Whippletree: Task-based Scheduling of Dynamic Workloads on the GPU

摘要

In this paper, we present Whippletree, a novel approach to schedulingrndynamic, irregular workloads on the GPU. We introduce a newrnprogramming model which offers the simplicity and expressivenessrnof task-based parallelism while retaining all aspects of the multilevelrnexecution hierarchy essential to unlocking the full potential of arnmodern GPU. At the same time, our programming model lends itselfrnto efficient implementation on the SIMD-based architecture typicalrnof a current GPU. We demonstrate the practical utility of our modelrnby providing a reference implementation on top of current CUDArnhardware. Furthermore, we show that our model compares favorablyrnto traditional approaches in terms of both performance as well asrnthe range of applications that can be covered. We demonstrate thernbenefits of our model for recursive Reyes rendering, procedural geometryrngeneration and volume rendering with concurrent irradiancerncaching.