Automatic Partitioning of Sequential Applications Driven by Domain-Independent Kernels

Abstract. The automatic parallelization of sequential applications is a great challenge for current compiler technology. The partitioning of a se-quential application into parallel programs that can be executed concur-rently on a given parallel architecture is a complex and time-consuming undertaking. In addition, the programmer is often responsible for defin-ing a good partitioning that takes into account the properties of both the program and the architecture. This paper proposes a new fully au-tomated partitioning algorithm driven by an intermediate representa-tion of the sequential application in terms of the domain-independent concept-level kernels (e.g., induction, reduction, recurrence) recognized by the XARK compiler framework. Such kernel-centric view of the ap-plication hides the complexity of the implementation details (e.g., proce-dure calls, pointers, global variables, complex control flows) and provides robustness against different codification styles. For illustrative purposes, we use inter-procedural implementations of the Sobel edge filter and the EQUAKE application of SPEC CPU2000.