pymomentum-cpu 0.1.78.post11__cp312-cp312-win_amd64.whl

pymomentum/backend/__init__.py

@@ -0,0 +1,16 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+PyMomentum Backend
+
+High-performance implementations for forward kinematics and linear blend skinning operations.
+"""
+
+# Make submodules available for import
+from . import skel_state_backend, trs_backend, utils
+
+__all__ = ["skel_state_backend", "trs_backend", "utils"]

pymomentum/backend/skel_state_backend.py

@@ -0,0 +1,614 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+"""
+Skeleton State Backend for PyMomentum
+
+This module provides efficient forward kinematics and skinning operations using
+the skeleton state representation (8 parameters per joint: translation, quaternion, scale).
+
+The skeleton state representation is more compact than the TRS backend and uses
+quaternions for rotation, making it suitable for applications requiring smooth
+interpolation and fewer parameters.
+
+Performance Notes:
+This backend matches the behavior of the C++ Momentum code, ensuring consistency
+with the reference implementation. However, the TRS backend may be 25-50% faster
+in PyTorch due to not requiring quaternion normalization operations, though the
+exact performance difference may vary depending on the specific use case.
+
+Key Functions:
+- global_skel_state_from_local_skel_state: Forward kinematics from local to global joint states
+- skin_points_from_skel_state: Linear blend skinning using skeleton states
+- local_skel_state_from_joint_params: Convert joint parameters to local states
+
+Related Modules:
+- trs_backend: Alternative backend using separate translation/rotation/scale tensors
+- skel_state: Core skeleton state operations and utilities
+- quaternion: Quaternion math operations used by this backend
+"""
+
+from typing import List, Tuple
+
+import torch as th
+
+from pymomentum import quaternion, skel_state, trs
+from pymomentum.backend.trs_backend import unpose_from_global_joint_state
+
+
+@th.jit.script
+def local_skel_state_from_joint_params(
+    joint_params: th.Tensor,
+    joint_offset: th.Tensor,
+    joint_quat_rotation: th.Tensor,
+) -> th.Tensor:
+    """
+    Convert joint parameters to local skeleton state representation.
+
+    This function transforms 7-parameter joint representation (translation, euler angles, log-scale)
+    into the 8-parameter skeleton state format (translation, quaternion, scale). The skeleton state
+    representation uses quaternions for rotation and linear scale factors.
+
+    Parameter Transformation:
+    - Translation: joint_params[:,:,:3] + joint_offset -> direct translation
+    - Rotation: euler_xyz angles -> quaternion, composed with joint pre-rotation
+    - Scale: log2(scale) -> linear scale via exp2() transformation
+
+    Args:
+        joint_params: Joint parameters, shape (batch_size, num_joints, 7).
+            Each joint has [tx, ty, tz, euler_x, euler_y, euler_z, log2_scale] parameters.
+        joint_offset: Per-joint translation offset, shape (num_joints, 3).
+            Static offset applied to each joint's local translation.
+        joint_quat_rotation: Per-joint rotation offset as quaternions, shape (num_joints, 4).
+            Static rotation [qx, qy, qz, qw] applied to each joint's local rotation.
+
+    Returns:
+        local_skel_state: Local skeleton state, shape (batch_size, num_joints, 8).
+            Each joint contains [tx, ty, tz, qx, qy, qz, qw, s] parameters.
+
+    Note:
+        The quaternion format follows [qx, qy, qz, qw] convention (vector-first).
+        Scale values are exponentiated from log2 space to linear space.
+    """
+    t = joint_offset[None, :] + joint_params[:, :, :3]
+    q = quaternion.multiply(
+        joint_quat_rotation[None],
+        quaternion.euler_xyz_to_quaternion(joint_params[:, :, 3:6]),
+    )
+    s = th.exp2(joint_params[:, :, 6:])
+
+    return th.cat([t, q, s], dim=-1)
+
+
+@th.jit.script
+def global_skel_state_from_local_skel_state_impl(
+    local_skel_state: th.Tensor,
+    prefix_mul_indices: List[th.Tensor],
+    save_intermediate_results: bool = True,
+    use_double_precision: bool = True,
+) -> Tuple[th.Tensor, List[th.Tensor]]:
+    """
+    Compute global skeleton state from local joint transformations using forward kinematics.
+
+    This function implements forward kinematics (FK) using prefix multiplication for efficient
+    parallel computation of joint transformations. Each joint's local transformation is composed
+    with its parent's global transformation to produce the joint's global transformation.
+
+    The skeleton state uses an 8-parameter representation per joint: [translation, quaternion, scale]
+    - translation (3D): local translation vector [tx, ty, tz]
+    - quaternion (4D): rotation quaternion [qx, qy, qz, qw] (normalized)
+    - scale (1D): uniform scale factor [s]
+
+    Forward Kinematics Algorithm:
+    For each joint j with parent p in the kinematic hierarchy:
+        global_state_j = global_state_p ⊙ local_state_j
+
+    Where ⊙ represents similarity transformation composition:
+    - t_global = t_parent + s_parent * rotate_by_quaternion(q_parent, t_local)
+    - q_global = quaternion_multiply(q_parent, q_local)
+    - s_global = s_parent * s_local
+
+    Args:
+        local_skel_state: Local joint transformations, shape (batch_size, num_joints, 8).
+            Each joint state contains [tx, ty, tz, qx, qy, qz, qw, s] parameters.
+        prefix_mul_indices: List of [child_index, parent_index] tensor pairs that define
+            the traversal order for the kinematic tree. This ordering enables efficient
+            parallel computation while respecting parent-child dependencies.
+        save_intermediate_results: If True, saves intermediate joint states during the
+            forward pass for use in backpropagation. Set to False for inference-only
+            computations to reduce memory usage.
+        use_double_precision: If True, performs computations in float64 for improved
+            numerical stability. Recommended for deep kinematic chains to minimize
+            accumulated floating-point errors.
+
+    Returns:
+        global_skel_state: Global joint transformations, shape (batch_size, num_joints, 8).
+            Each joint contains the composed transformation from root to that joint.
+        intermediate_results: List of intermediate joint states from the forward pass.
+            Required for efficient gradient computation during backpropagation.
+            Empty if save_intermediate_results=False.
+
+    Note:
+        This function is JIT-compiled for performance. The prefix multiplication approach
+        allows vectorized batch computation while maintaining kinematic chain dependencies.
+
+    See Also:
+        :func:`global_skel_state_from_local_skel_state`: User-facing wrapper function
+        :func:`local_skel_state_from_joint_params`: Convert joint parameters to local states
+    """
+    dtype = local_skel_state.dtype
+    intermediate_results: List[th.Tensor] = []
+    if use_double_precision:
+        global_skel_state = local_skel_state.clone().double()
+    else:
+        global_skel_state = local_skel_state.clone()
+    for prefix_mul_index in prefix_mul_indices:
+        source = prefix_mul_index[0]
+        target = prefix_mul_index[1]
+
+        state1 = global_skel_state.index_select(-2, target)
+        state2 = global_skel_state.index_select(-2, source)
+
+        if save_intermediate_results:
+            intermediate_results.append(state2.clone())
+
+        global_skel_state.index_copy_(-2, source, skel_state.multiply(state1, state2))
+
+    return (global_skel_state.to(dtype), intermediate_results)
+
+
+def global_skel_state_from_local_skel_state_no_grad(
+    local_skel_state: th.Tensor,
+    prefix_mul_indices: List[th.Tensor],
+    save_intermediate_results: bool = True,
+    use_double_precision: bool = True,
+) -> Tuple[th.Tensor, List[th.Tensor]]:
+    """
+    Compute global skeleton state without gradient tracking.
+
+    This is a convenience wrapper around global_skel_state_from_local_skel_state_impl
+    that explicitly disables gradient computation using torch.no_grad(). Useful for
+    inference-only forward passes to reduce memory usage.
+
+    Args:
+        local_skel_state: Local joint transformations, shape (batch_size, num_joints, 8)
+        prefix_mul_indices: List of [child_index, parent_index] tensor pairs
+        save_intermediate_results: Whether to save intermediate states for backprop
+        use_double_precision: Whether to use float64 for numerical stability
+
+    Returns:
+        global_skel_state: Global joint transformations, shape (batch_size, num_joints, 8)
+        intermediate_results: List of intermediate joint states from forward pass
+
+    See Also:
+        :func:`global_skel_state_from_local_skel_state_impl`: Implementation function
+    """
+    with th.no_grad():
+        outputs = global_skel_state_from_local_skel_state_impl(
+            local_skel_state,
+            prefix_mul_indices,
+            save_intermediate_results=save_intermediate_results,
+            use_double_precision=use_double_precision,
+        )
+    return outputs
+
+
+# @th.jit.script
+def global_skel_state_from_local_skel_state_backprop(
+    global_skel_state: th.Tensor,
+    grad_global_skel_state: th.Tensor,
+    prefix_mul_indices: List[th.Tensor],
+    intermediate_results: List[th.Tensor],
+    use_double_precision: bool = True,
+) -> th.Tensor:
+    """
+    Compute gradients for local skeleton state through backpropagation.
+
+    This function implements the backward pass for forward kinematics, computing
+    gradients of the loss with respect to local joint states given gradients
+    with respect to global joint states.
+
+    The backpropagation uses the intermediate results saved during the forward
+    pass to efficiently compute gradients without recomputing the full forward
+    kinematics chain.
+
+    Gradient Flow:
+    For each joint j with parent p, the backward pass computes:
+    ∂L/∂local_j = ∂L/∂global_j * ∂global_j/∂local_j
+
+    Where the Jacobian ∂global_j/∂local_j depends on the parent's global state
+    and is computed using the chain rule through similarity transformation composition.
+
+    Args:
+        global_skel_state: Global joint states from forward pass, shape (batch_size, num_joints, 8).
+            These states are modified in-place during backpropagation.
+        grad_global_skel_state: Gradients w.r.t. global joint states, shape (batch_size, num_joints, 8).
+            Input gradients from downstream computations (e.g., skinning loss).
+        prefix_mul_indices: List of [child_index, parent_index] tensor pairs that defined
+            the forward pass traversal order. Backprop processes these in reverse.
+        intermediate_results: List of intermediate joint states saved during forward pass.
+            Required to compute accurate gradients without numerical drift.
+        use_double_precision: If True, performs computations in float64 for numerical stability.
+            Recommended for deep kinematic chains and precise gradient computation.
+
+    Returns:
+        grad_local_skel_state: Gradients w.r.t. local joint states, shape (batch_size, num_joints, 8).
+            These gradients can be used to update joint parameters via optimization.
+
+    Note:
+        This function processes the kinematic chain in reverse order of the forward pass,
+        accumulating gradients from children to parents while reconstructing intermediate states.
+
+    See Also:
+        :func:`global_skel_state_from_local_skel_state_impl`: Forward pass implementation
+    """
+    dtype = global_skel_state.dtype
+    if use_double_precision:
+        grad_local_skel_state = grad_global_skel_state.clone().double()
+        global_skel_state = global_skel_state.clone().double()
+    else:
+        grad_local_skel_state = grad_global_skel_state.clone()
+        global_skel_state = global_skel_state.clone()
+
+    for prefix_mul_index, state_source_original in list(
+        zip(prefix_mul_indices, intermediate_results)
+    )[::-1]:
+        source = prefix_mul_index[0]
+        target = prefix_mul_index[1]
+
+        state_target = global_skel_state.index_select(-2, target)
+
+        # forward
+        # This is for checking that the numerical error is in the safe
+        # range, but it's not completely abnormal to see large numerical
+        # error during backpropagation at the beginning of training.
+        # state_source = global_skel_state[:, source]
+        # state_source_expect = sim3_multiplication(
+        #     state_target, state_source_original
+        # )
+        # assert th.allclose(
+        #     state_source, state_source_expect, atol=1e-5, rtol=1e-5
+        # )
+
+        # backward
+        grad_target_accum, grad_source_original = skel_state.multiply_backprop(
+            state1=state_target,
+            state2=state_source_original,
+            grad_state=grad_local_skel_state.index_select(-2, source),
+        )
+
+        # setup the reduced gradients
+        grad_local_skel_state.index_copy_(-2, source, grad_source_original)
+        grad_local_skel_state.index_add_(-2, target, grad_target_accum)
+        # setup the reduced KC
+        global_skel_state.index_copy_(-2, source, state_source_original)
+
+    return grad_local_skel_state.to(dtype)
+
+
+class GlobalSkelStateFromLocalSkelStateJIT(th.autograd.Function):
+    """
+    PyTorch autograd function for differentiable forward kinematics using skeleton states.
+
+    This class implements automatic differentiation for the forward kinematics operation,
+    allowing gradients to flow from global joint states back to local joint states.
+    The forward pass computes global states efficiently while saving intermediate results
+    for the backward pass.
+
+    Note:
+        This class is used internally by global_skel_state_from_local_skel_state when
+        not in JIT mode. It provides gradient computation capabilities that are not
+        available in pure JIT-compiled functions.
+    """
+
+    @staticmethod
+    # pyre-ignore[14]
+    def forward(
+        local_skel_state: th.Tensor,
+        prefix_mul_indices: List[th.Tensor],
+    ) -> Tuple[th.Tensor, List[th.Tensor]]:
+        return global_skel_state_from_local_skel_state_no_grad(
+            local_skel_state,
+            prefix_mul_indices,
+        )
+
+    @staticmethod
+    # pyre-ignore[14]
+    # pyre-ignore[2]
+    def setup_context(ctx, inputs, outputs) -> None:
+        (
+            _,
+            prefix_mul_indices,
+        ) = inputs
+        (
+            global_skel_state,
+            intermediate_results,
+        ) = outputs
+        # need to clone as it's modified in-place
+        ctx.save_for_backward(global_skel_state.clone())
+        ctx.intermediate_results = intermediate_results
+        ctx.prefix_mul_indices = prefix_mul_indices
+
+    @staticmethod
+    # pyre-ignore[14]
+    def backward(
+        # pyre-ignore[2]
+        ctx,
+        grad_global_skel_state: th.Tensor,
+        _0,
+    ) -> Tuple[th.Tensor, None]:
+        (global_skel_state,) = ctx.saved_tensors
+
+        intermediate_results = ctx.intermediate_results
+        prefix_mul_indices = ctx.prefix_mul_indices
+
+        grad_local_state = global_skel_state_from_local_skel_state_backprop(
+            global_skel_state,
+            grad_global_skel_state,
+            prefix_mul_indices,
+            intermediate_results,
+        )
+        return grad_local_state, None
+
+
+def global_skel_state_from_local_skel_state(
+    local_skel_state: th.Tensor,
+    prefix_mul_indices: List[th.Tensor],
+) -> th.Tensor:
+    """
+    Compute global skeleton state from local joint transformations (user-facing wrapper).
+
+    This is the main entry point for forward kinematics using skeleton states. It automatically
+    selects between JIT-compiled and autograd-enabled implementations based on the execution context.
+
+    Args:
+        local_skel_state: Local joint transformations, shape (batch_size, num_joints, 8).
+                         Each joint contains [tx, ty, tz, qx, qy, qz, qw, s] parameters.
+        prefix_mul_indices: List of [child_index, parent_index] tensor pairs defining
+                           the kinematic hierarchy traversal order.
+
+    Returns:
+        global_skel_state: Global joint transformations, shape (batch_size, num_joints, 8).
+                          Each joint contains the composed transformation from root to that joint.
+
+    Note:
+        When called within torch.jit.script or torch.jit.trace context, uses the JIT-compiled
+        implementation for maximum performance. Otherwise, uses the autograd-enabled version
+        for gradient computation.
+
+    See Also:
+        :func:`global_skel_state_from_local_skel_state_impl`: JIT implementation
+        :func:`local_skel_state_from_joint_params`: Convert joint parameters to local states
+    """
+
+    if th.jit.is_tracing() or th.jit.is_scripting():
+        global_skel_state, _ = global_skel_state_from_local_skel_state_impl(
+            local_skel_state, prefix_mul_indices
+        )
+        return global_skel_state
+    else:
+        global_skel_state, _ = GlobalSkelStateFromLocalSkelStateJIT.apply(
+            local_skel_state,
+            prefix_mul_indices,
+        )
+        return global_skel_state
+
+
+@th.jit.script
+def skin_points_from_skel_state(
+    template: th.Tensor,
+    global_skel_state: th.Tensor,
+    binded_skel_state_inv: th.Tensor,
+    skin_indices_flattened: th.Tensor,
+    skin_weights_flattened: th.Tensor,
+    vert_indices_flattened: th.Tensor,
+) -> th.Tensor:
+    """
+    Apply linear blend skinning (LBS) to points using skeleton state transformations.
+
+    This function deforms template points by blending joint transformations according
+    to skinning weights. Each point is influenced by multiple joints, with the final
+    position computed as a weighted average of joint-transformed positions.
+
+    Linear Blend Skinning Formula:
+    For each vertex v_i with position p_i in template:
+        skinned_p_i = Σ_j w_ij * T_j * (T_bind_j^(-1) * p_i)
+
+    Where:
+    - w_ij: skinning weight of joint j on vertex i (sum to 1 per vertex)
+    - T_j: global transformation of joint j
+    - T_bind_j^(-1): inverse bind pose transformation of joint j
+
+    The bind pose represents the joint configuration when skinning weights were defined.
+    The inverse bind pose transformation removes the original joint influence before
+    applying the current joint transformation.
+
+    Args:
+        template: Template vertex positions, shape (batch_size, num_vertices, 3).
+            These are the rest pose positions before deformation.
+        global_skel_state: Global joint transformations, shape (batch_size, num_joints, 8).
+            Each joint contains [tx, ty, tz, qx, qy, qz, qw, s] parameters.
+        binded_skel_state_inv: Inverse bind pose transformations, shape (num_joints, 8).
+            Static transformations that represent the joint configuration during binding.
+        skin_indices_flattened: Joint indices for skinning, shape (num_influences,).
+            Flattened list of joint indices that influence vertices.
+        skin_weights_flattened: Skinning weights, shape (num_influences,).
+            Corresponding weights for each joint influence (normalized per vertex).
+        vert_indices_flattened: Vertex indices for skinning, shape (num_influences,).
+            Flattened list of vertex indices corresponding to joint influences.
+
+    Returns:
+        skinned_points: Deformed vertex positions, shape (batch_size, num_vertices, 3).
+            Points transformed according to current joint poses and skinning weights.
+
+    Note:
+        This function is JIT-compiled for performance. The flattened indices allow
+        efficient vectorized computation of skinning influences across all vertices.
+
+    See Also:
+        :func:`skin_oriented_points_from_skel_state`: Skinning for oriented points (gaussians)
+        :func:`global_skel_state_from_local_skel_state`: Forward kinematics for joint states
+    """
+    assert template.shape[-1] == 3
+    while template.ndim < global_skel_state.ndim:
+        template = template.unsqueeze(0)
+
+    template = template.expand(
+        list(global_skel_state.shape[:-2]) + list(template.shape[-2:])
+    )
+
+    joint_state = skel_state.multiply(
+        global_skel_state,
+        binded_skel_state_inv,
+    )
+
+    skinned = th.zeros_like(template)
+    skinned = skinned.index_add(
+        -2,
+        vert_indices_flattened,
+        skel_state.transform_points(
+            th.index_select(joint_state, -2, skin_indices_flattened),
+            th.index_select(template, -2, vert_indices_flattened),
+        )
+        * skin_weights_flattened[None, :, None],
+    )
+    return skinned
+
+
+@th.jit.script
+def skin_oriented_points_from_skel_state(
+    means: th.Tensor,
+    quaternions: th.Tensor,
+    global_skel_state: th.Tensor,
+    binded_skel_state_inv: th.Tensor,
+    skin_indices_flattened: th.Tensor,
+    skin_weights_flattened: th.Tensor,
+    vert_indices_flattened: th.Tensor,
+) -> tuple[th.Tensor, th.Tensor]:
+    """
+    Apply linear blend skinning to oriented points (Gaussian splatting) using skeleton states.
+
+    This function extends standard LBS to handle oriented points, which have both position
+    and orientation. Each oriented point is influenced by multiple joints, with both its
+    position and orientation blended according to skinning weights.
+
+    For orientations, rotation matrices are linearly interpolated (LERP) and then converted
+    back to quaternions. This is different from spherical linear interpolation (SLERP) but
+    provides smoother blending for skinning applications.
+
+    Args:
+        means: Point positions, shape (batch_size, num_points, 3).
+            Center positions of the oriented points (e.g., Gaussian means).
+        quaternions: Point orientations, shape (batch_size, num_points, 4).
+            Orientations as quaternions in [qx, qy, qz, qw] format.
+        global_skel_state: Global joint transformations, shape (batch_size, num_joints, 8).
+            Each joint contains [tx, ty, tz, qx, qy, qz, qw, s] parameters.
+        binded_skel_state_inv: Inverse bind pose transformations, shape (num_joints, 8).
+            Static transformations representing the joint configuration during binding.
+        skin_indices_flattened: Joint indices for skinning, shape (num_influences,).
+            Flattened list of joint indices that influence points.
+        skin_weights_flattened: Skinning weights, shape (num_influences,).
+            Corresponding weights for each joint influence (normalized per point).
+        vert_indices_flattened: Point indices for skinning, shape (num_influences,).
+            Flattened list of point indices corresponding to joint influences.
+
+    Returns:
+        skinned_means: Deformed point positions, shape (batch_size, num_points, 3).
+            Transformed positions according to current joint poses and skinning weights.
+        skinned_quaternions: Blended point orientations, shape (batch_size, num_points, 4).
+            Orientations blended using rotation matrix interpolation and converted back to quaternions.
+
+    Note:
+        This function is commonly used for Gaussian splatting applications where each
+        Gaussian has both a mean position and orientation that need to be skinned.
+        The rotation blending uses linear interpolation in matrix space for stability.
+
+    See Also:
+        :func:`skin_points_from_skel_state`: Standard LBS for points without orientation
+        :func:`global_skel_state_from_local_skel_state`: Forward kinematics for joint states
+    """
+    assert means.shape[-1] == 3
+    assert quaternions.shape[-1] == 4
+    while means.ndim < global_skel_state.ndim:
+        means = means.unsqueeze(0)
+
+    means = means.expand(list(global_skel_state.shape[:-2]) + list(means.shape[-2:]))
+
+    joint_state = skel_state.multiply(
+        global_skel_state,
+        binded_skel_state_inv,
+    )
+
+    sim3_transforms = th.index_select(joint_state, -2, skin_indices_flattened)
+    t, q, s = skel_state.split(sim3_transforms)
+    r = quaternion.to_rotation_matrix(q)
+
+    means_flattened = th.index_select(means, -2, vert_indices_flattened)
+
+    skinned_means = th.zeros_like(means)
+    skinned_means = skinned_means.index_add(
+        -2,
+        vert_indices_flattened,
+        (s * quaternion.rotate_vector(q, means_flattened) + t)
+        * skin_weights_flattened[None, :, None],
+    )
+
+    lerp_rotations = th.zeros(
+        quaternions.shape[:-1] + (3, 3),
+        dtype=quaternions.dtype,
+        device=quaternions.device,
+    )
+    lerp_rotations = lerp_rotations.index_add(
+        -3,
+        vert_indices_flattened,
+        r * skin_weights_flattened[None, :, None, None],
+    )
+    lerp_quaternions = quaternion.from_rotation_matrix(lerp_rotations)
+    lerp_quaternions = quaternion.normalize(lerp_quaternions)
+    skinned_quaternions = quaternion.multiply_assume_normalized(
+        quaternions, lerp_quaternions
+    )
+    return skinned_means, skinned_quaternions
+
+
+def unpose_from_momentum_global_joint_state(
+    verts: th.Tensor,
+    global_joint_state: th.Tensor,
+    binded_joint_state_inv: th.Tensor,
+    skin_indices_flattened: th.Tensor,
+    skin_weights_flattened: th.Tensor,
+    vert_indices_flattened: th.Tensor,
+    with_high_precision: bool = True,
+) -> th.Tensor:
+    """
+    The inverse function of skinning().
+    WARNING: the precision is low...
+
+    Args:
+        verts: [batch_size, num_verts, 3]
+        global_joint_state (th.Tensor): (B, J, 8)
+        binded_joint_state (th.Tensor): (J, 8)
+        skin_indices_flattened: (N, ) LBS skinning nbr joint indices
+        skin_weights_flattened: (N, ) LBS skinning nbr joint weights
+        vert_indices_flattened: (N, ) LBS skinning nbr corresponding vertex indices
+        with_high_precision: if True, use high precision solver (LDLT),
+            but requires a cuda device sync
+    """
+    t, r, s = trs.from_skeleton_state(global_joint_state)
+    t0, r0, _ = trs.from_skeleton_state(binded_joint_state_inv)
+
+    return unpose_from_global_joint_state(
+        verts,
+        t,
+        r,
+        s,
+        t0,
+        r0,
+        skin_indices_flattened,
+        skin_weights_flattened,
+        vert_indices_flattened,
+        with_high_precision,
+    )