leo-formation

fp64 GPU DOP853 and leapfrog propagators for the LEO formation-flight setting discussed in arXiv:2511.19468v1:

• DOP853 — high-order Dormand-Prince DOP853 in NVIDIA Warp, fp64. Matches the SciPy DOP853 oracle to ≤1e-9 across a full orbit.
• Leapfrog — point-mass + J2 leapfrog in NVIDIA Warp, fp64, with optional CUDA-Graph capture.

This repo provides two GPU implementations for that setting: DOP853 for accuracy and leapfrog for throughput.

Implementation

leo_formation/warp_dop853.py — DOP853 in NVIDIA Warp. Includes the in-place kernel with embedded error (dop853_step), a zero-error variant (dop853_step_zero_error), and a pure-output variant (dop853_step_out). dop853_graph.py wraps per-step launches in wp.ScopedCapture.

leo_formation/warp_orbital.py — leapfrog (leapfrog_step, leapfrog_step_x4), point-mass + J2, fp64.

Reproducing

pytest

tests/test_dop853_parity.py checks DOP853 against the SciPy oracle. tests/test_graph_capture_parity.py checks leapfrog sequential vs captured execution.

---

Tools and hardware: ATTRIBUTION.md. Dependencies: pyproject.toml. License: Apache-2.0 (LICENSE).