Compilation techniques for automatic extraction of parallelism and locality in heterogeneous architectures
José M. Andión · LA Referencia (Red Federada de Repositorios Institucionales de Publicaciones Científicas) · 2015
[Abstract] High performance computing has become a key enabler for innovation in science and industry. This fact has unleashed a continuous demand of more computing power that the silicon industry has satisfied with parallel and heterogeneous architectures, and complex memory hierarchies. As a consequence, software developers have been challenged to write new codes and rewrite the old ones to be efficient in these new systems. Unfortunately, success cases are scarce and require huge investments in human workforce. Current compilers generate peak-peformance binary code in monocore architectures. Following this victory, this thesis explores new ideas in compiler design to overcome this challenge with the automatic extraction of parallelism and locality. First, we present a new compiler intermediate representation based on diKernels named KIR, which is insensitive to syntactic variations in the source code and exposes multiple levels of parallelism. On top of the KIR, we build a source-to-source approach that generates parallel code annotated with compiler directives: OpenMP for multicores and OpenHMPP for GPUs. Finally, we model program behavior from the point of view of the memory accesses through the reconstruction of affine loops for sequential and parallel codes. The experimental evaluations throughout the thesis corroborate the effectiveness and efficiency of the proposed solutions.