Automatic parallelization of stream programs for resource efficient embedded processors

This thesis considers how to exploit the specific characteristics of data streaming functions and multi-core processors to increase throughput through appropriate software process mappings. The hypothesis is that large numbers of low-power processors can achieve high throughput for streaming applications if a good mapping is provided. The innovation is to use compilation principles to guide the mapping, rather than heuristics. Three increasingly complex approaches are developed that focus on computational bottlenecks, then adds communication overheads, and lastly adds the costs of splitting and merging operations. Using this approach demonstrates that the successively more complex models can achieve correspondingly greater throughput