HyLAC: Hybrid linear assignment solver in CUDA

The Linear Assignment Problem (LAP) is a fundamental combinatorial optimization problem with a wide range of applications. Over the years, significant progress has been made in developing efficient algorithms to solve the LAP, particularly in the realm of high-performance computing, leading to remarkable reductions in computation time. In recent years, hardware improvements in General Purpose Graphics Processing Units (GPGPUs) have shown promise in meeting the ever-increasing compute bandwidth requirements. This has attracted researchers to develop GPU-accelerated algorithms to solve the LAP.

Recent work in the GPU domain has uncovered parallelism available in the problem structure to achieve significant performance improvements. However, each solution presented so far targets either sparse or dense instances of the problem and has some scope for improvement. The Hungarian algorithm is one of the most famous approaches to solving the LAP in polynomial time. Hungarian algorithm has classical $O\left(N^4\right)$ (Munkres') and tree based $O\left(N^3\right)$ (Lawler's) implementations. It is well established that the Munkres' implementation is faster for sparse LAP instances while the Lawler's implementation is faster for dense instances. In this work, we blend the GPU implementations of Munkres' and Lawler's to develop a Hybrid GPU-accelerated solver for LAP that switches between the two implementations based on available sparsity. Also, we improve the existing GPU implementations to reduce memory contention, minimize CPU-GPU synchronizations, and coalesced memory access. The resulting solver (HyLAC) works faster than existing CPU/GPU LAP solvers for sparse as well as dense problem instances. HyLAC achieves a speedup of up to 6.14× over existing state-of-the-art GPU implementation when run on the same hardware. We also develop an implementation to solve a list of small LAPs (tiled LAP), which is particularly useful in the optimization domain. This tiled LAP solver performs 22.59× faster than the existing implementation.