Section: New Results

Compilation for scalable on-chip parallelism

Participants : Antoniu Pop, Feng Li, Sven Verdoolaege, Govindarajan Ramaswamy, Albert Cohen.

Task-parallel programming models are getting increasingly popular. Many of them provide expressive mechanisms for inter-task synchronization. For example, OpenMP 4.0 will integrate data-driven execution semantics derived from the StarSs research language. Compared to data-parallel and fork-join models of parallelism, the advanced features being introduced into task-parallel models in turn enable improved scalability through load balancing, memory latency mitigation, mitigation of the pressure on memory bandwidth, and as a side effect, reduced power consumption.

We developed a systematic approach to compile a loop nest into concurrent, dependent tasks. We formulated a partitioning scheme based on the tile-to-tile dependences, represented as affine polyhedra. This scheme ensures at compilation time that tasks belonging to the same class have the same, fully explicit incoming and outgoing dependence patterns. This alleviates the burden of a full-blown dependence resolver to track the readiness of tasks at run time. We evaluated our approach and algorithms in the PPCG compiler, targeting OpenStream, our experimental data-flow task-parallel language with explicit inter-task dependences and a lightweight runtime. Experimental results demonstrate the effectiveness of the approach.

Part of this work also involved our colleagues from the POLYFLOW associate-team at the Indian Institute of Science, Bangalore, India.