Comparison with the current black-parrot implementation:
Note: This is a comparison for bp_cnt_sat_bits=2 and bh_indx_width_p=9, which is the default value currently used in black parrot. Several predictors might out-perform the bimodal predictor for slightly different bh_indx_width_p values (see here)
There are tree main observations:
- Saturating counters should probably have two or three bits and the get worse with more, most likely due to the longer learning/re-learning phase.
- The (default) behaviour of hash functions: the bigger the image size to map into, the less collisions and therefore the higher the accuracy.
- bp_n_hist=3 is slightly more optimal than bp_n_hist=6