Sobel Edge Detection using CUDA

This project implements Sobel Edge Detection on both CPU and CUDA-enabled GPU, and provides a detailed performance analysis including FPS, frame time, throughput, speedup, and roofline modeling using NVIDIA Nsight tools.

The goal is to demonstrate how GPU acceleration drastically improves performance for data-parallel image processing workloads, especially at high resolutions (up to 8K images).

---

📁 Project Structure

EDGE_DETECTION_USING_CUDA/
│
├── .venv/                  # Python virtual environment (plotting)
├── .vscode/                # VS Code settings
│
├── benchmarks/
│   ├── Plots/              # Generated performance plots
│   │   ├── fps_comparison.png
│   │   ├── frame_time_comparison.png
│   │   ├── speedup_vs_size.png
│   │   └── throughput_scaling.png
│   │
│   └── results/            # Benchmark outputs & profiling artifacts
│       ├── 512x512.png
│       ├── 1024x1024.png
│       ├── 1920x1080.png
│       ├── 5824x3264--6k.png
│       ├── 7680x4320--8k.png
│       ├── cpu_gpu_comparison.csv
│       ├── profile_nsight.png
│       ├── roofline_analysis.png
│       ├── roofline_report.ncu-rep
│       └── sobel_timeline.nsys-rep
│
├── build/                  # Build directory (CMake)
│   ├── edge_detect         # Executable
│   ├── CMakeFiles/
│   ├── CMakeCache.txt
│   └── Makefile
│
├── data/
│   ├── input/              # Input images
│   │   ├── lena.png
│   │   ├── kid.png
│   │   ├── city.png
│   │   ├── city-view.png
│   │   └── trade-center.png
│   │
│   └── output/             # Output images (CPU & GPU)
│       ├── 512×512/
│       ├── 1024×1024/
│       ├── 1920×1080/
│       ├── 5824×3264--6k/
│       └── 7680×4320--8k/
│
├── src/
│   ├── host/
│   │   ├── main.cpp        # Entry point & benchmarking
│   │   ├── sobel_cpu.cpp   # CPU Sobel implementation
│   │   └── sobel_cpu.hpp
│   │
│   ├── kernels/
│   │   ├── sobel_cuda_naive.hpp
│   │   └── sobel_shared.cu # Optimized CUDA kernel
│   │
│   └── utils/
│       ├── plot_results.py # Python plotting script
│       ├── cpu_gpu_comparison.csv
│       ├── fps_comparison.png
│       ├── frame_time_comparison.png
│       ├── speedup_vs_size.png
│       └── throughput_scaling.png
│
├── CMakeLists.txt
├── report.pdf              # Final performance report
└── README.md

---


## ⚙️ Requirements

### Hardware
- NVIDIA GPU with CUDA support

### Software
- **Ubuntu / WSL2 (recommended)**
- CUDA Toolkit (>= 11.x)
- CMake (>= 3.10)
- OpenCV (for image I/O)
- NVIDIA Nsight Systems & Nsight Compute (for profiling)

Check CUDA installation:
```bash
nvcc --version

---

🛠️ Build Instructions

From the project root:

mkdir build
cd build
cmake ..
make -j

This will generate the executable:

./edge_detect

---

▶️ How to Execute

Run Sobel on an image

./edge_detect ../data/input/lena.png

You can replace lena.png with any image:

./edge_detect ../data/input/kid.png
./edge_detect ../data/input/city.png
./edge_detect ../data/input/city-view.png
./edge_detect ../data/input/trade-center.png

---

📊 Runtime Output Explained

The program reports averaged metrics for both CPU and GPU:

• Warm-up runs
• Measured runs
• Average frame time (ms)
• FPS
• Throughput (MPixels/sec)
• Output correctness check

Example:

CPU Frame Time : 96.21 ms
GPU Frame Time : 0.502 ms
Speedup        : ~191x
CPU and GPU outputs MATCH ✓

---

🔬 Profiling & Analysis

Nsight Systems (Timeline)

nsys profile ./edge_detect ../data/input/city-view.png

Nsight Compute (Kernel Metrics & Roofline)

ncu ./edge_detect ../data/input/city-view.png

Key metrics analyzed:

• Achieved Occupancy (~91%)
• Kernel Runtime (~1.65 ms)
• Memory Throughput (~31.4 GB/s)
• Roofline utilization (~63%)

---

📈 Results Summary

• GPU achieves up to 246× speedup over CPU
• GPU frame time remains under 9 ms even for 8K images
• CPU throughput saturates (~20 MPixels/sec)
• GPU shows near-optimal occupancy and balanced compute/memory behavior

Full analysis available in the report: 📄 Report.pdf

---

🎓 Educational Value

This project demonstrates:

• CUDA kernel design
• Memory vs compute trade-offs
• Performance scaling
• Roofline modeling
• Professional GPU profiling methodology

---

🚀 Future Improvements

• Shared memory tiling
• Constant memory for Sobel masks
• Kernel fusion
• Multi-stream execution
• FP16 / Tensor Core exploration

---

👤 Author

Mohammad Salik Dev
CUDA & GPU Computing Enthusiast

---

If you have questions or want to extend this project, feel free to explore and experiment!