Question

I want to build a cuda plugin for an other project written in C++ using MPI.
I got the following situation:

• startingpoint.cpp is a c++ file including the cudaintegrator.h and
  includes datastructure.h
• cudaintegrator.h is the header file for a
  cuda file including datastructure.h
• cudaintegrator.cpp is the
  implemenation of cudaintegrator.h
• datastructure.h is a header file
  specifing some datastructures

the files are organized like that:

/trunk/
   /src
      /folder1
         /folder2
            /startingpoint.cpp
         /folder3 
            /cudaintegrator.h
            /cudaintegrator.cpp
      /folder4
         /folder5
            /datastructure.h

When I run nvcc from my trunk dir with the following command:

/usr/local/cuda/bin/nvcc   -c -o src/folder1/folder3/cudaintegrator.o src/folder1/folder3/cudaintegrator.cu

the included file `#include “folder4/folder5/datastructure.h” is not found:

src/folder1/folder2/cudaintegrator.h:12:33: fatal error: folder4/folder5/datastructure.h: No such file or directory

When I run mpic++ for the compilation of startingpoint.cpp, which uses the same include, the datastructure.h is found.

I suspect that the working directory of nvcc and mpic++ are somehow different(?) even both are run from the trunk directory.
When I change the include to ../../folder4/folder5/datastructure.h the file itself is found, but subsequent includes fail for the same reason. I can not change the includes in all subsequent files because they can’t be found by mpic++ in this case.
Does anyone have a suggestion how I should include the files in this situation correctly, or how to instruct nvcc to find the included files?

Answer 1

Providing include paths (-I) to the nvcc invocation should be enough (if not, it uses relative include paths only, that’s why adding ../.. works for a single include).