GVS Warp Integration API¶

Integration API

SoRoMoX users should call the public GVS kinematics and dynamics methods. The symbols on this page are intended for packages that reuse the GVS kinematics, dynamics, or actuation pipelines with preallocated Warp buffers.

Linear-threadlike actuation¶

GVSThreadlikeOperands.from_model(robot) exposes the ordered built-in linear routing data together with interior body quadrature, link strain bases and reference strains, pre-scaled link bases, physical quadrature coordinates and weights, segment spans, and link-local/global coordinate maps. GVSThreadlikeShapes specifies the shared caller-owned strain workspace (E, S, Q, 6), matrix output (E, D, A), and fused-force output (E, D).

The exact launch sequences are:

launch_gvs_threadlike_actuation_matrix(...) prepares every quadrature strain once in (E, S, Q, 6), then integrates and writes every matrix entry.
launch_gvs_threadlike_actuation_force(...) prepares the same workspace, applies coordinate_scale[path] * controls[e, path], and reduces paths directly into (E, D). CUDA uses one 128-lane block per environment and segment; Warp CPU uses a scalar-coordinate kernel.

Both second stages explicitly write inactive paths and joint-only coordinates as zero, so neither sequence requires a separate output-clear launch. Matrix and force calls may reuse the same strain workspace when they do not overlap.

Floating-point inputs, workspaces, and outputs are FP64; coordinate maps and path spans are INT32. Callers own and retain all Warp views and buffers. Neither launcher allocates, calls JAX, synchronizes, or transfers data, and both two-launch sequences may be captured in a CUDA graph. Environment, segment, link-DOF, quadrature, and path dimensions remain runtime-shaped. The normalization test is exactly norm > global_eps; a collapsed tangent contributes zero, matching JAX.

soromox.execution.warp.gvs ¶

Public Warp-native GVS kinematics and dynamics building blocks.

The kernels and launch functions in this module operate entirely on warp.array objects. They do not require JAX and do not allocate output buffers, making them suitable for external CUDA-graph-capturable integrators such as a Newton solver. Callers own the buffers and must preserve the shapes, dtypes, coordinate maps, and launch ordering described by each function.

Soromox system users should normally call :meth:GVS.dynamics_terms <soromox.systems.GVS.dynamics_terms> or :meth:GVS.forward_dynamics <soromox.systems.GVS.forward_dynamics> instead. The lower-level API is intended for integration packages that need to compose the kinematics and dynamics pipelines inside a larger Warp-native simulation loop.

getattr ¶

__getattr__(name: str) -> Any

Load a public GVS Warp symbol without making Warp a core dependency.

Parameters:

Name	Type	Description	Default
`name`	`str`	Attribute requested from this package.	required

Returns:

Type	Description
`Any`	The public operand type, kernel, or launch function named by `name`.

Raises:

Type	Description
`AttributeError`	If `name` is not part of the public GVS Warp API.
`ImportError`	If a kernel is requested without `warp-lang` installed.

dir ¶

__dir__() -> list[str]

Include lazily exported Warp symbols in interactive discovery.