The Case for Directive Programming for Accelerator Autotuner Optimization

TitleThe Case for Directive Programming for Accelerator Autotuner Optimization
Publication TypeTech Report
Year of Publication2017
AuthorsFayad, D., J. Kurzak, P. Luszczek, P. Wu, and J. Dongarra
Technical Report Series TitleInnovative Computing Laboratory Technical Report
Numbericl-ut-17-07
Date Published10-2017
InstitutionUniversity of Tennessee
Abstract

In this work, we present the use of compiler pragma directives for parallelizing autotuning of specialized compute kernels for hardware accelerators. A set of constructs, that include prallelizing a source code that prune a generated search space with a large number of constraints for an autotunning infrastructure. For a better performance we studied optimization aimed at minimization of the run time.We also studied the behavior of the parallel load balance and the speedup on four different machines: x86, Xeon Phi, ARMv8, and POWER8.

Project Tags: 
External Publication Flag: