Parallel Sobel Edge Detection (MPI, OpenMP, CUDA, Hybrid)

This project explores parallel implementations of the Sobel edge detection algorithm using multiple paradigms:

• MPI for distributed-memory CPU systems
• OpenMP for shared-memory CPU systems
• CUDA for GPU acceleration
• Hybrid MPI + OpenMP for node-level and core-level parallelism

---

Objective

The goal of this project is to analyze and compare parallel programming models for 2D image processing using Sobel edge detection.

We aim to:

• Apply and evaluate parallel strategies (MPI, OpenMP, CUDA)
• Compare execution performance across different architectures (multicore CPU vs GPU)
• Demonstrate hybrid CPU-based parallelism with MPI + OpenMP

---

Methodology

Sobel Edge Detection

We compute the image gradient magnitude using the Sobel operator, a common method for edge detection.

Parallelization Strategies

• MPI: Distribute image rows with boundary overlap, using MPI_Scatterv and MPI_Gatherv for data transfer.
• OpenMP: Use #pragma omp parallel for to parallelize row-wise Sobel computation.
• CUDA: Launch a 2D grid of GPU threads for pixel-wise Sobel filtering using coalesced memory access and shared memory.
• Hybrid: Each MPI process handles a chunk of the image and applies OpenMP for multi-threaded processing within its chunk.

---

Implementation Details

Directory Structure

parallel-edge-detection/
├── cuda/                # CUDA implementation
├── data/                # Input images
├── hybrid/              # Hybrid MPI + OpenMP
├── include/             # Header files (e.g., stb_image)
├── mpi/                 # MPI implementation
├── openmp/              # OpenMP implementation
├── output/              # Output images
├── README.md            # Documentation

Source Files

• mpi/mpi_sobel.c
• openmp/omp_sobel.c
• cuda/cuda_sobel.cu
• hybrid/hybrid_sobel.c

---

Build Instructions

Use the following commands to build each variant.

MPI

cd mpi
mpicc -std=c99 -O2 mpi_sobel.c -o mpi_sobel -lm

OpenMP

cd openmp
gcc -fopenmp -std=c99 -O2 omp_sobel.c -o omp_sobel -lm

Hybrid (MPI + OpenMP)

cd hybrid
mpicc -fopenmp -std=c99 -O2 hybrid_sobel.c -o hybrid_sobel -lm

CUDA

cd cuda
nvcc cuda_sobel.cu -o cuda_sobel

---

Execution Examples

MPI

mpirun --mca btl self,vader -np 4 ./mpi_sobel ../data/input_2048.jpg

OpenMP

export OMP_NUM_THREADS=4
./omp_sobel ../data/input_2048.jpg

Hybrid

export OMP_NUM_THREADS=4
mpirun --mca btl self,vader -np 2 ./hybrid_sobel ../data/input_2048.jpg

CUDA

./cuda_sobel ../data/input_2048.jpg

---

Benchmarking

Each implementation includes a benchmarking script:

• run_mpi_exp.sh
• run_omp_exp.sh
• run_cuda_exp.sh
• run_hybrid_exp.sh

To execute all:

bash run_mpi_exp.sh
bash run_omp_exp.sh
bash run_cuda_exp.sh
bash run_hybrid_exp.sh

---

Results and Analysis

• report.md: Summary
• Output images for verification:
  • output_mpi.jpg
  • output_omp.jpg
  • output_cuda.jpg
  • output_hybrid.jpg

---

Dependencies

• OpenMPI
• GCC with OpenMP
• CUDA Toolkit
• stb_image.h, stb_image_write.h

---

Clean Up

To clean the workspace:

rm mpi/mpi_sobel openmp/omp_sobel cuda/cuda_sobel hybrid/hybrid_sobel\
    mpi/*.log openmp/*.log cuda/*.log hybrid/*.log\
   output/output_*.jpg