Optimizing Streaming Parallelism on Heterogeneous Many-Core Architectures