This project includes a serial and a parallel implementation of the Worley Noise on GPU. The parallel version uses CUDA.

Worley noise, sometimes referred to as cell noise, is used to add details to the surface of rendered objects [1], a process known as procedural texturing. It can be used to create textures resembling stone, water, ice, intestines, crumpled paper, mountain ranges, cell noise, crocodile skin and clouds.

The basic principle of the Worley noise algorithm initially involves taking random fixed points in a space. This project only focuses on 2-dimensional space, although 3-dimensional space is also possible. Then, for every point, , in the plane, the distance to the th closest fixed point is calculated, . The distance is then normalized to create a shade of grey. The figures below show how 4 random points are transformed into Worley noise. The tiling (red boxed) are used as an optimizations as suggested by Rosén [2].

Random points	Noise from random points

1 Pixel/Tile	5 Pixels/Tile

Intensity = 1	Intensity = 1.5

Normal	Reverse

The serial implementation was tested on an Intel Core i7-7700HQ. The parallel implementation used these two GPUs: NVIDIA GeForce GTX1050 and NVIDIA TITAN Black.

The table below shows the average timings (outputted using the --performance flag) for a 7000x7000 image using tiles of size 512x512 pixels.

The speedups (compared to the serial implementation) for the same configurations, are shown below.

This project is built and packaged using Nsight Eclipse version 9.2 & Cuda version 10.2.

There are three build configurations: Release, Debug and Tests.

To run this project outside an Nsight environment, run make inside the /Demo folder to get an executable file.

Worley Noise GPU is distributed under the GPL-3.0 licence.

JButil for PGM image processing.