Page 335 - DSP Integrated Circuits
P. 335

320                                             Chapter 7 DSP System Design


        7.6.7 Maximum Spanning Tree Method
        The maximum spanning tree method, which was developed by Renfors and Neuvo
        [20], can be used to achieve rate optimal schedules. The method is based on graph-
        theoretical concepts. The starting point is the fully specified SFG. The SFG corre-
        sponds to a computation graph N, which is formed by inserting the operation
        delays into the SFG. Further, a new graph N' is formed from N by replacing the
        delay elements by negative delay elements (-T). In this graph we find the maxi-
        mum distance spanning tree—i.e., a spanning tree where the distance from the
        input to each node is maximal. Next, shimming delays are inserted in the link
        branches (i.e., the remaining branches) so that the total delay in the loops becomes
        zero. Finally, remove the negative delays. The remaining delays, apart from the
        operation delays, are the shimming delays. These concepts are described by an
        example.




        EXAMPLE 7.7
        Consider the second-order section as in Example 7.6 and derive a rate optimal
        schedule using the maximum spanning tree method.
           Figure 7.47 shows the computation
        graph N'—i.e., the computation graph with
        the operation delays inserted and T
        replaced by — T. Note that the graphs are of
        the type activity-on-arcs and that the delay
        of adders is assigned to the edge
        immediately after the addition. The
        maximal spanning tree is shown in Figure
        7.48 where edges belonging to the tree are
                                                Figure 7.47 Modified computation
        drawn with thick lines and the link                graph N'
        branches with thin lines. Next, insert link
        branches one by one and add shimming
        delays so that the total delay in the fundamental loops that are formed becomes
        zero. Finally, we remove the negative delays and arrive at the scheduled
        computation graph shown in Figure 7.49.
           Note that there are no shimming delays in the critical loops. All operations are
        scheduled as early as possible, which in general will be suboptimal from a resource















        Figure 7.48 Scheduled computation        Figure 7.49 The maximum spanning
                  graph                                    tree of AT
   330   331   332   333   334   335   336   337   338   339   340