Operation scheduling for parallel functional units using genetic algorithms

We describe a new and efficient approach to solve the scheduling problem for VLIW architectures. The scheduling times of the operations are used as the problems parameters. This in conjunction with a pruning technique based on critical path analysis leads to a significant reduction of search space complexity. A genetic algorithm is used to search for valid schedules of a given length. The genetic algorithm uses a fitness vector that guides the genetic operators crossover and mutation resulting in a fast convergence towards near perfect solutions. The proposed method is also applicable to the problem of register allocation by using a different fitness function. Another advantage of the genetic algorithm approach is that usually a great number of equally performing schedules is obtained allowing for further optimization subject to arbitrary constraints.