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

warp/native/vec.h

@@ -1,1133 +1,1177 @@
-/** 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
-{
-
-template<unsigned Length, typename Type>
-struct vec_t
-{
-    Type c[Length];
-
-    inline CUDA_CALLABLE vec_t()
-        : c()
-    {}
-    
-    inline CUDA_CALLABLE vec_t(Type s)
-    {
-        for( unsigned i=0; i < Length; ++i )
-        {
-            c[i] = s;
-        }
-    }
-    
-    template <typename OtherType>
-    inline explicit CUDA_CALLABLE vec_t(const vec_t<Length, OtherType>& other)
-    {
-        for( unsigned i=0; i < Length; ++i )
-        {
-            c[i] = static_cast<Type>(other[i]);
-        }
-    }
-
-    inline CUDA_CALLABLE vec_t(Type x, Type y)
-    {
-        assert(Length == 2);
-        c[0]=x;
-        c[1]=y;
-    }
-
-    inline CUDA_CALLABLE vec_t(Type x, Type y, Type z)
-    {
-        assert(Length == 3);
-        c[0]=x;
-        c[1]=y;
-        c[2]=z;
-    }
-
-
-    inline CUDA_CALLABLE vec_t(Type x, Type y, Type z, Type w)
-    {
-        assert(Length == 4);
-        c[0]=x;
-        c[1]=y;
-        c[2]=z;
-        c[3]=w;
-    }
-    
-    inline CUDA_CALLABLE vec_t(const initializer_array<Length, Type> &l)
-    {
-        for( unsigned i=0; i < Length; ++i )
-        {
-            c[i] = l[i];
-        }
-    }
-    
-    // special screw vector constructor for spatial_vectors:
-    inline CUDA_CALLABLE vec_t(vec_t<3,Type> w, vec_t<3,Type> v)
-    {
-        c[0] = w[0];
-        c[1] = w[1];
-        c[2] = w[2];
-        c[3] = v[0];
-        c[4] = v[1];
-        c[5] = v[2];
-    }
-
-    inline CUDA_CALLABLE Type operator[](int index) const
-    {
-        assert(index < Length);
-        return c[index];
-    }
-
-    inline CUDA_CALLABLE Type& operator[](int index)
-    {
-        assert(index < Length);
-        return c[index];
-    }
-};
-
-using vec2b = vec_t<2,int8>;
-using vec3b = vec_t<3,int8>;
-using vec4b = vec_t<4,int8>;
-using vec2ub = vec_t<2,uint8>;
-using vec3ub = vec_t<3,uint8>;
-using vec4ub = vec_t<4,uint8>;
-
-using vec2s = vec_t<2,int16>;
-using vec3s = vec_t<3,int16>;
-using vec4s = vec_t<4,int16>;
-using vec2us = vec_t<2,uint16>;
-using vec3us = vec_t<3,uint16>;
-using vec4us = vec_t<4,uint16>;
-
-using vec2i = vec_t<2,int32>;
-using vec3i = vec_t<3,int32>;
-using vec4i = vec_t<4,int32>;
-using vec2ui = vec_t<2,uint32>;
-using vec3ui = vec_t<3,uint32>;
-using vec4ui = vec_t<4,uint32>;
-
-using vec2l = vec_t<2,int64>;
-using vec3l = vec_t<3,int64>;
-using vec4l = vec_t<4,int64>;
-using vec2ul = vec_t<2,uint64>;
-using vec3ul = vec_t<3,uint64>;
-using vec4ul = vec_t<4,uint64>;
-
-using vec2h = vec_t<2,half>;
-using vec3h = vec_t<3,half>;
-using vec4h = vec_t<4,half>;
-
-using vec2 = vec_t<2,float>;
-using vec3 = vec_t<3,float>;
-using vec4 = vec_t<4,float>;
-
-using vec2f = vec_t<2,float>;
-using vec3f = vec_t<3,float>;
-using vec4f = vec_t<4,float>;
-
-using vec2d = vec_t<2,double>;
-using vec3d = vec_t<3,double>;
-using vec4d = vec_t<4,double>;
-
-//--------------
-// vec<Length, Type> methods
-
-// Should these accept const references as arguments? It's all
-// inlined so maybe it doesn't matter? Even if it does, it
-// probably depends on the Length of the vector...
-
-// negation:
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> operator - (vec_t<Length, Type> a)
-{
-    // NB: this constructor will initialize all ret's components to 0, which is
-    // unnecessary... 
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = -a[i];
-    }
-
-    // Wonder if this does a load of copying when it returns... hopefully not as it's inlined?
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-CUDA_CALLABLE inline vec_t<Length, Type> pos(const vec_t<Length, Type>& x)
-{
-    return x;
-}
-
-template<unsigned Length, typename Type>
-CUDA_CALLABLE inline vec_t<Length, Type> neg(const vec_t<Length, Type>& x)
-{
-    return -x;
-}
-
-template<typename Type>
-CUDA_CALLABLE inline vec_t<3, Type> neg(const vec_t<3, Type>& x)
-{
-    return vec_t<3, Type>(-x.c[0], -x.c[1], -x.c[2]);
-}
-
-template<typename Type>
-CUDA_CALLABLE inline vec_t<2, Type> neg(const vec_t<2, Type>& x)
-{
-    return vec_t<2, Type>(-x.c[0], -x.c[1]);
-}
-
-template<unsigned Length, typename Type>
-CUDA_CALLABLE inline void adj_neg(const vec_t<Length, Type>& x, vec_t<Length, Type>& adj_x, const vec_t<Length, Type>& adj_ret)
-{
-    adj_x -= adj_ret;
-}
-
-// equality:
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE bool operator ==(const vec_t<Length, Type>& a, const vec_t<Length, Type>& b)
-{
-    for( unsigned i=0; i < Length; ++i )
-    {
-        if(a[i] != b[i])
-        {
-            return false;
-        }
-    }
-    return true;
-}
-
-// scalar multiplication:
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> mul(vec_t<Length, Type> a, Type s)
-{
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = a[i] * s;
-    }
-    return ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<3, Type> mul(vec_t<3, Type> a, Type s)
-{
-    return vec_t<3, Type>(a.c[0]*s,a.c[1]*s,a.c[2]*s);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<2, Type> mul(vec_t<2, Type> a, Type s)
-{
-    return vec_t<2, Type>(a.c[0]*s,a.c[1]*s);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> mul(Type s, vec_t<Length, Type> a)
-{
-    return mul(a, s);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> operator*(Type s, vec_t<Length, Type> a)
-{
-    return mul(a, s);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> operator*(vec_t<Length, Type> a, Type s)
-{
-    return mul(a, s);
-}
-
-
-// component wise multiplication:
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> cw_mul(vec_t<Length, Type> a, vec_t<Length, Type> b)
-{
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = a[i] * b[i];
-    }
-    return ret;
-}
-
-// division
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> div(vec_t<Length, Type> a, Type s)
-{
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = a[i] / s;
-    }
-    return ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<3, Type> div(vec_t<3, Type> a, Type s)
-{
-    return vec_t<3, Type>(a.c[0]/s,a.c[1]/s,a.c[2]/s);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<2, Type> div(vec_t<2, Type> a, Type s)
-{
-    return vec_t<2, Type>(a.c[0]/s,a.c[1]/s);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> div(Type s, vec_t<Length, Type> a)
-{
-    vec_t<Length, Type> ret;
-    for (unsigned i=0; i < Length; ++i)
-    {
-        ret[i] = s / a[i];
-    }
-    return ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<3, Type> div(Type s, vec_t<3, Type> a)
-{
-    return vec_t<3, Type>(s/a.c[0],s/a.c[1],s/a.c[2]);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<2, Type> div(Type s, vec_t<2, Type> a)
-{
-    return vec_t<2, Type>(s/a.c[0],s/a.c[1]);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> operator / (vec_t<Length, Type> a, Type s)
-{
-    return div(a,s);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> operator / (Type s, vec_t<Length, Type> a)
-{
-    return div(s, a);
-}
-
-// component wise division
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> cw_div(vec_t<Length, Type> a, vec_t<Length, Type> b)
-{
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = a[i] / b[i];
-    }
-    return ret;
-}
-
-// addition
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> add(vec_t<Length, Type> a, vec_t<Length, Type> b)
-{
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = a[i] + b[i];
-    }
-    return ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<2, Type> add(vec_t<2, Type> a, vec_t<2, Type> b)
-{
-    return vec_t<2, Type>( a.c[0] + b.c[0], a.c[1] + b.c[1]);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<3, Type> add(vec_t<3, Type> a, vec_t<3, Type> b)
-{
-    return vec_t<3, Type>( a.c[0] + b.c[0], a.c[1] + b.c[1], a.c[2] + b.c[2]);
-}
-
-// subtraction
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> sub(vec_t<Length, Type> a, vec_t<Length, Type> b)
-{
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = Type(a[i] - b[i]);
-    }
-    return ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<2, Type> sub(vec_t<2, Type> a, vec_t<2, Type> b)
-{
-    return vec_t<2, Type>( a.c[0] - b.c[0], a.c[1] - b.c[1]);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<3, Type> sub(vec_t<3, Type> a, vec_t<3, Type> b)
-{
-    return vec_t<3, Type>( a.c[0] - b.c[0], a.c[1] - b.c[1], a.c[2] - b.c[2]);
-}
-
-// dot product:
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE Type dot(vec_t<Length, Type> a, vec_t<Length, Type> b)
-{
-    Type ret(0);
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret += a[i] * b[i];
-    }
-    return ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE Type dot(vec_t<2, Type> a, vec_t<2, Type> b)
-{
-    return a.c[0] * b.c[0] + a.c[1] * b.c[1];
-}
-
-template<typename Type>
-inline CUDA_CALLABLE Type dot(vec_t<3, Type> a, vec_t<3, Type> b)
-{
-    return a.c[0] * b.c[0] + a.c[1] * b.c[1] + a.c[2] * b.c[2];
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE Type tensordot(vec_t<Length, Type> a, vec_t<Length, Type> b)
-{
-    // corresponds to `np.tensordot()` with all axes being contracted
-    return dot(a, b);
-}
-
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE Type extract(const vec_t<Length, Type> & a, int idx)
-{
-#ifndef NDEBUG
-    if (idx < 0 || idx >= Length)
-    {
-        printf("vec index %d out of bounds at %s %d\n", idx, __FILE__, __LINE__);
-        assert(0);
-    }
-#endif
-
-    return a[idx];        
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE Type* index(vec_t<Length, Type>& v, int idx)
-{
-#ifndef NDEBUG
-    if (idx < 0 || idx >= Length)
-    {
-        printf("vec index %d out of bounds at %s %d\n", idx, __FILE__, __LINE__);
-        assert(0);
-    }
-#endif
-
-    return &v[idx];
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE Type* indexref(vec_t<Length, Type>* v, int idx)
-{
-#ifndef NDEBUG
-    if (idx < 0 || idx >= Length)
-    {
-        printf("vec store %d out of bounds at %s %d\n", idx, __FILE__, __LINE__);
-        assert(0);
-    }
-#endif
-
-    return &((*v)[idx]);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_index(vec_t<Length, Type>& v, int idx,
-                                       vec_t<Length, Type>& adj_v, int adj_idx, const Type& adj_value)
-{
-    // nop
-}
-
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_indexref(vec_t<Length, Type>* v, int idx, 
-                                       vec_t<Length, Type>& adj_v, int adj_idx, const Type& adj_value)
-{
-    // nop
-}
-
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE Type length(vec_t<Length, Type> a)
-{
-    return sqrt(dot(a, a));
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE Type length_sq(vec_t<Length, Type> a)
-{
-    return dot(a, a);
-}
-
-
-template<typename Type>
-inline CUDA_CALLABLE Type length(vec_t<2, Type> a)
-{
-    return sqrt(a.c[0] * a.c[0] + a.c[1] * a.c[1]);
-}
-
-template<typename Type>
-inline CUDA_CALLABLE Type length(vec_t<3, Type> a)
-{
-    return sqrt(a.c[0] * a.c[0] + a.c[1] * a.c[1] + a.c[2] * a.c[2]);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> normalize(vec_t<Length, Type> a)
-{
-    Type l = length(a);
-    if (l > Type(kEps))
-        return div(a,l);
-    else
-        return vec_t<Length, Type>();
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<2, Type> normalize(vec_t<2, Type> a)
-{
-    Type l = sqrt(a.c[0] * a.c[0] + a.c[1] * a.c[1]);
-    if (l > Type(kEps))
-        return vec_t<2, Type>(a.c[0]/l,a.c[1]/l);
-    else
-        return vec_t<2, Type>();
-}
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<3, Type> normalize(vec_t<3, Type> a)
-{
-    Type l = sqrt(a.c[0] * a.c[0] + a.c[1] * a.c[1] + a.c[2] * a.c[2]);
-    if (l > Type(kEps))
-        return vec_t<3, Type>(a.c[0]/l,a.c[1]/l,a.c[2]/l);
-    else
-        return vec_t<3, Type>();
-}
-
-
-template<typename Type>
-inline CUDA_CALLABLE vec_t<3,Type> cross(vec_t<3,Type> a, vec_t<3,Type> b)
-{
-    return {
-        Type(a[1]*b[2] - a[2]*b[1]),
-        Type(a[2]*b[0] - a[0]*b[2]),
-        Type(a[0]*b[1] - a[1]*b[0])
-    };
-}
-
-
-template<unsigned Length, typename Type>
-inline bool CUDA_CALLABLE isfinite(vec_t<Length, Type> x)
-{
-    for( unsigned i=0; i < Length; ++i )
-    {
-        if(!isfinite(x[i]))
-        {
-            return false;
-        }
-    }
-    return true;
-}
-
-// These two functions seem to compile very slowly
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length,Type> min(vec_t<Length,Type> a, vec_t<Length,Type> b)
-{
-    vec_t<Length,Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = a[i] < b[i] ? a[i] : b[i];
-    }
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length,Type> max(vec_t<Length,Type> a, vec_t<Length,Type> b)
-{
-    vec_t<Length,Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = a[i] > b[i] ? a[i] : b[i];
-    }
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE Type min(vec_t<Length,Type> v)
-{
-    Type ret = v[0];
-    for( unsigned i=1; i < Length; ++i )
-    {
-        if (v[i] < ret)
-            ret = v[i];
-    }
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE Type max(vec_t<Length,Type> v)
-{
-    Type ret = v[0];
-    for( unsigned i=1; i < Length; ++i )
-    {
-        if (v[i] > ret)
-            ret = v[i];
-    }
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE unsigned argmin(vec_t<Length,Type> v)
-{
-    unsigned ret = 0;
-    for( unsigned i=1; i < Length; ++i )
-    {
-        if (v[i] < v[ret])
-            ret = i;
-    }
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE unsigned argmax(vec_t<Length,Type> v)
-{
-    unsigned ret = 0;
-    for( unsigned i=1; i < Length; ++i )
-    {
-        if (v[i] > v[ret])
-            ret = i;
-    }
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void expect_near(const vec_t<Length, Type>& actual, const vec_t<Length, Type>& expected, const Type& tolerance)
-{
-    const Type diff(0);
-    for(size_t i=0; i<Length; ++i)
-    {
-        diff = max(diff,abs(actual[i] - expected[i]));
-    }
-    if (diff > tolerance)
-    {
-        printf("Error, expect_near() failed with tolerance "); print(tolerance);
-        printf("\t Expected: "); print(expected); 
-        printf("\t Actual: "); print(actual);
-    }
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_expect_near(const vec_t<Length, Type>& actual, const vec_t<Length, Type>& expected, Type tolerance, vec_t<Length, Type>& adj_actual, vec_t<Length, Type>& adj_expected, Type adj_tolerance)
-{
-    // nop
-}
-
-// adjoint for the initializer_array constructor:
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_vec_t(const initializer_array<Length, Type> &cmps, const initializer_array<Length, Type*> &adj_cmps, const vec_t<Length, Type>& adj_ret)
-{
-    for(unsigned i=0; i < Length; ++i)
-    {
-        *(adj_cmps[i]) += adj_ret[i];
-    }
-}
-
-
-// adjoint for the component constructors:
-template<typename Type>
-inline CUDA_CALLABLE void adj_vec_t(Type cmpx, Type cmpy, Type &adj_cmpx, Type &adj_cmpy, const vec_t<2, Type>& adj_ret)
-{
-    adj_cmpx += adj_ret.c[0];
-    adj_cmpy += adj_ret.c[1];
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_vec_t(Type cmpx, Type cmpy, Type cmpz, Type &adj_cmpx, Type &adj_cmpy, Type &adj_cmpz, const vec_t<3, Type>& adj_ret)
-{
-    adj_cmpx += adj_ret.c[0];
-    adj_cmpy += adj_ret.c[1];
-    adj_cmpz += adj_ret.c[2];
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_vec_t(Type cmpx, Type cmpy, Type cmpz, Type cmpw, Type &adj_cmpx, Type &adj_cmpy, Type &adj_cmpz, Type &adj_cmpw, const vec_t<4, Type>& adj_ret)
-{
-    adj_cmpx += adj_ret.c[0];
-    adj_cmpy += adj_ret.c[1];
-    adj_cmpz += adj_ret.c[2];
-    adj_cmpw += adj_ret.c[3];
-}
-
-// adjoint for the constant constructor:
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_vec_t(Type s, Type& adj_s, const vec_t<Length, Type>& adj_ret)
-{
-    for( unsigned i=0; i < Length; ++i )
-    {
-        adj_s += adj_ret[i];
-    }
-}
-
-// adjoint for the casting constructor
-template<unsigned Length, typename Type, typename OtherType>
-inline CUDA_CALLABLE void adj_vec_t(const vec_t<Length, OtherType>& other, vec_t<Length, OtherType>& adj_other, const vec_t<Length, Type>& adj_ret)
-{
-    for( unsigned i=0; i < Length; ++i )
-    {
-        adj_other[i] += static_cast<OtherType>(adj_ret[i]);
-    }
-}
-
-template<typename Type>
-CUDA_CALLABLE inline void adj_vec_t(const vec_t<3,Type>& w, const vec_t<3,Type>& v, vec_t<3,Type>& adj_w, vec_t<3,Type>& adj_v, const vec_t<6,Type>& adj_ret)
-{
-    adj_w[0] += adj_ret[0];
-    adj_w[1] += adj_ret[1];
-    adj_w[2] += adj_ret[2];
-    adj_v[0] += adj_ret[3];
-    adj_v[1] += adj_ret[4];
-    adj_v[2] += adj_ret[5];
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_mul(vec_t<Length, Type> a, Type s, vec_t<Length, Type>& adj_a, Type& adj_s, const vec_t<Length, Type>& adj_ret)
-{
-    for( unsigned i=0; i < Length; ++i )
-    {
-        adj_a[i] += s*adj_ret[i];
-    }
-
-    adj_s += dot(a, adj_ret);
-
-#if FP_CHECK
-    if (!isfinite(a) || !isfinite(s) || !isfinite(adj_a) || !isfinite(adj_s) || !isfinite(adj_ret))
-    {
-        // \TODO: How shall we implement this error message?
-        //printf("adj_mul((%f %f %f %f), %f, (%f %f %f %f), %f, (%f %f %f %f)\n", a.x, a.y, a.z, a.w, s, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_s, adj_ret.x, adj_ret.y, adj_ret.z, adj_ret.w);
-        assert(0);
-    }
-#endif
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_mul(Type s, vec_t<Length, Type> a, Type& adj_s, vec_t<Length, Type>& adj_a, const vec_t<Length, Type>& adj_ret)
-{
-    adj_mul(a, s, adj_a, adj_s, adj_ret);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_cw_mul(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const vec_t<Length, Type>& adj_ret)
-{
-  adj_a += cw_mul(b, adj_ret);
-  adj_b += cw_mul(a, adj_ret);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_div(vec_t<Length, Type> a, Type s, vec_t<Length, Type>& adj_a, Type& adj_s, const vec_t<Length, Type>& adj_ret)
-{
-
-    adj_s -= dot(a , adj_ret)/ (s * s); // - a / s^2
-
-    for( unsigned i=0; i < Length; ++i )
-    {
-        adj_a[i] += adj_ret[i] / s;
-    }
-
-#if FP_CHECK
-    if (!isfinite(a) || !isfinite(s) || !isfinite(adj_a) || !isfinite(adj_s) || !isfinite(adj_ret))
-    {
-        // \TODO: How shall we implement this error message?
-        // printf("adj_div((%f %f %f %f), %f, (%f %f %f %f), %f, (%f %f %f %f)\n", a.x, a.y, a.z, a.w, s, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_s, adj_ret.x, adj_ret.y, adj_ret.z, adj_ret.w);
-        assert(0);
-    }
-#endif
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_div(Type s, vec_t<Length, Type> a, Type& adj_s, vec_t<Length, Type>& adj_a, const vec_t<Length, Type>& adj_ret)
-{
-
-    adj_s -= dot(a , adj_ret)/ (s * s); // - a / s^2
-
-    for( unsigned i=0; i < Length; ++i )
-    {
-        adj_a[i] += s / adj_ret[i];
-    }
-
-#if FP_CHECK
-    if (!isfinite(a) || !isfinite(s) || !isfinite(adj_a) || !isfinite(adj_s) || !isfinite(adj_ret))
-    {
-        // \TODO: How shall we implement this error message?
-        // printf("adj_div((%f %f %f %f), %f, (%f %f %f %f), %f, (%f %f %f %f)\n", a.x, a.y, a.z, a.w, s, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_s, adj_ret.x, adj_ret.y, adj_ret.z, adj_ret.w);
-        assert(0);
-    }
-#endif
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_cw_div(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& ret, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const vec_t<Length, Type>& adj_ret) {
-  adj_a += cw_div(adj_ret, b);
-  adj_b -= cw_mul(adj_ret, cw_div(ret, b));
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_add(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const vec_t<Length, Type>& adj_ret)
-{
-    adj_a += adj_ret;
-    adj_b += adj_ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_add(vec_t<2, Type> a, vec_t<2, Type> b, vec_t<2, Type>& adj_a, vec_t<2, Type>& adj_b, const vec_t<2, Type>& adj_ret)
-{
-    adj_a.c[0] += adj_ret.c[0];
-    adj_a.c[1] += adj_ret.c[1];
-    adj_b.c[0] += adj_ret.c[0];
-    adj_b.c[1] += adj_ret.c[1];
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_add(vec_t<3, Type> a, vec_t<3, Type> b, vec_t<3, Type>& adj_a, vec_t<3, Type>& adj_b, const vec_t<3, Type>& adj_ret)
-{
-    adj_a.c[0] += adj_ret.c[0];
-    adj_a.c[1] += adj_ret.c[1];
-    adj_a.c[2] += adj_ret.c[2];
-    adj_b.c[0] += adj_ret.c[0];
-    adj_b.c[1] += adj_ret.c[1];
-    adj_b.c[2] += adj_ret.c[2];
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_sub(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const vec_t<Length, Type>& adj_ret)
-{
-    adj_a += adj_ret;
-    adj_b -= adj_ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_sub(vec_t<2, Type> a, vec_t<2, Type> b, vec_t<2, Type>& adj_a, vec_t<2, Type>& adj_b, const vec_t<2, Type>& adj_ret)
-{
-    adj_a.c[0] += adj_ret.c[0];
-    adj_a.c[1] += adj_ret.c[1];
-    adj_b.c[0] -= adj_ret.c[0];
-    adj_b.c[1] -= adj_ret.c[1];
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_sub(vec_t<3, Type> a, vec_t<3, Type> b, vec_t<3, Type>& adj_a, vec_t<3, Type>& adj_b, const vec_t<3, Type>& adj_ret)
-{
-    adj_a.c[0] += adj_ret.c[0];
-    adj_a.c[1] += adj_ret.c[1];
-    adj_a.c[2] += adj_ret.c[2];
-    adj_b.c[0] -= adj_ret.c[0];
-    adj_b.c[1] -= adj_ret.c[1];
-    adj_b.c[2] -= adj_ret.c[2];
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_dot(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const Type adj_ret)
-{
-    adj_a += b*adj_ret;
-    adj_b += a*adj_ret;
-
-#if FP_CHECK
-    if (!isfinite(a) || !isfinite(b) || !isfinite(adj_a) || !isfinite(adj_b) || !isfinite(adj_ret))
-    {
-        // \TODO: How shall we implement this error message?
-        //printf("adj_dot((%f %f %f %f), (%f %f %f %f), (%f %f %f %f), (%f %f %f %f), %f)\n", a.x, a.y, a.z, a.w, b.x, b.y, b.z, b.w, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_b.x, adj_b.y, adj_b.z, adj_b.w, adj_ret);
-        assert(0);
-    }
-#endif
-}
-
-
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_dot(vec_t<2, Type> a, vec_t<2, Type> b, vec_t<2, Type>& adj_a, vec_t<2, Type>& adj_b, const Type adj_ret)
-{
-    adj_a.c[0] += b.c[0]*adj_ret;
-    adj_a.c[1] += b.c[1]*adj_ret;
-
-    adj_b.c[0] += a.c[0]*adj_ret;
-    adj_b.c[1] += a.c[1]*adj_ret;
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_dot(vec_t<3, Type> a, vec_t<3, Type> b, vec_t<3, Type>& adj_a, vec_t<3, Type>& adj_b, const Type adj_ret)
-{
-    adj_a.c[0] += b.c[0]*adj_ret;
-    adj_a.c[1] += b.c[1]*adj_ret;
-    adj_a.c[2] += b.c[2]*adj_ret;
-
-    adj_b.c[0] += a.c[0]*adj_ret;
-    adj_b.c[1] += a.c[1]*adj_ret;
-    adj_b.c[2] += a.c[2]*adj_ret;
-}
-
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_extract(const vec_t<Length, Type> & a, int idx, vec_t<Length, Type> & adj_a, int & adj_idx, Type & adj_ret)
-{
-#ifndef NDEBUG
-    if (idx < 0 || idx > Length)
-    {
-        printf("Tvec2<Scalar> index %d out of bounds at %s %d\n", idx, __FILE__, __LINE__);
-        assert(0);
-    }
-#endif
-
-    adj_a[idx] += adj_ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_length(vec_t<Length, Type> a, Type ret, vec_t<Length, Type>& adj_a, const Type adj_ret)
-{
-    if (ret > Type(kEps))
-    {
-        adj_a += div(a, ret) * adj_ret;
-    }
-
-#if FP_CHECK
-    if (!isfinite(adj_a))
-    {
-        // \TODO: How shall we implement this error message?
-        //printf("%s:%d - adj_length((%f %f %f %f), (%f %f %f %f), (%f))\n", __FILE__, __LINE__, a.x, a.y, a.z, a.w, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_ret);
-        assert(0);
-    }
-#endif
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_length_sq(vec_t<Length, Type> a, vec_t<Length, Type>& adj_a, const Type adj_ret)
-{
-    adj_a += Type(2.0)*a*adj_ret;
-
-#if FP_CHECK
-    if (!isfinite(adj_a))
-    {
-        // \TODO: How shall we implement this error message?
-        //printf("%s:%d - adj_length((%f %f %f %f), (%f %f %f %f), (%f))\n", __FILE__, __LINE__, a.x, a.y, a.z, a.w, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_ret);
-        assert(0);
-    }
-#endif
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_normalize(vec_t<Length, Type> a, vec_t<Length, Type>& ret, vec_t<Length, Type>& adj_a, const vec_t<Length, Type>& adj_ret)
-{
-    Type d = length(a);
-    
-    if (d > Type(kEps))
-    {
-        Type invd = Type(1.0f)/d;
-
-        adj_a += (adj_ret*invd - ret*(dot(ret, adj_ret))*invd);
-
-#if FP_CHECK
-        if (!isfinite(adj_a))
-        {
-            // \TODO: How shall we implement this error message?
-            //printf("%s:%d - adj_normalize((%f %f %f %f), (%f %f %f %f), (%f, %f, %f, %f))\n", __FILE__, __LINE__, a.x, a.y, a.z, a.w, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_ret.x, adj_ret.y, adj_ret.z, adj_ret.w);
-            assert(0);
-        }
-#endif
-    }
-}
-
-template<typename Type>
-inline CUDA_CALLABLE void adj_cross(vec_t<3,Type> a, vec_t<3,Type> b, vec_t<3,Type>& adj_a, vec_t<3,Type>& adj_b, const vec_t<3,Type>& adj_ret)
-{
-    // todo: sign check
-    adj_a += cross(b, adj_ret);
-    adj_b -= cross(a, adj_ret);
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_min(const vec_t<Length,Type> &a, const vec_t<Length,Type> &b, vec_t<Length,Type>& adj_a, vec_t<Length,Type>& adj_b, const vec_t<Length,Type> &adj_ret)
-{
-    for( unsigned i=0; i < Length; ++i )
-    {
-        if (a[i] < b[i])
-            adj_a[i] += adj_ret[i];
-        else
-            adj_b[i] += adj_ret[i];
-    }
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_max(const vec_t<Length,Type> &a, const vec_t<Length,Type> &b, vec_t<Length,Type>& adj_a, vec_t<Length,Type>& adj_b, const vec_t<Length,Type> &adj_ret)
-{
-    for( unsigned i=0; i < Length; ++i )
-    {
-        if (a[i] > b[i])
-            adj_a[i] += adj_ret[i];
-        else
-            adj_b[i] += adj_ret[i];
-    }
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_min(const vec_t<Length,Type> &v, vec_t<Length,Type>& adj_v, const Type &adj_ret)
-{
-    unsigned i = argmin(v);
-    adj_v[i] += adj_ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_max(const vec_t<Length,Type> &v, vec_t<Length,Type>& adj_v, const Type &adj_ret)
-{
-    unsigned i = argmax(v);
-    adj_v[i] += adj_ret;
-}
-
-// Do I need to specialize these for different lengths?
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> atomic_add(vec_t<Length, Type> * addr, vec_t<Length, Type> value)
-{
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = atomic_add(&(addr -> c[i]), value[i]);
-    }
-
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> atomic_min(vec_t<Length, Type> * addr, vec_t<Length, Type> value)
-{
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = atomic_min(&(addr -> c[i]), value[i]);
-    }
-
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE vec_t<Length, Type> atomic_max(vec_t<Length, Type> * addr, vec_t<Length, Type> value)
-{
-    vec_t<Length, Type> ret;
-    for( unsigned i=0; i < Length; ++i )
-    {
-        ret[i] = atomic_max(&(addr -> c[i]), value[i]);
-    }
-
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-inline CUDA_CALLABLE void adj_atomic_minmax(
-    vec_t<Length,Type> *addr,
-    vec_t<Length,Type> *adj_addr,
-    const vec_t<Length,Type> &value,
-    vec_t<Length,Type> &adj_value)
-{
-    for (unsigned i=0; i < Length; ++i)
-        adj_atomic_minmax(&(addr->c[i]), &(adj_addr->c[i]), value[i], adj_value[i]);
-}
-
-// ok, the original implementation of this didn't take the absolute values.
-// I wouldn't consider this expected behavior. It looks like it's only
-// being used for bounding boxes at the moment, where this doesn't matter,
-// but you often use it for ray tracing where it does. Not sure if the
-// fabs() incurs a performance hit...
-template<unsigned Length, typename Type>
-CUDA_CALLABLE inline int longest_axis(const vec_t<Length, Type>& v)
-{
-    Type lmax = abs(v[0]);
-    int ret(0);
-    for( unsigned i=1; i < Length; ++i )
-    {
-        Type l = abs(v[i]);
-        if( l > lmax )
-        {
-            ret = i;
-            lmax = l;
-        }
-    }
-    return ret;
-}
-
-template<unsigned Length, typename Type>
-CUDA_CALLABLE inline vec_t<Length,Type> lerp(const vec_t<Length,Type>& a, const vec_t<Length,Type>& b, Type t)
-{
-    return a*(Type(1)-t) + b*t;
-}
-
-template<unsigned Length, typename Type>
-CUDA_CALLABLE inline void adj_lerp(const vec_t<Length,Type>& a, const vec_t<Length,Type>& b, Type t, vec_t<Length,Type>& adj_a, vec_t<Length,Type>& adj_b, Type& adj_t, const vec_t<Length,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 Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, int8>* buf, const vec_t<Length, int8> &value) { }
-template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, uint8>* buf, const vec_t<Length, uint8> &value) { }
-template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, int16>* buf, const vec_t<Length, int16> &value) { }
-template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, uint16>* buf, const vec_t<Length, uint16> &value) { }
-template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, int32>* buf, const vec_t<Length, int32> &value) { }
-template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, uint32>* buf, const vec_t<Length, uint32> &value) { }
-template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, int64>* buf, const vec_t<Length, int64> &value) { }
-template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, uint64>* buf, const vec_t<Length, uint64> &value) { }
-
-
-// adjoints for some of the constructors, used in intersect.h
-inline CUDA_CALLABLE void adj_vec2(float x, float y, float& adj_x, float& adj_y, const vec2& adj_ret)
-{
-    adj_x += adj_ret[0];
-    adj_y += adj_ret[1];
-}
-
-inline CUDA_CALLABLE void adj_vec3(float x, float y, float z, float& adj_x, float& adj_y, float& adj_z, const vec3& adj_ret)
-{
-    adj_x += adj_ret[0];
-    adj_y += adj_ret[1];
-    adj_z += adj_ret[2];
-}
-
-inline CUDA_CALLABLE void adj_vec4(float x, float y, float z, float w, float& adj_x, float& adj_y, float& adj_z, float& adj_w, const vec4& adj_ret)
-{
-    adj_x += adj_ret[0];
-    adj_y += adj_ret[1];
-    adj_z += adj_ret[2];
-    adj_w += adj_ret[3];
-}
-
-inline CUDA_CALLABLE void adj_vec3(float s, float& adj_s, const vec3& adj_ret)
-{
-    adj_vec_t(s, adj_s, adj_ret);
-}
-
-inline CUDA_CALLABLE void adj_vec4(float s, float& adj_s, const vec4& adj_ret)
-{
-    adj_vec_t(s, adj_s, adj_ret);
-}
-
-
-} // 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
+{
+
+template<unsigned Length, typename Type>
+struct vec_t
+{
+    Type c[Length];
+
+    inline CUDA_CALLABLE vec_t()
+        : c()
+    {}
+    
+    inline CUDA_CALLABLE vec_t(Type s)
+    {
+        for( unsigned i=0; i < Length; ++i )
+        {
+            c[i] = s;
+        }
+    }
+    
+    template <typename OtherType>
+    inline explicit CUDA_CALLABLE vec_t(const vec_t<Length, OtherType>& other)
+    {
+        for( unsigned i=0; i < Length; ++i )
+        {
+            c[i] = static_cast<Type>(other[i]);
+        }
+    }
+
+    inline CUDA_CALLABLE vec_t(Type x, Type y)
+    {
+        assert(Length == 2);
+        c[0]=x;
+        c[1]=y;
+    }
+
+    inline CUDA_CALLABLE vec_t(Type x, Type y, Type z)
+    {
+        assert(Length == 3);
+        c[0]=x;
+        c[1]=y;
+        c[2]=z;
+    }
+
+
+    inline CUDA_CALLABLE vec_t(Type x, Type y, Type z, Type w)
+    {
+        assert(Length == 4);
+        c[0]=x;
+        c[1]=y;
+        c[2]=z;
+        c[3]=w;
+    }
+    
+    inline CUDA_CALLABLE vec_t(const initializer_array<Length, Type> &l)
+    {
+        for( unsigned i=0; i < Length; ++i )
+        {
+            c[i] = l[i];
+        }
+    }
+    
+    // special screw vector constructor for spatial_vectors:
+    inline CUDA_CALLABLE vec_t(vec_t<3,Type> w, vec_t<3,Type> v)
+    {
+        c[0] = w[0];
+        c[1] = w[1];
+        c[2] = w[2];
+        c[3] = v[0];
+        c[4] = v[1];
+        c[5] = v[2];
+    }
+
+    inline CUDA_CALLABLE Type operator[](int index) const
+    {
+        assert(index < Length);
+        return c[index];
+    }
+
+    inline CUDA_CALLABLE Type& operator[](int index)
+    {
+        assert(index < Length);
+        return c[index];
+    }
+};
+
+using vec2b = vec_t<2,int8>;
+using vec3b = vec_t<3,int8>;
+using vec4b = vec_t<4,int8>;
+using vec2ub = vec_t<2,uint8>;
+using vec3ub = vec_t<3,uint8>;
+using vec4ub = vec_t<4,uint8>;
+
+using vec2s = vec_t<2,int16>;
+using vec3s = vec_t<3,int16>;
+using vec4s = vec_t<4,int16>;
+using vec2us = vec_t<2,uint16>;
+using vec3us = vec_t<3,uint16>;
+using vec4us = vec_t<4,uint16>;
+
+using vec2i = vec_t<2,int32>;
+using vec3i = vec_t<3,int32>;
+using vec4i = vec_t<4,int32>;
+using vec2ui = vec_t<2,uint32>;
+using vec3ui = vec_t<3,uint32>;
+using vec4ui = vec_t<4,uint32>;
+
+using vec2l = vec_t<2,int64>;
+using vec3l = vec_t<3,int64>;
+using vec4l = vec_t<4,int64>;
+using vec2ul = vec_t<2,uint64>;
+using vec3ul = vec_t<3,uint64>;
+using vec4ul = vec_t<4,uint64>;
+
+using vec2h = vec_t<2,half>;
+using vec3h = vec_t<3,half>;
+using vec4h = vec_t<4,half>;
+
+using vec2 = vec_t<2,float>;
+using vec3 = vec_t<3,float>;
+using vec4 = vec_t<4,float>;
+
+using vec2f = vec_t<2,float>;
+using vec3f = vec_t<3,float>;
+using vec4f = vec_t<4,float>;
+
+using vec2d = vec_t<2,double>;
+using vec3d = vec_t<3,double>;
+using vec4d = vec_t<4,double>;
+
+//--------------
+// vec<Length, Type> methods
+
+// Should these accept const references as arguments? It's all
+// inlined so maybe it doesn't matter? Even if it does, it
+// probably depends on the Length of the vector...
+
+// negation:
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> operator - (vec_t<Length, Type> a)
+{
+    // NB: this constructor will initialize all ret's components to 0, which is
+    // unnecessary... 
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = -a[i];
+    }
+
+    // Wonder if this does a load of copying when it returns... hopefully not as it's inlined?
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+CUDA_CALLABLE inline vec_t<Length, Type> pos(const vec_t<Length, Type>& x)
+{
+    return x;
+}
+
+template<unsigned Length, typename Type>
+CUDA_CALLABLE inline vec_t<Length, Type> neg(const vec_t<Length, Type>& x)
+{
+    return -x;
+}
+
+template<typename Type>
+CUDA_CALLABLE inline vec_t<3, Type> neg(const vec_t<3, Type>& x)
+{
+    return vec_t<3, Type>(-x.c[0], -x.c[1], -x.c[2]);
+}
+
+template<typename Type>
+CUDA_CALLABLE inline vec_t<2, Type> neg(const vec_t<2, Type>& x)
+{
+    return vec_t<2, Type>(-x.c[0], -x.c[1]);
+}
+
+template<unsigned Length, typename Type>
+CUDA_CALLABLE inline void adj_neg(const vec_t<Length, Type>& x, vec_t<Length, Type>& adj_x, const vec_t<Length, Type>& adj_ret)
+{
+    adj_x -= adj_ret;
+}
+
+// equality:
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE bool operator ==(const vec_t<Length, Type>& a, const vec_t<Length, Type>& b)
+{
+    for( unsigned i=0; i < Length; ++i )
+    {
+        if(a[i] != b[i])
+        {
+            return false;
+        }
+    }
+    return true;
+}
+
+// scalar multiplication:
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> mul(vec_t<Length, Type> a, Type s)
+{
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = a[i] * s;
+    }
+    return ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<3, Type> mul(vec_t<3, Type> a, Type s)
+{
+    return vec_t<3, Type>(a.c[0]*s,a.c[1]*s,a.c[2]*s);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<2, Type> mul(vec_t<2, Type> a, Type s)
+{
+    return vec_t<2, Type>(a.c[0]*s,a.c[1]*s);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> mul(Type s, vec_t<Length, Type> a)
+{
+    return mul(a, s);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> operator*(Type s, vec_t<Length, Type> a)
+{
+    return mul(a, s);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> operator*(vec_t<Length, Type> a, Type s)
+{
+    return mul(a, s);
+}
+
+
+// component wise multiplication:
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> cw_mul(vec_t<Length, Type> a, vec_t<Length, Type> b)
+{
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = a[i] * b[i];
+    }
+    return ret;
+}
+
+// division
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> div(vec_t<Length, Type> a, Type s)
+{
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = a[i] / s;
+    }
+    return ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<3, Type> div(vec_t<3, Type> a, Type s)
+{
+    return vec_t<3, Type>(a.c[0]/s,a.c[1]/s,a.c[2]/s);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<2, Type> div(vec_t<2, Type> a, Type s)
+{
+    return vec_t<2, Type>(a.c[0]/s,a.c[1]/s);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> div(Type s, vec_t<Length, Type> a)
+{
+    vec_t<Length, Type> ret;
+    for (unsigned i=0; i < Length; ++i)
+    {
+        ret[i] = s / a[i];
+    }
+    return ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<3, Type> div(Type s, vec_t<3, Type> a)
+{
+    return vec_t<3, Type>(s/a.c[0],s/a.c[1],s/a.c[2]);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<2, Type> div(Type s, vec_t<2, Type> a)
+{
+    return vec_t<2, Type>(s/a.c[0],s/a.c[1]);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> operator / (vec_t<Length, Type> a, Type s)
+{
+    return div(a,s);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> operator / (Type s, vec_t<Length, Type> a)
+{
+    return div(s, a);
+}
+
+// component wise division
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> cw_div(vec_t<Length, Type> a, vec_t<Length, Type> b)
+{
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = a[i] / b[i];
+    }
+    return ret;
+}
+
+// addition
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> add(vec_t<Length, Type> a, vec_t<Length, Type> b)
+{
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = a[i] + b[i];
+    }
+    return ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<2, Type> add(vec_t<2, Type> a, vec_t<2, Type> b)
+{
+    return vec_t<2, Type>( a.c[0] + b.c[0], a.c[1] + b.c[1]);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<3, Type> add(vec_t<3, Type> a, vec_t<3, Type> b)
+{
+    return vec_t<3, Type>( a.c[0] + b.c[0], a.c[1] + b.c[1], a.c[2] + b.c[2]);
+}
+
+// subtraction
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> sub(vec_t<Length, Type> a, vec_t<Length, Type> b)
+{
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = Type(a[i] - b[i]);
+    }
+    return ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<2, Type> sub(vec_t<2, Type> a, vec_t<2, Type> b)
+{
+    return vec_t<2, Type>( a.c[0] - b.c[0], a.c[1] - b.c[1]);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<3, Type> sub(vec_t<3, Type> a, vec_t<3, Type> b)
+{
+    return vec_t<3, Type>( a.c[0] - b.c[0], a.c[1] - b.c[1], a.c[2] - b.c[2]);
+}
+
+// dot product:
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE Type dot(vec_t<Length, Type> a, vec_t<Length, Type> b)
+{
+    Type ret(0);
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret += a[i] * b[i];
+    }
+    return ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE Type dot(vec_t<2, Type> a, vec_t<2, Type> b)
+{
+    return a.c[0] * b.c[0] + a.c[1] * b.c[1];
+}
+
+template<typename Type>
+inline CUDA_CALLABLE Type dot(vec_t<3, Type> a, vec_t<3, Type> b)
+{
+    return a.c[0] * b.c[0] + a.c[1] * b.c[1] + a.c[2] * b.c[2];
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE Type tensordot(vec_t<Length, Type> a, vec_t<Length, Type> b)
+{
+    // corresponds to `np.tensordot()` with all axes being contracted
+    return dot(a, b);
+}
+
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE Type extract(const vec_t<Length, Type> & a, int idx)
+{
+#ifndef NDEBUG
+    if (idx < 0 || idx >= Length)
+    {
+        printf("vec index %d out of bounds at %s %d\n", idx, __FILE__, __LINE__);
+        assert(0);
+    }
+#endif
+
+    return a[idx];        
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE Type* index(vec_t<Length, Type>& v, int idx)
+{
+#ifndef NDEBUG
+    if (idx < 0 || idx >= Length)
+    {
+        printf("vec index %d out of bounds at %s %d\n", idx, __FILE__, __LINE__);
+        assert(0);
+    }
+#endif
+
+    return &v[idx];
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE Type* indexref(vec_t<Length, Type>* v, int idx)
+{
+#ifndef NDEBUG
+    if (idx < 0 || idx >= Length)
+    {
+        printf("vec store %d out of bounds at %s %d\n", idx, __FILE__, __LINE__);
+        assert(0);
+    }
+#endif
+
+    return &((*v)[idx]);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_index(vec_t<Length, Type>& v, int idx,
+                                       vec_t<Length, Type>& adj_v, int adj_idx, const Type& adj_value)
+{
+    // nop
+}
+
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_indexref(vec_t<Length, Type>* v, int idx, 
+                                       vec_t<Length, Type>& adj_v, int adj_idx, const Type& adj_value)
+{
+    // nop
+}
+
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE Type length(vec_t<Length, Type> a)
+{
+    return sqrt(dot(a, a));
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE Type length_sq(vec_t<Length, Type> a)
+{
+    return dot(a, a);
+}
+
+
+template<typename Type>
+inline CUDA_CALLABLE Type length(vec_t<2, Type> a)
+{
+    return sqrt(a.c[0] * a.c[0] + a.c[1] * a.c[1]);
+}
+
+template<typename Type>
+inline CUDA_CALLABLE Type length(vec_t<3, Type> a)
+{
+    return sqrt(a.c[0] * a.c[0] + a.c[1] * a.c[1] + a.c[2] * a.c[2]);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> normalize(vec_t<Length, Type> a)
+{
+    Type l = length(a);
+    if (l > Type(kEps))
+        return div(a,l);
+    else
+        return vec_t<Length, Type>();
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<2, Type> normalize(vec_t<2, Type> a)
+{
+    Type l = sqrt(a.c[0] * a.c[0] + a.c[1] * a.c[1]);
+    if (l > Type(kEps))
+        return vec_t<2, Type>(a.c[0]/l,a.c[1]/l);
+    else
+        return vec_t<2, Type>();
+}
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<3, Type> normalize(vec_t<3, Type> a)
+{
+    Type l = sqrt(a.c[0] * a.c[0] + a.c[1] * a.c[1] + a.c[2] * a.c[2]);
+    if (l > Type(kEps))
+        return vec_t<3, Type>(a.c[0]/l,a.c[1]/l,a.c[2]/l);
+    else
+        return vec_t<3, Type>();
+}
+
+
+template<typename Type>
+inline CUDA_CALLABLE vec_t<3,Type> cross(vec_t<3,Type> a, vec_t<3,Type> b)
+{
+    return {
+        Type(a[1]*b[2] - a[2]*b[1]),
+        Type(a[2]*b[0] - a[0]*b[2]),
+        Type(a[0]*b[1] - a[1]*b[0])
+    };
+}
+
+
+template<unsigned Length, typename Type>
+inline bool CUDA_CALLABLE isfinite(vec_t<Length, Type> x)
+{
+    for( unsigned i=0; i < Length; ++i )
+    {
+        if(!isfinite(x[i]))
+        {
+            return false;
+        }
+    }
+    return true;
+}
+
+template<unsigned Length, typename Type>
+inline bool CUDA_CALLABLE isnan(vec_t<Length, Type> x)
+{
+    for( unsigned i=0; i < Length; ++i )
+    {
+        if(isnan(x[i]))
+        {
+            return true;
+        }
+    }
+    return false;
+}
+
+template<unsigned Length, typename Type>
+inline bool CUDA_CALLABLE isinf(vec_t<Length, Type> x)
+{
+    for( unsigned i=0; i < Length; ++i )
+    {
+        if(isinf(x[i]))
+        {
+            return true;
+        }
+    }
+    return false;
+}
+
+// These two functions seem to compile very slowly
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length,Type> min(vec_t<Length,Type> a, vec_t<Length,Type> b)
+{
+    vec_t<Length,Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = a[i] < b[i] ? a[i] : b[i];
+    }
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length,Type> max(vec_t<Length,Type> a, vec_t<Length,Type> b)
+{
+    vec_t<Length,Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = a[i] > b[i] ? a[i] : b[i];
+    }
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE Type min(vec_t<Length,Type> v)
+{
+    Type ret = v[0];
+    for( unsigned i=1; i < Length; ++i )
+    {
+        if (v[i] < ret)
+            ret = v[i];
+    }
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE Type max(vec_t<Length,Type> v)
+{
+    Type ret = v[0];
+    for( unsigned i=1; i < Length; ++i )
+    {
+        if (v[i] > ret)
+            ret = v[i];
+    }
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE unsigned argmin(vec_t<Length,Type> v)
+{
+    unsigned ret = 0;
+    for( unsigned i=1; i < Length; ++i )
+    {
+        if (v[i] < v[ret])
+            ret = i;
+    }
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE unsigned argmax(vec_t<Length,Type> v)
+{
+    unsigned ret = 0;
+    for( unsigned i=1; i < Length; ++i )
+    {
+        if (v[i] > v[ret])
+            ret = i;
+    }
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void expect_near(const vec_t<Length, Type>& actual, const vec_t<Length, Type>& expected, const Type& tolerance)
+{
+    const Type diff(0);
+    for(size_t i=0; i<Length; ++i)
+    {
+        diff = max(diff,abs(actual[i] - expected[i]));
+    }
+    if (diff > tolerance)
+    {
+        printf("Error, expect_near() failed with tolerance "); print(tolerance);
+        printf("\t Expected: "); print(expected); 
+        printf("\t Actual: "); print(actual);
+    }
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_expect_near(const vec_t<Length, Type>& actual, const vec_t<Length, Type>& expected, Type tolerance, vec_t<Length, Type>& adj_actual, vec_t<Length, Type>& adj_expected, Type adj_tolerance)
+{
+    // nop
+}
+
+// adjoint for the initializer_array constructor:
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_vec_t(const initializer_array<Length, Type> &cmps, const initializer_array<Length, Type*> &adj_cmps, const vec_t<Length, Type>& adj_ret)
+{
+    for(unsigned i=0; i < Length; ++i)
+    {
+        *(adj_cmps[i]) += adj_ret[i];
+    }
+}
+
+
+// adjoint for the component constructors:
+template<typename Type>
+inline CUDA_CALLABLE void adj_vec_t(Type cmpx, Type cmpy, Type &adj_cmpx, Type &adj_cmpy, const vec_t<2, Type>& adj_ret)
+{
+    adj_cmpx += adj_ret.c[0];
+    adj_cmpy += adj_ret.c[1];
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_vec_t(Type cmpx, Type cmpy, Type cmpz, Type &adj_cmpx, Type &adj_cmpy, Type &adj_cmpz, const vec_t<3, Type>& adj_ret)
+{
+    adj_cmpx += adj_ret.c[0];
+    adj_cmpy += adj_ret.c[1];
+    adj_cmpz += adj_ret.c[2];
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_vec_t(Type cmpx, Type cmpy, Type cmpz, Type cmpw, Type &adj_cmpx, Type &adj_cmpy, Type &adj_cmpz, Type &adj_cmpw, const vec_t<4, Type>& adj_ret)
+{
+    adj_cmpx += adj_ret.c[0];
+    adj_cmpy += adj_ret.c[1];
+    adj_cmpz += adj_ret.c[2];
+    adj_cmpw += adj_ret.c[3];
+}
+
+// adjoint for the constant constructor:
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_vec_t(Type s, Type& adj_s, const vec_t<Length, Type>& adj_ret)
+{
+    for( unsigned i=0; i < Length; ++i )
+    {
+        adj_s += adj_ret[i];
+    }
+}
+
+// adjoint for the casting constructor
+template<unsigned Length, typename Type, typename OtherType>
+inline CUDA_CALLABLE void adj_vec_t(const vec_t<Length, OtherType>& other, vec_t<Length, OtherType>& adj_other, const vec_t<Length, Type>& adj_ret)
+{
+    for( unsigned i=0; i < Length; ++i )
+    {
+        adj_other[i] += static_cast<OtherType>(adj_ret[i]);
+    }
+}
+
+template<typename Type>
+CUDA_CALLABLE inline void adj_vec_t(const vec_t<3,Type>& w, const vec_t<3,Type>& v, vec_t<3,Type>& adj_w, vec_t<3,Type>& adj_v, const vec_t<6,Type>& adj_ret)
+{
+    adj_w[0] += adj_ret[0];
+    adj_w[1] += adj_ret[1];
+    adj_w[2] += adj_ret[2];
+    adj_v[0] += adj_ret[3];
+    adj_v[1] += adj_ret[4];
+    adj_v[2] += adj_ret[5];
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_mul(vec_t<Length, Type> a, Type s, vec_t<Length, Type>& adj_a, Type& adj_s, const vec_t<Length, Type>& adj_ret)
+{
+    for( unsigned i=0; i < Length; ++i )
+    {
+        adj_a[i] += s*adj_ret[i];
+    }
+
+    adj_s += dot(a, adj_ret);
+
+#if FP_CHECK
+    if (!isfinite(a) || !isfinite(s) || !isfinite(adj_a) || !isfinite(adj_s) || !isfinite(adj_ret))
+    {
+        // \TODO: How shall we implement this error message?
+        //printf("adj_mul((%f %f %f %f), %f, (%f %f %f %f), %f, (%f %f %f %f)\n", a.x, a.y, a.z, a.w, s, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_s, adj_ret.x, adj_ret.y, adj_ret.z, adj_ret.w);
+        assert(0);
+    }
+#endif
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_mul(Type s, vec_t<Length, Type> a, Type& adj_s, vec_t<Length, Type>& adj_a, const vec_t<Length, Type>& adj_ret)
+{
+    adj_mul(a, s, adj_a, adj_s, adj_ret);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_cw_mul(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const vec_t<Length, Type>& adj_ret)
+{
+  adj_a += cw_mul(b, adj_ret);
+  adj_b += cw_mul(a, adj_ret);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_div(vec_t<Length, Type> a, Type s, vec_t<Length, Type>& adj_a, Type& adj_s, const vec_t<Length, Type>& adj_ret)
+{
+
+    adj_s -= dot(a , adj_ret)/ (s * s); // - a / s^2
+
+    for( unsigned i=0; i < Length; ++i )
+    {
+        adj_a[i] += adj_ret[i] / s;
+    }
+
+#if FP_CHECK
+    if (!isfinite(a) || !isfinite(s) || !isfinite(adj_a) || !isfinite(adj_s) || !isfinite(adj_ret))
+    {
+        // \TODO: How shall we implement this error message?
+        // printf("adj_div((%f %f %f %f), %f, (%f %f %f %f), %f, (%f %f %f %f)\n", a.x, a.y, a.z, a.w, s, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_s, adj_ret.x, adj_ret.y, adj_ret.z, adj_ret.w);
+        assert(0);
+    }
+#endif
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_div(Type s, vec_t<Length, Type> a, Type& adj_s, vec_t<Length, Type>& adj_a, const vec_t<Length, Type>& adj_ret)
+{
+
+    adj_s -= dot(a , adj_ret)/ (s * s); // - a / s^2
+
+    for( unsigned i=0; i < Length; ++i )
+    {
+        adj_a[i] += s / adj_ret[i];
+    }
+
+#if FP_CHECK
+    if (!isfinite(a) || !isfinite(s) || !isfinite(adj_a) || !isfinite(adj_s) || !isfinite(adj_ret))
+    {
+        // \TODO: How shall we implement this error message?
+        // printf("adj_div((%f %f %f %f), %f, (%f %f %f %f), %f, (%f %f %f %f)\n", a.x, a.y, a.z, a.w, s, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_s, adj_ret.x, adj_ret.y, adj_ret.z, adj_ret.w);
+        assert(0);
+    }
+#endif
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_cw_div(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& ret, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const vec_t<Length, Type>& adj_ret) {
+  adj_a += cw_div(adj_ret, b);
+  adj_b -= cw_mul(adj_ret, cw_div(ret, b));
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_add(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const vec_t<Length, Type>& adj_ret)
+{
+    adj_a += adj_ret;
+    adj_b += adj_ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_add(vec_t<2, Type> a, vec_t<2, Type> b, vec_t<2, Type>& adj_a, vec_t<2, Type>& adj_b, const vec_t<2, Type>& adj_ret)
+{
+    adj_a.c[0] += adj_ret.c[0];
+    adj_a.c[1] += adj_ret.c[1];
+    adj_b.c[0] += adj_ret.c[0];
+    adj_b.c[1] += adj_ret.c[1];
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_add(vec_t<3, Type> a, vec_t<3, Type> b, vec_t<3, Type>& adj_a, vec_t<3, Type>& adj_b, const vec_t<3, Type>& adj_ret)
+{
+    adj_a.c[0] += adj_ret.c[0];
+    adj_a.c[1] += adj_ret.c[1];
+    adj_a.c[2] += adj_ret.c[2];
+    adj_b.c[0] += adj_ret.c[0];
+    adj_b.c[1] += adj_ret.c[1];
+    adj_b.c[2] += adj_ret.c[2];
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_sub(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const vec_t<Length, Type>& adj_ret)
+{
+    adj_a += adj_ret;
+    adj_b -= adj_ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_sub(vec_t<2, Type> a, vec_t<2, Type> b, vec_t<2, Type>& adj_a, vec_t<2, Type>& adj_b, const vec_t<2, Type>& adj_ret)
+{
+    adj_a.c[0] += adj_ret.c[0];
+    adj_a.c[1] += adj_ret.c[1];
+    adj_b.c[0] -= adj_ret.c[0];
+    adj_b.c[1] -= adj_ret.c[1];
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_sub(vec_t<3, Type> a, vec_t<3, Type> b, vec_t<3, Type>& adj_a, vec_t<3, Type>& adj_b, const vec_t<3, Type>& adj_ret)
+{
+    adj_a.c[0] += adj_ret.c[0];
+    adj_a.c[1] += adj_ret.c[1];
+    adj_a.c[2] += adj_ret.c[2];
+    adj_b.c[0] -= adj_ret.c[0];
+    adj_b.c[1] -= adj_ret.c[1];
+    adj_b.c[2] -= adj_ret.c[2];
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_dot(vec_t<Length, Type> a, vec_t<Length, Type> b, vec_t<Length, Type>& adj_a, vec_t<Length, Type>& adj_b, const Type adj_ret)
+{
+    adj_a += b*adj_ret;
+    adj_b += a*adj_ret;
+
+#if FP_CHECK
+    if (!isfinite(a) || !isfinite(b) || !isfinite(adj_a) || !isfinite(adj_b) || !isfinite(adj_ret))
+    {
+        // \TODO: How shall we implement this error message?
+        //printf("adj_dot((%f %f %f %f), (%f %f %f %f), (%f %f %f %f), (%f %f %f %f), %f)\n", a.x, a.y, a.z, a.w, b.x, b.y, b.z, b.w, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_b.x, adj_b.y, adj_b.z, adj_b.w, adj_ret);
+        assert(0);
+    }
+#endif
+}
+
+
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_dot(vec_t<2, Type> a, vec_t<2, Type> b, vec_t<2, Type>& adj_a, vec_t<2, Type>& adj_b, const Type adj_ret)
+{
+    adj_a.c[0] += b.c[0]*adj_ret;
+    adj_a.c[1] += b.c[1]*adj_ret;
+
+    adj_b.c[0] += a.c[0]*adj_ret;
+    adj_b.c[1] += a.c[1]*adj_ret;
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_dot(vec_t<3, Type> a, vec_t<3, Type> b, vec_t<3, Type>& adj_a, vec_t<3, Type>& adj_b, const Type adj_ret)
+{
+    adj_a.c[0] += b.c[0]*adj_ret;
+    adj_a.c[1] += b.c[1]*adj_ret;
+    adj_a.c[2] += b.c[2]*adj_ret;
+
+    adj_b.c[0] += a.c[0]*adj_ret;
+    adj_b.c[1] += a.c[1]*adj_ret;
+    adj_b.c[2] += a.c[2]*adj_ret;
+}
+
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_extract(const vec_t<Length, Type> & a, int idx, vec_t<Length, Type> & adj_a, int & adj_idx, Type & adj_ret)
+{
+#ifndef NDEBUG
+    if (idx < 0 || idx > Length)
+    {
+        printf("Tvec2<Scalar> index %d out of bounds at %s %d\n", idx, __FILE__, __LINE__);
+        assert(0);
+    }
+#endif
+
+    adj_a[idx] += adj_ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_length(vec_t<Length, Type> a, Type ret, vec_t<Length, Type>& adj_a, const Type adj_ret)
+{
+    if (ret > Type(kEps))
+    {
+        adj_a += div(a, ret) * adj_ret;
+    }
+
+#if FP_CHECK
+    if (!isfinite(adj_a))
+    {
+        // \TODO: How shall we implement this error message?
+        //printf("%s:%d - adj_length((%f %f %f %f), (%f %f %f %f), (%f))\n", __FILE__, __LINE__, a.x, a.y, a.z, a.w, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_ret);
+        assert(0);
+    }
+#endif
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_length_sq(vec_t<Length, Type> a, vec_t<Length, Type>& adj_a, const Type adj_ret)
+{
+    adj_a += Type(2.0)*a*adj_ret;
+
+#if FP_CHECK
+    if (!isfinite(adj_a))
+    {
+        // \TODO: How shall we implement this error message?
+        //printf("%s:%d - adj_length((%f %f %f %f), (%f %f %f %f), (%f))\n", __FILE__, __LINE__, a.x, a.y, a.z, a.w, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_ret);
+        assert(0);
+    }
+#endif
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_normalize(vec_t<Length, Type> a, vec_t<Length, Type>& ret, vec_t<Length, Type>& adj_a, const vec_t<Length, Type>& adj_ret)
+{
+    Type d = length(a);
+    
+    if (d > Type(kEps))
+    {
+        Type invd = Type(1.0f)/d;
+
+        adj_a += (adj_ret*invd - ret*(dot(ret, adj_ret))*invd);
+
+#if FP_CHECK
+        if (!isfinite(adj_a))
+        {
+            // \TODO: How shall we implement this error message?
+            //printf("%s:%d - adj_normalize((%f %f %f %f), (%f %f %f %f), (%f, %f, %f, %f))\n", __FILE__, __LINE__, a.x, a.y, a.z, a.w, adj_a.x, adj_a.y, adj_a.z, adj_a.w, adj_ret.x, adj_ret.y, adj_ret.z, adj_ret.w);
+            assert(0);
+        }
+#endif
+    }
+}
+
+template<typename Type>
+inline CUDA_CALLABLE void adj_cross(vec_t<3,Type> a, vec_t<3,Type> b, vec_t<3,Type>& adj_a, vec_t<3,Type>& adj_b, const vec_t<3,Type>& adj_ret)
+{
+    // todo: sign check
+    adj_a += cross(b, adj_ret);
+    adj_b -= cross(a, adj_ret);
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_isfinite(const vec_t<Length, Type> &x, vec_t<Length,Type>& adj_x, const bool &adj_ret)
+{
+
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_isnan(const vec_t<Length, Type> &x, vec_t<Length,Type>& adj_x, const bool &adj_ret)
+{
+
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_isinf(const vec_t<Length, Type> &x, vec_t<Length,Type>& adj_x, const bool &adj_ret)
+{
+
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_min(const vec_t<Length,Type> &a, const vec_t<Length,Type> &b, vec_t<Length,Type>& adj_a, vec_t<Length,Type>& adj_b, const vec_t<Length,Type> &adj_ret)
+{
+    for( unsigned i=0; i < Length; ++i )
+    {
+        if (a[i] < b[i])
+            adj_a[i] += adj_ret[i];
+        else
+            adj_b[i] += adj_ret[i];
+    }
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_max(const vec_t<Length,Type> &a, const vec_t<Length,Type> &b, vec_t<Length,Type>& adj_a, vec_t<Length,Type>& adj_b, const vec_t<Length,Type> &adj_ret)
+{
+    for( unsigned i=0; i < Length; ++i )
+    {
+        if (a[i] > b[i])
+            adj_a[i] += adj_ret[i];
+        else
+            adj_b[i] += adj_ret[i];
+    }
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_min(const vec_t<Length,Type> &v, vec_t<Length,Type>& adj_v, const Type &adj_ret)
+{
+    unsigned i = argmin(v);
+    adj_v[i] += adj_ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_max(const vec_t<Length,Type> &v, vec_t<Length,Type>& adj_v, const Type &adj_ret)
+{
+    unsigned i = argmax(v);
+    adj_v[i] += adj_ret;
+}
+
+// Do I need to specialize these for different lengths?
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> atomic_add(vec_t<Length, Type> * addr, vec_t<Length, Type> value)
+{
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = atomic_add(&(addr -> c[i]), value[i]);
+    }
+
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> atomic_min(vec_t<Length, Type> * addr, vec_t<Length, Type> value)
+{
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = atomic_min(&(addr -> c[i]), value[i]);
+    }
+
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE vec_t<Length, Type> atomic_max(vec_t<Length, Type> * addr, vec_t<Length, Type> value)
+{
+    vec_t<Length, Type> ret;
+    for( unsigned i=0; i < Length; ++i )
+    {
+        ret[i] = atomic_max(&(addr -> c[i]), value[i]);
+    }
+
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+inline CUDA_CALLABLE void adj_atomic_minmax(
+    vec_t<Length,Type> *addr,
+    vec_t<Length,Type> *adj_addr,
+    const vec_t<Length,Type> &value,
+    vec_t<Length,Type> &adj_value)
+{
+    for (unsigned i=0; i < Length; ++i)
+        adj_atomic_minmax(&(addr->c[i]), &(adj_addr->c[i]), value[i], adj_value[i]);
+}
+
+// ok, the original implementation of this didn't take the absolute values.
+// I wouldn't consider this expected behavior. It looks like it's only
+// being used for bounding boxes at the moment, where this doesn't matter,
+// but you often use it for ray tracing where it does. Not sure if the
+// fabs() incurs a performance hit...
+template<unsigned Length, typename Type>
+CUDA_CALLABLE inline int longest_axis(const vec_t<Length, Type>& v)
+{
+    Type lmax = abs(v[0]);
+    int ret(0);
+    for( unsigned i=1; i < Length; ++i )
+    {
+        Type l = abs(v[i]);
+        if( l > lmax )
+        {
+            ret = i;
+            lmax = l;
+        }
+    }
+    return ret;
+}
+
+template<unsigned Length, typename Type>
+CUDA_CALLABLE inline vec_t<Length,Type> lerp(const vec_t<Length,Type>& a, const vec_t<Length,Type>& b, Type t)
+{
+    return a*(Type(1)-t) + b*t;
+}
+
+template<unsigned Length, typename Type>
+CUDA_CALLABLE inline void adj_lerp(const vec_t<Length,Type>& a, const vec_t<Length,Type>& b, Type t, vec_t<Length,Type>& adj_a, vec_t<Length,Type>& adj_b, Type& adj_t, const vec_t<Length,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 Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, int8>* buf, const vec_t<Length, int8> &value) { }
+template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, uint8>* buf, const vec_t<Length, uint8> &value) { }
+template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, int16>* buf, const vec_t<Length, int16> &value) { }
+template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, uint16>* buf, const vec_t<Length, uint16> &value) { }
+template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, int32>* buf, const vec_t<Length, int32> &value) { }
+template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, uint32>* buf, const vec_t<Length, uint32> &value) { }
+template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, int64>* buf, const vec_t<Length, int64> &value) { }
+template<unsigned Length> CUDA_CALLABLE inline void adj_atomic_add(vec_t<Length, uint64>* buf, const vec_t<Length, uint64> &value) { }
+
+
+// adjoints for some of the constructors, used in intersect.h
+inline CUDA_CALLABLE void adj_vec2(float x, float y, float& adj_x, float& adj_y, const vec2& adj_ret)
+{
+    adj_x += adj_ret[0];
+    adj_y += adj_ret[1];
+}
+
+inline CUDA_CALLABLE void adj_vec3(float x, float y, float z, float& adj_x, float& adj_y, float& adj_z, const vec3& adj_ret)
+{
+    adj_x += adj_ret[0];
+    adj_y += adj_ret[1];
+    adj_z += adj_ret[2];
+}
+
+inline CUDA_CALLABLE void adj_vec4(float x, float y, float z, float w, float& adj_x, float& adj_y, float& adj_z, float& adj_w, const vec4& adj_ret)
+{
+    adj_x += adj_ret[0];
+    adj_y += adj_ret[1];
+    adj_z += adj_ret[2];
+    adj_w += adj_ret[3];
+}
+
+inline CUDA_CALLABLE void adj_vec3(float s, float& adj_s, const vec3& adj_ret)
+{
+    adj_vec_t(s, adj_s, adj_ret);
+}
+
+inline CUDA_CALLABLE void adj_vec4(float s, float& adj_s, const vec4& adj_ret)
+{
+    adj_vec_t(s, adj_s, adj_ret);
+}
+
+
+} // namespace wp