<s>
A	O
Kahn	B-Application
process	I-Application
network	I-Application
(	O
KPN	O
,	O
or	O
process	O
network	O
)	O
is	O
a	O
distributed	B-Architecture
model	O
of	O
computation	O
in	O
which	O
a	O
group	O
of	O
deterministic	O
sequential	O
processes	B-Operating_System
communicate	O
through	O
unbounded	O
first	B-Operating_System
in	I-Operating_System
,	I-Operating_System
first	I-Operating_System
out	I-Operating_System
channels	O
.	O
</s>
<s>
The	O
model	O
requires	O
that	O
reading	O
from	O
a	O
channel	O
is	O
blocking	B-Operating_System
while	O
writing	O
is	O
non-blocking	B-Architecture
.	O
</s>
<s>
Due	O
to	O
these	O
key	O
restrictions	O
,	O
the	O
resulting	O
process	O
network	O
exhibits	O
deterministic	B-General_Concept
behavior	I-General_Concept
that	O
does	O
not	O
depend	O
on	O
the	O
timing	O
of	O
computation	O
nor	O
on	O
communication	B-Protocol
delays	I-Protocol
.	O
</s>
<s>
Kahn	B-Application
process	I-Application
networks	I-Application
were	O
originally	O
developed	O
for	O
modeling	O
parallel	B-Operating_System
programs	I-Operating_System
,	O
but	O
have	O
proven	O
convenient	O
for	O
modeling	O
embedded	B-Architecture
systems	I-Architecture
,	O
high-performance	B-Architecture
computing	I-Architecture
systems	O
,	O
signal	O
processing	O
systems	O
,	O
stream	B-Application
processing	I-Application
systems	I-Application
,	O
dataflow	B-Application
programming	I-Application
languages	I-Application
,	O
and	O
other	O
computational	O
tasks	O
.	O
</s>
<s>
KPN	O
is	O
a	O
common	O
model	O
for	O
describing	O
signal	O
processing	O
systems	O
where	O
infinite	O
streams	O
of	O
data	O
are	O
incrementally	O
transformed	O
by	O
processes	B-Operating_System
executing	O
in	O
sequence	O
or	O
parallel	O
.	O
</s>
<s>
Despite	O
parallel	O
processes	B-Operating_System
,	O
multitasking	B-Operating_System
or	O
parallelism	B-Operating_System
are	O
not	O
required	O
for	O
executing	O
this	O
model	O
.	O
</s>
<s>
In	O
a	O
KPN	O
,	O
processes	B-Operating_System
communicate	O
via	O
unbounded	O
FIFO	B-Operating_System
channels	O
.	O
</s>
<s>
Processes	B-Operating_System
read	O
and	O
write	O
atomic	O
data	O
elements	O
,	O
alternatively	O
called	O
tokens	O
,	O
from	O
and	O
to	O
channels	O
.	O
</s>
<s>
Writing	O
to	O
a	O
channel	O
is	O
non-blocking	B-Architecture
,	O
i.e.	O
</s>
<s>
it	O
always	O
succeeds	O
and	O
does	O
not	O
stall	O
the	O
process	O
,	O
while	O
reading	O
from	O
a	O
channel	O
is	O
blocking	B-Operating_System
,	O
i.e.	O
</s>
<s>
Processes	B-Operating_System
are	O
not	O
allowed	O
to	O
test	O
an	O
input	O
channel	O
for	O
existence	O
of	O
tokens	O
without	O
consuming	O
them	O
.	O
</s>
<s>
A	O
FIFO	B-Operating_System
cannot	O
be	O
consumed	O
by	O
multiple	O
processes	B-Operating_System
,	O
nor	O
can	O
multiple	O
processes	B-Operating_System
write	O
to	O
a	O
single	O
FIFO	B-Operating_System
.	O
</s>
<s>
Timing	O
or	O
execution	O
order	O
of	O
processes	B-Operating_System
must	O
not	O
affect	O
the	O
result	O
and	O
therefore	O
testing	O
input	O
channels	O
for	O
tokens	O
is	O
forbidden	O
.	O
</s>
<s>
Testing	O
input	O
channels	O
for	O
emptiness	O
(	O
or	O
non-blocking	B-Architecture
reads	O
)	O
could	O
be	O
allowed	O
for	O
optimization	O
purposes	O
,	O
but	O
it	O
should	O
not	O
affect	O
outputs	O
.	O
</s>
<s>
Assuming	O
process	O
P	O
in	O
the	O
KPN	O
above	O
is	O
constructed	O
so	O
that	O
it	O
first	O
reads	O
data	O
from	O
channel	O
A	O
,	O
then	O
channel	O
B	O
,	O
computes	O
something	O
and	O
then	O
writes	O
data	O
to	O
channel	O
C	O
,	O
the	O
execution	O
model	O
of	O
the	O
process	O
can	O
be	O
modeled	O
with	O
the	O
Petri	B-Operating_System
net	I-Operating_System
shown	O
on	O
the	O
right	O
.	O
</s>
<s>
When	O
data	O
arrives	O
at	O
channel	O
A	O
or	O
B	O
,	O
tokens	O
are	O
placed	O
into	O
places	O
FIFO	B-Operating_System
A	O
and	O
FIFO	B-Operating_System
B	O
respectively	O
.	O
</s>
<s>
The	O
transitions	O
of	O
the	O
Petri	B-Operating_System
net	I-Operating_System
are	O
associated	O
with	O
the	O
respective	O
I/O	O
operations	O
and	O
computation	O
.	O
</s>
<s>
The	O
number	O
of	O
unconsumed	O
tokens	O
depends	O
on	O
the	O
execution	O
order	O
(	O
scheduling	O
)	O
of	O
processes	B-Operating_System
.	O
</s>
<s>
A	O
spontaneous	O
data	O
source	O
could	O
produce	O
arbitrarily	O
many	O
tokens	O
into	O
a	O
channel	O
if	O
the	O
scheduler	O
would	O
not	O
execute	O
processes	B-Operating_System
consuming	O
those	O
tokens	O
.	O
</s>
<s>
FIFO	B-Operating_System
bounds	O
can	O
be	O
mathematically	O
derived	O
in	O
design	O
to	O
avoid	O
FIFO	B-Operating_System
overflows	O
.	O
</s>
<s>
FIFO	B-Operating_System
bounds	O
can	O
be	O
grown	O
on	O
demand	O
.	O
</s>
<s>
Blocking	B-Operating_System
writes	O
can	O
be	O
used	O
so	O
that	O
a	O
process	O
blocks	O
if	O
a	O
FIFO	B-Operating_System
is	O
full	O
.	O
</s>
<s>
Processes	B-Operating_System
that	O
have	O
no	O
input	O
channels	O
act	O
as	O
data	O
sources	O
and	O
processes	B-Operating_System
that	O
have	O
no	O
output	O
channels	O
act	O
as	O
data	O
sinks	O
.	O
</s>
<s>
Processes	B-Operating_System
of	O
a	O
KPN	O
are	O
deterministic	O
.	O
</s>
<s>
Processes	B-Operating_System
can	O
be	O
modeled	O
as	O
sequential	O
programs	O
that	O
do	O
reads	O
and	O
writes	O
to	O
ports	O
in	O
any	O
order	O
or	O
quantity	O
as	O
long	O
as	O
determinism	O
property	O
is	O
preserved	O
.	O
</s>
<s>
Hence	O
,	O
timing	O
of	O
the	O
processes	B-Operating_System
does	O
not	O
affect	O
outputs	O
of	O
the	O
system	O
.	O
</s>
<s>
KPN	O
processes	B-Operating_System
are	O
monotonic	O
.	O
</s>
<s>
This	O
KPN	O
could	O
,	O
for	O
example	O
,	O
be	O
used	O
to	O
map	O
the	O
KPN	O
onto	O
an	O
FPGA-based	O
platform	O
systematically	O
.	O
</s>
<s>
The	O
Ambric	B-Device
Am2045	O
massively	B-General_Concept
parallel	I-General_Concept
processor	I-General_Concept
array	I-General_Concept
is	O
a	O
KPN	O
implemented	O
in	O
actual	O
silicon	O
.	O
</s>
<s>
Thus	O
its	O
channels	O
are	O
strictly	O
bounded	O
with	O
blocking	B-Operating_System
writes	O
.	O
</s>
<s>
The	O
AI	O
Engine	O
's	O
in	O
some	O
AMD	O
Xilinx	O
Versals	O
are	O
building	O
blocks	O
of	O
a	O
Kahn	B-Application
Process	I-Application
Network	I-Application
.	O
</s>
