high performance computing on graphics processing units: hgpu.org

Heterogeneous processing systems have become the industry standard in almost every segment of the computing market from servers to mobile systems. In addition to employing shared/distributed memory processors, the current trend is to use hardware components such as field programmable gate arrays (FPGAs), single instruction multiple data (SIMD) engines and graphics processing units (GPUs) in heterogeneous systems. As a result of this trend, extracting maximum performance requires compilation to highly heterogeneous architectures that include parts with different memory and computation models. Although there has been significant amount of research on programing each of these architectures individually, targeting a heterogeneous system without specializing an application to each component separately is still an open problem. Besides performance, the portability of an application between different pieces of a system and retargetability to various heterogeneous architectures is a significant challenge for programmers. To efficiently exploit the heterogeneity, it is necessary to have a programming model that provides a higher-level of abstraction to the programmer and the related compilation framework. In this thesis, we first focus on enabling a write-once programming paradigm in the context of the stream programming model for various components of heterogeneous systems. We mainly focus on FPGAs, SIMD engines and GPUs as these architectures will play an important role in accelerating various parts of applications on heterogeneous systems. We introduce several compiler optimizations that facilitate portability and retargetability while achieving high performance. As a result of our compilation system, programmers can write a program once and efficiently run it on different components of a system. Second, we focus on an important challenge that arises in heterogeneous systems when there are dynamic resource changes. The ability to dynamically adapt a running application to a target architecture in the face of changes in resource availability (e.g., number of cores, available memory or bandwidth) is crucial to a wider adoption of heterogeneous architectures. In this work, we introduce a hybrid flexible compilation framework that facilitates dynamic adaption of applications to the changing characteristics of the underlying architecture.