J. W. Zhang
Y. L. Ma
P.O. Box 19, Northwestern Polytechnical Univ., Xian, 710072, China
An artificial neural network (ANN) can be implemented by many methods, such as electronic hardware or simulation on digital machines. Hardware implementation of an ANN cannot provide network size and computation flexibility because its topology is difficult to change once it is implemented in the hardware. Simulation of an ANN on commercial parallel computers is not cost effective because their architectures are not optimized for ANN computation, while software approaches need a large amount of execution time. The solution to these problems is to use the special parallel-processing architectures. Based on studying existing methods, this paper concentrates on the DSP-based virtual implementation of ANN. A parallel processing system composed of TMS320C30 has been designed and configured, which meets the needs of ANN application to acoustical signal processing in the real world. It is multiprocessor system with shared multiport memories. In this paper, the architecture of the system is described, and its performance is evaluated. The scalibility and communication method are also studied. The simulation results show that parallel efficiency of the system has reached a high level 80% when running a BP algorithm for classification of acoustical signals.