Page 310 - DSP Integrated Circuits
P. 310

7.5 Scheduling Formulations                                           295

        critical loop of the SFG contains
        more than one delay element, the
        critical path restricting the sample
        period can not be found. Also, non-
        uniform boundaries make it more
        difficult to evaluate the cost func-
        tion since operations belonging to
        different sample intervals are per-
        formed concurrently.               Figure 7.9 Schedule with nonuniform
                                                     boundaries

        EXAMPLE 7.1
        Figure 7.10 shows a second-
        order section in direct form II.
        The basic operations are
        multiplication and addition. A
        multiplication is assumed to
        take four time units and an
        addition takes one time unit.
        The required sample period is
        Tsample = 10 time units.
            The critical loop and the
        critical path of this algorithm
        are T mui t + 2T add = 6 and    Figure 7.10 Second-order section in direct form II
        2T mui t + 3T add = 11 time
        units, respectively, Hence, to
        achieve the desired sample rate, we must either duplicate some computational
        resources or introduce delays to divide the critical path into smaller pieces
        (pipelining). We choose to introduce pipelining. The pipelined structure is shown
        in Figure 7.11 and the computation graph is shown in Figure 7.12. The dots
        indicate the starting and finishing times of the operations.
            Hiquaiizmg ana smmming
        delays have been inserted in
        order to meet the specified
        sample period by extending
        all computational paths to 10
        time units. An equalizing
        delay of four time units has
        been placed after the last
        adder. A shimming delay, with
        unit time, has been placed
        after the multiplication by OQ-
        The output in the recursive
        loop has been delayed by an
        equalizing delay of four time
        units, so that the time differ-
        ence between the input and
        output is 10 time units. The   Figure 7.11 Second-order section with pipelining
   305   306   307   308   309   310   311   312   313   314   315