best would be probably to write out those patch colors into a texture and let the texture units do all the interpolation.

vertex colors have the disadvantages that you'll draw a patch using two triangles and the interpolation won't look properly, you will have similar issues like with gouraud shading. it depends on what you do and what limitations you have, on mobile platforms that might be quite ok, on consoles/pc you might expect more.




create a texture in memory first (basically a 2d array of colors)

you need to create a UV set for that texture (in a simple case you can just planar project your surface).

based on the UVs draw the color of the patch, for each patch, into the texture.

not urgently needed, but it can increase the quality if you have a dilation step that fills the yet unfilled pixels by the neighbour colors. (you can repeat this step several times).

some people additionally blur the texture, that can hide artifacts if you don't use a water tight patch (-> if partially occluded patches exist).

also not urgently needed, but it can improve the quality if you create the lightmap in a higher resolution and as last step scale it down to the desired resolution.

as last step, create a device texture and copy your local texture.




some old school tutorial that might help.

http://www.flipcode.com/archives/Light_Mapping_Theory_and_Implementation.shtml




you make me want to work on my radiosity calculator again http://twitpic.com/4khy8e

On a basic level, there's really not a ton of difference between storing your lighting in your vertices or in a texture. In both cases you have a sparse set of sample points covering a large set of surfaces, and you interpolate between those sample points to help alleviate the aliasing artifacts. However you need a high sample point resolution if you want to capture high-frequency lighting changes, which typically occur in shadows from direct lighting. The advantages of using textures for this are...

1. Your sample points are regularly spaced on a grid, while with vertices located based on the topology of the mesh
2. You can increase the sample resolution for certain surfaces without having to add more vertices, which means adding more position/normal/tangent/whatever data as well as increasing the number of triangles you need to render (it also means you use smaller triangles, which are less efficient)
3. You can compress textures
4. You can mipmap textures
5. Typically you can stream textures more easily than vertex data, or you can also use them in conjunction with virtual textures

Of course it's typically easier to get up and running with per-vertex data since you don't need to parametrize all of the meshes in your scene, or deal with lightmap atlasing/packing.

Also, the resolution needed for your lighting samples depends on what lighting you're calculating. For shadows or small-scale AO you'll need quite a bit of resolution for it to look good...even with lightmaps you typically end up with pretty blurry shadows that don't match your real-time shadows. However if you don't store direct lighting and only store indirect lighting or indirect + sky lighting, you can get pretty good results once you composite in some real-time lighting with shadows. Naughty Dog did this for a lot of their outdoor scenes in Uncharted 2, with lighting stored per-vertex. I also did this with my radiosity sample. If you look at the pictures you can see the indirect + sky lighting looks pretty messy on its own, but combined with the sun lighting + shadows it ends up looking pretty nice. Using a basis like H-basis also helps a lot, since you still get some high-frequency variation from the normal maps. One of the Uncharted 2 presentations also has some pictures like that and they have similar results.

EDIT: I forgot to mention this recent publication that attempts to address issues with interpolating vertex-baked samples: http://www.ppsloan.org/publications/vb-egsr11.pdf