Skip to content

Latest commit

 

History

History
40 lines (33 loc) · 3.68 KB

README.md

File metadata and controls

40 lines (33 loc) · 3.68 KB
Raw

DFG Partition Algorithm for FLEX

This repository is part of FLEX project. The scripts are used for dataflow graph (DFG) partitioning when the DFG is too large to be mapped onto a spatial CGRA. You can find our paper here 👇

FLEX: Introducing FLEXible Execution on CGRA with Spatio-Temporal Vector Dataflow
Thilini Kaushalya Bandara; Dan Wu; Rohan Juneja; Dhananjaya Wijerathne; Tulika Mitra; Li-Shiuan Peh
National University of Singapore

About the script

  1. Allows a big DFG to be partitioned into multiple cluster, with store and load nodes added to pass intermediate result for cross cluster edges.
  2. Ensures no cycles among clusters.
  3. Ensures a valid memory allocation for intermediate values. The memory node placement can be seen in 'log.txt' and 'memPE_alloc.txt"

Details

  1. code read xml file and memory allocation file to generate nx.Digraph with node attributes: 'bank' and 'op'. panorama.py -> load_graph()
  2. code partition the graph with num_cluster clusters main.py -> dfg_partition()
  3. code add virtual load and store nodes for cross-cluster edges. For multiple edges sharing the same source vertex, only one virtual store node will be added.
  4. code Repeatedly try merge two small clusters and generate memory allocation (to port) for all load\store nodes, virtual nodes included. main.py -> merge()
  5. In 2-4, cluster size, memory nodes in one cluster and cluster-level cycles are checked, if fails, increase the num_cluster and repeat 2-4. In 4, if there is no valid memory node mapping, increase the num_cluster and repeat 2-4.
  6. For a valid partition, add select node for each cluster (cluster inherently with select node are excluded), connect the select node with all load and store nodes, virtual included. main.py -> add_select(),
  7. code dump each cluster to xml file. main.py -> dump_to_xml()

Configs

  1. Shuffled the candidate ports during searching to avoid congestion in one memory bank. See map_one() in main.py
  2. To modify the output format, check main.py line 334, function ‘dump_to_xml’.
  3. To disable array splitting (randomly choose bank [variable_address\bank_size] or [variable_address\bank_size + 1]), go panorama.py, line 62. Comment out “+ random.randint(0,1)”
  4. As the clustering algorithm also has randomness, uncomment panorama.py, line 12-13, to ensure reproducibility. Also, you can comment out and run multiple times to get more results.

Citation

@inproceedings{bandara2023flex,
  title={FLEX: Introducing FLEXible Execution on CGRA with Spatio-Temporal Vector Dataflow},
  author={Bandara, Thilini Kaushalya and Wu, Dan and Juneja, Rohan and Wijerathne, Dhananjaya and Mitra, Tulika and Peh, Li-Shiuan},
  booktitle={2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD)},
  pages={1--9},
  year={2023},
  organization={IEEE}
}