Ian Parberry, "Amortized Noise", Journal of Computer Graphics Techniques, Vol. 3, No. 2, pp. 31-47, 2014. [Manuscript and BibTex at JCGT]

Abstract

Perlin noise is often used to compute a regularly spaced grid of noise values. The amortized noise algorithm takes advantage of this regular call pattern to amortize the computation cost of floating-point computations over interpolated points using dynamic programming techniques. The 2D amortized noise algorithm uses a factor of 17/3 ≈ 5.67 fewer floating-point multiplications than the 2D Perlin noise algorithm, resulting in a speedup by a factor of approximately 3.6-4.8 in practice on available desktop and laptop computing hardware. The 3D amortized noise algorithm uses a factor of 40/7 ≈ 5.71 fewer floating-point multiplications than the 3D Perlin noise algorithm; however, the increasing overhead for the initialization of tables limits the speedup factor achieved in practice to around 2.25. Improvements to both 2D Perlin noise and 2D amortized noise include making them infinite and non-repeating by replacing the permutation table with a perfect hash function, and making them smoother by using quintic splines instead of cubic splines. While these improvements slow down 2D Perlin noise down by a factor of approximately 32-92, they slow 2D amortized noise by a negligible amount.

 

These animated gifs were created by taking 128 slices across a 256x256x256 block of 3D noise with 5 octaves generated by the amortized 3D noise algorithm. The result is similar to 3D Perlin noise, but it does not repeat at regular intervals, and it is smoother since it uses a quintic spline instead of a cubic spline. The image on the left shows -1 as black, +1 as white, and values in between as shades of gray. The image on the right shows negative values as purple and non-negative values as green.

The source code for amortized noise is available for free download from https://github.com/Ian-Parberry/AmortizedNoise. It has been released under the GNU All-Permissive License and contains the following notice. If these conditions are unacceptable to you, then do not download them.

The source code is written in C++ and includes project files for Windows Visual Studio 2012 and a Unix makefile using g++.

Documentation

• 2D Evaluator (to reproduce performance evaluation in paper)
• 2D Generator (to generate and view 2D amortized noise)
• 3D Generator (to generate and view 3D amortized noise)

An earlier version of this paper was published as Technical Report LARC-2013-03, Laboratory for Recreational Computing, Dept. of Computer Science & Engineering, University of North Texas, September 2013.