Zusammenfassung
Manual program parallelization and optimization may be necessary
to reach a decent portion of the target architecture's peak
performance when automatic tools fail at choosing the best
strategy. While a broad range of languages and libraries provide
convenient ways to express parallelism, the difficult, time
consuming and error-prone parallelism identification and
extraction task is mostly left under the programmer's
responsibility. To address this issue, we introduce a
visualization-based approach to ease parallelism extraction and
expression that leverages polyhedral compilation technologies.
Our interactive tool, Clint, maps direct manipulation of the
visual representation to polyhedral program transformations with
real-time semantics preservation feedback. We conducted two user
studies showing that Clint's visualization can be accurately
understood by both experts and non-expert programmers, and that
the parallelism can be extracted better from Clint's
representation than from the source code in many cases.
Nutzer