Copyrights to these papers may be held by the publishers. The download files are preprints. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.

E. Tang, James C. Hoe and Franz Franchetti (Proc. SRC TECHCON, 2024)
Magic Memory: A Memory-Centric Declarative Programming Paradigm to Enable High Productivity on Heterogeneous Systems
Preprint (558 KB)
Bibtex

Heterogeneous computing systems involving CPUs, GPUs, FPGAs and even ASICs have already been shown to have immense computational power for various applications. However, programming these systems efficiently and effectively remains a significant challenge. Traditional imperative programming approaches result in complex programs that require expert-level understanding in order to optimize or maintain. The declarative programming model, Magic Memory, focuses on the behavior of the program and thus hides the nature of the underlying hardware that executes the program. In this model, computation is expressed using mathematical functions between arrays. This function can then be treated as an invariant that is always held true. During steady state, computation is performed by observing changes to these magic memory regions and then only recomputing the values that are affected by the change. As a first demonstration of this work, we utilize a CPU-FPGA system to compute the PageRank of a graph and recompute the PageRank after adding or removing an edge from the graph.

Keywords:
CPUs, FPGA, GPUs, Program generation