Automatic Compilation of Scene Generation Applications on Clusters with Graphics Coprocessors

Contemporary off-the-shelf graphics processing units (GPUs) are programmable multiprocessors-on-a-chip that surpass their host processors at raw computational capacity and far surpass them at computation capacity per dollar. Further, they are expected to continue to outpace general-purpose processors. Recently, GPUs have been used to accelerate non-graphics floating-point intensive tasks but they remain difficult to program, in part because their rapid evolution makes them a moving target. We propose to include clusters of standard computers with GPUs as a target for an existing compiler framework that supports a parameterized architecture description. The compiler framework exploits a "streaming" model of computation common to a class of multiprocessor-on-a-chip architectures under development in DARPA's Polymorphous Computing Architecture (PCA) program. The first phase of the proposed work would consist of experiments to quantify the applicability of GPUs to the scene generation problem and the ability of the compiler framework to model clusters with GPUs. The second phase would consist of compiler implementation work to add clusters with GPUs as a supported target to the compiler.