Gaussian-Blur

Parallel computing of Gaussian Blur for imaging processing

Several different methods

• fork
• OpenMP
• MPI
• OpenACC(not finished)

Run the code

Put the input images, code and Makefile in one directory, four size of input images are provided(500,1000,1500,2000)

make

The first parameter is the calcuation radius, the second is the size of input image

./serial 1 500 

./openmp 1 500

mpiexec -n 1 mpi 1 500

The first parameter is the calcuation radius, the second is the number of processes ,the third is the size of input image

./fork 1 1 500

Submit the code to HPC

Use SFTP to upload all file to HPC, the output will be generated as result.txt

sbatch output.slurm

Optional : The runtime of one single Slurm job is too long. Addtional Slurm is also provided

sbatch mpi.slurm
sbatch openmp.slurm

Output

The output images are stored in dir ./Output Images

For example, g1 means the calculation radius is 1

The runtimes are stored in dir ./Result

• result.txt is the total runtime
• optional: mpi.txt and openmp.txt is the result if you submit MPI or OpenMP slurm script only

MapReduce using HIPI

Setup Environment

Environment : Cloudera QuickStart VM

Setup HIPI

http://hipi.cs.virginia.edu/gettingstarted.html

Run the code

Put the examples folder inside hipi

Update setting.gradle

include ':core', ':tools:hibImport', ... ':examples:covar', ':examples:helloWorld'

Create input image in your home dir and put test images in

mkdir SampleImages 

tools/hibImport.sh ~/SampleImages images.hib

Compile the code

cd examples/helloWorld

gradle jar

hadoop jar build/libs/helloWorld.jar images.hib output1

See the result, the result will be average time of processing images using Gaussian Blur

hadoop fs -cat ouput1/part-r-00000

Sample code

The serial verison of C code is basd on this repo https://github.com/whdcumt/Gaussian_Blur/blob/master/code.c

The Java code of HIPI is based on sample code of computing the average pixel color from http://hipi.cs.virginia.edu/gettingstarted.html