Prophet/Critic Hybrid Branch Prediction

1. Prophet/Critic Hybrid Branch Prediction Falcon, Stark, Ramirez, Lai, Valero Presenter: Christian Wanamaker

2. Outline • Overview & Motivation • Hybrid Branch Prediction • The Prophet/Critic Branch Predictor • Results • Conclusions

3. Overview • Better Branch Prediction is a highly desirable technique because it does not require trade-offs between performance, power, and energy • Despite much research on Branch Prediction, it is by no means solved • Branch prediction is liable to become even more important as pipelines deepen and issue-widths increase

4. Overview (continued) • Issue width = # of uOps (micro ops) issued per clock cycle • Using the Branch Target Buffer to look ahead for branches • Perceptrons – simple neural network. Can look at a longer history than simpler counters.

5. Hybrid Branch Predictors • Use two or more different branch prediction techniques • One may override another (either based on a third selector or one may always override) • The predictions may be combined, for instance as a majority vote • Example: Tournament predictors (often a branch prediction buffer and a Correlating branch predictor with a third predictor choosing which of the two is used in this situation)

6. Prophet/Critic Branch Predictor • Basic Idea – the Prophet makes a series of predictions of future branches, the Critic critiques and if necessary alters them. • The Prophet makes predictions based on the history of the branch. • The Critic looks at the branches prediction of the Prophet after the prophet has predicted a certain number of steps ahead, then critiques the prophecy

7. Prophet Critic Basics • BTB looks ahead, Prophet predicts whether BTB branches are taken/not taken • Branch Outcome Register – the predictions that are in the critics “future” - the number of future bits allowed. • More future bits allow for more accurate viewing of the future

8. Prophet Critic Basics

9. Prophet Critic Basics • The branch predictions are kept in the Fetch Target Queue (FTQ) • Once the future bits are received, the critic makes it’s pronouncement • The critic overrides the prophet if it comes to a different conclusion • If so, the FTQ is purged of un-critiqued predictions, and the prophet is redirected to the path shown by the critic

10. Prophet/Critic Architecture (cont)

11. Prophet/Critic: How it works

12. Prophet Critic: How it works • The Prophet will mispredict A • The Critic will note that the Prophet mispredicted in this case the first time the misprediction occurs. • In the future, when the Critic sees the misprediction, it will correct it • More future bits increase the accuracy of prediction, but reduce the history, so there is an important tradeoff here.

13. Prophet/Critic Filtering • The critic can be limited by multiple branches contending for the same resources • In addition, the critic is not always correct • So, easy to predict branches should be filtered out • This is achieved with tags that are set when a mispredict occcurs • If not tagged, the critic’s critique is ignored.

14. Prophet/Critic Filtering

15. Testing and Result • Testing was done on a cycle accurate IA32 with Long Instruction Traces • The simulator had to follow bad branches, as otherwise the critic would not learn. • Branch Predictors: Gshare, 2bc-gskew, perceptron • uPC is uOps per cycle • misp/kuops is misses per thousand uops

16. Hardware Budgets and predictor types

17. Results of varying Future Bits

18. Results

19. Results

20. Results

21. Conclusions • Speedup of up to 8% with 12 future bits (using the same amount of branch prediction space) • The mispredict rate can be reduced up to 25-31% • Adding future bits helps, but more is not always better • Research suggests that the best future bits can be chosen dynamically

22. Thank you Any questions?