vidia Nsight Visual Edition User Guide has a topic "Attach Debugging to a
Running Cuda process". It does not work for me. After I select Nvidia debugger
in "Debug/Attach to process" menu, my Cuda process is sometimes grayed out.
Therefore, I develop and debug Cuda code in exe-project and then convert it
into a boost::python extension.
To debug Cuda code in exe-project, install a breakpoint in Cuda code and go to
the menu "Nsight/Start Cuda debugging". Do not start a regular debugger and
then try to step into a kernel call.