Page 373 - DSP Integrated Circuits

358 Chapter 8 DSP Architectures

We will discuss different types of interconnection networks in more detail in
section 8.6.1.
The difference between system
and subsystem architectures is
essentially due to the difference in
granularity of the processes exe- Figure 8.2 Cascade form
cuted. Process granularity tends to
be larger at the higher levels and
lower at tne subsystem levels. For example, a process at tne system level may be
an FFT and at the lowest level an arithmetic operation. In practice, the most com-
mon architectures at the higher system levels are the cascade form and the paral-
lel form, shown in Figures 8.2 and 8.3, respectively.
Other common system architectures have
feedback loops and may have several levels of
hierarchies. Large DSP systems are usually
organized as hybrids of these forms. From a
computational point of view, the cascade form
is well suited for pipelining while the sub-
systems can be executed concurrently in paral-
lel form.
Most decomposition approaches at the
Figure 8.3 Parallel form
higher system levels exploit the fact that many
DSP algorithms are naturally partitioned into
parts with small intercommunication require-
ments. The first design step is therefore to identify these parts. Internally these
parts may have high communication demands, for example, a recursive digital fil-
ter. Such tightly coupled parts of an algorithm must be implemented using shared-
memory architectures. These architectures will be discussed in section 8.9.
A message-based system
architecture, with the subsystems
as components, can be chosen if
the communication requirements
between the subsystems are
small. Otherwise, the system
must be implemented with a
shared-memory architecture. Mes-
sage-based architectures will be
discussed in section 8.6.
The subsystems can be imple-
mented using, for example, a stan-
dard digital signal processor
(uniprocessor architecture) or an
ASIC processor if the work load is
small, whereas a specialized multi-
processor architecture is required
for larger work loads. A complex
DSP system can be composed of
several architectural alternatives Figure 8.4 DSP system implemented using
as illustrated in Figure 8.4. several types of architectures