How to set up the repository and run the code

1. In general run everything from the main folder e.g. python3 utils/plot_script/foo.py
2. Download mnist.csv and place it as MNIST/mnist.csv (e.g. from https://www.kaggle.com/datasets/oddrationale/mnist-in-csv/data)
3. Install PAPI (papi7.1.0) to allow # Flops measurements, # Memory Movements, ...
4. Install STREAM to allow measurement of bandwidth

Run and plot an example

Prepare Dataset

1. Modify dimension of PCA on MNIST by editing utils/process_data/pca_mnist.py (OPTIONAL, default = 64)
2. Gather the MNIST dataset with PCA plotted features by running python3 utils/process_data/pca_mnist.py

Measure System information (OPTIONAL)

1. Compile STREAM with and measure: gcc -O3 -fopenmp -DSTREAM_ARRAY_SIZE=100000000 stream.c -o stream && ./stream, to measure bandwitdh Beta needed for roofline plot
2. Modify main.cpp to modify the benchmarks (OPTIONAL) and, accordingly, the data of YOUR machine in implementations/constants.h and utils/plot_script/plot_data.py with information on the system and on the run (e.g. add measured-bandwidth, cache sizes, ...)

Run benchmarks

1. Prepare the system by fixing frequency of the cores (single), disabling Turbo Boost Technology or similar, disable most background processes, and ensure enough warmup and iterations runs are used.
2. Run a benchmark by first compiling with g++ main.cpp -o main -lpapi -std=c++20 -O3 -march=native (or if local installation or papi: g++ main.cpp -o main -I/usr/local/papi/include -L/usr/local/papi/lib -lpapi -std=c++20 -O3 -march=native)
3. Run ./main and wait for the program to finish

Visualize results

1. Visualize the result: python3 utils/plot_script/tsne_plot.py
2. Plot a simple benchmarks with theoretical limit: python3 utils/plot_script/performance_plot.py
3. Plot a roofline plot of the run: python3 utils/plot_script/roofline_plot.py
4. Plot a speedup plot of the run: python3 utils/plot_script/speedup_plot.py