Abstract
Nowadays the basic graphic cards contain more than 300 compute cores, which is expected to significantly increase the performances of parallel applications. In the past few years, several works were made concerning the simulation of Multi-agent Systems on the Graphics Processing Unit (GPU). But they often use agents which have a position in the environment associated with an interaction radius. In this study we present a generic implementation on the GPU of reactive agents, which are formed of a set of positions. The communications between the agents are only indirect, and are possible by reading the trail left by other agents. In our system these trails diffuse in the environment. The problem posed is to compute the diffusion in an efficient way, otherwise the whole environment should be iterated. Our implementation takes advantage of the GPU by executing each agent and diffusion on one compute unit. Firstly, our implementation keeps the same dynamics in comparison to previous studies, where the agents were not executed in parallel. Secondly, the simulation time has a 45 × speed advantage over a 4 core CPU.