In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

Xiaonan Tian

will defend his dissertation

A Compiler Optimization Framework for Directive-Based GPU Computing

Abstract

Accelerator-based computing, commonly utilizing Graphics Processing Units (GPUs), have gained wide popularity on the road to exascale. However, significant advances in programming models are required for accelerator-based systems in order to close the gap between achievable and theoretical peak performance. Furthermore, these advances should not come at the cost of programmability. In particular, two challenges for programming models are addressed in this work: (1) they must support fine-grained parallel ism and locality-awareness within a chip, and (2) they should provide an incremental migration path for existing applications targeting future systems. This dissertation focuses on a directive-based solution which satisfies the second challenge by enabling incremental parallelization of existing codes. Directive-based programming models for accelerators, such as OpenACC and OpenMP, help non-expert programmers to parallelize applications productively and achieve portable performance. However, there is typically a performance gap between these high level directive-based approaches and using lower level programming interfaces (e.g. CUDA or OpenCL), indicating that these models have not reached maturity yet. To shrink this performance gap, more research is needed in developing optimized compiler implementations of these directives, as well as exploring additional directives. This dissertation focuses on paticular topics that are important for targeting GPUs : loop scheduling tranformatio ns and data locality optimizations.

Date: Friday, April 22, 2016
Time: 10:00 AM
Place: PGH 218
Advisor: Dr. Barbara Chapman

Faculty, students, and the general public are invited.