Meliora framework

Meliora is a code analysis infrastructure for machine learning-based performance model generation of arbitrary codes based on static analysis of intermediate language representations.

Requirements

LLVM9, CMake3.1+, networkx>=2, pynauty

Installation

The framework is majorly composed of metric extraction and graph learning components (based on PSCN).

Metric generation component

    $ cd datagen
    $ mkdir build
    $ cd build
    $ cmake ..
    $ make

Graph learning component

No additional step required. Users can choose to use the pynauty provided with the PSCN implementation.

    $ cd modelgen/pynauty-0.6.0
    $ make pynauty
    $ make user-ins pynauty
    $ pip install

Train the model

To train the model with provided data:

python3 modelgen/train_model.py -l data/data1/labels_train.txt data/data1/train_input2/ -o [specify the path for saving the model, optional. default will be current location]

Prediction (matching)

python3 modelgen/get_recom.py -l data/data1/labels_train.txt -v data/data1/labels_train.txt -m data/data1/static4_2_cfg_model.h5  data/data1/validate_input2

It will generate pred_results.csv. To generate the matching summary, run:

python3 datagen/script/predAnalyzer data/data1/pred_results.csv