BEGIN:VCALENDAR
PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN
VERSION:1.0
BEGIN:VEVENT
DTSTART:20131121T223000Z
DTEND:20131122T000000Z
LOCATION:601/603
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: With computer systems becoming increasingly parallel and heterogeneous, autotuning is turning out to be essential in achieving performance portability for parallel applications.=0A=0AMy current research focuses on input-aware autotuning, since autotuners that take characteristics of the input data along with other execution context into account can optimize more aggressively.  I describe a programmer-directed autotuning framework that facilitates autotuning of code variants, or alternative implementations of the same computation, and selects code variants at execution based on features of the input data set.=0A=0AIn real-world irregular benchmarks from sparse numerical methods and graph computations, variants selected using our system achieve up to 99.85% of the performance achieved by variants selected through exhaustive search.=0A=0AI also discuss my future research focus: how autotuning systems can be made more accessible to non-expert users by integrating them with higher-level programming models that decouple computation and implementation.
SUMMARY:A Framework for Input and Architecture Aware Code Variant Autotuning
PRIORITY:3
END:VEVENT
END:VCALENDAR