SaC: Off-the-shelf Support for Data-Parallelism on Multicores

Опубликовано 7 сентября 2016, 16:59

The advent of multicore processors has raised new demand for expressing and exploiting concurrency in the software mass market.  This talk summarizes previous work on the data parallel, functional array processing language Single Assignment C (SaC for short).  In its context we have developed sophisticated optimization techniques that transform high-level, generic array programs into highly efficient multi-threaded code executable on SMP architectures without any explicit concurrency annotations.  The recent introduction of such machines in the mass market renders SaC a ready to use technology for data parallel applications.  Its high-level of abstraction and its low-overhead C interface make it an excellent vehicle for exploiting data-parallelism by rewriting / embedding application kernels of existing legacy code in SaC. The talk focuses on the key compiler technologies that are crucial for the runtime achievements of SaC.