Description
1 Objectives
The objective is to get you started with gpgpu-sim simulator. This assignment
is meant to familiarize you with the simulator for PA3 and final project.
3 Description
Each task are independent, when you finish one task, please comment
the code you have modified firstly before you move to the next.
3.1 Functional Simulator
3.1.1 Task1: Instruction semantic
This task requires you to modify the functional simulator. You need to change
the semantics of float(32-bits) point add instruction to power operation. Suppose you have a float point add instruction fd = fa + fb. The register fa and
fb store 32-bits float point value. fa =2.0, fb=3.0, the original add instruction
will generate the new value 5.0 and store it to fd. Now the new semantic should
generate the value 8.0=2.0
3.0 and store it to fd. For this task, the test case
is vectorAdd, we have uploaded one vector add CUDA program on moodle.
1
Please use this one to test your code. The original instruction semantic should
generate the results as follows.
You can verify this result(original semantic) by run the vector add code on
Hydra.
The new instruction semantic should generate the results shown in the following figure.
You can verify the new semantic by running your code on GPGPU-SIM.
We will use two test case, one is vectorAdd provided on Moodle, another
one would be similar to vectorAdd but used as secrete test case.
Hints, for this task, you need to read the code in instructions.cc under the
cuda-sim directory carefully and understand how it works. In your report, have
a short explanation about your understanding and where you have modified.
3.2 Performance/Timing simulator
This task you need to work on performance/timing simulator. The maximum
instructions you should simulate is 100000000 for the following tasks. For
those benchmarks have less than 100000000 instructions, just let it run until it
finished.
3.2.1 Task2: Benchmark performance study
In this task, you need to chosse 6 benchmarks from ispass-2009 and run them
with SM6 TITANX and SM7 TITANV configurations and plot the IPCs of these
benchmarks. This study is helping you get familier with benchmarks cause you
will use these benchmarks in your final project.
Please download the compiled benchmarks on moodle. To run the benchmarks,
you can follow the instruction listed here.
a. create ”test” directory under gpgpusim (mkdir test).
b. copy the SM6 TITANX and SM7 TITANV to ”test” directory.
c. copy the benchmark to ”test” directory.
d. under each benchmark, create two soft links that point to gpgpusim.config
and config XX islip.icnt under SM6 TITANX and SM7 TITANV(using ln command), or just copy these two files to each benchmarks directory. e. run the
benchmark, please pay attention to the benchmark input, the input data has
been provided with the benchmarks. Do not delete these files.
Graph: Plot IPCs for each benchmarks, X-axis: benchmarks, Y-axis: IPCs.
You should have two lines for SM6 TITANX and SM7 TITANV respectively.
Have a short discussion about your plots.
2
In this task, you need to use the 5 benchmarks from ISPASS.zip and run them
3.2.2 Task 3: Branch instructions and divergence
For this task, you need to modify the performance/timing simulator code by
adding counters to report the following statistics.
(1)total number of warps that executed conditional branch instructions and have
divergence.
(2)total number of warps that executed conditional branches, no matter they
have divergence or not.
Hints: For this task, please read the scheduler unit in shader.cc and the simt stack
in abstract hardware model carefully and understand how it works. The benchmark you should use for this task is LPS. You need to dump out the counters
and report the counters in your report. The format you dump out the counters
is shown as follows
3.2.3 Task 4: Memory access space
For this task, you need to modify the performance/timing simulator code by
adding a counters to report the following statistics.
(1)The total number of global memory access.
(2)The total number of local memory access.
Hints:For this task, you need to read the code about load-store unit in share.cc
carefully and understand how it works. You need to dump out the counters and
report the counters in your report. The benchmark you should use for this task
is LPS. The format you dump out the counters is shown in the following figure.
4 Submission
(1)Create a directory named ”cuda-sim” and copy the files you modified for
task1 to this directory.
(2)Create a directory namde ”gpgpu-sim”, then create two directories named
”task3” and ”task4” and copy the files you modified for task3 and task4 respectively
(3)Write the report, for task1, task3 and task4, please have a short discussion
about where you modified. For task2, including all the plots and your analysis
in your report.
(4)zip ”cuda-sim”, ”gpgpu-sim” and your report as unity id.zip
3
For this task, you need to read the code about load-store unit in shader.cc
5 Grading
(1)task1: 25 points, 10 points for each test case(totally 2 test case), 5 points for
discussion in your report
(2)task2: 20 points, 15 points for the plots, 5 points for discussion
(3)task3: 30 points, 5 points for compiling your code successfully, 20 points for
dumping out the counters correctly, 5 points for reporting the counters in your
report.
(4)task4: 25 points, 5 points for compiling your code without any errors, 15
points for dumping out the counters correctly, 5 points for reporting the counters
in your report.
4