1. Gang Luo, Hongfei Guo {gangluo, guo}@cs.wisc.edu
Software-based and Hardware-based Branch Prediction Strategies and Performance Evaluation
Gang Luo, Hongfei Guo
{gangluo,

2. Contents Background Implementation based on SimpleScalar3.0
Simulation Model
Simulation Results
Conclusion and Future work

3. 1. Background Why branch prediction? Software-based branch predictors
Hardware-based branch predictors

4. Why need branch prediction
Avoid control hazard
Predict branch direction and branch address
Avoid stopping pipeline
Speedup program execution

5. Classification of Branch Prediction Strategies
Software-based static branch predictors
Hardware-based dynamic branch predictors

6. Software-based Strategy
Always Taken
Always Not Taken
Backward Taken, Forward Not Taken
All branches with certain operation codes Taken, others Not Taken

7. Hardware-based Strategy
One-bit counter
Two-bit counter
Gag
Pag
Pap
Branch Instruction Table

8. 2. Implementation based on SS3.0
Backward taken & forward not taken
one-bit prediction counter
fix a bug
---- Support command line configuration

9. current branch address Backward taken & forward not takenY
target address
Current address
>
Target address
Predicted address
Next inst. address
target address N
Backward taken & forward not taken
branch predictor

10. current branch address Predicted inst. address
One bit counter
(two bits counter)
HASH table
current branch address
HASH 1 ……
Lookup
Predicted direction
BTB
update
address2
target address2
Addressn-1
target address n-1
address1
target address1 ……
address n
target address n
Predicted target
2 to 1
MUX
Predicted inst. address
Next inst. address
Compare
update
New inst. address
Branch prediction counter

11. 3. Simulation Model Mode Level 1 Table Level 2 Table
# of Entries Width of SR # of Entries
One-bit
Bimode
GAg
PAg
PAp
BTB Associ

12. Spec95 Benchmarks Used
Gcc
Compress Li
M88k
Perl

13. 4. Simulation Results Results of software-based Strategies
Results of hardware-based Strategies
Software-based Strategies vs.
hardware-based Strategies
Explanations

14. Performance of Software-based Strategies

15. Analysis of Software-based Strategies
Always Taken is the best one
Always Not Taken & Backward Taken, Forward Not Taken are equally bad

16. Explanation
Backward Taken, Forward Not Taken is mainly pointed at the for loop and while loop statement
Most branches in the program are inclined to be taken

17. Performance of Hardware-based Strategies

18. Analysis of Hardware-based Strategies
Two-bit counter and Gag are the best two
One-bit counter, Pag, and Pap also achieve high prediction accuracy

19. Explanation One-bit counter: intrinsic shortcoming
Pag: all the branch histories share the same global two-bit counter tables, which leads to confusion
Pap: too many branch history pattern tables compared to the number of instructions executed, can’t get enough training

20. Software-based Strategies vs. Hardware-based Strategies
Hardware-based strategies always achieve high prediction accuracy, greater than 90%
Most software-based strategies achieve low prediction accuracy, less than 70%

21. Explanation
Different data sets will let the programs have different dynamic branch behaviors
Even for the same data set, the same branch instruction will exhibit different dynamic branch behaviors at different times in the same running

22. 5. Future Work
Use some AI algorithms to adjust the parameters of the branch prediction strategies, then we can ensure the most proper training time to achieve the highest performance.