repo759-final-project

Team Members

1. Arjun Anil
2. Aditi Chintawar

Proposed project brief

A study of how Otsu's method for Image Binarization can be sped up using CUDA and OpenMP multi-threading techniques.

---

Getting Started

Important

All SLURM scripts in this repository target the euler04 node on the instruction partition. You must have access to the instruction partition to submit jobs as-is. If you are running on a different cluster or partition, see Running on a different node or partition.

1. Generate the dataset

From the repo root:

cd dataset
make generate-dataset
cd ..

This compiles and runs the dataset generator, placing binary image files under dataset/out/.

2. Run the implementations

Each implementation has its own SLURM script. cd into the respective directories and queue using sbatch

# p1 -- single-threaded baseline
cd p1-baseline && sbatch slurm-p1.sh && cd ..

# p3 -- OpenMP parallel (scales 1, 2, 4, 8 threads)
cd p3-openmp && sbatch slurm-p3.sh && cd ..

# p4 -- cuBLAS
cd p4-cublas && sbatch slurm-p4.sh && cd ..

# p5 -- CUDA Thrust
cd p5-thrust && sbatch slurm-p5.sh && cd ..

# p6 -- CUDA CUB
cd p6-cub && sbatch slurm-p6.sh && cd ..

The slurm scripts will compile the required binary internally so no need to run an extra compilation/make command.
Output and error logs are written to <pN>.out / <pN>.err inside the respective implementation directory.

---

Running on a different node or partition

Each SLURM script contains two directives that tie it to a specific node and partition:

#SBATCH --partition=instruction
#SBATCH --nodelist=euler04

To run on a different setup, open the relevant script and:

• Change the partition -- replace instruction with your target partition name, or remove the --partition line entirely to use your cluster's default.
• Change or remove the node constraint -- replace euler04 with your target node name, or remove the --nodelist line entirely to let the scheduler choose any available node.

Alternatively, SLURM allows command-line flags to override script directives:

sbatch --partition=<your-partition> --nodelist=<your-node> p1-baseline/slurm-p1.sh