Lighting optimizations in version 2.31
Entry posted by Josh ·
2,697 views
I got the volume intersection tests working, and it won't be hard to extend this to make trigger and fluid volumes. The image below shows a point light divided up into six volumes. Each pyramidal volume represents the camera frustum of one face of the shadow cubemap. In version 2.3 and earlier, whenever an object moves within the spherical volume of the point light, the shadow is considered invalidated, and all six faces of the shadowmap are redrawn. This is why point lights are so expensive to render compared to spot lights, which only require a single draw when the shadow is refreshed. The shadow updating was originally designed this way because it was thought that a geometry shader on shader model 4 hardware would allow rendering all six faces in one pass and provide a performance improvement. However, geometry shaders didn't turn out to be a good solution.
In version 2.31 the engine will only redraw those faces that have been invalidated. Volume tests are used to determine which volumes an object intersects. The oil drum in the image would probably only invalidate one volume of the shadow if it moves. Of course if the light itself moves, all volumes are invalidated, but you already know a moving point light can be expensive to render.
There's another technique I want to implement, which I am calling double-buffered shadow maps. This involves storing two copies of a shadowmap, one for static geometry and the other for dynamic geometry. In the example below, the static room and equipment would be rendered to the static shadow map, and the oildrum and any other moving objects would be rendered to the dynamic shadow map. Whenever the dynamic shadowmap is redrawn, the contents of the static shadow map are first copied to it. If any static object moves, both the static and then the dynamic shadow map would be updated. This would make it so that instead of re-rendering all the objects around the point light, the engine would only have to copy the static shadowmap contents, and then render the dynamic objects on top of that. This will yield a huge boost in performance in scenes with lots of point lights.
I've had these ideas for a long time, but the reason I am devoting time to them right now is because a game is coming out relatively soon that uses Leadwerks Engine, and I want it to be the best it can be. You know who you are. 
6 Comments
Recommended Comments