GShare + H2P Hybrid Branch Predictor

Advanced Computer Architecture (ACA) Project

This repository contains a high-performance hybrid branch predictor designed for the ChampSim simulator. Our implementation combines the classic GShare predictor with a Hardness-to-Predict (H2P) table and a Perceptron backend to handle complex, non-linear branch behaviors.

---

🚀 Key Features

• Hybrid Architecture: Dynamic selection between GShare (for simple patterns) and Perceptron (for complex correlations).
• Confidence-Based Training: Implements mispred || low-confidence thresholding to optimize perceptron learning.
• Bias Weighting: Enhanced perceptron entries with dedicated bias weights for improved capture of branch leaning.
• Phase Adaptation: Periodic hardness-table decay to re-evaluate branch complexity over time.
• Instrumentation: Built-in counters for usage rates, training triggers, and hardness promotions.

---

📁 Project Structure

• src/
  • gshare_h2p.h/cc: Core predictor implementation.
• configs/
  • champsim_*.json: ChampSim configuration files for various predictor setups.
• results/
  • eval_results_final.csv: Comprehensive evaluation across 4 benchmark traces.
• LICENSE: MIT License.
• README.md: Project documentation.

---

🛠️ How to Build and Run

To use this predictor, you must have a ChampSim environment set up.

1. Integrate into ChampSim: Copy the files from src/ into your ChampSim branch/ directory.

2. Build:

   # Example build command in ChampSim
   ./config.sh configs/champsim_gshare_h2p.json
   make

3. Run Evaluation:

   # Run with a trace
   ./bin/champsim --warmup_instructions 5000000 --simulation_instructions 20000000 <trace_file>

---

📊 Performance Snapshot

Extracted from 5M Warmup / 20M Sim on Perlbench trace.

Predictor	IPC	Branch Accuracy	Branch MPKI
Bimodal	2.333	97.42%	3.601
GShare	2.370	98.02%	2.763
Perceptron	2.424	98.57%	1.993
GShare+H2P (Ours)	2.380	98.18%	2.546

Note

The GShare+H2P hybrid balances accuracy with resource efficiency, providing a significant edge over standard GShare. See results/ for full trace evaluations.

---

👥 Contributors

• The ACA Project Team (Advanced Computer Architecture Group)

---

© 2026 MIT License