Assignment #2, Due April 30th

For this assignment, you will use SimpleScalar to examine the performance of an application called ADPCM, and the potential performance gains if the computation kernel could be speed up by having a custom accelerator or co-processor.

Accounts and Machines

Everyone should have an account on the CSE architecture lab. See the instructor if this is not the case. This account will let you log into the following two machines in our lab (if your job takes a long time to run, please nice it at -19):

cycletron20.ucsd.edu
cycletron21.ucsd.edu

These are the two machines you can use to complete this assignment.

Software

To complete the assignment you'll need to use the following software:

ARM gcc, objdump, etc... The compiler tools are located at:
/home/cse291-sp02/project/arm/bin/
ARM SimpleScalar simulator. The source code is located at:
/home/cse291-sp02/project/arm/simplesim-arm/
The APDCM application to be optimized. The source code is located in:
/home/cse291-sp02/project/assign2/adpcm

You should put the compiler bin directory in your path, and copy the source tree for ARM SimpleScalar to your own directory for your SimpleScalar modifications. Compiling SimpleScalar will produce the ARM simulator called sim-outorder.

You will need to compile ADPCM with the ARM gcc compiler, and compile simplescalar with the linux x86 gcc compiler (/usr/local/bin/gcc). To run ADPCM on simplescalar execute the following command:

../sim-outorder -config xscale.cfg rawdaudio < clinton.adpcm > clinton.pcm

This will decode the clinton.adpcm file producing a PCM file. You will use the above configuration, program and input for all of the results for this assignment.

The xscale.cfg file is found in:
/home/cse291-sp02/project/arm/simplesim-arm/config/

The clinton.adpcm file is found in:
/home/cse291-sp02/project/assign2/adpcm/data

Questions to Answer

Once you are able to run sim-outorder on the ADPCM program, you will then need to modify SimpleScalar to answer the following questions:

How many cycles (using SimpleScalar) does it take to execute the above program with its input ("rawdaudio < clinton.adpcm > clinton.pcm") on an Xscale processor?
How many cycles from SimpleScalar are spent inside each call to routine "adpcm_decoder" on average, when running the rawdaudio with the above input? "adpcm_decoder" is the kernel in which all of the decoding is performed. You can use the program "arm-linux-objdump" provided with the ARM compiler to find out the static PCs for procedure calls and the starting and ending PCs of routines. Print the average number of cycles per call, and the total number of cycles spent in adpcm_decoder using the "stat_reg_counter" routine in SimpleScalar.
How many times does the routine "adpcm_decoder" get called?
How many loop iterations does the for loop in "adpcm_decoder" execute on average per call to the routine "adpcm_decoder"?
Assume you have built a coprocessor to specialize the execution of the routine adpcm_decoder. This coprocessor takes as input the params to adpcm_decoder and returns a buffer of results. Also assume that you've done a detailed analysis of the timing for this accelerator. This shows that each tall to the routine "adpcm_decoder" will execute 50% faster on the customized co-processor than on the XScale processor. How many cycles does it take to execute the program adding this co-processor to Xscale? What is the speedup for the execution of adpcm for this design?

What to turn in:

Answers to the above questions.
The 2-3 pages showing what code you changed/added in SimpleScalar, and where you made the modifications.

April 18, 2002