Fast block distributed CUDA implementation of the Hungarian algorithm

Abstract

The Hungarian algorithm solves the linear assignment problem in polynomial time. A GPU/CUDA implementation of this algorithm is proposed. GPUs are massive parallel machines. In this implementation, the alternating path search phase of the algorithm is distributed by several blocks in a way to minimize global device synchronization. This phase is very important and has a big contribution to the execution time. Other advanced features also implemented are: parallel graph traversal; the parallel detection of multiple alternating paths in a single iteration; a simplified and fast matrix compression that stores the zeros of the slack matrix, resulting in very fast graph traversal; highly optimized reductions for the initial slack matrix calculation and update. This results in a fast implementation for moderate size problems.