Branch Prediction 1 Introduction 2 Questions

CS510: Advanced Computer Architecture
Programming Assignment 1
Branch Prediction
Due: Oct. 31, 2007 (24:00)
1 Introduction
For this assignment you use simplescalar (sim-outoder) to evaluate several branch prediction algorithms.
• Simulator: sim-outorder
• Benchmarks: basicmath, lame, stringsearch, pgp, FFT (use small size input)
1.1 Parameter Description
You can specify the branch predictor by -bpred options. You can find the options in the help message
(sim-outorder -h). Followings are the options you need to do this assignment.
• -bpred string (default: bimod)
branch predictor type {nottaken|taken|perfect|bimod|2lev|comb}
• -bpred:bimod int (default: 2048)
bimodal predictor config (table size)
• -bpred:2lev int int int {0|1} (default: 1 1024 8 0)
2-level predictor config (level1 table size, level2 table size, history register width, xor)
1.2 Sample Predictors
These are configuration examples for the predictors in the lecture slide “Dynamic ILP(II)”
• 2-bit scheme for BHT with 1K entries
-bpred bimod
-bpred:bimod 1024
• (2,2) predictor using lower 4 bits of branch address
-bpred 2lev
-bpred:2lev 1 64 2 0 (1 2 (4+2) 2 0)
• gshare predictor using lower 8 bits of branch address
-bpred 2lev
-bpred:2lev 1 256 8 1 (1 2 8 8 1)
2 Questions
1. Program profile
• Profile the ratio of conditional/unconditional branches to total instructions. (by dynamic count)

• Profile the breakdown of taken/not-taken branches by following categories. You may need to
customize sim-outorder.
taken
not-taken
2. Bimodal predictor
B - forward B - backward BL
• Show the performance variation according to the table size.
(Draw a graph: Table size as x-axis, IPC as y-axis)
• Compare the results.
3. Branch Target Buffer
If a branch predictor predicts a branch instruction will be taken, it looks up the target address of the
branch in its branch target buffer(BTB). The size and associativity of BTB also affects IPC. You can
specify BTB by [-bpred:btb set-size associativity (default: 512 4)]
• Show the performance variation according to the BTB size. Use direct-mapped BTB and 512-
entry bimodal predictor.
• Complete the below table for each benchmark. Use 512-entry bimodal predictor.
Sets Assoc. IPC Addr-rate Dir-rate
512 1
256 2
128 4
32 16
8 64
1 512
• Explain the result.
4. 2-level predictor: gshare
• Evaluate IPC of the system with gshare(5-bit) predictor.
• Find the GAg configuration (e.g., -bpred:2lev 1 2 k
k 0) that yields the similar IPC.
• Estimate and compare the resource cost of the above two predictors.
3 Hand in
When you finish this assignment, submit the report and source files to seonggun.kim[at]arcs.kaist.ac.kr.
The quality of your discussion on each task is the most important factor in grading. Every result requires
your proper explanation.
Appendix A.
1. Installing SimpleScalar-ARM
• Download the SimpleScalar-ARM source tarball(simplesim-arm-0.2.tar.gz) from the following web
page. http://www.simplescalar.com/v4test.html
• Install the SimpleScalar-ARM in your workstation or server. For further information, please refer
to the installation guide provided from http://www.simplescalar.com/docs.html

2. Installing ARM cross compiler
• You will have to create your own ARM binary files during this assignment. In order to do this,
you have to download and install an ARM cross compiler in your server.
• For your convenience, visit the following web page and download handy script that performs the
tedious process automatically. http://www.kegel.com/crosstool/
3. Using the SimpleScalar simulator
• Download the source code of MiBench from http://www.eecs.umich.edu/mibench/. It is an
embedded benchmark suite made up of the several applications from six different program groups.
In this assignment, use following benchmarks with small size input:
⊲ basicmath (Automotive and Industrial group)
⊲ lame (Consumer group)
⊲ pgp (Security group)
⊲ stringsearch (Office group)
⊲ FFT (Telecomm group)
• Build ARM executable binaries of MiBench with the optimization level 2(-O2). You should give
-static (static linking) option to run the applications on SimpleScalar.
Notice: if you have some problem with installing an ARM cross compiler, you may use
the arm binaries deployed by the distributers of MiBench or by TA.