Page 350 - DSP Integrated Circuits

7.10 Interpolator, Cont. 335

We choose to use four PEs to perform the 22 adaptor operations. Hence,
1 sample = o ^ PS-
Figure 7.73 shows the ASAP schedule
folded modulo(T) = 6Tp£. The number of con-
current operations is: 8, 0, 7, 0, 7, and 0. Obvi-
ously, a better schedule can be found. We
therefore reschedule the operations using the
loop folding technique in order to reduce the
number of concurrent operations to only four.
We assume that these operations can be
assigned to the four PEs.
Note that the critical path is through 11
adaptors—i.e., the length is 22 time units. If
four PEs are to be used, we must introduce
four pipeline stages into the algorithm. The
new schedule, which is shown in Figure 7.74,
is 13Tp£ long. For the sake of simplicity, data
transfers between adaptors have not been
shown explicitly in Figure 7.74. Figure 7.73 Folded ASAP schedule

Figure 7.74 Improved schedule for the adaptor operations

We redo the scheduling modulo(T samp/g),
as shown in Figure 7.75. As wanted, there
are only four concurrent operations. This
indicates that four PEs may be enough.
However, the nonpreemption constraint on
the operations might mean that we will
need extra PEs. Here we have unit delay
operations only, and this will make four
PEs sufficient. This is due to the fact that
unit delay operations can not overlap opera-
tions starting at another time instance.
Only concurrent operations can overlap. Figure 7.75 Folded schedule