xref: /honee/qfunctions/utils_eigensolver_jacobi.h (revision b78d7c7d152a2530c4ff7c4fb0143fe9be02cbec)
1 // SPDX-FileCopyrightText: Copyright (c) 2017-2024, HONEE contributors.
2 // SPDX-License-Identifier: Apache-2.0 OR BSD-2-Clause
3 
4 /// @file
5 /// Eigen system solver for symmetric NxN matrices. Modified from the CC0 code provided at https://github.com/jewettaij/jacobi_pd
6 #pragma once
7 
8 #include <ceed/types.h>
9 #ifndef CEED_RUNNING_JIT_PASS
10 #include <math.h>
11 #include <stdbool.h>
12 #endif
13 
14 #include "utils.h"
15 
16 // @typedef choose the criteria for sorting eigenvalues and eigenvectors
17 typedef enum eSortCriteria {
18   SORT_NONE,
19   SORT_DECREASING_EVALS,
20   SORT_INCREASING_EVALS,
21   SORT_DECREASING_ABS_EVALS,
22   SORT_INCREASING_ABS_EVALS
23 } SortCriteria;
24 
25 ///@brief Find the off-diagonal index in row i whose absolute value is largest
26 ///
27 /// @param[in] *A matrix
28 /// @param[in] i row index
29 /// @returns   Index of absolute largest off-diagonal element in row i
MaxEntryRow(const CeedScalar * A,CeedInt N,CeedInt i)30 CEED_QFUNCTION_HELPER CeedInt MaxEntryRow(const CeedScalar *A, CeedInt N, CeedInt i) {
31   CeedInt j_max = i + 1;
32   for (CeedInt j = i + 2; j < N; j++)
33     if (fabs(A[i * N + j]) > fabs(A[i * N + j_max])) j_max = j;
34   return j_max;
35 }
36 
37 /// @brief Find the indices (i_max, j_max) marking the location of the
38 ///        entry in the matrix with the largest absolute value.  This
39 ///        uses the max_idx_row[] array to find the answer in O(n) time.
40 ///
41 /// @param[in]    *A    matrix
42 /// @param[inout] i_max row index
43 /// @param[inout] j_max column index
MaxEntry(const CeedScalar * A,CeedInt N,CeedInt * max_idx_row,CeedInt * i_max,CeedInt * j_max)44 CEED_QFUNCTION_HELPER void MaxEntry(const CeedScalar *A, CeedInt N, CeedInt *max_idx_row, CeedInt *i_max, CeedInt *j_max) {
45   *i_max               = 0;
46   *j_max               = max_idx_row[*i_max];
47   CeedScalar max_entry = fabs(A[*i_max * N + *j_max]);
48   for (CeedInt i = 1; i < N - 1; i++) {
49     CeedInt j = max_idx_row[i];
50     if (fabs(A[i * N + j]) > max_entry) {
51       max_entry = fabs(A[i * N + j]);
52       *i_max    = i;
53       *j_max    = j;
54     }
55   }
56 }
57 
58 /// @brief Calculate the components of a rotation matrix which performs a
59 ///        rotation in the i,j plane by an angle (θ) that (when multiplied on
60 ///        both sides) will zero the ij'th element of A, so that afterwards
61 ///        A[i][j] = 0.  The results will be stored in c, s, and t
62 ///        (which store cos(θ), sin(θ), and tan(θ), respectively).
63 ///
64 /// @param[in] *A matrix
65 /// @param[in] i row index
66 /// @param[in] j column index
CalcRot(const CeedScalar * A,CeedInt N,CeedInt i,CeedInt j,CeedScalar * rotmat_cst)67 CEED_QFUNCTION_HELPER void CalcRot(const CeedScalar *A, CeedInt N, CeedInt i, CeedInt j, CeedScalar *rotmat_cst) {
68   rotmat_cst[2]      = 1.0;  // = tan(θ)
69   CeedScalar A_jj_ii = (A[j * N + j] - A[i * N + i]);
70   if (A_jj_ii != 0.0) {
71     // kappa = (A[j][j] - A[i][i]) / (2*A[i][j])
72     CeedScalar kappa = A_jj_ii;
73     rotmat_cst[2]    = 0.0;
74     CeedScalar A_ij  = A[i * N + j];
75     if (A_ij != 0.0) {
76       kappa /= (2.0 * A_ij);
77       // t satisfies: t^2 + 2*t*kappa - 1 = 0
78       // (choose the root which has the smaller absolute value)
79       rotmat_cst[2] = 1.0 / (sqrt(1 + kappa * kappa) + fabs(kappa));
80       if (kappa < 0.0) rotmat_cst[2] = -rotmat_cst[2];
81     }
82   }
83   rotmat_cst[0] = 1.0 / sqrt(1 + rotmat_cst[2] * rotmat_cst[2]);
84   rotmat_cst[1] = rotmat_cst[0] * rotmat_cst[2];
85 }
86 
87 /// @brief  Perform a similarity transformation by multiplying matrix A on both
88 ///         sides by a rotation matrix (and its transpose) to eliminate A[i][j].
89 /// @details This rotation matrix performs a rotation in the i,j plane by
90 ///         angle θ.  This function assumes that c=cos(θ). s=sin(θ), t=tan(θ)
91 ///         have been calculated in advance (using the CalcRot() function).
92 ///         It also assumes that i<j.  The max_idx_row[] array is also updated.
93 ///         To save time, since the matrix is symmetric, the elements
94 ///         below the diagonal (ie. A[u][v] where u>v) are not computed.
95 /// @verbatim
96 ///   A' = R^T * A * R
97 /// where R the rotation in the i,j plane and ^T denotes the transpose.
98 ///                 i         j
99 ///       _                             _
100 ///      |  1                            |
101 ///      |    .                          |
102 ///      |      .                        |
103 ///      |        1                      |
104 ///      |          c   ...   s          |
105 ///      |          .  .      .          |
106 /// R  = |          .    1    .          |
107 ///      |          .      .  .          |
108 ///      |          -s  ...   c          |
109 ///      |                      1        |
110 ///      |                        .      |
111 ///      |                          .    |
112 ///      |_                           1 _|
113 /// @endverbatim
114 ///
115 /// Let A' denote the matrix A after multiplication by R^T and R.
116 /// The components of A' are:
117 ///
118 /// @verbatim
119 ///   A'_uv =  Σ_w  Σ_z   R_wu * A_wz * R_zv
120 /// @endverbatim
121 ///
122 /// Note that a the rotation at location i,j will modify all of the matrix
123 /// elements containing at least one index which is either i or j
124 /// such as: A[w][i], A[i][w], A[w][j], A[j][w].
125 /// Check and see whether these modified matrix elements exceed the
126 /// corresponding values in max_idx_row[] array for that row.
127 /// If so, then update max_idx_row for that row.
128 /// This is somewhat complicated by the fact that we must only consider
129 /// matrix elements in the upper-right triangle strictly above the diagonal.
130 /// (ie. matrix elements whose second index is > the first index).
131 /// The modified elements we must consider are marked with an "X" below:
132 ///
133 /// @verbatim
134 ///                 i         j
135 ///       _                             _
136 ///      |  .       X         X          |
137 ///      |    .     X         X          |
138 ///      |      .   X         X          |
139 ///      |        . X         X          |
140 ///      |          X X X X X 0 X X X X  |  i
141 ///      |            .       X          |
142 ///      |              .     X          |
143 /// A  = |                .   X          |
144 ///      |                  . X          |
145 ///      |                    X X X X X  |  j
146 ///      |                      .        |
147 ///      |                        .      |
148 ///      |                          .    |
149 ///      |_                           . _|
150 /// @endverbatim
151 ///
152 /// @param[in] *A matrix
153 /// @param[in] i row index
154 /// @param[in] j column index
ApplyRot(CeedScalar * A,CeedInt N,CeedInt i,CeedInt j,CeedInt * max_idx_row,CeedScalar * rotmat_cst)155 CEED_QFUNCTION_HELPER void ApplyRot(CeedScalar *A, CeedInt N, CeedInt i, CeedInt j, CeedInt *max_idx_row, CeedScalar *rotmat_cst) {
156   // Compute the diagonal elements of A which have changed:
157   A[i * N + i] -= rotmat_cst[2] * A[i * N + j];
158   A[j * N + j] += rotmat_cst[2] * A[i * N + j];
159   // Note: This is algebraically equivalent to:
160   // A[i][i] = c*c*A[i][i] + s*s*A[j][j] - 2*s*c*A[i][j]
161   // A[j][j] = s*s*A[i][i] + c*c*A[j][j] + 2*s*c*A[i][j]
162 
163   // Update the off-diagonal elements of A which will change (above the diagonal)
164 
165   A[i * N + j] = 0.0;
166 
167   // compute A[w][i] and A[i][w] for all w!=i,considering above-diagonal elements
168   for (CeedInt w = 0; w < i; w++) {                                              // 0 <= w <  i  <  j < N
169     A[i * N + w] = A[w * N + i];                                                 // backup the previous value. store below diagonal (i>w)
170     A[w * N + i] = rotmat_cst[0] * A[w * N + i] - rotmat_cst[1] * A[w * N + j];  // A[w][i], A[w][j] from previous iteration
171     if (i == max_idx_row[w]) max_idx_row[w] = MaxEntryRow(A, N, w);
172     else if (fabs(A[w * N + i]) > fabs(A[w * N + max_idx_row[w]])) max_idx_row[w] = i;
173   }
174   for (CeedInt w = i + 1; w < j; w++) {                                          // 0 <= i <  w  <  j < N
175     A[w * N + i] = A[i * N + w];                                                 // backup the previous value. store below diagonal (w>i)
176     A[i * N + w] = rotmat_cst[0] * A[i * N + w] - rotmat_cst[1] * A[w * N + j];  // A[i][w], A[w][j] from previous iteration
177   }
178   for (CeedInt w = j + 1; w < N; w++) {                                          // 0 <= i < j+1 <= w < N
179     A[w * N + i] = A[i * N + w];                                                 // backup the previous value. store below diagonal (w>i)
180     A[i * N + w] = rotmat_cst[0] * A[i * N + w] - rotmat_cst[1] * A[j * N + w];  // A[i][w], A[j][w] from previous iteration
181   }
182 
183   // now that we're done modifying row i, we can update max_idx_row[i]
184   max_idx_row[i] = MaxEntryRow(A, N, i);
185 
186   // compute A[w][j] and A[j][w] for all w!=j,considering above-diagonal elements
187   for (CeedInt w = 0; w < i; w++) {                                              // 0 <=  w  <  i <  j < N
188     A[w * N + j] = rotmat_cst[1] * A[i * N + w] + rotmat_cst[0] * A[w * N + j];  // A[i][w], A[w][j] from previous iteration
189     if (j == max_idx_row[w]) max_idx_row[w] = MaxEntryRow(A, N, w);
190     else if (fabs(A[w * N + j]) > fabs(A[w * N + max_idx_row[w]])) max_idx_row[w] = j;
191   }
192   for (CeedInt w = i + 1; w < j; w++) {                                          // 0 <= i+1 <= w <  j < N
193     A[w * N + j] = rotmat_cst[1] * A[w * N + i] + rotmat_cst[0] * A[w * N + j];  // A[w][i], A[w][j] from previous iteration
194     if (j == max_idx_row[w]) max_idx_row[w] = MaxEntryRow(A, N, w);
195     else if (fabs(A[w * N + j]) > fabs(A[w * N + max_idx_row[w]])) max_idx_row[w] = j;
196   }
197   for (CeedInt w = j + 1; w < N; w++) {                                          // 0 <=  i  <  j <  w < N
198     A[j * N + w] = rotmat_cst[1] * A[w * N + i] + rotmat_cst[0] * A[j * N + w];  // A[w][i], A[j][w] from previous iteration
199   }
200   // now that we're done modifying row j, we can update max_idx_row[j]
201   max_idx_row[j] = MaxEntryRow(A, N, j);
202 }
203 
204 ///@brief Multiply matrix A on the LEFT side by a transposed rotation matrix R^T
205 ///       This matrix performs a rotation in the i,j plane by angle θ  (where
206 ///       the arguments "s" and "c" refer to cos(θ) and sin(θ), respectively).
207 /// @verbatim
208 ///   A'_uv = Σ_w  R_wu * A_wv
209 /// @endverbatim
210 ///
211 /// @param[in] *A matrix
212 /// @param[in] i row index
213 /// @param[in] j column index
ApplyRotLeft(CeedScalar * A,CeedInt N,CeedInt i,CeedInt j,CeedScalar * rotmat_cst)214 CEED_QFUNCTION_HELPER void ApplyRotLeft(CeedScalar *A, CeedInt N, CeedInt i, CeedInt j, CeedScalar *rotmat_cst) {
215   // Recall that c = cos(θ) and s = sin(θ)
216   for (CeedInt v = 0; v < N; v++) {
217     CeedScalar Aiv = A[i * N + v];
218     A[i * N + v]   = rotmat_cst[0] * A[i * N + v] - rotmat_cst[1] * A[j * N + v];
219     A[j * N + v]   = rotmat_cst[1] * Aiv + rotmat_cst[0] * A[j * N + v];
220   }
221 }
222 
223 /// @brief Sort the rows in evec according to the numbers in v (also sorted)
224 ///
225 /// @param[inout] *eval vector containing the keys used for sorting
226 /// @param[inout] *evec matrix whose rows will be sorted according to v
227 /// @param[in]    n  size of the vector and matrix
228 /// @param[in]    s  sort decreasing order?
SortRows(CeedScalar * eval,CeedScalar * evec,CeedInt N,SortCriteria sort_criteria)229 CEED_QFUNCTION_HELPER void SortRows(CeedScalar *eval, CeedScalar *evec, CeedInt N, SortCriteria sort_criteria) {
230   if (sort_criteria == SORT_NONE) return;
231 
232   for (CeedInt i = 0; i < N - 1; i++) {
233     CeedInt i_max = i;
234     for (CeedInt j = i + 1; j < N; j++) {
235       // find the "maximum" element in the array starting at position i+1
236       switch (sort_criteria) {
237         case SORT_DECREASING_EVALS:
238           if (eval[j] > eval[i_max]) i_max = j;
239           break;
240         case SORT_INCREASING_EVALS:
241           if (eval[j] < eval[i_max]) i_max = j;
242           break;
243         case SORT_DECREASING_ABS_EVALS:
244           if (fabs(eval[j]) > fabs(eval[i_max])) i_max = j;
245           break;
246         case SORT_INCREASING_ABS_EVALS:
247           if (fabs(eval[j]) < fabs(eval[i_max])) i_max = j;
248           break;
249         default:
250           break;
251       }
252     }
253     SwapScalar(&eval[i], &eval[i_max]);
254     for (CeedInt k = 0; k < N; k++) SwapScalar(&evec[i * N + k], &evec[i_max * N + k]);
255   }
256 }
257 
258 /// @brief Calculate all the eigenvalues and eigevectors of a symmetric matrix
259 ///        using the Jacobi eigenvalue algorithm:
260 ///        https://en.wikipedia.org/wiki/Jacobi_eigenvalue_algorithm
261 /// @returns The number of Jacobi iterations attempted, which should be > 0.
262 ///          If the return value is not strictly > 0 then convergence failed.
263 /// @note  To reduce the computation time further, set calc_evecs=false.
264 ///        Additionally, note that the output evecs should be normalized. It
265 ///        simply takes the Identity matrix and performs (isometric) rotations
266 ///        on it, so divergence from normalized is due to finite-precision
267 ///        arithmetic of the rotations.
268 //
269 // @param[in]  A              the matrix you wish to diagonalize (size NxN)
270 // @param[in]  N              size of the matrix
271 // @param[out] eval           store the eigenvalues here (size N)
272 // @param[out] evec           store the eigenvectors here (in rows, size NxN)
273 // @param[out] max_idx_row    work vector of size N
274 // @param[in]  sort_criteria  sort results?
275 // @param[in]  calc_evecs     calculate the eigenvectors?
276 // @param[in]  max_num_sweeps maximum number of iterations = max_num_sweeps * number of off-diagonals (N*(N-1)/2)
Diagonalize(CeedScalar * A,CeedInt N,CeedScalar * eval,CeedScalar * evec,CeedInt * max_idx_row,SortCriteria sort_criteria,bool calc_evec,const CeedInt max_num_sweeps)277 CEED_QFUNCTION_HELPER CeedInt Diagonalize(CeedScalar *A, CeedInt N, CeedScalar *eval, CeedScalar *evec, CeedInt *max_idx_row,
278                                           SortCriteria sort_criteria, bool calc_evec, const CeedInt max_num_sweeps) {
279   CeedScalar rotmat_cst[3] = {0.};  // cos(θ), sin(θ), and tan(θ),
280 
281   if (calc_evec)
282     for (CeedInt i = 0; i < N; i++)
283       for (CeedInt j = 0; j < N; j++) evec[i * N + j] = (i == j) ? 1.0 : 0.0;  // Set evec equal to the identity matrix
284 
285   for (CeedInt i = 0; i < N - 1; i++) max_idx_row[i] = MaxEntryRow(A, N, i);
286 
287   // -- Iteration --
288   CeedInt n_iters;
289   CeedInt max_num_iters = max_num_sweeps * N * (N - 1) / 2;
290   for (n_iters = 1; n_iters <= max_num_iters; n_iters++) {
291     CeedInt i, j;
292     MaxEntry(A, N, max_idx_row, &i, &j);
293 
294     // If A[i][j] is small compared to A[i][i] and A[j][j], set it to 0.
295     if ((A[i * N + i] + A[i * N + j] == A[i * N + i]) && (A[j * N + j] + A[i * N + j] == A[j * N + j])) {
296       A[i * N + j]   = 0.0;
297       max_idx_row[i] = MaxEntryRow(A, N, i);
298     }
299 
300     if (A[i * N + j] == 0.0) break;
301 
302     CalcRot(A, N, i, j, rotmat_cst);                // Calculate the parameters of the rotation matrix.
303     ApplyRot(A, N, i, j, max_idx_row, rotmat_cst);  // Apply this rotation to the A matrix.
304     if (calc_evec) ApplyRotLeft(evec, N, i, j, rotmat_cst);
305   }
306 
307   for (CeedInt i = 0; i < N; i++) eval[i] = A[i * N + i];
308 
309   // Optional: Sort results by eigenvalue.
310   SortRows(eval, evec, N, sort_criteria);
311 
312   if ((n_iters > max_num_iters) && (N > 1))  // If we exceeded max_num_iters,
313     return 0;                                // indicate an error occured.
314 
315   return n_iters;
316 }
317 
318 // @brief Interface to Diagonalize for 3x3 systems
Diagonalize3(CeedScalar A[3][3],CeedScalar eval[3],CeedScalar evec[3][3],CeedInt max_idx_row[3],SortCriteria sort_criteria,bool calc_evec,const CeedInt max_num_sweeps)319 CEED_QFUNCTION_HELPER CeedInt Diagonalize3(CeedScalar A[3][3], CeedScalar eval[3], CeedScalar evec[3][3], CeedInt max_idx_row[3],
320                                            SortCriteria sort_criteria, bool calc_evec, const CeedInt max_num_sweeps) {
321   return Diagonalize((CeedScalar *)A, 3, (CeedScalar *)eval, (CeedScalar *)evec, (CeedInt *)max_idx_row, sort_criteria, calc_evec, max_num_sweeps);
322 }
323