They are mentioned in many articles and i have no idea what they are...
Example:
"The textures applied on large areas show repeating patterns, but it gets remove by the modulation. The trick is indeed in the additional processing, in computing parallel fractal channels and using them for color modulation, normal perturbation etc."
Think about a typical image, which has 3-4 channels of data, comprising the R, G, B and optionally Alpha channels, which are combined together produce the final image. When you are working with procedural generation, you typically use many layers (or "channels") of noise, and combine those to produce the final result.
You'll probably have one noise generator that produces a height/normal-map, another generator that you use to perturb the normal map to breakup any visible regularity, a generator that you use to breakup the regularity of texture transitions, and so forth...
So you precompute the noise function and store it in a texture?
No. Though you can, if your use case makes it advantageous to do so.
Think of "channels" as meaning "layers" of noise in this context.
Oh ok. And how are the layers represented in the shader exactly? Because he mentioned that he was "running out of channels". That's what made me think of textures..
Because he mentioned that he was "running out of channels".
Reading between the lines, "running out" tends to mean that one is over one's CPU/GPU processing budget.
You could also try pinging cameni (who wrote Outerra) and see if he'll respond himself.
So you precompute the noise function and store it in a texture?
No. Though you can, if your use case makes it advantageous to do so.
Think of "channels" as meaning "layers" of noise in this context.
This.
Basically, we are generating 1+4 independent layers of fractal noise, the first one used for terrain height modulation, and four less precise ones (each with a specific spectrum) for various tasks within the generator. The generated noise is cached in textures for perf reasons, though you don't have to, as swiftcoder wrote.
Running out of the channels here means that for some (new) stuff we don't have enough of these layers, and would need more, with different characteristics maybe. But that means there would be more GPU memory needed for the cached data, and that is not very good since Outerra consumes quite a lot of it already.
There are some tricks that I hope to use, like combining several noise channels of different spectra into one, and then separating them with a filter. That would reduce the precision though, so I need to experiment.
