SafeChoice : A Novel Clustering Algorithm for Wirelength -Driven Placement

1. SafeChoice: A Novel Clustering Algorithm for Wirelength-Driven Placement Jackey Z. Yan Chris Chu Wai-Kei Mak Department of ECE Department of CS Iowa State Univ. National Tsing Hua Univ.

2. Modern VLSI Placement • Millions of standard cells (e.g., logic gates) • Hundreds of big macros (e.g., IP cores)

3. What is Clustering? • Form small objects into bigger clusters and send the clustered netlist to the placer • Essential part of modern placers • Cut down problem size • If clusters are formed correctly, it will guide the placer to generate better layout

4. Previous Clustering Algorithms

5. a b c Common Problem in Previous Work • Based on a heuristic assumption Cluster highly directly connected objects ? Minimize placement wirelength Indirect Connections • mislead placers to a low-quality solution

6. SafeChoice (SC) Overview • Specifically designed for wirelength-driven placement • Priority queue (PQ) based • Handling hyperedges directly • Main Features Guarantee • Safe Clustering • Safe Condition • Selective Enumeration • Smart Stopping Criterion • Best clusters for HPWL(SC vs. FC, BC & NC) • Best HPWL(SCPlace vs. all state-of-the-art placers)

7. If the optimal wirelength in the clustered netlist is equal to the optimal wirelength in the original netlist, then it is safe to cluster For ,if can be moved to the same location without increasing the wirelength,then it is safe to cluster Concept of Safe Clustering ignore overlap issue in the cluster

8. y x Assumptions: • ignore overlap issue Horizontally / Vertically Safe Horizontally Safe Safe • ignore fixed objects Vertically Safe

9. HPWL change of HPWL change of a is moving towards b b is moving towards a c d a b a c d b c a d b c d a b Wirelength Gradient Function increase decrease no change

10. a b no increase no increase increase Total Wirelength Gradient Function Find out the smaller total HPWL change between either moving a towards b or moving b towards a

11. It is safe to cluster a and b if a b c no increase Impractical Safe Condition for Vc={a,b}

12. Only the placement p with the maximum matters then would not affect the safe condition, and thus can be ignored in the enumeration Principle in Selective Enumeration

13. Enumeration Size: ignore left central right ignore the central interval a b fixing 3 categories of vertices Enumeration Size Cut Down use one position to substitute various positions in each interval

14. Flow of Selective Enumeration Yes would not consider clustering a and b No Enumerate Placements Yes Safeto cluster a and b No Unsafeto cluster a and b

15. SafeChoice Algorithm • Consider both safeness and area • Priority queue (PQ) based • Three operation modes

16. Experiments Setup • All experiments run on a Linux machine (Intel Xeon 2.83 GHz CPU and 32 GB RAM) • ISPD05/06 Placement Benchmarks (use scaled HPWL for ISPD06 circuits) • Two categories of experiments • Compare SafeChoice with clustering algorithms • Compare SCPlace with placement algorithms

17. Flow without clustering (flat-mPL6) Flat-mPL6 (GP+DP) FastDP Experimental Flow for Clustering • Compared with FirstChoice (FC), BestChoice (BC), NetCluster (NC) • mPL6 as placement engine Normalized Flow with clustering Clustering Flat-mPL6 (GP+DP) Unclustering Flat-mPL6 (DP) + FastDP

18. SC-R vs. FC, BC & NC Average Clustering Raito: 0.57

21. SC vs. Multilevel mPL6

22. SCPlace vs. RQL & mPL6 SCPlace: Two-level placement algorithm based on SafeChoice and flat-mPL6 (SCPlace will be described in the future journal paper)

23. SCPlace vs. Best Approach Results of other placers are cited from RQL paper [N.Viswanathan et.al, DAC 2007]

24. Conclusion • SafeChoice – a novel clustering algorithm for wirelength-driven placement • Safe Clustering • Safe Condition • Selective Enumeration • Smart Stopping Criterion • Promising experimental results • Best clusters for HPWL(SC VS. FC, BC & NC) • Best HPWL(SCPlace VS. all state-of-the-art placers)

25. Future Work • Derive the safe condition for • Develop our own placer based on SafeChoice • Integrate into other algorithms (e.g., partitioning) • SafeChoice source code is publicly available atwww.public.iastate.edu/~zijunyan/ Thank Thank You You

26. Backup

27. Enumeration Size: ignore left central right ignore the central interval a b fixing 3 categories of vertices Enumeration Size Cut Down use one position to substitute various positions in each interval

28. Enumeration Size: ignore ignore various position in each interval for left central right e e ignore positions in central interval a b fixing 3 categories of vertices Enumeration Size Cut Down

29. SC-G vs. FC & BC Average Clustering Raito: 0.70 Some unsafe clusters have better area than safe clusters

30. Clustering for Various Clustering Raito • Compare with FC and BC • Target clustering ratio

32. Level 1 Level 2 SC-Based Two-Level Placement Flow of SCPlace • Based on physical SafeChoice SafeChoice Flat-mPL6 (GP) Unclustering So Simple! Physical SafeChoice Incremental Flat-mPL6 Details of physical SafeChoice and SCPlace can be found in the future journal paper Unclustering FastDP