1. André Seznec Caps Team IRISA/INRIA Design tradeoffs for the Alpha EV8 Conditional Branch Predictor André Seznec, IRISA/INRIA Stephen Felix, Intel Venkata Krishnan, Stargen Inc Yiannakis Sazeides, University of Cyprus

2. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Alpha EV8 (cancelled june 2001)  SMT: 4 threads  wide-issue superscalar processor:  8-way issue Single process performance is the goal Multithreaded performance is a bonus 5-10 % overhead for SMT

3. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Challenges on the EV8 conditional branch predictor  High accuracy is needed:  14 cycles minimum miss penalty  Up to 16 predictions per cycle:  from two non-contiguous fetch blocks!  Various implementation constraints:  master the number of physical memory arrays  use of single-ported memory cells  timing constraints

4. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa instruction fetch blocks on EV8 br taken not taken br not taken not taken

5. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Alpha EV8 front-end pipeline  Fetches up to two, 8-instruction blocks per cycle from the I-cache:  a block ends either on an aligned 8-instruction end or on a taken control flow  up to 16 conditional branches fetched and predicted per cycle  Next two block addresses must be predicted in a single cycle:  critical path: use of a line predictor backed with a complex PC address generator: conditional branch predictor, RAS, jump predictor..

6. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa PC address generation pipeline Cycle 1Cycle 2Cycle 3 Line prediction is completed Prediction table read is completed PC address generation is completed C and DA and BY and Z

7. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa EV8 predictor: (derived from) (2Bc-gskew) e-gskew

8. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa 2Bc-gskew: degrees of freedom partial update policy  on correct predictions, only updates correct components:  do not destroy other predictions  better accuracy !  On correct predictions:  prediction bit is only read  hysteresis bit is only written USE OF DISTINCT PREDICTION AND HYSTERESIS ARRAYS !! No reason for same size for hysteresis and prediction arrays

9. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa EV8 predictor: leveraging degrees of freedom Different history lengths Smaller bimodal table

10. André Seznec Caps Team IRISA/INRIA Dealing with implementation constraints

11. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Issues on global history Blocks A and BBlocks Y and Z Blocks C and D Branch infos from C, B and A are not valid to predict D! On each cycle, upto 16 branch are predicted: 0 to 16 bits to be inserted in the history vector !?

12. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Block compressed history lghist  Incorporate at most one bit in the history per fetch block:  0, 1 or 2 bits to be incorporated in history vector per cycle  Which bit ?  Direction of the last conditional branch in the block previous ones are not taken  XORed with position (1st half/ 2nd half) in the block more uniform distribution of the history vectors

13. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa instruction fetch blocks on EV8 br taken 1 is inserted br takennot taken 0 is inserted

14. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa The EV8 branch predictor information vector  History information is not available on the three previous blocks A, B, and C  but, addresses are available !! Information vector to index the predictor: 1. Instruction address 2. Lghist (3-blocks-old history + path) 3. Path info on the last three blocks

15. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Using single-ported memory arrays The challenge: 16 predictions to be performed per cycle from two non-contiguous blocks ! 8 updates per cycle: for two non-contiguous blocks ! But single-ported arrays are highly desirable :-)

16. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Bank-interleaved or double-ported branch predictor ?  Reads of predictions for two 8-instructions blocks:  double-porting: memory cells twice as large losing half of the entries ?  bank-interleaving: need for arbitration longer critical electrical path losing throughput short loops fitting in a single 8-instruction block !? ????????

17. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Conflict free interleaved bank predictor Key idea: Force adjacent predictions to lie in distinct banks Bank for A is determined by Y and Z if (y6,y5)== Bz then Ba =(y6,y5+1) else Ba = (y6,y5) 4-way interleaved:

18. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Conflict free bank-interleaved predictor (2)  Conflicts are avoided by construction  Bank number is computed one cycle ahead  not on the critical path Single ported bank-interleaved memory arrays !

19. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa « Logical view » vs real implementation  4 tables * 4 banks * 2 (pred. +hyst.):  32 memory arrays  Indexing functions are computed, then arrays are accessed  4 banks * 2 (pred. + hyst.)  4 tables in a single array  8 memory arrays  No time to lose:  start access and compute part of the index in //

20. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Reading the branch prediction tables Bank selection 1 out of 4 MetaG0G1 BIM Wordline selection 1 out 64 Column selection: 8 out of 256 Unshuffle: 8 to 8

21. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Reading the branch prediction tables (2)  Span over 5 cycle phases:  Cycle -1: bank number computation bank selection  Cycle 0: phase 0: wordline selection phase 1: column selection  Cycle 1: phase 0: unshuffle permutation

22. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Constraints for indices composition  Strong: Wordline bits:  immediate availability  common to the four logical tables  Medium: Column bits  a single 2-entry XOR gate  Weak: Unshuffle bits:  near complete freedom, a full tree of XOR gates if needed

23. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Designing the indexing functions (1) 6 wordline bits  Must be available at the beginning of the cycle:  block address bits  3-block old lghist bits  path bits  Tradeoff:  address bits for emphasizing bimodal component behavior  lghist bits are more uniformly distributed 4 lghist bits + 2 address bits

24. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Designing the indexing functions (2) Column selection and unshuffle  Favor independance of the four indexing functions:  if two (address,history) pairs conflict on a table then try to avoid repeating the conflict on an other table  Guarantee that for a single address, two histories that differ by only one or two bits will not map on the same entry  Favor usage of the whole table:  lghist bits are more uniformly distributed than address bits XORing 2 lghist bits for column bits a XOR tree with up to 11 bits for unshuffle

25. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa EV8 branch predictor configuration  208 Kbits for prediction and 144 Kbits for hysteresis  «BIM»: 16 K + 16 K, 4 lghist bits (+ 3-block path)  G0: 64 K + 32 K, 13 lghist bits  G1: 64 K + 64 K, 21 lghist bits  Meta: 64 K + 32 K, 17 lghist bits  4 prediction banks and 4 hysteresis banks

26. André Seznec Caps Team IRISA/INRIA Performance evaluation Sorry, SPEC 95 :-)

27. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Benchmarks characteristics  Highly optimized SPECint 95:  much more not-taken than taken  ratio lghist/ghist length: from 1.12 to 1.59  from 8.9 to 16.2 branches per 100 instructions

28. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa 2Bc-gskew vs other global history predictors

29. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Quality of information vector

30. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Reducing some table sizes no significant impact

31. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Quality of indexing functions

32. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Conclusion  Design of a real branch predictor leads to challenges ignored in most academic studies:  3-block old history vector  inability to maintain a complete history  simultaneous accesses to the predictor  minimization of the number of memory arrays  timing constraints on the indexing functions We overcame these difficulties and adapted a state of the art academic branch predictor to real world constraints.

33. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Summary of the contributions  Efficient information vector can be built with mixing path and compressed history:  don’t focus on the info vector, use what is convenient!  Use of different table sizes, history lengths in the predictor.  Sharing of hysteresis bits  Conflict free parallel access scheme for the predictor  Engineering of indexing functions

34. The Alpha EV8 Conditional Branch Predictor André Seznec Caps Team Irisa Acknowledgements To the whole EV8 design team Special mention to: Ta-chung Chang, George Chrysos, John Edmondson, Joel Emer, Tryggve Fossum, Glenn Giacalone, Balakrishnan Iyer, Manickavelu Balasubramanian, Harish Patil, George Tien and James Vash.