Craig Macomber 1232953626|%e %b %Y, %H:%M %Z|agohover
I have been thinking about planet graphics. I had a plan, and it was pretty good. I had made the decision to try and do planets in our current engine, which basically means no normal/bump maps and no shaders. Because of this I decided that we would have to do texturing on the CPU. I designed a system of hierarchical layers that would allow detail across all scales (from continents to grass). This system is powerful, and not fractal to implement in a shader on the GPU which we couldn't use anyway. It would be a matter of zoom and wait for texture detail.
I still like that system, but today I realized that if we were to use a graphics engine where we have much more control over the GPU (like OpenGL), I could implement the same system, but much faster, and better.
First of all, the reason it can not be done in a shader is that when you zoom in, there are so many possible layers. In reality, very few will ever be used at once, but the pixel shader can not take advantage of the hierarchical nature to eliminate almost all of them. Also, even the amount of layers is use will be so large, and the amount of computation per sample in each layer so larger that it simply would never be close to fast enough.
Solution: Suppose we were to have an already textured planet. Once you zoom in, you get to the point where there is no detail left from the texture. At this zoom, suppose you implemented the appropriate sub layers (trees in the forests, dunes in the deserts) in shaders. They would add in the detail!
Now, you zoom in further, and its time for the shader's sub layers. We use the GPU to convert the current shaders into textures which we draw on top of the old lower detail ones, and generate new shaders for the sub layers (branches on the trees or what ever).
This allows per pixel detail at all zoom levels, and keeps the speed up. There is simply 1 shader drawing a few layers over 1 texture. That should be fast! Getting the GPU to generate the textures that will be combined will be somewhat hard, but apparently it's possible. Also, it would allow high quality effects like per layer bump mapping with proper lighting!
The slightly easier version is so skip the detail shaders and just use generated textures (generate them to higher detail). This would probably be done first, and detail shaders would be added if needed.
Also, the mesh height data could be generated on the GPU as height map textures which would speed stuff up a lot in that department. (Yep, that means fast detailed mesh with fast detailed textures with good lighting!)
What this means to you: If we go forward with this, it will mean some delays in the short run, but planets will look much better, and so will everything else (because of a better graphics engine)
I should note that I have never done anything like this, so it's all very speculative.
Basic design:
