Reduce Eigen allocations with fixed-size matrix and vector types by schaubh · Pull Request #1502 · AVSLab/basilisk

schaubh · 2026-07-31T17:35:03Z

Review: By commit
Merge strategy: Merge (no squash)

Description

This PR reduces Eigen heap allocations and unnecessary copies by replacing dynamically sized matrices and vectors with fixed-size or partially fixed-size types where dimensions are known at compile time.

The commits are organized by subsystem:

Eigen/C-array conversions and geodetic utilities
Common numerical utilities
Gauss-Markov propagation
SimpleNav and PlanetNav
Lambert planning and validation
Small-body navigation filters
Effector-control algorithms
Remaining three-dimensional simulation and control paths
Release-note documentation

The changes preserve existing numerical behavior and message interfaces. The SWIG-visible dynamic Eigen conversion functions remain available, while C++ overloads avoid converting fixed-size Eigen expressions to temporary MatrixXd objects.

Verification

Completed a full incremental C++ and SWIG build.
Passed 18 Eigen conversion and geodetic CTests.
Passed 8 targeted state-architecture, small-body EKF, small-body UKF, and MTB Python tests.
Confirmed git diff --check passes.

No tests were added or re-baselined because the changes optimize internal storage and expression evaluation without changing expected results. Existing conversion and module tests exercise the affected behavior.

Documentation

No module input/output documentation or user-facing configuration was invalidated.

Inline comments describing row-major and column-major conversions were updated for the new implementations. A release-note snippet documents the reduction in Eigen heap allocation across the affected utilities and modules.

Future work

No required follow-up is anticipated.

Use row-major Eigen maps to avoid transpose temporaries, add overloads that preserve fixed-size Eigen expressions, and use fixed-size 3x3 matrices for geodetic frame conversions.

Pass BSpline, Discretize, and Saturate inputs by reference, avoid unnecessary vector copies during discretization, and use fixed two-column storage for saturation bounds.

Preallocate random, noise, and propagated-state vectors to avoid per-step allocations, and pass configuration and state data by reference.

Represent the SimpleNav and PlanetNav error transition matrices with compile-time dimensions and keep per-step propagation in fixed-size storage.

Represent six-element states and covariance transformations with compile-time Eigen types throughout Lambert planning, validation, and Runge-Kutta propagation.

Encode EKF and UKF state, covariance, sigma-point, and measurement dimensions in their Eigen types, and remove the redundant runtime dimension counters.

Constrain mass-matrix blocks and VSCMG steering matrices to their fixed three-row or three-column dimensions and avoid copying stored momentum into a dynamic vector.

Replace dynamic three-dimensional matrices and vectors with fixed-size types or direct maps in waypoint control, magnetic torque bars, gravity, radiation, and ground mapping.