Download Dataflow Networks, Intuitive and Formal Semantics - Lecture Slides | CPSC 489 and more Study notes Computer Science in PDF only on Docsity! 1 1 Dataflow Networks • Dataflow Networks • Syntax and Semantics – actor, tokens and firing • Static scheduling • Other dataflow models slide -ASV-UCB 2
Dataflow network
a A data-flow network is a collection of functional nodes
which are connected ‘and,communicate over
unbounded FIFO queues
a Nodes are commonly called-actors
a The bits of information that are communicated over
the queues are commonly called tokens
slide -ASV-UCB
Intuitive Semantics
a Unbounded FIFOs perform communication via
sequences-of tokens carrying values
« integer, (loa, fixed point
+ matrix of integer, float, fixed point,
« image of pixels
a Slate implemented as-self-loop.
a Determinacy:
« unique output sequences given unique input sequences
« Sutticient condition: blvcking read
(process cannot test input queues for emptiness)
slide -ASV-UCB
+ singh
+ single o
slide -ASV-UCB
« singl
« single o
¢ o(n) =cl ifn
10
slide -ASV-UCB
« sing]
+ single vo
¢ o(n) =cl ifn
slide -ASV-UCB
+ singh
+ single o1
+ o(n) =cl i
2
slide -ASV-UCB
¢ singh
a Exal
+ single o1
+ o(n) = el ifn
13
slide -ASV-UCB
* singl
+ single o
« o(n) = cl i(n!
“
slide -ASV-UCB
Chains of-sequences
Consider the set Sof all finite and infinite sequences of
tokens
a This set is partially orderedsby the prefix order
a A subset C of Sis called.a chaimill all pairs of elements
of C are comparable
If Cis a chain, then it must be.a linear order inside S
(hence the name chain)
mw Example: { [ x, ], [ x, x, ], Pp X93 |, ... } is a chain
a Example: { [ x, ], [ x, x, ],[ x4], «.. } is not a chain
slide -ASV-UCB 19
(Least) Upper. Bound
Givena suhset Y of Span upper bound-of ¥ is an
element of S such thatz.is larger than all elements of
¥
Consider now-the set -4,(subsetiof' S) of allthe upper
bounds of Y
If Z has a least element, then u is:called the least
upper bound (lub) of Y
a The least upper bound, if it exists, is unique
a Note: u might not he in Y (if itis; then it is the largest
value of Y)
slide -ASV-UCB 20
10
Complete Partia-Order
a Every chainin S has@least upper bound
a Because of this property, 'S is called‘a Complete Partial
Order
a Notation: if Cisachain, weindicate the leastupper
bound of’ C by lubCC}
a Note: the least upper bound may be thought of as the
limit of the the chain
slide -ASV-UCB 21
Processes
= Process: function from a p-tuple of sequences to a q-
tuple of sequences
F : Sp=pes4
Tuples have the induced pointwise order:
Y=(Yp..s¥ph Y°S(Y pastay’y) in SP:
¥ <= Y’ iff y; <=y?; forall [<=i<= p
a Given a chain C in SP, FCC) may or may not be a
chain in S4
a We are interested in conditions that make that true
slide -ASV-UCB 2
11
12 slide -ASV-UCB 23 Mahapatra- Texas A&M- Fall00 24 Summary of function class • Summary of function class and their relationship for the function F: Sp → Sq Sequential ⇒ continuous ⇒ monotonic
Static scheduling-of DF
= Key property. of DFnctworks: outputscquences do not
dependon timeof firing-of actors
a SDF networkscambe statically scheduled abcompile-
time
+ execute an actor when itis.knewnto. be fireable
* no overhead due fosequencing of coneurrency
« static buffer sizing
a Different schedules yield-different
+ code size
« buffer size
slide -ASV-UCB 2»
Static scheduling of. SDF
Based only on process graph (ignores-tunctionality)
Network state: numberof tokens in FIFOs
Objective: tind schedule that is valid, isec
+ admissible
(only fires actors when fireable).
perivdic
(brings network back4o initial state by firing each actor at
least once)
Optimize cost function overadmissible schedules
slide -ASV-UCB 30
15
—_==
—=_ ~
—_==
m Repetitions (or firing f schedule S: number
of firings of each actor in Ss ~~
>
m yo(A) n= v,(B) n, —
must be satisfied for each edge ~~
slide -ASV-UCB 31
a Balance for each edges -
+ 39(A)- YD) <0
+ ¥B)-¥(C)=0 0
+ 2vfA)- v(C) =0 —
. . 7
# 2v(A)- v0) =0 =
ww
slide -ASV-UCB 32
16
17 slide -ASV-UCB 33 Mahapatra-Texas A&M-Fall’00 34 Tagged Token DFM • Arvind and Gostelow 80’s • Each token has a tag, firing is enabled when tokens have matching tags. – No need of FIFO discipline in the channel