A Novel Heterogeneous Framework for Local Dependency Dynamic Programming Problems

Dynamic programming is a well known technique to solve combinatorial search and optimization problems. LDDP (Local Dependency DP) problems are those DP problems where each update to an entry in the DP table is determined by the contents of its neighboring cells. The amount of parallelism available for a given LDDP problem may vary depending on the position and number of neigh boring cells involved in the computation. In this paper, we develop a parallel heterogeneous (CPU+GPU) framework for solving LDDP problems. We use GPU and CPU specific optimizations and investigate if the given problem is a good candidate for heterogeneous computation. We also devise techniques to distribute the workload between CPU and GPU. We show how our framework can be used by considering several case study problems.