Future Technologies Colloquium Series

ABSTRACT:

New representations of linear algebra calculations in a high-level optimizer expand the scope of programs that can be optimized and the range of hardware targets for such optimizations. A polyhedral abstraction exposes parallelism to a greater degree than available using previous high-level compiler abstractions, and allows ILP formulations of optimizations that balance parallelism with locality of reference. The resulting form can be rendered to emerging architectures with explicitly managed communication and local memories, as well as to classic shared memory multiprocessors. The R-Stream compiler provides an implementation of these abstractions and optimizations linked to a powerful C-language infrastructure to allow using this tool through a conventional programming language without special extensions. This seminar will discuss the polyhedral abstractions, the implementation, current results, and opportunities enabled by this new compiler technology.