cvxgenrust Benchmarks

This reository contains the code for reproducing the numerical results in the paper "Generation of custom solvers in Rust for convex optimization".

Dependencies

• Python 3.12 or newer
• Rust and Cargo
• CMake and a C/C++ build toolchain

Setup

We use uv to manage the Python environment. Once you have uv installed, create the environment and install dependencies:

make sync

This runs uv sync --frozen, which should leave the environment exactly at the versions recorded in uv.lock.

Run the full benchmark

Run the complete configured experiment:

make experiment

This target:

1. generates cvxgenrust and CVXPYgen projects under artifacts/generated;
2. builds the generated Rust and C drivers;
3. runs all configured warmups and timing samples;
4. writes results/solve_times.csv, results/environment.json, results/summary.md, and plots under results/plots;
5. validates row counts, solver statuses, and objective-error tolerances.

The default configuration is benchmarks.toml, which uses 3 warmups, 30 timing samples, seed 10015, and the CLARABEL solver settings listed in that file.

Other commands

If results/solve_times.csv already exists, regenerate the Markdown summary and plots without rerunning benchmarks:

make report

For plots only:

make plot

Reuse existing generated projects and skip generation:

make run-skip-generate

Remove generated artifacts and result files:

make clean

The paper-style PDF plot in results/plots/solve_times.pdf is generated by:

uv run python src/paper.py