Image Processing Pipeline: Architectural Profiling & Parallelization

Overview

This project implements and benchmarks three parallelization strategies for a computationally intensive image processing pipeline (Gaussian blur + Sobel edge detection):

1. Sequential Baseline - Single-threaded reference implementation
2. CPU Multi-threading - OpenMP with block, tiled, and SIMD decompositions
3. GPU Acceleration - CUDA with shared memory optimization

Repository Structure

• src/ - source code for sequential, OpenMP, and CUDA pipelines
• include/ - shared headers and pipeline declarations
• datasets/input/ - input image data for experiments
• datasets/output/ - generated output images and intermediate data
• build/ - local build artifacts created by CMake
• results/ - benchmark CSV and summary files
• plots_*/ - generated performance plots
• report_*/ - generated analysis reports and visualizations
• scripts/ - helper scripts for plotting and profiling

Input and Output Data

• Input datasets are stored in datasets/input/.
• Processed images and benchmark outputs are written to datasets/output/ or the generated report directories.
• This repository tracks the source code, while generated artifacts such as build files, temporary .ppm files, benchmark plots, and report outputs are ignored by .gitignore.

Report Contents

The full report output is saved under directories named report_<timestamp>/. Each report directory typically contains:

• report.html - human-readable benchmark summary and analysis
• summary.txt - plain-text performance summary
• performance_data.txt - raw performance measurements
• images/ - sample output images from the pipeline
• plots/ - generated benchmarking plots and charts

How to Run

1. Build the project with CMake or make.
2. Run the sequential, OpenMP, or CUDA pipeline targets.
3. Use generate_report.sh to produce the report and plots.
4. Inspect generated results in report_<timestamp>/ and results/.

Demo Run and Results

A sample demo was run using the repository inputs and generated output images plus a full report.

• Input images used:
  • datasets/input/4k.jpg
  • datasets/input/8k.jpg
  • datasets/input/test_4k.ppm
  • datasets/input/test_8k.ppm
• Generated output images:
  • datasets/output/4k_seq.jpg
  • datasets/output/4k_cuda.jpg
  • datasets/output/4k_omp_16.jpg
  • datasets/output/8k_seq.jpg
  • datasets/output/8k_cuda.jpg
  • datasets/output/8k_omp_16.jpg
• Result image examples from the report:
  • report_20260620_153314/images/4k_cuda.jpg
  • report_20260620_153314/images/8k_cuda.jpg

Report Output

The demo produced a full report in report_20260620_153314/, including:

• report_20260620_153314/report.html
• report_20260620_153314/summary.txt
• report_20260620_153314/performance_data.txt
• report_20260620_153314/images/
• report_20260620_153314/plots/

Output Folder Contents

A selection of generated images and plots was copied into output/ for easier review:

• output/4k_cuda.jpg
• output/8k_cuda.jpg
• output/4k_omp_16.jpg
• output/8k_omp_16.jpg
• output/test.jpg
• output/cpu_vs_gpu.png
• output/efficiency_heatmap.png
• output/execution_time.png
• output/speedup_comparison.png
• output/strong_scaling.png

Demo Result Summary

The sample run showed strong acceleration from the parallel versions:

• CUDA speedup: 4.17x for 8K input, 1.24x for 4K input
• Best OpenMP speedup: 4.98x with 32 threads (8K), 4.79x with 32 threads (4K)
• OpenMP (16 threads) speedup: 4.35x (8K) and 3.84x (4K)