warp-lang 1.0.2__py3-none-manylinux2014_x86_64.whl → 1.1.0__py3-none-manylinux2014_x86_64.whl

warp/native/spatial.h

@@ -1,631 +1,631 @@
-/** Copyright (c) 2022 NVIDIA CORPORATION.  All rights reserved.
- * NVIDIA CORPORATION and its licensors retain all intellectual property
- * and proprietary rights in and to this software, related documentation
- * and any modifications thereto.  Any use, reproduction, disclosure or
- * distribution of this software and related documentation without an express
- * license agreement from NVIDIA CORPORATION is strictly prohibited.
- */
-
-#pragma once
-
-namespace wp
-{
-
-//---------------------------------------------------------------------------------
-// Represents a twist in se(3)
-template <typename Type>
-using spatial_vector_t = vec_t<6,Type>;
-
-template<typename Type>
-CUDA_CALLABLE inline Type spatial_dot(const spatial_vector_t<Type>& a, const spatial_vector_t<Type>& b)
-{
-    return dot(a, b);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline vec_t<3,Type> &w_vec( spatial_vector_t<Type>& a )
-{
-    return *(vec_t<3,Type>*)(&a);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline vec_t<3,Type> &v_vec( spatial_vector_t<Type>& a )
-{
-    return *(vec_t<3,Type>*)(&a.c[3]);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline const vec_t<3,Type> &w_vec( const spatial_vector_t<Type>& a )
-{
-    spatial_vector_t<Type> &non_const_vec = *(spatial_vector_t<Type>*)(const_cast<Type*>(&a.c[0]));
-    return w_vec(non_const_vec);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline const vec_t<3,Type> &v_vec( const spatial_vector_t<Type>& a )
-{
-    spatial_vector_t<Type> &non_const_vec = *(spatial_vector_t<Type>*)(const_cast<Type*>(&a.c[0]));
-    return v_vec(non_const_vec);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline spatial_vector_t<Type> spatial_cross(const spatial_vector_t<Type>& a,  const spatial_vector_t<Type>& b)
-{
-    vec_t<3,Type> w = cross(w_vec(a), w_vec(b));
-    vec_t<3,Type> v = cross(v_vec(a), w_vec(b)) + cross(w_vec(a), v_vec(b));
-    
-    return spatial_vector_t<Type>({w[0], w[1], w[2], v[0], v[1], v[2]});
-}
-
-template<typename Type>
-CUDA_CALLABLE inline spatial_vector_t<Type> spatial_cross_dual(const spatial_vector_t<Type>& a,  const spatial_vector_t<Type>& b)
-{
-    vec_t<3,Type> w = cross(w_vec(a), w_vec(b)) + cross(v_vec(a), v_vec(b));
-    vec_t<3,Type> v = cross(w_vec(a), v_vec(b));
-
-    return spatial_vector_t<Type>({w[0], w[1], w[2], v[0], v[1], v[2]});
-}
-
-template<typename Type>
-CUDA_CALLABLE inline vec_t<3,Type> spatial_top(const spatial_vector_t<Type>& a)
-{
-    return w_vec(a);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline vec_t<3,Type> spatial_bottom(const spatial_vector_t<Type>& a)
-{
-    return v_vec(a);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_spatial_dot(const spatial_vector_t<Type>& a, const spatial_vector_t<Type>& b, spatial_vector_t<Type>& adj_a, spatial_vector_t<Type>& adj_b, const Type& adj_ret)
-{
-    adj_dot(a, b, adj_a, adj_b, adj_ret);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_spatial_cross(const spatial_vector_t<Type>& a,  const spatial_vector_t<Type>& b, spatial_vector_t<Type>& adj_a,  spatial_vector_t<Type>& adj_b, const spatial_vector_t<Type>& adj_ret)
-{
-    adj_cross(w_vec(a), w_vec(b), w_vec(adj_a), w_vec(adj_b), w_vec(adj_ret));
-    
-    adj_cross(v_vec(a), w_vec(b), v_vec(adj_a), w_vec(adj_b), v_vec(adj_ret));
-    adj_cross(w_vec(a), v_vec(b), w_vec(adj_a), v_vec(adj_b), v_vec(adj_ret));
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_spatial_cross_dual(const spatial_vector_t<Type>& a,  const spatial_vector_t<Type>& b, spatial_vector_t<Type>& adj_a, spatial_vector_t<Type>& adj_b, const spatial_vector_t<Type>& adj_ret)
-{
-    adj_cross(w_vec(a), w_vec(b), w_vec(adj_a), w_vec(adj_b), w_vec(adj_ret));
-    adj_cross(v_vec(a), v_vec(b), v_vec(adj_a), v_vec(adj_b), w_vec(adj_ret));
-
-    adj_cross(w_vec(a), v_vec(b), w_vec(adj_a), v_vec(adj_b), v_vec(adj_ret));
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_spatial_top(const spatial_vector_t<Type>& a, spatial_vector_t<Type>& adj_a, const vec_t<3,Type>& adj_ret)
-{
-    w_vec(adj_a) += adj_ret;
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_spatial_bottom(const spatial_vector_t<Type>& a, spatial_vector_t<Type>& adj_a, const vec_t<3,Type>& adj_ret)
-{
-    v_vec(adj_a) += adj_ret;
-}
-
-
-//---------------------------------------------------------------------------------
-// Represents a rigid body transform<Type>ation
-
-template<typename Type>
-struct transform_t
-{
-    vec_t<3,Type> p;
-    quat_t<Type> q;
-
-    CUDA_CALLABLE inline transform_t(vec_t<3,Type> p=vec_t<3,Type>(), quat_t<Type> q=quat_t<Type>()) : p(p), q(q) {}
-    CUDA_CALLABLE inline transform_t(Type)  {}  // helps uniform initialization
-
-    CUDA_CALLABLE inline Type operator[](int index) const
-    {
-        assert(index < 7);
-
-        return p.c[index];
-    }
-
-    CUDA_CALLABLE inline Type& operator[](int index)
-    {
-        assert(index < 7);
-
-        return p.c[index];
-    }    
-};
-
-template<typename Type=float32>
-CUDA_CALLABLE inline transform_t<Type> transform_identity()
-{
-    return transform_t<Type>(vec_t<3,Type>(), quat_identity<Type>());
-}
-
-template<typename Type>
-inline CUDA_CALLABLE bool operator==(const transform_t<Type>& a, const transform_t<Type>& b)
-{
-    return a.p == b.p && a.q == b.q;
-}
-
-
-template<typename Type>
-inline bool CUDA_CALLABLE isfinite(const transform_t<Type>& t)
-{
-    return isfinite(t.p) && isfinite(t.q);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline vec_t<3,Type> transform_get_translation(const transform_t<Type>& t)
-{
-    return t.p;
-}
-
-template<typename Type>
-CUDA_CALLABLE inline quat_t<Type> transform_get_rotation(const transform_t<Type>& t)
-{
-    return t.q;
-}
-
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> transform_multiply(const transform_t<Type>& a, const transform_t<Type>& b)
-{
-    return { quat_rotate(a.q, b.p) + a.p, mul(a.q, b.q) };
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_transform_multiply(const transform_t<Type>& a, const transform_t<Type>& b, transform_t<Type>& adj_a, transform_t<Type>& adj_b, const transform_t<Type>& adj_ret)
-{
-    // translational part
-    adj_quat_rotate(a.q, b.p, adj_a.q, adj_b.p, adj_ret.p);
-    adj_a.p += adj_ret.p;
-
-    // rotational part
-    adj_mul(a.q, b.q, adj_a.q, adj_b.q, adj_ret.q);
-}
-
-
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> transform_inverse(const transform_t<Type>& t)
-{
-    quat_t<Type> q_inv = quat_inverse(t.q);
-    return transform_t<Type>(-quat_rotate(q_inv, t.p), q_inv);
-}
-
-    
-template<typename Type>
-CUDA_CALLABLE inline vec_t<3,Type> transform_vector(const transform_t<Type>& t, const vec_t<3,Type>& x)
-{
-    return quat_rotate(t.q, x);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline vec_t<3,Type> transform_point(const transform_t<Type>& t, const vec_t<3,Type>& x)
-{
-    return t.p + quat_rotate(t.q, x);
-}
-
-// not totally sure why you'd want to do this seeing as adding/subtracting two rotation
-// quats doesn't seem to do anything meaningful
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> add(const transform_t<Type>& a, const transform_t<Type>& b)
-{
-    return { a.p + b.p, a.q + b.q };
-}
-
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> sub(const transform_t<Type>& a, const transform_t<Type>& b)
-{
-    return { a.p - b.p, a.q - b.q };
-}
-
-// also not sure why you'd want to do this seeing as the quat would end up unnormalized
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> mul(const transform_t<Type>& a, Type s)
-{
-    return { a.p*s, a.q*s };
-}
-
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> mul(Type s, const transform_t<Type>& a)
-{
-    return mul(a, s);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> mul(const transform_t<Type>& a, const transform_t<Type>& b)
-{
-    return transform_multiply(a, b);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> operator*(const transform_t<Type>& a, Type s)
-{
-    return mul(a, s);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> operator*(Type s, const transform_t<Type>& a)
-{
-    return mul(a, s);
-}
-
-
-template<typename Type>
-inline CUDA_CALLABLE Type tensordot(const transform_t<Type>& a, const transform_t<Type>& b)
-{
-    // corresponds to `np.tensordot()` with all axes being contracted
-    return tensordot(a.p, b.p) + tensordot(a.q, b.q);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE Type extract(const transform_t<Type>& t, int i)
-{
-    return t[i];
-}
-
-template<typename Type>
-inline void CUDA_CALLABLE adj_extract(const transform_t<Type>& t, int i, transform_t<Type>& adj_t, int& adj_i, Type adj_ret)
-{
-    adj_t[i] += adj_ret;
-}
-
-
-// adjoint methods
-template<typename Type>
-CUDA_CALLABLE inline void adj_add(const transform_t<Type>& a, const transform_t<Type>& b, transform_t<Type>& adj_a, transform_t<Type>& adj_b, const transform_t<Type>& adj_ret)
-{
-    adj_add(a.p, b.p, adj_a.p, adj_b.p, adj_ret.p);
-    adj_add(a.q, b.q, adj_a.q, adj_b.q, adj_ret.q);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_sub(const transform_t<Type>& a, const transform_t<Type>& b, transform_t<Type>& adj_a, transform_t<Type>& adj_b, const transform_t<Type>& adj_ret)
-{
-    adj_sub(a.p, b.p, adj_a.p, adj_b.p, adj_ret.p);
-    adj_sub(a.q, b.q, adj_a.q, adj_b.q, adj_ret.q);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_mul(const transform_t<Type>& a, Type s, transform_t<Type>& adj_a, Type& adj_s, const transform_t<Type>& adj_ret)
-{
-    adj_mul(a.p, s, adj_a.p, adj_s, adj_ret.p);
-    adj_mul(a.q, s, adj_a.q, adj_s, adj_ret.q);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_mul(Type s, const transform_t<Type>& a, Type& adj_s, transform_t<Type>& adj_a, const transform_t<Type>& adj_ret)
-{
-    adj_mul(a, s, adj_a, adj_s, adj_ret);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_mul(const transform_t<Type>& a, const transform_t<Type>& b, transform_t<Type>& adj_a, transform_t<Type>& adj_b, const transform_t<Type>& adj_ret)
-{
-    adj_transform_multiply(a, b, adj_a, adj_b, adj_ret);
-}
-
-
-template<typename Type>
-inline CUDA_CALLABLE transform_t<Type> atomic_add(transform_t<Type>* addr, const transform_t<Type>& value) 
-{   
-    vec_t<3,Type> p = atomic_add(&addr->p, value.p);
-    quat_t<Type> q = atomic_add(&addr->q, value.q);
-
-    return transform_t<Type>(p, q);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_transform_t(const vec_t<3,Type>& p, const quat_t<Type>& q, vec_t<3,Type>& adj_p, quat_t<Type>& adj_q, const transform_t<Type>& adj_ret)
-{
-    adj_p += adj_ret.p;
-    adj_q += adj_ret.q;
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_transform_get_translation(const transform_t<Type>& t, transform_t<Type>& adj_t, const vec_t<3,Type>& adj_ret)
-{
-    adj_t.p += adj_ret;
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_transform_get_rotation(const transform_t<Type>& t, transform_t<Type>& adj_t, const quat_t<Type>& adj_ret)
-{
-    adj_t.q += adj_ret;
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_transform_inverse(const transform_t<Type>& t, transform_t<Type>& adj_t, const transform_t<Type>& adj_ret)
-{
-
-    // forward
-    quat_t<Type> q_inv = quat_inverse(t.q); 
-    vec_t<3,Type> p = quat_rotate(q_inv, t.p);
-    vec_t<3,Type> np = -p;
-    // transform<Type> t = transform<Type>(np, q_inv)
-
-    // backward
-    quat_t<Type> adj_q_inv(0.0f);
-    quat_t<Type> adj_q(0.0f);
-    vec_t<3,Type> adj_p(0.0f);
-    vec_t<3,Type> adj_np(0.0f);
-
-    adj_transform_t(np, q_inv, adj_np, adj_q_inv, adj_ret);
-    adj_p = -adj_np;
-    adj_quat_rotate(q_inv, t.p, adj_q_inv, adj_t.p, adj_p);
-    adj_quat_inverse(t.q, adj_t.q, adj_q_inv);
-    
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_transform_vector(const transform_t<Type>& t, const vec_t<3,Type>& x, transform_t<Type>& adj_t, vec_t<3,Type>& adj_x, const vec_t<3,Type>& adj_ret)
-{
-    adj_quat_rotate(t.q, x, adj_t.q, adj_x, adj_ret);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_transform_point(const transform_t<Type>& t, const vec_t<3,Type>& x, transform_t<Type>& adj_t, vec_t<3,Type>& adj_x, const vec_t<3,Type>& adj_ret)
-{
-    adj_quat_rotate(t.q, x, adj_t.q, adj_x, adj_ret);
-    adj_t.p += adj_ret;
-}
-
-
-template<typename Type>
-CUDA_CALLABLE void print(transform_t<Type> t);
-
-template<typename Type>
-CUDA_CALLABLE inline transform_t<Type> lerp(const transform_t<Type>& a, const transform_t<Type>& b, Type t)
-{
-    return a*(Type(1)-t) + b*t;
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_lerp(const transform_t<Type>& a, const transform_t<Type>& b, Type t, transform_t<Type>& adj_a, transform_t<Type>& adj_b, Type& adj_t, const transform_t<Type>& adj_ret)
-{
-    adj_a += adj_ret*(Type(1)-t);
-    adj_b += adj_ret*t;
-    adj_t += tensordot(b, adj_ret) - tensordot(a, adj_ret);
-}
-
-template<typename Type>
-using spatial_matrix_t = mat_t<6,6,Type>;
-
-template<typename Type>
-inline CUDA_CALLABLE spatial_matrix_t<Type> spatial_adjoint(const mat_t<3,3,Type>& R, const mat_t<3,3,Type>& S)
-{    
-    spatial_matrix_t<Type> adT;
-
-    // T = [Rah,   0]
-    //     [S  R]
-
-    // diagonal blocks    
-    for (int i=0; i < 3; ++i)
-    {
-        for (int j=0; j < 3; ++j)
-        {
-            adT.data[i][j] = R.data[i][j];
-            adT.data[i+3][j+3] = R.data[i][j];
-        }
-    }
-
-    // lower off diagonal
-    for (int i=0; i < 3; ++i)
-    {
-        for (int j=0; j < 3; ++j)
-        {
-            adT.data[i+3][j] = S.data[i][j];
-        }
-    }
-
-    return adT;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_spatial_adjoint(const mat_t<3,3,Type>& R, const mat_t<3,3,Type>& S, mat_t<3,3,Type>& adj_R, mat_t<3,3,Type>& adj_S, const spatial_matrix_t<Type>& adj_ret)
-{
-    // diagonal blocks    
-    for (int i=0; i < 3; ++i)
-    {
-        for (int j=0; j < 3; ++j)
-        {
-            adj_R.data[i][j] += adj_ret.data[i][j];
-            adj_R.data[i][j] += adj_ret.data[i+3][j+3];
-        }
-    }
-
-    // lower off diagonal
-    for (int i=0; i < 3; ++i)
-    {
-        for (int j=0; j < 3; ++j)
-        {
-            adj_S.data[i][j] += adj_ret.data[i+3][j];
-        }
-    }
-}
-
-
-CUDA_CALLABLE inline int row_index(int stride, int i, int j)
-{
-    return i*stride + j;
-}
-
-// builds spatial Jacobian J which is an (joint_count*6)x(dof_count) matrix
-template<typename Type>
-CUDA_CALLABLE inline void spatial_jacobian(
-    const spatial_vector_t<Type>* S,
-    const int* joint_parents, 
-    const int* joint_qd_start, 
-    int joint_start,    // offset of the first joint for the articulation
-    int joint_count,    
-    int J_start,
-    Type* J)
-{
-    const int articulation_dof_start = joint_qd_start[joint_start];
-    const int articulation_dof_end = joint_qd_start[joint_start + joint_count];
-    const int articulation_dof_count = articulation_dof_end-articulation_dof_start;
-
-	// shift output pointers
-	const int S_start = articulation_dof_start;
-
-	S += S_start;
-	J += J_start;
-	
-    for (int i=0; i < joint_count; ++i)
-    {
-        const int row_start = i * 6;
-
-        int j = joint_start + i;
-        while (j != -1)
-        {
-            const int joint_dof_start = joint_qd_start[j];
-            const int joint_dof_end = joint_qd_start[j+1];
-            const int joint_dof_count = joint_dof_end-joint_dof_start;
-
-            // fill out each row of the Jacobian walking up the tree
-            //for (int col=dof_start; col < dof_end; ++col)
-            for (int dof=0; dof < joint_dof_count; ++dof)
-            {
-                const int col = (joint_dof_start-articulation_dof_start) + dof;
-
-                J[row_index(articulation_dof_count, row_start+0, col)] = S[col].w[0];
-                J[row_index(articulation_dof_count, row_start+1, col)] = S[col].w[1];
-                J[row_index(articulation_dof_count, row_start+2, col)] = S[col].w[2];
-                J[row_index(articulation_dof_count, row_start+3, col)] = S[col].v[0];
-                J[row_index(articulation_dof_count, row_start+4, col)] = S[col].v[1];
-                J[row_index(articulation_dof_count, row_start+5, col)] = S[col].v[2];
-            }
-
-            j = joint_parents[j];
-        }
-    }
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_spatial_jacobian(
-    const spatial_vector_t<Type>* S, 
-    const int* joint_parents, 
-    const int* joint_qd_start, 
-    const int joint_start,
-    const int joint_count, 
-    const int J_start, 
-    const Type* J,
-    // adjs
-    spatial_vector_t<Type>* adj_S, 
-    int* adj_joint_parents, 
-    int* adj_joint_qd_start, 
-    int& adj_joint_start,
-    int& adj_joint_count, 
-    int& adj_J_start, 
-    const Type* adj_J)
-{   
-    const int articulation_dof_start = joint_qd_start[joint_start];
-    const int articulation_dof_end = joint_qd_start[joint_start + joint_count];
-    const int articulation_dof_count = articulation_dof_end-articulation_dof_start;
-
-	// shift output pointers
-	const int S_start = articulation_dof_start;
-
-	S += S_start;
-	J += J_start;
-
-    adj_S += S_start;
-    adj_J += J_start;
-	
-    for (int i=0; i < joint_count; ++i)
-    {
-        const int row_start = i * 6;
-
-        int j = joint_start + i;
-        while (j != -1)
-        {
-            const int joint_dof_start = joint_qd_start[j];
-            const int joint_dof_end = joint_qd_start[j+1];
-            const int joint_dof_count = joint_dof_end-joint_dof_start;
-
-            // fill out each row of the Jacobian walking up the tree
-            //for (int col=dof_start; col < dof_end; ++col)
-            for (int dof=0; dof < joint_dof_count; ++dof)
-            {
-                const int col = (joint_dof_start-articulation_dof_start) + dof;
-
-                adj_S[col].w[0] += adj_J[row_index(articulation_dof_count, row_start+0, col)];
-                adj_S[col].w[1] += adj_J[row_index(articulation_dof_count, row_start+1, col)];
-                adj_S[col].w[2] += adj_J[row_index(articulation_dof_count, row_start+2, col)];
-                adj_S[col].v[0] += adj_J[row_index(articulation_dof_count, row_start+3, col)];
-                adj_S[col].v[1] += adj_J[row_index(articulation_dof_count, row_start+4, col)];
-                adj_S[col].v[2] += adj_J[row_index(articulation_dof_count, row_start+5, col)];
-            }
-
-            j = joint_parents[j];
-        }
-    }
-}
-
-
-template<typename Type>
-CUDA_CALLABLE inline void spatial_mass(const spatial_matrix_t<Type>* I_s, int joint_start, int joint_count, int M_start, Type* M)
-{
-    const int stride = joint_count*6;
-
-    for (int l=0; l < joint_count; ++l)
-    {
-        for (int i=0; i < 6; ++i)
-        {
-            for (int j=0; j < 6; ++j)
-            {
-                M[M_start + row_index(stride, l*6 + i, l*6 + j)] = I_s[joint_start + l].data[i][j];
-            }
-        }
-    } 
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_spatial_mass(
-    const spatial_matrix_t<Type>* I_s, 
-    const int joint_start,
-    const int joint_count, 
-    const int M_start,
-    const Type* M,
-    spatial_matrix_t<Type>* adj_I_s, 
-    int& adj_joint_start,
-    int& adj_joint_count, 
-    int& adj_M_start,
-    const Type* adj_M)
-{
-    const int stride = joint_count*6;
-
-    for (int l=0; l < joint_count; ++l)
-    {
-        for (int i=0; i < 6; ++i)
-        {
-            for (int j=0; j < 6; ++j)
-            {
-                adj_I_s[joint_start + l].data[i][j] += adj_M[M_start + row_index(stride, l*6 + i, l*6 + j)];
-            }
-        }
-    } 
-}
-
-using transform = transform_t<float>;
-using transformh = transform_t<half>;
-using transformf = transform_t<float>;
-using transformd = transform_t<double>;
-
-using spatial_vector = spatial_vector_t<float>;
-using spatial_vectorh = spatial_vector_t<half>;
-using spatial_vectorf = spatial_vector_t<float>;
-using spatial_vectord = spatial_vector_t<double>;
-
-using spatial_matrix = spatial_matrix_t<float>;
-using spatial_matrixh = spatial_matrix_t<half>;
-using spatial_matrixf = spatial_matrix_t<float>;
-using spatial_matrixd = spatial_matrix_t<double>;
-
+/** Copyright (c) 2022 NVIDIA CORPORATION.  All rights reserved.
+ * NVIDIA CORPORATION and its licensors retain all intellectual property
+ * and proprietary rights in and to this software, related documentation
+ * and any modifications thereto.  Any use, reproduction, disclosure or
+ * distribution of this software and related documentation without an express
+ * license agreement from NVIDIA CORPORATION is strictly prohibited.
+ */
+
+#pragma once
+
+namespace wp
+{
+
+//---------------------------------------------------------------------------------
+// Represents a twist in se(3)
+template <typename Type>
+using spatial_vector_t = vec_t<6,Type>;
+
+template<typename Type>
+CUDA_CALLABLE inline Type spatial_dot(const spatial_vector_t<Type>& a, const spatial_vector_t<Type>& b)
+{
+    return dot(a, b);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline vec_t<3,Type> &w_vec( spatial_vector_t<Type>& a )
+{
+    return *(vec_t<3,Type>*)(&a);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline vec_t<3,Type> &v_vec( spatial_vector_t<Type>& a )
+{
+    return *(vec_t<3,Type>*)(&a.c[3]);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline const vec_t<3,Type> &w_vec( const spatial_vector_t<Type>& a )
+{
+    spatial_vector_t<Type> &non_const_vec = *(spatial_vector_t<Type>*)(const_cast<Type*>(&a.c[0]));
+    return w_vec(non_const_vec);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline const vec_t<3,Type> &v_vec( const spatial_vector_t<Type>& a )
+{
+    spatial_vector_t<Type> &non_const_vec = *(spatial_vector_t<Type>*)(const_cast<Type*>(&a.c[0]));
+    return v_vec(non_const_vec);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline spatial_vector_t<Type> spatial_cross(const spatial_vector_t<Type>& a,  const spatial_vector_t<Type>& b)
+{
+    vec_t<3,Type> w = cross(w_vec(a), w_vec(b));
+    vec_t<3,Type> v = cross(v_vec(a), w_vec(b)) + cross(w_vec(a), v_vec(b));
+    
+    return spatial_vector_t<Type>({w[0], w[1], w[2], v[0], v[1], v[2]});
+}
+
+template<typename Type>
+CUDA_CALLABLE inline spatial_vector_t<Type> spatial_cross_dual(const spatial_vector_t<Type>& a,  const spatial_vector_t<Type>& b)
+{
+    vec_t<3,Type> w = cross(w_vec(a), w_vec(b)) + cross(v_vec(a), v_vec(b));
+    vec_t<3,Type> v = cross(w_vec(a), v_vec(b));
+
+    return spatial_vector_t<Type>({w[0], w[1], w[2], v[0], v[1], v[2]});
+}
+
+template<typename Type>
+CUDA_CALLABLE inline vec_t<3,Type> spatial_top(const spatial_vector_t<Type>& a)
+{
+    return w_vec(a);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline vec_t<3,Type> spatial_bottom(const spatial_vector_t<Type>& a)
+{
+    return v_vec(a);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_spatial_dot(const spatial_vector_t<Type>& a, const spatial_vector_t<Type>& b, spatial_vector_t<Type>& adj_a, spatial_vector_t<Type>& adj_b, const Type& adj_ret)
+{
+    adj_dot(a, b, adj_a, adj_b, adj_ret);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_spatial_cross(const spatial_vector_t<Type>& a,  const spatial_vector_t<Type>& b, spatial_vector_t<Type>& adj_a,  spatial_vector_t<Type>& adj_b, const spatial_vector_t<Type>& adj_ret)
+{
+    adj_cross(w_vec(a), w_vec(b), w_vec(adj_a), w_vec(adj_b), w_vec(adj_ret));
+    
+    adj_cross(v_vec(a), w_vec(b), v_vec(adj_a), w_vec(adj_b), v_vec(adj_ret));
+    adj_cross(w_vec(a), v_vec(b), w_vec(adj_a), v_vec(adj_b), v_vec(adj_ret));
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_spatial_cross_dual(const spatial_vector_t<Type>& a,  const spatial_vector_t<Type>& b, spatial_vector_t<Type>& adj_a, spatial_vector_t<Type>& adj_b, const spatial_vector_t<Type>& adj_ret)
+{
+    adj_cross(w_vec(a), w_vec(b), w_vec(adj_a), w_vec(adj_b), w_vec(adj_ret));
+    adj_cross(v_vec(a), v_vec(b), v_vec(adj_a), v_vec(adj_b), w_vec(adj_ret));
+
+    adj_cross(w_vec(a), v_vec(b), w_vec(adj_a), v_vec(adj_b), v_vec(adj_ret));
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_spatial_top(const spatial_vector_t<Type>& a, spatial_vector_t<Type>& adj_a, const vec_t<3,Type>& adj_ret)
+{
+    w_vec(adj_a) += adj_ret;
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_spatial_bottom(const spatial_vector_t<Type>& a, spatial_vector_t<Type>& adj_a, const vec_t<3,Type>& adj_ret)
+{
+    v_vec(adj_a) += adj_ret;
+}
+
+
+//---------------------------------------------------------------------------------
+// Represents a rigid body transform<Type>ation
+
+template<typename Type>
+struct transform_t
+{
+    vec_t<3,Type> p;
+    quat_t<Type> q;
+
+    CUDA_CALLABLE inline transform_t(vec_t<3,Type> p=vec_t<3,Type>(), quat_t<Type> q=quat_t<Type>()) : p(p), q(q) {}
+    CUDA_CALLABLE inline transform_t(Type)  {}  // helps uniform initialization
+
+    CUDA_CALLABLE inline Type operator[](int index) const
+    {
+        assert(index < 7);
+
+        return p.c[index];
+    }
+
+    CUDA_CALLABLE inline Type& operator[](int index)
+    {
+        assert(index < 7);
+
+        return p.c[index];
+    }    
+};
+
+template<typename Type=float32>
+CUDA_CALLABLE inline transform_t<Type> transform_identity()
+{
+    return transform_t<Type>(vec_t<3,Type>(), quat_identity<Type>());
+}
+
+template<typename Type>
+inline CUDA_CALLABLE bool operator==(const transform_t<Type>& a, const transform_t<Type>& b)
+{
+    return a.p == b.p && a.q == b.q;
+}
+
+
+template<typename Type>
+inline bool CUDA_CALLABLE isfinite(const transform_t<Type>& t)
+{
+    return isfinite(t.p) && isfinite(t.q);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline vec_t<3,Type> transform_get_translation(const transform_t<Type>& t)
+{
+    return t.p;
+}
+
+template<typename Type>
+CUDA_CALLABLE inline quat_t<Type> transform_get_rotation(const transform_t<Type>& t)
+{
+    return t.q;
+}
+
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> transform_multiply(const transform_t<Type>& a, const transform_t<Type>& b)
+{
+    return { quat_rotate(a.q, b.p) + a.p, mul(a.q, b.q) };
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_transform_multiply(const transform_t<Type>& a, const transform_t<Type>& b, transform_t<Type>& adj_a, transform_t<Type>& adj_b, const transform_t<Type>& adj_ret)
+{
+    // translational part
+    adj_quat_rotate(a.q, b.p, adj_a.q, adj_b.p, adj_ret.p);
+    adj_a.p += adj_ret.p;
+
+    // rotational part
+    adj_mul(a.q, b.q, adj_a.q, adj_b.q, adj_ret.q);
+}
+
+
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> transform_inverse(const transform_t<Type>& t)
+{
+    quat_t<Type> q_inv = quat_inverse(t.q);
+    return transform_t<Type>(-quat_rotate(q_inv, t.p), q_inv);
+}
+
+    
+template<typename Type>
+CUDA_CALLABLE inline vec_t<3,Type> transform_vector(const transform_t<Type>& t, const vec_t<3,Type>& x)
+{
+    return quat_rotate(t.q, x);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline vec_t<3,Type> transform_point(const transform_t<Type>& t, const vec_t<3,Type>& x)
+{
+    return t.p + quat_rotate(t.q, x);
+}
+
+// not totally sure why you'd want to do this seeing as adding/subtracting two rotation
+// quats doesn't seem to do anything meaningful
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> add(const transform_t<Type>& a, const transform_t<Type>& b)
+{
+    return { a.p + b.p, a.q + b.q };
+}
+
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> sub(const transform_t<Type>& a, const transform_t<Type>& b)
+{
+    return { a.p - b.p, a.q - b.q };
+}
+
+// also not sure why you'd want to do this seeing as the quat would end up unnormalized
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> mul(const transform_t<Type>& a, Type s)
+{
+    return { a.p*s, a.q*s };
+}
+
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> mul(Type s, const transform_t<Type>& a)
+{
+    return mul(a, s);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> mul(const transform_t<Type>& a, const transform_t<Type>& b)
+{
+    return transform_multiply(a, b);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> operator*(const transform_t<Type>& a, Type s)
+{
+    return mul(a, s);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> operator*(Type s, const transform_t<Type>& a)
+{
+    return mul(a, s);
+}
+
+
+template<typename Type>
+inline CUDA_CALLABLE Type tensordot(const transform_t<Type>& a, const transform_t<Type>& b)
+{
+    // corresponds to `np.tensordot()` with all axes being contracted
+    return tensordot(a.p, b.p) + tensordot(a.q, b.q);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE Type extract(const transform_t<Type>& t, int i)
+{
+    return t[i];
+}
+
+template<typename Type>
+inline void CUDA_CALLABLE adj_extract(const transform_t<Type>& t, int i, transform_t<Type>& adj_t, int& adj_i, Type adj_ret)
+{
+    adj_t[i] += adj_ret;
+}
+
+
+// adjoint methods
+template<typename Type>
+CUDA_CALLABLE inline void adj_add(const transform_t<Type>& a, const transform_t<Type>& b, transform_t<Type>& adj_a, transform_t<Type>& adj_b, const transform_t<Type>& adj_ret)
+{
+    adj_add(a.p, b.p, adj_a.p, adj_b.p, adj_ret.p);
+    adj_add(a.q, b.q, adj_a.q, adj_b.q, adj_ret.q);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_sub(const transform_t<Type>& a, const transform_t<Type>& b, transform_t<Type>& adj_a, transform_t<Type>& adj_b, const transform_t<Type>& adj_ret)
+{
+    adj_sub(a.p, b.p, adj_a.p, adj_b.p, adj_ret.p);
+    adj_sub(a.q, b.q, adj_a.q, adj_b.q, adj_ret.q);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_mul(const transform_t<Type>& a, Type s, transform_t<Type>& adj_a, Type& adj_s, const transform_t<Type>& adj_ret)
+{
+    adj_mul(a.p, s, adj_a.p, adj_s, adj_ret.p);
+    adj_mul(a.q, s, adj_a.q, adj_s, adj_ret.q);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_mul(Type s, const transform_t<Type>& a, Type& adj_s, transform_t<Type>& adj_a, const transform_t<Type>& adj_ret)
+{
+    adj_mul(a, s, adj_a, adj_s, adj_ret);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_mul(const transform_t<Type>& a, const transform_t<Type>& b, transform_t<Type>& adj_a, transform_t<Type>& adj_b, const transform_t<Type>& adj_ret)
+{
+    adj_transform_multiply(a, b, adj_a, adj_b, adj_ret);
+}
+
+
+template<typename Type>
+inline CUDA_CALLABLE transform_t<Type> atomic_add(transform_t<Type>* addr, const transform_t<Type>& value) 
+{   
+    vec_t<3,Type> p = atomic_add(&addr->p, value.p);
+    quat_t<Type> q = atomic_add(&addr->q, value.q);
+
+    return transform_t<Type>(p, q);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_transform_t(const vec_t<3,Type>& p, const quat_t<Type>& q, vec_t<3,Type>& adj_p, quat_t<Type>& adj_q, const transform_t<Type>& adj_ret)
+{
+    adj_p += adj_ret.p;
+    adj_q += adj_ret.q;
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_transform_get_translation(const transform_t<Type>& t, transform_t<Type>& adj_t, const vec_t<3,Type>& adj_ret)
+{
+    adj_t.p += adj_ret;
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_transform_get_rotation(const transform_t<Type>& t, transform_t<Type>& adj_t, const quat_t<Type>& adj_ret)
+{
+    adj_t.q += adj_ret;
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_transform_inverse(const transform_t<Type>& t, transform_t<Type>& adj_t, const transform_t<Type>& adj_ret)
+{
+
+    // forward
+    quat_t<Type> q_inv = quat_inverse(t.q); 
+    vec_t<3,Type> p = quat_rotate(q_inv, t.p);
+    vec_t<3,Type> np = -p;
+    // transform<Type> t = transform<Type>(np, q_inv)
+
+    // backward
+    quat_t<Type> adj_q_inv(0.0f);
+    quat_t<Type> adj_q(0.0f);
+    vec_t<3,Type> adj_p(0.0f);
+    vec_t<3,Type> adj_np(0.0f);
+
+    adj_transform_t(np, q_inv, adj_np, adj_q_inv, adj_ret);
+    adj_p = -adj_np;
+    adj_quat_rotate(q_inv, t.p, adj_q_inv, adj_t.p, adj_p);
+    adj_quat_inverse(t.q, adj_t.q, adj_q_inv);
+    
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_transform_vector(const transform_t<Type>& t, const vec_t<3,Type>& x, transform_t<Type>& adj_t, vec_t<3,Type>& adj_x, const vec_t<3,Type>& adj_ret)
+{
+    adj_quat_rotate(t.q, x, adj_t.q, adj_x, adj_ret);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_transform_point(const transform_t<Type>& t, const vec_t<3,Type>& x, transform_t<Type>& adj_t, vec_t<3,Type>& adj_x, const vec_t<3,Type>& adj_ret)
+{
+    adj_quat_rotate(t.q, x, adj_t.q, adj_x, adj_ret);
+    adj_t.p += adj_ret;
+}
+
+
+template<typename Type>
+CUDA_CALLABLE void print(transform_t<Type> t);
+
+template<typename Type>
+CUDA_CALLABLE inline transform_t<Type> lerp(const transform_t<Type>& a, const transform_t<Type>& b, Type t)
+{
+    return a*(Type(1)-t) + b*t;
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_lerp(const transform_t<Type>& a, const transform_t<Type>& b, Type t, transform_t<Type>& adj_a, transform_t<Type>& adj_b, Type& adj_t, const transform_t<Type>& adj_ret)
+{
+    adj_a += adj_ret*(Type(1)-t);
+    adj_b += adj_ret*t;
+    adj_t += tensordot(b, adj_ret) - tensordot(a, adj_ret);
+}
+
+template<typename Type>
+using spatial_matrix_t = mat_t<6,6,Type>;
+
+template<typename Type>
+inline CUDA_CALLABLE spatial_matrix_t<Type> spatial_adjoint(const mat_t<3,3,Type>& R, const mat_t<3,3,Type>& S)
+{    
+    spatial_matrix_t<Type> adT;
+
+    // T = [Rah,   0]
+    //     [S  R]
+
+    // diagonal blocks    
+    for (int i=0; i < 3; ++i)
+    {
+        for (int j=0; j < 3; ++j)
+        {
+            adT.data[i][j] = R.data[i][j];
+            adT.data[i+3][j+3] = R.data[i][j];
+        }
+    }
+
+    // lower off diagonal
+    for (int i=0; i < 3; ++i)
+    {
+        for (int j=0; j < 3; ++j)
+        {
+            adT.data[i+3][j] = S.data[i][j];
+        }
+    }
+
+    return adT;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_spatial_adjoint(const mat_t<3,3,Type>& R, const mat_t<3,3,Type>& S, mat_t<3,3,Type>& adj_R, mat_t<3,3,Type>& adj_S, const spatial_matrix_t<Type>& adj_ret)
+{
+    // diagonal blocks    
+    for (int i=0; i < 3; ++i)
+    {
+        for (int j=0; j < 3; ++j)
+        {
+            adj_R.data[i][j] += adj_ret.data[i][j];
+            adj_R.data[i][j] += adj_ret.data[i+3][j+3];
+        }
+    }
+
+    // lower off diagonal
+    for (int i=0; i < 3; ++i)
+    {
+        for (int j=0; j < 3; ++j)
+        {
+            adj_S.data[i][j] += adj_ret.data[i+3][j];
+        }
+    }
+}
+
+
+CUDA_CALLABLE inline int row_index(int stride, int i, int j)
+{
+    return i*stride + j;
+}
+
+// builds spatial Jacobian J which is an (joint_count*6)x(dof_count) matrix
+template<typename Type>
+CUDA_CALLABLE inline void spatial_jacobian(
+    const spatial_vector_t<Type>* S,
+    const int* joint_parents, 
+    const int* joint_qd_start, 
+    int joint_start,    // offset of the first joint for the articulation
+    int joint_count,    
+    int J_start,
+    Type* J)
+{
+    const int articulation_dof_start = joint_qd_start[joint_start];
+    const int articulation_dof_end = joint_qd_start[joint_start + joint_count];
+    const int articulation_dof_count = articulation_dof_end-articulation_dof_start;
+
+	// shift output pointers
+	const int S_start = articulation_dof_start;
+
+	S += S_start;
+	J += J_start;
+	
+    for (int i=0; i < joint_count; ++i)
+    {
+        const int row_start = i * 6;
+
+        int j = joint_start + i;
+        while (j != -1)
+        {
+            const int joint_dof_start = joint_qd_start[j];
+            const int joint_dof_end = joint_qd_start[j+1];
+            const int joint_dof_count = joint_dof_end-joint_dof_start;
+
+            // fill out each row of the Jacobian walking up the tree
+            //for (int col=dof_start; col < dof_end; ++col)
+            for (int dof=0; dof < joint_dof_count; ++dof)
+            {
+                const int col = (joint_dof_start-articulation_dof_start) + dof;
+
+                J[row_index(articulation_dof_count, row_start+0, col)] = S[col].w[0];
+                J[row_index(articulation_dof_count, row_start+1, col)] = S[col].w[1];
+                J[row_index(articulation_dof_count, row_start+2, col)] = S[col].w[2];
+                J[row_index(articulation_dof_count, row_start+3, col)] = S[col].v[0];
+                J[row_index(articulation_dof_count, row_start+4, col)] = S[col].v[1];
+                J[row_index(articulation_dof_count, row_start+5, col)] = S[col].v[2];
+            }
+
+            j = joint_parents[j];
+        }
+    }
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_spatial_jacobian(
+    const spatial_vector_t<Type>* S, 
+    const int* joint_parents, 
+    const int* joint_qd_start, 
+    const int joint_start,
+    const int joint_count, 
+    const int J_start, 
+    const Type* J,
+    // adjs
+    spatial_vector_t<Type>* adj_S, 
+    int* adj_joint_parents, 
+    int* adj_joint_qd_start, 
+    int& adj_joint_start,
+    int& adj_joint_count, 
+    int& adj_J_start, 
+    const Type* adj_J)
+{   
+    const int articulation_dof_start = joint_qd_start[joint_start];
+    const int articulation_dof_end = joint_qd_start[joint_start + joint_count];
+    const int articulation_dof_count = articulation_dof_end-articulation_dof_start;
+
+	// shift output pointers
+	const int S_start = articulation_dof_start;
+
+	S += S_start;
+	J += J_start;
+
+    adj_S += S_start;
+    adj_J += J_start;
+	
+    for (int i=0; i < joint_count; ++i)
+    {
+        const int row_start = i * 6;
+
+        int j = joint_start + i;
+        while (j != -1)
+        {
+            const int joint_dof_start = joint_qd_start[j];
+            const int joint_dof_end = joint_qd_start[j+1];
+            const int joint_dof_count = joint_dof_end-joint_dof_start;
+
+            // fill out each row of the Jacobian walking up the tree
+            //for (int col=dof_start; col < dof_end; ++col)
+            for (int dof=0; dof < joint_dof_count; ++dof)
+            {
+                const int col = (joint_dof_start-articulation_dof_start) + dof;
+
+                adj_S[col].w[0] += adj_J[row_index(articulation_dof_count, row_start+0, col)];
+                adj_S[col].w[1] += adj_J[row_index(articulation_dof_count, row_start+1, col)];
+                adj_S[col].w[2] += adj_J[row_index(articulation_dof_count, row_start+2, col)];
+                adj_S[col].v[0] += adj_J[row_index(articulation_dof_count, row_start+3, col)];
+                adj_S[col].v[1] += adj_J[row_index(articulation_dof_count, row_start+4, col)];
+                adj_S[col].v[2] += adj_J[row_index(articulation_dof_count, row_start+5, col)];
+            }
+
+            j = joint_parents[j];
+        }
+    }
+}
+
+
+template<typename Type>
+CUDA_CALLABLE inline void spatial_mass(const spatial_matrix_t<Type>* I_s, int joint_start, int joint_count, int M_start, Type* M)
+{
+    const int stride = joint_count*6;
+
+    for (int l=0; l < joint_count; ++l)
+    {
+        for (int i=0; i < 6; ++i)
+        {
+            for (int j=0; j < 6; ++j)
+            {
+                M[M_start + row_index(stride, l*6 + i, l*6 + j)] = I_s[joint_start + l].data[i][j];
+            }
+        }
+    } 
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_spatial_mass(
+    const spatial_matrix_t<Type>* I_s, 
+    const int joint_start,
+    const int joint_count, 
+    const int M_start,
+    const Type* M,
+    spatial_matrix_t<Type>* adj_I_s, 
+    int& adj_joint_start,
+    int& adj_joint_count, 
+    int& adj_M_start,
+    const Type* adj_M)
+{
+    const int stride = joint_count*6;
+
+    for (int l=0; l < joint_count; ++l)
+    {
+        for (int i=0; i < 6; ++i)
+        {
+            for (int j=0; j < 6; ++j)
+            {
+                adj_I_s[joint_start + l].data[i][j] += adj_M[M_start + row_index(stride, l*6 + i, l*6 + j)];
+            }
+        }
+    } 
+}
+
+using transform = transform_t<float>;
+using transformh = transform_t<half>;
+using transformf = transform_t<float>;
+using transformd = transform_t<double>;
+
+using spatial_vector = spatial_vector_t<float>;
+using spatial_vectorh = spatial_vector_t<half>;
+using spatial_vectorf = spatial_vector_t<float>;
+using spatial_vectord = spatial_vector_t<double>;
+
+using spatial_matrix = spatial_matrix_t<float>;
+using spatial_matrixh = spatial_matrix_t<half>;
+using spatial_matrixf = spatial_matrix_t<float>;
+using spatial_matrixd = spatial_matrix_t<double>;
+
  } // namespace wp