SAGE: Structure-Aware Genetic Compiler Flag Optimisation

SAGE is a system for automatically optimising compiler flags using a genetic algorithm. It targets C benchmarks (e.g. PolyBench) and searches for flag combinations that outperform the standard -O3 optimisation level.

---

Features

• Genetic algorithm-based compiler flag optimisation
• Automatic baseline comparison against -O3
• Support for PolyBench/C benchmarks
• Configurable via config.json or command-line arguments
• Hyperparameter tuning using OpenTuner
• Reproducible and extensible design

---

Prerequisites

Ensure the following are installed:

• Python 3.11+
• GCC 15.2.0 (or compatible)
• OpenTuner:

  pip install opentuner
  

---

Configuration

All GA settings are specified in config.json in the project root. Edit this file to change the target benchmark, hyperparameters, or random seed without modifying source code.

{
  "benchmark": "algorithm/benchmarks/matmul.cpp",
  "compiler": "g++-15",
  "population_size": 64,
  "crossover_type": "one_point",
  "crossover_rate": 0.471,
  "mutation_type": "gauss_by_center",
  "mutation_rate": 0.187,
  "selection_type": "ranking",
  "elitism_ratio": 0.300,
  "parents_portion": 0.553,
  "max_generations": 33,
  "max_no_improvement": 25,
  "num_runs": 5,
  "random_seed": 42
}

---

Running SAGE

1. Clone the repository:

git clone https://github.com/yarakmk/GeneticAlgorithmProject.git
cd GeneticAlgorithmProject

2. Set the target benchmark in config.json.

3. Set the compiler commad in config.json.

4. Run the optimiser:

python3 main.py

SAGE will print per-generation progress including best fitness, average fitness, and elapsed time, and output the best flag configuration discovered on completion.

---

Hyperparameter Tuning

To tune the GA hyperparameters for your benchmark and hardware using OpenTuner:

python3 hyperparameter_tuner.py --test-limit 52

Results are saved to opentuner.db/<ip-address>.db. The best configuration can then be manually copied into config.json. Note that opentuner.db and opentuner.log are runtime-generated and excluded from version control.

---

Running Tests

Run the full test suite using pytest:

pytest tests/

To include a coverage report:

pytest tests/ --cov=. --cov-report=term-missing

---

Reproducibility

A fixed random seed is set in config.json under "random_seed". This seeds both Python's random module and numpy at startup. Note that the primary experimental results in the accompanying report were produced prior to the introduction of the fixed seed and may differ slightly across runs due to stochastic variation in the GA.

---

Further Details

For more details on how to configure SAGE, please refer to the User Guide in Appendix B in the report.

Platform

• OS: macOS 14.0 (Sonoma)
• Architecture: ARM64 (Apple M2)
• GCC: 15.2.0
• Python: 3.12