layout symmetry annotation for analog circuits with graph
TRANSCRIPT
Layout Symmetry Annotation for Analog Circuits with Graph Neural Networks
Xiaohan Gao1, Chenhui Deng2, Mingjie Liu3, Zhiru Zhang2, David Z. Pan3, Yibo Lin1
1Department of Computer Science, Peking University
2ECE Department, Cornell University
3ECE Department, The University of Texas at Austin
Outline2
Introduction and Problem Formulation
SymDetect: Our Proposed Framework
Experimental Results
Conclusions and Future Work
Analog Layout Automation3
Analog/mixed signal IC design still relies heavily on manual design
Automatic constraint annotation is a critical step to existing analog layout automation tools, such as ALIGN [DACβ19, Kunal+], MAGICAL [ICCADβ19, Xu+]
Symmetry Constraint Annotation4
Symmetry Constraint
β Symmetry constraint is one of the substantial and representative constraints for layout design
Comparator Schematic Comparator Layout
Prior Works and Challenges5
Categorized into two major types
β Circuit analysis
β Graph matching
Circuit analysis
β Traditional circuit analysis [ICCADβ93, Charbon+]
Graph matching
β Analog constraints extraction based on the signal flow analysis [ASICONβ05, Zhou+]
β S3DET: detecting symmetry constraint by [ASPDACβ11, Liu+]
Prior Works and Challenges6
Challenges
β The performance highly-correlated to empirical parameters
β The surrounding structures of symmetry constraints varies considerably
β No incorporation of designer expertise
Graph Neural Networks (GNNs)7
Graph Neural Networks
β Graph neural networks are powerful in modeling implicit features of graph structure data
β Analog circuits β hypergraphs
Transductive and inductive settings
β Transductive
GCN [ICLRβ17, Kipf+]
β Inductive
GraphSage [NeurIPSβ17, Hamilton+]
β Transductive and inductive
GAT [ICLRβ18, Velickovic+]
GNNs: Node Embedding 8
Node Embedding
β Map each node in a graph into a low-dimensional space
β Capture the insight from local graph structure and node features
Sample and aggregation scheme
β Compute node embeddings by sampling neighbor nodes and aggregating their embeddings
Problem Formulation9
Input
β netlist
β labeled symmetry pairs
Output
β For a new circuit
to predict symmetry pairs
Objective
β to build a model with:
maximum annotation accuracy
1
2
3
SymDetect: Graph Representation11
Graph Representation
β Both device instances and device pins are recognized as graph nodes
β Edges connecting pin nodes with their corresponding device nodes
β The pin nodes form a clique if sharing the same net
SymDetect: Feature Extraction12
Feature extraction
β Type-based feature
β Path-based feature
Type-based feature
β Type information as part of node feature
β A 2-dimension vector to indicate
whether a node is a device or a pin
[0, 1] for a device, [1, 0] for a pin
β A 13-dimension one-hot vector
device types: capacitor, resistor, diode, NMOS, PMOS, IO
pin types: source, drain, gate, substrate, passive, cathode of a diode, anode of a diode
[1, 0]
[0, 1]
[0 β¦ β¦ 0]
nmos[0 β¦ β¦ 0]
source
SymDetect: Feature Extraction13
Feature Extraction
β Type-based feature
β Path-based feature
Path-based feature
β Characterize the global position of
each node by computing the length
of VSS/GND-sourced paths
0
1 1
2
SymDetect: Sample and Aggregate Model15
Sage
β Aggregate by aggregator function
β βπ£π = π(π β {βπ£
πβ1βππΈπ΄π(βπ’
πβ1, βπ’ β π΅(π£))})
0
SymDetect: Binary Classification17
Binary Classification
β We designate a label π¦ for each node pair (π£1, π£2)
for symmetric, for non-symmetric
β We calculate the probability of label π¦ = by bilinear scoring function:
ππππ = π π§1πππ§2
where π(β ) is an activation function, π§1 and π§2 are node embeddings
β We apply binary cross entropy loss to train our GNN model:
πππ π = β1
πΟπ=1π π¦π β log πππππ + β π¦π β log( β πππππ)
SymDetect: Post Processing18
Post Processing
β To reduce false alarms
Rule-based filter
β The sizes and types of two symmetric nodes are supposed to be identical
Probability-based filter
β A device node appears in no more than one symmetry pair
β The filter sorts all candidate pairs by predicted probabilities
β Pick the pair with higher probability
Experimental Setup19
Baseline methods
β S3DET [ASPDACβ20, Liu+]
β signal flow analysis (SFA) [ICCADβ19, Xu+]
Dataset statistics
Experimental Setup20
Evaluation metric
β We adopt: true positive rate(TPR), false positive rate(FPR), and F1-score and ROC
β πππ =ππ
π
β πΉππ =πΉπ
π
β πΉ β π ππππ =2ππ
2ππ+πΉπ+πΉπ
β ROC plots the TPR against the FPR
TP FN
TNFP
true\predicted
positive
negative
positive negative
confusion matrix
Experimental Results23
Experimental Results
β ROC:
β Runtime:
β The one-shot training time costs less than 2 minutes
β The inference time per circuit (about 0.1s) is comparable to the other two approaches
Conclusions and Future Work24
Conclusions
β A graph learning based framework for layout symmetry annotation in analog circuits
β Node feature extraction from both local information (device/pin type) and global graph structure (path length from VSS/GND)
β A training technique to learn the node similarity from the imbalanced data
β The proposed approach outperforms previous symmetry annotation
Future work
β Extend our GNN-based approach to system symmetry constraint
β We can also include other constraints such as matching constraint or shielding constraint