HIP Image FX

HIP Image FX is a high-performance, GPU-accelerated image processing toolkit built with HIP, supporting both AMD (ROCm) and NVIDIA (CUDA via hipcc) GPUs.
It features automatic kernel configuration tuning to maximize performance across hardware, with a transparent CPU fallback for comparison and validation.

The project is designed as both a practical image-processing tool and a reusable framework for GPU kernel autotuning in real-world workloads.

About the Project

This project was built to explore and demonstrate:

• Cross-vendor GPU compute using HIP
• Where GPU acceleration meaningfully outperforms CPU execution
• Automatic kernel tuning strategies for real workloads
• Writing portable, testable, and benchmark-driven C++20 systems

Rather than focusing on a single optimized kernel, HIP Image FX emphasizes reusability, measurement, and adaptability across different GPU architectures.

Features

• GPU-accelerated image filters:
  • grayscale
  • negative
  • gaussian-blur
• Automatic GPU configuration tuning
  (~12–18% speedup with transparent caching)
• Batch processing for image directories
• CPU fallback path for comparison and testing
• Reusable autotuning framework for custom HIP kernels
• CLI-driven workflow suitable for scripting and automation

Documentation

• Complete Documentation – Full guides and references
• Autotuning Guide – Framework design, API, and usage
• Testing Guide – Test structure and execution

Quick Start

Prerequisites

• AMD GPUs: ROCm 5.0+ with HIP runtime
• NVIDIA GPUs: CUDA 11.0+ (HIP compiled via hipcc)
• Meson build system
• C++20-compatible compiler

Build

meson setup build --native-file native/hip.ini --reconfigure
ninja -C build

Usage

# Single image
./build/hip-img-fx --input photo.jpg --output result.jpg --filter grayscale

# Batch processing (directory)
./build/hip-img-fx --input images/ --output results/ --filter gaussian-blur --batch-size 64

# CPU fallback
./build/hip-img-fx --input photo.jpg --output result.jpg --filter negative --use-cpu

Run ./build/hip-img-fx --help for the full list of options.

Benchmarking

Run the full benchmark suite and generate performance analysis:

./scripts/run_benchmark.sh
./scripts/run_analysis.sh bench/results/benchmark_*.csv

Detailed benchmark results and analysis are available in:
bench/results/benchmark_report.md

The benchmarks compare:

• GPU vs CPU execution
• Tuned vs untuned kernels
• Performance scaling across batch sizes

Testing

# Run all tests
meson test -C build --print-errorlogs

# CPU-only tests (no GPU required)
meson test -C build --suite cpu

See tests/README.md for detailed testing documentation.

Using HIP Image FX as a Library

HIP Image FX can be installed and reused as a library, including its autotuning framework:

meson install -C build
pkg-config --cflags hip-img-fx

This allows you to integrate GPU autotuning and image-processing kernels into existing HIP-based pipelines without building custom tuning infrastructure from scratch.

Refer to docs/AUTOTUNING_GUIDE.md for integration examples.

Roadmap

• Additional image filters and convolution kernels
• Expanded autotuning strategies and heuristics
• CI validation across multiple GPU architectures
• Improved profiling and visualization tooling

About the Author

This project is developed and maintained by Victor Anderssén.

HIP Image FX was built to demonstrate experience in:

• GPU programming with HIP (AMD ROCm & NVIDIA CUDA backends)
• Performance analysis, benchmarking, and kernel autotuning
• Modern C++ (C++20), Meson-based builds, and testable system design
• Designing reusable frameworks rather than one-off demos

If you are evaluating this project for professional use, or are interested in collaboration, consulting, or full-time work related to GPU compute or high-performance systems, feel free to reach out.

Contact

• Portfolio: https://victoranderssen.com
• GitHub: https://github.com/Avicted
• LinkedIn: https://www.linkedin.com/in/victoranderssen/

License

Apache License 2.0 - see LICENSE