Coarse-Grained Reconfigurable Array Architectures - Signal Processing Systems - page 556

a b

pred
mux

src 0
mux

src 1
mux

opt src
3 mux

src 0
mux

src 1
mux

optional
pipe latch

optional
pipe latch

optional
pipe latch

optional
pipe latch

LOCAL RF
4 registers

ALU + MULT

flag

any combination
of operations feedback

latch

shift

optional
pipe latch

latch

ADRES issue slot

MorphoSys issue slot

c

ALU + MULT ALU + MULT

d

LOCAL
RF

IS 0
ALU

IS 1
ALU

IS 2
ALU

IS 3
ALU

src 0
mux

src 1
mux

src 0
mux

src 1
mux

shared
MUL

sw sw sw sw

ALU + MULT ALU + MULT

a single issue slot as proposed by Galains et al. issue slots that share a multiplier

Fig. 7 Four different structures of ISs proposed in the literature. Part ( a ) displays a fixed
MorphoSys IS, including its local RF. Part ( b ) displays the fully customizable ADRES IS, that can
connect to shared or non-shared local RFs. Part ( c ) depicts the IS structure proposed by Galanis
et al. [ 27 ] , and ( d ) depicts a row of four RSPA ISs that share a multiplier [ 33 ]

those that generate code for the predetermined homogeneous ISs of, e.g., the
MorphoSys CGRA. Given the state of the art in compiler technology for this type
of architecture, the advantages of this freedom are (1) the possibility to design
different instances optimized for certain application domains, (2) the knowledge
that the features of those designs will be exploited, and (3) the ability to compile
loops that feature other forms of ILP than DLP. DLP can still be supported, of
course, by simply incorporating SIMD-capable (a.k.a. subword parallel) ISs of,
e.g., 4

16 bits wide. The drawback of this design freedom is that, at least with
the current compiler techniques, these techniques are so generic that they miss
optimization opportunities because they do not exploit regularity in the designed
architectures. They do not exploit it for speeding up the compilation, nor do they
for producing better schedules. An other potential problem of specializing the ISs
for an application domain is overspecialization. While extensive specialization will
typically benefit performance, it can also have negative effects, in particular on
energy consumption [ 72 ] .
Figure 7 c depicts the IS proposed by Galanis et al. [ 27 ] . Again, all ISs are
identical. In contrast to the MorphoSys design, however, these ISs consist of several

×