KB4CT: Knowledge Base for Compiler Tuning

Project Overview

KB4CT is a knowledge-based compiler tuning system that optimizes LLVM pass sequences through a combination of offline empirical prototype discovery and an online knowledge-guided personalized evolutionary algorithm. The system operates in two main stages: Offline Knowledge Base Construction and Online Knowledge-Guided Personalized Optimization.

Dataset Preparation

Extract Dataset from Supplementary Material

Extract the LLVM IR datasets from the supplementary material. Place the extracted datasets in the project root directory, ensuring the structure is as follows:

KB4CT/
├── dataset/
│   ├── train/          
│   │   ├── dataset1/
│   │   ├── dataset2/
│   │   └── ...
│   └── test/           
│       ├── dataset1/
│       ├── dataset2/
│       └── ...
│── DLL
├── LLVMEnv/
├── llvm_tools/
├── output/
└── KB4CT.py

Running Instructions

Basic Execution

python KB4CT.py

Configuration

You can configure the parameters in the if name == 'main': section of KB4CT.py.

Offline GA Parameters

"offline_ga_params": {
    "seq_len": 100,           # Sequence length
    "population_size": 100,   # Population size
    "generations": 30,        # Number of generations
    "elite_size": 20,         # Number of elite individuals
    "crossover_rate": 0.8,    # Crossover probability
    "mutation_rate": 0.8      # Mutation probability
}

Online GA Parameters

"online_ga_params": {
    "population_size": 50,     # Population size
    "generations": 5,          # Number of generations
    "elite_size": 10,          # Number of elite individuals
    "crossover_rate": 0.8,     # Crossover probability
    "mutation_rate": 0.99      # Mutation probability
}

Output Results

After execution, the results will be saved in the output/ directory:

• pass_embeddings_visualization.png: Visualization of pass embeddings
• pass_clusters_visualization.png: Visualization of pass clusters
• ablation_study_*.png: Ablation study result figures
• ablation_study_report.txt: Ablation study report
• ablation_detailed_results.json: Detailed result data

Ablation Study Modes

The system supports the following ablation study modes:

1. full: Full knowledge-guided method
2. no_knowledge_crossover: No knowledge-guided crossover
3. no_knowledge_mutation: No knowledge-guided mutation
4. random_init: Random population initialization
5. no_knowledge: Standard GA without any knowledge guidance