Code Generation for Process Network Models onto Parallel Architectures

With multi-core and many-core architectures becoming the current focus of research and development, and as vast varieties of architectures and programming models emerging in research, the design space for applications is becoming enormous. From the number of cores, the memory hierarchy, the interconnect to even the programming model and language used are all design choices that need to be optimized for applications in order to fully benefit from parallel architectures. We propose a code generation framework targeting rapid design space exploration and prototyping. From the high level design, code for specific architectures and mappings can be generated and used for comparison. We choose Khan Process Networks[11] as our current specification language, because of its inherit parallelism and expressiveness. Our code generator take advantage of Message Passing Interface (MPI) [6] as the API for implementing message passing across platforms. We show the scalability of the generated MPI code and the ability to extend our framework to allow for tuning and optimization.