This proposal represents a one-year extension to the current X-TUNE project in the DOE Exascale Software Stack program. X-TUNE has been developing domain-specific compiler optimization and autotuning systems to address the looming programming challenges of exascale systems and deploy this technology in important DOE applications and representative computation kernels. Autotuning empirically evaluates a search space of possible implementations of a computation to identify the implementation that best meets its optimization criteria (performance, power, or both). In an exascale regime where processors and memory systems are anticipated to be heterogeneous and hierarchical, exascale applications will need to partition computation across different processors and manage data placement and data transfer horizontally and vertically in the memory hierarchy. Autotuning is expected to play a crucial role in managing the unprecedented complexity of programming exascale systems. Through automation, X-TUNE has demonstrated the capability of generating implementations that are superior to what a programmer can derive on his or her own, resulting in a higher-level, portable code base and with much reduced development time. The promise of autotuning has already been demonstrated by expert tool developers and performance experts targeting petascale architectures, and in particular the team assembled for this proposal has played a leadership role in advancing autotuning technology. Nevertheless, much work is needed to ready this technology for exascale and deliver it to the DOE application community. In this additional year of the project, we plan to focus our efforts on Geometric Multigrid (GMG) Computations, demonstrating end-to-end optimization for a variety of configurations of a GMG solver as represented by the miniGMG and HPGMG mini-applications.
