I'd like some advice on how best to go about this. I'm making a 2d isometric RPG and I've implemented a simple 2d lighting solution using GLSL that allows you to put simple spherical spot lights in your scene. The lights can move and this is important for gameplay reasons, since ideally I would like the player to be able to equip torches or other light sources and carry them around.

This is a party based RPG, which is a problem because in principle if one party member is carrying a torch then I can't really rule out them all carrying one. This makes it hard to predict the maximum number of spot lights that will be relevant to each tile.

Each spot light is specified in some uniform arrays for the position, colour and attenuation data. At first I thought I could just make the number of lights itself a uniform variable within the shader. I think this doesn't work however, since at least on older hardware if you have a for loop within the fragment shader it must be over a constant.

So I have two different approaches I'm thinking about:

1. Determine the total number of lights relevant to the current scene at the beginning of the frame, have different shaders defined for different possible numbers (within some reasonable range) and activate the correct one. Set the lighting data once and use this for the whole scene.

2. Pick a fixed smaller number of lights for the shader and then try to determine for each tile which lights have the most influence, set the lighting data uniforms for each tile (or whenever it changes from tile to tile).

So the second option will in some cases consider a smaller number of lights and will have less calculations per pixel, but you will need to set the uniforms more often. The first one will need different shader versions which means additional management, (perhaps there is some macro magic that can help avoid maintaining 5 versions of the same file?) I don't have much intuition for the likely trade-offs in performance. The calculations for each light are the usual distance and dot-product and so on.

Hi mr. Prune,

Your description made me very curious, - would you mind showing your lighting off a bit with demo screenshots?

As a possible solution to your problem; encode all your light data as a texture, and then upload that. You can change fragment colours by rendering to this texture in a pre pass for each dynamic light, or setting the pixel value from your cpu program.

You get 4 bytes precision which should be enough for r, g, b to match the x, y, z values in your world, alpha for radius, and then another pixel below that can hold the actual color information.

Thus, you can hold 256^2 / 2, 32768 lights in a 256 by 256 px texture.

Obviously, you should slice up the update events
into multiple frames if you intend all of them to be dynamic, but I hope you won't do that.

The easiest way would be to do multiple passes. You render the scene once for each light and have it add the colors together each pass. This results in some overhead from rendering everything multiple times, but this can be reduced with a z-prepass or by using deferred rendering.

Yes, I suppose that could be quite slow for large numbers of lights, but if it only happens in extreme cases then it could work. This is probably a dumb question, but can you do the multiple passes using a simple additive blend function? Or something more complicated.

The only problem with this is that for game-play reasons I need to be able to use translucency. (This is for showing partially transparent character sprites when they are hidden behind the environment. Think infinity engine style.) I'm not sure if that can work, I need to think about it.

Your description made me very curious, - would you mind showing your lighting off a bit with demo screenshots?

Yes I can do, it's really nothing special and I have only stock art to test it with. I've generalised the spotlight formula a bit to use 3 attenuation factors in the normal way, and a range falloff on top of that so that the range can be know precisely. The two circles are debug drawings of the light position in my game editor.

[sharedmedia=gallery:images:2886]

Putting the data in a texture is interesting, I hadn't considered that. As I said I don't have much intuition for the performance costs of various things in opengl, so I don't know how this would compare with using a large number of uniforms. For example, how important is it to minimise the setting of uniforms in shaders? And how does it vary by the amount of data? (i.e is setting an array of 10 vec4 much worse than simply setting 3 ints or similar.)