eecs 219b spring 2001
DESCRIPTION
EECS 219B Spring 2001. Timing Optimization Andreas Kuehlmann. Restructuring for Timing Optimization. Outline: Definitions and problem statement Overview of techniques (motivated by adders) Tree height reduction (THR) Generalized bypass transform (GBX) Generalized select transform (GST) - PowerPoint PPT PresentationTRANSCRIPT
1
EECS 219BEECS 219BSpring 2001Spring 2001
Timing Optimization
Andreas Kuehlmann
2
Restructuring for Timing OptimizationRestructuring for Timing Optimization
Outline:• Definitions and problem statement• Overview of techniques (motivated by adders)
– Tree height reduction (THR)– Generalized bypass transform (GBX)– Generalized select transform (GST)– Partial collapsing
3
Circuit RestructuringCircuit Restructuring
Approaches:
Local: • Mimic optimization techniques in adders
– Carry lookahead (THR tree height reduction)– Conditional sum (GST transformation)– Carry bypass (GBX transformation)
Global:• Reduce depth of entire circuit
– Partial collapsing– Boolean simplification
4
Tree-Height Reduction (THR)Tree-Height Reduction (THR)
n
l m
i j
h
k
3
6
5 5
1 4
1
0 0 0 0 2 0 0
a b c d e f g
i
1
0 0
a b
m
j
h
k
3
41
0 0 2 0 0
c d e f g
n’Duplicatedlogic
12
00
5Criticalregion
CollapsedCritical region
Singh’88:
5
Tree-Height ReductionTree-Height Reduction
i
1
0 0
a b
m
j
h
k
3
41
0 0 2 0 0
c d e f g
n’Duplicatedlogic
12
00
5
i
1
0 0
a b
m
j
h
k
3
41
0 0 2 0 0
c d e f g
12
0
35
n’
2
1
0
4
CollapsedCritical region
New delay = 5
6
Generalized bypass transform (GBX)Generalized bypass transform (GBX)
• Make critical path false– Speed up the circuit
• Bypass logic of critical path(s)
fm=f fm+1 fn=g…
fm =f fm+1 fn=g… 0
1
g’
dg__df
Booleandifference
s-a-0 redundant
McGeer’91:
7
GBX and KMS transformGBX and KMS transform
GBX gives little area increase, BUT creates an untestable fault
(on control input to multiplexer)
KMS transform: (remove false paths without increasing delay)
• fk is last node on false path that fans out.
• Duplicate false path {f1,…, fk} -> {f’1, … , f’k}
• f’j fans out to every fanout of fj except fj+1, and fj just fans out to fj+1
• Set f0 input to f1 to controlling value and propagate constant (can do because path is false and does not fanout)
KMS results
– Function of every node, except f1, … ,fk is unchanged
– Added k nodes– Area added in linear in size of length of false paths; in practice small area increase.
8
KMSKMS
fm+1 fm+2 fn…fm+k fm+k+1
f’m+1 f’m+2 f’m+k
fm+1 fm+2 fn…fm+k fm+k+10
…Delay is notincreased
Keutzer, Malik, Saldanha’90:
9
Generalized select transform (GST)Generalized select transform (GST)
Berman’90: Late signal feeds multiplexor
c d e f g
a
b
out
c d e f g
b
c d e f g
b
a=0
a=1
out0
1
a
10
GST vs GBXGST vs GBX
… 0
1g’
dh__da
a
a
b
cg
h
c d e f gb
c d e f gb
a=0
a=1
out0
1
a
GST
c d e f gb
c d e f gb
a=0
a=1
… 0
1g’
a cg
b
GBX
a
h
Boolean
diffe
Note:
rence =
a a a
hh h
GBX
11
GST vs GBXGST vs GBX
• Select transform appears to be more area efficient• But Boolean difference generally more efficiently formed in
practice• No delay/speedup advantage for either transform• Can reuse parts of the critical paths for multiple fanouts on GST
c d e f gb
c d e f gb
a=0
a=1
out10
1
a
GST out20
1
a
12
Technology Independent Delay ReductionsTechnology Independent Delay Reductions
Generally THR, GBX, GST (critical path based methods) work OK, – but very greedy and computationally expensive
Why are technology independent delay reductions hard?
Lack of fast and accurate delay models– # levels, fast but crude– # levels + correction term (fanout, wires,… ): a little better,
but still crude (what coefficients to use?)– Technology mapped: reasonable, but very slow– Place and route: better but extremely slow– Silicon: best, but infeasibly slow (except for FPGAs)
better
slower
13
Clustering/partial-collapseClustering/partial-collapse
Traditional critical-path based methods require– Well defined critical path– Good delay/slack information
Problems:– Good delay information comes from mapper and layout– Delay estimates and models are weak
Possible solutions:– Better delay modeling at technology independent level– Make speedup, insensitive to actual critical paths and
mapped delays
14
Clustering/Partial CollapseClustering/Partial Collapse
Two-level circuits are fast– Collapse circuit to 2-level - but
• Huge area penalty• Huge capacitive loading on inputs (can be much slower)
To avoid huge area penalty– Identify clusters of nodes
• Each cluster has some fixed size– Perform collapse of each cluster– Simplify each node
Details– How to choose the clusters?– How to choose cluster size?– How to simplify each node?
15
Lawler’s Clustering AlgorithmLawler’s Clustering Algorithm
• Optimal in delay:– For a given clustering size
• May duplicate nodes (hence possible area penalty)– Not optimal w.r.t duplication– Use a heuristic
• Fast: O(m x k)– m = number of edges in network– k = maximum cluster size
16
Clustering Algorithm - OverviewClustering Algorithm - Overview
• Label phase: (k is cluster size)– If node u is an input, label(u) := L := 0
• Else L := max label of fanin of u– If (# nodes in TFI(u) with (label = L) >= k)
label(u) := L+1• Cluster phase: (outputs to inputs)
– If node u is an output, L := infinity• Else L := max label of fanouts of u
– If (label(u) < L) then create a new cluster with “root” u and with members all the nodes in TFI(u) with label = label(u)
• Collapse phase:– Collapse all nodes in a cluster into a single node– Note: a node may be in several clusters (causes area increase
17
Example of ClusteringExample of Clustering
0
0
0
0 0
0
1
1
1
1
2
01
12
0
0
Result: Lawler’s algorithmgives minimum depth circuit
Typically, 1. we decompose initial circuit
into 2-input NANDs and invertors.
2. then cluster size k reflects # 2-input NANDs to be collapsed together.
k = 3
18
Choosing Choosing kk
• I(k): number of levels, given k
• d(k): duplication ratio
– Number of gates in cluster network divided by number of gates in original network
• Determine k0 where k0/d(k0)~2.0
• For every k from 2 to k0, compute d(k), I(k)
– Use exhaustive enumeration: label and cluster (without collapse) for each k.
– Each iteration is O(|E|k)
• Choose k such that
– I(k) is minimized
• Break ties using d(k)
– Minimize d(k)
d(k)
I(k)
1 2 k0
19
Area RecoveryArea Recovery
Area increase is due to node duplication – this occurs when node is in multiple clusters
Two solutions:– Break clusters into smaller pieces off critical path– After cluster and collapse, recover area
20
Relabeling Procedure:Relabeling Procedure:
Attempt to increase node labels without exceeding cluster size
In reverse topological order
Start : assign
Increase label(u) if• new-label(u) <= label(v) for each fanout v and• new-label(u) = new-label(v) for each fanout v only if
label(u) = label(v) before relabeling, and• no cluster size is violated
- ( ) max ( )i jj PO
new label O label O
21
Relabeling ExampleRelabeling Example
00
00
00
00 00
00
11
11
11
22
22
00
00
00
00 00
00
11
11
11
11
22
before
after
22
Post-Collapse Area RecoveryPost-Collapse Area Recovery
• Algebraic factorization, but– Undo factorization if depth increases
• Optimization using DC’s• Redundancy removal
23
Conclusions Conclusions
• Variety of methods for delay optimization– No single technique dominates – When applied to ripple-carry adder get– Carry-lookahead adder (THR)– Carry-bypass adder (GBX)– Carry-select adder (GST)– Clustering/Partial collapse
• All techniques ignore false paths when assessing the delay and critical regions– Can use KMS transform to eliminate false paths without
increasing delay (area increase however).