A New Data Reduction Approach over the Stream Processor Architecture

In order to solve the parallel processing problem of mass data, stream processor develops quickly in recent years. The essence of stream processor is getting high performance according to executing thousands of threads parallelly. It is widely used in scientific computing, data mining, video quality analysis and other fields. But mass threads means the high costs of thread synchronization. Mass computing time is wasted in the reduction of thread data. This paper provides a binary reduction algorithm based on three basic reduction algorithms according to CUDA parallel computing model. Experiments show that the performance of binary reduction algorithm is better than three basic reduction algorithms and it has great scalability.