CUDA based Parallel Derivation of Parametric L-system

This paper proposes an approach to derive a parametric L-system in parallel based on Compute Unified Device Architecture (CUDA). It consists of a host program running on CPU and a device program running on CUDA enabled GPU. The host program is used to transfer data between CPU and GPU, pre-allocate host and device memory, and launch the device program. The device program takes charge of deriving an L-system by computing module strings with kernel functions which represent productions of the L-system and are executed by CUDA threads in parallel. Unlike most traditional L-system derivation algorithms based on string operations, our algorithms, which use newly defined data structures to represent an L-system, can be implemented without using any string operations that are not supported by current CUDA compute capability. The given experiments illustrate that this method has better computing performance than the method using serial derivation. The parameters which affect the performance, such as the number of module strings computed by each CUDA thread and the size of each thread block, are also investigated through these experiments.