⚡ High-Performance Image Convolution Engine (CUDA)

"635x Faster than CPU." A bare-metal CUDA implementation of Gaussian Blur using Shared Memory Tiling and Memory Coalescing.

🎯 The Engineering Challenge

In high-performance computing (HPC) and Computer Vision, Memory Bandwidth is the bottleneck. A naive GPU implementation reads from VRAM (Global Memory) redundantly, choking the memory controller.

This project implements a Tiled Convolution Kernel that manually manages the L1 Cache (Shared Memory) to reduce Global Memory reads by 90%.

🛠️ Tech Stack

• Language: C++ (Host), CUDA C (Device)
• Optimization: Shared Memory Tiling, Loop Unrolling
• Libraries: None (Built from scratch using stb_image for I/O)
• Hardware: Optimized for NVIDIA Pascal/Ampere/Ada Architectures

---

📊 Performance Benchmarks

Tested on a 4K Image (3840 x 2160 pixels).

Implementation	Execution Time	Speedup vs CPU	Engineering Note
CPU (Single Thread)	947.05 ms	1x	Baseline. Loop-heavy, cache-inefficient.
GPU (Naive Global Mem)	1.79 ms	529x	Massively parallel, but Memory Bound.
GPU (Shared Mem Tiled)	1.49 ms	635x	Optimized. Uses Shared Memory to handle "Halo" pixels.

---

🧠 Key Optimizations Implemented

1. Shared Memory Tiling

Instead of every thread reading 9 pixels from VRAM (Global Memory), threads in a block cooperate to load a 16x16 Tile (plus border) into Shared Memory once.

• Result: Reduced Global Memory transactions by ~9x.

2. Handling the "Halo" (Ghost Cells)

Convolution requires neighbor pixels. Threads at the edge of a block need data from the next block.

• Solution: Implemented "Apron" loading logic where edge threads perform double-duty to load the halo into the Shared Memory cache.

3. Memory Coalescing

Input data is loaded in a coalesced manner (consecutive threads read consecutive memory addresses) to maximize memory bus utilization.

---

🏃 Quick Start

Prerequisites: NVIDIA GPU, CUDA Toolkit 11+.

1. Clone:

   git clone https://github.com/iamartyaa/cuda-convolution.git
   cd cuda-convolution

2. Compile:

   nvcc cuda_blur.cu -o gpu_blur

3. Run:

   ./gpu_blur

---

📈 Visualizing the Speedup

Run the included Python visualization script to see a simulated side-by-side comparison:

python visualize_video.py

(This renders a real-time comparison window showing the scanline difference between CPU and GPU speeds.)

---

Author

Amartya Yadav GPU & High-Performance Computing Enthusiast