Conditional Branch Predictor Based on Weightless Neural Networks

Code and algorithms used in my Master's main project in 2020-2023. You can find the final text here

Requirements:

• C++14
• Python 3.9.4

Datasets:

• The "Dataset_pc_decimal" folder provides the 6 datasets used in this work.
• In each file, the first column represents the PC in decimal and the second column represents the real outcome of the actual branch

In order to run experiments:

• In linux terminal (ubuntu) make the apropriate compilation:

  $ sh compiler.sh
  

• Then, run in tterminal using command line arguments:

  $ ./bpu-wisard Dataset_pc_decimal/[name of dataset] [n] [a] [b] [c] [d0] [d1] [d2] [d3] [d4] [e]
  

  where:

  • [name of dataset]: is the full name of one of the files located in the "Dataset_pc_decimal" folder
  • [n]: is the size of the n-tuple
  • [a], [b], [c], [d0], [d1], [d2], [d3], [d4], [e]: are the parameters discussed in final text of the dissertation

  For example, after compiling, run:

  $ ./bpu-wisard Dataset_pc_decimal/M1.txt 32 2 2 2 2 2 3 4 1 1
  

  The terminal will show something similar to this:

    ----- Results ------
  
    Predicted  branches: 282678
  
    Not predicted branches: 17322
  
    Accuracy: 94.226000
   
    ------ Size of ntuple (address_size): 32 -----
  
    ------ Size of each input: 348 -----
  

• The "Specialized-preditors" directory collets the results of the specialized predictor for each dataset. Run the python scripts to print the corresponding bar plots.

Important insights

• Increase the accuracy of deployed machine learning technology by more than 2% compared to state-of-the-art.
• Reduce the input size of the algorithm by 31% compared to traditional methods based on traditional neural networks.

Articles

Some of this work was done in collaboration with researchers from other institutions. Our efforts have been recognized in two publications:

• Neurocomputing
• ESANN

License

MIT License