Solving the scheduling problem in high level synthesis using a normalized mean field neural network

A neural network solution to the time constrained scheduling problem in high level synthesis of digital circuits is proposed. The mean field theory neural network with graded neurons, proposed by Peterson and Soderberg (1989) is adapted and renamed as normalized mean field net. With the use of ASAP and ALAP schedules as limiting constraints, the number of neural variables is kept to a scalable size resulting in a fast and efficient implementation. An extension to include multi-cycle operations is presented. The proposed network is simulated and tested on two examples including a fairly large benchmark circuit from the 1988 High Level Synthesis Workshop. In all cases, the network is able to find optimal solutions in the first trial