This document describes the structure of the Gematria repository.
The Python code is split into modules according to its function; C++ code
follows the same directory structure. By convention, we put all C++ files to
gematria/{package_name}/, and the Python libraries under
gematria/{package_name}/python.
Contains code and data structures related to in-memory representation of basic blocks. While we use protos to store the data in rest, most of the code uses a lightweight data structure that can is easy to share between Python and C++.
Implementation of the GRANITE model and base classes for building models with Graph neural networks. The graph construction from basic block data is implemented in C++ for efficiency reasons.
Notable files:
granite/python/run_granite_model.py: the main module for running the GRANITE model.
Input/output utilities.
Contains base classes for building models and the necessary support code for training and inference (the implementation of the training loop, an inference loop, and a generic main function for running Gematria models).
Notable files:
model/python/model_base.py: the base class for all Gematria models. Contains most of the model-independent code, like the training and inference loops, cost definition, ...model/python/main_function.py: contains model-independent code needed to launch Gematria models from command-line: definitions of command-line flags, and a generic main() function.
Protocol buffer definitions used in the project.
Base classes and implementation of models that treat the basic block as a sequence of instructions. In particular, contains implementation of the Ithemal model and the Ithemal+ model described in the Granite paper.
Notable files:
sequence/python/run_sequence_model.py: the main module for running the Ithemal and Ithemal+ models.
Contains helper classes and functions for testing the model and a tiny data set of basic blocks for testing.
Contains various utilities that do not fit into other packages.