inline void operator() (ustring name, Dual2<R> &result,
                         const Dual2<S> &s, const Dual2<T> &t,
                         ShaderGlobals *sg, const NoiseParams *opt) const {
     if (name == Strings::uperlin || name == Strings::noise) {
         Noise noise;
         noise(result, s, t);
     } else if (name == Strings::perlin || name == Strings::snoise) {
         SNoise snoise;
         snoise(result, s, t);
     } else if (name == Strings::simplexnoise || name == Strings::simplex) {
         SimplexNoise simplexnoise;
         simplexnoise(result, s, t);
     } else if (name == Strings::usimplexnoise || name == Strings::usimplex) {
         USimplexNoise usimplexnoise;
         usimplexnoise(result, s, t);
     } else if (name == Strings::cell) {
         CellNoise cellnoise;
         cellnoise(result.val(), s.val(), t.val());
         result.clear_d();
     } else if (name == Strings::gabor) {
         GaborNoise gnoise;
         gnoise (name, result, s, t, sg, opt);
     } else {
         ((ShadingContext *)sg->context)->error ("Unknown noise type \"%s\"", name.c_str());
     }
 }
 inline void operator() (ustring name, Dual2<R> &result,
                         const Dual2<S> &s, const Dual2<T> &t,
                         const S &sp, const T &tp,
                         ShaderGlobals *sg, const NoiseParams *opt) const {
     if (name == Strings::uperlin || name == Strings::noise) {
         PeriodicNoise noise;
         noise(result, s, t, sp, tp);
     } else if (name == Strings::perlin || name == Strings::snoise) {
         PeriodicSNoise snoise;
         snoise(result, s, t, sp, tp);
     } else if (name == Strings::cell) {
         PeriodicCellNoise cellnoise;
         cellnoise(result.val(), s.val(), t.val(), sp, tp);
         result.clear_d();
     } else if (name == Strings::gabor) {
         GaborPNoise gnoise;
         gnoise (name, result, s, t, sp, tp, sg, opt);
     } else if (name == Strings::hash) {
         PeriodicHashNoise hashnoise;
         hashnoise(result.val(), s.val(), t.val(), sp, tp);
         result.clear_d();
     } else {
         ((ShadingContext *)sg->context)->error ("Unknown noise type \"%s\"", name.c_str());
     }
 }
// set up the filter matrix
static void
gabor_setup_filter (const Dual2<Vec3> &P, GaborParams &gp)
{
    // Make texture-space normal, tangent, bitangent
    Vec3 n, t, b;
    n = P.dx().cross (P.dy());  // normal to P
    if (n.dot(n) < 1.0e-6f) {  /* length of deriv < 1/1000 */
        // No way to do filter if we have no derivs, and no reason to
        // do it if it's too small to have any effect.
        gp.do_filter = false;
        return;   // we won't need anything else if filtering is off
    }
    make_orthonormals (n, t, b);

    // Rotations from tangent<->texture space
    Matrix33 Mtex_to_tan = make_matrix33_cols (t, b, n);  // M3_local
    Matrix33 Mscreen_to_tex = make_matrix33_cols (P.dx(), P.dy(), Vec3(0.0f,0.0f,0.0f));
    Matrix33 Mscreen_to_tan = Mscreen_to_tex * Mtex_to_tan;  // M3_scr_tan
    Matrix22 M_scr_tan (Mscreen_to_tan[0][0], Mscreen_to_tan[0][1],
                        Mscreen_to_tan[1][0], Mscreen_to_tan[1][1]);
    float sigma_f_scr = 0.5f;
    Matrix22 Sigma_f_scr (sigma_f_scr * sigma_f_scr, 0.0f,
                          0.0f, sigma_f_scr * sigma_f_scr);
    Matrix22 M_scr_tan_t = M_scr_tan.transposed();
    Matrix22 Sigma_f_tan = M_scr_tan_t * Sigma_f_scr * M_scr_tan;

    gp.N = n;
    gp.filter = Sigma_f_tan;
    gp.det_filter = determinant(Sigma_f_tan);
    gp.local  = Mtex_to_tan;
    if (gp.det_filter < 1.0e-18f) {
        gp.do_filter = false;
        // Turn off filtering when tiny values will lead to numerical
        // errors later if we filter.  Yes, it's kind of arbitrary.
    }
}
inline Dual2<float> safe_fmod (const Dual2<float> &a, const Dual2<float> &b) {
    return Dual2<float> (safe_fmod (a.val(), b.val()), a.dx(), a.dy());
}
inline Dual2<float> fabsf (const Dual2<float> &x) {
    return x.val() >= 0 ? x : -x;
}
 OSL_HOSTDEVICE inline void operator() (Dual2<Vec3> &result, const Dual2<Vec3> &p, const Dual2<float> &t) const { result.set (v(), v(), v()); }
 OSL_HOSTDEVICE inline void operator() (Dual2<float> &result, const Dual2<Vec3> &p,
                                        const Dual2<float> &t, int seed=0) const { result.set (0.5f, 0.5f, 0.5f); }
 OSL_HOSTDEVICE inline void operator() (Dual2<float> &result, const Dual2<float> &x,
                                        const Dual2<float> &y, int seed=0) const { result.set (0.0f, 0.0f, 0.0f); }
// Helper function: per-component 'floor' of a Dual2<Vec3>.
inline Vec3
floor (const Dual2<Vec3> &vd)
{
    const Vec3 &v (vd.val());
    return Vec3 (floorf(v[0]), floorf(v[1]), floorf(v[2]));
}
// Evaluate the summed contribution of all gabor impulses within the
// cell whose corner is c_i.  x_c_i is vector from x (the point
// we are trying to evaluate noise at) and c_i.
Dual2<float>
gabor_cell (GaborParams &gp, const Vec3 &c_i, const Dual2<Vec3> &x_c_i,
            int seed = 0)
{
    fast_rng rng (gp.periodic ? Vec3(wrap(c_i,gp.period)) : c_i, seed);
    int n_impulses = rng.poisson (gp.lambda * gp.radius3);
    Dual2<float> sum = 0;

    for (int i = 0; i < n_impulses; i++) {
        // OLD code: Vec3 x_i_c (rng(), rng(), rng());
        // Turned out that C++ spec says order of args are unspecified.
        // gcc appeared to do right-to-left, so to make sure our noise
        // function is locked down (and works identically for clang,
        // which evaluates left-to-right), we ask for the rng() calls
        // one at a time and match the way it looked before.
        float z_rng = rng(), y_rng = rng(), x_rng = rng();
        Vec3 x_i_c (x_rng, y_rng, z_rng);
        Dual2<Vec3> x_k_i = gp.radius * (x_c_i - x_i_c);        
        float phi_i;
        Vec3 omega_i;
        gabor_sample (gp, c_i, rng, omega_i, phi_i);
        if (x_k_i.val().length2() < gp.radius2) {
            if (! gp.do_filter) {
                // N.B. if determinant(gp.filter) is too small, we will
                // run into numerical problems.  But the filtering isn't
                // needed in that case anyway, so just don't filter.
                // This seems to only come up when the filter region is
                // tiny.
                sum += gabor_kernel (gp.weight, omega_i, phi_i, gp.a, x_k_i);  // 3D
            } else {
                // Transform the impulse's anisotropy into tangent space
                Vec3 omega_i_t;
                multMatrix (gp.local, omega_i, omega_i_t);

                // Slice to get a 2D kernel
                Dual2<float> d_i = -dot(gp.N, x_k_i);
                Dual2<float> w_i_t_s;
                Vec2 omega_i_t_s;
                Dual2<float> phi_i_t_s;
                slice_gabor_kernel_3d (d_i, gp.weight, gp.a,
                                       omega_i_t, phi_i,
                                       w_i_t_s, omega_i_t_s, phi_i_t_s);

                // Filter the 2D kernel
                Dual2<float> w_i_t_s_f;
                float a_i_t_s_f;
                Vec2 omega_i_t_s_f;
                Dual2<float> phi_i_t_s_f;
                filter_gabor_kernel_2d (gp.filter, w_i_t_s, gp.a, omega_i_t_s, phi_i_t_s, w_i_t_s_f, a_i_t_s_f, omega_i_t_s_f, phi_i_t_s_f);

                // Now evaluate the 2D filtered kernel
                Dual2<Vec3> xkit;
                multMatrix (gp.local, x_k_i, xkit);
                Dual2<Vec2> x_k_i_t = make_Vec2 (comp(xkit,0), comp(xkit,1));
                Dual2<float> gk = gabor_kernel (w_i_t_s_f, omega_i_t_s_f, phi_i_t_s_f, a_i_t_s_f, x_k_i_t); // 2D
                if (! OIIO::isfinite(gk.val())) {
                    // Numeric failure of the filtered version.  Fall
                    // back on the unfiltered.
                    gk = gabor_kernel (gp.weight, omega_i, phi_i, gp.a, x_k_i);  // 3D
                }
                sum += gk;
            }
        }
    }

    return sum;
}