FastMatMult

Simple benchmark for a naive dense matrix multiplication kernel in C++. The program multiplies two random square matrices for a set of sizes, captures the fastest runtime per size, and writes time/GFLOP data to a CSV file that can be plotted from Python.

Build

CMake

cmake -S . -B build
cmake --build build

Single-file build

g++ -O3 -std=c++17 -march=native -DNDEBUG main.cpp -o fast_mat_mult

Usage

./fast_mat_mult [--output file.csv] [--repetitions count] [--sizes n1,n2,...] [--seed value] [--algo naive|goto|both] [--verify]

• --output / -o: CSV file to be generated (default gflops_data.csv)
• --repetitions / -r: how many runs per matrix size (best run is recorded, default 3)
• --sizes / -s: comma-separated list of matrix dimensions (default 64,128,256,384,512)
• --seed: RNG seed for reproducible inputs
• --algo: choose the naive implementation, the (future) optimized Goto-style path, or run both for side-by-side data
• --verify: optional correctness check that runs naive vs Goto once per size (not timed)

Each CSV record has size,algorithm,time_seconds,gflops. Example execution:

./fast_mat_mult --sizes 64,128,256,512 --repetitions 5 --algo both --verify

Plotting in Python

Install dependencies (e.g., pip install -r requirements.txt), then run:

python plot_gflops.py baseline_naive.csv  # or pass any CSV produced by the benchmark

The script plots GFLOP/s versus matrix size for every algorithm found in the CSV and opens an interactive window; use --title to customize the figure title.