1. And Branch Prediction Machine Learning You can follow along at:
Machine Learning
And Branch Prediction
IN ORDER OF SPEAKING:
Guru Nishok Radhakrishnan, Crystin Rodrick, and Anuj Wali

2. INTRODUCTION Introduction Previous research Simulation
OVERVIEW
Introduction
Previous research
Simulation
Machine learning
Design and Implementation
Future Work

3. INTRODUCTION Different number of branch predictors.
Decreasing HW costs , Feasible branch predictors.
Machine Learning – Recognizing patterns within the data.

4. PERFORMANCE OVERVIEW Bimodal Branch Prediction:
Used in non-MMX Intel Pentium  processor
Saturates at 93.5% correct
REF [4]

5. LOCAL BRANCH PREDICTION: Separate history buffer.
Pattern history table may or may not be shared.
Used in Pentium II, Pentium III.
Saturates at 97.1% correct.
REF [4]

6. GLOBAL BRANCH PREDICTION:
Shared history records
Advantage: Correlation between the different conditional jumps is a part of making the predictions.
Disadvantage: History gives irrelevant information if the conditional jumps are uncorrelated.
Used in Intel Pentium M, Core ,Core 2.

7. Saturates at 96% correct.
REF [4]

8. Neural Branch Predictor:
Advantage: Exploit long histories which requires linear resource growth.
Disadvantage: Perceptron predictor has a high Latency.
Used in AMD Ryzen Processor

9. PREVIOUS RESEARCH:
“Dynamic Branch Prediction using Machine Learning Algorithms”, Kedar Bellare, Pallika Kanani and Shiraj Sen Department of Computer Science, University of Massachusetts Amherst May 17, 2006.
p=rep1&type=pdf
“Branch Prediction with Bayesian Networks”, Jeremy Singer, Gavin Brown, and Ian Watson ,University of Manchester, UK
7b1fa3d3.pdf

10. “A 2-Clock-Cycle Naïve Bayes classifier for Dynamic branch prediction in pipelined RISC Microprocessors” , Itaru Hida, Masayuki Ikebe, Tetsuya Asai, and Masato Motomura, Graduate School of Information Science and Technology.

11. MOTIVATION Two level predictor - High overhead.
Neural Predictors is already commercially available.
Overcome the issues through machine learning concepts.

12. PARSING THE TRACE FILES
SIMULATION
PARSING THE TRACE FILES
BRANCH
HISTORY
BRANCH
PREDICTION

13. WHAT IS MACHINE LEARNING?
ARTIFICIAL INTELLIGENCE
No rule set, they create their own rules.
DEEP LEARNING
Multiple layers
MACHINE LEARNING
Training sets

14. Types of Machines Learning
ALGORITHMS
Reinforcement Learning
Unsupervised Learning
Semi-Supervised Learning
Supervised Learning

15. Algorithm
Naive Bayes

16. LoGISTIC REGRESSION
Created in the 50’s.
Directly learns P(Y|X)

17. AlGORITHM hw(x) =g(wTx) = g(w0x0 +w1x1 +···+wnxn) g(z) = 1/ (1 + e-z)
Logistic Regression
hw(x) =g(wTx) = g(w0x0 +w1x1 +···+wnxn)
g(z) = 1/ (1 + e-z)
Thus, hw(x)=1/1+e(-wT x)
w T x should be large negative values for negative instances   w T x should be large positive values for positive instances.
If we assume a threshold of 0.5, then we predict 1, else 0.

18. Logistic FUNction J(w) = − y(i) log(hw(x)) + (1 − y(i)) log(1 − hw(x))
wj =wj − α (∂/∂wj)*J(w)

19. DUCK TYPING VS

20. EXAMPLE DATASET
QUACK
WADDLE
TYPE 1
DUCK
TOY

21. PROBABILITIES P(QUACK=1|DUCK) = 5/7 = 0.71 P(QUACK=1|TOY) = 2/7 = 0.29
P(WADDLE=1|DUCK) = 2/5 = 0.4
P(WADDLE=1|TOY) = 3/5 = 0.6
P(DUCK) = 5/10 = 0.5
P(TOY) = 5/10 = 0.5

22. NAIVE BAYES PREDICT BY FINDING ARGMAX OF P(Y|WADDLE=1,QUACK=1) DUCK
SOLUTION
PREDICT BY FINDING ARGMAX OF P(Y|WADDLE=1,QUACK=1)
DUCK
P(0.5)*P(0.71)*P(0.4) = 0.14
TOY
P(0.5)*P(0.29)*P(0.6) = 0.09
THEREFORE, WE PREDICT DUCK

23. LOGISTIC REGRESSION INITIALIZE W, 0.33 SPREAD OVER ALL WEIGHTS
RUN THROUGH
INITIALIZE W, 0.33 SPREAD OVER ALL WEIGHTS
AFTER FIRST ITERATION
1 / (1 + e(-( (1) + 0.3*(1))
wj =wj − α (∂/∂wj)*J(w)
Once you get the optimal weights, plug in the X values to get your y

24. DESIGN
Naive Bayes
Why Naive Bayes?
Faster than other regression/classification models
Relatively less complex circuitry
Amount of physical space needed is lower

25. DESIGN What do we need to build one? Set of known attributes
Naive Bayes
What do we need to build one?
Set of known attributes
Conditional Probability Table
Posterior Probability

26. DESIGN Set of known attributes: Last 30 branch outcomes
Naive Bayes
Set of known attributes:
Last 30 branch outcomes
Maintained as a queue

27. DESIGN CPT TABLE Bimodal Counters instead of Probabilities
Naive Bayes
CPT TABLE
Bimodal Counters
instead of Probabilities
Updated after each branch

28. DESIGN Posterior Probabilities P(y = 0 | X) P(y = 1 | X)
Naive Bayes
Posterior Probabilities
P(y = 0 | X)
P(y = 1 | X)
Made faster using Look-Up-Table

29. DESIGN
Naive Bayes
PROCESS FLOW

30. DESIGN PROCESSOR ARCHITECTURE LatticeMico32 6 stage pipeline
Naive Bayes
PROCESSOR ARCHITECTURE
LatticeMico32
6 stage pipeline
WHERE DOES IT FIT IN THE PIPELINE?

31. DESIGN Naive Bayes EXPECTED MODEL PARAMETERS Counter Size
History Length
Figures taken from [1]

32. DESIGN Naive Bayes EXPECTED RESULTS
Performance Measurement: Misprediction Rate (%)
Performs better than others
Figure taken from [1]

33. POSSIBLE WORK ISSUES Not much research available Feasibility
Logistic Regression
ISSUES
Not much research available
Feasibility
POSSIBLE IMPLEMENTATION MODEL
Same Attribute Set
Performance similar to Perceptron

34. REFERENCES
1) “A 2-Clock-Cycle Naïve Bayes classifier for Dynamic branch prediction in pipelined RISC Microprocessors” , Itaru Hida, Masayuki Ikebe, Tetsuya Asai, and Masato Motomura, Graduate School of Information Science and Technology.
2) “Dynamic Branch Prediction using Machine Learning Algorithms”, Kedar Bellare, Pallika Kanani and Shiraj Sen Department of Computer Science, University of Massachusetts Amherst May 17, 2006.
3) “Branch Prediction with Bayesian Networks”, Jeremy Singer, Gavin Brown, and Ian Watson ,University of Manchester, UK.
4)”Combining Branch Predictors”, Scott McFarling , June 1993.

35. THANK YOU