high performance computing on graphics processing units: hgpu.org

Modern graphics processing units (GPUs) provide high-performance general purpose computation abilities. They have massive parallel architectures that are suitable for executing parallel algorithms and operations. They are also throughput-oriented devices that are optimized to achieve high throughput for stream processing. Designing efficient GPU programs is a notoriously difficult task. The ForSyDe methodology is suitable to ease the difficulties of GPU programming. The methodology encourages software development from a high level of abstraction and then transforming the abstract model to an implementation through a series of formal methods. The existing ForSyDe models support the synchronous data flow (SDF) model of computation (MoC) which is suitable for modeling stream computations and is good for synthesizing efficient stream processing programs. There also exists high-level design models named parallel patterns that are suitable to represent parallel algorithms and operations. The thesis studies the method of modeling parallel algorithms using parallel patterns, and explores the way to synthesize efficient OpenCL implementation on GPUs for parallel patterns. The thesis also tries to enable the integration of parallel patterns into the ForSyDe SDF model in order to model stream parallel operations. An automation library that helps designing stream programs for parallel algorithms targeting GPUs is purposed in the thesis project. Several experiments are performed to evaluate the effectiveness of the proposed library regarding implementations of the high-level model.