warp-lang 1.0.2__py3-none-manylinux2014_x86_64.whl → 1.2.0__py3-none-manylinux2014_x86_64.whl

warp/native/mat.h

@@ -1,1498 +1,1545 @@
-/** 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
-
-#include "initializer_array.h"
-
-namespace wp
-{
-
-//----------------------------------------------------------
-// mat
-template<typename T>
-struct quat_t;
-
-template<unsigned Rows, unsigned Cols, typename Type>
-struct mat_t
-{
-    inline CUDA_CALLABLE mat_t()
-        : data()
-    {}
-
-    inline CUDA_CALLABLE mat_t(Type s)
-    {
-        for (unsigned i=0; i < Rows; ++i)
-            for (unsigned j=0; j < Cols; ++j)
-                data[i][j] = s;
-    }
-    
-    template <typename OtherType>
-    inline explicit CUDA_CALLABLE mat_t(const mat_t<Rows, Cols, OtherType>& other)
-    {
-        for (unsigned i=0; i < Rows; ++i)
-            for (unsigned j=0; j < Cols; ++j)
-                data[i][j] = other.data[i][j];
-    }
-    
-    inline CUDA_CALLABLE mat_t(vec_t<2,Type> c0, vec_t<2,Type> c1)
-    {
-        data[0][0] = c0[0];
-        data[1][0] = c0[1];
-
-        data[0][1] = c1[0];
-        data[1][1] = c1[1];
-    }
-    
-    inline CUDA_CALLABLE mat_t(vec_t<3,Type> c0, vec_t<3,Type> c1, vec_t<3,Type> c2)
-    {
-        data[0][0] = c0[0];
-        data[1][0] = c0[1];
-        data[2][0] = c0[2];
-
-        data[0][1] = c1[0];
-        data[1][1] = c1[1];
-        data[2][1] = c1[2];
-
-        data[0][2] = c2[0];
-        data[1][2] = c2[1];
-        data[2][2] = c2[2];
-    }
-
-    inline CUDA_CALLABLE mat_t(vec_t<4,Type> c0, vec_t<4,Type> c1, vec_t<4,Type> c2, vec_t<4,Type> c3)
-    {
-        data[0][0] = c0[0];
-        data[1][0] = c0[1];
-        data[2][0] = c0[2];
-        data[3][0] = c0[3];
-
-        data[0][1] = c1[0];
-        data[1][1] = c1[1];
-        data[2][1] = c1[2];
-        data[3][1] = c1[3];
-
-        data[0][2] = c2[0];
-        data[1][2] = c2[1];
-        data[2][2] = c2[2];
-        data[3][2] = c2[3];
-
-        data[0][3] = c3[0];
-        data[1][3] = c3[1];
-        data[2][3] = c3[2];
-        data[3][3] = c3[3];
-    }
-
-    inline CUDA_CALLABLE mat_t(Type m00, Type m01, Type m10, Type m11) 
-    {
-        data[0][0] = m00;
-        data[1][0] = m10;
-        data[0][1] = m01;
-        data[1][1] = m11;
-    }
-    
-    inline CUDA_CALLABLE mat_t(
-        Type m00, Type m01, Type m02,
-        Type m10, Type m11, Type m12,
-        Type m20, Type m21, Type m22)
-    {
-        data[0][0] = m00;
-        data[1][0] = m10;
-        data[2][0] = m20;
-
-        data[0][1] = m01;
-        data[1][1] = m11;
-        data[2][1] = m21;
-
-        data[0][2] = m02;
-        data[1][2] = m12;
-        data[2][2] = m22;
-    }
-
-    inline CUDA_CALLABLE mat_t(
-                 Type m00, Type m01, Type m02, Type m03,
-                 Type m10, Type m11, Type m12, Type m13,
-                 Type m20, Type m21, Type m22, Type m23,
-                 Type m30, Type m31, Type m32, Type m33)
-    {
-        data[0][0] = m00;
-        data[1][0] = m10;
-        data[2][0] = m20;
-        data[3][0] = m30;
-
-        data[0][1] = m01;
-        data[1][1] = m11;
-        data[2][1] = m21;
-        data[3][1] = m31;
-
-        data[0][2] = m02;
-        data[1][2] = m12;
-        data[2][2] = m22;
-        data[3][2] = m32;
-
-        data[0][3] = m03;
-        data[1][3] = m13;
-        data[2][3] = m23;
-        data[3][3] = m33;
-    }
-
-    // implemented in quat.h
-    inline CUDA_CALLABLE mat_t(const vec_t<3,Type>& pos, const quat_t<Type>& rot, const vec_t<3,Type>& scale);
-
-
-    inline CUDA_CALLABLE mat_t(const initializer_array<Rows * Cols, Type> &l)
-    {
-        for (unsigned i=0; i < Rows; ++i)
-        {
-            for (unsigned j=0; j < Cols; ++j)
-            {
-                data[i][j] = l[i * Cols + j];
-            }
-        }
-    }
-
-    inline CUDA_CALLABLE mat_t(const initializer_array<Cols, vec_t<Rows,Type> > &l)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            for (unsigned i=0; i < Rows; ++i)
-            {
-                data[i][j] = l[j][i];
-            }
-        }
-    }
-
-    CUDA_CALLABLE vec_t<Cols,Type> get_row(int index) const
-    {
-        return (vec_t<Cols,Type>&)data[index]; 
-    }
-
-    CUDA_CALLABLE void set_row(int index, const vec_t<Cols,Type>& v)
-    {
-        (vec_t<Cols,Type>&)data[index] = v;
-    }
-
-    CUDA_CALLABLE vec_t<Rows,Type> get_col(int index) const
-    {
-        vec_t<Rows,Type> ret;
-        for( unsigned i=0;i < Rows; ++i )
-        {
-            ret[i] = data[i][index];
-        }
-        return ret;
-    }
-
-    CUDA_CALLABLE void set_col(int index, const vec_t<Rows,Type>& v)
-    {
-        for( unsigned i=0;i < Rows; ++i )
-        {
-            data[i][index] = v[i];
-        }
-    }
-
-    // row major storage assumed to be compatible with PyTorch
-    Type data[Rows][Cols];
-};
-
-
-template<unsigned Rows, typename Type>
-inline CUDA_CALLABLE mat_t<Rows, Rows, Type> identity()
-{
-    mat_t<Rows, Rows, Type> m;
-    for( unsigned i=0; i < Rows; ++i )
-    {
-        m.data[i][i] = Type(1);
-    }
-    return m;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE bool operator==(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
-{
-    for (unsigned i=0; i < Rows; ++i)
-        for (unsigned j=0; j < Cols; ++j)
-            if (a.data[i][j] != b.data[i][j])
-                return false;
-
-    return true;
-}
-
-
-// negation:
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> operator - (mat_t<Rows,Cols,Type> a)
-{
-    // NB: this constructor will initialize all ret's components to 0, which is
-    // unnecessary... 
-    mat_t<Rows,Cols,Type> ret;
-    for (unsigned i=0; i < Rows; ++i)
-        for (unsigned j=0; j < Cols; ++j)
-            ret.data[i][j] = -a.data[i][j];
-
-    // Wonder if this does a load of copying when it returns... hopefully not as it's inlined?
-    return ret;
-}
-
-
-template<unsigned Rows, unsigned Cols, typename Type>
-CUDA_CALLABLE inline mat_t<Rows,Cols,Type> pos(const mat_t<Rows,Cols,Type>& x)
-{
-    return x;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-CUDA_CALLABLE inline void adj_pos(const mat_t<Rows,Cols,Type>& x, mat_t<Rows,Cols,Type>& adj_x, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-    adj_x += adj_ret;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-CUDA_CALLABLE inline mat_t<Rows,Cols,Type> neg(const mat_t<Rows,Cols,Type>& x)
-{
-    return -x;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-CUDA_CALLABLE inline void adj_neg(const mat_t<Rows,Cols,Type>& x, mat_t<Rows,Cols,Type>& adj_x, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-    adj_x -= adj_ret;
-}
-
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> atomic_add(mat_t<Rows,Cols,Type> * addr, mat_t<Rows,Cols,Type> value) 
-{
-    mat_t<Rows,Cols,Type> m;
-    
-    for (unsigned i=0; i < Rows; ++i)
-        for (unsigned j=0; j < Cols; ++j)
-            m.data[i][j] = atomic_add(&addr->data[i][j], value.data[i][j]);
-
-    return m;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> atomic_min(mat_t<Rows,Cols,Type> * addr, mat_t<Rows,Cols,Type> value) 
-{
-    mat_t<Rows,Cols,Type> m;
-    
-    for (unsigned i=0; i < Rows; ++i)
-        for (unsigned j=0; j < Cols; ++j)
-            m.data[i][j] = atomic_min(&addr->data[i][j], value.data[i][j]);
-
-    return m;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> atomic_max(mat_t<Rows,Cols,Type> * addr, mat_t<Rows,Cols,Type> value) 
-{
-    mat_t<Rows,Cols,Type> m;
-    
-    for (unsigned i=0; i < Rows; ++i)
-        for (unsigned j=0; j < Cols; ++j)
-            m.data[i][j] = atomic_max(&addr->data[i][j], value.data[i][j]);
-
-    return m;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_atomic_minmax(
-    mat_t<Rows,Cols,Type> *addr,
-    mat_t<Rows,Cols,Type> *adj_addr,
-    const mat_t<Rows,Cols,Type> &value,
-    mat_t<Rows,Cols,Type> &adj_value)
-{
-    for (unsigned i=0; i < Rows; ++i)
-        for (unsigned j=0; j < Cols; ++j)
-            adj_atomic_minmax(&addr->data[i][j], &adj_addr->data[i][j], value.data[i][j], adj_value.data[i][j]);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE vec_t<Cols,Type> extract(const mat_t<Rows,Cols,Type>& m, int row)
-{
-    vec_t<Cols,Type> ret;
-    for(unsigned i=0; i < Cols; ++i)
-    {
-        ret.c[i] = m.data[row][i];
-    }
-    return ret;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE Type extract(const mat_t<Rows,Cols,Type>& m, int row, int col)
-{
-#ifndef NDEBUG
-    if (row < 0 || row >= Rows)
-    {
-        printf("mat row index %d out of bounds at %s %d\n", row, __FILE__, __LINE__);
-        assert(0);
-    }
-    if (col < 0 || col >= Cols)
-    {
-        printf("mat col index %d out of bounds at %s %d\n", col, __FILE__, __LINE__);
-        assert(0);
-    }
-#endif
-    return m.data[row][col];
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE vec_t<Cols, Type>* index(mat_t<Rows,Cols,Type>& m, int row)
-{
-#ifndef NDEBUG
-    if (row < 0 || row >= Rows)
-    {
-        printf("mat row index %d out of bounds at %s %d\n", row, __FILE__, __LINE__);
-        assert(0);
-    }
-#endif
-
-    return reinterpret_cast<vec_t<Cols, Type>*>(&m.data[row]);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE Type* index(mat_t<Rows,Cols,Type>& m, int row, int col)
-{
-#ifndef NDEBUG
-    if (row < 0 || row >= Rows)
-    {
-        printf("mat row index %d out of bounds at %s %d\n", row, __FILE__, __LINE__);
-        assert(0);
-    }
-    if (col < 0 || col >= Cols)
-    {
-        printf("mat col index %d out of bounds at %s %d\n", col, __FILE__, __LINE__);
-        assert(0);
-    }
-#endif
-
-    return &m.data[row][col];
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_index(const mat_t<Rows,Cols,Type>& m, int row,
-                                       const mat_t<Rows,Cols,Type>& adj_m, int adj_row, const vec_t<Cols, Type>& adj_value)
-{
-    // nop
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_index(const mat_t<Rows,Cols,Type>& m, int row, int col,
-                                       const mat_t<Rows,Cols,Type>& adj_m, int adj_row, int adj_col, Type adj_value)
-{
-    // nop
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline bool CUDA_CALLABLE isfinite(const mat_t<Rows,Cols,Type>& m)
-{
-    for (unsigned i=0; i < Rows; ++i)
-        for (unsigned j=0; j < Cols; ++j)
-            if (!isfinite(m.data[i][j]))
-                return false;
-    return true;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> add(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
-{
-    mat_t<Rows,Cols,Type> t;
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            t.data[i][j] = a.data[i][j] + b.data[i][j];
-        }
-    }
-
-    return t;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> sub(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
-{
-    mat_t<Rows,Cols,Type> t;
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            t.data[i][j] = a.data[i][j] - b.data[i][j];
-        }
-    }
-
-    return t;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> div(const mat_t<Rows,Cols,Type>& a, Type b)
-{
-    mat_t<Rows,Cols,Type> t;
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            t.data[i][j] = a.data[i][j]/b;
-        }
-    }
-
-    return t;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> div(Type b, const mat_t<Rows,Cols,Type>& a)
-{
-    mat_t<Rows,Cols,Type> t;
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            t.data[i][j] = b / a.data[i][j];
-        }
-    }
-
-    return t;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> mul(const mat_t<Rows,Cols,Type>& a, Type b)
-{
-    mat_t<Rows,Cols,Type> t;
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            t.data[i][j] = a.data[i][j]*b;
-        }
-    }
-
-    return t;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> mul(Type b, const mat_t<Rows,Cols,Type>& a)
-{
-    return mul(a,b);
-}
-
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> operator*(Type b, const mat_t<Rows,Cols,Type>& a)
-{
-    return mul(a,b);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> operator*( const mat_t<Rows,Cols,Type>& a, Type b)
-{
-    return mul(a,b);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE vec_t<Rows,Type> mul(const mat_t<Rows,Cols,Type>& a, const vec_t<Cols,Type>& b)
-{
-    vec_t<Rows,Type> r = a.get_col(0)*b[0];
-    for( unsigned i=1; i < Cols; ++i )
-    {
-        r += a.get_col(i)*b[i];
-    }
-    return r;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE vec_t<Cols,Type> mul(const vec_t<Rows,Type>& b, const mat_t<Rows,Cols,Type>& a)
-{
-    vec_t<Cols,Type> r = a.get_row(0)*b[0];
-    for( unsigned i=1; i < Rows; ++i )
-    {
-        r += a.get_row(i)*b[i];
-    }
-    return r;
-}
-
-template<unsigned Rows, unsigned Cols, unsigned ColsOut, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,ColsOut,Type> mul(const mat_t<Rows,Cols,Type>& a, const mat_t<Cols,ColsOut,Type>& b)
-{
-    mat_t<Rows,ColsOut,Type> t(0);
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < ColsOut; ++j)
-        {
-            for (unsigned k=0; k < Cols; ++k)
-            {
-                t.data[i][j] += a.data[i][k]*b.data[k][j];
-            }
-        }
-    }
-    
-    return t;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE Type ddot(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
-{
-    // double dot product between a and b:
-    Type r(0);
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            r += a.data[i][j] * b.data[i][j];
-        }
-    }
-    return r;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE Type tensordot(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
-{
-    // corresponds to `np.tensordot()` with all axes being contracted
-    return ddot(a, b);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Cols,Rows,Type> transpose(const mat_t<Rows,Cols,Type>& a)
-{
-    mat_t<Cols,Rows,Type> t;
-    for (unsigned i=0; i < Cols; ++i)
-    {
-        for (unsigned j=0; j < Rows; ++j)
-        {
-            t.data[i][j] = a.data[j][i];
-        }
-    }
-
-    return t;
-}
-
-// Only implementing determinants for 2x2, 3x3 and 4x4 matrices for now...
-template<typename Type>
-inline CUDA_CALLABLE Type determinant(const mat_t<2,2,Type>& m)
-{
-    return m.data[0][0]*m.data[1][1] - m.data[1][0]*m.data[0][1];
-}
-
-template<typename Type>
-inline CUDA_CALLABLE Type determinant(const mat_t<3,3,Type>& m)
-{
-    return dot(
-        vec_t<3,Type>(m.data[0][0],m.data[0][1],m.data[0][2]),
-        cross(
-            vec_t<3,Type>(m.data[1][0],m.data[1][1],m.data[1][2]),
-            vec_t<3,Type>(m.data[2][0],m.data[2][1],m.data[2][2])
-        )
-    );
-}
-
-template<typename Type>
-inline CUDA_CALLABLE Type determinant(const mat_t<4,4,Type>& m)
-{
-    // adapted from USD GfMatrix4f::Inverse()
-    Type x00, x01, x02, x03;
-    Type x10, x11, x12, x13;
-    Type x20, x21, x22, x23;
-    Type x30, x31, x32, x33;
-    double y01, y02, y03, y12, y13, y23;
-    Type z00, z10, z20, z30;
-
-    // Pickle 1st two columns of matrix into registers
-    x00 = m.data[0][0];
-    x01 = m.data[0][1];
-    x10 = m.data[1][0];
-    x11 = m.data[1][1];
-    x20 = m.data[2][0];
-    x21 = m.data[2][1];
-    x30 = m.data[3][0];
-    x31 = m.data[3][1];
-
-    // Compute all six 2x2 determinants of 1st two columns
-    y01 = x00*x11 - x10*x01;
-    y02 = x00*x21 - x20*x01;
-    y03 = x00*x31 - x30*x01;
-    y12 = x10*x21 - x20*x11;
-    y13 = x10*x31 - x30*x11;
-    y23 = x20*x31 - x30*x21;
-
-    // Pickle 2nd two columns of matrix into registers
-    x02 = m.data[0][2];
-    x03 = m.data[0][3];
-    x12 = m.data[1][2];
-    x13 = m.data[1][3];
-    x22 = m.data[2][2];
-    x23 = m.data[2][3];
-    x32 = m.data[3][2];
-    x33 = m.data[3][3];
-
-    // Compute all six 2x2 determinants of 2nd two columns
-    y01 = x02*x13 - x12*x03;
-    y02 = x02*x23 - x22*x03;
-    y03 = x02*x33 - x32*x03;
-    y12 = x12*x23 - x22*x13;
-    y13 = x12*x33 - x32*x13;
-    y23 = x22*x33 - x32*x23;
-
-    // Compute all 3x3 cofactors for 1st two columns
-    z30 = x11*y02 - x21*y01 - x01*y12;
-    z20 = x01*y13 - x11*y03 + x31*y01;
-    z10 = x21*y03 - x31*y02 - x01*y23;
-    z00 = x11*y23 - x21*y13 + x31*y12;
-
-    // compute 4x4 determinant & its reciprocal
-    double det = x30*z30 + x20*z20 + x10*z10 + x00*z00;
-    return det;
-}
-
-template<unsigned Rows, typename Type>
-inline CUDA_CALLABLE Type trace(const mat_t<Rows,Rows,Type>& m)
-{
-    Type ret = m.data[0][0];
-    for( unsigned i=1; i < Rows; ++i )
-    {
-        ret += m.data[i][i];
-    }
-    return ret;
-}
-
-template<unsigned Rows, typename Type>
-inline CUDA_CALLABLE vec_t<Rows, Type> get_diag(const mat_t<Rows,Rows,Type>& m)
-{
-    vec_t<Rows, Type> ret;
-    for( unsigned i=0; i < Rows; ++i )
-    {
-        ret[i] = m.data[i][i];
-    }
-    return ret;
-}
-
-// Only implementing inverses for 2x2, 3x3 and 4x4 matrices for now...
-template<typename Type>
-inline CUDA_CALLABLE mat_t<2,2,Type> inverse(const mat_t<2,2,Type>& m)
-{
-    Type det = determinant(m);
-    if (det > Type(kEps) || det < -Type(kEps))
-    {
-        return mat_t<2,2,Type>( m.data[1][1], -m.data[0][1],
-                     -m.data[1][0],  m.data[0][0])*(Type(1.0f)/det);
-    }
-    else
-    {
-        return mat_t<2,2,Type>();
-    }
-}
-
-template<typename Type>
-inline CUDA_CALLABLE mat_t<3,3,Type> inverse(const mat_t<3,3,Type>& m)
-{
-	Type det = determinant(m);
-
-	if (det != Type(0.0f))
-	{
-		mat_t<3,3,Type> b;
-		
-		b.data[0][0] = m.data[1][1]*m.data[2][2] - m.data[1][2]*m.data[2][1]; 
-		b.data[1][0] = m.data[1][2]*m.data[2][0] - m.data[1][0]*m.data[2][2]; 
-		b.data[2][0] = m.data[1][0]*m.data[2][1] - m.data[1][1]*m.data[2][0]; 
-		
-        b.data[0][1] = m.data[0][2]*m.data[2][1] - m.data[0][1]*m.data[2][2]; 
-        b.data[1][1] = m.data[0][0]*m.data[2][2] - m.data[0][2]*m.data[2][0]; 
-        b.data[2][1] = m.data[0][1]*m.data[2][0] - m.data[0][0]*m.data[2][1]; 
-
-        b.data[0][2] = m.data[0][1]*m.data[1][2] - m.data[0][2]*m.data[1][1];
-        b.data[1][2] = m.data[0][2]*m.data[1][0] - m.data[0][0]*m.data[1][2];
-        b.data[2][2] = m.data[0][0]*m.data[1][1] - m.data[0][1]*m.data[1][0];
-
-		return b*(Type(1.0f)/det);
-	}
-	else
-	{
-		return mat_t<3,3,Type>();
-	}
-}
-
-template<typename Type>
-inline CUDA_CALLABLE mat_t<4,4,Type> inverse(const mat_t<4,4,Type>& m)
-{
-    // adapted from USD GfMatrix4f::Inverse()
-    Type x00, x01, x02, x03;
-    Type x10, x11, x12, x13;
-    Type x20, x21, x22, x23;
-    Type x30, x31, x32, x33;
-    double y01, y02, y03, y12, y13, y23;
-    Type z00, z10, z20, z30;
-    Type z01, z11, z21, z31;
-    double z02, z03, z12, z13, z22, z23, z32, z33;
-
-    // Pickle 1st two columns of matrix into registers
-    x00 = m.data[0][0];
-    x01 = m.data[0][1];
-    x10 = m.data[1][0];
-    x11 = m.data[1][1];
-    x20 = m.data[2][0];
-    x21 = m.data[2][1];
-    x30 = m.data[3][0];
-    x31 = m.data[3][1];
-
-    // Compute all six 2x2 determinants of 1st two columns
-    y01 = x00*x11 - x10*x01;
-    y02 = x00*x21 - x20*x01;
-    y03 = x00*x31 - x30*x01;
-    y12 = x10*x21 - x20*x11;
-    y13 = x10*x31 - x30*x11;
-    y23 = x20*x31 - x30*x21;
-
-    // Pickle 2nd two columns of matrix into registers
-    x02 = m.data[0][2];
-    x03 = m.data[0][3];
-    x12 = m.data[1][2];
-    x13 = m.data[1][3];
-    x22 = m.data[2][2];
-    x23 = m.data[2][3];
-    x32 = m.data[3][2];
-    x33 = m.data[3][3];
-
-    // Compute all 3x3 cofactors for 2nd two columns */
-    z33 = x02*y12 - x12*y02 + x22*y01;
-    z23 = x12*y03 - x32*y01 - x02*y13;
-    z13 = x02*y23 - x22*y03 + x32*y02;
-    z03 = x22*y13 - x32*y12 - x12*y23;
-    z32 = x13*y02 - x23*y01 - x03*y12;
-    z22 = x03*y13 - x13*y03 + x33*y01;
-    z12 = x23*y03 - x33*y02 - x03*y23;
-    z02 = x13*y23 - x23*y13 + x33*y12;
-
-    // Compute all six 2x2 determinants of 2nd two columns
-    y01 = x02*x13 - x12*x03;
-    y02 = x02*x23 - x22*x03;
-    y03 = x02*x33 - x32*x03;
-    y12 = x12*x23 - x22*x13;
-    y13 = x12*x33 - x32*x13;
-    y23 = x22*x33 - x32*x23;
-
-    // Compute all 3x3 cofactors for 1st two columns
-    z30 = x11*y02 - x21*y01 - x01*y12;
-    z20 = x01*y13 - x11*y03 + x31*y01;
-    z10 = x21*y03 - x31*y02 - x01*y23;
-    z00 = x11*y23 - x21*y13 + x31*y12;
-    z31 = x00*y12 - x10*y02 + x20*y01;
-    z21 = x10*y03 - x30*y01 - x00*y13;
-    z11 = x00*y23 - x20*y03 + x30*y02;
-    z01 = x20*y13 - x30*y12 - x10*y23;
-
-    // compute 4x4 determinant & its reciprocal
-    double det = x30*z30 + x20*z20 + x10*z10 + x00*z00;
-    
-    if(fabs(det) > kEps) 
-    {
-        mat_t<4,4,Type> invm;
-
-        double rcp = 1.0 / det;
-
-        // Multiply all 3x3 cofactors by reciprocal & transpose
-        invm.data[0][0] = Type(z00*rcp);
-        invm.data[0][1] = Type(z10*rcp);
-        invm.data[1][0] = Type(z01*rcp);
-        invm.data[0][2] = Type(z20*rcp);
-        invm.data[2][0] = Type(z02*rcp);
-        invm.data[0][3] = Type(z30*rcp);
-        invm.data[3][0] = Type(z03*rcp);
-        invm.data[1][1] = Type(z11*rcp);
-        invm.data[1][2] = Type(z21*rcp);
-        invm.data[2][1] = Type(z12*rcp);
-        invm.data[1][3] = Type(z31*rcp);
-        invm.data[3][1] = Type(z13*rcp);
-        invm.data[2][2] = Type(z22*rcp);
-        invm.data[2][3] = Type(z32*rcp);
-        invm.data[3][2] = Type(z23*rcp);
-        invm.data[3][3] = Type(z33*rcp);
-
-        return invm;
-    }
-    else 
-    {
-        return mat_t<4,4,Type>();
-    }
-}
-
-template<unsigned Rows,typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Rows,Type> diag(const vec_t<Rows,Type>& d)
-{
-    mat_t<Rows,Rows,Type> ret(Type(0));
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        ret.data[i][i] = d[i];
-    }
-    return ret;
-}
-
-template<unsigned Rows,unsigned Cols,typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> outer(const vec_t<Rows,Type>& a, const vec_t<Cols,Type>& b)
-{
-    // col 0 = a * b[0] etc...
-    mat_t<Rows,Cols,Type> ret;
-    for (unsigned row=0; row < Rows; ++row)
-    {
-        for (unsigned col=0; col < Cols; ++col) // columns
-        {
-            ret.data[row][col] = a[row] * b[col];
-        }
-    }
-    return ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE mat_t<3,3,Type> skew(const vec_t<3,Type>& a)
-{
-    mat_t<3,3,Type> out(
-        Type(0), -a[2],   a[1],
-        a[2],   Type(0), -a[0],
-        -a[1],   a[0],   Type(0)
-    );
-
-    return out;
-}
-
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> cw_mul(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
-{
-    mat_t<Rows,Cols,Type> t;
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            t.data[i][j] = a.data[i][j] * b.data[i][j];
-        }
-    }
-
-    return t;
-}
-
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE mat_t<Rows,Cols,Type> cw_div(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
-{
-    mat_t<Rows,Cols,Type> t;
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            t.data[i][j] = a.data[i][j] / b.data[i][j];
-        }
-    }
-
-    return t;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<3,Type> transform_point(const mat_t<4,4,Type>& m, const vec_t<3,Type>& v)
-{
-    vec_t<4,Type> out = mul(m, vec_t<4,Type>(v[0], v[1], v[2], Type(1)));
-    return vec_t<3,Type>(out[0], out[1], out[2]);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<3,Type> transform_vector(const mat_t<4,4,Type>& m, const vec_t<3,Type>& v)
-{
-    vec_t<4,Type> out = mul(m, vec_t<4,Type>(v[0], v[1], v[2], 0.f));
-    return vec_t<3,Type>(out[0], out[1], out[2]);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_extract(const mat_t<Rows,Cols,Type>& m, int row, mat_t<Rows,Cols,Type>& adj_m, int& adj_row, const vec_t<Cols,Type>& adj_ret)
-{
-    for( unsigned col=0; col < Cols; ++col )
-        adj_m.data[row][col] += adj_ret[col];
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline void CUDA_CALLABLE adj_extract(const mat_t<Rows,Cols,Type>& m, int row, int col, mat_t<Rows,Cols,Type>& adj_m, int& adj_row, int& adj_col, Type adj_ret)
-{
-#ifndef NDEBUG
-    if (row < 0 || row > Rows)
-    {
-        printf("mat row index %d out of bounds at %s %d\n", row, __FILE__, __LINE__);
-        assert(0);
-    }
-    if (col < 0 || col > Cols)
-    {
-        printf("mat col index %d out of bounds at %s %d\n", col, __FILE__, __LINE__);
-        assert(0);
-    }
-#endif
-    adj_m.data[row][col] += adj_ret;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_outer(const vec_t<Rows,Type>& a, const vec_t<Cols,Type>& b, vec_t<Rows,Type>& adj_a, vec_t<Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-  adj_a += mul(adj_ret, b);
-  adj_b += mul(transpose(adj_ret), a);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_add(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            adj_a.data[i][j] += adj_ret.data[i][j];
-            adj_b.data[i][j] += adj_ret.data[i][j];
-        }
-    }
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_sub(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            adj_a.data[i][j] += adj_ret.data[i][j];
-            adj_b.data[i][j] -= adj_ret.data[i][j];
-        }
-    }
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_div(const mat_t<Rows,Cols,Type>& a, Type s, mat_t<Rows,Cols,Type>& adj_a, Type& adj_s, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-    adj_s -= tensordot(a , adj_ret)/ (s * s); // - a / s^2
-
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            adj_a.data[i][j] += adj_ret.data[i][j] / s;
-        }
-    }
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_div(Type s, const mat_t<Rows,Cols,Type>& a, Type& adj_s, mat_t<Rows,Cols,Type>& adj_a, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-    adj_s -= tensordot(a , adj_ret)/ (s * s); // - a / s^2
-
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            adj_a.data[i][j] += s / adj_ret.data[i][j];
-        }
-    }
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_mul(const mat_t<Rows,Cols,Type>& a, Type b, mat_t<Rows,Cols,Type>& adj_a, Type& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            adj_a.data[i][j] += b*adj_ret.data[i][j];
-            adj_b += a.data[i][j]*adj_ret.data[i][j];
-        }
-    }
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_mul(Type b, const mat_t<Rows,Cols,Type>& a, Type& adj_b, mat_t<Rows,Cols,Type>& adj_a, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-    adj_mul(a, b, adj_a, adj_b, adj_ret);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_ddot(mat_t<Rows,Cols,Type> a, mat_t<Rows,Cols,Type> b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const Type adj_ret)
-{
-    adj_a += b*adj_ret;
-    adj_b += a*adj_ret;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_mul(const mat_t<Rows,Cols,Type>& a, const vec_t<Cols,Type>& b, mat_t<Rows,Cols,Type>& adj_a, vec_t<Cols,Type>& adj_b, const vec_t<Rows,Type>& adj_ret)
-{
-    adj_a += outer(adj_ret, b);
-    adj_b += mul(transpose(a), adj_ret);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_mul(const vec_t<Rows,Type>& b, const mat_t<Rows,Cols,Type>& a, vec_t<Rows,Type>& adj_b, mat_t<Rows,Cols,Type>& adj_a, const vec_t<Cols,Type>& adj_ret)
-{
-    adj_a += outer(b, adj_ret);
-    adj_b += mul(adj_ret, transpose(a));
-}
-
-template<unsigned Rows, unsigned Cols, unsigned ColsOut, typename Type>
-inline CUDA_CALLABLE void adj_mul(const mat_t<Rows,Cols,Type>& a, const mat_t<Cols,ColsOut,Type>& b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Cols,ColsOut,Type>& adj_b, const mat_t<Rows,ColsOut,Type>& adj_ret)
-{
-    adj_a += mul(adj_ret, transpose(b));
-    adj_b += mul(transpose(a), adj_ret);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_transpose(const mat_t<Rows,Cols,Type>& a, mat_t<Rows,Cols,Type>& adj_a, const mat_t<Cols,Rows,Type>& adj_ret)
-{
-    adj_a += transpose(adj_ret);
-}
-
-template<unsigned Rows, typename Type>
-inline CUDA_CALLABLE void adj_trace(const mat_t<Rows,Rows,Type>& m, mat_t<Rows,Rows,Type>& adj_m, Type adj_ret)
-{
-    for (unsigned i=0; i < Rows; ++i)
-        adj_m.data[i][i] += adj_ret;
-}
-
-template<unsigned Rows, typename Type>
-inline CUDA_CALLABLE void adj_diag(const vec_t<Rows,Type>& d, vec_t<Rows,Type>& adj_d, const mat_t<Rows,Rows,Type>& adj_ret)
-{
-    for (unsigned i=0; i < Rows; ++i)
-        adj_d[i] += adj_ret.data[i][i];
-}
-
-template<unsigned Rows, typename Type>
-inline CUDA_CALLABLE void adj_get_diag(const mat_t<Rows,Rows,Type>& m, mat_t<Rows,Rows,Type>& adj_m, const vec_t<Rows,Type>& adj_ret)
-{
-    for (unsigned i=0; i < Rows; ++i)
-        adj_m.data[i][i] += adj_ret[i];
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_determinant(const mat_t<2,2,Type>& m, mat_t<2,2,Type>& adj_m, Type adj_ret)
-{
-    adj_m.data[0][0] += m.data[1][1]*adj_ret;
-    adj_m.data[1][1] += m.data[0][0]*adj_ret;
-    adj_m.data[0][1] -= m.data[1][0]*adj_ret;
-    adj_m.data[1][0] -= m.data[0][1]*adj_ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_determinant(const mat_t<3,3,Type>& m, mat_t<3,3,Type>& adj_m, Type adj_ret)
-{
-    (vec_t<3,Type>&)adj_m.data[0] += cross(m.get_row(1), m.get_row(2))*adj_ret;
-    (vec_t<3,Type>&)adj_m.data[1] += cross(m.get_row(2), m.get_row(0))*adj_ret;
-    (vec_t<3,Type>&)adj_m.data[2] += cross(m.get_row(0), m.get_row(1))*adj_ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_determinant(const mat_t<4,4,Type>& m, mat_t<4,4,Type>& adj_m, Type adj_ret)
-{
-    // adapted from USD GfMatrix4f::Inverse()
-    Type x00, x01, x02, x03;
-    Type x10, x11, x12, x13;
-    Type x20, x21, x22, x23;
-    Type x30, x31, x32, x33;
-    double y01, y02, y03, y12, y13, y23;
-    Type z00, z10, z20, z30;
-    Type z01, z11, z21, z31;
-    double z02, z03, z12, z13, z22, z23, z32, z33;
-
-    // Pickle 1st two columns of matrix into registers
-    x00 = m.data[0][0];
-    x01 = m.data[0][1];
-    x10 = m.data[1][0];
-    x11 = m.data[1][1];
-    x20 = m.data[2][0];
-    x21 = m.data[2][1];
-    x30 = m.data[3][0];
-    x31 = m.data[3][1];
-
-    // Compute all six 2x2 determinants of 1st two columns
-    y01 = x00*x11 - x10*x01;
-    y02 = x00*x21 - x20*x01;
-    y03 = x00*x31 - x30*x01;
-    y12 = x10*x21 - x20*x11;
-    y13 = x10*x31 - x30*x11;
-    y23 = x20*x31 - x30*x21;
-
-    // Pickle 2nd two columns of matrix into registers
-    x02 = m.data[0][2];
-    x03 = m.data[0][3];
-    x12 = m.data[1][2];
-    x13 = m.data[1][3];
-    x22 = m.data[2][2];
-    x23 = m.data[2][3];
-    x32 = m.data[3][2];
-    x33 = m.data[3][3];
-
-    // Compute all 3x3 cofactors for 2nd two columns */
-    z33 = x02*y12 - x12*y02 + x22*y01;
-    z23 = x12*y03 - x32*y01 - x02*y13;
-    z13 = x02*y23 - x22*y03 + x32*y02;
-    z03 = x22*y13 - x32*y12 - x12*y23;
-    z32 = x13*y02 - x23*y01 - x03*y12;
-    z22 = x03*y13 - x13*y03 + x33*y01;
-    z12 = x23*y03 - x33*y02 - x03*y23;
-    z02 = x13*y23 - x23*y13 + x33*y12;
-
-    // Compute all six 2x2 determinants of 2nd two columns
-    y01 = x02*x13 - x12*x03;
-    y02 = x02*x23 - x22*x03;
-    y03 = x02*x33 - x32*x03;
-    y12 = x12*x23 - x22*x13;
-    y13 = x12*x33 - x32*x13;
-    y23 = x22*x33 - x32*x23;
-
-    // Compute all 3x3 cofactors for 1st two columns
-    z30 = x11*y02 - x21*y01 - x01*y12;
-    z20 = x01*y13 - x11*y03 + x31*y01;
-    z10 = x21*y03 - x31*y02 - x01*y23;
-    z00 = x11*y23 - x21*y13 + x31*y12;
-    z31 = x00*y12 - x10*y02 + x20*y01;
-    z21 = x10*y03 - x30*y01 - x00*y13;
-    z11 = x00*y23 - x20*y03 + x30*y02;
-    z01 = x20*y13 - x30*y12 - x10*y23;
-
-    // Multiply all 3x3 cofactors by adjoint & transpose
-    adj_m.data[0][0] += Type(z00*adj_ret);
-    adj_m.data[1][0] += Type(z10*adj_ret);
-    adj_m.data[0][1] += Type(z01*adj_ret);
-    adj_m.data[2][0] += Type(z20*adj_ret);
-    adj_m.data[0][2] += Type(z02*adj_ret);
-    adj_m.data[3][0] += Type(z30*adj_ret);
-    adj_m.data[0][3] += Type(z03*adj_ret);
-    adj_m.data[1][1] += Type(z11*adj_ret);
-    adj_m.data[2][1] += Type(z21*adj_ret);
-    adj_m.data[1][2] += Type(z12*adj_ret);
-    adj_m.data[3][1] += Type(z31*adj_ret);
-    adj_m.data[1][3] += Type(z13*adj_ret);
-    adj_m.data[2][2] += Type(z22*adj_ret);
-    adj_m.data[3][2] += Type(z32*adj_ret);
-    adj_m.data[2][3] += Type(z23*adj_ret);
-    adj_m.data[3][3] += Type(z33*adj_ret);
-}
-
-template<unsigned Rows, typename Type>
-inline CUDA_CALLABLE void adj_inverse(const mat_t<Rows,Rows,Type>& m, mat_t<Rows,Rows,Type>& ret, mat_t<Rows,Rows,Type>& adj_m, const mat_t<Rows,Rows,Type>& adj_ret)
-{
-    // todo: how to cache this from the forward pass?
-    mat_t<Rows,Rows,Type> invt = transpose(ret);
-
-    // see https://people.maths.ox.ac.uk/gilesm/files/NA-08-01.pdf 2.2.3
-    adj_m -= mul(mul(invt, adj_ret), invt);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_transform_point(const mat_t<4,4,Type>& m, const vec_t<3,Type>& v, mat_t<4,4,Type>& adj_m, vec_t<3,Type>& adj_v, const vec_t<3,Type>& adj_ret)
-{
-    vec_t<4,Type> out = vec_t<4,Type>(v[0], v[1], v[2], 1.f);
-    adj_m = add(adj_m, transpose(mat_t<4,4,Type>(adj_ret[0] * out, adj_ret[1] * out, adj_ret[2] * out, vec_t<4,Type>())));
-    adj_v[0] += dot(vec_t<3,Type>(m.data[0][0], m.data[1][0], m.data[2][0]), adj_ret);
-    adj_v[1] += dot(vec_t<3,Type>(m.data[0][1], m.data[1][1], m.data[2][1]), adj_ret);
-    adj_v[2] += dot(vec_t<3,Type>(m.data[0][2], m.data[1][2], m.data[2][2]), adj_ret);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_transform_vector(const mat_t<4,4,Type>& m, const vec_t<3,Type>& v, mat_t<4,4,Type>& adj_m, vec_t<3,Type>& adj_v, const vec_t<3,Type>& adj_ret)
-{
-    vec_t<4,Type> out = vec_t<4,Type>(v[0], v[1], v[2], 0.f);
-    adj_m = add(adj_m, transpose(mat_t<4,4,Type>(adj_ret[0] * out, adj_ret[1] * out, adj_ret[2] * out, vec_t<4,Type>())));
-    adj_v[0] += dot(vec_t<3,Type>(m.data[0][0], m.data[1][0], m.data[2][0]), adj_ret);
-    adj_v[1] += dot(vec_t<3,Type>(m.data[0][1], m.data[1][1], m.data[2][1]), adj_ret);
-    adj_v[2] += dot(vec_t<3,Type>(m.data[0][2], m.data[1][2], m.data[2][2]), adj_ret);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_skew(const vec_t<3,Type>& a, vec_t<3,Type>& adj_a, const mat_t<3,3,Type>& adj_ret)
-{
-    adj_a[0] += adj_ret.data[2][1] - adj_ret.data[1][2];
-    adj_a[1] += adj_ret.data[0][2] - adj_ret.data[2][0];
-    adj_a[2] += adj_ret.data[1][0] - adj_ret.data[0][1];
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_cw_mul(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-  adj_a += cw_mul(b, adj_ret);
-  adj_b += cw_mul(a, adj_ret);
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_cw_div(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b, mat_t<Rows,Cols,Type>& ret, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
-{
-  adj_a += cw_div(adj_ret, b);
-  adj_b -= cw_mul(adj_ret, cw_div(ret, b));
-}
-
-// adjoint for the constant constructor:
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_mat_t(Type s, Type& adj_s, const mat_t<Rows, Cols, Type>& adj_ret)
-{
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            adj_s += adj_ret.data[i][j];
-        }
-    }
-}
-
-// adjoint for the casting constructor:
-template<unsigned Rows, unsigned Cols, typename Type, typename OtherType>
-inline CUDA_CALLABLE void adj_mat_t(const mat_t<Rows, Cols, OtherType>& other, mat_t<Rows, Cols, OtherType>& adj_other, const mat_t<Rows, Cols, Type>& adj_ret)
-{
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            adj_other.data[i][j] += adj_ret.data[i][j];
-        }
-    }
-}
-
-// adjoint for the initializer_array scalar constructor:
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_mat_t(const initializer_array<Rows * Cols, Type> &cmps, const initializer_array<Rows * Cols, Type*> &adj_cmps, const mat_t<Rows, Cols, Type>& adj_ret)
-{
-    for (unsigned i=0; i < Rows; ++i)
-    {
-        for (unsigned j=0; j < Cols; ++j)
-        {
-            *adj_cmps[i * Cols + j] += adj_ret.data[i][j];
-        }
-    }
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_mat_t(Type m00, Type m01, Type m10, Type m11, Type& adj_m00, Type& adj_m01, Type& adj_m10, Type& adj_m11, const mat_t<2, 2, Type>& adj_ret)
-{
-    adj_m00 += adj_ret.data[0][0];
-    adj_m01 += adj_ret.data[0][1];
-    adj_m10 += adj_ret.data[1][0];
-    adj_m11 += adj_ret.data[1][1];
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_mat_t(Type m00, Type m01, Type m02,
-                      Type m10, Type m11, Type m12,
-                      Type m20, Type m21, Type m22,
-                      Type& a00, Type& a01, Type& a02,
-                      Type& a10, Type& a11, Type& a12,
-                      Type& a20, Type& a21, Type& a22,
-                      const mat_t<3, 3, Type>& adj_ret)
-{
-    a00 += adj_ret.data[0][0];
-    a01 += adj_ret.data[0][1];
-    a02 += adj_ret.data[0][2];
-    a10 += adj_ret.data[1][0];
-    a11 += adj_ret.data[1][1];
-    a12 += adj_ret.data[1][2];
-    a20 += adj_ret.data[2][0];
-    a21 += adj_ret.data[2][1];
-    a22 += adj_ret.data[2][2];
-}
-
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_mat_t(Type m00, Type m01, Type m02, Type m03,
-                      Type m10, Type m11, Type m12, Type m13,
-                      Type m20, Type m21, Type m22, Type m23,
-                      Type m30, Type m31, Type m32, Type m33,
-                      Type& a00, Type& a01, Type& a02, Type& a03,
-                      Type& a10, Type& a11, Type& a12, Type& a13,
-                      Type& a20, Type& a21, Type& a22, Type& a23,
-                      Type& a30, Type& a31, Type& a32, Type& a33,
-                      const mat_t<4, 4, Type>& adj_ret)
-{
-    a00 += adj_ret.data[0][0];
-    a01 += adj_ret.data[0][1];
-    a02 += adj_ret.data[0][2];
-    a03 += adj_ret.data[0][3];
-
-    a10 += adj_ret.data[1][0];
-    a11 += adj_ret.data[1][1];
-    a12 += adj_ret.data[1][2];
-    a13 += adj_ret.data[1][3];
-
-    a20 += adj_ret.data[2][0];
-    a21 += adj_ret.data[2][1];
-    a22 += adj_ret.data[2][2];
-    a23 += adj_ret.data[2][3];
-
-    a30 += adj_ret.data[3][0];
-    a31 += adj_ret.data[3][1];
-    a32 += adj_ret.data[3][2];
-    a33 += adj_ret.data[3][3];
-}
-
-
-
-// adjoint for the initializer_array vector constructor:
-template<unsigned Rows, unsigned Cols, typename Type>
-inline CUDA_CALLABLE void adj_mat_t(const initializer_array<Cols, vec_t<Rows,Type> > &cmps, const initializer_array<Cols, vec_t<Rows,Type>* > &adj_cmps, const mat_t<Rows, Cols, Type>& adj_ret)
-{
-    for (unsigned j=0; j < Cols; ++j)
-    {
-        for (unsigned i=0; i < Rows; ++i)
-        {
-            (*adj_cmps[j])[i] += adj_ret.data[i][j];
-        }
-    }
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_mat_t(const vec_t<2,Type> &cmps0, const vec_t<2,Type> &cmps1, vec_t<2,Type> &adj_cmps0, vec_t<2,Type> &adj_cmps1, const mat_t<2, 2, Type>& adj_ret)
-{
-    for (unsigned i=0; i < 2; ++i)
-    {
-        adj_cmps0[i] += adj_ret.data[i][0];
-        adj_cmps1[i] += adj_ret.data[i][1];
-    }
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_mat_t(const vec_t<3,Type> &cmps0, const vec_t<3,Type> &cmps1, const vec_t<3,Type> &cmps2, vec_t<3,Type> &adj_cmps0, vec_t<3,Type> &adj_cmps1, vec_t<3,Type> &adj_cmps2, const mat_t<3, 3, Type>& adj_ret)
-{
-    for (unsigned i=0; i < 3; ++i)
-    {
-        adj_cmps0[i] += adj_ret.data[i][0];
-        adj_cmps1[i] += adj_ret.data[i][1];
-        adj_cmps2[i] += adj_ret.data[i][2];
-    }
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_mat_t(const vec_t<4,Type> &cmps0, const vec_t<4,Type> &cmps1, const vec_t<4,Type> &cmps2, const vec_t<4,Type> &cmps3, vec_t<4,Type> &adj_cmps0, vec_t<4,Type> &adj_cmps1, vec_t<4,Type> &adj_cmps2, vec_t<4,Type> &adj_cmps3, const mat_t<4, 4, Type>& adj_ret)
-{
-    for (unsigned i=0; i < 4; ++i)
-    {
-        adj_cmps0[i] += adj_ret.data[i][0];
-        adj_cmps1[i] += adj_ret.data[i][1];
-        adj_cmps2[i] += adj_ret.data[i][2];
-        adj_cmps3[i] += adj_ret.data[i][3];
-    }
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-CUDA_CALLABLE inline mat_t<Rows, Cols, Type> lerp(const mat_t<Rows, Cols, Type>& a, const mat_t<Rows, Cols, Type>& b, Type t)
-{
-    return a*(Type(1)-t) + b*t;
-}
-
-template<unsigned Rows, unsigned Cols, typename Type>
-CUDA_CALLABLE inline void adj_lerp(const mat_t<Rows, Cols, Type>& a, const mat_t<Rows, Cols, Type>& b, Type t, mat_t<Rows, Cols, Type>& adj_a, mat_t<Rows, Cols, Type>& adj_b, Type& adj_t, const mat_t<Rows, Cols, 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);
-}
-
-// for integral types we do not accumulate gradients
-template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, int8>* buf, const mat_t<Rows, Cols, int8> &value) { }
-template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, uint8>* buf, const mat_t<Rows, Cols, uint8> &value) { }
-template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, int16>* buf, const mat_t<Rows, Cols, int16> &value) { }
-template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, uint16>* buf, const mat_t<Rows, Cols, uint16> &value) { }
-template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, int32>* buf, const mat_t<Rows, Cols, int32> &value) { }
-template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, uint32>* buf, const mat_t<Rows, Cols, uint32> &value) { }
-template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, int64>* buf, const mat_t<Rows, Cols, int64> &value) { }
-template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, uint64>* buf, const mat_t<Rows, Cols, uint64> &value) { }
-
-using mat22h = mat_t<2,2,half>;
-using mat33h = mat_t<3,3,half>;
-using mat44h = mat_t<4,4,half>;
-
-using mat22 = mat_t<2,2,float>;
-using mat33 = mat_t<3,3,float>;
-using mat44 = mat_t<4,4,float>;
-
-using mat22f = mat_t<2,2,float>;
-using mat33f = mat_t<3,3,float>;
-using mat44f = mat_t<4,4,float>;
-
-using mat22d = mat_t<2,2,double>;
-using mat33d = mat_t<3,3,double>;
-using mat44d = mat_t<4,4,double>;
-
-inline CUDA_CALLABLE void adj_mat22(vec2 c0, vec2 c1,
-                      vec2& a0, vec2& a1,
-                      const mat22& adj_ret)
-{
-    a0 += adj_ret.get_col(0);
-    a1 += adj_ret.get_col(1);
-}
-
-inline CUDA_CALLABLE void adj_mat22(float m00, float m01, float m10, float m11, float& adj_m00, float& adj_m01, float& adj_m10, float& adj_m11, const mat22& adj_ret)
-{
-    adj_m00 += adj_ret.data[0][0];
-    adj_m01 += adj_ret.data[0][1];
-    adj_m10 += adj_ret.data[1][0];
-    adj_m11 += adj_ret.data[1][1];
-}
-
-inline CUDA_CALLABLE void adj_mat33(vec3 c0, vec3 c1, vec3 c2,
-                      vec3& a0, vec3& a1, vec3& a2,
-                      const mat33& adj_ret)
-{
-    // column constructor
-    a0 += adj_ret.get_col(0);
-    a1 += adj_ret.get_col(1);
-    a2 += adj_ret.get_col(2);
-
-}
-
-inline CUDA_CALLABLE void adj_mat33(float m00, float m01, float m02,
-                      float m10, float m11, float m12,
-                      float m20, float m21, float m22,
-                      float& a00, float& a01, float& a02,
-                      float& a10, float& a11, float& a12,
-                      float& a20, float& a21, float& a22,
-                      const mat33& adj_ret)
-{
-    a00 += adj_ret.data[0][0];
-    a01 += adj_ret.data[0][1];
-    a02 += adj_ret.data[0][2];
-    a10 += adj_ret.data[1][0];
-    a11 += adj_ret.data[1][1];
-    a12 += adj_ret.data[1][2];
-    a20 += adj_ret.data[2][0];
-    a21 += adj_ret.data[2][1];
-    a22 += adj_ret.data[2][2];
-}
-
-inline CUDA_CALLABLE void adj_mat44(
-    vec4 c0, vec4 c1, vec4 c2, vec4 c3,
-    vec4& a0, vec4& a1, vec4& a2, vec4& a3,
-    const mat44& adj_ret)
-{
-    // column constructor
-    a0 += adj_ret.get_col(0);
-    a1 += adj_ret.get_col(1);
-    a2 += adj_ret.get_col(2);
-    a3 += adj_ret.get_col(3);
-}
-
-inline CUDA_CALLABLE void adj_mat44(float m00, float m01, float m02, float m03,
-                      float m10, float m11, float m12, float m13,
-                      float m20, float m21, float m22, float m23,
-                      float m30, float m31, float m32, float m33,
-                      float& a00, float& a01, float& a02, float& a03,
-                      float& a10, float& a11, float& a12, float& a13,
-                      float& a20, float& a21, float& a22, float& a23,
-                      float& a30, float& a31, float& a32, float& a33,
-                      const mat44& adj_ret)
-{
-    a00 += adj_ret.data[0][0];
-    a01 += adj_ret.data[0][1];
-    a02 += adj_ret.data[0][2];
-    a03 += adj_ret.data[0][3];
-
-    a10 += adj_ret.data[1][0];
-    a11 += adj_ret.data[1][1];
-    a12 += adj_ret.data[1][2];
-    a13 += adj_ret.data[1][3];
-
-    a20 += adj_ret.data[2][0];
-    a21 += adj_ret.data[2][1];
-    a22 += adj_ret.data[2][2];
-    a23 += adj_ret.data[2][3];
-
-    a30 += adj_ret.data[3][0];
-    a31 += adj_ret.data[3][1];
-    a32 += adj_ret.data[3][2];
-    a33 += adj_ret.data[3][3];
-}
-
-} // namespace wp
+/** 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
+
+#include "initializer_array.h"
+
+namespace wp
+{
+
+//----------------------------------------------------------
+// mat
+template<typename T>
+struct quat_t;
+
+template<unsigned Rows, unsigned Cols, typename Type>
+struct mat_t
+{
+    inline CUDA_CALLABLE mat_t()
+        : data()
+    {}
+
+    inline CUDA_CALLABLE mat_t(Type s)
+    {
+        for (unsigned i=0; i < Rows; ++i)
+            for (unsigned j=0; j < Cols; ++j)
+                data[i][j] = s;
+    }
+    
+    template <typename OtherType>
+    inline explicit CUDA_CALLABLE mat_t(const mat_t<Rows, Cols, OtherType>& other)
+    {
+        for (unsigned i=0; i < Rows; ++i)
+            for (unsigned j=0; j < Cols; ++j)
+                data[i][j] = other.data[i][j];
+    }
+    
+    inline CUDA_CALLABLE mat_t(vec_t<2,Type> c0, vec_t<2,Type> c1)
+    {
+        data[0][0] = c0[0];
+        data[1][0] = c0[1];
+
+        data[0][1] = c1[0];
+        data[1][1] = c1[1];
+    }
+    
+    inline CUDA_CALLABLE mat_t(vec_t<3,Type> c0, vec_t<3,Type> c1, vec_t<3,Type> c2)
+    {
+        data[0][0] = c0[0];
+        data[1][0] = c0[1];
+        data[2][0] = c0[2];
+
+        data[0][1] = c1[0];
+        data[1][1] = c1[1];
+        data[2][1] = c1[2];
+
+        data[0][2] = c2[0];
+        data[1][2] = c2[1];
+        data[2][2] = c2[2];
+    }
+
+    inline CUDA_CALLABLE mat_t(vec_t<4,Type> c0, vec_t<4,Type> c1, vec_t<4,Type> c2, vec_t<4,Type> c3)
+    {
+        data[0][0] = c0[0];
+        data[1][0] = c0[1];
+        data[2][0] = c0[2];
+        data[3][0] = c0[3];
+
+        data[0][1] = c1[0];
+        data[1][1] = c1[1];
+        data[2][1] = c1[2];
+        data[3][1] = c1[3];
+
+        data[0][2] = c2[0];
+        data[1][2] = c2[1];
+        data[2][2] = c2[2];
+        data[3][2] = c2[3];
+
+        data[0][3] = c3[0];
+        data[1][3] = c3[1];
+        data[2][3] = c3[2];
+        data[3][3] = c3[3];
+    }
+
+    inline CUDA_CALLABLE mat_t(Type m00, Type m01, Type m10, Type m11) 
+    {
+        data[0][0] = m00;
+        data[1][0] = m10;
+        data[0][1] = m01;
+        data[1][1] = m11;
+    }
+    
+    inline CUDA_CALLABLE mat_t(
+        Type m00, Type m01, Type m02,
+        Type m10, Type m11, Type m12,
+        Type m20, Type m21, Type m22)
+    {
+        data[0][0] = m00;
+        data[1][0] = m10;
+        data[2][0] = m20;
+
+        data[0][1] = m01;
+        data[1][1] = m11;
+        data[2][1] = m21;
+
+        data[0][2] = m02;
+        data[1][2] = m12;
+        data[2][2] = m22;
+    }
+
+    inline CUDA_CALLABLE mat_t(
+                 Type m00, Type m01, Type m02, Type m03,
+                 Type m10, Type m11, Type m12, Type m13,
+                 Type m20, Type m21, Type m22, Type m23,
+                 Type m30, Type m31, Type m32, Type m33)
+    {
+        data[0][0] = m00;
+        data[1][0] = m10;
+        data[2][0] = m20;
+        data[3][0] = m30;
+
+        data[0][1] = m01;
+        data[1][1] = m11;
+        data[2][1] = m21;
+        data[3][1] = m31;
+
+        data[0][2] = m02;
+        data[1][2] = m12;
+        data[2][2] = m22;
+        data[3][2] = m32;
+
+        data[0][3] = m03;
+        data[1][3] = m13;
+        data[2][3] = m23;
+        data[3][3] = m33;
+    }
+
+    // implemented in quat.h
+    inline CUDA_CALLABLE mat_t(const vec_t<3,Type>& pos, const quat_t<Type>& rot, const vec_t<3,Type>& scale);
+
+
+    inline CUDA_CALLABLE mat_t(const initializer_array<Rows * Cols, Type> &l)
+    {
+        for (unsigned i=0; i < Rows; ++i)
+        {
+            for (unsigned j=0; j < Cols; ++j)
+            {
+                data[i][j] = l[i * Cols + j];
+            }
+        }
+    }
+
+    inline CUDA_CALLABLE mat_t(const initializer_array<Cols, vec_t<Rows,Type> > &l)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            for (unsigned i=0; i < Rows; ++i)
+            {
+                data[i][j] = l[j][i];
+            }
+        }
+    }
+
+    CUDA_CALLABLE vec_t<Cols,Type> get_row(int index) const
+    {
+        return (vec_t<Cols,Type>&)data[index]; 
+    }
+
+    CUDA_CALLABLE void set_row(int index, const vec_t<Cols,Type>& v)
+    {
+        (vec_t<Cols,Type>&)data[index] = v;
+    }
+
+    CUDA_CALLABLE vec_t<Rows,Type> get_col(int index) const
+    {
+        vec_t<Rows,Type> ret;
+        for( unsigned i=0;i < Rows; ++i )
+        {
+            ret[i] = data[i][index];
+        }
+        return ret;
+    }
+
+    CUDA_CALLABLE void set_col(int index, const vec_t<Rows,Type>& v)
+    {
+        for( unsigned i=0;i < Rows; ++i )
+        {
+            data[i][index] = v[i];
+        }
+    }
+
+    // row major storage assumed to be compatible with PyTorch
+    Type data[Rows][Cols];
+};
+
+
+template<unsigned Rows, typename Type>
+inline CUDA_CALLABLE mat_t<Rows, Rows, Type> identity()
+{
+    mat_t<Rows, Rows, Type> m;
+    for( unsigned i=0; i < Rows; ++i )
+    {
+        m.data[i][i] = Type(1);
+    }
+    return m;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE bool operator==(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
+{
+    for (unsigned i=0; i < Rows; ++i)
+        for (unsigned j=0; j < Cols; ++j)
+            if (a.data[i][j] != b.data[i][j])
+                return false;
+
+    return true;
+}
+
+
+// negation:
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> operator - (mat_t<Rows,Cols,Type> a)
+{
+    // NB: this constructor will initialize all ret's components to 0, which is
+    // unnecessary... 
+    mat_t<Rows,Cols,Type> ret;
+    for (unsigned i=0; i < Rows; ++i)
+        for (unsigned j=0; j < Cols; ++j)
+            ret.data[i][j] = -a.data[i][j];
+
+    // Wonder if this does a load of copying when it returns... hopefully not as it's inlined?
+    return ret;
+}
+
+
+template<unsigned Rows, unsigned Cols, typename Type>
+CUDA_CALLABLE inline mat_t<Rows,Cols,Type> pos(const mat_t<Rows,Cols,Type>& x)
+{
+    return x;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+CUDA_CALLABLE inline void adj_pos(const mat_t<Rows,Cols,Type>& x, mat_t<Rows,Cols,Type>& adj_x, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+    adj_x += adj_ret;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+CUDA_CALLABLE inline mat_t<Rows,Cols,Type> neg(const mat_t<Rows,Cols,Type>& x)
+{
+    return -x;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+CUDA_CALLABLE inline void adj_neg(const mat_t<Rows,Cols,Type>& x, mat_t<Rows,Cols,Type>& adj_x, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+    adj_x -= adj_ret;
+}
+
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> atomic_add(mat_t<Rows,Cols,Type> * addr, mat_t<Rows,Cols,Type> value) 
+{
+    mat_t<Rows,Cols,Type> m;
+    
+    for (unsigned i=0; i < Rows; ++i)
+        for (unsigned j=0; j < Cols; ++j)
+            m.data[i][j] = atomic_add(&addr->data[i][j], value.data[i][j]);
+
+    return m;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> atomic_min(mat_t<Rows,Cols,Type> * addr, mat_t<Rows,Cols,Type> value) 
+{
+    mat_t<Rows,Cols,Type> m;
+    
+    for (unsigned i=0; i < Rows; ++i)
+        for (unsigned j=0; j < Cols; ++j)
+            m.data[i][j] = atomic_min(&addr->data[i][j], value.data[i][j]);
+
+    return m;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> atomic_max(mat_t<Rows,Cols,Type> * addr, mat_t<Rows,Cols,Type> value) 
+{
+    mat_t<Rows,Cols,Type> m;
+    
+    for (unsigned i=0; i < Rows; ++i)
+        for (unsigned j=0; j < Cols; ++j)
+            m.data[i][j] = atomic_max(&addr->data[i][j], value.data[i][j]);
+
+    return m;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_atomic_minmax(
+    mat_t<Rows,Cols,Type> *addr,
+    mat_t<Rows,Cols,Type> *adj_addr,
+    const mat_t<Rows,Cols,Type> &value,
+    mat_t<Rows,Cols,Type> &adj_value)
+{
+    for (unsigned i=0; i < Rows; ++i)
+        for (unsigned j=0; j < Cols; ++j)
+            adj_atomic_minmax(&addr->data[i][j], &adj_addr->data[i][j], value.data[i][j], adj_value.data[i][j]);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE vec_t<Cols,Type> extract(const mat_t<Rows,Cols,Type>& m, int row)
+{
+    vec_t<Cols,Type> ret;
+    for(unsigned i=0; i < Cols; ++i)
+    {
+        ret.c[i] = m.data[row][i];
+    }
+    return ret;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE Type extract(const mat_t<Rows,Cols,Type>& m, int row, int col)
+{
+#ifndef NDEBUG
+    if (row < 0 || row >= Rows)
+    {
+        printf("mat row index %d out of bounds at %s %d\n", row, __FILE__, __LINE__);
+        assert(0);
+    }
+    if (col < 0 || col >= Cols)
+    {
+        printf("mat col index %d out of bounds at %s %d\n", col, __FILE__, __LINE__);
+        assert(0);
+    }
+#endif
+    return m.data[row][col];
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE vec_t<Cols, Type>* index(mat_t<Rows,Cols,Type>& m, int row)
+{
+#ifndef NDEBUG
+    if (row < 0 || row >= Rows)
+    {
+        printf("mat row index %d out of bounds at %s %d\n", row, __FILE__, __LINE__);
+        assert(0);
+    }
+#endif
+
+    return reinterpret_cast<vec_t<Cols, Type>*>(&m.data[row]);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE Type* index(mat_t<Rows,Cols,Type>& m, int row, int col)
+{
+#ifndef NDEBUG
+    if (row < 0 || row >= Rows)
+    {
+        printf("mat row index %d out of bounds at %s %d\n", row, __FILE__, __LINE__);
+        assert(0);
+    }
+    if (col < 0 || col >= Cols)
+    {
+        printf("mat col index %d out of bounds at %s %d\n", col, __FILE__, __LINE__);
+        assert(0);
+    }
+#endif
+
+    return &m.data[row][col];
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_index(const mat_t<Rows,Cols,Type>& m, int row,
+                                       const mat_t<Rows,Cols,Type>& adj_m, int adj_row, const vec_t<Cols, Type>& adj_value)
+{
+    // nop
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_index(const mat_t<Rows,Cols,Type>& m, int row, int col,
+                                       const mat_t<Rows,Cols,Type>& adj_m, int adj_row, int adj_col, Type adj_value)
+{
+    // nop
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline bool CUDA_CALLABLE isfinite(const mat_t<Rows,Cols,Type>& m)
+{
+    for (unsigned i=0; i < Rows; ++i)
+        for (unsigned j=0; j < Cols; ++j)
+            if (!isfinite(m.data[i][j]))
+                return false;
+    return true;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline void CUDA_CALLABLE adj_isfinite(const mat_t<Rows,Cols,Type>& m, mat_t<Rows,Cols,Type>& adj_m, const bool &adj_ret)
+{
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline bool CUDA_CALLABLE isnan(const mat_t<Rows,Cols,Type>& m)
+{
+    for (unsigned i=0; i < Rows; ++i)
+        for (unsigned j=0; j < Cols; ++j)
+            if (isnan(m.data[i][j]))
+                return true;
+    return false;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline void CUDA_CALLABLE adj_isnan(const mat_t<Rows,Cols,Type>& m, mat_t<Rows,Cols,Type>& adj_m, const bool &adj_ret)
+{
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline bool CUDA_CALLABLE isinf(const mat_t<Rows,Cols,Type>& m)
+{
+    for (unsigned i=0; i < Rows; ++i)
+        for (unsigned j=0; j < Cols; ++j)
+            if (isinf(m.data[i][j]))
+                return true;
+    return false;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline void CUDA_CALLABLE adj_isinf(const mat_t<Rows,Cols,Type>& m, mat_t<Rows,Cols,Type>& adj_m, const bool &adj_ret)
+{
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> add(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
+{
+    mat_t<Rows,Cols,Type> t;
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            t.data[i][j] = a.data[i][j] + b.data[i][j];
+        }
+    }
+
+    return t;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> sub(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
+{
+    mat_t<Rows,Cols,Type> t;
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            t.data[i][j] = a.data[i][j] - b.data[i][j];
+        }
+    }
+
+    return t;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> div(const mat_t<Rows,Cols,Type>& a, Type b)
+{
+    mat_t<Rows,Cols,Type> t;
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            t.data[i][j] = a.data[i][j]/b;
+        }
+    }
+
+    return t;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> div(Type b, const mat_t<Rows,Cols,Type>& a)
+{
+    mat_t<Rows,Cols,Type> t;
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            t.data[i][j] = b / a.data[i][j];
+        }
+    }
+
+    return t;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> mul(const mat_t<Rows,Cols,Type>& a, Type b)
+{
+    mat_t<Rows,Cols,Type> t;
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            t.data[i][j] = a.data[i][j]*b;
+        }
+    }
+
+    return t;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> mul(Type b, const mat_t<Rows,Cols,Type>& a)
+{
+    return mul(a,b);
+}
+
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> operator*(Type b, const mat_t<Rows,Cols,Type>& a)
+{
+    return mul(a,b);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> operator*( const mat_t<Rows,Cols,Type>& a, Type b)
+{
+    return mul(a,b);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE vec_t<Rows,Type> mul(const mat_t<Rows,Cols,Type>& a, const vec_t<Cols,Type>& b)
+{
+    vec_t<Rows,Type> r = a.get_col(0)*b[0];
+    for( unsigned i=1; i < Cols; ++i )
+    {
+        r += a.get_col(i)*b[i];
+    }
+    return r;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE vec_t<Cols,Type> mul(const vec_t<Rows,Type>& b, const mat_t<Rows,Cols,Type>& a)
+{
+    vec_t<Cols,Type> r = a.get_row(0)*b[0];
+    for( unsigned i=1; i < Rows; ++i )
+    {
+        r += a.get_row(i)*b[i];
+    }
+    return r;
+}
+
+template<unsigned Rows, unsigned Cols, unsigned ColsOut, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,ColsOut,Type> mul(const mat_t<Rows,Cols,Type>& a, const mat_t<Cols,ColsOut,Type>& b)
+{
+    mat_t<Rows,ColsOut,Type> t(0);
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < ColsOut; ++j)
+        {
+            for (unsigned k=0; k < Cols; ++k)
+            {
+                t.data[i][j] += a.data[i][k]*b.data[k][j];
+            }
+        }
+    }
+    
+    return t;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE Type ddot(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
+{
+    // double dot product between a and b:
+    Type r(0);
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            r += a.data[i][j] * b.data[i][j];
+        }
+    }
+    return r;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE Type tensordot(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
+{
+    // corresponds to `np.tensordot()` with all axes being contracted
+    return ddot(a, b);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Cols,Rows,Type> transpose(const mat_t<Rows,Cols,Type>& a)
+{
+    mat_t<Cols,Rows,Type> t;
+    for (unsigned i=0; i < Cols; ++i)
+    {
+        for (unsigned j=0; j < Rows; ++j)
+        {
+            t.data[i][j] = a.data[j][i];
+        }
+    }
+
+    return t;
+}
+
+// Only implementing determinants for 2x2, 3x3 and 4x4 matrices for now...
+template<typename Type>
+inline CUDA_CALLABLE Type determinant(const mat_t<2,2,Type>& m)
+{
+    return m.data[0][0]*m.data[1][1] - m.data[1][0]*m.data[0][1];
+}
+
+template<typename Type>
+inline CUDA_CALLABLE Type determinant(const mat_t<3,3,Type>& m)
+{
+    return dot(
+        vec_t<3,Type>(m.data[0][0],m.data[0][1],m.data[0][2]),
+        cross(
+            vec_t<3,Type>(m.data[1][0],m.data[1][1],m.data[1][2]),
+            vec_t<3,Type>(m.data[2][0],m.data[2][1],m.data[2][2])
+        )
+    );
+}
+
+template<typename Type>
+inline CUDA_CALLABLE Type determinant(const mat_t<4,4,Type>& m)
+{
+    // adapted from USD GfMatrix4f::Inverse()
+    Type x00, x01, x02, x03;
+    Type x10, x11, x12, x13;
+    Type x20, x21, x22, x23;
+    Type x30, x31, x32, x33;
+    double y01, y02, y03, y12, y13, y23;
+    Type z00, z10, z20, z30;
+
+    // Pickle 1st two columns of matrix into registers
+    x00 = m.data[0][0];
+    x01 = m.data[0][1];
+    x10 = m.data[1][0];
+    x11 = m.data[1][1];
+    x20 = m.data[2][0];
+    x21 = m.data[2][1];
+    x30 = m.data[3][0];
+    x31 = m.data[3][1];
+
+    // Compute all six 2x2 determinants of 1st two columns
+    y01 = x00*x11 - x10*x01;
+    y02 = x00*x21 - x20*x01;
+    y03 = x00*x31 - x30*x01;
+    y12 = x10*x21 - x20*x11;
+    y13 = x10*x31 - x30*x11;
+    y23 = x20*x31 - x30*x21;
+
+    // Pickle 2nd two columns of matrix into registers
+    x02 = m.data[0][2];
+    x03 = m.data[0][3];
+    x12 = m.data[1][2];
+    x13 = m.data[1][3];
+    x22 = m.data[2][2];
+    x23 = m.data[2][3];
+    x32 = m.data[3][2];
+    x33 = m.data[3][3];
+
+    // Compute all six 2x2 determinants of 2nd two columns
+    y01 = x02*x13 - x12*x03;
+    y02 = x02*x23 - x22*x03;
+    y03 = x02*x33 - x32*x03;
+    y12 = x12*x23 - x22*x13;
+    y13 = x12*x33 - x32*x13;
+    y23 = x22*x33 - x32*x23;
+
+    // Compute all 3x3 cofactors for 1st two columns
+    z30 = x11*y02 - x21*y01 - x01*y12;
+    z20 = x01*y13 - x11*y03 + x31*y01;
+    z10 = x21*y03 - x31*y02 - x01*y23;
+    z00 = x11*y23 - x21*y13 + x31*y12;
+
+    // compute 4x4 determinant & its reciprocal
+    double det = x30*z30 + x20*z20 + x10*z10 + x00*z00;
+    return det;
+}
+
+template<unsigned Rows, typename Type>
+inline CUDA_CALLABLE Type trace(const mat_t<Rows,Rows,Type>& m)
+{
+    Type ret = m.data[0][0];
+    for( unsigned i=1; i < Rows; ++i )
+    {
+        ret += m.data[i][i];
+    }
+    return ret;
+}
+
+template<unsigned Rows, typename Type>
+inline CUDA_CALLABLE vec_t<Rows, Type> get_diag(const mat_t<Rows,Rows,Type>& m)
+{
+    vec_t<Rows, Type> ret;
+    for( unsigned i=0; i < Rows; ++i )
+    {
+        ret[i] = m.data[i][i];
+    }
+    return ret;
+}
+
+// Only implementing inverses for 2x2, 3x3 and 4x4 matrices for now...
+template<typename Type>
+inline CUDA_CALLABLE mat_t<2,2,Type> inverse(const mat_t<2,2,Type>& m)
+{
+    Type det = determinant(m);
+    if (det > Type(kEps) || det < -Type(kEps))
+    {
+        return mat_t<2,2,Type>( m.data[1][1], -m.data[0][1],
+                     -m.data[1][0],  m.data[0][0])*(Type(1.0f)/det);
+    }
+    else
+    {
+        return mat_t<2,2,Type>();
+    }
+}
+
+template<typename Type>
+inline CUDA_CALLABLE mat_t<3,3,Type> inverse(const mat_t<3,3,Type>& m)
+{
+	Type det = determinant(m);
+
+	if (det != Type(0.0f))
+	{
+		mat_t<3,3,Type> b;
+		
+		b.data[0][0] = m.data[1][1]*m.data[2][2] - m.data[1][2]*m.data[2][1]; 
+		b.data[1][0] = m.data[1][2]*m.data[2][0] - m.data[1][0]*m.data[2][2]; 
+		b.data[2][0] = m.data[1][0]*m.data[2][1] - m.data[1][1]*m.data[2][0]; 
+		
+        b.data[0][1] = m.data[0][2]*m.data[2][1] - m.data[0][1]*m.data[2][2]; 
+        b.data[1][1] = m.data[0][0]*m.data[2][2] - m.data[0][2]*m.data[2][0]; 
+        b.data[2][1] = m.data[0][1]*m.data[2][0] - m.data[0][0]*m.data[2][1]; 
+
+        b.data[0][2] = m.data[0][1]*m.data[1][2] - m.data[0][2]*m.data[1][1];
+        b.data[1][2] = m.data[0][2]*m.data[1][0] - m.data[0][0]*m.data[1][2];
+        b.data[2][2] = m.data[0][0]*m.data[1][1] - m.data[0][1]*m.data[1][0];
+
+		return b*(Type(1.0f)/det);
+	}
+	else
+	{
+		return mat_t<3,3,Type>();
+	}
+}
+
+template<typename Type>
+inline CUDA_CALLABLE mat_t<4,4,Type> inverse(const mat_t<4,4,Type>& m)
+{
+    // adapted from USD GfMatrix4f::Inverse()
+    Type x00, x01, x02, x03;
+    Type x10, x11, x12, x13;
+    Type x20, x21, x22, x23;
+    Type x30, x31, x32, x33;
+    double y01, y02, y03, y12, y13, y23;
+    Type z00, z10, z20, z30;
+    Type z01, z11, z21, z31;
+    double z02, z03, z12, z13, z22, z23, z32, z33;
+
+    // Pickle 1st two columns of matrix into registers
+    x00 = m.data[0][0];
+    x01 = m.data[0][1];
+    x10 = m.data[1][0];
+    x11 = m.data[1][1];
+    x20 = m.data[2][0];
+    x21 = m.data[2][1];
+    x30 = m.data[3][0];
+    x31 = m.data[3][1];
+
+    // Compute all six 2x2 determinants of 1st two columns
+    y01 = x00*x11 - x10*x01;
+    y02 = x00*x21 - x20*x01;
+    y03 = x00*x31 - x30*x01;
+    y12 = x10*x21 - x20*x11;
+    y13 = x10*x31 - x30*x11;
+    y23 = x20*x31 - x30*x21;
+
+    // Pickle 2nd two columns of matrix into registers
+    x02 = m.data[0][2];
+    x03 = m.data[0][3];
+    x12 = m.data[1][2];
+    x13 = m.data[1][3];
+    x22 = m.data[2][2];
+    x23 = m.data[2][3];
+    x32 = m.data[3][2];
+    x33 = m.data[3][3];
+
+    // Compute all 3x3 cofactors for 2nd two columns */
+    z33 = x02*y12 - x12*y02 + x22*y01;
+    z23 = x12*y03 - x32*y01 - x02*y13;
+    z13 = x02*y23 - x22*y03 + x32*y02;
+    z03 = x22*y13 - x32*y12 - x12*y23;
+    z32 = x13*y02 - x23*y01 - x03*y12;
+    z22 = x03*y13 - x13*y03 + x33*y01;
+    z12 = x23*y03 - x33*y02 - x03*y23;
+    z02 = x13*y23 - x23*y13 + x33*y12;
+
+    // Compute all six 2x2 determinants of 2nd two columns
+    y01 = x02*x13 - x12*x03;
+    y02 = x02*x23 - x22*x03;
+    y03 = x02*x33 - x32*x03;
+    y12 = x12*x23 - x22*x13;
+    y13 = x12*x33 - x32*x13;
+    y23 = x22*x33 - x32*x23;
+
+    // Compute all 3x3 cofactors for 1st two columns
+    z30 = x11*y02 - x21*y01 - x01*y12;
+    z20 = x01*y13 - x11*y03 + x31*y01;
+    z10 = x21*y03 - x31*y02 - x01*y23;
+    z00 = x11*y23 - x21*y13 + x31*y12;
+    z31 = x00*y12 - x10*y02 + x20*y01;
+    z21 = x10*y03 - x30*y01 - x00*y13;
+    z11 = x00*y23 - x20*y03 + x30*y02;
+    z01 = x20*y13 - x30*y12 - x10*y23;
+
+    // compute 4x4 determinant & its reciprocal
+    double det = x30*z30 + x20*z20 + x10*z10 + x00*z00;
+    
+    if(fabs(det) > kEps) 
+    {
+        mat_t<4,4,Type> invm;
+
+        double rcp = 1.0 / det;
+
+        // Multiply all 3x3 cofactors by reciprocal & transpose
+        invm.data[0][0] = Type(z00*rcp);
+        invm.data[0][1] = Type(z10*rcp);
+        invm.data[1][0] = Type(z01*rcp);
+        invm.data[0][2] = Type(z20*rcp);
+        invm.data[2][0] = Type(z02*rcp);
+        invm.data[0][3] = Type(z30*rcp);
+        invm.data[3][0] = Type(z03*rcp);
+        invm.data[1][1] = Type(z11*rcp);
+        invm.data[1][2] = Type(z21*rcp);
+        invm.data[2][1] = Type(z12*rcp);
+        invm.data[1][3] = Type(z31*rcp);
+        invm.data[3][1] = Type(z13*rcp);
+        invm.data[2][2] = Type(z22*rcp);
+        invm.data[2][3] = Type(z32*rcp);
+        invm.data[3][2] = Type(z23*rcp);
+        invm.data[3][3] = Type(z33*rcp);
+
+        return invm;
+    }
+    else 
+    {
+        return mat_t<4,4,Type>();
+    }
+}
+
+template<unsigned Rows,typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Rows,Type> diag(const vec_t<Rows,Type>& d)
+{
+    mat_t<Rows,Rows,Type> ret(Type(0));
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        ret.data[i][i] = d[i];
+    }
+    return ret;
+}
+
+template<unsigned Rows,unsigned Cols,typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> outer(const vec_t<Rows,Type>& a, const vec_t<Cols,Type>& b)
+{
+    // col 0 = a * b[0] etc...
+    mat_t<Rows,Cols,Type> ret;
+    for (unsigned row=0; row < Rows; ++row)
+    {
+        for (unsigned col=0; col < Cols; ++col) // columns
+        {
+            ret.data[row][col] = a[row] * b[col];
+        }
+    }
+    return ret;
+}
+
+template<unsigned Cols,typename Type>
+inline CUDA_CALLABLE vec_t<Cols,Type> outer(Type a, const vec_t<Cols,Type>& b)
+{
+    return mul(a, b);
+}
+
+template<unsigned Rows,typename Type>
+inline CUDA_CALLABLE vec_t<Rows,Type> outer(const vec_t<Rows,Type>& a, Type b)
+{
+    return mul(a, b);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE mat_t<3,3,Type> skew(const vec_t<3,Type>& a)
+{
+    mat_t<3,3,Type> out(
+        Type(0), -a[2],   a[1],
+        a[2],   Type(0), -a[0],
+        -a[1],   a[0],   Type(0)
+    );
+
+    return out;
+}
+
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> cw_mul(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
+{
+    mat_t<Rows,Cols,Type> t;
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            t.data[i][j] = a.data[i][j] * b.data[i][j];
+        }
+    }
+
+    return t;
+}
+
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE mat_t<Rows,Cols,Type> cw_div(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b)
+{
+    mat_t<Rows,Cols,Type> t;
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            t.data[i][j] = a.data[i][j] / b.data[i][j];
+        }
+    }
+
+    return t;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<3,Type> transform_point(const mat_t<4,4,Type>& m, const vec_t<3,Type>& v)
+{
+    vec_t<4,Type> out = mul(m, vec_t<4,Type>(v[0], v[1], v[2], Type(1)));
+    return vec_t<3,Type>(out[0], out[1], out[2]);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<3,Type> transform_vector(const mat_t<4,4,Type>& m, const vec_t<3,Type>& v)
+{
+    vec_t<4,Type> out = mul(m, vec_t<4,Type>(v[0], v[1], v[2], 0.f));
+    return vec_t<3,Type>(out[0], out[1], out[2]);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_extract(const mat_t<Rows,Cols,Type>& m, int row, mat_t<Rows,Cols,Type>& adj_m, int& adj_row, const vec_t<Cols,Type>& adj_ret)
+{
+    for( unsigned col=0; col < Cols; ++col )
+        adj_m.data[row][col] += adj_ret[col];
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline void CUDA_CALLABLE adj_extract(const mat_t<Rows,Cols,Type>& m, int row, int col, mat_t<Rows,Cols,Type>& adj_m, int& adj_row, int& adj_col, Type adj_ret)
+{
+#ifndef NDEBUG
+    if (row < 0 || row > Rows)
+    {
+        printf("mat row index %d out of bounds at %s %d\n", row, __FILE__, __LINE__);
+        assert(0);
+    }
+    if (col < 0 || col > Cols)
+    {
+        printf("mat col index %d out of bounds at %s %d\n", col, __FILE__, __LINE__);
+        assert(0);
+    }
+#endif
+    adj_m.data[row][col] += adj_ret;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_outer(const vec_t<Rows,Type>& a, const vec_t<Cols,Type>& b, vec_t<Rows,Type>& adj_a, vec_t<Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+  adj_a += mul(adj_ret, b);
+  adj_b += mul(transpose(adj_ret), a);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_add(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            adj_a.data[i][j] += adj_ret.data[i][j];
+            adj_b.data[i][j] += adj_ret.data[i][j];
+        }
+    }
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_sub(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            adj_a.data[i][j] += adj_ret.data[i][j];
+            adj_b.data[i][j] -= adj_ret.data[i][j];
+        }
+    }
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_div(const mat_t<Rows,Cols,Type>& a, Type s, mat_t<Rows,Cols,Type>& adj_a, Type& adj_s, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+    adj_s -= tensordot(a , adj_ret)/ (s * s); // - a / s^2
+
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            adj_a.data[i][j] += adj_ret.data[i][j] / s;
+        }
+    }
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_div(Type s, const mat_t<Rows,Cols,Type>& a, Type& adj_s, mat_t<Rows,Cols,Type>& adj_a, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+    adj_s -= tensordot(a , adj_ret)/ (s * s); // - a / s^2
+
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            adj_a.data[i][j] += s / adj_ret.data[i][j];
+        }
+    }
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_mul(const mat_t<Rows,Cols,Type>& a, Type b, mat_t<Rows,Cols,Type>& adj_a, Type& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            adj_a.data[i][j] += b*adj_ret.data[i][j];
+            adj_b += a.data[i][j]*adj_ret.data[i][j];
+        }
+    }
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_mul(Type b, const mat_t<Rows,Cols,Type>& a, Type& adj_b, mat_t<Rows,Cols,Type>& adj_a, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+    adj_mul(a, b, adj_a, adj_b, adj_ret);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_ddot(mat_t<Rows,Cols,Type> a, mat_t<Rows,Cols,Type> b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const Type adj_ret)
+{
+    adj_a += b*adj_ret;
+    adj_b += a*adj_ret;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_mul(const mat_t<Rows,Cols,Type>& a, const vec_t<Cols,Type>& b, mat_t<Rows,Cols,Type>& adj_a, vec_t<Cols,Type>& adj_b, const vec_t<Rows,Type>& adj_ret)
+{
+    adj_a += outer(adj_ret, b);
+    adj_b += mul(transpose(a), adj_ret);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_mul(const vec_t<Rows,Type>& b, const mat_t<Rows,Cols,Type>& a, vec_t<Rows,Type>& adj_b, mat_t<Rows,Cols,Type>& adj_a, const vec_t<Cols,Type>& adj_ret)
+{
+    adj_a += outer(b, adj_ret);
+    adj_b += mul(adj_ret, transpose(a));
+}
+
+template<unsigned Rows, unsigned Cols, unsigned ColsOut, typename Type>
+inline CUDA_CALLABLE void adj_mul(const mat_t<Rows,Cols,Type>& a, const mat_t<Cols,ColsOut,Type>& b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Cols,ColsOut,Type>& adj_b, const mat_t<Rows,ColsOut,Type>& adj_ret)
+{
+    adj_a += mul(adj_ret, transpose(b));
+    adj_b += mul(transpose(a), adj_ret);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_transpose(const mat_t<Rows,Cols,Type>& a, mat_t<Rows,Cols,Type>& adj_a, const mat_t<Cols,Rows,Type>& adj_ret)
+{
+    adj_a += transpose(adj_ret);
+}
+
+template<unsigned Rows, typename Type>
+inline CUDA_CALLABLE void adj_trace(const mat_t<Rows,Rows,Type>& m, mat_t<Rows,Rows,Type>& adj_m, Type adj_ret)
+{
+    for (unsigned i=0; i < Rows; ++i)
+        adj_m.data[i][i] += adj_ret;
+}
+
+template<unsigned Rows, typename Type>
+inline CUDA_CALLABLE void adj_diag(const vec_t<Rows,Type>& d, vec_t<Rows,Type>& adj_d, const mat_t<Rows,Rows,Type>& adj_ret)
+{
+    for (unsigned i=0; i < Rows; ++i)
+        adj_d[i] += adj_ret.data[i][i];
+}
+
+template<unsigned Rows, typename Type>
+inline CUDA_CALLABLE void adj_get_diag(const mat_t<Rows,Rows,Type>& m, mat_t<Rows,Rows,Type>& adj_m, const vec_t<Rows,Type>& adj_ret)
+{
+    for (unsigned i=0; i < Rows; ++i)
+        adj_m.data[i][i] += adj_ret[i];
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_determinant(const mat_t<2,2,Type>& m, mat_t<2,2,Type>& adj_m, Type adj_ret)
+{
+    adj_m.data[0][0] += m.data[1][1]*adj_ret;
+    adj_m.data[1][1] += m.data[0][0]*adj_ret;
+    adj_m.data[0][1] -= m.data[1][0]*adj_ret;
+    adj_m.data[1][0] -= m.data[0][1]*adj_ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_determinant(const mat_t<3,3,Type>& m, mat_t<3,3,Type>& adj_m, Type adj_ret)
+{
+    (vec_t<3,Type>&)adj_m.data[0] += cross(m.get_row(1), m.get_row(2))*adj_ret;
+    (vec_t<3,Type>&)adj_m.data[1] += cross(m.get_row(2), m.get_row(0))*adj_ret;
+    (vec_t<3,Type>&)adj_m.data[2] += cross(m.get_row(0), m.get_row(1))*adj_ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_determinant(const mat_t<4,4,Type>& m, mat_t<4,4,Type>& adj_m, Type adj_ret)
+{
+    // adapted from USD GfMatrix4f::Inverse()
+    Type x00, x01, x02, x03;
+    Type x10, x11, x12, x13;
+    Type x20, x21, x22, x23;
+    Type x30, x31, x32, x33;
+    double y01, y02, y03, y12, y13, y23;
+    Type z00, z10, z20, z30;
+    Type z01, z11, z21, z31;
+    double z02, z03, z12, z13, z22, z23, z32, z33;
+
+    // Pickle 1st two columns of matrix into registers
+    x00 = m.data[0][0];
+    x01 = m.data[0][1];
+    x10 = m.data[1][0];
+    x11 = m.data[1][1];
+    x20 = m.data[2][0];
+    x21 = m.data[2][1];
+    x30 = m.data[3][0];
+    x31 = m.data[3][1];
+
+    // Compute all six 2x2 determinants of 1st two columns
+    y01 = x00*x11 - x10*x01;
+    y02 = x00*x21 - x20*x01;
+    y03 = x00*x31 - x30*x01;
+    y12 = x10*x21 - x20*x11;
+    y13 = x10*x31 - x30*x11;
+    y23 = x20*x31 - x30*x21;
+
+    // Pickle 2nd two columns of matrix into registers
+    x02 = m.data[0][2];
+    x03 = m.data[0][3];
+    x12 = m.data[1][2];
+    x13 = m.data[1][3];
+    x22 = m.data[2][2];
+    x23 = m.data[2][3];
+    x32 = m.data[3][2];
+    x33 = m.data[3][3];
+
+    // Compute all 3x3 cofactors for 2nd two columns */
+    z33 = x02*y12 - x12*y02 + x22*y01;
+    z23 = x12*y03 - x32*y01 - x02*y13;
+    z13 = x02*y23 - x22*y03 + x32*y02;
+    z03 = x22*y13 - x32*y12 - x12*y23;
+    z32 = x13*y02 - x23*y01 - x03*y12;
+    z22 = x03*y13 - x13*y03 + x33*y01;
+    z12 = x23*y03 - x33*y02 - x03*y23;
+    z02 = x13*y23 - x23*y13 + x33*y12;
+
+    // Compute all six 2x2 determinants of 2nd two columns
+    y01 = x02*x13 - x12*x03;
+    y02 = x02*x23 - x22*x03;
+    y03 = x02*x33 - x32*x03;
+    y12 = x12*x23 - x22*x13;
+    y13 = x12*x33 - x32*x13;
+    y23 = x22*x33 - x32*x23;
+
+    // Compute all 3x3 cofactors for 1st two columns
+    z30 = x11*y02 - x21*y01 - x01*y12;
+    z20 = x01*y13 - x11*y03 + x31*y01;
+    z10 = x21*y03 - x31*y02 - x01*y23;
+    z00 = x11*y23 - x21*y13 + x31*y12;
+    z31 = x00*y12 - x10*y02 + x20*y01;
+    z21 = x10*y03 - x30*y01 - x00*y13;
+    z11 = x00*y23 - x20*y03 + x30*y02;
+    z01 = x20*y13 - x30*y12 - x10*y23;
+
+    // Multiply all 3x3 cofactors by adjoint & transpose
+    adj_m.data[0][0] += Type(z00*adj_ret);
+    adj_m.data[1][0] += Type(z10*adj_ret);
+    adj_m.data[0][1] += Type(z01*adj_ret);
+    adj_m.data[2][0] += Type(z20*adj_ret);
+    adj_m.data[0][2] += Type(z02*adj_ret);
+    adj_m.data[3][0] += Type(z30*adj_ret);
+    adj_m.data[0][3] += Type(z03*adj_ret);
+    adj_m.data[1][1] += Type(z11*adj_ret);
+    adj_m.data[2][1] += Type(z21*adj_ret);
+    adj_m.data[1][2] += Type(z12*adj_ret);
+    adj_m.data[3][1] += Type(z31*adj_ret);
+    adj_m.data[1][3] += Type(z13*adj_ret);
+    adj_m.data[2][2] += Type(z22*adj_ret);
+    adj_m.data[3][2] += Type(z32*adj_ret);
+    adj_m.data[2][3] += Type(z23*adj_ret);
+    adj_m.data[3][3] += Type(z33*adj_ret);
+}
+
+template<unsigned Rows, typename Type>
+inline CUDA_CALLABLE void adj_inverse(const mat_t<Rows,Rows,Type>& m, mat_t<Rows,Rows,Type>& ret, mat_t<Rows,Rows,Type>& adj_m, const mat_t<Rows,Rows,Type>& adj_ret)
+{
+    // todo: how to cache this from the forward pass?
+    mat_t<Rows,Rows,Type> invt = transpose(ret);
+
+    // see https://people.maths.ox.ac.uk/gilesm/files/NA-08-01.pdf 2.2.3
+    adj_m -= mul(mul(invt, adj_ret), invt);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_transform_point(const mat_t<4,4,Type>& m, const vec_t<3,Type>& v, mat_t<4,4,Type>& adj_m, vec_t<3,Type>& adj_v, const vec_t<3,Type>& adj_ret)
+{
+    vec_t<4,Type> out = vec_t<4,Type>(v[0], v[1], v[2], 1.f);
+    adj_m = add(adj_m, transpose(mat_t<4,4,Type>(adj_ret[0] * out, adj_ret[1] * out, adj_ret[2] * out, vec_t<4,Type>())));
+    adj_v[0] += dot(vec_t<3,Type>(m.data[0][0], m.data[1][0], m.data[2][0]), adj_ret);
+    adj_v[1] += dot(vec_t<3,Type>(m.data[0][1], m.data[1][1], m.data[2][1]), adj_ret);
+    adj_v[2] += dot(vec_t<3,Type>(m.data[0][2], m.data[1][2], m.data[2][2]), adj_ret);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_transform_vector(const mat_t<4,4,Type>& m, const vec_t<3,Type>& v, mat_t<4,4,Type>& adj_m, vec_t<3,Type>& adj_v, const vec_t<3,Type>& adj_ret)
+{
+    vec_t<4,Type> out = vec_t<4,Type>(v[0], v[1], v[2], 0.f);
+    adj_m = add(adj_m, transpose(mat_t<4,4,Type>(adj_ret[0] * out, adj_ret[1] * out, adj_ret[2] * out, vec_t<4,Type>())));
+    adj_v[0] += dot(vec_t<3,Type>(m.data[0][0], m.data[1][0], m.data[2][0]), adj_ret);
+    adj_v[1] += dot(vec_t<3,Type>(m.data[0][1], m.data[1][1], m.data[2][1]), adj_ret);
+    adj_v[2] += dot(vec_t<3,Type>(m.data[0][2], m.data[1][2], m.data[2][2]), adj_ret);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_skew(const vec_t<3,Type>& a, vec_t<3,Type>& adj_a, const mat_t<3,3,Type>& adj_ret)
+{
+    adj_a[0] += adj_ret.data[2][1] - adj_ret.data[1][2];
+    adj_a[1] += adj_ret.data[0][2] - adj_ret.data[2][0];
+    adj_a[2] += adj_ret.data[1][0] - adj_ret.data[0][1];
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_cw_mul(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+  adj_a += cw_mul(b, adj_ret);
+  adj_b += cw_mul(a, adj_ret);
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_cw_div(const mat_t<Rows,Cols,Type>& a, const mat_t<Rows,Cols,Type>& b, mat_t<Rows,Cols,Type>& ret, mat_t<Rows,Cols,Type>& adj_a, mat_t<Rows,Cols,Type>& adj_b, const mat_t<Rows,Cols,Type>& adj_ret)
+{
+  adj_a += cw_div(adj_ret, b);
+  adj_b -= cw_mul(adj_ret, cw_div(ret, b));
+}
+
+// adjoint for the constant constructor:
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_mat_t(Type s, Type& adj_s, const mat_t<Rows, Cols, Type>& adj_ret)
+{
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            adj_s += adj_ret.data[i][j];
+        }
+    }
+}
+
+// adjoint for the casting constructor:
+template<unsigned Rows, unsigned Cols, typename Type, typename OtherType>
+inline CUDA_CALLABLE void adj_mat_t(const mat_t<Rows, Cols, OtherType>& other, mat_t<Rows, Cols, OtherType>& adj_other, const mat_t<Rows, Cols, Type>& adj_ret)
+{
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            adj_other.data[i][j] += adj_ret.data[i][j];
+        }
+    }
+}
+
+// adjoint for the initializer_array scalar constructor:
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_mat_t(const initializer_array<Rows * Cols, Type> &cmps, const initializer_array<Rows * Cols, Type*> &adj_cmps, const mat_t<Rows, Cols, Type>& adj_ret)
+{
+    for (unsigned i=0; i < Rows; ++i)
+    {
+        for (unsigned j=0; j < Cols; ++j)
+        {
+            *adj_cmps[i * Cols + j] += adj_ret.data[i][j];
+        }
+    }
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_mat_t(Type m00, Type m01, Type m10, Type m11, Type& adj_m00, Type& adj_m01, Type& adj_m10, Type& adj_m11, const mat_t<2, 2, Type>& adj_ret)
+{
+    adj_m00 += adj_ret.data[0][0];
+    adj_m01 += adj_ret.data[0][1];
+    adj_m10 += adj_ret.data[1][0];
+    adj_m11 += adj_ret.data[1][1];
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_mat_t(Type m00, Type m01, Type m02,
+                      Type m10, Type m11, Type m12,
+                      Type m20, Type m21, Type m22,
+                      Type& a00, Type& a01, Type& a02,
+                      Type& a10, Type& a11, Type& a12,
+                      Type& a20, Type& a21, Type& a22,
+                      const mat_t<3, 3, Type>& adj_ret)
+{
+    a00 += adj_ret.data[0][0];
+    a01 += adj_ret.data[0][1];
+    a02 += adj_ret.data[0][2];
+    a10 += adj_ret.data[1][0];
+    a11 += adj_ret.data[1][1];
+    a12 += adj_ret.data[1][2];
+    a20 += adj_ret.data[2][0];
+    a21 += adj_ret.data[2][1];
+    a22 += adj_ret.data[2][2];
+}
+
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_mat_t(Type m00, Type m01, Type m02, Type m03,
+                      Type m10, Type m11, Type m12, Type m13,
+                      Type m20, Type m21, Type m22, Type m23,
+                      Type m30, Type m31, Type m32, Type m33,
+                      Type& a00, Type& a01, Type& a02, Type& a03,
+                      Type& a10, Type& a11, Type& a12, Type& a13,
+                      Type& a20, Type& a21, Type& a22, Type& a23,
+                      Type& a30, Type& a31, Type& a32, Type& a33,
+                      const mat_t<4, 4, Type>& adj_ret)
+{
+    a00 += adj_ret.data[0][0];
+    a01 += adj_ret.data[0][1];
+    a02 += adj_ret.data[0][2];
+    a03 += adj_ret.data[0][3];
+
+    a10 += adj_ret.data[1][0];
+    a11 += adj_ret.data[1][1];
+    a12 += adj_ret.data[1][2];
+    a13 += adj_ret.data[1][3];
+
+    a20 += adj_ret.data[2][0];
+    a21 += adj_ret.data[2][1];
+    a22 += adj_ret.data[2][2];
+    a23 += adj_ret.data[2][3];
+
+    a30 += adj_ret.data[3][0];
+    a31 += adj_ret.data[3][1];
+    a32 += adj_ret.data[3][2];
+    a33 += adj_ret.data[3][3];
+}
+
+
+
+// adjoint for the initializer_array vector constructor:
+template<unsigned Rows, unsigned Cols, typename Type>
+inline CUDA_CALLABLE void adj_mat_t(const initializer_array<Cols, vec_t<Rows,Type> > &cmps, const initializer_array<Cols, vec_t<Rows,Type>* > &adj_cmps, const mat_t<Rows, Cols, Type>& adj_ret)
+{
+    for (unsigned j=0; j < Cols; ++j)
+    {
+        for (unsigned i=0; i < Rows; ++i)
+        {
+            (*adj_cmps[j])[i] += adj_ret.data[i][j];
+        }
+    }
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_mat_t(const vec_t<2,Type> &cmps0, const vec_t<2,Type> &cmps1, vec_t<2,Type> &adj_cmps0, vec_t<2,Type> &adj_cmps1, const mat_t<2, 2, Type>& adj_ret)
+{
+    for (unsigned i=0; i < 2; ++i)
+    {
+        adj_cmps0[i] += adj_ret.data[i][0];
+        adj_cmps1[i] += adj_ret.data[i][1];
+    }
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_mat_t(const vec_t<3,Type> &cmps0, const vec_t<3,Type> &cmps1, const vec_t<3,Type> &cmps2, vec_t<3,Type> &adj_cmps0, vec_t<3,Type> &adj_cmps1, vec_t<3,Type> &adj_cmps2, const mat_t<3, 3, Type>& adj_ret)
+{
+    for (unsigned i=0; i < 3; ++i)
+    {
+        adj_cmps0[i] += adj_ret.data[i][0];
+        adj_cmps1[i] += adj_ret.data[i][1];
+        adj_cmps2[i] += adj_ret.data[i][2];
+    }
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_mat_t(const vec_t<4,Type> &cmps0, const vec_t<4,Type> &cmps1, const vec_t<4,Type> &cmps2, const vec_t<4,Type> &cmps3, vec_t<4,Type> &adj_cmps0, vec_t<4,Type> &adj_cmps1, vec_t<4,Type> &adj_cmps2, vec_t<4,Type> &adj_cmps3, const mat_t<4, 4, Type>& adj_ret)
+{
+    for (unsigned i=0; i < 4; ++i)
+    {
+        adj_cmps0[i] += adj_ret.data[i][0];
+        adj_cmps1[i] += adj_ret.data[i][1];
+        adj_cmps2[i] += adj_ret.data[i][2];
+        adj_cmps3[i] += adj_ret.data[i][3];
+    }
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+CUDA_CALLABLE inline mat_t<Rows, Cols, Type> lerp(const mat_t<Rows, Cols, Type>& a, const mat_t<Rows, Cols, Type>& b, Type t)
+{
+    return a*(Type(1)-t) + b*t;
+}
+
+template<unsigned Rows, unsigned Cols, typename Type>
+CUDA_CALLABLE inline void adj_lerp(const mat_t<Rows, Cols, Type>& a, const mat_t<Rows, Cols, Type>& b, Type t, mat_t<Rows, Cols, Type>& adj_a, mat_t<Rows, Cols, Type>& adj_b, Type& adj_t, const mat_t<Rows, Cols, 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);
+}
+
+// for integral types we do not accumulate gradients
+template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, int8>* buf, const mat_t<Rows, Cols, int8> &value) { }
+template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, uint8>* buf, const mat_t<Rows, Cols, uint8> &value) { }
+template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, int16>* buf, const mat_t<Rows, Cols, int16> &value) { }
+template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, uint16>* buf, const mat_t<Rows, Cols, uint16> &value) { }
+template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, int32>* buf, const mat_t<Rows, Cols, int32> &value) { }
+template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, uint32>* buf, const mat_t<Rows, Cols, uint32> &value) { }
+template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, int64>* buf, const mat_t<Rows, Cols, int64> &value) { }
+template<unsigned Rows, unsigned Cols> CUDA_CALLABLE inline void adj_atomic_add(mat_t<Rows, Cols, uint64>* buf, const mat_t<Rows, Cols, uint64> &value) { }
+
+using mat22h = mat_t<2,2,half>;
+using mat33h = mat_t<3,3,half>;
+using mat44h = mat_t<4,4,half>;
+
+using mat22 = mat_t<2,2,float>;
+using mat33 = mat_t<3,3,float>;
+using mat44 = mat_t<4,4,float>;
+
+using mat22f = mat_t<2,2,float>;
+using mat33f = mat_t<3,3,float>;
+using mat44f = mat_t<4,4,float>;
+
+using mat22d = mat_t<2,2,double>;
+using mat33d = mat_t<3,3,double>;
+using mat44d = mat_t<4,4,double>;
+
+inline CUDA_CALLABLE void adj_mat22(vec2 c0, vec2 c1,
+                      vec2& a0, vec2& a1,
+                      const mat22& adj_ret)
+{
+    a0 += adj_ret.get_col(0);
+    a1 += adj_ret.get_col(1);
+}
+
+inline CUDA_CALLABLE void adj_mat22(float m00, float m01, float m10, float m11, float& adj_m00, float& adj_m01, float& adj_m10, float& adj_m11, const mat22& adj_ret)
+{
+    adj_m00 += adj_ret.data[0][0];
+    adj_m01 += adj_ret.data[0][1];
+    adj_m10 += adj_ret.data[1][0];
+    adj_m11 += adj_ret.data[1][1];
+}
+
+inline CUDA_CALLABLE void adj_mat33(vec3 c0, vec3 c1, vec3 c2,
+                      vec3& a0, vec3& a1, vec3& a2,
+                      const mat33& adj_ret)
+{
+    // column constructor
+    a0 += adj_ret.get_col(0);
+    a1 += adj_ret.get_col(1);
+    a2 += adj_ret.get_col(2);
+
+}
+
+inline CUDA_CALLABLE void adj_mat33(float m00, float m01, float m02,
+                      float m10, float m11, float m12,
+                      float m20, float m21, float m22,
+                      float& a00, float& a01, float& a02,
+                      float& a10, float& a11, float& a12,
+                      float& a20, float& a21, float& a22,
+                      const mat33& adj_ret)
+{
+    a00 += adj_ret.data[0][0];
+    a01 += adj_ret.data[0][1];
+    a02 += adj_ret.data[0][2];
+    a10 += adj_ret.data[1][0];
+    a11 += adj_ret.data[1][1];
+    a12 += adj_ret.data[1][2];
+    a20 += adj_ret.data[2][0];
+    a21 += adj_ret.data[2][1];
+    a22 += adj_ret.data[2][2];
+}
+
+inline CUDA_CALLABLE void adj_mat44(
+    vec4 c0, vec4 c1, vec4 c2, vec4 c3,
+    vec4& a0, vec4& a1, vec4& a2, vec4& a3,
+    const mat44& adj_ret)
+{
+    // column constructor
+    a0 += adj_ret.get_col(0);
+    a1 += adj_ret.get_col(1);
+    a2 += adj_ret.get_col(2);
+    a3 += adj_ret.get_col(3);
+}
+
+inline CUDA_CALLABLE void adj_mat44(float m00, float m01, float m02, float m03,
+                      float m10, float m11, float m12, float m13,
+                      float m20, float m21, float m22, float m23,
+                      float m30, float m31, float m32, float m33,
+                      float& a00, float& a01, float& a02, float& a03,
+                      float& a10, float& a11, float& a12, float& a13,
+                      float& a20, float& a21, float& a22, float& a23,
+                      float& a30, float& a31, float& a32, float& a33,
+                      const mat44& adj_ret)
+{
+    a00 += adj_ret.data[0][0];
+    a01 += adj_ret.data[0][1];
+    a02 += adj_ret.data[0][2];
+    a03 += adj_ret.data[0][3];
+
+    a10 += adj_ret.data[1][0];
+    a11 += adj_ret.data[1][1];
+    a12 += adj_ret.data[1][2];
+    a13 += adj_ret.data[1][3];
+
+    a20 += adj_ret.data[2][0];
+    a21 += adj_ret.data[2][1];
+    a22 += adj_ret.data[2][2];
+    a23 += adj_ret.data[2][3];
+
+    a30 += adj_ret.data[3][0];
+    a31 += adj_ret.data[3][1];
+    a32 += adj_ret.data[3][2];
+    a33 += adj_ret.data[3][3];
+}
+
+} // namespace wp