1. make an OpenGL pixel buffer
2. bind it to CUDA
3. run the raytracer kernel which fills the buffer
4. give it back to OpenGL
5. generate a texture from the buffer
6. set up orthogonal view
7. render a quad with the texture on it

The results:

and to show that the camera works (somewhat):

For those who want to combine CUDA and OpenGL is highly recommend watching this video.

The raytracer loads a mesh from a 3DS Max  file, puts a ‘room’ around it and renders the scene using only primary rays. Here is an example:

Raytracer output

The scene above consists of 1034 triangles and renders in 0.693 seconds on my Geforce 8600GT. My CPU needs around 10 seconds to render the same scene with the same detail so using the GPU is much faster.

Right now I’m in the process of implementing a BHV acceleration structure to make rendering faster. What it basically does is, it puts a bounding box around the mesh and then divides the mesh into two smaller bounding boxes. This process is repeated until some conditions are met (i.e. maximum depth reached) or when each bounding box contains a single traingle. This speeds up rendering because now the rays are checked against the bounding boxes instead of the triangles, and the bounding boxes that do not intersect with the ray don’t have to be checked.