fuzzy logic placement emily blem ece556 final project december 14, 2004 reference: e. kang, r.b....
DESCRIPTION
Methods(1) Fuzzy set: a group of objects with different levels of membership Objects may partially belong to a set Operations: 1 and 2 = max( 1, 2 ) 1 or 2 = min( 1, 2 ) (image from Kang et. al. 1994)TRANSCRIPT
Fuzzy Logic Placement
Emily BlemECE556 Final ProjectDecember 14, 2004
Reference: E. Kang, R.B. Lin, and E. Shragowitz. “Fuzzy Logic Approach to VLSI Placement.” IEEE Trans. On VLSI Systems. V. 2, No. 4. December 1994.
Objectives
Multiple placement objectives: timing, chip size, interconnection length, etc.
Need a framework in which to resolve multiple objectives
Not well addressed by most algorithms
Methods(1) Fuzzy set: a group of objects with different levels of
membership Objects may partially belong to a set Operations: 1 and 2 = max(1, 2)
1 or 2 = min(1, 2)
(image from Kang et. al. 1994)
Methods (2) Can be applied to iterative or constructive design In an iterative design, reduce number of criteria for
each objective to 1 or 2 due to time constraints Constructive algorithm:
Top level: place cells in feasible regions based on timing requirements
Middle level: assign cells to feasible intervals Bottom level: assign each cell to a position within its
feasible interval Assignment completed row by row based on fuzzy logic
decision maker (FZDM)
Methods(3) Small chip area rules:
If a candidate cell provides good utilization of existing feed through pins and a small # of rows is used for each net connected to it, then a small # of feed through cells will be added
If a candidate cell adds a small # of feed through cells and produces almost equal row length, then small chip area will be generated
For large designs, criterion 1/2/1 has a not so strong preference over criterion 1/2/2
In early stages of placement, criterion 1/2/1/1 has a strong preference over criterion 1/2/1/2
In middle stages of placement, criterion 1/2/1/2 has a mild preference over criterion 1/2/1/1
Results Ability to tune solution a key feature of fuzzy
placement In paper, (balanced) fuzzy placement consistently
outperformed TimberWolf6.1 and OASIS using same routers after placement
design name Fract Struct Biomed
placer TW6.1 Fuzzy TW6.1 Fuzzy TW6.1 Fuzzy
propagation delay (ns) 1.12 0.65 6.15 5.32 13.32 10.9
chip area (mm2) 0.53 0.53 6.89 6.80 51.34 51.40
(data from Kang et. al. 1994)
Conclusions FZDM avoids issues of greedy placer in constructive
placement In iterative placement, CPU time issues make FZDM
into a weighted cost function According to paper, achieves impressive results Fuzzy logic structure makes it easy to tune solution
for different goals and achieve multiple objectives