1. SEMI-SYNTHETIC CIRCUIT GENERATION FOR TESTING INCREMENTAL PLACE AND ROUTE TOOLS David GrantGuy Lemieux University of British Columbia Vancouver, BC

2. Overview  Introduction Circuit Generation Method 1 (FPL) doesn’t work well Circuit Generation Method 2 (FPT) simple, works very well  Circuit Scaling Extension Conclusions 2

3. Introduction Problem  (Free) Large digital circuits are rare  FPGA CAD tools need large circuits Solution  Generate random circuits  But is random realistic? Goal  Generate realistic random circuits (synthetic)  Useful for testing place and route tools 3

4. Introduction Past approaches  Generate a complete synthetic circuit  Tools: ccirc+cgen, gnl New problem  Real world development is iterative, incremental  FPGA tools commonly used in incremental mode  Need circuits with incremental changes 4

5. Introduction Our Approach Start with a real circuit Generate a small synthetic change Result is a semi-synthetic circuit Four steps to generate a semi-synthetic circuit 1. Identify 2. Remove 3. Generate 4. Stitch (difficult?) 5

6. Introduction Problem: Blindly stitching a circuit can create combinational loops Cannot arbitrarily insert new latches to break loops 6

7. Overview Introduction  Circuit Generation Method 1 (FPL) Circuit Generation Method 2 (FPT)  Circuit Scaling Conclusions 7

8. Method 1: Introduction Four steps to generate a semi-synthetic circuit 1. Identify sub-circuit (T-VPack) 2. Remove sub-circuit 3. Generate clone (ccirc+gnl, ccirc+cgen) 4. Stitch Steps 1, 2, 3 are easy… Stitching is difficult !! 8

9. Method 1: Loop Graph, L Ideal Stitching  Determine which outputs connect back to inputs through combinational logic  If synthetic circuit generators could use L, we could stop here 1 2 3 4 5 1 2 3 4 5 Graph L 9

10. Method 1: Stitching Real Stitching takes 4 sub-steps… a) Generate a dependence graph, D –Calculated precisely from synthetic clone (subcircuit) b) Generate a permissible graph, P –Calculated imprecisely from loop graph of host circuit c) Solve the monomorphism problem –Find a monomorphism mapping D into P d) Stitch 10

11. Method 1: a) Dependence Graph, D D is computed from synthetic clone Captures all combinational paths through clone 1 2 3 4 5 1 2 3 4 5 Graph D 11

12. Method 1: b) Permissible Graph, P P computed from L using heuristic Problem is NP-complete in general  Start with loop graph L, consisting of only back-edges  Add all forward edges to L, creating cycles  call this P  Find forward edge in cycle, remove it, repeat until P is cycle-free  Remove all back-edges from P 1 2 3 4 5 Graph P 1 2 3 4 5 12 1 2 3 4 5 Graph L

13. Method 1: c) Monomorphism Monomorphism is like Isomorphism  Except that number of edges need not be identical Find 1:1 vertex mapping of D to P  D: forward combinational paths in synthetic clone  P: permissible forward combinational paths 1 2 3 4 5 Graph D 1 2 3 4 5 Graph P (1,2) (2,1) (3,3) (4,4) (5,5) Mapping 13

14. Method 1: d) Stitch Take the mapping solution and merge the clone into the hole left in the original circuit Mapping: (2,1), (1,2), (3,3), (4,4), (5,5) 14

15. Method 1: Discussion Good results for small cutout regions Unacceptably long run times for subcircuits > 50 I/Os Monomorphism solver is non-deterministic Preserved most key circuit characteristics  Logic depth increases by factor 2-3x …because method is unconstrained ! Conclusion  Stitching is not a trivial problem to solve (in a cycle- free way) 15

16. Overview Introduction Circuit Generation Method 1 (FPL)  Circuit Generation Method 2 (FPT)  Circuit Scaling Conclusions 16

17. Method 2: Introduction Four steps to generate a semi-synthetic circuit 1. Identify sub-circuit (T-VPack) 2. Remove sub-circuit 3. Generate clone (perturber) 4. Stitch The perturber does not “profile and generate”  Instead, it “perturbs” the existing circuit Stitching is trivialized using this method 17

18. Method 2: Perturbing a Circuit Perturber Algorithm a) Levelize the complete circuit b) Select a random edge in the sub-circuit (1 source and 1 sink) c) Select a second edge in sub-circuit under certain constraints * d) Swap the sinks e) Repeat 18

19. Method 2: Perturbing a Circuit Perturber Algorithm (cont'd) * Constraints guiding second edge selection 1. Source and sink levels must match 2. Source node cannot be sub-circuit input 19

20. Method 2: Advantages of Perturber Algorithm Key circuit characteristics are exactly preserved  Number of nodes and edges  Fanout distribution  Depth profile Strengths  No combinational loops are created Because the levelization is preserved  Very fast execution time  Very simple approach 20

21. Method 2: Initial Results Good results for small cutouts Less-than-impressive results for large cut-outs Cause  Locality is lost during perturbation Independent buses are cross-connected Regular features of the circuit are lost Connections swapped across large regions of chip  Need locality control !! 21

22. Method 2: Ancestor Control Method to preserve locality Add additional edge selection constraint 1. Source and sink levels must match 2. Source node may not be sub-circuit input 3. Both edges must share a common ancestor through combinational logic within a certain ancestor depth Stop at sub-circuit inputs and flops 22

23. Method 2: Testing Know all the key circuit characteristics are preserved  Focus comparisons on post-routing results Test 20 largest MCNC circuits  Metrics: channel width, delay, and wirelength  goodness of result == closeness to original MCNC result Test 1: Sanity test, full-circuit compare w/ ccirc+cgen Test 2: Incremental semi-synthetic circuits 23

24. Method 2: Sanity Check Results Channel Width of complete synthetic circuit Overall cgen: 20% error perturber: 14% error 24

25. Method 2: Sanity Check Results Delay of complete synthetic circuit Overall cgen: 7% error perturber: 9% error 25

26. Method 2: Sanity Check Results Wirelength for a complete synthetic circuit Overall cgen: 24 % error perturber: 17 % error 26

27. Method 2: Testing Test 2: Incremental Circuit Results  Generate semi-synthetic incremental circuit  Change only 5%, 10%, 20% of real circuit  No previous work to compare against 27

28. Method 2: Incremental Results Channel Width for various cutouts Overall 5%: 2.6 % error 10%: 3.2 % error 20% 6.7 % error 28

29. Method 2: Incremental Results Delay for various cutouts Overall 5%: 5.5 % error 10%: 6.4 % error 20% 8.7 % error 29

30. Method 2: Incremental Results Wirelength for various cutouts Overall 5%: 1.4 % error 10%: 2.2 % error 20% 3.9 % error 30

31. Method 2: Discussion Sanity Check  Complete circuit result error < ccirc+cgen error (error reduced by ~1/3) Incremental Circuits  Close to original (1%-6%) using 5%, 10% cutouts  More deviation (4%-9%) at 20% cutout 31

32. Overview Introduction Circuit Generation Method 1 (FPL) Circuit Generation Method 2 (FPT)  Circuit Scaling Conclusions 32

33. Circuit Scaling Scaling  Reduce the size of the cutout region so the tools have to fill in holes  Increase the size of the cutout region so the tools have to make room for larger circuit Approach  Scale a circuit (mutator), then perturb it 33

34. Circuit Scaling Reduction  Shotgun approach: delete nodes at random Sometimes need to delete chains of logic Enlargement  Duplicate circuit in parallel  Multiplex inputs and outputs with LUTs  Doubles (…, triples, etc) circuit Enlargement+reduction to achieve arbitrary scaling 34

35. Circuit Scaling: Results Test 5%, 10%, 20% cutout Scale cutout size to 50%, 75%, 200%, 400% original No unexpected changes in post-routing characteristics from MCNC original Some expected changes  Eg, wirelength increased as cutout size enlarged 35

36. Conclusions Have shown how to create a semi-synthetic circuit Method 2 is a superior method over past approaches  Runtime, simplicity  Preservation of key circuit properties Preserve them first, find ways to alter them later…  Quality of result Scaling is able to change the circuit size without changing the post-routing characteristics in unexpected ways 36

37. Future Work Critical Path  Lengthen the critical path to force the tools to shuffle nodes along the critical path Mutate other circuit properties?  Node depth profile  Fanout distribution  Fanin distribution  Wire lengths 37

38. It's Over Questions? Comments? Concerns? 38