Reducing the Complexity of Heterogeneous Computing: A Unified Approach for Application Development and Runtime Optimization

Heterogeneous systems with accelerators promise considerable performance improvements at a lower cost than homogeneous CPU-only systems. However, to benefit from this potential, considerable work is required from developers to integrate them efficiently in an application. This work contributes a new framework implemented with an online-learning runtime system that simplifies development and makes applications more portable, efficient and reliable across different systems.