Awesome

Data Collection

To collect data to train RouteGNN, we first need to collect placement result in different epochs when running placer and save them as the structure below

data---|---collect---netlist_name--|---1---|---placement results
                                   |
                                   |---2---|---placement results
                                   |
                                   |---epochs---|---placement results
                                   ...
                                   |
                                   |---2000---|---placement results

The placement results include the file as below

• hdm.npy vdm.npy: initial horizon/vertical utilization map
• iter_{epochs}_bad_cmap_h.npy iter_{epochs}_bad_cmap_v.npy: horizon/vertical rudy map
• iter_{epochs}_cmap_h.npy iter_{epochs}_cmap_v.npy: horizon/vertical overflow map
• iter_{epochs}_pos.npy: all cell position, shape:[num_movable_cells, 2]
• iter_{epochs}_x.npy, iter_{epochs}_y.npy: movable cell x-position/y-position
• iter_fix_hori_cap.npy iter_fix_verti_cap.npy: horizon/vertical capacity for each routing grid
• iter_fix_mapper.npy: map each cell to its original cell in circuits
• iter_fix_nctu_cap.npy: capacities for NCTUgr
• iter_fix_pd.npy: pin direction
• iter_fix_pin2edge.npy: map each pin to its connecting net
• iter_fix_pin2node.npy: map each pin to its cell
• iter_fix_po_x.npy iter_fix_po_y.npy: pin offset for each pins
• iter_fix_sizes_x.npy iter_fix_sizes_y.npy: width/height of each cells

After collecting the data, run the command below to process the original data for training in the "RouteGraph/data" path

python script_process.py

In script_process.py, you can appoint to process which netlist in "netlist_names" variable.

Train

Run the command to train model in the "RouteGraph" path

python script_train_congestion_optimize.py --name {your_model_name}

The model parameter will store in param/{your_model_name}.pkl

Run RoutePlacer

Use DREAMPlace Framework

Clone DREAMPlace repository. Copy the file in RoutePlacer repository to DREAMPlace and replace the file.

Complie DREAMPlace following DREAMPlace README

Run Benchmark

Copy DAC2012 and ISPD2011 benchmark to "DREAMPlace/benchmark". The file structure is as below

DREAMPlace---|---benchmark---|---ISPD2011---|---netlist_name---|---circuit_files
                             |              |
                             |              |---netlist_name---|---circuit_files
                             |
                             |
						   |---DAC2012----|---netlist_name---|---circuit_files
						   			     |
						   			     |
						   			     |---netlist_name---|---circuit_files

circuit files include the files below

netlist_name.aux
netlist_name.nets
netlist_name.nodes
netlist_name.pl
netlist_name.route
netlist_name.scl
netlist_name.shapes
netlist_name.wts

move trained model file into "DREAMPlace/param"

run command to get benchmark

python script_run_ours.py --name {dac12/ispd11} --model {model_name}

the result will store in "DREAMPlace/results"