|
1 /* glpmat.c */ |
|
2 |
|
3 /*********************************************************************** |
|
4 * This code is part of GLPK (GNU Linear Programming Kit). |
|
5 * |
|
6 * Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, |
|
7 * 2009, 2010 Andrew Makhorin, Department for Applied Informatics, |
|
8 * Moscow Aviation Institute, Moscow, Russia. All rights reserved. |
|
9 * E-mail: <mao@gnu.org>. |
|
10 * |
|
11 * GLPK is free software: you can redistribute it and/or modify it |
|
12 * under the terms of the GNU General Public License as published by |
|
13 * the Free Software Foundation, either version 3 of the License, or |
|
14 * (at your option) any later version. |
|
15 * |
|
16 * GLPK is distributed in the hope that it will be useful, but WITHOUT |
|
17 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY |
|
18 * or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public |
|
19 * License for more details. |
|
20 * |
|
21 * You should have received a copy of the GNU General Public License |
|
22 * along with GLPK. If not, see <http://www.gnu.org/licenses/>. |
|
23 ***********************************************************************/ |
|
24 |
|
25 #include "glpenv.h" |
|
26 #include "glpmat.h" |
|
27 #include "glpqmd.h" |
|
28 #include "amd/amd.h" |
|
29 #include "colamd/colamd.h" |
|
30 |
|
31 /*---------------------------------------------------------------------- |
|
32 -- check_fvs - check sparse vector in full-vector storage format. |
|
33 -- |
|
34 -- SYNOPSIS |
|
35 -- |
|
36 -- #include "glpmat.h" |
|
37 -- int check_fvs(int n, int nnz, int ind[], double vec[]); |
|
38 -- |
|
39 -- DESCRIPTION |
|
40 -- |
|
41 -- The routine check_fvs checks if a given vector of dimension n in |
|
42 -- full-vector storage format has correct representation. |
|
43 -- |
|
44 -- RETURNS |
|
45 -- |
|
46 -- The routine returns one of the following codes: |
|
47 -- |
|
48 -- 0 - the vector is correct; |
|
49 -- 1 - the number of elements (n) is negative; |
|
50 -- 2 - the number of non-zero elements (nnz) is negative; |
|
51 -- 3 - some element index is out of range; |
|
52 -- 4 - some element index is duplicate; |
|
53 -- 5 - some non-zero element is out of pattern. */ |
|
54 |
|
55 int check_fvs(int n, int nnz, int ind[], double vec[]) |
|
56 { int i, t, ret, *flag = NULL; |
|
57 /* check the number of elements */ |
|
58 if (n < 0) |
|
59 { ret = 1; |
|
60 goto done; |
|
61 } |
|
62 /* check the number of non-zero elements */ |
|
63 if (nnz < 0) |
|
64 { ret = 2; |
|
65 goto done; |
|
66 } |
|
67 /* check vector indices */ |
|
68 flag = xcalloc(1+n, sizeof(int)); |
|
69 for (i = 1; i <= n; i++) flag[i] = 0; |
|
70 for (t = 1; t <= nnz; t++) |
|
71 { i = ind[t]; |
|
72 if (!(1 <= i && i <= n)) |
|
73 { ret = 3; |
|
74 goto done; |
|
75 } |
|
76 if (flag[i]) |
|
77 { ret = 4; |
|
78 goto done; |
|
79 } |
|
80 flag[i] = 1; |
|
81 } |
|
82 /* check vector elements */ |
|
83 for (i = 1; i <= n; i++) |
|
84 { if (!flag[i] && vec[i] != 0.0) |
|
85 { ret = 5; |
|
86 goto done; |
|
87 } |
|
88 } |
|
89 /* the vector is ok */ |
|
90 ret = 0; |
|
91 done: if (flag != NULL) xfree(flag); |
|
92 return ret; |
|
93 } |
|
94 |
|
95 /*---------------------------------------------------------------------- |
|
96 -- check_pattern - check pattern of sparse matrix. |
|
97 -- |
|
98 -- SYNOPSIS |
|
99 -- |
|
100 -- #include "glpmat.h" |
|
101 -- int check_pattern(int m, int n, int A_ptr[], int A_ind[]); |
|
102 -- |
|
103 -- DESCRIPTION |
|
104 -- |
|
105 -- The routine check_pattern checks the pattern of a given mxn matrix |
|
106 -- in storage-by-rows format. |
|
107 -- |
|
108 -- RETURNS |
|
109 -- |
|
110 -- The routine returns one of the following codes: |
|
111 -- |
|
112 -- 0 - the pattern is correct; |
|
113 -- 1 - the number of rows (m) is negative; |
|
114 -- 2 - the number of columns (n) is negative; |
|
115 -- 3 - A_ptr[1] is not 1; |
|
116 -- 4 - some column index is out of range; |
|
117 -- 5 - some column indices are duplicate. */ |
|
118 |
|
119 int check_pattern(int m, int n, int A_ptr[], int A_ind[]) |
|
120 { int i, j, ptr, ret, *flag = NULL; |
|
121 /* check the number of rows */ |
|
122 if (m < 0) |
|
123 { ret = 1; |
|
124 goto done; |
|
125 } |
|
126 /* check the number of columns */ |
|
127 if (n < 0) |
|
128 { ret = 2; |
|
129 goto done; |
|
130 } |
|
131 /* check location A_ptr[1] */ |
|
132 if (A_ptr[1] != 1) |
|
133 { ret = 3; |
|
134 goto done; |
|
135 } |
|
136 /* check row patterns */ |
|
137 flag = xcalloc(1+n, sizeof(int)); |
|
138 for (j = 1; j <= n; j++) flag[j] = 0; |
|
139 for (i = 1; i <= m; i++) |
|
140 { /* check pattern of row i */ |
|
141 for (ptr = A_ptr[i]; ptr < A_ptr[i+1]; ptr++) |
|
142 { j = A_ind[ptr]; |
|
143 /* check column index */ |
|
144 if (!(1 <= j && j <= n)) |
|
145 { ret = 4; |
|
146 goto done; |
|
147 } |
|
148 /* check for duplication */ |
|
149 if (flag[j]) |
|
150 { ret = 5; |
|
151 goto done; |
|
152 } |
|
153 flag[j] = 1; |
|
154 } |
|
155 /* clear flags */ |
|
156 for (ptr = A_ptr[i]; ptr < A_ptr[i+1]; ptr++) |
|
157 { j = A_ind[ptr]; |
|
158 flag[j] = 0; |
|
159 } |
|
160 } |
|
161 /* the pattern is ok */ |
|
162 ret = 0; |
|
163 done: if (flag != NULL) xfree(flag); |
|
164 return ret; |
|
165 } |
|
166 |
|
167 /*---------------------------------------------------------------------- |
|
168 -- transpose - transpose sparse matrix. |
|
169 -- |
|
170 -- *Synopsis* |
|
171 -- |
|
172 -- #include "glpmat.h" |
|
173 -- void transpose(int m, int n, int A_ptr[], int A_ind[], |
|
174 -- double A_val[], int AT_ptr[], int AT_ind[], double AT_val[]); |
|
175 -- |
|
176 -- *Description* |
|
177 -- |
|
178 -- For a given mxn sparse matrix A the routine transpose builds a nxm |
|
179 -- sparse matrix A' which is a matrix transposed to A. |
|
180 -- |
|
181 -- The arrays A_ptr, A_ind, and A_val specify a given mxn matrix A to |
|
182 -- be transposed in storage-by-rows format. The parameter A_val can be |
|
183 -- NULL, in which case numeric values are not copied. The arrays A_ptr, |
|
184 -- A_ind, and A_val are not changed on exit. |
|
185 -- |
|
186 -- On entry the arrays AT_ptr, AT_ind, and AT_val must be allocated, |
|
187 -- but their content is ignored. On exit the routine stores a resultant |
|
188 -- nxm matrix A' in these arrays in storage-by-rows format. Note that |
|
189 -- if the parameter A_val is NULL, the array AT_val is not used. |
|
190 -- |
|
191 -- The routine transpose has a side effect that elements in rows of the |
|
192 -- resultant matrix A' follow in ascending their column indices. */ |
|
193 |
|
194 void transpose(int m, int n, int A_ptr[], int A_ind[], double A_val[], |
|
195 int AT_ptr[], int AT_ind[], double AT_val[]) |
|
196 { int i, j, t, beg, end, pos, len; |
|
197 /* determine row lengths of resultant matrix */ |
|
198 for (j = 1; j <= n; j++) AT_ptr[j] = 0; |
|
199 for (i = 1; i <= m; i++) |
|
200 { beg = A_ptr[i], end = A_ptr[i+1]; |
|
201 for (t = beg; t < end; t++) AT_ptr[A_ind[t]]++; |
|
202 } |
|
203 /* set up row pointers of resultant matrix */ |
|
204 pos = 1; |
|
205 for (j = 1; j <= n; j++) |
|
206 len = AT_ptr[j], pos += len, AT_ptr[j] = pos; |
|
207 AT_ptr[n+1] = pos; |
|
208 /* build resultant matrix */ |
|
209 for (i = m; i >= 1; i--) |
|
210 { beg = A_ptr[i], end = A_ptr[i+1]; |
|
211 for (t = beg; t < end; t++) |
|
212 { pos = --AT_ptr[A_ind[t]]; |
|
213 AT_ind[pos] = i; |
|
214 if (A_val != NULL) AT_val[pos] = A_val[t]; |
|
215 } |
|
216 } |
|
217 return; |
|
218 } |
|
219 |
|
220 /*---------------------------------------------------------------------- |
|
221 -- adat_symbolic - compute S = P*A*D*A'*P' (symbolic phase). |
|
222 -- |
|
223 -- *Synopsis* |
|
224 -- |
|
225 -- #include "glpmat.h" |
|
226 -- int *adat_symbolic(int m, int n, int P_per[], int A_ptr[], |
|
227 -- int A_ind[], int S_ptr[]); |
|
228 -- |
|
229 -- *Description* |
|
230 -- |
|
231 -- The routine adat_symbolic implements the symbolic phase to compute |
|
232 -- symmetric matrix S = P*A*D*A'*P', where P is a permutation matrix, |
|
233 -- A is a given sparse matrix, D is a diagonal matrix, A' is a matrix |
|
234 -- transposed to A, P' is an inverse of P. |
|
235 -- |
|
236 -- The parameter m is the number of rows in A and the order of P. |
|
237 -- |
|
238 -- The parameter n is the number of columns in A and the order of D. |
|
239 -- |
|
240 -- The array P_per specifies permutation matrix P. It is not changed on |
|
241 -- exit. |
|
242 -- |
|
243 -- The arrays A_ptr and A_ind specify the pattern of matrix A. They are |
|
244 -- not changed on exit. |
|
245 -- |
|
246 -- On exit the routine stores the pattern of upper triangular part of |
|
247 -- matrix S without diagonal elements in the arrays S_ptr and S_ind in |
|
248 -- storage-by-rows format. The array S_ptr should be allocated on entry, |
|
249 -- however, its content is ignored. The array S_ind is allocated by the |
|
250 -- routine itself which returns a pointer to it. |
|
251 -- |
|
252 -- *Returns* |
|
253 -- |
|
254 -- The routine returns a pointer to the array S_ind. */ |
|
255 |
|
256 int *adat_symbolic(int m, int n, int P_per[], int A_ptr[], int A_ind[], |
|
257 int S_ptr[]) |
|
258 { int i, j, t, ii, jj, tt, k, size, len; |
|
259 int *S_ind, *AT_ptr, *AT_ind, *ind, *map, *temp; |
|
260 /* build the pattern of A', which is a matrix transposed to A, to |
|
261 efficiently access A in column-wise manner */ |
|
262 AT_ptr = xcalloc(1+n+1, sizeof(int)); |
|
263 AT_ind = xcalloc(A_ptr[m+1], sizeof(int)); |
|
264 transpose(m, n, A_ptr, A_ind, NULL, AT_ptr, AT_ind, NULL); |
|
265 /* allocate the array S_ind */ |
|
266 size = A_ptr[m+1] - 1; |
|
267 if (size < m) size = m; |
|
268 S_ind = xcalloc(1+size, sizeof(int)); |
|
269 /* allocate and initialize working arrays */ |
|
270 ind = xcalloc(1+m, sizeof(int)); |
|
271 map = xcalloc(1+m, sizeof(int)); |
|
272 for (jj = 1; jj <= m; jj++) map[jj] = 0; |
|
273 /* compute pattern of S; note that symbolically S = B*B', where |
|
274 B = P*A, B' is matrix transposed to B */ |
|
275 S_ptr[1] = 1; |
|
276 for (ii = 1; ii <= m; ii++) |
|
277 { /* compute pattern of ii-th row of S */ |
|
278 len = 0; |
|
279 i = P_per[ii]; /* i-th row of A = ii-th row of B */ |
|
280 for (t = A_ptr[i]; t < A_ptr[i+1]; t++) |
|
281 { k = A_ind[t]; |
|
282 /* walk through k-th column of A */ |
|
283 for (tt = AT_ptr[k]; tt < AT_ptr[k+1]; tt++) |
|
284 { j = AT_ind[tt]; |
|
285 jj = P_per[m+j]; /* j-th row of A = jj-th row of B */ |
|
286 /* a[i,k] != 0 and a[j,k] != 0 ergo s[ii,jj] != 0 */ |
|
287 if (ii < jj && !map[jj]) ind[++len] = jj, map[jj] = 1; |
|
288 } |
|
289 } |
|
290 /* now (ind) is pattern of ii-th row of S */ |
|
291 S_ptr[ii+1] = S_ptr[ii] + len; |
|
292 /* at least (S_ptr[ii+1] - 1) locations should be available in |
|
293 the array S_ind */ |
|
294 if (S_ptr[ii+1] - 1 > size) |
|
295 { temp = S_ind; |
|
296 size += size; |
|
297 S_ind = xcalloc(1+size, sizeof(int)); |
|
298 memcpy(&S_ind[1], &temp[1], (S_ptr[ii] - 1) * sizeof(int)); |
|
299 xfree(temp); |
|
300 } |
|
301 xassert(S_ptr[ii+1] - 1 <= size); |
|
302 /* (ii-th row of S) := (ind) */ |
|
303 memcpy(&S_ind[S_ptr[ii]], &ind[1], len * sizeof(int)); |
|
304 /* clear the row pattern map */ |
|
305 for (t = 1; t <= len; t++) map[ind[t]] = 0; |
|
306 } |
|
307 /* free working arrays */ |
|
308 xfree(AT_ptr); |
|
309 xfree(AT_ind); |
|
310 xfree(ind); |
|
311 xfree(map); |
|
312 /* reallocate the array S_ind to free unused locations */ |
|
313 temp = S_ind; |
|
314 size = S_ptr[m+1] - 1; |
|
315 S_ind = xcalloc(1+size, sizeof(int)); |
|
316 memcpy(&S_ind[1], &temp[1], size * sizeof(int)); |
|
317 xfree(temp); |
|
318 return S_ind; |
|
319 } |
|
320 |
|
321 /*---------------------------------------------------------------------- |
|
322 -- adat_numeric - compute S = P*A*D*A'*P' (numeric phase). |
|
323 -- |
|
324 -- *Synopsis* |
|
325 -- |
|
326 -- #include "glpmat.h" |
|
327 -- void adat_numeric(int m, int n, int P_per[], |
|
328 -- int A_ptr[], int A_ind[], double A_val[], double D_diag[], |
|
329 -- int S_ptr[], int S_ind[], double S_val[], double S_diag[]); |
|
330 -- |
|
331 -- *Description* |
|
332 -- |
|
333 -- The routine adat_numeric implements the numeric phase to compute |
|
334 -- symmetric matrix S = P*A*D*A'*P', where P is a permutation matrix, |
|
335 -- A is a given sparse matrix, D is a diagonal matrix, A' is a matrix |
|
336 -- transposed to A, P' is an inverse of P. |
|
337 -- |
|
338 -- The parameter m is the number of rows in A and the order of P. |
|
339 -- |
|
340 -- The parameter n is the number of columns in A and the order of D. |
|
341 -- |
|
342 -- The matrix P is specified in the array P_per, which is not changed |
|
343 -- on exit. |
|
344 -- |
|
345 -- The matrix A is specified in the arrays A_ptr, A_ind, and A_val in |
|
346 -- storage-by-rows format. These arrays are not changed on exit. |
|
347 -- |
|
348 -- Diagonal elements of the matrix D are specified in the array D_diag, |
|
349 -- where D_diag[0] is not used, D_diag[i] = d[i,i] for i = 1, ..., n. |
|
350 -- The array D_diag is not changed on exit. |
|
351 -- |
|
352 -- The pattern of the upper triangular part of the matrix S without |
|
353 -- diagonal elements (previously computed by the routine adat_symbolic) |
|
354 -- is specified in the arrays S_ptr and S_ind, which are not changed on |
|
355 -- exit. Numeric values of non-diagonal elements of S are stored in |
|
356 -- corresponding locations of the array S_val, and values of diagonal |
|
357 -- elements of S are stored in locations S_diag[1], ..., S_diag[n]. */ |
|
358 |
|
359 void adat_numeric(int m, int n, int P_per[], |
|
360 int A_ptr[], int A_ind[], double A_val[], double D_diag[], |
|
361 int S_ptr[], int S_ind[], double S_val[], double S_diag[]) |
|
362 { int i, j, t, ii, jj, tt, beg, end, beg1, end1, k; |
|
363 double sum, *work; |
|
364 work = xcalloc(1+n, sizeof(double)); |
|
365 for (j = 1; j <= n; j++) work[j] = 0.0; |
|
366 /* compute S = B*D*B', where B = P*A, B' is a matrix transposed |
|
367 to B */ |
|
368 for (ii = 1; ii <= m; ii++) |
|
369 { i = P_per[ii]; /* i-th row of A = ii-th row of B */ |
|
370 /* (work) := (i-th row of A) */ |
|
371 beg = A_ptr[i], end = A_ptr[i+1]; |
|
372 for (t = beg; t < end; t++) |
|
373 work[A_ind[t]] = A_val[t]; |
|
374 /* compute ii-th row of S */ |
|
375 beg = S_ptr[ii], end = S_ptr[ii+1]; |
|
376 for (t = beg; t < end; t++) |
|
377 { jj = S_ind[t]; |
|
378 j = P_per[jj]; /* j-th row of A = jj-th row of B */ |
|
379 /* s[ii,jj] := sum a[i,k] * d[k,k] * a[j,k] */ |
|
380 sum = 0.0; |
|
381 beg1 = A_ptr[j], end1 = A_ptr[j+1]; |
|
382 for (tt = beg1; tt < end1; tt++) |
|
383 { k = A_ind[tt]; |
|
384 sum += work[k] * D_diag[k] * A_val[tt]; |
|
385 } |
|
386 S_val[t] = sum; |
|
387 } |
|
388 /* s[ii,ii] := sum a[i,k] * d[k,k] * a[i,k] */ |
|
389 sum = 0.0; |
|
390 beg = A_ptr[i], end = A_ptr[i+1]; |
|
391 for (t = beg; t < end; t++) |
|
392 { k = A_ind[t]; |
|
393 sum += A_val[t] * D_diag[k] * A_val[t]; |
|
394 work[k] = 0.0; |
|
395 } |
|
396 S_diag[ii] = sum; |
|
397 } |
|
398 xfree(work); |
|
399 return; |
|
400 } |
|
401 |
|
402 /*---------------------------------------------------------------------- |
|
403 -- min_degree - minimum degree ordering. |
|
404 -- |
|
405 -- *Synopsis* |
|
406 -- |
|
407 -- #include "glpmat.h" |
|
408 -- void min_degree(int n, int A_ptr[], int A_ind[], int P_per[]); |
|
409 -- |
|
410 -- *Description* |
|
411 -- |
|
412 -- The routine min_degree uses the minimum degree ordering algorithm |
|
413 -- to find a permutation matrix P for a given sparse symmetric positive |
|
414 -- matrix A which minimizes the number of non-zeros in upper triangular |
|
415 -- factor U for Cholesky factorization P*A*P' = U'*U. |
|
416 -- |
|
417 -- The parameter n is the order of matrices A and P. |
|
418 -- |
|
419 -- The pattern of the given matrix A is specified on entry in the arrays |
|
420 -- A_ptr and A_ind in storage-by-rows format. Only the upper triangular |
|
421 -- part without diagonal elements (which all are assumed to be non-zero) |
|
422 -- should be specified as if A were upper triangular. The arrays A_ptr |
|
423 -- and A_ind are not changed on exit. |
|
424 -- |
|
425 -- The permutation matrix P is stored by the routine in the array P_per |
|
426 -- on exit. |
|
427 -- |
|
428 -- *Algorithm* |
|
429 -- |
|
430 -- The routine min_degree is based on some subroutines from the package |
|
431 -- SPARSPAK (see comments in the module glpqmd). */ |
|
432 |
|
433 void min_degree(int n, int A_ptr[], int A_ind[], int P_per[]) |
|
434 { int i, j, ne, t, pos, len; |
|
435 int *xadj, *adjncy, *deg, *marker, *rchset, *nbrhd, *qsize, |
|
436 *qlink, nofsub; |
|
437 /* determine number of non-zeros in complete pattern */ |
|
438 ne = A_ptr[n+1] - 1; |
|
439 ne += ne; |
|
440 /* allocate working arrays */ |
|
441 xadj = xcalloc(1+n+1, sizeof(int)); |
|
442 adjncy = xcalloc(1+ne, sizeof(int)); |
|
443 deg = xcalloc(1+n, sizeof(int)); |
|
444 marker = xcalloc(1+n, sizeof(int)); |
|
445 rchset = xcalloc(1+n, sizeof(int)); |
|
446 nbrhd = xcalloc(1+n, sizeof(int)); |
|
447 qsize = xcalloc(1+n, sizeof(int)); |
|
448 qlink = xcalloc(1+n, sizeof(int)); |
|
449 /* determine row lengths in complete pattern */ |
|
450 for (i = 1; i <= n; i++) xadj[i] = 0; |
|
451 for (i = 1; i <= n; i++) |
|
452 { for (t = A_ptr[i]; t < A_ptr[i+1]; t++) |
|
453 { j = A_ind[t]; |
|
454 xassert(i < j && j <= n); |
|
455 xadj[i]++, xadj[j]++; |
|
456 } |
|
457 } |
|
458 /* set up row pointers for complete pattern */ |
|
459 pos = 1; |
|
460 for (i = 1; i <= n; i++) |
|
461 len = xadj[i], pos += len, xadj[i] = pos; |
|
462 xadj[n+1] = pos; |
|
463 xassert(pos - 1 == ne); |
|
464 /* construct complete pattern */ |
|
465 for (i = 1; i <= n; i++) |
|
466 { for (t = A_ptr[i]; t < A_ptr[i+1]; t++) |
|
467 { j = A_ind[t]; |
|
468 adjncy[--xadj[i]] = j, adjncy[--xadj[j]] = i; |
|
469 } |
|
470 } |
|
471 /* call the main minimimum degree ordering routine */ |
|
472 genqmd(&n, xadj, adjncy, P_per, P_per + n, deg, marker, rchset, |
|
473 nbrhd, qsize, qlink, &nofsub); |
|
474 /* make sure that permutation matrix P is correct */ |
|
475 for (i = 1; i <= n; i++) |
|
476 { j = P_per[i]; |
|
477 xassert(1 <= j && j <= n); |
|
478 xassert(P_per[n+j] == i); |
|
479 } |
|
480 /* free working arrays */ |
|
481 xfree(xadj); |
|
482 xfree(adjncy); |
|
483 xfree(deg); |
|
484 xfree(marker); |
|
485 xfree(rchset); |
|
486 xfree(nbrhd); |
|
487 xfree(qsize); |
|
488 xfree(qlink); |
|
489 return; |
|
490 } |
|
491 |
|
492 /**********************************************************************/ |
|
493 |
|
494 void amd_order1(int n, int A_ptr[], int A_ind[], int P_per[]) |
|
495 { /* approximate minimum degree ordering (AMD) */ |
|
496 int k, ret; |
|
497 double Control[AMD_CONTROL], Info[AMD_INFO]; |
|
498 /* get the default parameters */ |
|
499 amd_defaults(Control); |
|
500 #if 0 |
|
501 /* and print them */ |
|
502 amd_control(Control); |
|
503 #endif |
|
504 /* make all indices 0-based */ |
|
505 for (k = 1; k < A_ptr[n+1]; k++) A_ind[k]--; |
|
506 for (k = 1; k <= n+1; k++) A_ptr[k]--; |
|
507 /* call the ordering routine */ |
|
508 ret = amd_order(n, &A_ptr[1], &A_ind[1], &P_per[1], Control, Info) |
|
509 ; |
|
510 #if 0 |
|
511 amd_info(Info); |
|
512 #endif |
|
513 xassert(ret == AMD_OK || ret == AMD_OK_BUT_JUMBLED); |
|
514 /* retsore 1-based indices */ |
|
515 for (k = 1; k <= n+1; k++) A_ptr[k]++; |
|
516 for (k = 1; k < A_ptr[n+1]; k++) A_ind[k]++; |
|
517 /* patch up permutation matrix */ |
|
518 memset(&P_per[n+1], 0, n * sizeof(int)); |
|
519 for (k = 1; k <= n; k++) |
|
520 { P_per[k]++; |
|
521 xassert(1 <= P_per[k] && P_per[k] <= n); |
|
522 xassert(P_per[n+P_per[k]] == 0); |
|
523 P_per[n+P_per[k]] = k; |
|
524 } |
|
525 return; |
|
526 } |
|
527 |
|
528 /**********************************************************************/ |
|
529 |
|
530 static void *allocate(size_t n, size_t size) |
|
531 { void *ptr; |
|
532 ptr = xcalloc(n, size); |
|
533 memset(ptr, 0, n * size); |
|
534 return ptr; |
|
535 } |
|
536 |
|
537 static void release(void *ptr) |
|
538 { xfree(ptr); |
|
539 return; |
|
540 } |
|
541 |
|
542 void symamd_ord(int n, int A_ptr[], int A_ind[], int P_per[]) |
|
543 { /* approximate minimum degree ordering (SYMAMD) */ |
|
544 int k, ok; |
|
545 int stats[COLAMD_STATS]; |
|
546 /* make all indices 0-based */ |
|
547 for (k = 1; k < A_ptr[n+1]; k++) A_ind[k]--; |
|
548 for (k = 1; k <= n+1; k++) A_ptr[k]--; |
|
549 /* call the ordering routine */ |
|
550 ok = symamd(n, &A_ind[1], &A_ptr[1], &P_per[1], NULL, stats, |
|
551 allocate, release); |
|
552 #if 0 |
|
553 symamd_report(stats); |
|
554 #endif |
|
555 xassert(ok); |
|
556 /* restore 1-based indices */ |
|
557 for (k = 1; k <= n+1; k++) A_ptr[k]++; |
|
558 for (k = 1; k < A_ptr[n+1]; k++) A_ind[k]++; |
|
559 /* patch up permutation matrix */ |
|
560 memset(&P_per[n+1], 0, n * sizeof(int)); |
|
561 for (k = 1; k <= n; k++) |
|
562 { P_per[k]++; |
|
563 xassert(1 <= P_per[k] && P_per[k] <= n); |
|
564 xassert(P_per[n+P_per[k]] == 0); |
|
565 P_per[n+P_per[k]] = k; |
|
566 } |
|
567 return; |
|
568 } |
|
569 |
|
570 /*---------------------------------------------------------------------- |
|
571 -- chol_symbolic - compute Cholesky factorization (symbolic phase). |
|
572 -- |
|
573 -- *Synopsis* |
|
574 -- |
|
575 -- #include "glpmat.h" |
|
576 -- int *chol_symbolic(int n, int A_ptr[], int A_ind[], int U_ptr[]); |
|
577 -- |
|
578 -- *Description* |
|
579 -- |
|
580 -- The routine chol_symbolic implements the symbolic phase of Cholesky |
|
581 -- factorization A = U'*U, where A is a given sparse symmetric positive |
|
582 -- definite matrix, U is a resultant upper triangular factor, U' is a |
|
583 -- matrix transposed to U. |
|
584 -- |
|
585 -- The parameter n is the order of matrices A and U. |
|
586 -- |
|
587 -- The pattern of the given matrix A is specified on entry in the arrays |
|
588 -- A_ptr and A_ind in storage-by-rows format. Only the upper triangular |
|
589 -- part without diagonal elements (which all are assumed to be non-zero) |
|
590 -- should be specified as if A were upper triangular. The arrays A_ptr |
|
591 -- and A_ind are not changed on exit. |
|
592 -- |
|
593 -- The pattern of the matrix U without diagonal elements (which all are |
|
594 -- assumed to be non-zero) is stored on exit from the routine in the |
|
595 -- arrays U_ptr and U_ind in storage-by-rows format. The array U_ptr |
|
596 -- should be allocated on entry, however, its content is ignored. The |
|
597 -- array U_ind is allocated by the routine which returns a pointer to it |
|
598 -- on exit. |
|
599 -- |
|
600 -- *Returns* |
|
601 -- |
|
602 -- The routine returns a pointer to the array U_ind. |
|
603 -- |
|
604 -- *Method* |
|
605 -- |
|
606 -- The routine chol_symbolic computes the pattern of the matrix U in a |
|
607 -- row-wise manner. No pivoting is used. |
|
608 -- |
|
609 -- It is known that to compute the pattern of row k of the matrix U we |
|
610 -- need to merge the pattern of row k of the matrix A and the patterns |
|
611 -- of each row i of U, where u[i,k] is non-zero (these rows are already |
|
612 -- computed and placed above row k). |
|
613 -- |
|
614 -- However, to reduce the number of rows to be merged the routine uses |
|
615 -- an advanced algorithm proposed in: |
|
616 -- |
|
617 -- D.J.Rose, R.E.Tarjan, and G.S.Lueker. Algorithmic aspects of vertex |
|
618 -- elimination on graphs. SIAM J. Comput. 5, 1976, 266-83. |
|
619 -- |
|
620 -- The authors of the cited paper show that we have the same result if |
|
621 -- we merge row k of the matrix A and such rows of the matrix U (among |
|
622 -- rows 1, ..., k-1) whose leftmost non-diagonal non-zero element is |
|
623 -- placed in k-th column. This feature signficantly reduces the number |
|
624 -- of rows to be merged, especially on the final steps, where rows of |
|
625 -- the matrix U become quite dense. |
|
626 -- |
|
627 -- To determine rows, which should be merged on k-th step, for a fixed |
|
628 -- time the routine uses linked lists of row numbers of the matrix U. |
|
629 -- Location head[k] contains the number of a first row, whose leftmost |
|
630 -- non-diagonal non-zero element is placed in column k, and location |
|
631 -- next[i] contains the number of a next row with the same property as |
|
632 -- row i. */ |
|
633 |
|
634 int *chol_symbolic(int n, int A_ptr[], int A_ind[], int U_ptr[]) |
|
635 { int i, j, k, t, len, size, beg, end, min_j, *U_ind, *head, *next, |
|
636 *ind, *map, *temp; |
|
637 /* initially we assume that on computing the pattern of U fill-in |
|
638 will double the number of non-zeros in A */ |
|
639 size = A_ptr[n+1] - 1; |
|
640 if (size < n) size = n; |
|
641 size += size; |
|
642 U_ind = xcalloc(1+size, sizeof(int)); |
|
643 /* allocate and initialize working arrays */ |
|
644 head = xcalloc(1+n, sizeof(int)); |
|
645 for (i = 1; i <= n; i++) head[i] = 0; |
|
646 next = xcalloc(1+n, sizeof(int)); |
|
647 ind = xcalloc(1+n, sizeof(int)); |
|
648 map = xcalloc(1+n, sizeof(int)); |
|
649 for (j = 1; j <= n; j++) map[j] = 0; |
|
650 /* compute the pattern of matrix U */ |
|
651 U_ptr[1] = 1; |
|
652 for (k = 1; k <= n; k++) |
|
653 { /* compute the pattern of k-th row of U, which is the union of |
|
654 k-th row of A and those rows of U (among 1, ..., k-1) whose |
|
655 leftmost non-diagonal non-zero is placed in k-th column */ |
|
656 /* (ind) := (k-th row of A) */ |
|
657 len = A_ptr[k+1] - A_ptr[k]; |
|
658 memcpy(&ind[1], &A_ind[A_ptr[k]], len * sizeof(int)); |
|
659 for (t = 1; t <= len; t++) |
|
660 { j = ind[t]; |
|
661 xassert(k < j && j <= n); |
|
662 map[j] = 1; |
|
663 } |
|
664 /* walk through rows of U whose leftmost non-diagonal non-zero |
|
665 is placed in k-th column */ |
|
666 for (i = head[k]; i != 0; i = next[i]) |
|
667 { /* (ind) := (ind) union (i-th row of U) */ |
|
668 beg = U_ptr[i], end = U_ptr[i+1]; |
|
669 for (t = beg; t < end; t++) |
|
670 { j = U_ind[t]; |
|
671 if (j > k && !map[j]) ind[++len] = j, map[j] = 1; |
|
672 } |
|
673 } |
|
674 /* now (ind) is the pattern of k-th row of U */ |
|
675 U_ptr[k+1] = U_ptr[k] + len; |
|
676 /* at least (U_ptr[k+1] - 1) locations should be available in |
|
677 the array U_ind */ |
|
678 if (U_ptr[k+1] - 1 > size) |
|
679 { temp = U_ind; |
|
680 size += size; |
|
681 U_ind = xcalloc(1+size, sizeof(int)); |
|
682 memcpy(&U_ind[1], &temp[1], (U_ptr[k] - 1) * sizeof(int)); |
|
683 xfree(temp); |
|
684 } |
|
685 xassert(U_ptr[k+1] - 1 <= size); |
|
686 /* (k-th row of U) := (ind) */ |
|
687 memcpy(&U_ind[U_ptr[k]], &ind[1], len * sizeof(int)); |
|
688 /* determine column index of leftmost non-diagonal non-zero in |
|
689 k-th row of U and clear the row pattern map */ |
|
690 min_j = n + 1; |
|
691 for (t = 1; t <= len; t++) |
|
692 { j = ind[t], map[j] = 0; |
|
693 if (min_j > j) min_j = j; |
|
694 } |
|
695 /* include k-th row into corresponding linked list */ |
|
696 if (min_j <= n) next[k] = head[min_j], head[min_j] = k; |
|
697 } |
|
698 /* free working arrays */ |
|
699 xfree(head); |
|
700 xfree(next); |
|
701 xfree(ind); |
|
702 xfree(map); |
|
703 /* reallocate the array U_ind to free unused locations */ |
|
704 temp = U_ind; |
|
705 size = U_ptr[n+1] - 1; |
|
706 U_ind = xcalloc(1+size, sizeof(int)); |
|
707 memcpy(&U_ind[1], &temp[1], size * sizeof(int)); |
|
708 xfree(temp); |
|
709 return U_ind; |
|
710 } |
|
711 |
|
712 /*---------------------------------------------------------------------- |
|
713 -- chol_numeric - compute Cholesky factorization (numeric phase). |
|
714 -- |
|
715 -- *Synopsis* |
|
716 -- |
|
717 -- #include "glpmat.h" |
|
718 -- int chol_numeric(int n, |
|
719 -- int A_ptr[], int A_ind[], double A_val[], double A_diag[], |
|
720 -- int U_ptr[], int U_ind[], double U_val[], double U_diag[]); |
|
721 -- |
|
722 -- *Description* |
|
723 -- |
|
724 -- The routine chol_symbolic implements the numeric phase of Cholesky |
|
725 -- factorization A = U'*U, where A is a given sparse symmetric positive |
|
726 -- definite matrix, U is a resultant upper triangular factor, U' is a |
|
727 -- matrix transposed to U. |
|
728 -- |
|
729 -- The parameter n is the order of matrices A and U. |
|
730 -- |
|
731 -- Upper triangular part of the matrix A without diagonal elements is |
|
732 -- specified in the arrays A_ptr, A_ind, and A_val in storage-by-rows |
|
733 -- format. Diagonal elements of A are specified in the array A_diag, |
|
734 -- where A_diag[0] is not used, A_diag[i] = a[i,i] for i = 1, ..., n. |
|
735 -- The arrays A_ptr, A_ind, A_val, and A_diag are not changed on exit. |
|
736 -- |
|
737 -- The pattern of the matrix U without diagonal elements (previously |
|
738 -- computed with the routine chol_symbolic) is specified in the arrays |
|
739 -- U_ptr and U_ind, which are not changed on exit. Numeric values of |
|
740 -- non-diagonal elements of U are stored in corresponding locations of |
|
741 -- the array U_val, and values of diagonal elements of U are stored in |
|
742 -- locations U_diag[1], ..., U_diag[n]. |
|
743 -- |
|
744 -- *Returns* |
|
745 -- |
|
746 -- The routine returns the number of non-positive diagonal elements of |
|
747 -- the matrix U which have been replaced by a huge positive number (see |
|
748 -- the method description below). Zero return code means the matrix A |
|
749 -- has been successfully factorized. |
|
750 -- |
|
751 -- *Method* |
|
752 -- |
|
753 -- The routine chol_numeric computes the matrix U in a row-wise manner |
|
754 -- using standard gaussian elimination technique. No pivoting is used. |
|
755 -- |
|
756 -- Initially the routine sets U = A, and before k-th elimination step |
|
757 -- the matrix U is the following: |
|
758 -- |
|
759 -- 1 k n |
|
760 -- 1 x x x x x x x x x x |
|
761 -- . x x x x x x x x x |
|
762 -- . . x x x x x x x x |
|
763 -- . . . x x x x x x x |
|
764 -- k . . . . * * * * * * |
|
765 -- . . . . * * * * * * |
|
766 -- . . . . * * * * * * |
|
767 -- . . . . * * * * * * |
|
768 -- . . . . * * * * * * |
|
769 -- n . . . . * * * * * * |
|
770 -- |
|
771 -- where 'x' are elements of already computed rows, '*' are elements of |
|
772 -- the active submatrix. (Note that the lower triangular part of the |
|
773 -- active submatrix being symmetric is not stored and diagonal elements |
|
774 -- are stored separately in the array U_diag.) |
|
775 -- |
|
776 -- The matrix A is assumed to be positive definite. However, if it is |
|
777 -- close to semi-definite, on some elimination step a pivot u[k,k] may |
|
778 -- happen to be non-positive due to round-off errors. In this case the |
|
779 -- routine uses a technique proposed in: |
|
780 -- |
|
781 -- S.J.Wright. The Cholesky factorization in interior-point and barrier |
|
782 -- methods. Preprint MCS-P600-0596, Mathematics and Computer Science |
|
783 -- Division, Argonne National Laboratory, Argonne, Ill., May 1996. |
|
784 -- |
|
785 -- The routine just replaces non-positive u[k,k] by a huge positive |
|
786 -- number. This involves non-diagonal elements in k-th row of U to be |
|
787 -- close to zero that, in turn, involves k-th component of a solution |
|
788 -- vector to be close to zero. Note, however, that this technique works |
|
789 -- only if the system A*x = b is consistent. */ |
|
790 |
|
791 int chol_numeric(int n, |
|
792 int A_ptr[], int A_ind[], double A_val[], double A_diag[], |
|
793 int U_ptr[], int U_ind[], double U_val[], double U_diag[]) |
|
794 { int i, j, k, t, t1, beg, end, beg1, end1, count = 0; |
|
795 double ukk, uki, *work; |
|
796 work = xcalloc(1+n, sizeof(double)); |
|
797 for (j = 1; j <= n; j++) work[j] = 0.0; |
|
798 /* U := (upper triangle of A) */ |
|
799 /* note that the upper traingle of A is a subset of U */ |
|
800 for (i = 1; i <= n; i++) |
|
801 { beg = A_ptr[i], end = A_ptr[i+1]; |
|
802 for (t = beg; t < end; t++) |
|
803 j = A_ind[t], work[j] = A_val[t]; |
|
804 beg = U_ptr[i], end = U_ptr[i+1]; |
|
805 for (t = beg; t < end; t++) |
|
806 j = U_ind[t], U_val[t] = work[j], work[j] = 0.0; |
|
807 U_diag[i] = A_diag[i]; |
|
808 } |
|
809 /* main elimination loop */ |
|
810 for (k = 1; k <= n; k++) |
|
811 { /* transform k-th row of U */ |
|
812 ukk = U_diag[k]; |
|
813 if (ukk > 0.0) |
|
814 U_diag[k] = ukk = sqrt(ukk); |
|
815 else |
|
816 U_diag[k] = ukk = DBL_MAX, count++; |
|
817 /* (work) := (transformed k-th row) */ |
|
818 beg = U_ptr[k], end = U_ptr[k+1]; |
|
819 for (t = beg; t < end; t++) |
|
820 work[U_ind[t]] = (U_val[t] /= ukk); |
|
821 /* transform other rows of U */ |
|
822 for (t = beg; t < end; t++) |
|
823 { i = U_ind[t]; |
|
824 xassert(i > k); |
|
825 /* (i-th row) := (i-th row) - u[k,i] * (k-th row) */ |
|
826 uki = work[i]; |
|
827 beg1 = U_ptr[i], end1 = U_ptr[i+1]; |
|
828 for (t1 = beg1; t1 < end1; t1++) |
|
829 U_val[t1] -= uki * work[U_ind[t1]]; |
|
830 U_diag[i] -= uki * uki; |
|
831 } |
|
832 /* (work) := 0 */ |
|
833 for (t = beg; t < end; t++) |
|
834 work[U_ind[t]] = 0.0; |
|
835 } |
|
836 xfree(work); |
|
837 return count; |
|
838 } |
|
839 |
|
840 /*---------------------------------------------------------------------- |
|
841 -- u_solve - solve upper triangular system U*x = b. |
|
842 -- |
|
843 -- *Synopsis* |
|
844 -- |
|
845 -- #include "glpmat.h" |
|
846 -- void u_solve(int n, int U_ptr[], int U_ind[], double U_val[], |
|
847 -- double U_diag[], double x[]); |
|
848 -- |
|
849 -- *Description* |
|
850 -- |
|
851 -- The routine u_solve solves an linear system U*x = b, where U is an |
|
852 -- upper triangular matrix. |
|
853 -- |
|
854 -- The parameter n is the order of matrix U. |
|
855 -- |
|
856 -- The matrix U without diagonal elements is specified in the arrays |
|
857 -- U_ptr, U_ind, and U_val in storage-by-rows format. Diagonal elements |
|
858 -- of U are specified in the array U_diag, where U_diag[0] is not used, |
|
859 -- U_diag[i] = u[i,i] for i = 1, ..., n. All these four arrays are not |
|
860 -- changed on exit. |
|
861 -- |
|
862 -- The right-hand side vector b is specified on entry in the array x, |
|
863 -- where x[0] is not used, and x[i] = b[i] for i = 1, ..., n. On exit |
|
864 -- the routine stores computed components of the vector of unknowns x |
|
865 -- in the array x in the same manner. */ |
|
866 |
|
867 void u_solve(int n, int U_ptr[], int U_ind[], double U_val[], |
|
868 double U_diag[], double x[]) |
|
869 { int i, t, beg, end; |
|
870 double temp; |
|
871 for (i = n; i >= 1; i--) |
|
872 { temp = x[i]; |
|
873 beg = U_ptr[i], end = U_ptr[i+1]; |
|
874 for (t = beg; t < end; t++) |
|
875 temp -= U_val[t] * x[U_ind[t]]; |
|
876 xassert(U_diag[i] != 0.0); |
|
877 x[i] = temp / U_diag[i]; |
|
878 } |
|
879 return; |
|
880 } |
|
881 |
|
882 /*---------------------------------------------------------------------- |
|
883 -- ut_solve - solve lower triangular system U'*x = b. |
|
884 -- |
|
885 -- *Synopsis* |
|
886 -- |
|
887 -- #include "glpmat.h" |
|
888 -- void ut_solve(int n, int U_ptr[], int U_ind[], double U_val[], |
|
889 -- double U_diag[], double x[]); |
|
890 -- |
|
891 -- *Description* |
|
892 -- |
|
893 -- The routine ut_solve solves an linear system U'*x = b, where U is a |
|
894 -- matrix transposed to an upper triangular matrix. |
|
895 -- |
|
896 -- The parameter n is the order of matrix U. |
|
897 -- |
|
898 -- The matrix U without diagonal elements is specified in the arrays |
|
899 -- U_ptr, U_ind, and U_val in storage-by-rows format. Diagonal elements |
|
900 -- of U are specified in the array U_diag, where U_diag[0] is not used, |
|
901 -- U_diag[i] = u[i,i] for i = 1, ..., n. All these four arrays are not |
|
902 -- changed on exit. |
|
903 -- |
|
904 -- The right-hand side vector b is specified on entry in the array x, |
|
905 -- where x[0] is not used, and x[i] = b[i] for i = 1, ..., n. On exit |
|
906 -- the routine stores computed components of the vector of unknowns x |
|
907 -- in the array x in the same manner. */ |
|
908 |
|
909 void ut_solve(int n, int U_ptr[], int U_ind[], double U_val[], |
|
910 double U_diag[], double x[]) |
|
911 { int i, t, beg, end; |
|
912 double temp; |
|
913 for (i = 1; i <= n; i++) |
|
914 { xassert(U_diag[i] != 0.0); |
|
915 temp = (x[i] /= U_diag[i]); |
|
916 if (temp == 0.0) continue; |
|
917 beg = U_ptr[i], end = U_ptr[i+1]; |
|
918 for (t = beg; t < end; t++) |
|
919 x[U_ind[t]] -= U_val[t] * temp; |
|
920 } |
|
921 return; |
|
922 } |
|
923 |
|
924 /* eof */ |