ANSA: a new neural net based scheduling algorithm for high level synthesis

In this paper, we expand our earlier neural network method for solving the scheduling problem in high level synthesis. The new algorithm, ANSA, operates in three phases. The first phase is a normalized mean field net algorithm with a new energy function which incorporates weighting of different operation types to create deeper basins of attraction. Other novelties include a fast and deterministic noise generation scheme and a new melting technique to determine the starting temperature. The next two stages provide a mechanism for finding nonuniformly distributed optimal schedules. The second phase uses the same energy function as the first but with a bias in favor of aligned operations. The third stage is a probabilistic correction algorithm for cases with highly irregular subschedules. ANSA was tested on five benchmark examples including large ones such as the discrete cosine transform for all possible schedule lengths with and without pipelining. It achieved a 100% convergence rate to optimal solutions in all cases.