PipelinedWormhole Routers - Microarchitecture of Network-on-Chip Routers - page 82

credit
update

outPort outAvailable[0...N-1]

ready out[0...N-1]

CC

RC

SU

en

1

req

SA

outLock 1

dst

en

head

sets on head
resets on tail credit update

dequeue

1

1

valid
data

valid
data

valid

ST

Input #i

Output #j

Fig. 5.11 The organization of the pipelined router that pipelines the grants of the SA unit and
keeps them for direct use by the body and tail flits of the same packet, thus breaking the dependency
across the request and the delayed arrival of the corresponding grants

0 1 2 3 4 5 6 7

cc
RC-SA H LT-BW DQ-ST LT-BW

su

cc
s SA -DQ-ST

B LT-BW LT-BW

cc
s SA -DQ-ST

T LT-BW LT-BW

cc
RC-SA H LT-BW DQ-ST LT-BW

su

Fig. 5.12 The operation of the alternative SA pipelined router, where body and tail flits move
directly to ST once they have the necessary credits by inheriting the grants produced by the head
flit of the same packet and stored in the corresponding pipeline register

the output buffer will receive via the output multiplexer body/tail flits that have not
been qualified for the necessary credits. To resolve this issue a mutiplexer is added
in the forward direction, as shown in Fig. 5.11 , that selects the qualified valid signal
produced by the request generation logic in the case of body/tail flits instead of the
normal valid signal.
The cycle-by-cycle operation of this alternative SA pipelined organization is
presented in Fig. 5.12 . A head flit is written to the input buffer in cycle 0 and issues
a request to SA in cycle 1 after having completed RC in the same cycle. The grants
from SA return in cycle 2 and saved for later use by the body and tail flits of the
same packet. In cycle 2, the head flit accepts the grant and gets dequeued moving
to its destined output via the output multiplexer (ST). In cycle 3, the body flit of the
same packet arrives at the frontmost position of the input buffer. Without sending
any request to the SA (requests are actually generated only for the purpose of ready
qualification) once the body flit has at least one credit for its selected output it gets
dequeued and moves to ST after consuming one credit. The ST stage will switch