Joint Code and Data Minimization

Abstract

This chapter and the next discuss the problem of computing a single appearance schedule that minimizes the buffer memory requirement over all valid single appearance schedules. Thus, given the model of buffer implementation defined in Section 4.2, we wish to construct a software implementation that minimizes the data memory requirement over all minimum code-size implementations. Even for chain-structured SDF graphs, the number of distinct valid single appearance schedules increases combinatorially with the number of actors [Murt94a], and thus exhaustive evaluation is not, in a general, a feasible means to find the single appearance schedule that minimizes the buffer memory requirement. Section 6.3 develops an efficient dynamic programming algorithm that computes an optimal hierarchy of loops given a lexical ordering of the actors. For well-ordered graphs, where there is only one topological sorting of the vertices, the schedule that results from applying the dynamic programming algorithm is guaranteed to be the optimal one. For graphs that have more than one topological sort, we develop heuristics in Section 6.4 and Chapter 7 for generating suitable topological sorts. These are nested optimally using the dynamic programming algorithm.