Marching cube taken from https://github.com/smistad/GPU-Marching-Cubes and ported to cuda.

For building you need to have Cuda, OpenGL, GLEW and GLUT installed.
To compile (assuming you are in the top directory):
mkdir build
cd build
cmake ..
make

After compiling you can run the program:
./mc path/to/voxel/data.raw

Please note, that RAW files need a second file, which contains the dimensions of the RAW file. If voxeldata.raw is a RAW file, the other file would be called voxeldata.raw.dim . An example file would look like this:
x = 64
y = 32
z = 90

For other example RAW files see: https://bitbucket.org/cg2/opencl/src/c2de329bf761/src/voxeldata

In the original opencl code the base layer of the histogrampyramid consisted of
uchar4 voxels, which saved in the z coordinate the original voxel value of the
voxeldata and did not use the w coordinate. the author said, that it was a cache
optimization, but for me there is no difference in speed, when ushing uchar2 and
reading the voxelvalue from another array, with which the base layer of the
histogrampyramid was constructed. with uchar4 the voxelvolume would have been
saved twice on the GPU. maybe I have some other bottleneck in my implementation,
so these cache optimization doesn't affect me now.