Case Study

structured

sequential program

Analysis tools andmethods

partitiontimingdata.

Performancepredictor

Performanceanalyser

event-trace file

Monitored

configuration

parallelisedcode

sequentialcode sections

High-level softwaretemplates and configurationtool

prototypeimplementation.convolvewith22Gabor filters

in 5 x 5 x 21neighbourhood

demodulate& apply4 point central differencing todifferentiateresponse

across all filters

form phasethroughdifferentiatedresponse andthresholdacross allfilters

calculatecomponent (phase-normal)velocities forunthresholdedphasesacrossall filters

minimumleast squarescalculationbased on normalconstraintacross 6 estimates

thresholdfull velocitiesthroughconditioningnumber

robustextension

temporal filtering

spatial filteringcore algorithmconfidence testing

multi-resolutionextension

propagation

Pipelineon targetparallelmachine

Mapping 1

Prototypeparallelpipeline

Mapping 2

DSPstageFPGADSP stage

target frames are placed in message buffer

t

correct output order isestablished by linked-listintervening storage.

2

size of separablefilter depends on

x

s.d. of Gaussian.

x

middle targetframe.

5 x 5 x 5 separable filter

t

y1043access orderdeterminedby tag array

zeroised-borderregion.

y

target frame forspatiotemporal filter

circular time buffer

Stage One

Gaussian (normal) smoothingfilter.

Stage Two

Derivative Filtering

Spatial measurement matrix

Use eigenvector test asconfidence measure

derivative images

Combine to formvelocity vector

strip of local neighbourhoodstyx

Temporal measurement matrix

Stage Three

Form least (or total) squares measurement matrices.

Stage Four

Eigenvector and velocity vector

calculations.

Input message

2

border regionmiddle frame

5neighbourhood strip

t

zeroisedoffset

2zeroisedoffset

border region

extract derivatives by means ofweighted difference kernel

Intermediate strips

tyx5

target lineuse 5 x 5 neighbourhoods to find(x, y) velocity vectors for target line.

Output message

full velocitypartial velocityraw velocityresidual

nullified offsetnullified offset

Begin processing

Receive rowfrom farm 1

Correct

order ?

yes

no

Store in linked-list data structure

no

Wait for worker

Outputbuffer full ?no

yes

request and outputbuffer.

yes

Outputbuffer full ?

Unpack from

no

Stored messages ?

yes

linked-list and free list node.