Abstract

We are in the midst of an explosion of new parallel architectures. Hardware like GPUs, FPG As, and Intel’s Xeon Phi can provide tremendous boosts in performance over standards CPUs for certain classes of computation, but programming this increasingly diverse array of devices is painful and often requires expert knowledge.

High-level declarative languages offer a tantalizing approach to exploiting these platforms by allowing the programmer to state what is to be computed without specifying all the low-level implementation details. But this seemingly turns a difficult problem for the programmer into an even more difficult problem for the compiler writer.

My secret weapon is metaprogramming. In this talk I will describe several systems I have designed and built that use metaprogramming to provide high-level declarative language interfaces to a number of exotic devices that nonetheless compile to efficient, low-level implementations. The combination of high-level declarative languages and metaprogramming offers a promising way forward to taming the parallel zoo.