Ask for details about the data generation part of GAMORA

[Jedges]

Hello author, I would like to ask whether it is possible to export the generated AIG and the labels in the GAMORA task correspond to the nodes in the AIG when generating multipliers of different bit widths through abc, because the graph structure currently generated does not contain NOT nodes, and I need this information. At the same time, how is the node encoding performed after Technology Mapping, because I noticed that in the corresponding node-feat.csv file, the dimension of the node encoding is no longer a simple 3 dimensions.

[Jedges]

Also note that in the code provided, the implementation for generating the booth multiplier is TBD, so how can I generate a booth multiplier like the one in your dataset through abc?

[ycunxi]

Also note that in the code provided, the implementation for generating the booth multiplier is TBD, so how can I generate a booth multiplier like the one in your dataset through abc?

You will see booth.bar in this link. The booth multiplier datasets have been released. Regarding generation, you can check ABC to modify the command to generate booth encoded multiplier, or directly load Booth multipliers (readable by abc of course).

https://huggingface.co/datasets/yucx0626/Gamora-CSA-Multiplier

[ycunxi]

Hello author, I would like to ask whether it is possible to export the generated AIG and the labels in the GAMORA task correspond to the nodes in the AIG when generating multipliers of different bit widths through abc, because the graph structure currently generated does not contain NOT nodes, and I need this information. At the same time, how is the node encoding performed after Technology Mapping, because I noticed that in the corresponding node-feat.csv file, the dimension of the node encoding is no longer a simple 3 dimensions.

A simple method would be just convert the AIG into a AND-NOT network at the edgelist level. You can try networkx which should do the job.

[lydiawunan]

1. Regarding getting the NOT node information, I would recommend reconstructing the network based on node features and the edge list.
2. Due to some engineering issues, the current node features have 4 dimensions, but the first two are duplicated, making it 3 unique features. The first two dimensions indicate whether the node is PI/PO or intermediate. The third and fourth dimensions indicate whether the input to this node is negated or not.
3. We generate Booth multipliers manually. Here is an example of how to generate it.
   • First, write the verilog file "mult8.v" of a multiplier:

     
      module comp (a,b,c);
          input wire [7:0] a, b;
          output wire [15:0] z;
          assign z = a * b;
      endmodule

   • Second, use abc to generate the "booth_mult8.blif" file

     ./abc -c "%read mult8.v;%blast -b;&put;ps;write booth_mult8.blif;"

   • Third, use python script to generate the dataset

     python ABC_dataset_generation_map.py --datagen 2 --designfile booth_mult8

[Jedges]

Thanks very much for your reply. I will carry out the next experiment according to your suggestion.