Functional Dataflow Programming Toolchain for Portable Parallel Code

Modern computer systems often incorporate multiple computational devices such as conventional processors, graphics processing units (GPUs), specialized accelerators, and field-programmable gate arrays (FPGAs). Programmers are able to exploit these new architectures for improved performance and reduced power consumption, but at the cost of increased programing complexity. Further complicating matters, architectural differences often require completely distinct kernel implementations to attain acceptable relative performance, even within the same family of devices, resulting in code that must be constantly maintained. Therefore, it is highly desirable to have a unified programming language and paradigm of sufficient abstraction such that software can be written effectively for the current target but can also be compiled for a multitude of possible other architectures. To address this need, we propose a portable, highly parallel, set of tools based on C++. Programs written in our specialized embedded domain-specific language (EDSL) can be used to rewrite portions of their programs, which will automatically be optimized for their target platforms. Using this EDSL leads to software that contains complete dependency information for each function within a program, allowing automated dataflow analysis, optimization, and code generation to be feasibly performed on a large scale.