ILGA: An Improved Local Greedy Algorithm for Optimal Parameters Searches

In program optimization, architecture-aware optimizations are often used to solve problems for best parameters of the same program in different configurations and architectures are different. Because the large size of the search space and the cost of evaluating the performance of object codes with different parameters, the search process is critical for iterative compilation. As a result, these problems above make it infeasible to find the true optimal value of the optimization parameter by brute force. In this paper, we focus on finding better searching algorithms to find the best parameters. An improved local greedy algorithm to find the near optimal pa-rameter is firstly proposed. Then its development with moving probabilities and many improvement policies is proposed. In experiments, we compared them with many other searching algorithms. The results show that ILGA can generate better performance compared with the random algorithm, genetic algorithm, and traditional local greedy algorithm.