Abstract

Skeletal parallel programming ease parallel programming by providing efficient ready-made skeletons. However, nested use of skeletons are difficult to be implemented in an efficiently and load unbalanced way. In this paper, we propose a novel transformation theorem called the segmented diffusion theorem, an extension of the diffusion theorem, which can flatten many nested uses of skeletons into more efficient skeletal parallel programs. This theorem is not only useful for guiding construction of efficient skeletal parallel programs, but also suitable for optimizing skeletal parallel programs in compilers.

Citations