Skip to content

python_analysis.md

Latest commit

 

History

History
114 lines (82 loc) · 4.19 KB

python_analysis.md

File metadata and controls

114 lines (82 loc) · 4.19 KB
Raw

Python Analysis

Install dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,test]"

The Python package uses a src/ layout. Installing the project editable makes commands such as python -m python.analysis.plot_snapshots work from the source checkout. If you prefer not to install the package, set PYTHONPATH=src before running module commands.

Plot and Render Snapshots

Plot the latest snapshot and diagnostics:

python -m python.analysis.plot_snapshots --input experiments/validation/smoke_test --output smoke_snapshot.png

Render an animation:

python -m python.animation.render_snapshots --input experiments/validation/smoke_test --mode scatter3d --camera-orbit --output smoke_collision.mp4

Render a density projection:

python -m python.animation.render_snapshots --input experiments/validation/smoke_test --mode density --projection camera --output smoke_density.mp4

Render a static density projection:

python -m python.analysis.plot_snapshots --input experiments/validation/smoke_test --density-output smoke_density.png --no-diagnostics

Create a self-contained interactive browser viewer:

python -m python.animation.interactive_viewer --input experiments/validation/smoke_test --output viewer.html

Regenerate README Artifacts

Regenerate the README collision GIF from the dedicated 1000-body config:

./build/fmm_galaxy_sim --config configs/readme_1000_body_collision.toml
python -m python.animation.render_scientific_gif --input experiments/validation/readme_1000_body_collision --output docs/assets/galaxy_collision_3d_1000.gif --mode density
python -m python.analysis.plot_snapshots --input experiments/validation/readme_1000_body_collision --snapshot experiments/validation/readme_1000_body_collision/snapshot_000149.csv --output docs/assets/readme_snapshot_step149.png --density-output docs/assets/readme_density_step149.png --no-diagnostics
python scripts/run_benchmarks.py --executable build-readme-gif/fmm_galaxy_sim.exe --particles 250 500 1000 --steps 20 --repetitions 3
python scripts/run_benchmarks.py --executable build/fmm_galaxy_sim --solvers cuda-tree cuda-fmm --particles 10000 50000 100000 --steps 10 --repetitions 3 --output-format none --expansion-order 0

Benchmarks and Sweeps

Compare output formats on the same benchmark cases:

python scripts/run_benchmarks.py --executable build/fmm_galaxy_sim --solvers direct --particles 10000 --steps 10 --output-formats csv parquet

Generate the standard direct-reference force-error suite:

python scripts/run_force_error_benchmarks.py --executable build/fmm_galaxy_sim

For CI-scale validation, use the smoke profile:

python scripts/run_force_error_benchmarks.py --executable build/fmm_galaxy_sim --smoke

The suite writes experiments/accuracy/force_error_summary.csv, force_error_summary.md, force_error_vs_n.png, force_error_vs_theta.png, energy_drift.png, and momentum_drift.png. It compares step-0 accelerations against direct summation and reports drift from each solver's diagnostics over a short integration window.

Launch a generic YAML-defined parameter sweep:

python scripts/sweep.py --grid configs/sweeps/theta_leaf_order.yaml

The sweep runner generates per-run TOML configs, raw logs, simulator output directories, sweep_summary.csv, optional sweep_summary.parquet, and sweep_metadata.json. Use --dry-run to only materialize planned configs, --resume to skip completed runs with metadata, and --jobs N for local parallel execution.

Solver Crossover Analysis

Generate solver crossover plots and tables from runtime and accuracy benchmark CSVs:

python -m python.analysis.solver_crossover \
  --runtime-csv docs/benchmarks/local_cpu_benchmark.csv \
  --accuracy-csv experiments/accuracy/force_error_summary.csv

For fresh runtime inputs, scripts/run_benchmarks.py --crossover-suite runs a wider particle-count sweep with both snapshot output disabled and CSV output enabled. The crossover analysis writes runtime_vs_n.png, particle_steps_vs_n.png, force_error_vs_runtime.png, best_solver_by_n.csv, target_accuracy_summary.csv, and solver_crossover_summary.md.