COSPIM: a program optimization system for tightly-coupled heterogeneous environments

Processor-in-memory is a new class of computer architecture designed for reducing the performance gap between the processor and the memory. This architecture provides a tightly-coupled heterogeneous environment by integrating different processors in a system. An efficient parallelization and optimization mechanism is necessary for this system to transform the existed applications to achieve better performance. In this paper, we propose a comprehensive framework, COSPIM, based on the statement viewpoint in our early SAGE system. It integrates program decomposition, ETC (expected time to compute) evaluation and scheduling mechanisms together. We describe how COSPIM splits statements and produces schedule to execute on the host processor and the coprocessor simultaneously. The experimental results of this approach are also discussed.