Hello Worley

If you are into procedural things you probably heard about Perlin noise. This noise function can be used to recreate many of the patterns you find in nature, like clouds, large mountain ranges, marble. While it is not the only way to produce them, it is probably the most efficient way to achieve very good looking results.

There is another type of noise that is as useful as Perlin noise: Worley noise, which is also called cellular noise. Here you can see how it looks like, compared to Perlin:

Perlin is great for phenomena that originated from mixing different entities, like how moisture permeates a wall or mold grows on a tree. Worley is best for self-organizing patterns, things that break down into clear boundaries like cells, pebbles and even man-made objects like a stone wall.

The idea behind this noise function is simple. The function field is scattered with random points, then for any point in space the function determines how this point relates to the nearby random points.

Unlike Perlin's function, which you may take as a magical black box, it helps to learn the implementation details of the Worley function. Many interesting effects can be achieved by using different criteria. One example would be to return the distance to the closest random point. This is the used in the image before. As you can see it outputs the Voronoi diagram for the field.

I had delayed including Worley noise in my terrain generator for too long. All the screenshots and videos I had posted so far were produced using only some form of Perlin-like noise. It was a long weekend, so while the wife and twins were taking a nap I pulled out the holy book for procedural content and went straight to Worley's chapter.

It only took a few hours to have his noise ported to OpenCL. Here are screenshots of a large rock formation out of Worley noise:

Here it can be seen in more detail:

Worley's algorithm is perfectly suited for parallelization. I only changed the criteria on which neighbor cells are visited. For a CPU implementation, Worley goes into an elaborate scheme to avoid testing points inside a cell if it can be early discarded. My parallel superstitious instinct told me this type of optimization would not carry well into OpenCL. If at least one of the points in the current wavefront hits the worst case scenario, then it was as if all the points were worst case scenario too. So I removed all conditions and went with the brute force approach. I do not have any numbers to back this up, however. I could be wrong.

Now I have this new toy and I keep thinking of shiny things I could build. It is that warm feeling a new noise function will give you.