5-point stencil

This code implements a weighted 5-point stencil on a rectangular grid. The value in each cell of the grid is updated based on an average of the values in the neighbouring North, South, East and West cells.

The grid is initialised into a checkerboard pattern, with each square of the checkerboard being 64x64 pixels. The stencil operation reads from one grid and writes to a temporary grid. The stencil is run twice for every iteration, with the final result being held in the original array. The results are quantised to integers in the range [0,255] and output as a binary image file that may be viewed graphically.

The only output of each run is the runtime of the iteration loop of the program. Initialisation and output are not timed.

Building and running the code

A simple Makefile is provided to build the code using the GCC compiler. Just type make to build the code. A binary named stencil is produced on a successful build.

A job script for BlueCrystal Phase 4 is also provided in stencil.job. This will request time on 1 node of the veryshort queue and execute the stencil binary on one of the input problems. Submit this to the queue with the following command:

sbatch stencil.job

There are three input problems tested, representing different grid sizes. The inputs are all set as command line arguments to the executable in the following format:

./stencil nx ny niters

The inputs required are:

nx	ny	niters	Command	Serial Time on BCp4
1024	1024	100	./stencil 1024 1024 100	5.908341s
4096	4096	100	./stencil 4096 4096 100	130.196475s
8000	8000	100	./stencil 8000 8000 100	561.118133s

Checking results

The program will have executed correctly if the output image matches the provided reference output images with a small tolerance of +/- 1. A Python check script is provided to check the values.

python check.py --ref-stencil-file stencil_1024_1024_100.pgm --stencil-file stencil.pgm

If any errors are found, the script can be rerun with the addition of the --verbose flag to provide information about where the errors occur in the grid.

The reference input files for the different problems are named:

nx	ny	niters	Reference file
1024	1024	100	stencil_1024_1024_100.pgm
4096	4096	100	stencil_4096_4096_100.pgm
8000	8000	100	stencil_8000_8000_100.pgm

MPI Advice

For the MPI assignment you'll want to start running stencil over multiple cores and over multiple nodes. To do this on BCp4 you need to change the parameters at the top of the stencil.job job file as follows:

• --nodes controls the number of nodes you want to run on
• --ntasks-per-node controls the number of MPI processes per node

If these values are correctly set, then srun will pick up on them and run the required number of processes for your job.

Important: For Intel MPI to work with srun you need the following line in the job submit script:

export I_MPI_PMI_LIBRARY=/usr/lib64/libpmi.so

More info: https://www.acrc.bris.ac.uk/protected/bc4-docs/scheduler/index.html#running-parallel