high performance computing on graphics processing units: hgpu.org

Graphics processing units (GPUs) provide high performance at low power consumption as long as resources are well utilized. Thread block size is one factor in determining a kernel’s occupancy, which is a metric for measuring GPU utilization. A general guideline is to find the block size that leads to the highest occupancy. However, many combinations of block and grid sizes can provide highest occupancy, but performance can vary significantly between different configurations. This is because variation in thread structure yields different utilization of hardware resources. Thus, optimizing for occupancy alone is insufficient and thread structure must also be considered. It is the programmer’s responsibility to set block size, but selecting the right size is not always intuitive. In this paper, we propose using machine learning to automatically select profitable block sizes. Additionally, we show that machine learning techniques coupled with performance counters can provide insight into the underlying reasons for performance variance between different configurations.