Compiling Stream Kernels for Polymorphous Computing Architectures

Polymorphous Computing Architectures (PCA) have multiple modes of operation and can reassign resources allocated to these modes during program execution. Such architectures enable a single computational fabric to meet the diverse computing needs of complex applications that previously required multiple, distinct HW/SW solutions integrated into a system solution. The MONARCH chip is a PCA capable of supporting dataflow, SIMD, and RISC execution models. This paper describes the parametric compilation of stream kernels to the MONARCH dataflow architecture. We focus on the use of a new architecture description language, the introduction of a dataflow intermediate representation and the code generation and optimization framework.