|
| 1 | +#ifndef UNARY_SVD_OP_H |
| 2 | +#define UNARY_SVD_OP_H |
| 3 | + |
| 4 | +#include "Fastor/meta/meta.h" |
| 5 | +#include "Fastor/backend/inner.h" |
| 6 | +#include "Fastor/backend/lufact.h" |
| 7 | +#include "Fastor/simd_vector/SIMDVector.h" |
| 8 | +#include "Fastor/tensor/AbstractTensor.h" |
| 9 | +#include "Fastor/tensor/Aliasing.h" |
| 10 | +#include "Fastor/tensor/Tensor.h" |
| 11 | +#include "Fastor/tensor/TensorTraits.h" |
| 12 | +#include "Fastor/expressions/expression_traits.h" |
| 13 | +#include "Fastor/expressions/linalg_ops/linalg_computation_types.h" |
| 14 | +#include "Fastor/expressions/linalg_ops/unary_det_op.h" |
| 15 | +#include "Fastor/expressions/linalg_ops/unary_trans_op.h" |
| 16 | + |
| 17 | + |
| 18 | +namespace Fastor { |
| 19 | + |
| 20 | +// SVD |
| 21 | +template<typename T, size_t M, enable_if_t_<M==2, bool> = false > |
| 22 | +FASTOR_INLINE void svd(const Tensor<T,M,M> &A, Tensor<T,M,M> &U, Tensor<T,M,M> &S, Tensor<T,M,M> &V) { |
| 23 | + |
| 24 | + constexpr T Epsilon_v = std::numeric_limits<T>::epsilon(); |
| 25 | + |
| 26 | + const T f00 = A(0, 0); |
| 27 | + const T f01 = A(0, 1); |
| 28 | + const T f10 = A(1, 0); |
| 29 | + const T f11 = A(1, 1); |
| 30 | + |
| 31 | + // If matrix is diagonal, SVD is trivial |
| 32 | + if (std::abs(f01 - f10) < Epsilon_v && std::abs(f01) < Epsilon_v) |
| 33 | + { |
| 34 | + // Compute U |
| 35 | + U(0,0) = f00 < 0 ? -1. : 1.; |
| 36 | + U(0,1) = 0.; |
| 37 | + U(1,0) = 0.; |
| 38 | + U(1,1) = f11 < 0. ? -1. : 1.; |
| 39 | + |
| 40 | + // Compute S |
| 41 | + S(0,0) = std::abs(f00); |
| 42 | + S(0,1) = 0; |
| 43 | + S(1,0) = 0; |
| 44 | + S(1,1) = std::abs(f11); |
| 45 | + |
| 46 | + // Compute V |
| 47 | + V.eye2(); |
| 48 | + } |
| 49 | + // Otherwise, we need to compute A^T*A |
| 50 | + else |
| 51 | + { |
| 52 | + T j = f00 * f00 + f01 * f01; |
| 53 | + T k = f10 * f10 + f11 * f11; |
| 54 | + T v_c = f00 * f10 + f01 * f11; |
| 55 | + // Check to see if A^T*A is diagonal |
| 56 | + if (std::abs(v_c) < Epsilon_v) |
| 57 | + { |
| 58 | + // Compute S |
| 59 | + T s1 = std::sqrt(j); |
| 60 | + T s2 = std::abs(j - k) < Epsilon_v ? s1 : std::sqrt(k); |
| 61 | + S(0,0) = s1; |
| 62 | + S(0,1) = 0; |
| 63 | + S(1,0) = 0; |
| 64 | + S(1,1) = s2; |
| 65 | + |
| 66 | + // Compute U |
| 67 | + U.eye2(); |
| 68 | + |
| 69 | + // Compute V |
| 70 | + V(0,0) = f00 / s1; |
| 71 | + V(0,1) = f10 / s2; |
| 72 | + V(1,0) = f01 / s1; |
| 73 | + V(1,1) = f11 / s2; |
| 74 | + } |
| 75 | + // Otherwise, solve quadratic equation for eigenvalues |
| 76 | + else |
| 77 | + { |
| 78 | + T jmk = j - k; |
| 79 | + T jpk = j + k; |
| 80 | + T root = std::sqrt(jmk * jmk + 4. * v_c * v_c); |
| 81 | + T eig1 = (jpk + root) * 0.5; |
| 82 | + T eig2 = (jpk - root) * 0.5; |
| 83 | + |
| 84 | + // Compute S |
| 85 | + T s1 = std::sqrt(eig1); |
| 86 | + T s2 = std::abs(root) < Epsilon_v ? s1 : ( eig2 > 0 ? std::sqrt(eig2) : Epsilon_v); |
| 87 | + S(0,0) = s1; |
| 88 | + S(0,1) = 0; |
| 89 | + S(1,0) = 0; |
| 90 | + S(1,1) = s2; |
| 91 | + |
| 92 | + // Compute U - use eigenvectors of A^T*A as U |
| 93 | + T v_s = eig1 - j; |
| 94 | + T len = std::max(std::sqrt(v_s * v_s + v_c * v_c), Epsilon_v); |
| 95 | + v_c /= len; |
| 96 | + v_s /= len; |
| 97 | + |
| 98 | + U(0,0) = v_c; |
| 99 | + U(0,1) = -v_s; |
| 100 | + U(1,0) = v_s; |
| 101 | + U(1,1) = v_c; |
| 102 | + |
| 103 | + // Compute V - as A * U / s |
| 104 | + const T cc = (f00 * v_c + f10 * v_s) / s1; |
| 105 | + const T cs = (f01 * v_c + f11 * v_s) / s1; |
| 106 | + if (std::abs(s2) > Epsilon_v) |
| 107 | + { |
| 108 | + V(0,0) = cc; |
| 109 | + V(0,1) = (f10* v_c - f00 * v_s) / s2; |
| 110 | + V(1,0) = cs; |
| 111 | + V(1,1) = (f11 * v_c - f01 * v_s) / s2; |
| 112 | + } |
| 113 | + else |
| 114 | + { |
| 115 | + V(0,0) = cc; |
| 116 | + V(0,1) = cs; |
| 117 | + V(1,0) = cs; |
| 118 | + V(1,1) = -cc; |
| 119 | + } |
| 120 | + } |
| 121 | + } |
| 122 | +} |
| 123 | + |
| 124 | + |
| 125 | + |
| 126 | +// Signed SVD |
| 127 | +template<typename T, size_t M> |
| 128 | +FASTOR_INLINE void ssvd(const Tensor<T,M,M> &A, Tensor<T,M,M> &U, Tensor<T,M,M> &S, Tensor<T,M,M> &V) { |
| 129 | + |
| 130 | + // Same as above but avoiding the L matrix |
| 131 | + svd(A, U, S, V); |
| 132 | + |
| 133 | + // See where to pull the reflection out of |
| 134 | + const T detU = determinant(U); |
| 135 | + const T detV = determinant(V); |
| 136 | + |
| 137 | + if (detU >= 0 && detV >= 0) |
| 138 | + { |
| 139 | + // No reflection svd == svd_rv, return |
| 140 | + return; |
| 141 | + } |
| 142 | + |
| 143 | + Tensor<T, M, M> L = matmul(U, transpose(V)); |
| 144 | + const T lastColumn = determinant(L); |
| 145 | + |
| 146 | + if (detU < 0 && detV > 0) |
| 147 | + { |
| 148 | + U(all, M - 1) *= lastColumn; |
| 149 | + } |
| 150 | + else if (detU > 0 && detV < 0) |
| 151 | + { |
| 152 | + V(all, M - 1) *= lastColumn; |
| 153 | + } |
| 154 | + |
| 155 | + // Push the reflection to the diagonal |
| 156 | + S(M - 1, M - 1) *= lastColumn; |
| 157 | +} |
| 158 | +//-----------------------------------------------------------------------------------------------------------// |
| 159 | + |
| 160 | +} // end of namespace Fastor |
| 161 | + |
| 162 | + |
| 163 | +#endif // UNARY_SVD_OP_H |
0 commit comments