17. Branch Prediction

Static

• always not taken

• always taken

• BTFN (backward taken, forward not taken)

  • predict T when the branch goes backward (loops for ex)
  • predict N when the branch goes forward (jump to section further in the code)

• FTBN (forwards taken, backwards not taken)

  • opposite of BTFN

• Profile based (likely direction)

• program analysis based (likely direction)

The compiler can optimise for this if it’s architecture aware.

Dynamic Branch Prediction

we can either have a GHR (global history) that indexes into a PHT

• so if we globally last saw TT → index into PHT with that and get 01 or something back

or we can have local branch history, with a PHT for each branch index

2-bit Counter

This means that we just take the last branch outcome and use it as the FSM input. No PHT needed.

Two-level Global Branch Prediction

Instead of only indexing into the PHT via two last seen states, we instead use the previous branch direction as well.

Example of what a table in use looks like:

• TT → 01 means that when it sees TT prev access, it’s weakly not taken.

The architecture of this looks like:

Example:

Adding more Context

We can also index using branch address as well, to add more “context”.

Local Branch History

Implementation:

Aside - Taxonomy

Hybrid Predictors

There is heterogeneity → ifs, loops, etc…
so we use multiple predictors and predict which predictor to use.

More Advanced

Perceptron

TAGE

Different History lengths for different branches, chosen dynamically.