S.T. Pan, K. So, and J.T. Rahmeh, "Improving the Accuracy of Dynamic Branch Prediction Using Branch Correlation," Proc. 5th Int'l Conf. Architectural Support for Programming Languages and Operating Systems (ASPLOS-V), IEEE CS Press, Los Alamitos, Calif., 1992, pp. 76-84.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....cannot make use of chained lookups. This rules out any scheme that requires precise branch history because only the address of the first block (used to find branch history) is known at lookup time and that block may be in another entry. Precise history is important for local history predictors [17,10] but is not needed for others such as bimodal and gshare [8] Yeh et al. do show how to precisely implement a global history predictor, GAg [18] in which a pattern history table holding 2 bit counters is indexed by a global history register. At the beginning of each cycle d lookups are made in a ....

Shien-Tai Pan, Kimming So, and Joseph T. Rahmeh, "Improving the accuracy of dynamic branch prediction using branch correlation," ACM Computer Architecture News, vol. 20, pp. 76-84, October 1992.

....body does not contain any conditional branches, loop prediction can be trivial. If the loop body exhibits a complex behavior containing multiple branches within its loop body, accurate loop execution prediction becomes more di#cult. Fortunately, conditional branch behavior tends to be correlated [19]. With this in mind, we propose to predict entire loop executions based on correlating path to loop branch behavior with path in loop behavior. 2 Related work The identification of delays caused by the computation of branch addresses was first observed in the 1960s and formally discussed at the ....

S.-T. Pan, K. So, and J. T. Rahmeh. Improving the accuracy of dynamic branch prediction using branch correlation. In Proc. of the fifth International Conference on Architectural Support for Programming Languages and Operating System, volume 27-9, pages 76--84, New York, NY, 1992. ACM Press.

....and record the number of times this pattern repeats. In essence, we capture the set of path in iteration states that appear during a single loop visit. We then use this history to predict future loop visit behavior. Past research has shown that conditional branch behavior tends to be correlated [10]. Our own past work has also shown that the behavior of conditional branches leading up to a loop body (the path toloop) is a good indicator of the path in loop pattern that will be followed [1] With this in mind, we propose to predict and unroll entire loop executions based on path to loop ....

S.-T. Pan, K. So, and J. T. Rahmeh. Improving the Accuracy of Dynamic Branch Prediction Using Branch Correlation. In Proc. of the 5th International Conference on Architectural Support for Programming Languages and Operating System, pages 76--84, 1992.

....4.2 Correlation Profiling By correlation we mean a consecutive sequence of events. Correlation profiling is the counting of such correlations in a program. The concept of correlation has been used in various aspects of perfbrmance optimization in microprocessors. For example, Pan et al. [12] used branch correlation to im prove the accuracy of dynamic branch prediction. Their results show that, as compared with the traditional twobit counter based prediction scheme, the correlation based branch prediction achieves up to 11 additional accuracy. Mowry et al. 9] used control flow ....

....to be effective in modeling the extra energy Q as well. Using this idea, we rewrite the energy estimate as = ClR1 q C2R2 q . 11) where Rj s are the counters fbr the correlation events and cj s, again, are the regression coefficients to be determined. Different types of correlation events [9, 12] have been inves tigated previously in the context of branch and cache miss predictions. In this paper, we concentrate on one particular type, called the control flow correlation. The basic form of control flow correlation we use is similar to the form used in [9] We call it basic block ....

S. Pan, K. So, and J. T. Rahmeh. Improving the accuracy of dynamic branch prediction using branch correlation. In Proc. ASPLOS-V, pages 76 84, Oct. 1992.

....ability to exploit even wider issue capabilities. Indeed, Jouppi and Ranganathan [17] claim that branch prediction will be the most restrictive bottleneck in processors by 2010, worse even than stalls from the increasingly high performance memory system. In pursuing better prediction, two level [26], 41] and hybrid [24] predictors , which explicitly track prior branch history, have received special attention. Most of this attention has been focused on reducing aliasing errors (conflict mispredictions) which arise when unrelated branches happen to collide in a particular branch predictor ....

....and hybrid predictor with different predictor size 29 0 2 4 6 8 10 12 14 Average misprediction rate improvement by Mshare overHybr6 predictor Mshare Alloyed Predictor Fig. 19. Average misprediction rate improvement by mshare over hybrid predictor at different predictor size et al. [26] observed that a tradeoff exists between including history and address bits in the PHT index, and introduced GAs. Yeh and Patt [41] found that global history schemes suffer more from aliasing and are more strongly sensitive to both history length and address length than are localhistory schemes. ....

S.-T. Pan, K. So, and J. T. Rahmeh. Improving the accuracy of dynamic branch prediction using branch correlation. In Proceedings of the Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, pages 76--84, Oct. 1992.

.... v4 [27] 4 K entry is the same size as that in the Alpha 21064 [7] and is at the point of diminishing returns for bimodal predictors, although the 21164 used an 8 K entry predictor [8] The gshare predictor [17] shown in Figure 4a, is a variation on the two level GAg GAs global history predictor [18, 29]. The advantage of global history is that it can detect and predict sequences of correlated branches. In a conventional global history predictor (GAs) a history (the global branch history register or GBHR) of the outcomes of the most recent branches is concatenated with some bits of the branch PC ....

S.-T. Pan, K. So, and J. T. Rahmeh. Improving the accuracy of dynamic branch prediction using branch correlation. In Proceedings of the Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, pages 76--84, Oct. 1992.

....counter is a cell of a branch prediction table (BPT) which could be accessed in different ways. The simplest BPT index is a portion of the branch address. More complex two level predictors combine the branch address or its part with shift register representing the history of branch outcomes [5][6][7] Global two level predictors benefit from correlation between subsequent branches during program execution (global branch history) while local predictors are based on correlation between subsequent executions of the same branch (local branch history) Hybrid branch predictors can include both ....

S-T. Pan, K. So, J.T. Rahmeh, "Improving the Accuracy of Dynamic Branch Prediction Using Branch Correlation," ASPLOS V, 1992, pp. 76-84.

....counters. The Alpha 21164 [8] used a larger PHT of 8 K entries, but we conservatively choose the smaller PHT to make it more difficult to show benefits from decay. The gshare predictor [13] shown in the left hand portion of Figure 1, is a variation on the two level global history predictor [14, 19]. The advantage of global history is that it can detect and predict sequences of correlated branches. In a conventional global history predictor, a history (the global branch history register or GBHR) of the outcomes of the # most recent branches is concatenated with some bits of the branch PC to ....

S.-T. Pan, K. So, and J. T. Rahmeh. Improving the accuracy of dynamic branch prediction using branch correlation. In Proceedings of the Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, pages 76--84, Oct. 1992.

....a prediction scheme is as well as how much more we can improve the existing predictors. To establish a theoretical basis, we first introduce a conceptual system model to characterize components in a branch prediction process. Using this model, we notice that many of the best prediction schemes [Pan92, Yeh92, Yeh93, McFarling92, Chang94, Nair95b] use predictors similar to a two level adaptive branch predictor [Yeh91] Then, we demonstrate that these twolevel like predictors are, in fact, simplified implementations of an optimal predictor in data compression, Prediction by Partial Matching (PPM) Cleary84, Moffat90] This establishes a ....

....used for each prediction scheme. For the information processor, it describes both the information used by the selector and the way dispatcher maps information. Finally, the predictor used in each prediction scheme is also listed. From this table, we notice that many of the best prediction schemes [Pan92, Yeh92, Yeh93, McFarling92, Chang94, Nair95b] use Markov predictors. We will explain this further in Section 4. Source Information processor Prediction scheme modification predictor or profiling selector dispatcher forward not taken, no (address target) many to one constant backward taken one to one, mapped with 2 bit counter no ....

Pan, S-T., So, K. and Rahmeh, J. T. Improving the accuracy of dynamic branch prediction using branch correlation. Proceedings of the 5th International Conference on Architectural Support for Programming Languages and Operating Systems, 76- 84, 1992.

....to instruction uniqueness, an instruction s setting within the dynamic flow of the program is important. Previous research shows that capturing information along the path that leads to an instruction, combined with the PC, can provide a more distinguishable behavior pattern than using just the PC [25, 26, 27, 40, 41]. The tints are stored in trace format, providing the associated history with some context. Traveling Speculations also provide the opportunity to dynamically acquire per instruction information, a task sometimes relegated to static profiling. It has been demonstrated that many predictors are ....

S. T. Pan, K. So, and J. T. Rahmeh. Improving the accuracy of dynamic branch prediction using branch correlation. In 5th International Conference on Architectural Support for Programming Languages and Operating Systems, pages 76--84, Oct 1992.

....prediction is later found to be incorrect, the effects of the instructions executed after the branch are nullified, and execution continues along the correct path. Branch prediction techniques have now improved to the point of getting about 90 96 prediction accuracies for non numeric programs [9] [14] One problem that cannot be solved by highaccuracy predictions alone is that a branch s identity and target are not known until the branch instruction is decoded. If the prediction is fall through, then this is not a problem; if the branch is predicted to be taken, then this causes a ....

S-T. Pan, K. So, and J. T. Rahmeh, "Improving the Accuracy of Dynamic Branch Prediction Using Branch Correlation," Proceedings of 5th International Conference on Architectural Support for Programming Languages and Operating Sys- tems (ASPLOS V), pp. 76-84, 1992.

.... prediction approaches use both profiling and compile time analysis techniques to statically predict the branch instructions [2] Dynamic branch predictors use run time information to improve their predictions [3] One of the fundamental ideas is to exploit correlation between branch outcomes [4]. Recently, an approach for conditional branch folding for embedded applications has been proposed [5] It is based on the detection and resolution of a subset of short loop branches. In the customization arena, various approaches have been researched, but none that target the microarchitecture, ....

S. T. Pan, K. So and J. T. Rahmeh, "Improving the accuracy of dynamic branch prediction using branch correlation", in ASPLOS V, pp. 76--84, October 1992.

....by the branch address to provide the direction prediction. Figure 2 shows the finite state machine for incrementing the two bit counters and the bimodal predictor organization respectively. GAs and PAs are types of two level predictors proposed by Yeh and Patt [27] and also Pan, So and Rahmeh [18]. Two level predictors address the inability of bimodal predictors to identify branch pattern behavior or inter branch correlation. The GAs prediction scheme uses global history to make direction predictions. It consists of a global history register (GBHR) which maintains a history of the outcomes ....

S.-T. Pan, K. So, and J. T. Rahmeh. Improving the accuracy of dynamic branch prediction using branch correlation. In Proceedings of the Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, pages 76--84, Oct. 1992.

....a combination of both DHT BHT (DHT BHT) 1] can be classified as one level dynamic prediction schemes as the prediction depends only on the history of the branch under consideration. In schemes such as two level adaptive training branch prediction [2] 4] and prediction using branch correlation [3][4] a more complex approach is used in which the prediction is not only based on the past history of the branch under consideration but also the outcomes of recently executed branches in the instruction stream. We classify these as two level dynamic branch prediction schemes. We simulated all of ....

....of the branch under consideration but also the outcomes of recently executed branches in the instruction stream. 2.2. 1 Correlation Based Prediction (BPT) Pan, et al. assert that in many integer workloads, controlflows are complex thus correlating sequential branches in an instruction stream [3]. To take advantage of this observation, they propose correlation based prediction (BPT) Here the prediction not only depends on the branch address but also the path which was taken by the instruction stream to reach the branch. This path is recorded in a global shift register which selects the ....

S.-T. Pan, K. So, and J.T. Rahmeh, "Improving the accuracy of dynamic branch prediction using branch correlation," Fifth International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS-V), vol. 27, Boston, MA, USA, 1992, p. 76-84

....scheme, proposed by Smith [22] contains a table of two bit saturating counters or pattern history table (PHT) which is indexed by the branch address to provide the direction prediction. GAs and PAs are types of twolevel predictors proposed by Yeh and Patt [27] and also Pan, So and Rahmeh [18]. Two level predictors address the inability of bimodal predictors to identify branch pattern behavior or inter branch correlation. The GAs prediction scheme uses global history to make direction predictions. It consists of a global history register (GBHR) which maintains a history of the outcomes ....

S.-T. Pan, K. So, and J. T. Rahmeh. Improving the accuracy of dynamic branch prediction using branch correlation. In Proceedings of the Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, pages 76--84, Oct. 1992.

No context found.

S.T. Pan, K. So, and J.T. Rahmeh, "Improving the Accuracy of Dynamic Branch Prediction Using Branch Correlation," Proc. 5th Int'l Conf. Architectural Support for Programming Languages and Operating Systems (ASPLOS-V), IEEE CS Press, Los Alamitos, Calif., 1992, pp. 76-84.

No context found.

S. T. Pan, K. So, and J. T. Rahmeh. "Improving the accuracy of dynamic branch prediction using branch correlation" Proceedings of ASPLOS V, pages 76--84, Boston, MA, October 1992.

No context found.

S. T. Pan, K. So, and J. T. Rahmeh. "Improving the accuracy of dynamic branch prediction using branch correlation" Proceedings of ASPLOS V, pages 76--84, Boston, MA, October 1992.

No context found.

S.-T. Pan, K. So, and J. T. Rahmeh, "Improving the accuracy of dynamic branch prediction using branch correlation," in Proceedings of the 5th International Conference on Architectural Support for Programming Languages and Operating System (ASPLOSV) , pp. 76--84, 1992. 24

No context found.

S. T. Pan, K. So, and J. T. Rahmeh. "Improving the accuracy of dynamic branch prediction using branch correlation" In Proceedings of ASPLOS V, pages 76--84, Boston, MA, October 1992.

No context found.

S.-T. Pan, K. So and J. Rahmeh, " Improving the Accuracy of Dynamic Branch Prediction Using Branch Correlation," in Proceedings of the 5th International Conference on Architecture Support for Programming Languages and Operating Systems, pp. 76-84, October 1992.

No context found.

S.T. Pan, K. So, and J.T. Rahmeh. Improving the accuracy of dynamic branch prediction using branch correlation. In Proceedings of the 24th Annual ACM/IEEE International Symposium and Workshop on Microarchitecture, pages 76--84, October 1992.

No context found.

S. T. Pan, K. So, and J. T. Rahmeh, "Improving the Accuracy of Dynamic Branch Prediction using Branch Correlation," in Proceedings of the Symposium on Architectural Support for Programming Languages and Operating Systems, (Boston, MA, USA), pp. 12--15, October 1992.

No context found.

Pan ST, So K, Rahmeh JT. Improving the Accuracy of Dynamic Branch Prediction Using Branch Correlation. In Proceedings of the ASPLOS V, 1992; 76-84.

No context found.

S-T Pan, K. So, and J.T. Rahmeh, "Improving the Accuracy of Dynamic Branch Prediction Using Branch Correlation," Proc. of the 5 International Conference on Architectural Support for Programming Languages and Operating Systems, 1992, pp. 76-84.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC