2008 ICFP DEFUN - Functional in the Field

8/8/2019 2008 ICFP DEFUN - Functional in the Field

1/22

+

Functional in thefieldHow we locate animals with functionalprograms

Ryan Newton


2/22

+

App2: birds

What Im advocating: 3 things

n Functional programming

n Type/memory safety

n Polymorphism (generics)

n Factoring control patterns, functions/closures

n Stream programming

n Well understood parallel paradigm

n As library/middleware, but better with compiler support

n Domain Specific Languages

n StreamIt, Brook, WaveScript

Audio0

Audi

o1

Audio2

Audio3

DetectEvent

sy

nc

App1: marmots


3/22

Evaluation

+Streaming: similarities to functional

A u d ioIn p u t Detect

erverSideML DOA

e n d R a wD a ta

Eventassociation-seudolikelihoodMap

B u ffe r> 3se c?

un LocalML DOA

N

e n d A M LO u tp u t Y

ensor Node

erver Side

N o d eLo ca tio n s sn4


4/22

+ Why streamlanguages A big mess of hardware

No convenient way to writeprograms that use all pieces

Def. not high perf ones (Erlang,Java)

No phone/pc p2p services

Would re uire mi ratin

/D esktop Serve r

P h o n

e

/TC

PI

P

,In fra stru ctu reSensors

N o t/TC P IP

Ja v a M

E S y m b ia n

B re

w

iP h o n eS D K

A n d ro i

d

Tin y O

S


5/22

+Vision:A language for a streaming era

n Cross platform andfastn

n Convenience automate what can be automatedn Network transparency, scheduling, distribution

(sometimes)

n But recognize limitsn Compiler can help but cant do everything

n Parameterization, design exploration

P h o n

e


6/22

Execution Model

+

Execute!

Execution Model

iterate run code oneach element

One input element perexecution, zero or moreoutput elements

State not shared

One input, one outputstream, multiplesubscribers

ite rate

m erg e

= { ( ( ))S2 iterate xin S 1 if iseve n x{ ;emit x emit

; }x; }e lse e m it x= ( , )S 3 m e rg e S 1 S 2

S 1 S 2

M ust b e of the sam e typ e

Lift in to a u n io n ty p e


7/22

+Stream parallelism

n Explicit: between tasks

n Implicit: between consecutive stream elements

n Good when everything we need is in a contiguoussequence of stream elements

n Best when data rates are known (synchronous dataflow)


8/22

Optimization

+Stream Graph Optimizations

Fu sio n

/Fissio n R e p lica tio n

( )inliningM ust b e

!sta te le ss

/Fissio n A rra y b u ild in g

A ll so u rce to so u rce tra n sfo rm a tio n s


9/22

+ Case Study 1: LocatingMarmots

,G o th ic C O d e p lo y m e n t A u g u s t 2 0 0 7 V ox n e t P la tfo rm

, ,2 x P X A 2 5 5 6 4 M B R A M 8 G B, . ,Flash 80 2 11 B M ica2

, + ,sup e rvisor Li b a tte ryC h a rg e co n tro lle r

: , -S e n so rs 4 x 4 8 K H z a u d io 3, ,a xis acce l G PS In te rn a lte m


10/22


11/22

Evaluation

+Combine DOA pseudo-likelihoodsto generate 2D likelihood map

100

104

110105


12/22

+ Easy to move node/server boundary

A u d ioIn p u t Detect

erverSideML DOA

e n d R a wD a ta

Eventassociation-seudolikelihoodMap

B u ffe r> 3se c?

un LocalML DOA

N

e n d A M LO u tp u t Y

ensor Node

erver Side

N o d eLo ca tio n s sn4


13/22

Evaluation

+WaveScript Speedup

nFor this app: speedup in embeddedwith less code in core algorithmimplementation

C WS Speedup

VoxNetDOA

3.00s 2.18s 1.38

VoxNetDetect

87.9% 56.5% 1.56

LOC / WS LOC / C

Detector 92 252

DOA 124 239

:W in s fo r th is a p p

Po rta b le H LL N etw ork transparent W h o le p ro g ra m co m p ile

S tre am g rap h op ts


14/22

+Case Study 2: backgroundsubtraction for birdclassification


15/22

+Case Study 2: backgroundsubtraction for birdclassification


16/22

+Case Study 2: backgroundsubtraction for birdclassificationn Application in use by UCLA Sensing/Computer

Vision groupn Not experts in embedded computation

n Not parallel in obvious, streaming wayn Stateful saves model around each pixel

nParallel access to matrix but overlapping tilesn

n Not easy at all in, say, StreamIt

p ixe l

n eig h b o rh o od

Pe r p ixe l sto re d sta te


17/22

+Case Study 2: backgroundsubtraction for birdclassificationn Original code massive duplication (1300 lines of C++)

n Step 1: port verbatim

n Step 2: factor duplicated code, use HOF

n Step 3: remove unnecessary floating point

n Step 4: parameterize design, expose parallelism

n(and further simplify in the process!)


18/22

+Abstracting control patterns

n Map, fold (reduce) are cleaner than for loops

n Complex nested for loops

n Obligation for the block of code within:use indices to find the data we need, and likewise

find where to put results, do computation

n Map/fold for relevant data structures

n Operate on one data element at a time

nBetter still: factor out relevant controlstructures for your application

pixel

neighborhood

( = ; < ; ++)for i 0 i n i{

( =for j ( =for k

}

. ( ( )..Array map fun x. ( ( ) ..List fold fun y

. ( Matrix foreach fun)))

( = ; < ; ++) {for i 0 i n i

}

. ( ( ) )Array map fun x

(pixel_kernel fun( ) { ( , ) + ( , ) }, )nbrhd nbrhd 0 0 nbrhd 1 1 strm(pixel_kernel fun( , ) { }, , )nbrhd st init_state strm


19/22

+Case Study: backgroundsubtraction application

n Original code massive duplication (1300 lines of C++)

n Step 1: port verbatim

n Step 2: factor duplicated code, use HOF

n Step 3: remove unnecessary floating point

n Step 4: parameterize design, expose parallelism

n

(and further simplify in the process!)


20/22

+

Image

Parallelization


21/22

Evaluation

+Server side: Parallel Speedup

nWithoutstream graph opts 60 threads, lock free fifos

:W in s fo r th is a p p

&Typ e s h ig h e r o rd e r fu n ctio n s e n a b le a fu lly p o ly m o r


22/22

+ Conclusion: FP enhancesstreaming Contributions:

Reusable Libraries: Parallel matrix library

WaveScope: compiler and runtime

Contact: Ryan Newton - [email protected]

:// . . .h ttp w a ve scop e csailm it ed u

Documents

2008 ICFP DEFUN - Functional in the Field