The International Conference for High Performance Computing, Networking, Storage and Analysis
MetaMorph: A Modular Library for Democratizing the Acceleration of Parallel Computing Across Heterogeneous Devices.
Authors: Paul D. Sathre (Virginia Polytechnic Institute and State University), Wu-chun Feng (Virginia Polytechnic Institute and State University)
Abstract: Heterogeneity in computing has become ubiquitous. Computing systems ranging from smartphones to supercomputers now consist of multiple types of computing brains, typically at least one multi-core CPU and a many-core GPU. Such systems offer the promise of increased performance and energy efficiency over CPUs if the resources within such systems are used judiciously.
Alas, extracting the full performance potential from accelerator devices requires architectural expertise, expertise that is in short supply. Thus, there exists a need for software tools and ecosystems to support the development, maintenance, and upgrading of accelerated codes by non-expert domain scientists. To address this need, we present our initial prototype of a modular library of drop-in accelerated functions, which abstracts current and future accelerator backends behind a single, unified API, known as MetaMorph. By incrementally replacing computational primitives with calls to the MetaMorph API, domain scientists can transparently migrate code to current and future accelerator devices.