I'm working on light management in my forward rendering pipeline. I've looked all over for tutorials or examples, but everything is just a single directional light, or 8 or 16 point lights. The main problem is that I want to have a large number of lights in the scene of different types (point, spot, directional). For opaque objects I'm solving this with deferred shading, but for transparent objects I need a forward rendering pipeline so I'm writing it first. Basically I need to determine which 8 or 16 lights are going to affect each pixel. I tried sorting lights based on the influence on the center point of the object being rendered (mostly distance based) but the problem with that is that large objects like houses don't get proper shading. I cull point and spot lights that are outside of the view frustum and don't even bother considering them for the current frame, so for large scenes that cuts out potentially hundreds of lights all at once from consideration. The only thing now is that I need to be able to pick out which of these last few lights will shade a particular fragment. If I have to do it per-object on the CPU then that will have to work, but is there any fast way to pick out lights in the vertex shader and pass on indices to the fragment shader? That way I could send a list of like 128 lights and only accumulate like 8 or 16 for each pixel. If anyone has approaches they've used that were successful I'd love to hear, because I can't find anything on Google about light management.
Forward rendering light management
the way we solve it nowadays is called forward+ shading, or tiled lighting, or...
you can find a great collection of the main papers at http://aras-p.info/blog/2012/03/27/tiled-forward-shading-links/
But if you can master "tile based lighting" right away, I would go for that (for me personally that's a bridge to far for now).
I get pretty good results by sorting lights based on how much they shine on each mesh being rendered, culling lights that don't intersect the mesh (at some threshold intensity, ~0.005). I compute the light intensity at the center of the bounding box, then sort the lights by decreasing intensity. I then use the first N lights in the sorted list to send to the shader for rendering that mesh. Unless your entire world mesh is rendered in 1 draw call (unlikely), there is usually enough granularity of the meshes in the scene to get by with 4-8 lights, since there can be many lights visible, but only a few can shine on each object. The number of lights per mesh is determined by the shader's inputs.
Basically I need to determine which 8 or 16 lights are going to affect each pixel.
No, you determine which several lights are going to affect each object. There is no per-pixel information in forward rendering.
Each point light should have a bounding sphere.
Each directional light effects all objects (though with portals etc. you can make “indoor” areas not affected by “outdoor” directional lights).
Each spotlight has a bounding cone.
Each object in the scene has an AABB or bounding sphere or both.
When you gather objects in the frustum, that should include point lights and spotlights, except that they go into a second pool.
Upon rendering, you have a relatively small pool of influential light sources (perhaps 10 influencing light sources out of 500 in the whole map).
From there, further per-object (not per-renderable/per-submesh) checks against that smaller pool of lights (using optimized octree tests) becomes very fast and for each object a list of potential influential light-sources is very easy to build.
If that number goes over your maximum per-pass light count then as Hodgman mentioned you make another pass with the remaining lights and additive blending.
L. Spiro
So far I like the idea of tile based forward rendering. Although it sounds awfully complex to implement. The thing I don't understand from the papers I've read so far though is what you do once you have the list of lights for each tile. The tiles are screen space regions of pixels, so how do you know what tile you're in when rendering the object in the fragment shader after building the lists of lights for each tile?
Use VPOS / SV_POSITION / gl_FragCoord to tell which pixel is being processed, then divide by your grid size to tell which grid-cell (tile) is being processed, then fetch the light list for that tile.
Alternatively, instead of rending your lighting shader as a full-screen quad, you could render one quad per tile. - that's only valid for tiled-deferred.
Oh... That makes sense. Thanks for the advice everyone. Now I've got enough ideas to run well on powerful graphics cards, and some fall backs for weaker systems.
