SHF: Small: High-Level Programming Models for GPUs

点击翻译按钮获取中文摘要

Modern Graphics-Processor Units (GPUs) are capable of performance that, just a few years ago, would have been classified as supercomputer-level. With the trend of integrating GPU cores into heterogeneous multicore processors, GPUs are becoming an important source of future performance growth in mainstream processors. Unfortunately, GPUs are notoriously hard to program, especially for irregular parallel computations. This project aims to address the challenges of programming GPUs by supporting higher-level programming models with advanced compilation techniques. The intellectual merits of the proposed work are that it advances the state of the art in compilation techniques and programming models for GPUs and other accelerator architectures. The broader impact of the project is to widen the applicability of GPUs to a wider range of computational problems and, in turn, to help make GPUs useful to a broader community of users by supporting higher-level programming models for GPUs that are easier to program.The project focuses on the use of Nested Data Parallelism (NDP) and the supporting global flattening transformation, which supports irregular parallelism by compiling it down to flat data parallelism. While NDP provides a high-level elegant programming model for many kinds of irregular parallel computations, a straightforward implementation is not competitive with hand-written GPU code. The goal of this project it to develop and evaluate a collection of techniques for compiling NDP code to GPU with the objective of making NDP competitive with hand-written CUDA and OpenCL code. The work will be carried out in the context of a compiler for Blelloch's NESL language, which is a small first-order functional language that embodies the core concepts of NDP. NESL provides a small, but expressive, context for the proposed research. The work is evaluated by benchmarking against handwritten CUDA and OpenCL solutions for various irregular parallel algorithms.