1. 12/4/2018
A Regularity-Driven Fast Gridless Detailed Router for High Frequency Datapath Designs By
Sabyasachi Das (Intel Corporation)
Sunil P. Khatri (Univ. of Colorado, Boulder)

2. Outline Motivation Proposed approach Extracting Net-Clusters
12/4/2018
Outline
Motivation
Proposed approach
Extracting Net-Clusters
Selecting a representative bit-slice
Routing same-bit and cross-bit nets
Propagating the routes
Advantages of our approach
Experimental results
Conclusions & Future Work

3. Datapath Designs
Datapath Designs are characterized by unique regularity present across bit-slices
One of the most critical parts in any circuit
Most commonly found in Microprocessor, DSP, graphic ICs.
Effective use of this regularity is crucial for efficient design of datapath
Datapath placement techniques are available
Not many results are available on datapath routing

4. Regularity in Datapath Circuits
An important task is to efficiently route all the nets by using the regularity inherent in datapath circuits

5. Overall Routing Flow

6. Net-Cluster Extraction
Net-cluster is a collection of nets (spread over multiple bit-slices), in which all nets have similar connections.
Different algorithms to extract net-clusters:
Footprint-driven clustering
Instance-driven clustering
Cluster-merging

7. Footprint-driven Clustering
Footprints are of two types:
Global footprint contains the pin-names and master-cells (of connecting instances) names.
Detailed footprint contains the instance names and the net-name.
This clustering technique has two steps:
Groups are created from global footprint information.
By using Detailed footprint in each Group, net-clusters are obtained.

8. Footprint-driven Clustering
After Global Footprint, two Groups are obtained:
G1 = {LB[3:0], SB[3:0]}
G2 = {SM[3:0]}
After Detailed Footprint, three net-clusters are obtained:
From G1:
NC1 = {LB[3:0]}
NC2 = {SB[3:0]}
From G2:
NC3 = {SM[3:0]}

9. Instance-driven Clustering
Uses the location of connecting pins (does not need the uniform naming scheme).
After instance-clustering, two net-clusters are obtained (from same Group):
NC1 = {AB, CD, EF, GH}
NC2 = {KL, MN, RS, TV}

10. Cluster Merging Multiple non-full clusters are merged.
FDC found 3 clusters:
NC1 = {LB[3:0]}
NC2 = {SM[1:0]}
NC3 = {SB[3:0]}
IDC found 1 cluster:
NC4 = {ABC, DEF}
Cluster merging technique merged NC2 and NC4
NC-New = {ABC, DEF, SM[1], SM[0]}

11. Selecting Representative Bit-Slice
We conceptually extend datapath to have an infinite number of bit-slices.
Then, any slice can be chosen as representative bit-slice for routing.
While routing, we need to take care of:
Same-bit nets
Backward cross-bit nets
Forward cross-bit nets

12. Routing Nets in Representative Bit-Slice
Our routing approach is a combination of pattern-based routing and maze routing
We call our routing as strap-based routing, where a strap is defined as a straight segment, which can be vertical or horizontal.
Our router uses minimal memory, because it loads the design data of only representative bit-slice.

13. Routing Nets in Representative Bit-Slice
The router routes nets in a sequential manner:
The nets connected to topmost (Y-wise) pins are routed first.
Long nets are given preference.
Router tries to find direct-routes (only vertical or only horizontal strap or VTH/HTV strap) for nets.
In case of conflict, rip-up and re-route technique is used.
If strap-based router cannot find a route, maze router is used.

14. Routing Cross-Bit Nets
We model a single cross-bit net (spread over multiple bit-slices), as a combination of multiple smaller sub-nets, each confined within a single bit-slice.
We virtually instantiate all those sub-nets into the representative bit-slice.
If maximum backward/forward connectivity is k, then we need to find k different routes (usually, value of k is not very high).

15. Propagating the Routes
After routing nets in representative bit-slice, similar routes are propagated to other nets in same net-cluster.
Following two cases are handled separately:
For same-bit nets: Routes are propagated identically
For cross-bit nets: If maximum degree of forward/backward connectivity is p, we find p different routes for that net & propagate them accordingly to the correct bit-slice’s net.

16. Advantages of Our Approach
Speed of Routing: Only a portion of nets are routed, making it really faster.
Easy Incremental Routing: Rip-up all the routes for a given net-cluster and again route them identically. This can be done multiple times.
Predictable Routes: Wiring parasitics are similar for different nets in a net-cluster. So, timing estimation is also similar.
Better Debuggability & Timing convergence: Easy to find and fix poorly routed nets.

17. Experimental Results (Circuits Used)
# Instance
# Connections
# Bits
Industry-1
1056
5504 32
Industry-2
2368
12672
Industry-3
4672
29184 64
Industry-4
6208
48128
These four industrial datapath designs are used

18. Experimental Results (run-time)
Circuit
Ind. Router
Our Router
Ratio
Industry-1 70 12
0.17
Industry-2
145 26
0.18
Industry-3
264 36
0.14
Industry-4
394 42
0.11
Average
0.15
Our router is at least 7X faster.

19. Experimental Results (wire-length)
Circuit
Ind. Router
Our Router
% Improvement
Industry-1
47458
46790
1.41
Industry-2
97453
99476
-2.07
Industry-3
276456
282569
-2.21
Industry-4
589679
564568
4.25
Average
0.35
Average gain of our method is minimal.

20. Experimental Results (via-count)
Circuit
Ind. Router
Our Router
% Improvement
Industry-1
6856
5678
17.18
Industry-2
25876
22336
13.68
Industry-3
39568
39452
0.28
Industry-4
44568
42976
3.57
Average
8.68
Less number of vias are used because of strap-based nature of our router.

21. Conclusions & Future Work
Regular routing technique is very useful for fast automated datapath design.
In future, we plan to focus on
Crosstalk issues in datapath routing
Handling irregular connectivity