Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San...
-
date post
22-Dec-2015 -
Category
Documents
-
view
229 -
download
4
Transcript of Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San...
![Page 1: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/1.jpg)
Routing-Aware Scan Chain Ordering
Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA,
USA.),Stefanus Mantik
(Cadence Design Systems Inc., San Jose, CA, USA.) Email: { [email protected], [email protected],
Supported by MARCO Gigascale Research Center and Cadence Design Systems
![Page 2: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/2.jpg)
Outline Introduction and Previous Work Motivation for Routing Awareness Nature of the New TSP TSP Solvers Routing Aware Scan Chain Ordering Experiments Results and Conclusions
![Page 3: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/3.jpg)
Introduction
Scan chains commonly used to enhance testability. All flip-flops chained to form a shift register.
Minimizing wirelength overhead of scan increases routability and improves timing by reducing capacitive loading on nets that share register pins with the scan chain.
We give a new scan chain ordering method which gives upto 85% wirelength improvements over commercial tools and previously reported methods.
SI Q
FFA
SI Q
FFB
SI Q
FFCPI PO
![Page 4: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/4.jpg)
Previous Work Scan chain ordering has been
modeled as a Traveling Salesman Problem (TSP) [FeuerK83]
Previous works did placement-based ordering using cell-to-cell distance (AB) or pin-to-pin (A’B’) Manhattan distance as the TSP distance metrics
Modified 2-opt and 3-opt heuristics for the almost symmetric pin-to-pin TSP were given by [BoeseKT94, KobayashiEK99]
Q Q
SI SIA B
Q
SI SI
Q
A’
B’
Cell-to-cell distancefrom FFB to FFA
Pin-to-pin distance from FFB to FFA
![Page 5: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/5.jpg)
Outline Introduction and Previous Work Motivation for Routing Awareness Nature of the New TSP TSP Solvers Routing Aware Scan Chain Ordering Experiments Results and Conclusions
![Page 6: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/6.jpg)
Motivation for Routing Awareness
Scan chain TSP costs should be based on wirelength estimate for the scan connection.
A FF output pin will have a fanout routing tree.
True routing distance to connect QB to SIA will be A”B”.
A routing aware scan chain ordering is likely to be different than one based on placement.
Q Q
SI SI
A” B”
![Page 7: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/7.jpg)
Outline Introduction and Previous Work Motivation for Routing Awareness Nature of the New TSP TSP Solvers Routing Aware Scan Chain Ordering Experiments Results and Conclusions
![Page 8: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/8.jpg)
Nature of the New TSP
Cost of connections: • QA to SIB: w = AB
• QB to SIC: x = BC
• QA to SIC: z = AC
• QB to SIA: y = BA
Note that
• AB BA
• AB + BC < AC
Q
SI
Q
SI
Q
SI FFA FFB
FFC
wx
yz
![Page 9: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/9.jpg)
Nature of the New TSP Asymmetry
E.g. AB BA Cell-to-cell distance metric was completely
symmetric while pin-to-pin metric was almost symmetric
Non-metricity Triangle inequality is not obeyed. E.g. AB + BC <
AC Cell-to-cell metric was metric while pin-to-pin
metric was almost metric The new TSP formulation can be highly asymmetric
and very non-metric
![Page 10: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/10.jpg)
Outline Introduction and Previous Work Motivation for Routing Awareness Nature of the New TSP TSP Solvers Routing Aware Scan Chain Ordering Experiments Results and Conclusions
![Page 11: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/11.jpg)
TSP Solvers Due to large asymmetry and non-metricity of the
TSP instance, standard symmetric TSP solvers do not give good results
Eighth DIMACS implementation challenge for ATSP ended in 2002 [JohnsonGM02] Iterated Lin-Kernighan based LKH [Helsgaun00]
was reported to give best tours We use iterated ScanOpt from the GSRC Bookshelf
which has results comparable to LKH-1.2 and is tailored to scan chain optimization
![Page 12: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/12.jpg)
TSP Solvers: ScanOpt
Large step Markov chain (LSMC) methods for solving TSP alternately apply a local optimization procedure Descent followed by a “kick move” which perturbs the local minimum to obtain the starting solution for the next Descent application
ScanOpt is a LSMC implementation based on the restricted 2,3-opt moves of [BoeseKT94] for solving an ATSP
Test Case
Tour Cost (m)
Run Time (sec.)
ScanOpt
LKH ScanOpt
LKH
A (pin-to-pin)A (pin-to-net)
216099297
206327511
14412149
56702717
![Page 13: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/13.jpg)
Outline Introduction and Previous Work Motivation for Routing Awareness Nature of the New TSP TSP Solvers Routing Aware Scan Chain Ordering Experiments Results and Conclusions
![Page 14: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/14.jpg)
Routing Aware Scan Chain Ordering
Incremental routing cost based on existing or anticipated routing
Considers both Q and Q’ outputs for the minimum wirelength connection
Driven by global routing or trial detailed routing
We calculate the scan connection cost from the routed segments in the detailed routed DEF netlist
Q’
FFA
SI FFB
Q
d(Q,SI)
d(Q’,SI)
![Page 15: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/15.jpg)
Outline Introduction and Previous Work Motivation for Routing Awareness Nature of the New TSP TSP Solvers Routing Aware Scan Chain Ordering Experiments Results and Conclusions
![Page 16: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/16.jpg)
Experiments: Tools Placement: Cadence Qplace v5.1.68 Detailed Routing: Cadence Wroute v2.2.31 Our TSP Solver: ScanOpt Industry standard scan chain ordering:
QPlace or Cadence Silicon Ensemble v5.3.125 (SE)
We do not use incremental routing due to poor results of running WRoute in incremental or ECO mode: an observation confirmed in [KahngM00]
![Page 17: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/17.jpg)
Experiments: Commercial Scan Chain Ordering To confirm absence of routing
awareness in SE or QPlace scan stitching, we use the tools to order scan chains before and after detailed routing
We then extract the scan orders from the routed DEFs
Pre and post route scan chain orderings by the tools is exactly the same. Hence we infer that these tools (and to best of our knowledge, all others) do not use any routing information
![Page 18: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/18.jpg)
Experiments: Flows1. The baseline place&route flow w/o scan
insertion2. Placement based scan chain ordering by SE3. Placement based ordering by QPlace4. Placement based ordering by ScanOpt5. Routing driven scan chain ordering using
ScanOpt Timing-driven and non-timing-driven versions
of all place and route flows
![Page 19: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/19.jpg)
Experiments: Routing Driven Flow
1. Trial route placed DEF netlist2. Construct the ATSP cost matrix by
computing pairwise minimum pin-to-net distances
3. Find the TSP tour using ScanOpt4. Input the scan chain order into the placed
DEF using the ORDERED construct of DEF5. Attach scan nets6. Final route the placed DEF with scan nets
![Page 20: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/20.jpg)
Experiments: Testcases
Industry LEF/DEF testcases Xswap derived from X by random swapping of FF
placements Xexpand obtained from X by expanding the site-map by
20%
Test Case
No. ofCells
No. of Scan FFs
#ScanChains
Die Areamm2
# Metal Layers
A/Aswap
Aexpand
B/Bswap
Bexpand
C/Cswap
Cexpand
6390639040350403503423534235
122612261975197545504550
22111010
0.5260.6326.8758.3733.8465.611
444444
![Page 21: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/21.jpg)
Outline Introduction and Previous Work Motivation for Routing Awareness Nature of the New TSP TSP Solvers Routing Aware Scan Chain Ordering Experiments Results and Conclusions
![Page 22: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/22.jpg)
Results: Distance Metrics
Asymmetry measure = ( average(|d(i,j) – d(j,i)|) )/( average(|d(i,j) + d(j,i)|) )
Metricity measure = average( (d(i,j) -min(d(i,j),min(d(i,k)+d(k,j))) /d(i,j) )
Asymmetry and non-metricity for cell-to-cell distances is zero
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
A B C
Testcase
Asy
mm
etry
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
A B C
Testcase
Non-M
etr
icit
y
Pin-to-pinPin-to-net
![Page 23: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/23.jpg)
Results: Wirelength
Flow VII (routing driven) and Flow IV (placement driven) both use ScanOpt
Routing driven ordering consistently gives much smaller scan wirelength than any of the industry flows
QPlace ordering is better than SE ordering
0
1
2
3
4
5
6
A Aswap Aexpand B Bswap Bexpand C Cswap CexpandTestcase
% S
can O
verh
ead
SE-place
QP-place
ScanOpt-place
Routing Driven
![Page 24: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/24.jpg)
Results: Timing
We measure quality of timing by minimum slack and the number of timing violations
Our aim is wirelength reduction Timing is not worse than the other flows
0
5
10
15
20
25
30
C Cswap Cexpand
Testcase
Min
imum
Sla
ck
010002000300040005000600070008000
C Cswap Cexpand
Testcase
No. of Vio
lati
ons
SE-place
QP-place
ScanOpt-place
Routing Driven
![Page 25: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/25.jpg)
Results: Runtime
Router runtimes normalized to 143MHz Sun Ultra-I are reported CPU time for Flow VII is sum of initial trial and final router runs For Flow VII routing is done from scratch to route the scan
chain. No incremental routing is used
0
2000
4000
6000
8000
10000
12000
14000
16000
A Aswap Aexpand B Bswap Bexpand C Cswap Cexpand
Testcase
Route
r R
unti
me (
sec.
)
Baseline
SE-place
QP-place
ScanOpt-place
Routing Driven
![Page 26: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/26.jpg)
Conclusions A substantial reduction in wirelength (20%-85%)
impact of scan is achieved by routing aware scan chain ordering
Despite being timing oblivious, routing-aware flow does not significantly worsen the timing
Runtime overheads of routing awareness can be reduced substantially if industry routers are able to deal better with incremental optimizations
Timing aware extensions are possible but rely on controllability of the router. See the ISQED’03 publication “A Proposal for Routing-Based Timing-Driven Scan Chain Ordering”
![Page 27: Routing-Aware Scan Chain Ordering Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA, USA.), Stefanus Mantik (Cadence Design.](https://reader034.fdocuments.in/reader034/viewer/2022050802/56649d7f5503460f94a63803/html5/thumbnails/27.jpg)
Routing-Aware Scan Chain Ordering
Puneet Gupta and Andrew B. Kahng (Univ. of California at San Diego, La Jolla, CA,
USA.),Stefanus Mantik
(Cadence Design Systems Inc., San Jose, CA, USA.) Email: { [email protected], [email protected],
Supported by MARCO Gigascale Research Center and Cadence Design Systems