Data-centric compilation
Induprakas Kodukula · 1998
This dissertation presents a novel compiling strategy called data-centric compilation for generating highly optimized implementations of programs on modern high-performance computers. The goal of data-centric compiling is to optimize programs for machines with memory hierarchies by reducing the flow of data between different levels of this hierarchy. In addition, it can be used to parallelize semi-structured applications. Experiments on Octane workstations from Silicon Graphics are used to validate this approach in practice. These results have persuaded a major workstation vendor to adopt data-centric compilation in their compiler product line.