleton-like or some globally synchronized initializer wrapper for the application)?

Should I expect a trivial time loss per call always?

Does cuda driver API has an internal counter for it so that if somehow driver API unloads, it automatically re-loads it so I don't have to run it again later?

What if this C++ is a DLL and will be called from C#, Java, Phyton? Can other third-party libraries initialize it before my application? I know GPU does context switching between processes(is driver API independent per process?) but does this initialization command hinder others' work by accident(for example, in a cloud computer where N other users run CUDA too)?

Answer 1

cuInit() should be called once per application, before any other cuda driver API calls are used .

I think the fact that it does not show up in the timeline is expected behavior.

There will be some time cost associated with cuInit() . I wouldn't call it "trivial". It will vary based on a variety of system configuration parameters.

There is no concept of the driver API "unloading".

If you use any driver API calls in a library, this call must precede them as well, in library code.

I think you might be getting this confused with a CUDA context . cuInit() does not create a context (ignoring the primary context). If you have a context created, it is possible to pass that context to other routines, even if they are in a dynamically linked library.

This call has no bearing on CUDA running in another process.