A Shadow Dynamic Finite State Machine for Branch Prediction: An Alternative for the 2-bit Saturating Counter

In order to meet high performance demands, modern processor architectures exploit varieties of dynamic branch prediction topologies ([4]-[6] provide an excellent introduction and research coverage) to increase instruction-level parallelism (ILP). Dynamic branch predictors use run-time branch execution history to predict branch direction. Most previous techniques use a branch pattern history table (known as PHTs, BHTs, or BPHTs) to record past branch behavior (e.g., global and/or local) and these tables are indexed using a function/subset of the branch address. Nearly all dynamic branch predictors explored in the last 10 years have been based on tables containing 2-bit saturating counters [7][8]. Extensive simulations of branch predictors reveal that the 2-bit saturating counter performs the best on average [9][10], and thus are used in modern commercial processors. In recent years, research has explored more advanced branch prediction techniques such as neural networks [11][12] and other forms of machine learning. Despite their impressive simulation accuracy, to the best of our knowledge no commercial efforts have publicly announced incorporating such branch predictors because these branch predictors are commonly known to exhibit high prediction latency and long training periods with increased area and energy per prediction [13]. In order to provide increased branch prediction accuracy with low area and power overheads, in this paper we propose a novel adaptive learning machinebased shadow dynamic finite state machine (SDFSM). The SDFSM learns/predicts an application’s unique branching pattern using the prediction values (taken/not taken) stored in each state. Upon branch execution, state transition is input independent and the value of the target state predicts the branch outcome. Each state has a corresponding shadow state, which contains the alternate branch prediction value. In the event of a mispredicted branch, the SDFSM performs self-modification by swapping the current state with the current state’s shadow state, which contains the correctly predicted branch outcome. This method of state swapping dynamically records unique branch patterns, thus specializing the branch predictor to the needs of an application. Extensive experimental results compare the SDFSM prediction accuracy to the commonly used bimodal [1][2] counter-based predictor and reveal that, for a subset of benchmarks, an SDFSM with six shadow states provides more accurate predictions than counterbase predictors with one-to-one prediction latency. The remainder of this paper is organized as follows. Section 2 describes the proposed SDFSM as an alternative replacement for 2-bit saturating counters and presents the SDFSM architecture. Section 3 and Section 4 present our simulation methodology setup and branch Figure 1: The proposed shadow dynamic finite state machine (SDFSM) using four states A Shadow Dynamic Finite State Machine for Branch Prediction: An alternative for the 2-bit Saturating Counter Saleh Abdel-Hafeez, Ann Gordon-Ross, Asem Albosul, Ahmad shatnawi and Shadi Harb Department of Computer Engineering, Jordan University of Science & Technology Irbid, Jordan 21110 [email protected] Department of Electrical & Computer Engineering University of Florida, Gainesville, FL 32611, USA [email protected] *Also with the NSF Center for High Performance Reconfigurable Computing (CHREC) at UF