high performance computing on graphics processing units: hgpu.org

The idea behind this project was to create a simulation of the evolution of life in CUDA. In this simulation the creatures are individual entities that can interact with the world. Each has its own set of state information and DNA representing it. Through this DNA the creatures evolve via division and mating. The evolution of the DNA during reproduction utilizes crossover and mutation but does not have any sort of fitness function consideration. Each creature itself chooses when and how to reproduce, thus the creatures that live longer and/or reproduce faster will tend to survive better or be more fit. The creatures control their actions and interactions with the world via a recurrent neural network. The structure and weights of this network are encoded into the DNA as well as everything else about the creature. The structure of this report is as follows. It starts with an overview of CUDA and the different facets of it that need to be considered when programing in CUDA. This is followed by a detailed description of the evolutionary model used in the simulation. It continues into the general structure of the data used to represent the simulation. The implementation of the simulation follows and covers both the high and low levels of the code. This section also discusses some of the design considerations with respect to the specific pieces of code being described. Lastly, the results of running the simulation will be discussed and final conclusions will be made.