phidian Introduction Related work Machine learning methodology - - PowerPoint PPT Presentation

How deep learning can drive physical synthesis towards more predictable legalization

Renan Netto, Sheiny Fabre, Tiago Augusto Fontana, Vinicius Livramento, La´ ercio Pilla, Jos´ e Lu´ ıs G¨ untzel

Embedded Computing Lab (ECL) — Dept. of Computer Science and Statistics Federal University of Santa Catarina (UFSC) ISPD - April 14-17 - San Francisco, CA

phidian

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Outline

1

Introduction

2

Related work

3

Machine learning methodology

4

Physical design integration

5

Experimental results

6

Conclusions

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Machine learning applications in physical design

Clock data Layout data

Routing data

j i j i

Timing data ML models

Routing Violations CTS Result Signoff Timing

Renan Netto ISPD - April 14-17 - San Francisco, CA 3

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Machine learning applications in physical design

Clock data Layout data

Routing data

j i j i

Timing data ML models

Routing Violations CTS Result Signoff Timing

It has not been used to predict legalization yet!

Renan Netto ISPD - April 14-17 - San Francisco, CA 3

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Machine learning applications in legalization

1) Choosing among different legalization algorithms

illegal placement solution A solution B

legalization algorithm A legalization algorithm B Renan Netto ISPD - April 14-17 - San Francisco, CA 4

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Machine learning applications in legalization

2) Guiding an incremental placement technique

Renan Netto ISPD - April 14-17 - San Francisco, CA 5

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Machine learning applications in legalization

3) Guiding an circuit partitioning strategy

Renan Netto ISPD - April 14-17 - San Francisco, CA 6

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Machine learning applications in legalization

3) Guiding an circuit partitioning strategy This work explores mainly option (3), but also partially explores option (2)

Renan Netto ISPD - April 14-17 - San Francisco, CA 6

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Contributions

Feature extraction strategy for training machine learning models. Evaluation of different ML models in order to select the best one for this problem. We employed the best ML model as a pruning mechanism for a circuit partitioning strategy.

Renan Netto ISPD - April 14-17 - San Francisco, CA 7

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Table of related works

Work Prediction Features ML model Kahng et al. CTS outcome clock data non-convolutional Han et al. signoff timing timing data non-convolutional Zhou et al. # of DRCs layout data, routing data non-convolutional Chan et al. locations of DRCs layout data, routing data non-convolutional Fabrizi et al. short violations layout data, routing data non-convolutional Xie et al. # and location of DRCs circuit snapshot convolutional This work legalization quality layout data, circuit snapshot non-convolutional, convolutional

Renan Netto ISPD - April 14-17 - San Francisco, CA 8

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Table of related works

Work Prediction Features ML model Kahng et al. CTS outcome clock data non-convolutional Han et al. signoff timing timing data non-convolutional Zhou et al. # of DRCs layout data, routing data non-convolutional Chan et al. locations of DRCs layout data, routing data non-convolutional Fabrizi et al. short violations layout data, routing data non-convolutional Xie et al. # and location of DRCs circuit snapshot convolutional This work legalization quality layout data, circuit snapshot non-convolutional, convolutional

Renan Netto ISPD - April 14-17 - San Francisco, CA 9

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Table of related works

Work Prediction Features ML model Kahng et al. CTS outcome clock data non-convolutional Han et al. signoff timing timing data non-convolutional Zhou et al. # of DRCs layout data, routing data non-convolutional Chan et al. locations of DRCs layout data, routing data non-convolutional Fabrizi et al. short violations layout data, routing data non-convolutional Xie et al. # and location of DRCs circuit snapshot convolutional This work legalization quality layout data, circuit snapshot non-convolutional, convolutional

Renan Netto ISPD - April 14-17 - San Francisco, CA 10

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Table of related works

Work Prediction Features ML model Kahng et al. CTS outcome clock data non-convolutional Han et al. signoff timing timing data non-convolutional Zhou et al. # of DRCs layout data, routing data non-convolutional Chan et al. locations of DRCs layout data, routing data non-convolutional Fabrizi et al. short violations layout data, routing data non-convolutional Xie et al. # and location of DRCs circuit snapshot convolutional This work legalization quality layout data, circuit snapshot non-convolutional, convolutional

Renan Netto ISPD - April 14-17 - San Francisco, CA 11

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Methodology overview

Layout data

ML model

Disp > threshold

Renan Netto ISPD - April 14-17 - San Francisco, CA 12

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Training data generation

Circuit partitioning using k-d tree (height = 2 in the example)

Renan Netto ISPD - April 14-17 - San Francisco, CA 13

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Training data generation

Sample 1 Sample 2 Sample 3 Sample 4

Circuit partitioning using k-d tree (height = 2 in the example)

Renan Netto ISPD - April 14-17 - San Francisco, CA 13

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Training data generation

Displacement 1 Displacement 2 Displacement 3 Displacement 4

Circuit partitioning using k-d tree (height = 2 in the example)

Renan Netto ISPD - April 14-17 - San Francisco, CA 13

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Handling partitions of different sizes

Sample 1 Sample 2

Renan Netto ISPD - April 14-17 - San Francisco, CA 14

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Handling partitions of different sizes

Sample 1 Sample 2 Sample 3 Sample 4

Renan Netto ISPD - April 14-17 - San Francisco, CA 14

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Handling partitions of different sizes

Sample 1 Sample 2 Sample 3 Sample 4

Actual values: 1 ≤ height ≤ 9 and 1024 samples for each height.

Renan Netto ISPD - April 14-17 - San Francisco, CA 15

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Feature selection: non-convolutional ML models

D = 0.49

Density of the partition area

A = [64, 48]

Area occupied by cells of each height

Renan Netto ISPD - April 14-17 - San Francisco, CA 16

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Feature selection: non-convolutional ML models

Renan Netto ISPD - April 14-17 - San Francisco, CA 17

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Feature selection: non-convolutional ML models

0.2 0.4 0.6 0.8 1 5 4 3 2 1

Renan Netto ISPD - April 14-17 - San Francisco, CA 17

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Feature selection: non-convolutional ML models

0.2 0.4 0.6 0.8 1 0.1 0.2 0.3 0.4 0.5

Renan Netto ISPD - April 14-17 - San Francisco, CA 17

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Feature selection: non-convolutional ML models

0.2 0.4 0.6 0.8 1 4 2 6 8

Renan Netto ISPD - April 14-17 - San Francisco, CA 18

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Feature selection: non-convolutional ML models

0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8

Renan Netto ISPD - April 14-17 - San Francisco, CA 18

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Feature selection: convolutional model

Partition snapshot

Fixed cells: blue Movable cells: shades of pink

Renan Netto ISPD - April 14-17 - San Francisco, CA 19

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Circuit partitioning strategy

Partitions circuit using a k-d tree data structure Each partition is legalized independently

Renan Netto ISPD - April 14-17 - San Francisco, CA 20

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Proposal: merge partitions that result in large displacement

Max Disp: 12 Max Disp: 28 Max Disp: 14 Max Disp: 23

Example with displacement threshold of 25

Renan Netto ISPD - April 14-17 - San Francisco, CA 21

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Proposal: merge partitions that result in large displacement

Max Disp: 14 Max Disp: 23

Example with displacement threshold of 25

Renan Netto ISPD - April 14-17 - San Francisco, CA 21

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Proposal: merge partitions that result in large displacement

Max Disp: 13 Max Disp: 14 Max Disp: 23

Example with displacement threshold of 25

Renan Netto ISPD - April 14-17 - San Francisco, CA 21

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Circuit partitioning strategy

Max Disp: 13 Max Disp: 14 Max Disp: 23

Two ways of verifying partitions with large displacement:

Running the legalization algorithm Using ML model to predict those partitions

Renan Netto ISPD - April 14-17 - San Francisco, CA 22

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Experimental infrastructure

Benchmarks

Training set

pci_bridge32_b_md3: 29K cells fft_2_md2: 32K cells fft_a_md2: 31K cells fft_a_md3: 31K cells des_perf_a_md1: 108K cells des_perf_a_md2: 108K cells des_perf_1: 113K cells des_perf_b_md1: 113K cells des_perf_b_md2: 113K cells edit_dist_1_md1: 131K cells edit_dist_a_md2: 127K cells edit_dist_a_md3: 127K cells

Validation set

pci_bridge32_a_md1: 29K cells pci_bridge32_a_md2: 29K cells pci_bridge32_b_md1: 29K cells pci_bridge32_b_md2: 29K cells

T est set

superblue18: 768M superblue4: 795M superblue16: 981M superblue5: 1086M superblue1: 1209M superblue3: 1213M superblue10: 1876M superblue7: 1931M

Three ML models:

Artificial Neural Network (ANN) Decision Tree (DT) Convolutional Neural Network (CNN)

Three different displacement thresholds: 5, 10 and 15 rows Metrics: accuracy, precision, recall, F-score

Renan Netto ISPD - April 14-17 - San Francisco, CA 23

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Evaluation of ML models: threshold of 5 rows

ANN DT CNN ML model 0.0 0.2 0.4 0.6 0.8 1.0 Result accuracy precision recall fscore

Similar accuracy between ANN and CNN CNN achieves better F-score DT is the worst model

Renan Netto ISPD - April 14-17 - San Francisco, CA 24

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Evaluation of ML models: threshold of 10 rows

ANN DT CNN ML model 0.0 0.2 0.4 0.6 0.8 1.0 Result accuracy precision recall fscore

Slight accuracy increase for ANN and DT Precision reduction on all models Recall affected only on DT

Renan Netto ISPD - April 14-17 - San Francisco, CA 25

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Evaluation of ML models: threshold of 15 rows

ANN DT CNN ML model 0.0 0.2 0.4 0.6 0.8 1.0 Result accuracy precision recall fscore

Even more unbalanced data F-score reduction on both ANN and CNN Chosen model: ANN

Higher accuracy and F-score for threshold of 15 rows Lower model complexity

Renan Netto ISPD - April 14-17 - San Francisco, CA 26

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Results of the pruning strategy: threshold of 5 rows

Avg disp Max disp HPWL 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Mean LEG ANN

Ratio of result using pruning strategy by original result (lower than 1 is better) Two ways of identifying partitions to merge:

Using the legalization algorithm (LEG) Using ML model (ANN)

Renan Netto ISPD - April 14-17 - San Francisco, CA 27

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Results of the pruning strategy: threshold of 5 rows

Avg disp Max disp HPWL 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Mean LEG ANN

ANN achieves same results as LEG Greater reduction on max displacement Avg displacement reduction was more relevant on small circuits No significant difference to threshold

• f 10 rows

Renan Netto ISPD - April 14-17 - San Francisco, CA 28

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Results of the pruning strategy: threshold of 15 rows

Avg disp Max disp HPWL 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Mean LEG ANN

Slight increase on max displacement for ANN Still great improvement compared to

• riginal solution

Renan Netto ISPD - April 14-17 - San Francisco, CA 29

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Number of calls to legalization algorithm

ANN / LEG

Reduction for all circuits Greater reduction for larger circuits and threshold of 5 rows

Renan Netto ISPD - April 14-17 - San Francisco, CA 30

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Legalization time speedup

Speedup is negligible for small circuits ANN is faster for almost all cases Higher speedup for lower threshold Exceptions are cases where ANN failed to identify some partitions

Renan Netto ISPD - April 14-17 - San Francisco, CA 31

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Conclusions and future work

We evaluated three ML models to improve predictability of legalization algorithms

Artificial Neural Network Decision Tree Convolutional Neural Network

Best model was used as pruning mechanism of partitioning strategy:

Greatly reduces maximum displacement Avoids up to 99% of calls to legalization algorithm

Future work:

Predicting resulting displacement itself Evaluating on different legalization algorithms Improvements on the CNN model to compensate its higher complexity

Renan Netto ISPD - April 14-17 - San Francisco, CA 32

How deep learning can drive physical synthesis towards more predictable legalization

Questions?

Renan Netto

renan.netto@posgrad.ufsc.br

phidian

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Speedup for best case

LEG: threshold of 15 rows ANN: threshold of 5 rows ANN is still faster

Renan Netto ISPD - April 14-17 - San Francisco, CA 34

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Accuracy along epochs

1 2 3 4 5 6 7 8 9 10 Epoch 79 80 81 82 83 84 85 86 87 Accuracy (%) training validation

Renan Netto ISPD - April 14-17 - San Francisco, CA 35

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Experiments with other models: threshold of 5 rows

RF: Random forest GT: Gradient boosted tree

ANN RF GT ML model 0.0 0.2 0.4 0.6 0.8 1.0 Result accuracy precision recall fscore

Renan Netto ISPD - April 14-17 - San Francisco, CA 36

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Experiments with other models: threshold of 10 rows

RF: Random forest GT: Gradient boosted tree

ANN RF GT ML model 0.0 0.2 0.4 0.6 0.8 1.0 Result accuracy precision recall fscore

Renan Netto ISPD - April 14-17 - San Francisco, CA 37

Introduction Related work Machine learning methodology Physical design integration Experimental results Conclusions

Experiments with other models: threshold of 15 rows

RF: Random forest GT: Gradient boosted tree

ANN RF GT ML model 0.0 0.2 0.4 0.6 0.8 1.0 Result accuracy precision recall fscore

Renan Netto ISPD - April 14-17 - San Francisco, CA 38