Matrix transposition : from sequential to parallel with CUDA

The following repository contains all the material related to both the homeworks on Matrix Transposition assigned during the GPU computing course : University of Trento (Italy) a.y. 2023/2024.

To see the report and better understand what this work is about, click Here

How to use

Download the directory

git clone https://github.com/LuCazzola/cudaMatrixTranspose.git

Here follows the Hierarchy of relevant project's files :

.
├── bin                         # final executables
│    └── ...
├── obj                         # intermediate object files
│    └── ...
└── src                         # source code
│    ├── headers                # header files
│    │    └── ...                         
│    ├── benchmark.c            # produce an output file according to options in "run_benchmark.sh"
│    ├── benchmark_gpu.cu       # produce an output file according to options in "launch_benchmark.sh"
│    .
│    ├── main.c                 # test the functions according to options in "run_main.sh"
│    ├── main_gpu.cu            # test the functions according to options in "launch_main.sh"
│    .
│    ├── transpose.c            # functions to compute the transpose of a given matrix
│    ├── transpose_gpu.c 
│    .
│    ├── matrix.c               # definition of methods to handle matrices
│    ├── opt_parser.c           # command line parameter parsing
│    .
│    └── common_cuda.cu         # defines some common functions for cuda methods
│
├── run_benchmark.sh            # set parameters related to "benchmark.c" and run the script
├── run_main.sh                 # set parameters related to "main.c" and run the script
├── run_cache_benchmark.sh      # run cachegrind to benchmark cache miss % on specified function
│   .
├── launch_benchmark.sh         # set parameters related to "benchmark_gpu.cu" and run the script on SLURM system
├── launch_main.sh              # set parameters related to "main_gpu.cu" and run the script on SLURM system
│   .
├── data                        # data gathered via "run_benchmark.sh" & "launch_benchmark.sh"
│    └── ...
├── plot_data.py                # generates graphs using the data stored in "data" folder
│
├── Makefile
└── ...

Main commands

Makefile defines 4 rules :

• make : builds object files and homework-1 + homework-2 executables
• make debug : builds object files and ALL executables adding debugging flags
• make benchmark : builds object files and benchmark + benchmark_gpu executable
• make clean : cleans all object files

There are many pre-set scripts to choose from :
>> CPU scripts section ( Homework-1 )
>> GPU scripts section ( Homework-2 )

---

CPU test commands ( Homework-1 )

NOTE

Go first inside the repository before running the scripts

cd cudaMatrixTranspose

---

COMMANDS

"run_main.sh" script sets parameters related to homework-1 executable and runs it.
To change run parameters and have a better understanding of its functionalities see : run_main.sh

make
./run_main.sh

"run_benchmark.sh" script sets parameters related to benchmark executable and runs it.
extracted data can be found on the data folder
To change run parameters and have a better understanding of its functionalities see : run_benchmark.sh

make benchmark
./run_benchmark.sh

"run_cache_benchmark.sh" script sets parameters related to homework-1 and runs Cachegrind on it, extracting localized informations about cache misses inside transpose_naive() or transpose_blocks() functions (according to the chosen parameter "method")
To change run parameters and have a better understanding of its functionalities see : run_cache_benchmark.sh

make clean
make debug
./run_cache_benchmark.sh

---

GPU test commands ( Homework-2 )

NOTE

Please consider that the following commands are supposed to be ran on the Marzola DISI cluster, modify the launch_main.sh & launch_benchmark.sh scripts if needed to change partition or SLURM system.

Outside the cloned project folder upload the project's directory to the login node

scp -r cudaMatrixTranspose <YOUR USERNAME>@marzola.disi.unitn.it:/home/<YOUR USERNAME>

Then login and go inside the project's folder

cd cudaMatrixTranspose
module load cuda

---

COMMANDS

"launch_main.sh" script sets parameters related to homework-2 executable and runs it.
To change run parameters and have a better understanding of its functionalities see : launch_main.sh

make
sbatch launch_main.sh

To visualize the results, once the node returns do:

cat output.out

"launch_benchmark.sh" script sets parameters related to benchmark_gpu executable and runs it.
extracted data can be found on the data folder
To change run parameters and have a better understanding of its functionalities see : launch_benchmark.sh

make benchmark
sbatch launch_benchmark.sh

---

Graph Plotting

Inside the project's directory there's also a python script which take's the content of data folder and generates 2 types of graphs

• x : Matrix size - y : Mean execution time
• x : Matrix size - y : Mean effective bandwidth

Test it by running (on you own device) :

python3 plot_data.py

You can customize what information to plot inside the script

---

Extra Customization

It's also possible to change some other parameters at compilation level (optimization level and matrix element data type) by changing some variables in the makefile) :