Memory Model

Understanding GPU memory management in Mini-OpenCV.

Overview

Mini-OpenCV uses a RAII-based memory model for automatic GPU memory management.

DeviceBuffer

The DeviceBuffer class provides RAII-style GPU memory management:

class DeviceBuffer {
public:
    explicit DeviceBuffer(size_t size);
    ~DeviceBuffer();  // Automatically calls cudaFree
    
    void* data();
    size_t size() const;
    
private:
    void* d_data_;
    size_t size_;
};

Key Features

• Automatic Deallocation: Memory freed when object goes out of scope
• Move Semantics: Efficient transfer of ownership
• No Copy: Prevents accidental deep copies

GpuImage

GpuImage wraps a DeviceBuffer with image metadata:

class GpuImage {
public:
    int width() const;
    int height() const;
    int channels() const;
    size_t bufferSize() const;
    
    DeviceBuffer& buffer();
    
private:
    int width_, height_, channels_;
    std::unique_ptr<DeviceBuffer> buffer_;
};

Memory Pool

For performance-critical applications, memory pools reduce allocation overhead:

class MemoryPool {
public:
    DeviceBuffer* allocate(size_t size);
    void release(DeviceBuffer* buffer);
    
private:
    std::vector<std::unique_ptr<DeviceBuffer>> pool_;
};

Best Practices

1. Reuse Buffers: Keep GpuImage objects alive for multiple operations
2. Batch Operations: Process multiple images in a single pipeline
3. Use Pipelines: PipelineProcessor reuses internal buffers

Next Steps

• CUDA Streams - Async execution
• Design Decisions - Architecture choices