GAP - general autonomous parallelizer for CUDA environment

Abstract

Writing efficient general-purpose programs for Graphics Processing Units (GPU’s) is a complex task. In order to be able to program these processors efficiently, one has to understand their intricate architecture, memory subsystem as well as the interaction with the Central Processing Unit (CPU). The paper presents the GAP - an automatic parallelizer designed to translate sequential ANSI C code to parallel CUDA C programs. The general processing architecture of GAP is presented. Developed and implemented compiler was tested on the series of ANSI C programs. The generated code performed very well, achieving significant speed-ups for the programs that expose high degree of data-parallelism. The results show that the idea of applying the automatic parallelization for generating the CUDA C code is feasible and realistic.