Automatic Parallelization of Tiled Stencil Loop Nests on GPUs

Download files
Access & Terms of Use
open access
Copyright: Di, Peng
Altmetric
Abstract
This thesis attempts to design and implement a compiler framework based on the polyhedral model. The compiler automatically parallelizes loop nests; especially stencil kernels, into efficient GPU code by loop tiling transformations which the polyhedral model describes. To enhance parallel performance, we introduce three practically efficient techniques to process different types of loop nests. The experimental results of our compiler framework have demonstrated that these advanced techniques can outperform previous approaches. Firstly, we aim to find efficient tiling transformations without violating data dependences. How to select a tile's shape and size is an open issue that is performance-critical and influenced by GPU's hardware constraints. We propose an approach to determine the tile shapes out of consideration for improving two-level parallelism of GPUs. The new approach finds appropriate tiling hyperplanes by embedding parallelism-enhancing constraints into the polyhedral model to maximize intra-tile, i.e., intra-SM parallelism. This improves the load balance among the streaming processors (SPs), which execute a wavefront of loop iterations within a tile. We eliminate parallelism-hindering false dependences to optimize inter-tile, i.e., inter-SM parallelism. This improves the load balance among the streaming multiprocessors (SMs), which execute a wavefront of tiles. Furthermore, to avoid combinatorial explosion of tile size's configurations, we present a model-driven approach to automating tile size selection that is performance-critical for loop tiling transformations, especially for DOACROSS loop nests. Our tile size selection model accurately estimates the execution times of tiled loop nests running on GPUs. The selected tile sizes lead to the performance results that are close to the best observed for a range of problem sizes tested. Finally, to address the difficulty and low-performance of parallelizing widely used SOR stencil loop nests, we present a new tiled parallel SOR method, called MLSOR, which admits more efficient data-parallel SIMD execution on GPUs. Unlike the previous two approaches that are dependence-preserving, the basic idea is to algorithmically restructure a stencil kernel based on a non-dependence-preserving parallelization scheme to avoid pipelining for higher parallelism. The new approach can be implemented in compilers through a pattern matching pass to optimize SOR-like DOACROSS loop nests on GPUs.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Author(s)
Di, Peng
Supervisor(s)
Xue, Jingling
Creator(s)
Editor(s)
Translator(s)
Curator(s)
Designer(s)
Arranger(s)
Composer(s)
Recordist(s)
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
2013
Resource Type
Thesis
Degree Type
PhD Doctorate
UNSW Faculty
Files
download public version.pdf 2.69 MB Adobe Portable Document Format
Related dataset(s)