Fused Two-Level Branch Prediction with Ahead Calculation

In this paper, we propose a Fused Two-Level (FTL) branch predictor combined with an Ahead Calculation method. The FTL predictor is derived from the fusion hybrid predictor. It achieves high accuracy by adopting PAp-base Geometrical History Length (p-GEHL) prediction, which is an effective prediction scheme exploiting local histories. The p-GEHL predictor has several prediction tables indexed from independent functions of the local branch histories and branch addresses. The prediction is computed through the summation of values read from the prediction tables. This approach effectively uses limited budget and allows accurate predictions. The Ahead Calculation is an effective implementation scheme for neural predictors exploiting local histories such as the p-GEHL predictor. This scheme is the so-called pre-calculation method. The prediction result is computed when a previous branch with the same address was predicted, and the result is stored in a RAM, which is called Local Prediction Cache (LPC). The LPC reduces prediction latency since the predictor only has to read the RAM by branch address instead of computing the prediction through adder trees. We optimized our FTL branch predictor for the CBP-2 realistic track infrastructure. This optimized-FTL branch predictor with Ahead Calculation achieved 3.466 MPKI with a 262,400-bit budget.