Implementation of RePart: Efficient Hypergraph Partitioning with Logic Replication Optimization for Multi-FPGA System. An advanced hypergraph partitioning implementation designed for deploying circuit hypergraphs onto multi-FPGA systems. Features:
- Logic replication for communication minimization
- Algorithmic enhancements for partitioning quality in coarsening, assignment and refinement phases of the partitioning algorithm
- End-to-end implementation considering FPGA-specific constraints
Some data structure implementations are derived from KaHyPar .
-
We use Titan23 benchmark dataset from [Timing Driven Titan: Enabling Large Benchmarks and Exploring the Gap Between Academic and Commercial CAD]. The hypergraphs of them can be found from Google Drive provided by TopoPart, which can also be found in the
./TiTan23\ Benchmarksfolder. We can usegenerateDataset.pyto convert origin format datasets to Integrated Circuit EDA Elite Challenge Contest Problem 3 format. -
The MFS56 is "synopsys02" from problem B in ICCAD 2019 Contest
$ python generateDataset.py 'Input Dataset Folder Path' 'FPGA(MFS) File Path' 'Output Dataset Folder Path'For example
$ python ./generateDataset.py ./Titan23\ Benchmarks/ ./MFS56 ./Titan23_TransformedThe imbalance factor generateDataset.py
balance_factor = 0.2
hop_max = 5 # max hop-distance (hop count + 1) of a legal edge10 cases from Integrated Circuit EDA Elite Challenge Contest Problem 3.
Preloaded test cases are in ./testcase.
$ cd ./RePart
$ g++ -Ofast -DNDEBUG -o partitioner partitioner.cpp \
-I../boost_1_86_0/include \
-L../boost_1_86_0/lib \
-static -lboost_thread -lboost_system -pthread$ ./partitioner -t ../Titan23_Transformed/neuron/ -s ../Titan23_Transformed/neuron/design.fpga.outNote: Replace Titan23_Transformed with your output folder name.
Output results are saved to Titan23_Transformed/neuron/design.fpga.out
Example (Case03):
$ ./partitioner -t ../testcase/case03 -s ../testcase/case03/design.fpga.outOutput results are saved to testcase/case03/design.fpga.out.
If you use RePart in your research or project, please cite our paper:
@misc{fu2026repartefficienthypergraphpartitioning,
title={RePart: Efficient Hypergraph Partitioning with Logic Replication Optimization for Multi-FPGA System},
author={Zizhuo Fu and Yifan Zhou and Zhaoxin Lu and Guangyu Sun and Runsheng Wang and Meng Li and Yibo Lin},
year={2026},
eprint={2604.00780},
archivePrefix={arXiv},
primaryClass={cs.AR},
url={https://arxiv.org/abs/2604.00780},
}