| Summary: | 碩士 === 國立交通大學 === 資訊工程學系 === 84 === The purpose of this thesis is to design a loop
paralellization techniquefor non-uniform data dependence loops.
In the past studies, most loop parallelizationtechniques are
focused on the problem of uniform dependence loops. However,
whenthe loop is non-uniform dependent these methods are either
fail or inefficient[Dart 94][Lamp 74][Dhol 92][Berg 87].
According to the survey of the relationbetween array subscripts
and data dependence by Zhiyu[Shen 89], there are almost40% of
the linear array subscripts are coupled array subscripts which
is tightlyrelated to non-uniform dependence relations. Due to
the motivation mentioned above,we focus our loop parallelization
technique on non-uniform dependence loops. Inour thesis, we
proposed a further improved method based on Punyamurtula's[Puny
94]to expore more parallelism. In our method, compiler will
choose proper loop transformations to make the benefit for loop
parallelization(Tiling) according tothe four cases derived form
all dependence relations in loop body so as to exporeparallelism
as much as possible. Our approach improves the parallelism of
case II and case III of method [Puny 94]. Furthermore, it also
expand the problem domainto a more general loop model which
includes multiple set of dependence relations.Besides, we build
our method on a parallel compilation environment(SUIF) and
evaluate the results on a parallel simulation environment which
developed by ourlab. in another project. These preliminary
results can be referenced for non-uniform dependence loop
parallelization techniques.
|
|---|
