Branch Prediction using Advanced Neural Methods

1. Branch Prediction using Advanced Neural Methods Sunghoon Kim CS252 Project

2. Introduction • Dynanmic Branch Prediction • No doubt about its importance in Speculation performance • Given history of branch behaviors, predict branch behaviors at the next step • Common solutions: gshare, bimode,hybrid… • Replace saturating counters with neural methods?

3. Neural methods • Capable of classification (predicting into which set of classes a particular instance will fall) • Learns correlations between inputs and output, and generalized learning to other inputs • Potential to solve problems of most two-levels predictors

4. Simulation Models - Gshare • 20-bit Global history shift register • Per-address history table with 2-bit saturation counter GHT Predict 2-bit saturation counter BrachPC>>2 Update counters PHT

5. Simulation Models - Perceptron PHT GHT • 14-bit Global history shift register • Per-address history table with 8-bit weights and bias • Indexed by Gshare or BranchPC alone BrachPC>>2 OR Predict BrachPC>>2 Training weights & bias

6. Simulation Models - Backpropagation GHT • 10-bit GHR • Sigmoid transfer function • Floating point computation • Floating point weights and biases • 20 neurons one hidden layer BrachPC>>2 OR Predict BrachPC>>2 Training weights & bias PHT

7. Simulation Models – Radial Basis Networks GHT PHT • Transfer function for radial basis neuron exp(-n2) • Distance function between an input vector and a weight vector BrachPC>>2 OR Predict BrachPC>>2 Training weights & bias

8. Simulation Models – Elman Networks GHT PHT • Feedback from the hidden layer outputs to the first layer BrachPC>>2 OR Predict BrachPC>>2 Training weights & bias

9. Simulation Models – Learning Vector Quantization Networks GHT PHT • Distance function as Radial but without biases • Competitive function gives one only to an winning input (biggest value) and zeroe to the other BrachPC>>2 OR Predict BrachPC>>2 Training weights & bias

10. Simulation Environment • SimpleScalar Tool • Some of SPEC2000 benchmarks • Execute 100,000,000 instructions and dump conditional branch histories • 5000 branch instructions are used for training • Make all budgets for PHTs the same • Floating point is 4 byte

11. Results

12. Hardware constraints • Predictors must predict within a (few) cycle • Gshare : easy to achieve • Perceptron : Integer adders, possible alternative, more accurate if more layers • Other advanced neural net : Hard to implement, Floating point functional units,

13. Future Works • Replace floating point weights and biases with scaled integer ones? • Replace floating point function with approximately equivalent integer function, using Taylor’s series? • Without budget consideration, what will be the best performance of advanced neural network methods? • Look at codes carefully if there are mistakes

14. Conclusions • There is not much of benefit using advanced neural networks on the same budget as Gshare, sometimes worse • Elman Networks method is the best • Hard to implement in hardware unless floating point computation are easy to do • NN can be alternative predictors if well designed