DAPO: Design Structure Aware Pass Ordering in High-Level Synthesis with Graph Contrastive and Reinforcement Learning

DAPO has been accepted to DATE'26. The pre-print is available on arXiv: https://arxiv.org/abs/2512.11342.

DAPO provides an end-to-end toolchain that learns design-aware optimization pass schedules for high-level synthesis. The workflow couples (i) heterogeneous graph contrastive learning for program embeddings, (ii) reinforcement learning for pass sequencing, and (iii) an LLVM-based quality-of-result (QoR) estimator used for fast reward evaluation.

1. Prerequisites

• LLVM/Clang 18.1.x installed and discoverable on PATH
• CMake ≥ 3.20 and a C++20-capable compiler
• Python ≥ 3.9
• Recommended: GPU with CUDA support for faster model training

If you are new to LLVM toolchain setup, refer to our tutorial: https://github.com/gjskywalker/LLVM-18-Tutorials.

Install Python dependencies from the repository root:

pip install -r requirements.txt

2. Build Analysis Tools

DAPO uses two custom LLVM passes: a feature extractor and a QoR (cycle) estimator. Build and stage both executables via the provided script:

cd /path/to/DAPO
./build_analysis.sh

The script configures Analysis/ with CMake, builds the passes, and copies the resulting binaries to Analysis_tools/feature_exctractor/feature_exctractor and Analysis_tools/qor_estimator/qor_estimator. These locations are consumed directly by the Python utilities.

3. Datasets & Intermediate Artifacts

3.1 Benchmark Sources

Benchmark C files are organized under Dataset/:

• Dataset/trainingset/: randomized kernels used by RL training.
• Dataset/Cross_Validation/: folds for GNN/RL model evaluation.
• Dataset/HLSDesigns/: representative real-world HLS programs.

3.2 Generate LLVM IR

Convert source benchmarks into LLVM IR (placed in Raw_IRs/) before graph construction:

python generate_ir.py \
	--Source Dataset/trainingset \
	--Output Raw_IRs

Repeat for other benchmark folders as needed.

4. Dataset Generation & GNN Models

All GNN-related scripts live under Models/GNNModels/.

4.1 Generate Graph Datasets

cd Models/GNNModels

# Convert LLVM IR files into heterogeneous graphs
python Build_Dataset.py \
	-d ../../Raw_IRs \
	-o Raw_Graphs \
	-t NGraph

# Optional: compute HGED distances for cross-program analysis
python HGED_Calculate.py

4.2 Train and Evaluate GNNs

# Train all configured GNN variants (GCN, RGCN, PNA, etc.)
python Train_All_Models.py \
    --epochs 200 \
    --patience 3

# Run inference 
python Infer_Pytorch.py 

Key outputs are placed under Models/GNNModels/models/ and Models/GNNModels/HGED_Result_Dapo/.

5. Reinforcement Learning Environment

RL scripts are located in Models/RLModels/. The environment consumes the feature extractor and QoR estimator staged earlier, so no manual path editing is required.

5.1 Prepare Training Assets

• Ensure Dataset/trainingset/ contains the C benchmarks (e.g., random*.cc).
• Place trained GNN checkpoints in Models/GNNModels/models/ (or update paths when launching RL).

5.2 Launch RL Training & Inference

cd ../RLModels

# PPO training with a single environment
python ppo_ray.py \
	--feature-type rgcn \
	--run-name rgcn_single \
	--model-path Embedding_model_RGCNConv.pth

# PPO training with multiple parallel environments
python ppo_ray_multienv.py \
	--feature-type rgcn \
	--run-name rgcn_multi \
	--model-path Embedding_model_RGCNConv.pth

Command-line arguments mirror the options defined in each script (see the --help flag for full listings). Generated feature traces and cycle logs are stored under Models/RLModels/gym_env/envs/Feature_Cycles_Tests/.

6. Repository Layout

Analysis/                 # C++ sources for feature extractor and QoR estimator
Analysis_tools/           # Staged LLVM tool binaries (created by build_analysis.sh)
Dataset/                  # Training and evaluation C benchmarks
Models/
	├── GNNModels/          # Graph dataset pipeline and GNN training scripts
	└── RLModels/           # Reinforcement learning environments and PPO drivers
Raw_IRs/                  # Sample LLVM IR inputs for dataset generation

7. Troubleshooting

• qor_estimator not found: rerun ./build_analysis.sh and verify LLVM 18 is on your PATH.
• clang failed to generate bitcode: confirm your LLVM clang binary is version 18 and accessible.
• RL training stalls: ensure the feature extractor outputs CSV files in Feature_Cycles_Tests/; delete stale contents if rerunning.

For questions or reproducibility details, please open an issue or contact the authors listed in the paper.