The introduction of additional mjData arrays for the implicit integrator caused in a shift in data alignment in memory. In certain combinations of machines and models, this change resulted in ~5% performance regression that appears to be attributable to L2 cache misses.
This change aligns all arrays in mjData to the boundary of a typical L2 cache line size. Benchmarking shows that this resolves the performance regression on a Broadwell machine.
PiperOrigin-RevId: 451652783
Change-Id: I4ed2c2d03453a83a0b4a9b8e224e2fd7c0dadea9
This PR added necessary annotations, following what we did for mjData
and mjModel for length of the given array.
This is useful because my parser used to generate Swift interface:
https://github.com/liuliu/swift-mujoco/blob/main/Sources/codegen/main.swift#L297
such that the access to underlying array is bound-checked.
Please let me know if this makes sense. After this change, I can
successfully generate all mj* struct interfaces in Swift without problem
(with notable exception of mjui): https://github.com/liuliu/swift-mujoco/tree/main/Sources
Added analytic derivatives of smooth (unconstrained) dynamics forces, with respect to velocities:
- Centripetal and Coriolis forces computed by the Recursive Newton-Euler algorithm.
- Damping and fluid-drag passive forces.
- Actuation forces.
A new implicit-in-velocity integrator is implemented using the analytic derivatives. This integrator lies between the Euler and Runge Kutta integrators in terms of both stability and computational cost.
PiperOrigin-RevId: 450377010
Change-Id: Ie192b441876c22e732fb749333926f296e0a09cc
Kevin Zakka