This directory contains a test where several proof-of-prinicple
tests are done:

1. An actual routine (VectAdd) is compiled for and run on
   the GPU.

2. JANA and Hall-D code are linked in a used in the same
   executable that has a routine that is run on the GPU.

3. Multiple source files are used. One compiled with nvcc
   and the other with generic g++ and the resulting objects
	linked in a single executable.

4. A routine in the g++ compiled program calls one in the
   nvcc compiled program that dispatches it to a GPU. This
	also shows that the name-mangling is performed by nvcc
	in a way consistent with g++ since nothing is declared
	extern "C".
	

In addition, a potential convention is established where:

- CUDA kernels (i.e. routines that are to run on the GPU)
  are named with a "kernel_" prepended to the name.

- kernels are all contained in files ending in ".cu" and
  compiled with nvcc (necessarily).

- The ".cu" files also contain a wrapper routine with the
  same basename as the kernel only "gpu_" is prepended.

This convention would make it easy to see which routines
are actual kernel routines and which are wrappers. The
wrappers also make it easy to expose the GPU enabled
routines to the g++ compiled parts so they don't have to
use the nvcc compiler.