A Computing Origami: Folding Streams in FPGAs
S. M. FarhadPhD Student
University of Sydney
DAC 2009, California, USA
2
Outline
Motivation Stream programming FPGA Problem
Stream Folding Results Conclusion
2
Stream Programming Paradigm
Programs expressed as stream graphs Streams: Sequence of data elements Actor: Functions applied to streams
Independent actors with explicit communication
Regular and repeating computation
3
Actor/Filter
Streams
Streams
FPGA
FPGAs are widely available as programmable coprocessors
Opportunities to exploit FPGA-based acceleration Multimedia, networking, graphics, and security codes
4
Problem
Maximizing throughput subject to Area and latency constraints
Resolving bottleneck actors The replicated filters do not require resynthesis
5
Motivating Example
6
Motivating Example
7
Motivating Example
8
9
Outline
Motivation Stream programming FPGA Problem
Stream Folding Results Conclusion
9
Area/Throughput Design Folding
1 foreach Filter f in S do2 workFactor[f] = f.latency.S.runs(f);3 designPointArea + = f.area.workFactor[f];4 scaleLimit = minf.hasState (1/workFactor[f]); 5 scaling = min(AREA/designPointArea, scaleLimit);6 foreach Filter f in S do7 replication[f] = workFactor[f].scaling;8 while area(replication) > AREA do9 replication = reduceThroughput(replication);
10
Calculating Throughput
11
)(
)()(
i
iout Flatency
Fpushit
)(
)().()(
j
j
njiout
Pout Flatency
Fpushitit
)(min1
itt Pout
ni
Pout
i
njj
outni
SJout w
w
itt 1).(min
))(..(min
))(.(min)(min
itCrt
itCitt
outiiSi
Sout
outiSi
Sout
Si
Sout
Calculating Latency
FPGAs that are coupled to host processors Initiation interval (DMA) Replication improves throughput, it often
increases the latency! Major factors for latency variation
Non-periodic data arrival Data-token reordering Local congestion
12
Latency constrained design folding
1 latConf= null ; T = ∞;2 while throughput(thrConf) ≤ T do3 if feasibleImprovement(thrConf) then4 candidates = simAnnealing(thrConf, T);5 foreach candidate in candidates do6 if throughput(candidate) < T then7 latConf = candidate;8 T = throughput(latConf);9 thrConf = reduceThroughput(thrConf);10 return latConf
13
Results
Benchmark
Minimum area Best throughput Constrained design
LUTs Latency II LUTs Latency II LUTs Latency IIConstraint
Run time
MatrixMult 1498 480 19 7618 185 3 4558 175 7
Latency ≤ 175 1.14s
Serpent 3028 1027 4 3878 773 2 3053 901 4Latency ≤ 910 0.73s
FFT2 37610 1199 3 43370 764 2 39530 868 7AREA ≤ 40000 34.7s
FMRadio 37458 371 39 87564 371 13 62511 371 20AREA ≤ 65000 1.01s
DCT 45752 349 3 137256 349 1 91504 349 2AREA ≤ 120000 0.73s
BitonicSort 43920 1042 3 131760 1042 1 47400 1282 2
AREA ≤ 50000 18.3s
Synthetic 350 309 135 15990 504 2 1490 309 47
AREA ≤ 1500 0.43s
14
Questions?
