
Have a look at the source. I *am* finding the 2D convex hull every frame, using an algorithm called QuickHull. (Plus the implementation I did was written quickly using the stl, and as such has room for some heavy optimisation)
Well, actually that's not entirely true, I *was* finding the convex hull every frame until I realised that for objects with smooth normals (i.e. not face normals) this produces some horrible lighting artefacts for faces that turn away from the light before the normals do. So the spheres now have a specialcase shadow volume construction that does a more accurate shadow volume without considering the actual points of the mesh.
Wrt finding convex hulls for more convex objects, I'd agree that in general this is a Bad Idea (TM), but this demo was written knowing that I'd only ever throw around convex solids. For more complex scenes (with a more general topology on objects) then something else should be used to find the silhouettes and construct the shadow volumes. As you no doubt already know, nvidia have some good documents on more "robust" shadowing techniques.
