ABSTRACT
Modern real-time rendering engines need to support many dynamic light sources in a scene. Meeting this requirement with traditional forward rendering is problematic. Typically, a forward-rendered engine culls lights on the CPU for each batch of scene geometry to be drawn, and changing the set of lights in use requires a separate draw call. Thus, there is an undesirable tradeoff between using smaller pieces of the scene for more efficient light culling versus using larger batches and more instancing for fewer total draw calls. The intersection tests required for light culling can also be a performance burden for the CPU.
