MINES ParisTech

Decoupled Software Pipelining (DSWP) is a program partitioning method enabling compilers to extract pipeline parallelism from se-quential programs. Parallel Stage DSWP (PS-DSWP) is an exten-sion that also exploits the data parallelism within pipeline filters. This paper presents the preliminary design of a new PS-DSWP method capable of handling arbitrary structured control flow, a slightly better algorithmic complexity, the natural exploitation of nested parallelism with communications across arbitrary levels, with a seamless integration with data-flow parallel programming environments. It is inspired by loop-distribution and supports nest-ed/structured partitioning along with the hierarchy of control de-pendences. The method relies on a data-flow streaming extension of OpenMP. These advances are made possible thanks to progresses in com-piler intermediate representation. We describe our usage of the Static Single Assignment (SSA) form, how we extend it to the con-text of concurrent streaming tasks, and we discuss the benefits and challenges for PS-DSWP