Principle:Google deepmind Mujoco Linear Algebra
| Knowledge Sources | Domains | Last Updated |
|---|---|---|
| Google DeepMind MuJoCo | Mathematics, Linear Algebra | 2025-02-15 |
Overview
Description: BLAS-like linear algebra operations optimized for the vector and matrix sizes commonly encountered in multibody dynamics.
Context: MuJoCo implements its own set of linear algebra primitives rather than depending on external BLAS libraries. These operations are tuned for the small-to-medium matrix sizes typical in robotics (3x3 rotations, 6x6 spatial inertias, and nv x nv mass matrices where nv is often tens to hundreds).
Theoretical Basis
The linear algebra utilities provide dense vector and matrix operations including dot products, matrix-vector multiplies, matrix-matrix multiplies, outer products, norms, and various specialized operations. By implementing these in-house, MuJoCo avoids the overhead of external library calls (function pointer indirection, thread synchronization) that can dominate runtime for small matrix sizes. The implementations use cache-friendly access patterns and, where beneficial, loop unrolling and SIMD intrinsics to maximize throughput on modern CPUs.
Related Pages
Implementations
- Implementation:Google_deepmind_Mujoco_Engine_Util_Blas
- Implementation:Google_deepmind_Mujoco_Engine_Util_Blas_Header
Workflows
- (none yet)