Architecture Overview

Mini-OpenCV uses a three-layer architecture designed for performance, modularity, and ease of use.

Three-Layer Design

Layer Responsibilities

1. Application Layer

The top-level API that users interact with:

Component	Purpose
ImageProcessor	Main entry point for image operations
PipelineProcessor	Chain multiple operations with async execution

2. Operator Layer

CUDA kernels implementing image processing algorithms:

Category	Operations	CUDA Technique
Pixel	Invert, grayscale, brightness	Per-pixel parallelism
Convolution	Gaussian blur, Sobel, custom kernels	Shared memory tiling
Histogram	Calculation, equalization	Atomic operations + reduction
Geometric	Resize, rotate, flip, affine	Bilinear interpolation
Morphology	Erosion, dilation, open/close	Custom structuring elements
Threshold	Global, adaptive, Otsu	Histogram-driven
Color Space	RGB/HSV/YUV conversion	Matrix operations
Filters	Median, bilateral, sharpen	Edge-preserving filters

3. Infrastructure Layer

Core utilities for GPU computing:

Component	Purpose
DeviceBuffer	RAII GPU memory management
GpuImage	Image container with GPU memory
CudaError	Error handling and checking
ImageIO	Image file I/O (JPEG, PNG, BMP)
StreamManager	CUDA stream pool for async execution

Data Flow

Memory Model

Zero-Copy Optimization

Key optimizations:

1. Lazy Allocation: Memory allocated on first use
2. Buffer Reuse: Memory pool for temporary buffers
3. Async Transfer: Overlap compute and transfer using CUDA streams

CUDA Stream Pipeline

Multi-stream execution enables overlapping operations:

Supported GPU Architectures

Architecture	Compute Capability	Example GPUs
Turing	SM 75	RTX 20 series, T4
Ampere	SM 80/86	A100, RTX 30 series
Ada Lovelace	SM 89	RTX 40 series, L4
Hopper	SM 90	H100

Next Steps

• Memory Model - Deep dive into GPU memory management
• CUDA Streams - Async execution details
• Design Decisions - Architecture decision records