Heuristic Optimization Methods for Improving Performance of Recursive General Purpose Applications on GPUs

摘要

Due to the demand of high definition graphics presentation in gaming and video market, Graphics Processing Units (GPUs) have drastically increased their computational capacities. General-Purpose computation on GPUs uses the fragment shader multicore of these processing units to concurrently process data streams. However, the I/O overheads in recursive GPGPU applications have a negative impact in the performance of those systems. This paper proposes the remap method to improve the performance of general purpose recursive applications on GPUs, by decreasing the I/O overheads imposed by the VRAM/GPU interface. It is shown that significant performance improvements are achieved by applying the remap method to realistic recursive applications.