This repo is the home of GNN part in LISA.

This repo has the following directories.

• data: store generated DFGs, dumped DFGs for a particular framework, generated labels, and training dataset.
• dfg_generatror: generate DFGs, and dump DFGs for a particular framework. All the generated data is stored in data folder.
• lisa_gnn_model: all the GNN models.

Instrcutions to use LISA (currently, the default framework is CGRA_ME. Will add a parameter later).

For the following examples, we use platform to represent compiler framework name such as cgra_me, and use arch_name to represent architecture name such as cgra_me_4_4 that means a 4x4 CGRA.

• First, activate lisa environment.

  conda activate lisa

• generate traninng graph (in dfg_generator)

  python dfg_generator.py -s 0 - n 10 -d platform

  -s: the start index of graph. Optional, the deafult value is 0
  -n: the number of generated graphed. Optional, the default value is 10
  This generates a folder platform under "data", and generate two folders platform and graph under "data/platform/".
  data/platform/graph: contain the standarded graph.
  data/platform/platform: contain the converted graph from the standarded graph to let the compiler to take as input DFG.

• transform graph (in dfg_generator)

  python dfg_transformer.py -s graph -d transformered_graph --home_directory platform

  --home_directory: the home directory of other directories. This folder is under "data" -s: the source dir of graphes. Optional, the deafult value is "graph"
  -d: the destination dir of transformed graphes. Optional, the default value is "transformered_graph"
  

• label fiter and generate training dataset (in data)

  python label_filter.py -r transformered_graph -l arch_name

  -r: raw graph, i.e, the directory of transformed graph (data/platform/transformered_graph)
  -l: the directory of labels (data/labels/arch_name).
  This command will generate a folder in training_dataset. The folder name is the same as label_dir*. This folder has the training dataset

• run training (in lisa_gnn_model)

  bash run_training.sh arch_name

  This will train all the models for this arch and save all the model with name "arch_name"

• run inference (in lisa_gnn_model)

  python gnn_inference.py graph_name arch_name

  The graph name is a file in "dara/infer/". For example, for a graph file "3c4c.txt", the input should be 3c4c.

Some useful commands in git:

• check whether the submodule has been pushed before pushing the main project

git push --recurse-submodules=check