Optimizing Two-Dimensional Continuous Dynamic Programming for Cell Broadband Engine Processors

Two-dimensional continuous dynamic programming (2DCDP), a specialized DP matching method for image recognition, can be applied to many applications such as object tracking, pattern matching, etc. However, the execution time is large, and the current general purpose processor does not achieve performance in real-time. In this paper, we present our approach to real-time image recognition using a cell broadband engine processor (Cell processor). We optimize 2DCDP for the cell processor by vectorizing with SIMD instructions, parallelizing with multiple SPEs, dynamic branch prediction in assembly level, and so on. Finally, the performance on the Cell processor is achieved over 15 times faster than the performance on an Intel Xeon 5160 processor.