Use half of the baselines instead of hardcoded 128. This provides better performance for large numbers of baselines.