How google-research/circuit_training works
AlphaChip generates chip floorplans through clustered netlist preprocessing, reinforcement-learning macro placement, placement-cost evaluation, and optional DREAMPlace standard-cell refinement. Users provide a netlist; grouping produces clustered data, the environment exposes placement observations and rewards, PPO actors learn placement policies, and the resulting floorplan can be refined and exported. The graph preserves documented congestion, density, spacing, alignment, fixed-macro, orientation, and distributed-training capabilities. Some orchestration paths are inferred from module roles because only selected source excerpts were available.
Main entry points: environment.py Placement Environment · grouper.py Netlist Grouper · plc_client.py Placement Cost Client · train_ppo_lib.py PPO Training
Components 19
Netlist Preparation
- Input Netlist
- Netlist Grouper
circuit_training/grouping/grouper.py - Grouping Logic
circuit_training/grouping/grouping.py - Clustered Netlist
Placement Environment
- Placement Environment
circuit_training/environment/environment.py - Observation Extractor
circuit_training/environment/observation_extractor.py - Placement Cost Client
circuit_training/environment/plc_client.py
Distributed Learning
- PPO Data Collection
circuit_training/learning/ppo_collect_lib.py - Replay Service
circuit_training/learning/ppo_reverb_server_lib.py - Policy Value Models
circuit_training/model/model.py - Graph Neural Model
circuit_training/model/model_lib.py - PPO Agent
circuit_training/learning/agent.py - PPO Training
circuit_training/learning/train_ppo_lib.py
Placement Backends
- DREAMPlace Refinement
circuit_training/dreamplace/dreamplace_core.py - DREAMPlace Converter
circuit_training/dreamplace/plc_converter.py - Coordinate Descent
circuit_training/environment/coordinate_descent_placer.py
Floorplan Results
- Final Floorplan
Other components
- Chip Designer
- Placement Cost Service
Connections 22
- Chip Designer → Input Netlist: provides
- Input Netlist → Netlist Grouper: reads
- Netlist Grouper → Grouping Logic: partitions
- Netlist Grouper → Clustered Netlist: writes
- Netlist Grouper → Placement Cost Client: evaluates
- Placement Cost Client → Placement Cost Service: calls
- Clustered Netlist → Placement Environment: initializes
- Placement Environment → Observation Extractor: extracts
- Placement Environment → Placement Cost Client: queries
- Placement Environment → PPO Data Collection: feeds episodes
- PPO Data Collection → Replay Service: publishes trajectories
- PPO Training → Replay Service: reads trajectories
- PPO Training → PPO Agent: trains
- PPO Agent → Policy Value Models: runs networks
- Policy Value Models → Graph Neural Model: constructs
- PPO Training → Policy Value Models: saves policy
- Placement Environment → Coordinate Descent: optimizes coordinates
- DREAMPlace Refinement → DREAMPlace Converter: converts data
- DREAMPlace Converter → Placement Cost Client: reads placement
- DREAMPlace Refinement → Final Floorplan: outputs refinement
- Coordinate Descent → Final Floorplan: outputs placement
- Chip Designer → Final Floorplan: receives