Compiling CUDA projects

I find that a lot of new CUDA developers have this tendency to use the CUDA sdk makefiles (include common.mk), and having loads of dependencies on the cuda utils that the sdk provides. This, in the long run does not seem like a good idea, as it involves depending on what the sdk makefile does, and hence depending on the (heavy) sdk itself. Here is a small tutorial on how to compile stand alone CUDA program with multiple h, cpp and cu files, and a few external headers/libs:

Lets say we have 4 files: main.cpp, test.cpp, test.h, kernel.cu

The idea is that we have to compile the cpp files with g++, and the cu files with nvcc.
Also, we specify that only compilation should take place (no linking), so that if we have any dependencies between files, we won't get zillions of errors.

To do this, compile the cpp files this way:

g++ -c *.cpp

Then compile the cu files:

nvcc -c *.cu

Note that nvcc may not be in your default path. When you install the CUDA toolkit, it is placed in the cuda/bin/nvcc path. For example if you installed cuda toolkit in /usr/local/cuda, then nvcc is in /usr/local/cuda/bin/nvcc. Adding this to your $PATH variable will fix things.

The above steps should produce 3 object files: main.o, test.o, and kernel.o

Finally, link these together and make the final executable:

g++ -o runme *.o

This will link all the object files into a single executable called "runme", which you can run as usual.

To include any specific headers / libraries, simply pass them as arguments during the compile / link phases. For example:

g++ -c -I/usr/local/cuda/include *.cpp
g++ -o runme -L/usr/local/cuda/lib -lcuda -lcudart *.o

While running CUDA programs, you may get a library not found error. To fix this, add the CUDA libraries to your $LD_LIBRARY_PATH environment variable. The libraries are libcuda.so, libcudart.so and will be present in your CUDA toolkit install path.

All this can be condensed into a nice makefile for convenience.

Finally, the part about actually writing CUDA programs that are independent of the SDK. Note that the SDK uses macros like CUT_DEVICE_INIT() which are in practice not needed at all. Simply include the cuda header files, and start making cuda calls.