I am a backend C/C++ CUDA engineer. Often, the main program is written in Python and we use C / C++ extension to call parts of the program written in C/C++ in the Python function. However, whenever I have to debug such program, I cannot use “cuda-gdb” nor “gdb python”. Does anybody have a debugger that can stepthrough each line and check memory / register values for a program that is partially written in python, C / C++ CUDA?
hmmm… are you sure it is applicable to CUDA? It works fine when I am on the CPU, but as soon as I move over to a GPU kernel, it stops working ( hangs ).
Ho right, you can only step in the cpu code. But it can run cuda code fine even though you cannot break into it.
I did not knew you could put breakpoints directly into the kernels. But the fact that you use python should not be a problem. nvvp works quite fine to inspect python code.