BVH would give you fast membership classification, something you already do, I think. BVH aims at decomposing your object hierarchically for fast inclusion tests. In contrast, the dexel/voxel models are space decompositions aimed at capturing the region occupied by a shape. For representation of dexels, some extra research might be necessary. It will be interesting to know how ModuleWorks represent them, although I'm not sure they delve into such details in their articles.I am guessing a bvh tree would work..
if(time <= 0) return REAL_MAX;
if (time < -2 * Precision::Confusion()) return REAL_MAX;
return ( U < 0 || V < 0 || V+U > 1.0) ? REAL_MAX : time;
return (U < -2 * Precision::Confusion() || V < -2 * Precision::Confusion() || U + V > 1.0) ? REAL_MAX : time;