ATPlace2.5D Public Package

This repository contains a compact public package for ATPlace2.5D.

Contents

• cases/: ten clean case directories. Each case contains Bookshelf-style input files, WL-driven.json, and Thermal-aware.json.
• src/: encrypted ATPlace layout kernel and its runtime.
• thermal/: HotSpot files used by thermal evaluation.
• utils/: parsers for case input files.
• Thermal.py: optional readable HotSpot helper for users who want to inspect or reuse the floorplan/layer/power-trace generation logic.
• reproduce.sh: shell entry for selecting a case and placement mode and starting the encrypted layout kernel.

Usage

Create a Python environment with the package dependencies used by the encrypted kernel. gurobipy is the recommended solver interface, and the parameter files set ILPsolver to grb.

Each case directory contains Bookshelf input files and two layout parameter files.

• WL-driven.json is for wirelength-driven placement.
• Thermal-aware.json is for thermal-aware placement.

pulp may work in some environments, but it is not guaranteed to produce strictly identical legal placement or numerical values. Use gurobipy when reproducing reported layouts.

On the 231 server, activate the reproduction environment first:

conda activate ATPlan

Run one case as:

bash reproduce.sh Case1 wl
bash reproduce.sh Case1 thermal

The first argument is one of Case1 through Case10. The second argument is wl or thermal.

The command selects:

• input files from cases/<CaseX>/;
• WL-driven.json for wl;
• Thermal-aware.json for thermal;
• output directory results/<CaseX>_<mode>_<timestamp>/ unless ATPLACE_OUT_DIR is set.

Then it calls reproduce.py, which loads the selected Bookshelf files and parameter file and invokes the encrypted ATPLACE.PlaceFlow.placeflow_core kernel. A successful execution writes:

• summary.json: selected case, mode, parameter file, output directory, HPWL/TWL, runtime, and whether a final layout was returned;
• layout.json: final chiplet coordinates and chiplet sizes when the encrypted kernel returns a final layout.

Optional environment variables:

PYTHON=/path/to/python bash reproduce.sh Case1 wl
ATPLACE_OUT_DIR=/tmp/case1_wl bash reproduce.sh Case1 wl
ATPLACE_THERMAL_DIR=/path/to/thermal bash reproduce.sh Case1 thermal
ATPLACE_DRY_RUN=1 bash reproduce.sh Case1 wl

The case parameter files are self-contained layout settings. Do not change interposer geometry unless the target design itself changes.

Thermal-Aware Mode

Thermal-aware placement is enabled by temp_aware_opt in Thermal-aware.json.

reproduce.py builds the case, creates a compact analytic thermal-model object, and passes that object into the encrypted ATPLACE.PlaceFlow.placeflow_core kernel. The public wrapper does not run a separate visible thermal-model training script and does not write a thermal-model checkpoint before placement. It is therefore expected that a thermal run may print the same early placement messages as a wirelength run after a short initialization step.

Thermal.py is not imported by reproduce.py in the current public wrapper. Prints added to Thermal.py will not appear during bash reproduce.sh Case1 thermal unless a user imports and calls that helper directly. The encrypted kernel may call its own internal thermal evaluation path. The readable Thermal.py file is kept as an optional reference implementation for generating HotSpot .flp, .lcf, .ptrace, and configuration files from a chiplet layout.

thermal/ contains the HotSpot executable and default HotSpot configuration used by public thermal evaluation resources. Runtime thermal directory selection is controlled by ATPLACE_THERMAL_DIR.

The main thermal-aware layout controls are in cases/<CaseX>/Thermal-aware.json:

• temp_aware_opt: enables the thermal-aware objective;
• temp_weight_init, temp_threshold, and gamma: control the thermal penalty;
• wl_weight: controls the wirelength term in the thermal-aware objective;
• target_density, density_weight_init, overflow_init, and eta: control placement density and optimization behavior;
• floorplan_stages[0].learning_rate: position and angle learning rates.

Use these JSON files for ordinary parameter changes. Keep each case's interposer_size unchanged unless the physical case definition changes.

Changing Thermal Configuration

For a new thermal configuration, copy the thermal directory and point the run to it:

cp -R thermal thermal_custom
# edit thermal_custom/hotspot.config
ATPLACE_THERMAL_DIR="$PWD/thermal_custom" bash reproduce.sh Case1 thermal

Editing Thermal.py changes only that readable helper. It does not change reproduce.py or the encrypted placement kernel unless those files are modified to import and call the helper.

Changing layer material constants, layer thicknesses, bump parameters, or generated .flp/.lcf/.ptrace content requires a wrapper or evaluation script that explicitly uses Thermal.py.

Running the Full Reproduction Matrix

To run all ten cases in both modes, with three repeats per case-mode element and at most five cases active at the same time:

python reproduce_matrix.py --repeats 3 --case-workers 5

The script writes:

• repro_runs/<timestamp>/commands.json: all commands launched;
• repro_runs/<timestamp>/<CaseX>/<mode>/rep*/stdout.log and stderr.log;
• repro_runs/<timestamp>/summary.csv: all completed repeats and metrics parsed from summary.json;
• repro_runs/<timestamp>/best.csv: best repeat per case-mode element, ranked by the smallest twl_m.

For a new case, create cases/<NewCase>/ with matching Bookshelf-style files:

<NewCase>.blocks
<NewCase>.nets
<NewCase>.pl
<NewCase>.power
WL-driven.json
Thermal-aware.json

Then either add the case interposer size to CASE_INTERPOSER_SIZE in reproduce.py, or include interposer_size in the case parameter JSON.

Layout Visualization

After a successful run, generate a PNG from the exported layout:

python visualize_layout.py results/Case1_wl_YYYYMMDD_HHMMSS/layout.json
python visualize_layout.py results/Case1_thermal_YYYYMMDD_HHMMSS/layout.json --out case1_thermal.png

The visualizer reads only layout.json. It does not call the encrypted placement kernel.

Citation

@inproceedings{wang2024atplace2,
  title={ATPlace2. 5D: Analytical thermal-aware chiplet placement framework for large-scale 2.5 D-IC},
  author={Wang, Qipan and Li, Xueqing and Jia, Tianyu and Lin, Yibo and Wang, Runsheng and Huang, Ru},
  booktitle={Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design},
  pages={1--9},
  year={2024}
}