We use a GNN to try to learn to predict the performance difference of an application compiled with two set of optimizations using the LLVM compiler. The dataset used is described bellow.
For that, we use the network described in the following scheme:
The implementation of the network is in GNN.py and it uses the Spektral framework (https://github.com/danielegrattarola/spektral)
To test it:
pip3 install tensorflow spektral matplotlib sklearn numpy
python3 GNN.pyThis implementation (and Spektral) requires Tensorflow 2.
In data/ there is a serialized NumPy array named small150.npz. This file includes three NumPy arrays. One for pairs of adjacent matrixes representing Control Flow Graphs (CFGs) for same applications compiled twice with different optimization plains. Another contains features for each node in the CFGs, extract from the number of x86_64 instructions for each CFG node in the code binary. The last contains the relative difference of performance from the two pairs of CFGs. The goal is to learn the relative performance using the CFG structure and the binary instructions counters. They have the following shapes:
- Pairs of CFGs: (2, 25500, 150, 150)
- CFGs features: (2, 25500, 150, 94)
- Relative speedup: (25500, 1)
There are 25500 pairs of CFGs collected from a set of programs compiled with LLVM with different optimization plains.
The goal is to predict the performance difference of an application compiled with two different compilation plains using information about the amount of x86_64 final total of instruction and the CFG instruction.
This code is under MIT license and the dataset in Creative Commons Attribution 4.0 International.