Software Support for Heterogeneous Computing

Abstract

Heterogeneous computing, materialized in the form of multiprocessor system-on-chips (MPSoC) comprising of various processing elements such as general-purpose cores with differing characteristics, GPUs, DSPs, non-programmable accelerators, and reconfigurable computing, are expected to dominate the current and the future consumer device landscape. The heterogeneity enables a computational kernel with specific requirements to be paired with the processing element(s) ideally suited to perform that computation, leading to substantially improved performance and energy-efficiency. While heterogeneous computing is an attractive proposition in theory, considerable software support at all levels is essential to fully realize its promises. The system software needs to orchestrate the different on-chip compute resources in a synergistic manner with minimal engagement from the application developers. We present compiler time and runtime techniques to unleash the full potential of heterogeneous multi-cores towards high-performance energy-efficient computing on consumer devices.