[Architecture] quantitative evaluation of the performance of the transfer Estimator

For the design of several schedulers, see the previous blog (here)

Quantitative Evaluation of prediction accuracy of several schedulers

The analysis and evaluation process of transfer prediction is mainly based on the trace of the given command sequence. The input text file will provide a command sequence trace, which contains N rows, each row contains three data, the commands are the current command PC, the next command PC, and the current command type (0 indicates the transfer command, and 1 indicates other commands). The following is a specific example:

0x0EA31410 0x0EA31414 10x0EA31414 0x0EA31470 00x0EA31470 0x0EA31474 10x0EA31474 0x0EA31478 00x0EA31478 0x0EA3147C 1

The GCC. Log, compress. Log, craft. Log, Gzip. Log, organized. Log, parser. log, and VPR. Log test data are used in the experiment. The test results are as follows:

Based on the example in the book, draw the result of the inaccuracy of the three schedulers:

The experiment results clearly show that the tournament prediction tool has a better prediction structure. On the other hand, under the same space limit, using the [] associated prediction tool with the relevant branch prediction information is not as good as simply predicting local historical tables. In fact, the branch history table can be viewed as a [1, 2] Join estimator. Analysis may be caused by excessive branch information. Reducing the number of related branches may improve the prediction accuracy. In this experiment, [2, 2], [4, 2], [6, 2], and [8, 2] (the corresponding addresses are respectively 12, 10, 8, and 6 bits, and the space is 32 K ), the result is as follows:

We can see that the average accuracy of the related branch estimator is the highest. Over-increasing the number of related branches may reduce the prediction accuracy.

Experiment conclusion

The 2-bit branch History Table forecaster can provide better branch prediction accuracy. Using Global Information and Related Branch Prediction Technologies can also improve the accuracy of prediction branches. By combining local information and global historical information, the tournament predicator provides the highest average prediction accuracy in the same space.

(Reprinted please indicate the author and Source: http://blog.csdn.net/xiaowei_cqu is not allowed for commercial use)