rev |
line source |
alpar@9
|
1 /* ========================================================================= */
|
alpar@9
|
2 /* === AMD_2 =============================================================== */
|
alpar@9
|
3 /* ========================================================================= */
|
alpar@9
|
4
|
alpar@9
|
5 /* ------------------------------------------------------------------------- */
|
alpar@9
|
6 /* AMD, Copyright (c) Timothy A. Davis, */
|
alpar@9
|
7 /* Patrick R. Amestoy, and Iain S. Duff. See ../README.txt for License. */
|
alpar@9
|
8 /* email: davis at cise.ufl.edu CISE Department, Univ. of Florida. */
|
alpar@9
|
9 /* web: http://www.cise.ufl.edu/research/sparse/amd */
|
alpar@9
|
10 /* ------------------------------------------------------------------------- */
|
alpar@9
|
11
|
alpar@9
|
12 /* AMD_2: performs the AMD ordering on a symmetric sparse matrix A, followed
|
alpar@9
|
13 * by a postordering (via depth-first search) of the assembly tree using the
|
alpar@9
|
14 * AMD_postorder routine.
|
alpar@9
|
15 */
|
alpar@9
|
16
|
alpar@9
|
17 #include "amd_internal.h"
|
alpar@9
|
18
|
alpar@9
|
19 /* ========================================================================= */
|
alpar@9
|
20 /* === clear_flag ========================================================== */
|
alpar@9
|
21 /* ========================================================================= */
|
alpar@9
|
22
|
alpar@9
|
23 static Int clear_flag (Int wflg, Int wbig, Int W [ ], Int n)
|
alpar@9
|
24 {
|
alpar@9
|
25 Int x ;
|
alpar@9
|
26 if (wflg < 2 || wflg >= wbig)
|
alpar@9
|
27 {
|
alpar@9
|
28 for (x = 0 ; x < n ; x++)
|
alpar@9
|
29 {
|
alpar@9
|
30 if (W [x] != 0) W [x] = 1 ;
|
alpar@9
|
31 }
|
alpar@9
|
32 wflg = 2 ;
|
alpar@9
|
33 }
|
alpar@9
|
34 /* at this point, W [0..n-1] < wflg holds */
|
alpar@9
|
35 return (wflg) ;
|
alpar@9
|
36 }
|
alpar@9
|
37
|
alpar@9
|
38
|
alpar@9
|
39 /* ========================================================================= */
|
alpar@9
|
40 /* === AMD_2 =============================================================== */
|
alpar@9
|
41 /* ========================================================================= */
|
alpar@9
|
42
|
alpar@9
|
43 GLOBAL void AMD_2
|
alpar@9
|
44 (
|
alpar@9
|
45 Int n, /* A is n-by-n, where n > 0 */
|
alpar@9
|
46 Int Pe [ ], /* Pe [0..n-1]: index in Iw of row i on input */
|
alpar@9
|
47 Int Iw [ ], /* workspace of size iwlen. Iw [0..pfree-1]
|
alpar@9
|
48 * holds the matrix on input */
|
alpar@9
|
49 Int Len [ ], /* Len [0..n-1]: length for row/column i on input */
|
alpar@9
|
50 Int iwlen, /* length of Iw. iwlen >= pfree + n */
|
alpar@9
|
51 Int pfree, /* Iw [pfree ... iwlen-1] is empty on input */
|
alpar@9
|
52
|
alpar@9
|
53 /* 7 size-n workspaces, not defined on input: */
|
alpar@9
|
54 Int Nv [ ], /* the size of each supernode on output */
|
alpar@9
|
55 Int Next [ ], /* the output inverse permutation */
|
alpar@9
|
56 Int Last [ ], /* the output permutation */
|
alpar@9
|
57 Int Head [ ],
|
alpar@9
|
58 Int Elen [ ], /* the size columns of L for each supernode */
|
alpar@9
|
59 Int Degree [ ],
|
alpar@9
|
60 Int W [ ],
|
alpar@9
|
61
|
alpar@9
|
62 /* control parameters and output statistics */
|
alpar@9
|
63 double Control [ ], /* array of size AMD_CONTROL */
|
alpar@9
|
64 double Info [ ] /* array of size AMD_INFO */
|
alpar@9
|
65 )
|
alpar@9
|
66 {
|
alpar@9
|
67
|
alpar@9
|
68 /*
|
alpar@9
|
69 * Given a representation of the nonzero pattern of a symmetric matrix, A,
|
alpar@9
|
70 * (excluding the diagonal) perform an approximate minimum (UMFPACK/MA38-style)
|
alpar@9
|
71 * degree ordering to compute a pivot order such that the introduction of
|
alpar@9
|
72 * nonzeros (fill-in) in the Cholesky factors A = LL' is kept low. At each
|
alpar@9
|
73 * step, the pivot selected is the one with the minimum UMFAPACK/MA38-style
|
alpar@9
|
74 * upper-bound on the external degree. This routine can optionally perform
|
alpar@9
|
75 * aggresive absorption (as done by MC47B in the Harwell Subroutine
|
alpar@9
|
76 * Library).
|
alpar@9
|
77 *
|
alpar@9
|
78 * The approximate degree algorithm implemented here is the symmetric analog of
|
alpar@9
|
79 * the degree update algorithm in MA38 and UMFPACK (the Unsymmetric-pattern
|
alpar@9
|
80 * MultiFrontal PACKage, both by Davis and Duff). The routine is based on the
|
alpar@9
|
81 * MA27 minimum degree ordering algorithm by Iain Duff and John Reid.
|
alpar@9
|
82 *
|
alpar@9
|
83 * This routine is a translation of the original AMDBAR and MC47B routines,
|
alpar@9
|
84 * in Fortran, with the following modifications:
|
alpar@9
|
85 *
|
alpar@9
|
86 * (1) dense rows/columns are removed prior to ordering the matrix, and placed
|
alpar@9
|
87 * last in the output order. The presence of a dense row/column can
|
alpar@9
|
88 * increase the ordering time by up to O(n^2), unless they are removed
|
alpar@9
|
89 * prior to ordering.
|
alpar@9
|
90 *
|
alpar@9
|
91 * (2) the minimum degree ordering is followed by a postordering (depth-first
|
alpar@9
|
92 * search) of the assembly tree. Note that mass elimination (discussed
|
alpar@9
|
93 * below) combined with the approximate degree update can lead to the mass
|
alpar@9
|
94 * elimination of nodes with lower exact degree than the current pivot
|
alpar@9
|
95 * element. No additional fill-in is caused in the representation of the
|
alpar@9
|
96 * Schur complement. The mass-eliminated nodes merge with the current
|
alpar@9
|
97 * pivot element. They are ordered prior to the current pivot element.
|
alpar@9
|
98 * Because they can have lower exact degree than the current element, the
|
alpar@9
|
99 * merger of two or more of these nodes in the current pivot element can
|
alpar@9
|
100 * lead to a single element that is not a "fundamental supernode". The
|
alpar@9
|
101 * diagonal block can have zeros in it. Thus, the assembly tree used here
|
alpar@9
|
102 * is not guaranteed to be the precise supernodal elemination tree (with
|
alpar@9
|
103 * "funadmental" supernodes), and the postordering performed by this
|
alpar@9
|
104 * routine is not guaranteed to be a precise postordering of the
|
alpar@9
|
105 * elimination tree.
|
alpar@9
|
106 *
|
alpar@9
|
107 * (3) input parameters are added, to control aggressive absorption and the
|
alpar@9
|
108 * detection of "dense" rows/columns of A.
|
alpar@9
|
109 *
|
alpar@9
|
110 * (4) additional statistical information is returned, such as the number of
|
alpar@9
|
111 * nonzeros in L, and the flop counts for subsequent LDL' and LU
|
alpar@9
|
112 * factorizations. These are slight upper bounds, because of the mass
|
alpar@9
|
113 * elimination issue discussed above.
|
alpar@9
|
114 *
|
alpar@9
|
115 * (5) additional routines are added to interface this routine to MATLAB
|
alpar@9
|
116 * to provide a simple C-callable user-interface, to check inputs for
|
alpar@9
|
117 * errors, compute the symmetry of the pattern of A and the number of
|
alpar@9
|
118 * nonzeros in each row/column of A+A', to compute the pattern of A+A',
|
alpar@9
|
119 * to perform the assembly tree postordering, and to provide debugging
|
alpar@9
|
120 * ouput. Many of these functions are also provided by the Fortran
|
alpar@9
|
121 * Harwell Subroutine Library routine MC47A.
|
alpar@9
|
122 *
|
alpar@9
|
123 * (6) both int and UF_long versions are provided. In the descriptions below
|
alpar@9
|
124 * and integer is and int or UF_long depending on which version is
|
alpar@9
|
125 * being used.
|
alpar@9
|
126
|
alpar@9
|
127 **********************************************************************
|
alpar@9
|
128 ***** CAUTION: ARGUMENTS ARE NOT CHECKED FOR ERRORS ON INPUT. ******
|
alpar@9
|
129 **********************************************************************
|
alpar@9
|
130 ** If you want error checking, a more versatile input format, and a **
|
alpar@9
|
131 ** simpler user interface, use amd_order or amd_l_order instead. **
|
alpar@9
|
132 ** This routine is not meant to be user-callable. **
|
alpar@9
|
133 **********************************************************************
|
alpar@9
|
134
|
alpar@9
|
135 * ----------------------------------------------------------------------------
|
alpar@9
|
136 * References:
|
alpar@9
|
137 * ----------------------------------------------------------------------------
|
alpar@9
|
138 *
|
alpar@9
|
139 * [1] Timothy A. Davis and Iain Duff, "An unsymmetric-pattern multifrontal
|
alpar@9
|
140 * method for sparse LU factorization", SIAM J. Matrix Analysis and
|
alpar@9
|
141 * Applications, vol. 18, no. 1, pp. 140-158. Discusses UMFPACK / MA38,
|
alpar@9
|
142 * which first introduced the approximate minimum degree used by this
|
alpar@9
|
143 * routine.
|
alpar@9
|
144 *
|
alpar@9
|
145 * [2] Patrick Amestoy, Timothy A. Davis, and Iain S. Duff, "An approximate
|
alpar@9
|
146 * minimum degree ordering algorithm," SIAM J. Matrix Analysis and
|
alpar@9
|
147 * Applications, vol. 17, no. 4, pp. 886-905, 1996. Discusses AMDBAR and
|
alpar@9
|
148 * MC47B, which are the Fortran versions of this routine.
|
alpar@9
|
149 *
|
alpar@9
|
150 * [3] Alan George and Joseph Liu, "The evolution of the minimum degree
|
alpar@9
|
151 * ordering algorithm," SIAM Review, vol. 31, no. 1, pp. 1-19, 1989.
|
alpar@9
|
152 * We list below the features mentioned in that paper that this code
|
alpar@9
|
153 * includes:
|
alpar@9
|
154 *
|
alpar@9
|
155 * mass elimination:
|
alpar@9
|
156 * Yes. MA27 relied on supervariable detection for mass elimination.
|
alpar@9
|
157 *
|
alpar@9
|
158 * indistinguishable nodes:
|
alpar@9
|
159 * Yes (we call these "supervariables"). This was also in the MA27
|
alpar@9
|
160 * code - although we modified the method of detecting them (the
|
alpar@9
|
161 * previous hash was the true degree, which we no longer keep track
|
alpar@9
|
162 * of). A supervariable is a set of rows with identical nonzero
|
alpar@9
|
163 * pattern. All variables in a supervariable are eliminated together.
|
alpar@9
|
164 * Each supervariable has as its numerical name that of one of its
|
alpar@9
|
165 * variables (its principal variable).
|
alpar@9
|
166 *
|
alpar@9
|
167 * quotient graph representation:
|
alpar@9
|
168 * Yes. We use the term "element" for the cliques formed during
|
alpar@9
|
169 * elimination. This was also in the MA27 code. The algorithm can
|
alpar@9
|
170 * operate in place, but it will work more efficiently if given some
|
alpar@9
|
171 * "elbow room."
|
alpar@9
|
172 *
|
alpar@9
|
173 * element absorption:
|
alpar@9
|
174 * Yes. This was also in the MA27 code.
|
alpar@9
|
175 *
|
alpar@9
|
176 * external degree:
|
alpar@9
|
177 * Yes. The MA27 code was based on the true degree.
|
alpar@9
|
178 *
|
alpar@9
|
179 * incomplete degree update and multiple elimination:
|
alpar@9
|
180 * No. This was not in MA27, either. Our method of degree update
|
alpar@9
|
181 * within MC47B is element-based, not variable-based. It is thus
|
alpar@9
|
182 * not well-suited for use with incomplete degree update or multiple
|
alpar@9
|
183 * elimination.
|
alpar@9
|
184 *
|
alpar@9
|
185 * Authors, and Copyright (C) 2004 by:
|
alpar@9
|
186 * Timothy A. Davis, Patrick Amestoy, Iain S. Duff, John K. Reid.
|
alpar@9
|
187 *
|
alpar@9
|
188 * Acknowledgements: This work (and the UMFPACK package) was supported by the
|
alpar@9
|
189 * National Science Foundation (ASC-9111263, DMS-9223088, and CCR-0203270).
|
alpar@9
|
190 * The UMFPACK/MA38 approximate degree update algorithm, the unsymmetric analog
|
alpar@9
|
191 * which forms the basis of AMD, was developed while Tim Davis was supported by
|
alpar@9
|
192 * CERFACS (Toulouse, France) in a post-doctoral position. This C version, and
|
alpar@9
|
193 * the etree postorder, were written while Tim Davis was on sabbatical at
|
alpar@9
|
194 * Stanford University and Lawrence Berkeley National Laboratory.
|
alpar@9
|
195
|
alpar@9
|
196 * ----------------------------------------------------------------------------
|
alpar@9
|
197 * INPUT ARGUMENTS (unaltered):
|
alpar@9
|
198 * ----------------------------------------------------------------------------
|
alpar@9
|
199
|
alpar@9
|
200 * n: The matrix order. Restriction: n >= 1.
|
alpar@9
|
201 *
|
alpar@9
|
202 * iwlen: The size of the Iw array. On input, the matrix is stored in
|
alpar@9
|
203 * Iw [0..pfree-1]. However, Iw [0..iwlen-1] should be slightly larger
|
alpar@9
|
204 * than what is required to hold the matrix, at least iwlen >= pfree + n.
|
alpar@9
|
205 * Otherwise, excessive compressions will take place. The recommended
|
alpar@9
|
206 * value of iwlen is 1.2 * pfree + n, which is the value used in the
|
alpar@9
|
207 * user-callable interface to this routine (amd_order.c). The algorithm
|
alpar@9
|
208 * will not run at all if iwlen < pfree. Restriction: iwlen >= pfree + n.
|
alpar@9
|
209 * Note that this is slightly more restrictive than the actual minimum
|
alpar@9
|
210 * (iwlen >= pfree), but AMD_2 will be very slow with no elbow room.
|
alpar@9
|
211 * Thus, this routine enforces a bare minimum elbow room of size n.
|
alpar@9
|
212 *
|
alpar@9
|
213 * pfree: On input the tail end of the array, Iw [pfree..iwlen-1], is empty,
|
alpar@9
|
214 * and the matrix is stored in Iw [0..pfree-1]. During execution,
|
alpar@9
|
215 * additional data is placed in Iw, and pfree is modified so that
|
alpar@9
|
216 * Iw [pfree..iwlen-1] is always the unused part of Iw.
|
alpar@9
|
217 *
|
alpar@9
|
218 * Control: A double array of size AMD_CONTROL containing input parameters
|
alpar@9
|
219 * that affect how the ordering is computed. If NULL, then default
|
alpar@9
|
220 * settings are used.
|
alpar@9
|
221 *
|
alpar@9
|
222 * Control [AMD_DENSE] is used to determine whether or not a given input
|
alpar@9
|
223 * row is "dense". A row is "dense" if the number of entries in the row
|
alpar@9
|
224 * exceeds Control [AMD_DENSE] times sqrt (n), except that rows with 16 or
|
alpar@9
|
225 * fewer entries are never considered "dense". To turn off the detection
|
alpar@9
|
226 * of dense rows, set Control [AMD_DENSE] to a negative number, or to a
|
alpar@9
|
227 * number larger than sqrt (n). The default value of Control [AMD_DENSE]
|
alpar@9
|
228 * is AMD_DEFAULT_DENSE, which is defined in amd.h as 10.
|
alpar@9
|
229 *
|
alpar@9
|
230 * Control [AMD_AGGRESSIVE] is used to determine whether or not aggressive
|
alpar@9
|
231 * absorption is to be performed. If nonzero, then aggressive absorption
|
alpar@9
|
232 * is performed (this is the default).
|
alpar@9
|
233
|
alpar@9
|
234 * ----------------------------------------------------------------------------
|
alpar@9
|
235 * INPUT/OUPUT ARGUMENTS:
|
alpar@9
|
236 * ----------------------------------------------------------------------------
|
alpar@9
|
237 *
|
alpar@9
|
238 * Pe: An integer array of size n. On input, Pe [i] is the index in Iw of
|
alpar@9
|
239 * the start of row i. Pe [i] is ignored if row i has no off-diagonal
|
alpar@9
|
240 * entries. Thus Pe [i] must be in the range 0 to pfree-1 for non-empty
|
alpar@9
|
241 * rows.
|
alpar@9
|
242 *
|
alpar@9
|
243 * During execution, it is used for both supervariables and elements:
|
alpar@9
|
244 *
|
alpar@9
|
245 * Principal supervariable i: index into Iw of the description of
|
alpar@9
|
246 * supervariable i. A supervariable represents one or more rows of
|
alpar@9
|
247 * the matrix with identical nonzero pattern. In this case,
|
alpar@9
|
248 * Pe [i] >= 0.
|
alpar@9
|
249 *
|
alpar@9
|
250 * Non-principal supervariable i: if i has been absorbed into another
|
alpar@9
|
251 * supervariable j, then Pe [i] = FLIP (j), where FLIP (j) is defined
|
alpar@9
|
252 * as (-(j)-2). Row j has the same pattern as row i. Note that j
|
alpar@9
|
253 * might later be absorbed into another supervariable j2, in which
|
alpar@9
|
254 * case Pe [i] is still FLIP (j), and Pe [j] = FLIP (j2) which is
|
alpar@9
|
255 * < EMPTY, where EMPTY is defined as (-1) in amd_internal.h.
|
alpar@9
|
256 *
|
alpar@9
|
257 * Unabsorbed element e: the index into Iw of the description of element
|
alpar@9
|
258 * e, if e has not yet been absorbed by a subsequent element. Element
|
alpar@9
|
259 * e is created when the supervariable of the same name is selected as
|
alpar@9
|
260 * the pivot. In this case, Pe [i] >= 0.
|
alpar@9
|
261 *
|
alpar@9
|
262 * Absorbed element e: if element e is absorbed into element e2, then
|
alpar@9
|
263 * Pe [e] = FLIP (e2). This occurs when the pattern of e (which we
|
alpar@9
|
264 * refer to as Le) is found to be a subset of the pattern of e2 (that
|
alpar@9
|
265 * is, Le2). In this case, Pe [i] < EMPTY. If element e is "null"
|
alpar@9
|
266 * (it has no nonzeros outside its pivot block), then Pe [e] = EMPTY,
|
alpar@9
|
267 * and e is the root of an assembly subtree (or the whole tree if
|
alpar@9
|
268 * there is just one such root).
|
alpar@9
|
269 *
|
alpar@9
|
270 * Dense variable i: if i is "dense", then Pe [i] = EMPTY.
|
alpar@9
|
271 *
|
alpar@9
|
272 * On output, Pe holds the assembly tree/forest, which implicitly
|
alpar@9
|
273 * represents a pivot order with identical fill-in as the actual order
|
alpar@9
|
274 * (via a depth-first search of the tree), as follows. If Nv [i] > 0,
|
alpar@9
|
275 * then i represents a node in the assembly tree, and the parent of i is
|
alpar@9
|
276 * Pe [i], or EMPTY if i is a root. If Nv [i] = 0, then (i, Pe [i])
|
alpar@9
|
277 * represents an edge in a subtree, the root of which is a node in the
|
alpar@9
|
278 * assembly tree. Note that i refers to a row/column in the original
|
alpar@9
|
279 * matrix, not the permuted matrix.
|
alpar@9
|
280 *
|
alpar@9
|
281 * Info: A double array of size AMD_INFO. If present, (that is, not NULL),
|
alpar@9
|
282 * then statistics about the ordering are returned in the Info array.
|
alpar@9
|
283 * See amd.h for a description.
|
alpar@9
|
284
|
alpar@9
|
285 * ----------------------------------------------------------------------------
|
alpar@9
|
286 * INPUT/MODIFIED (undefined on output):
|
alpar@9
|
287 * ----------------------------------------------------------------------------
|
alpar@9
|
288 *
|
alpar@9
|
289 * Len: An integer array of size n. On input, Len [i] holds the number of
|
alpar@9
|
290 * entries in row i of the matrix, excluding the diagonal. The contents
|
alpar@9
|
291 * of Len are undefined on output.
|
alpar@9
|
292 *
|
alpar@9
|
293 * Iw: An integer array of size iwlen. On input, Iw [0..pfree-1] holds the
|
alpar@9
|
294 * description of each row i in the matrix. The matrix must be symmetric,
|
alpar@9
|
295 * and both upper and lower triangular parts must be present. The
|
alpar@9
|
296 * diagonal must not be present. Row i is held as follows:
|
alpar@9
|
297 *
|
alpar@9
|
298 * Len [i]: the length of the row i data structure in the Iw array.
|
alpar@9
|
299 * Iw [Pe [i] ... Pe [i] + Len [i] - 1]:
|
alpar@9
|
300 * the list of column indices for nonzeros in row i (simple
|
alpar@9
|
301 * supervariables), excluding the diagonal. All supervariables
|
alpar@9
|
302 * start with one row/column each (supervariable i is just row i).
|
alpar@9
|
303 * If Len [i] is zero on input, then Pe [i] is ignored on input.
|
alpar@9
|
304 *
|
alpar@9
|
305 * Note that the rows need not be in any particular order, and there
|
alpar@9
|
306 * may be empty space between the rows.
|
alpar@9
|
307 *
|
alpar@9
|
308 * During execution, the supervariable i experiences fill-in. This is
|
alpar@9
|
309 * represented by placing in i a list of the elements that cause fill-in
|
alpar@9
|
310 * in supervariable i:
|
alpar@9
|
311 *
|
alpar@9
|
312 * Len [i]: the length of supervariable i in the Iw array.
|
alpar@9
|
313 * Iw [Pe [i] ... Pe [i] + Elen [i] - 1]:
|
alpar@9
|
314 * the list of elements that contain i. This list is kept short
|
alpar@9
|
315 * by removing absorbed elements.
|
alpar@9
|
316 * Iw [Pe [i] + Elen [i] ... Pe [i] + Len [i] - 1]:
|
alpar@9
|
317 * the list of supervariables in i. This list is kept short by
|
alpar@9
|
318 * removing nonprincipal variables, and any entry j that is also
|
alpar@9
|
319 * contained in at least one of the elements (j in Le) in the list
|
alpar@9
|
320 * for i (e in row i).
|
alpar@9
|
321 *
|
alpar@9
|
322 * When supervariable i is selected as pivot, we create an element e of
|
alpar@9
|
323 * the same name (e=i):
|
alpar@9
|
324 *
|
alpar@9
|
325 * Len [e]: the length of element e in the Iw array.
|
alpar@9
|
326 * Iw [Pe [e] ... Pe [e] + Len [e] - 1]:
|
alpar@9
|
327 * the list of supervariables in element e.
|
alpar@9
|
328 *
|
alpar@9
|
329 * An element represents the fill-in that occurs when supervariable i is
|
alpar@9
|
330 * selected as pivot (which represents the selection of row i and all
|
alpar@9
|
331 * non-principal variables whose principal variable is i). We use the
|
alpar@9
|
332 * term Le to denote the set of all supervariables in element e. Absorbed
|
alpar@9
|
333 * supervariables and elements are pruned from these lists when
|
alpar@9
|
334 * computationally convenient.
|
alpar@9
|
335 *
|
alpar@9
|
336 * CAUTION: THE INPUT MATRIX IS OVERWRITTEN DURING COMPUTATION.
|
alpar@9
|
337 * The contents of Iw are undefined on output.
|
alpar@9
|
338
|
alpar@9
|
339 * ----------------------------------------------------------------------------
|
alpar@9
|
340 * OUTPUT (need not be set on input):
|
alpar@9
|
341 * ----------------------------------------------------------------------------
|
alpar@9
|
342 *
|
alpar@9
|
343 * Nv: An integer array of size n. During execution, ABS (Nv [i]) is equal to
|
alpar@9
|
344 * the number of rows that are represented by the principal supervariable
|
alpar@9
|
345 * i. If i is a nonprincipal or dense variable, then Nv [i] = 0.
|
alpar@9
|
346 * Initially, Nv [i] = 1 for all i. Nv [i] < 0 signifies that i is a
|
alpar@9
|
347 * principal variable in the pattern Lme of the current pivot element me.
|
alpar@9
|
348 * After element me is constructed, Nv [i] is set back to a positive
|
alpar@9
|
349 * value.
|
alpar@9
|
350 *
|
alpar@9
|
351 * On output, Nv [i] holds the number of pivots represented by super
|
alpar@9
|
352 * row/column i of the original matrix, or Nv [i] = 0 for non-principal
|
alpar@9
|
353 * rows/columns. Note that i refers to a row/column in the original
|
alpar@9
|
354 * matrix, not the permuted matrix.
|
alpar@9
|
355 *
|
alpar@9
|
356 * Elen: An integer array of size n. See the description of Iw above. At the
|
alpar@9
|
357 * start of execution, Elen [i] is set to zero for all rows i. During
|
alpar@9
|
358 * execution, Elen [i] is the number of elements in the list for
|
alpar@9
|
359 * supervariable i. When e becomes an element, Elen [e] = FLIP (esize) is
|
alpar@9
|
360 * set, where esize is the size of the element (the number of pivots, plus
|
alpar@9
|
361 * the number of nonpivotal entries). Thus Elen [e] < EMPTY.
|
alpar@9
|
362 * Elen (i) = EMPTY set when variable i becomes nonprincipal.
|
alpar@9
|
363 *
|
alpar@9
|
364 * For variables, Elen (i) >= EMPTY holds until just before the
|
alpar@9
|
365 * postordering and permutation vectors are computed. For elements,
|
alpar@9
|
366 * Elen [e] < EMPTY holds.
|
alpar@9
|
367 *
|
alpar@9
|
368 * On output, Elen [i] is the degree of the row/column in the Cholesky
|
alpar@9
|
369 * factorization of the permuted matrix, corresponding to the original row
|
alpar@9
|
370 * i, if i is a super row/column. It is equal to EMPTY if i is
|
alpar@9
|
371 * non-principal. Note that i refers to a row/column in the original
|
alpar@9
|
372 * matrix, not the permuted matrix.
|
alpar@9
|
373 *
|
alpar@9
|
374 * Note that the contents of Elen on output differ from the Fortran
|
alpar@9
|
375 * version (Elen holds the inverse permutation in the Fortran version,
|
alpar@9
|
376 * which is instead returned in the Next array in this C version,
|
alpar@9
|
377 * described below).
|
alpar@9
|
378 *
|
alpar@9
|
379 * Last: In a degree list, Last [i] is the supervariable preceding i, or EMPTY
|
alpar@9
|
380 * if i is the head of the list. In a hash bucket, Last [i] is the hash
|
alpar@9
|
381 * key for i.
|
alpar@9
|
382 *
|
alpar@9
|
383 * Last [Head [hash]] is also used as the head of a hash bucket if
|
alpar@9
|
384 * Head [hash] contains a degree list (see the description of Head,
|
alpar@9
|
385 * below).
|
alpar@9
|
386 *
|
alpar@9
|
387 * On output, Last [0..n-1] holds the permutation. That is, if
|
alpar@9
|
388 * i = Last [k], then row i is the kth pivot row (where k ranges from 0 to
|
alpar@9
|
389 * n-1). Row Last [k] of A is the kth row in the permuted matrix, PAP'.
|
alpar@9
|
390 *
|
alpar@9
|
391 * Next: Next [i] is the supervariable following i in a link list, or EMPTY if
|
alpar@9
|
392 * i is the last in the list. Used for two kinds of lists: degree lists
|
alpar@9
|
393 * and hash buckets (a supervariable can be in only one kind of list at a
|
alpar@9
|
394 * time).
|
alpar@9
|
395 *
|
alpar@9
|
396 * On output Next [0..n-1] holds the inverse permutation. That is, if
|
alpar@9
|
397 * k = Next [i], then row i is the kth pivot row. Row i of A appears as
|
alpar@9
|
398 * the (Next[i])-th row in the permuted matrix, PAP'.
|
alpar@9
|
399 *
|
alpar@9
|
400 * Note that the contents of Next on output differ from the Fortran
|
alpar@9
|
401 * version (Next is undefined on output in the Fortran version).
|
alpar@9
|
402
|
alpar@9
|
403 * ----------------------------------------------------------------------------
|
alpar@9
|
404 * LOCAL WORKSPACE (not input or output - used only during execution):
|
alpar@9
|
405 * ----------------------------------------------------------------------------
|
alpar@9
|
406 *
|
alpar@9
|
407 * Degree: An integer array of size n. If i is a supervariable, then
|
alpar@9
|
408 * Degree [i] holds the current approximation of the external degree of
|
alpar@9
|
409 * row i (an upper bound). The external degree is the number of nonzeros
|
alpar@9
|
410 * in row i, minus ABS (Nv [i]), the diagonal part. The bound is equal to
|
alpar@9
|
411 * the exact external degree if Elen [i] is less than or equal to two.
|
alpar@9
|
412 *
|
alpar@9
|
413 * We also use the term "external degree" for elements e to refer to
|
alpar@9
|
414 * |Le \ Lme|. If e is an element, then Degree [e] is |Le|, which is the
|
alpar@9
|
415 * degree of the off-diagonal part of the element e (not including the
|
alpar@9
|
416 * diagonal part).
|
alpar@9
|
417 *
|
alpar@9
|
418 * Head: An integer array of size n. Head is used for degree lists.
|
alpar@9
|
419 * Head [deg] is the first supervariable in a degree list. All
|
alpar@9
|
420 * supervariables i in a degree list Head [deg] have the same approximate
|
alpar@9
|
421 * degree, namely, deg = Degree [i]. If the list Head [deg] is empty then
|
alpar@9
|
422 * Head [deg] = EMPTY.
|
alpar@9
|
423 *
|
alpar@9
|
424 * During supervariable detection Head [hash] also serves as a pointer to
|
alpar@9
|
425 * a hash bucket. If Head [hash] >= 0, there is a degree list of degree
|
alpar@9
|
426 * hash. The hash bucket head pointer is Last [Head [hash]]. If
|
alpar@9
|
427 * Head [hash] = EMPTY, then the degree list and hash bucket are both
|
alpar@9
|
428 * empty. If Head [hash] < EMPTY, then the degree list is empty, and
|
alpar@9
|
429 * FLIP (Head [hash]) is the head of the hash bucket. After supervariable
|
alpar@9
|
430 * detection is complete, all hash buckets are empty, and the
|
alpar@9
|
431 * (Last [Head [hash]] = EMPTY) condition is restored for the non-empty
|
alpar@9
|
432 * degree lists.
|
alpar@9
|
433 *
|
alpar@9
|
434 * W: An integer array of size n. The flag array W determines the status of
|
alpar@9
|
435 * elements and variables, and the external degree of elements.
|
alpar@9
|
436 *
|
alpar@9
|
437 * for elements:
|
alpar@9
|
438 * if W [e] = 0, then the element e is absorbed.
|
alpar@9
|
439 * if W [e] >= wflg, then W [e] - wflg is the size of the set
|
alpar@9
|
440 * |Le \ Lme|, in terms of nonzeros (the sum of ABS (Nv [i]) for
|
alpar@9
|
441 * each principal variable i that is both in the pattern of
|
alpar@9
|
442 * element e and NOT in the pattern of the current pivot element,
|
alpar@9
|
443 * me).
|
alpar@9
|
444 * if wflg > W [e] > 0, then e is not absorbed and has not yet been
|
alpar@9
|
445 * seen in the scan of the element lists in the computation of
|
alpar@9
|
446 * |Le\Lme| in Scan 1 below.
|
alpar@9
|
447 *
|
alpar@9
|
448 * for variables:
|
alpar@9
|
449 * during supervariable detection, if W [j] != wflg then j is
|
alpar@9
|
450 * not in the pattern of variable i.
|
alpar@9
|
451 *
|
alpar@9
|
452 * The W array is initialized by setting W [i] = 1 for all i, and by
|
alpar@9
|
453 * setting wflg = 2. It is reinitialized if wflg becomes too large (to
|
alpar@9
|
454 * ensure that wflg+n does not cause integer overflow).
|
alpar@9
|
455
|
alpar@9
|
456 * ----------------------------------------------------------------------------
|
alpar@9
|
457 * LOCAL INTEGERS:
|
alpar@9
|
458 * ----------------------------------------------------------------------------
|
alpar@9
|
459 */
|
alpar@9
|
460
|
alpar@9
|
461 Int deg, degme, dext, lemax, e, elenme, eln, i, ilast, inext, j,
|
alpar@9
|
462 jlast, jnext, k, knt1, knt2, knt3, lenj, ln, me, mindeg, nel, nleft,
|
alpar@9
|
463 nvi, nvj, nvpiv, slenme, wbig, we, wflg, wnvi, ok, ndense, ncmpa,
|
alpar@9
|
464 dense, aggressive ;
|
alpar@9
|
465
|
alpar@9
|
466 unsigned Int hash ; /* unsigned, so that hash % n is well defined.*/
|
alpar@9
|
467
|
alpar@9
|
468 /*
|
alpar@9
|
469 * deg: the degree of a variable or element
|
alpar@9
|
470 * degme: size, |Lme|, of the current element, me (= Degree [me])
|
alpar@9
|
471 * dext: external degree, |Le \ Lme|, of some element e
|
alpar@9
|
472 * lemax: largest |Le| seen so far (called dmax in Fortran version)
|
alpar@9
|
473 * e: an element
|
alpar@9
|
474 * elenme: the length, Elen [me], of element list of pivotal variable
|
alpar@9
|
475 * eln: the length, Elen [...], of an element list
|
alpar@9
|
476 * hash: the computed value of the hash function
|
alpar@9
|
477 * i: a supervariable
|
alpar@9
|
478 * ilast: the entry in a link list preceding i
|
alpar@9
|
479 * inext: the entry in a link list following i
|
alpar@9
|
480 * j: a supervariable
|
alpar@9
|
481 * jlast: the entry in a link list preceding j
|
alpar@9
|
482 * jnext: the entry in a link list, or path, following j
|
alpar@9
|
483 * k: the pivot order of an element or variable
|
alpar@9
|
484 * knt1: loop counter used during element construction
|
alpar@9
|
485 * knt2: loop counter used during element construction
|
alpar@9
|
486 * knt3: loop counter used during compression
|
alpar@9
|
487 * lenj: Len [j]
|
alpar@9
|
488 * ln: length of a supervariable list
|
alpar@9
|
489 * me: current supervariable being eliminated, and the current
|
alpar@9
|
490 * element created by eliminating that supervariable
|
alpar@9
|
491 * mindeg: current minimum degree
|
alpar@9
|
492 * nel: number of pivots selected so far
|
alpar@9
|
493 * nleft: n - nel, the number of nonpivotal rows/columns remaining
|
alpar@9
|
494 * nvi: the number of variables in a supervariable i (= Nv [i])
|
alpar@9
|
495 * nvj: the number of variables in a supervariable j (= Nv [j])
|
alpar@9
|
496 * nvpiv: number of pivots in current element
|
alpar@9
|
497 * slenme: number of variables in variable list of pivotal variable
|
alpar@9
|
498 * wbig: = INT_MAX - n for the int version, UF_long_max - n for the
|
alpar@9
|
499 * UF_long version. wflg is not allowed to be >= wbig.
|
alpar@9
|
500 * we: W [e]
|
alpar@9
|
501 * wflg: used for flagging the W array. See description of Iw.
|
alpar@9
|
502 * wnvi: wflg - Nv [i]
|
alpar@9
|
503 * x: either a supervariable or an element
|
alpar@9
|
504 *
|
alpar@9
|
505 * ok: true if supervariable j can be absorbed into i
|
alpar@9
|
506 * ndense: number of "dense" rows/columns
|
alpar@9
|
507 * dense: rows/columns with initial degree > dense are considered "dense"
|
alpar@9
|
508 * aggressive: true if aggressive absorption is being performed
|
alpar@9
|
509 * ncmpa: number of garbage collections
|
alpar@9
|
510
|
alpar@9
|
511 * ----------------------------------------------------------------------------
|
alpar@9
|
512 * LOCAL DOUBLES, used for statistical output only (except for alpha):
|
alpar@9
|
513 * ----------------------------------------------------------------------------
|
alpar@9
|
514 */
|
alpar@9
|
515
|
alpar@9
|
516 double f, r, ndiv, s, nms_lu, nms_ldl, dmax, alpha, lnz, lnzme ;
|
alpar@9
|
517
|
alpar@9
|
518 /*
|
alpar@9
|
519 * f: nvpiv
|
alpar@9
|
520 * r: degme + nvpiv
|
alpar@9
|
521 * ndiv: number of divisions for LU or LDL' factorizations
|
alpar@9
|
522 * s: number of multiply-subtract pairs for LU factorization, for the
|
alpar@9
|
523 * current element me
|
alpar@9
|
524 * nms_lu number of multiply-subtract pairs for LU factorization
|
alpar@9
|
525 * nms_ldl number of multiply-subtract pairs for LDL' factorization
|
alpar@9
|
526 * dmax: the largest number of entries in any column of L, including the
|
alpar@9
|
527 * diagonal
|
alpar@9
|
528 * alpha: "dense" degree ratio
|
alpar@9
|
529 * lnz: the number of nonzeros in L (excluding the diagonal)
|
alpar@9
|
530 * lnzme: the number of nonzeros in L (excl. the diagonal) for the
|
alpar@9
|
531 * current element me
|
alpar@9
|
532
|
alpar@9
|
533 * ----------------------------------------------------------------------------
|
alpar@9
|
534 * LOCAL "POINTERS" (indices into the Iw array)
|
alpar@9
|
535 * ----------------------------------------------------------------------------
|
alpar@9
|
536 */
|
alpar@9
|
537
|
alpar@9
|
538 Int p, p1, p2, p3, p4, pdst, pend, pj, pme, pme1, pme2, pn, psrc ;
|
alpar@9
|
539
|
alpar@9
|
540 /*
|
alpar@9
|
541 * Any parameter (Pe [...] or pfree) or local variable starting with "p" (for
|
alpar@9
|
542 * Pointer) is an index into Iw, and all indices into Iw use variables starting
|
alpar@9
|
543 * with "p." The only exception to this rule is the iwlen input argument.
|
alpar@9
|
544 *
|
alpar@9
|
545 * p: pointer into lots of things
|
alpar@9
|
546 * p1: Pe [i] for some variable i (start of element list)
|
alpar@9
|
547 * p2: Pe [i] + Elen [i] - 1 for some variable i
|
alpar@9
|
548 * p3: index of first supervariable in clean list
|
alpar@9
|
549 * p4:
|
alpar@9
|
550 * pdst: destination pointer, for compression
|
alpar@9
|
551 * pend: end of memory to compress
|
alpar@9
|
552 * pj: pointer into an element or variable
|
alpar@9
|
553 * pme: pointer into the current element (pme1...pme2)
|
alpar@9
|
554 * pme1: the current element, me, is stored in Iw [pme1...pme2]
|
alpar@9
|
555 * pme2: the end of the current element
|
alpar@9
|
556 * pn: pointer into a "clean" variable, also used to compress
|
alpar@9
|
557 * psrc: source pointer, for compression
|
alpar@9
|
558 */
|
alpar@9
|
559
|
alpar@9
|
560 /* ========================================================================= */
|
alpar@9
|
561 /* INITIALIZATIONS */
|
alpar@9
|
562 /* ========================================================================= */
|
alpar@9
|
563
|
alpar@9
|
564 /* Note that this restriction on iwlen is slightly more restrictive than
|
alpar@9
|
565 * what is actually required in AMD_2. AMD_2 can operate with no elbow
|
alpar@9
|
566 * room at all, but it will be slow. For better performance, at least
|
alpar@9
|
567 * size-n elbow room is enforced. */
|
alpar@9
|
568 ASSERT (iwlen >= pfree + n) ;
|
alpar@9
|
569 ASSERT (n > 0) ;
|
alpar@9
|
570
|
alpar@9
|
571 /* initialize output statistics */
|
alpar@9
|
572 lnz = 0 ;
|
alpar@9
|
573 ndiv = 0 ;
|
alpar@9
|
574 nms_lu = 0 ;
|
alpar@9
|
575 nms_ldl = 0 ;
|
alpar@9
|
576 dmax = 1 ;
|
alpar@9
|
577 me = EMPTY ;
|
alpar@9
|
578
|
alpar@9
|
579 mindeg = 0 ;
|
alpar@9
|
580 ncmpa = 0 ;
|
alpar@9
|
581 nel = 0 ;
|
alpar@9
|
582 lemax = 0 ;
|
alpar@9
|
583
|
alpar@9
|
584 /* get control parameters */
|
alpar@9
|
585 if (Control != (double *) NULL)
|
alpar@9
|
586 {
|
alpar@9
|
587 alpha = Control [AMD_DENSE] ;
|
alpar@9
|
588 aggressive = (Control [AMD_AGGRESSIVE] != 0) ;
|
alpar@9
|
589 }
|
alpar@9
|
590 else
|
alpar@9
|
591 {
|
alpar@9
|
592 alpha = AMD_DEFAULT_DENSE ;
|
alpar@9
|
593 aggressive = AMD_DEFAULT_AGGRESSIVE ;
|
alpar@9
|
594 }
|
alpar@9
|
595 /* Note: if alpha is NaN, this is undefined: */
|
alpar@9
|
596 if (alpha < 0)
|
alpar@9
|
597 {
|
alpar@9
|
598 /* only remove completely dense rows/columns */
|
alpar@9
|
599 dense = n-2 ;
|
alpar@9
|
600 }
|
alpar@9
|
601 else
|
alpar@9
|
602 {
|
alpar@9
|
603 dense = alpha * sqrt ((double) n) ;
|
alpar@9
|
604 }
|
alpar@9
|
605 dense = MAX (16, dense) ;
|
alpar@9
|
606 dense = MIN (n, dense) ;
|
alpar@9
|
607 AMD_DEBUG1 (("\n\nAMD (debug), alpha %g, aggr. "ID"\n",
|
alpar@9
|
608 alpha, aggressive)) ;
|
alpar@9
|
609
|
alpar@9
|
610 for (i = 0 ; i < n ; i++)
|
alpar@9
|
611 {
|
alpar@9
|
612 Last [i] = EMPTY ;
|
alpar@9
|
613 Head [i] = EMPTY ;
|
alpar@9
|
614 Next [i] = EMPTY ;
|
alpar@9
|
615 /* if separate Hhead array is used for hash buckets: *
|
alpar@9
|
616 Hhead [i] = EMPTY ;
|
alpar@9
|
617 */
|
alpar@9
|
618 Nv [i] = 1 ;
|
alpar@9
|
619 W [i] = 1 ;
|
alpar@9
|
620 Elen [i] = 0 ;
|
alpar@9
|
621 Degree [i] = Len [i] ;
|
alpar@9
|
622 }
|
alpar@9
|
623
|
alpar@9
|
624 #ifndef NDEBUG
|
alpar@9
|
625 AMD_DEBUG1 (("\n======Nel "ID" initial\n", nel)) ;
|
alpar@9
|
626 AMD_dump (n, Pe, Iw, Len, iwlen, pfree, Nv, Next, Last,
|
alpar@9
|
627 Head, Elen, Degree, W, -1) ;
|
alpar@9
|
628 #endif
|
alpar@9
|
629
|
alpar@9
|
630 /* initialize wflg */
|
alpar@9
|
631 wbig = Int_MAX - n ;
|
alpar@9
|
632 wflg = clear_flag (0, wbig, W, n) ;
|
alpar@9
|
633
|
alpar@9
|
634 /* --------------------------------------------------------------------- */
|
alpar@9
|
635 /* initialize degree lists and eliminate dense and empty rows */
|
alpar@9
|
636 /* --------------------------------------------------------------------- */
|
alpar@9
|
637
|
alpar@9
|
638 ndense = 0 ;
|
alpar@9
|
639
|
alpar@9
|
640 for (i = 0 ; i < n ; i++)
|
alpar@9
|
641 {
|
alpar@9
|
642 deg = Degree [i] ;
|
alpar@9
|
643 ASSERT (deg >= 0 && deg < n) ;
|
alpar@9
|
644 if (deg == 0)
|
alpar@9
|
645 {
|
alpar@9
|
646
|
alpar@9
|
647 /* -------------------------------------------------------------
|
alpar@9
|
648 * we have a variable that can be eliminated at once because
|
alpar@9
|
649 * there is no off-diagonal non-zero in its row. Note that
|
alpar@9
|
650 * Nv [i] = 1 for an empty variable i. It is treated just
|
alpar@9
|
651 * the same as an eliminated element i.
|
alpar@9
|
652 * ------------------------------------------------------------- */
|
alpar@9
|
653
|
alpar@9
|
654 Elen [i] = FLIP (1) ;
|
alpar@9
|
655 nel++ ;
|
alpar@9
|
656 Pe [i] = EMPTY ;
|
alpar@9
|
657 W [i] = 0 ;
|
alpar@9
|
658
|
alpar@9
|
659 }
|
alpar@9
|
660 else if (deg > dense)
|
alpar@9
|
661 {
|
alpar@9
|
662
|
alpar@9
|
663 /* -------------------------------------------------------------
|
alpar@9
|
664 * Dense variables are not treated as elements, but as unordered,
|
alpar@9
|
665 * non-principal variables that have no parent. They do not take
|
alpar@9
|
666 * part in the postorder, since Nv [i] = 0. Note that the Fortran
|
alpar@9
|
667 * version does not have this option.
|
alpar@9
|
668 * ------------------------------------------------------------- */
|
alpar@9
|
669
|
alpar@9
|
670 AMD_DEBUG1 (("Dense node "ID" degree "ID"\n", i, deg)) ;
|
alpar@9
|
671 ndense++ ;
|
alpar@9
|
672 Nv [i] = 0 ; /* do not postorder this node */
|
alpar@9
|
673 Elen [i] = EMPTY ;
|
alpar@9
|
674 nel++ ;
|
alpar@9
|
675 Pe [i] = EMPTY ;
|
alpar@9
|
676
|
alpar@9
|
677 }
|
alpar@9
|
678 else
|
alpar@9
|
679 {
|
alpar@9
|
680
|
alpar@9
|
681 /* -------------------------------------------------------------
|
alpar@9
|
682 * place i in the degree list corresponding to its degree
|
alpar@9
|
683 * ------------------------------------------------------------- */
|
alpar@9
|
684
|
alpar@9
|
685 inext = Head [deg] ;
|
alpar@9
|
686 ASSERT (inext >= EMPTY && inext < n) ;
|
alpar@9
|
687 if (inext != EMPTY) Last [inext] = i ;
|
alpar@9
|
688 Next [i] = inext ;
|
alpar@9
|
689 Head [deg] = i ;
|
alpar@9
|
690
|
alpar@9
|
691 }
|
alpar@9
|
692 }
|
alpar@9
|
693
|
alpar@9
|
694 /* ========================================================================= */
|
alpar@9
|
695 /* WHILE (selecting pivots) DO */
|
alpar@9
|
696 /* ========================================================================= */
|
alpar@9
|
697
|
alpar@9
|
698 while (nel < n)
|
alpar@9
|
699 {
|
alpar@9
|
700
|
alpar@9
|
701 #ifndef NDEBUG
|
alpar@9
|
702 AMD_DEBUG1 (("\n======Nel "ID"\n", nel)) ;
|
alpar@9
|
703 if (AMD_debug >= 2)
|
alpar@9
|
704 {
|
alpar@9
|
705 AMD_dump (n, Pe, Iw, Len, iwlen, pfree, Nv, Next,
|
alpar@9
|
706 Last, Head, Elen, Degree, W, nel) ;
|
alpar@9
|
707 }
|
alpar@9
|
708 #endif
|
alpar@9
|
709
|
alpar@9
|
710 /* ========================================================================= */
|
alpar@9
|
711 /* GET PIVOT OF MINIMUM DEGREE */
|
alpar@9
|
712 /* ========================================================================= */
|
alpar@9
|
713
|
alpar@9
|
714 /* ----------------------------------------------------------------- */
|
alpar@9
|
715 /* find next supervariable for elimination */
|
alpar@9
|
716 /* ----------------------------------------------------------------- */
|
alpar@9
|
717
|
alpar@9
|
718 ASSERT (mindeg >= 0 && mindeg < n) ;
|
alpar@9
|
719 for (deg = mindeg ; deg < n ; deg++)
|
alpar@9
|
720 {
|
alpar@9
|
721 me = Head [deg] ;
|
alpar@9
|
722 if (me != EMPTY) break ;
|
alpar@9
|
723 }
|
alpar@9
|
724 mindeg = deg ;
|
alpar@9
|
725 ASSERT (me >= 0 && me < n) ;
|
alpar@9
|
726 AMD_DEBUG1 (("=================me: "ID"\n", me)) ;
|
alpar@9
|
727
|
alpar@9
|
728 /* ----------------------------------------------------------------- */
|
alpar@9
|
729 /* remove chosen variable from link list */
|
alpar@9
|
730 /* ----------------------------------------------------------------- */
|
alpar@9
|
731
|
alpar@9
|
732 inext = Next [me] ;
|
alpar@9
|
733 ASSERT (inext >= EMPTY && inext < n) ;
|
alpar@9
|
734 if (inext != EMPTY) Last [inext] = EMPTY ;
|
alpar@9
|
735 Head [deg] = inext ;
|
alpar@9
|
736
|
alpar@9
|
737 /* ----------------------------------------------------------------- */
|
alpar@9
|
738 /* me represents the elimination of pivots nel to nel+Nv[me]-1. */
|
alpar@9
|
739 /* place me itself as the first in this set. */
|
alpar@9
|
740 /* ----------------------------------------------------------------- */
|
alpar@9
|
741
|
alpar@9
|
742 elenme = Elen [me] ;
|
alpar@9
|
743 nvpiv = Nv [me] ;
|
alpar@9
|
744 ASSERT (nvpiv > 0) ;
|
alpar@9
|
745 nel += nvpiv ;
|
alpar@9
|
746
|
alpar@9
|
747 /* ========================================================================= */
|
alpar@9
|
748 /* CONSTRUCT NEW ELEMENT */
|
alpar@9
|
749 /* ========================================================================= */
|
alpar@9
|
750
|
alpar@9
|
751 /* -----------------------------------------------------------------
|
alpar@9
|
752 * At this point, me is the pivotal supervariable. It will be
|
alpar@9
|
753 * converted into the current element. Scan list of the pivotal
|
alpar@9
|
754 * supervariable, me, setting tree pointers and constructing new list
|
alpar@9
|
755 * of supervariables for the new element, me. p is a pointer to the
|
alpar@9
|
756 * current position in the old list.
|
alpar@9
|
757 * ----------------------------------------------------------------- */
|
alpar@9
|
758
|
alpar@9
|
759 /* flag the variable "me" as being in Lme by negating Nv [me] */
|
alpar@9
|
760 Nv [me] = -nvpiv ;
|
alpar@9
|
761 degme = 0 ;
|
alpar@9
|
762 ASSERT (Pe [me] >= 0 && Pe [me] < iwlen) ;
|
alpar@9
|
763
|
alpar@9
|
764 if (elenme == 0)
|
alpar@9
|
765 {
|
alpar@9
|
766
|
alpar@9
|
767 /* ------------------------------------------------------------- */
|
alpar@9
|
768 /* construct the new element in place */
|
alpar@9
|
769 /* ------------------------------------------------------------- */
|
alpar@9
|
770
|
alpar@9
|
771 pme1 = Pe [me] ;
|
alpar@9
|
772 pme2 = pme1 - 1 ;
|
alpar@9
|
773
|
alpar@9
|
774 for (p = pme1 ; p <= pme1 + Len [me] - 1 ; p++)
|
alpar@9
|
775 {
|
alpar@9
|
776 i = Iw [p] ;
|
alpar@9
|
777 ASSERT (i >= 0 && i < n && Nv [i] >= 0) ;
|
alpar@9
|
778 nvi = Nv [i] ;
|
alpar@9
|
779 if (nvi > 0)
|
alpar@9
|
780 {
|
alpar@9
|
781
|
alpar@9
|
782 /* ----------------------------------------------------- */
|
alpar@9
|
783 /* i is a principal variable not yet placed in Lme. */
|
alpar@9
|
784 /* store i in new list */
|
alpar@9
|
785 /* ----------------------------------------------------- */
|
alpar@9
|
786
|
alpar@9
|
787 /* flag i as being in Lme by negating Nv [i] */
|
alpar@9
|
788 degme += nvi ;
|
alpar@9
|
789 Nv [i] = -nvi ;
|
alpar@9
|
790 Iw [++pme2] = i ;
|
alpar@9
|
791
|
alpar@9
|
792 /* ----------------------------------------------------- */
|
alpar@9
|
793 /* remove variable i from degree list. */
|
alpar@9
|
794 /* ----------------------------------------------------- */
|
alpar@9
|
795
|
alpar@9
|
796 ilast = Last [i] ;
|
alpar@9
|
797 inext = Next [i] ;
|
alpar@9
|
798 ASSERT (ilast >= EMPTY && ilast < n) ;
|
alpar@9
|
799 ASSERT (inext >= EMPTY && inext < n) ;
|
alpar@9
|
800 if (inext != EMPTY) Last [inext] = ilast ;
|
alpar@9
|
801 if (ilast != EMPTY)
|
alpar@9
|
802 {
|
alpar@9
|
803 Next [ilast] = inext ;
|
alpar@9
|
804 }
|
alpar@9
|
805 else
|
alpar@9
|
806 {
|
alpar@9
|
807 /* i is at the head of the degree list */
|
alpar@9
|
808 ASSERT (Degree [i] >= 0 && Degree [i] < n) ;
|
alpar@9
|
809 Head [Degree [i]] = inext ;
|
alpar@9
|
810 }
|
alpar@9
|
811 }
|
alpar@9
|
812 }
|
alpar@9
|
813 }
|
alpar@9
|
814 else
|
alpar@9
|
815 {
|
alpar@9
|
816
|
alpar@9
|
817 /* ------------------------------------------------------------- */
|
alpar@9
|
818 /* construct the new element in empty space, Iw [pfree ...] */
|
alpar@9
|
819 /* ------------------------------------------------------------- */
|
alpar@9
|
820
|
alpar@9
|
821 p = Pe [me] ;
|
alpar@9
|
822 pme1 = pfree ;
|
alpar@9
|
823 slenme = Len [me] - elenme ;
|
alpar@9
|
824
|
alpar@9
|
825 for (knt1 = 1 ; knt1 <= elenme + 1 ; knt1++)
|
alpar@9
|
826 {
|
alpar@9
|
827
|
alpar@9
|
828 if (knt1 > elenme)
|
alpar@9
|
829 {
|
alpar@9
|
830 /* search the supervariables in me. */
|
alpar@9
|
831 e = me ;
|
alpar@9
|
832 pj = p ;
|
alpar@9
|
833 ln = slenme ;
|
alpar@9
|
834 AMD_DEBUG2 (("Search sv: "ID" "ID" "ID"\n", me,pj,ln)) ;
|
alpar@9
|
835 }
|
alpar@9
|
836 else
|
alpar@9
|
837 {
|
alpar@9
|
838 /* search the elements in me. */
|
alpar@9
|
839 e = Iw [p++] ;
|
alpar@9
|
840 ASSERT (e >= 0 && e < n) ;
|
alpar@9
|
841 pj = Pe [e] ;
|
alpar@9
|
842 ln = Len [e] ;
|
alpar@9
|
843 AMD_DEBUG2 (("Search element e "ID" in me "ID"\n", e,me)) ;
|
alpar@9
|
844 ASSERT (Elen [e] < EMPTY && W [e] > 0 && pj >= 0) ;
|
alpar@9
|
845 }
|
alpar@9
|
846 ASSERT (ln >= 0 && (ln == 0 || (pj >= 0 && pj < iwlen))) ;
|
alpar@9
|
847
|
alpar@9
|
848 /* ---------------------------------------------------------
|
alpar@9
|
849 * search for different supervariables and add them to the
|
alpar@9
|
850 * new list, compressing when necessary. this loop is
|
alpar@9
|
851 * executed once for each element in the list and once for
|
alpar@9
|
852 * all the supervariables in the list.
|
alpar@9
|
853 * --------------------------------------------------------- */
|
alpar@9
|
854
|
alpar@9
|
855 for (knt2 = 1 ; knt2 <= ln ; knt2++)
|
alpar@9
|
856 {
|
alpar@9
|
857 i = Iw [pj++] ;
|
alpar@9
|
858 ASSERT (i >= 0 && i < n && (i == me || Elen [i] >= EMPTY));
|
alpar@9
|
859 nvi = Nv [i] ;
|
alpar@9
|
860 AMD_DEBUG2 ((": "ID" "ID" "ID" "ID"\n",
|
alpar@9
|
861 i, Elen [i], Nv [i], wflg)) ;
|
alpar@9
|
862
|
alpar@9
|
863 if (nvi > 0)
|
alpar@9
|
864 {
|
alpar@9
|
865
|
alpar@9
|
866 /* ------------------------------------------------- */
|
alpar@9
|
867 /* compress Iw, if necessary */
|
alpar@9
|
868 /* ------------------------------------------------- */
|
alpar@9
|
869
|
alpar@9
|
870 if (pfree >= iwlen)
|
alpar@9
|
871 {
|
alpar@9
|
872
|
alpar@9
|
873 AMD_DEBUG1 (("GARBAGE COLLECTION\n")) ;
|
alpar@9
|
874
|
alpar@9
|
875 /* prepare for compressing Iw by adjusting pointers
|
alpar@9
|
876 * and lengths so that the lists being searched in
|
alpar@9
|
877 * the inner and outer loops contain only the
|
alpar@9
|
878 * remaining entries. */
|
alpar@9
|
879
|
alpar@9
|
880 Pe [me] = p ;
|
alpar@9
|
881 Len [me] -= knt1 ;
|
alpar@9
|
882 /* check if nothing left of supervariable me */
|
alpar@9
|
883 if (Len [me] == 0) Pe [me] = EMPTY ;
|
alpar@9
|
884 Pe [e] = pj ;
|
alpar@9
|
885 Len [e] = ln - knt2 ;
|
alpar@9
|
886 /* nothing left of element e */
|
alpar@9
|
887 if (Len [e] == 0) Pe [e] = EMPTY ;
|
alpar@9
|
888
|
alpar@9
|
889 ncmpa++ ; /* one more garbage collection */
|
alpar@9
|
890
|
alpar@9
|
891 /* store first entry of each object in Pe */
|
alpar@9
|
892 /* FLIP the first entry in each object */
|
alpar@9
|
893 for (j = 0 ; j < n ; j++)
|
alpar@9
|
894 {
|
alpar@9
|
895 pn = Pe [j] ;
|
alpar@9
|
896 if (pn >= 0)
|
alpar@9
|
897 {
|
alpar@9
|
898 ASSERT (pn >= 0 && pn < iwlen) ;
|
alpar@9
|
899 Pe [j] = Iw [pn] ;
|
alpar@9
|
900 Iw [pn] = FLIP (j) ;
|
alpar@9
|
901 }
|
alpar@9
|
902 }
|
alpar@9
|
903
|
alpar@9
|
904 /* psrc/pdst point to source/destination */
|
alpar@9
|
905 psrc = 0 ;
|
alpar@9
|
906 pdst = 0 ;
|
alpar@9
|
907 pend = pme1 - 1 ;
|
alpar@9
|
908
|
alpar@9
|
909 while (psrc <= pend)
|
alpar@9
|
910 {
|
alpar@9
|
911 /* search for next FLIP'd entry */
|
alpar@9
|
912 j = FLIP (Iw [psrc++]) ;
|
alpar@9
|
913 if (j >= 0)
|
alpar@9
|
914 {
|
alpar@9
|
915 AMD_DEBUG2 (("Got object j: "ID"\n", j)) ;
|
alpar@9
|
916 Iw [pdst] = Pe [j] ;
|
alpar@9
|
917 Pe [j] = pdst++ ;
|
alpar@9
|
918 lenj = Len [j] ;
|
alpar@9
|
919 /* copy from source to destination */
|
alpar@9
|
920 for (knt3 = 0 ; knt3 <= lenj - 2 ; knt3++)
|
alpar@9
|
921 {
|
alpar@9
|
922 Iw [pdst++] = Iw [psrc++] ;
|
alpar@9
|
923 }
|
alpar@9
|
924 }
|
alpar@9
|
925 }
|
alpar@9
|
926
|
alpar@9
|
927 /* move the new partially-constructed element */
|
alpar@9
|
928 p1 = pdst ;
|
alpar@9
|
929 for (psrc = pme1 ; psrc <= pfree-1 ; psrc++)
|
alpar@9
|
930 {
|
alpar@9
|
931 Iw [pdst++] = Iw [psrc] ;
|
alpar@9
|
932 }
|
alpar@9
|
933 pme1 = p1 ;
|
alpar@9
|
934 pfree = pdst ;
|
alpar@9
|
935 pj = Pe [e] ;
|
alpar@9
|
936 p = Pe [me] ;
|
alpar@9
|
937
|
alpar@9
|
938 }
|
alpar@9
|
939
|
alpar@9
|
940 /* ------------------------------------------------- */
|
alpar@9
|
941 /* i is a principal variable not yet placed in Lme */
|
alpar@9
|
942 /* store i in new list */
|
alpar@9
|
943 /* ------------------------------------------------- */
|
alpar@9
|
944
|
alpar@9
|
945 /* flag i as being in Lme by negating Nv [i] */
|
alpar@9
|
946 degme += nvi ;
|
alpar@9
|
947 Nv [i] = -nvi ;
|
alpar@9
|
948 Iw [pfree++] = i ;
|
alpar@9
|
949 AMD_DEBUG2 ((" s: "ID" nv "ID"\n", i, Nv [i]));
|
alpar@9
|
950
|
alpar@9
|
951 /* ------------------------------------------------- */
|
alpar@9
|
952 /* remove variable i from degree link list */
|
alpar@9
|
953 /* ------------------------------------------------- */
|
alpar@9
|
954
|
alpar@9
|
955 ilast = Last [i] ;
|
alpar@9
|
956 inext = Next [i] ;
|
alpar@9
|
957 ASSERT (ilast >= EMPTY && ilast < n) ;
|
alpar@9
|
958 ASSERT (inext >= EMPTY && inext < n) ;
|
alpar@9
|
959 if (inext != EMPTY) Last [inext] = ilast ;
|
alpar@9
|
960 if (ilast != EMPTY)
|
alpar@9
|
961 {
|
alpar@9
|
962 Next [ilast] = inext ;
|
alpar@9
|
963 }
|
alpar@9
|
964 else
|
alpar@9
|
965 {
|
alpar@9
|
966 /* i is at the head of the degree list */
|
alpar@9
|
967 ASSERT (Degree [i] >= 0 && Degree [i] < n) ;
|
alpar@9
|
968 Head [Degree [i]] = inext ;
|
alpar@9
|
969 }
|
alpar@9
|
970 }
|
alpar@9
|
971 }
|
alpar@9
|
972
|
alpar@9
|
973 if (e != me)
|
alpar@9
|
974 {
|
alpar@9
|
975 /* set tree pointer and flag to indicate element e is
|
alpar@9
|
976 * absorbed into new element me (the parent of e is me) */
|
alpar@9
|
977 AMD_DEBUG1 ((" Element "ID" => "ID"\n", e, me)) ;
|
alpar@9
|
978 Pe [e] = FLIP (me) ;
|
alpar@9
|
979 W [e] = 0 ;
|
alpar@9
|
980 }
|
alpar@9
|
981 }
|
alpar@9
|
982
|
alpar@9
|
983 pme2 = pfree - 1 ;
|
alpar@9
|
984 }
|
alpar@9
|
985
|
alpar@9
|
986 /* ----------------------------------------------------------------- */
|
alpar@9
|
987 /* me has now been converted into an element in Iw [pme1..pme2] */
|
alpar@9
|
988 /* ----------------------------------------------------------------- */
|
alpar@9
|
989
|
alpar@9
|
990 /* degme holds the external degree of new element */
|
alpar@9
|
991 Degree [me] = degme ;
|
alpar@9
|
992 Pe [me] = pme1 ;
|
alpar@9
|
993 Len [me] = pme2 - pme1 + 1 ;
|
alpar@9
|
994 ASSERT (Pe [me] >= 0 && Pe [me] < iwlen) ;
|
alpar@9
|
995
|
alpar@9
|
996 Elen [me] = FLIP (nvpiv + degme) ;
|
alpar@9
|
997 /* FLIP (Elen (me)) is now the degree of pivot (including
|
alpar@9
|
998 * diagonal part). */
|
alpar@9
|
999
|
alpar@9
|
1000 #ifndef NDEBUG
|
alpar@9
|
1001 AMD_DEBUG2 (("New element structure: length= "ID"\n", pme2-pme1+1)) ;
|
alpar@9
|
1002 for (pme = pme1 ; pme <= pme2 ; pme++) AMD_DEBUG3 ((" "ID"", Iw[pme]));
|
alpar@9
|
1003 AMD_DEBUG3 (("\n")) ;
|
alpar@9
|
1004 #endif
|
alpar@9
|
1005
|
alpar@9
|
1006 /* ----------------------------------------------------------------- */
|
alpar@9
|
1007 /* make sure that wflg is not too large. */
|
alpar@9
|
1008 /* ----------------------------------------------------------------- */
|
alpar@9
|
1009
|
alpar@9
|
1010 /* With the current value of wflg, wflg+n must not cause integer
|
alpar@9
|
1011 * overflow */
|
alpar@9
|
1012
|
alpar@9
|
1013 wflg = clear_flag (wflg, wbig, W, n) ;
|
alpar@9
|
1014
|
alpar@9
|
1015 /* ========================================================================= */
|
alpar@9
|
1016 /* COMPUTE (W [e] - wflg) = |Le\Lme| FOR ALL ELEMENTS */
|
alpar@9
|
1017 /* ========================================================================= */
|
alpar@9
|
1018
|
alpar@9
|
1019 /* -----------------------------------------------------------------
|
alpar@9
|
1020 * Scan 1: compute the external degrees of previous elements with
|
alpar@9
|
1021 * respect to the current element. That is:
|
alpar@9
|
1022 * (W [e] - wflg) = |Le \ Lme|
|
alpar@9
|
1023 * for each element e that appears in any supervariable in Lme. The
|
alpar@9
|
1024 * notation Le refers to the pattern (list of supervariables) of a
|
alpar@9
|
1025 * previous element e, where e is not yet absorbed, stored in
|
alpar@9
|
1026 * Iw [Pe [e] + 1 ... Pe [e] + Len [e]]. The notation Lme
|
alpar@9
|
1027 * refers to the pattern of the current element (stored in
|
alpar@9
|
1028 * Iw [pme1..pme2]). If aggressive absorption is enabled, and
|
alpar@9
|
1029 * (W [e] - wflg) becomes zero, then the element e will be absorbed
|
alpar@9
|
1030 * in Scan 2.
|
alpar@9
|
1031 * ----------------------------------------------------------------- */
|
alpar@9
|
1032
|
alpar@9
|
1033 AMD_DEBUG2 (("me: ")) ;
|
alpar@9
|
1034 for (pme = pme1 ; pme <= pme2 ; pme++)
|
alpar@9
|
1035 {
|
alpar@9
|
1036 i = Iw [pme] ;
|
alpar@9
|
1037 ASSERT (i >= 0 && i < n) ;
|
alpar@9
|
1038 eln = Elen [i] ;
|
alpar@9
|
1039 AMD_DEBUG3 ((""ID" Elen "ID": \n", i, eln)) ;
|
alpar@9
|
1040 if (eln > 0)
|
alpar@9
|
1041 {
|
alpar@9
|
1042 /* note that Nv [i] has been negated to denote i in Lme: */
|
alpar@9
|
1043 nvi = -Nv [i] ;
|
alpar@9
|
1044 ASSERT (nvi > 0 && Pe [i] >= 0 && Pe [i] < iwlen) ;
|
alpar@9
|
1045 wnvi = wflg - nvi ;
|
alpar@9
|
1046 for (p = Pe [i] ; p <= Pe [i] + eln - 1 ; p++)
|
alpar@9
|
1047 {
|
alpar@9
|
1048 e = Iw [p] ;
|
alpar@9
|
1049 ASSERT (e >= 0 && e < n) ;
|
alpar@9
|
1050 we = W [e] ;
|
alpar@9
|
1051 AMD_DEBUG4 ((" e "ID" we "ID" ", e, we)) ;
|
alpar@9
|
1052 if (we >= wflg)
|
alpar@9
|
1053 {
|
alpar@9
|
1054 /* unabsorbed element e has been seen in this loop */
|
alpar@9
|
1055 AMD_DEBUG4 ((" unabsorbed, first time seen")) ;
|
alpar@9
|
1056 we -= nvi ;
|
alpar@9
|
1057 }
|
alpar@9
|
1058 else if (we != 0)
|
alpar@9
|
1059 {
|
alpar@9
|
1060 /* e is an unabsorbed element */
|
alpar@9
|
1061 /* this is the first we have seen e in all of Scan 1 */
|
alpar@9
|
1062 AMD_DEBUG4 ((" unabsorbed")) ;
|
alpar@9
|
1063 we = Degree [e] + wnvi ;
|
alpar@9
|
1064 }
|
alpar@9
|
1065 AMD_DEBUG4 (("\n")) ;
|
alpar@9
|
1066 W [e] = we ;
|
alpar@9
|
1067 }
|
alpar@9
|
1068 }
|
alpar@9
|
1069 }
|
alpar@9
|
1070 AMD_DEBUG2 (("\n")) ;
|
alpar@9
|
1071
|
alpar@9
|
1072 /* ========================================================================= */
|
alpar@9
|
1073 /* DEGREE UPDATE AND ELEMENT ABSORPTION */
|
alpar@9
|
1074 /* ========================================================================= */
|
alpar@9
|
1075
|
alpar@9
|
1076 /* -----------------------------------------------------------------
|
alpar@9
|
1077 * Scan 2: for each i in Lme, sum up the degree of Lme (which is
|
alpar@9
|
1078 * degme), plus the sum of the external degrees of each Le for the
|
alpar@9
|
1079 * elements e appearing within i, plus the supervariables in i.
|
alpar@9
|
1080 * Place i in hash list.
|
alpar@9
|
1081 * ----------------------------------------------------------------- */
|
alpar@9
|
1082
|
alpar@9
|
1083 for (pme = pme1 ; pme <= pme2 ; pme++)
|
alpar@9
|
1084 {
|
alpar@9
|
1085 i = Iw [pme] ;
|
alpar@9
|
1086 ASSERT (i >= 0 && i < n && Nv [i] < 0 && Elen [i] >= 0) ;
|
alpar@9
|
1087 AMD_DEBUG2 (("Updating: i "ID" "ID" "ID"\n", i, Elen[i], Len [i]));
|
alpar@9
|
1088 p1 = Pe [i] ;
|
alpar@9
|
1089 p2 = p1 + Elen [i] - 1 ;
|
alpar@9
|
1090 pn = p1 ;
|
alpar@9
|
1091 hash = 0 ;
|
alpar@9
|
1092 deg = 0 ;
|
alpar@9
|
1093 ASSERT (p1 >= 0 && p1 < iwlen && p2 >= -1 && p2 < iwlen) ;
|
alpar@9
|
1094
|
alpar@9
|
1095 /* ------------------------------------------------------------- */
|
alpar@9
|
1096 /* scan the element list associated with supervariable i */
|
alpar@9
|
1097 /* ------------------------------------------------------------- */
|
alpar@9
|
1098
|
alpar@9
|
1099 /* UMFPACK/MA38-style approximate degree: */
|
alpar@9
|
1100 if (aggressive)
|
alpar@9
|
1101 {
|
alpar@9
|
1102 for (p = p1 ; p <= p2 ; p++)
|
alpar@9
|
1103 {
|
alpar@9
|
1104 e = Iw [p] ;
|
alpar@9
|
1105 ASSERT (e >= 0 && e < n) ;
|
alpar@9
|
1106 we = W [e] ;
|
alpar@9
|
1107 if (we != 0)
|
alpar@9
|
1108 {
|
alpar@9
|
1109 /* e is an unabsorbed element */
|
alpar@9
|
1110 /* dext = | Le \ Lme | */
|
alpar@9
|
1111 dext = we - wflg ;
|
alpar@9
|
1112 if (dext > 0)
|
alpar@9
|
1113 {
|
alpar@9
|
1114 deg += dext ;
|
alpar@9
|
1115 Iw [pn++] = e ;
|
alpar@9
|
1116 hash += e ;
|
alpar@9
|
1117 AMD_DEBUG4 ((" e: "ID" hash = "ID"\n",e,hash)) ;
|
alpar@9
|
1118 }
|
alpar@9
|
1119 else
|
alpar@9
|
1120 {
|
alpar@9
|
1121 /* external degree of e is zero, absorb e into me*/
|
alpar@9
|
1122 AMD_DEBUG1 ((" Element "ID" =>"ID" (aggressive)\n",
|
alpar@9
|
1123 e, me)) ;
|
alpar@9
|
1124 ASSERT (dext == 0) ;
|
alpar@9
|
1125 Pe [e] = FLIP (me) ;
|
alpar@9
|
1126 W [e] = 0 ;
|
alpar@9
|
1127 }
|
alpar@9
|
1128 }
|
alpar@9
|
1129 }
|
alpar@9
|
1130 }
|
alpar@9
|
1131 else
|
alpar@9
|
1132 {
|
alpar@9
|
1133 for (p = p1 ; p <= p2 ; p++)
|
alpar@9
|
1134 {
|
alpar@9
|
1135 e = Iw [p] ;
|
alpar@9
|
1136 ASSERT (e >= 0 && e < n) ;
|
alpar@9
|
1137 we = W [e] ;
|
alpar@9
|
1138 if (we != 0)
|
alpar@9
|
1139 {
|
alpar@9
|
1140 /* e is an unabsorbed element */
|
alpar@9
|
1141 dext = we - wflg ;
|
alpar@9
|
1142 ASSERT (dext >= 0) ;
|
alpar@9
|
1143 deg += dext ;
|
alpar@9
|
1144 Iw [pn++] = e ;
|
alpar@9
|
1145 hash += e ;
|
alpar@9
|
1146 AMD_DEBUG4 ((" e: "ID" hash = "ID"\n",e,hash)) ;
|
alpar@9
|
1147 }
|
alpar@9
|
1148 }
|
alpar@9
|
1149 }
|
alpar@9
|
1150
|
alpar@9
|
1151 /* count the number of elements in i (including me): */
|
alpar@9
|
1152 Elen [i] = pn - p1 + 1 ;
|
alpar@9
|
1153
|
alpar@9
|
1154 /* ------------------------------------------------------------- */
|
alpar@9
|
1155 /* scan the supervariables in the list associated with i */
|
alpar@9
|
1156 /* ------------------------------------------------------------- */
|
alpar@9
|
1157
|
alpar@9
|
1158 /* The bulk of the AMD run time is typically spent in this loop,
|
alpar@9
|
1159 * particularly if the matrix has many dense rows that are not
|
alpar@9
|
1160 * removed prior to ordering. */
|
alpar@9
|
1161 p3 = pn ;
|
alpar@9
|
1162 p4 = p1 + Len [i] ;
|
alpar@9
|
1163 for (p = p2 + 1 ; p < p4 ; p++)
|
alpar@9
|
1164 {
|
alpar@9
|
1165 j = Iw [p] ;
|
alpar@9
|
1166 ASSERT (j >= 0 && j < n) ;
|
alpar@9
|
1167 nvj = Nv [j] ;
|
alpar@9
|
1168 if (nvj > 0)
|
alpar@9
|
1169 {
|
alpar@9
|
1170 /* j is unabsorbed, and not in Lme. */
|
alpar@9
|
1171 /* add to degree and add to new list */
|
alpar@9
|
1172 deg += nvj ;
|
alpar@9
|
1173 Iw [pn++] = j ;
|
alpar@9
|
1174 hash += j ;
|
alpar@9
|
1175 AMD_DEBUG4 ((" s: "ID" hash "ID" Nv[j]= "ID"\n",
|
alpar@9
|
1176 j, hash, nvj)) ;
|
alpar@9
|
1177 }
|
alpar@9
|
1178 }
|
alpar@9
|
1179
|
alpar@9
|
1180 /* ------------------------------------------------------------- */
|
alpar@9
|
1181 /* update the degree and check for mass elimination */
|
alpar@9
|
1182 /* ------------------------------------------------------------- */
|
alpar@9
|
1183
|
alpar@9
|
1184 /* with aggressive absorption, deg==0 is identical to the
|
alpar@9
|
1185 * Elen [i] == 1 && p3 == pn test, below. */
|
alpar@9
|
1186 ASSERT (IMPLIES (aggressive, (deg==0) == (Elen[i]==1 && p3==pn))) ;
|
alpar@9
|
1187
|
alpar@9
|
1188 if (Elen [i] == 1 && p3 == pn)
|
alpar@9
|
1189 {
|
alpar@9
|
1190
|
alpar@9
|
1191 /* --------------------------------------------------------- */
|
alpar@9
|
1192 /* mass elimination */
|
alpar@9
|
1193 /* --------------------------------------------------------- */
|
alpar@9
|
1194
|
alpar@9
|
1195 /* There is nothing left of this node except for an edge to
|
alpar@9
|
1196 * the current pivot element. Elen [i] is 1, and there are
|
alpar@9
|
1197 * no variables adjacent to node i. Absorb i into the
|
alpar@9
|
1198 * current pivot element, me. Note that if there are two or
|
alpar@9
|
1199 * more mass eliminations, fillin due to mass elimination is
|
alpar@9
|
1200 * possible within the nvpiv-by-nvpiv pivot block. It is this
|
alpar@9
|
1201 * step that causes AMD's analysis to be an upper bound.
|
alpar@9
|
1202 *
|
alpar@9
|
1203 * The reason is that the selected pivot has a lower
|
alpar@9
|
1204 * approximate degree than the true degree of the two mass
|
alpar@9
|
1205 * eliminated nodes. There is no edge between the two mass
|
alpar@9
|
1206 * eliminated nodes. They are merged with the current pivot
|
alpar@9
|
1207 * anyway.
|
alpar@9
|
1208 *
|
alpar@9
|
1209 * No fillin occurs in the Schur complement, in any case,
|
alpar@9
|
1210 * and this effect does not decrease the quality of the
|
alpar@9
|
1211 * ordering itself, just the quality of the nonzero and
|
alpar@9
|
1212 * flop count analysis. It also means that the post-ordering
|
alpar@9
|
1213 * is not an exact elimination tree post-ordering. */
|
alpar@9
|
1214
|
alpar@9
|
1215 AMD_DEBUG1 ((" MASS i "ID" => parent e "ID"\n", i, me)) ;
|
alpar@9
|
1216 Pe [i] = FLIP (me) ;
|
alpar@9
|
1217 nvi = -Nv [i] ;
|
alpar@9
|
1218 degme -= nvi ;
|
alpar@9
|
1219 nvpiv += nvi ;
|
alpar@9
|
1220 nel += nvi ;
|
alpar@9
|
1221 Nv [i] = 0 ;
|
alpar@9
|
1222 Elen [i] = EMPTY ;
|
alpar@9
|
1223
|
alpar@9
|
1224 }
|
alpar@9
|
1225 else
|
alpar@9
|
1226 {
|
alpar@9
|
1227
|
alpar@9
|
1228 /* --------------------------------------------------------- */
|
alpar@9
|
1229 /* update the upper-bound degree of i */
|
alpar@9
|
1230 /* --------------------------------------------------------- */
|
alpar@9
|
1231
|
alpar@9
|
1232 /* the following degree does not yet include the size
|
alpar@9
|
1233 * of the current element, which is added later: */
|
alpar@9
|
1234
|
alpar@9
|
1235 Degree [i] = MIN (Degree [i], deg) ;
|
alpar@9
|
1236
|
alpar@9
|
1237 /* --------------------------------------------------------- */
|
alpar@9
|
1238 /* add me to the list for i */
|
alpar@9
|
1239 /* --------------------------------------------------------- */
|
alpar@9
|
1240
|
alpar@9
|
1241 /* move first supervariable to end of list */
|
alpar@9
|
1242 Iw [pn] = Iw [p3] ;
|
alpar@9
|
1243 /* move first element to end of element part of list */
|
alpar@9
|
1244 Iw [p3] = Iw [p1] ;
|
alpar@9
|
1245 /* add new element, me, to front of list. */
|
alpar@9
|
1246 Iw [p1] = me ;
|
alpar@9
|
1247 /* store the new length of the list in Len [i] */
|
alpar@9
|
1248 Len [i] = pn - p1 + 1 ;
|
alpar@9
|
1249
|
alpar@9
|
1250 /* --------------------------------------------------------- */
|
alpar@9
|
1251 /* place in hash bucket. Save hash key of i in Last [i]. */
|
alpar@9
|
1252 /* --------------------------------------------------------- */
|
alpar@9
|
1253
|
alpar@9
|
1254 /* NOTE: this can fail if hash is negative, because the ANSI C
|
alpar@9
|
1255 * standard does not define a % b when a and/or b are negative.
|
alpar@9
|
1256 * That's why hash is defined as an unsigned Int, to avoid this
|
alpar@9
|
1257 * problem. */
|
alpar@9
|
1258 hash = hash % n ;
|
alpar@9
|
1259 ASSERT (((Int) hash) >= 0 && ((Int) hash) < n) ;
|
alpar@9
|
1260
|
alpar@9
|
1261 /* if the Hhead array is not used: */
|
alpar@9
|
1262 j = Head [hash] ;
|
alpar@9
|
1263 if (j <= EMPTY)
|
alpar@9
|
1264 {
|
alpar@9
|
1265 /* degree list is empty, hash head is FLIP (j) */
|
alpar@9
|
1266 Next [i] = FLIP (j) ;
|
alpar@9
|
1267 Head [hash] = FLIP (i) ;
|
alpar@9
|
1268 }
|
alpar@9
|
1269 else
|
alpar@9
|
1270 {
|
alpar@9
|
1271 /* degree list is not empty, use Last [Head [hash]] as
|
alpar@9
|
1272 * hash head. */
|
alpar@9
|
1273 Next [i] = Last [j] ;
|
alpar@9
|
1274 Last [j] = i ;
|
alpar@9
|
1275 }
|
alpar@9
|
1276
|
alpar@9
|
1277 /* if a separate Hhead array is used: *
|
alpar@9
|
1278 Next [i] = Hhead [hash] ;
|
alpar@9
|
1279 Hhead [hash] = i ;
|
alpar@9
|
1280 */
|
alpar@9
|
1281
|
alpar@9
|
1282 Last [i] = hash ;
|
alpar@9
|
1283 }
|
alpar@9
|
1284 }
|
alpar@9
|
1285
|
alpar@9
|
1286 Degree [me] = degme ;
|
alpar@9
|
1287
|
alpar@9
|
1288 /* ----------------------------------------------------------------- */
|
alpar@9
|
1289 /* Clear the counter array, W [...], by incrementing wflg. */
|
alpar@9
|
1290 /* ----------------------------------------------------------------- */
|
alpar@9
|
1291
|
alpar@9
|
1292 /* make sure that wflg+n does not cause integer overflow */
|
alpar@9
|
1293 lemax = MAX (lemax, degme) ;
|
alpar@9
|
1294 wflg += lemax ;
|
alpar@9
|
1295 wflg = clear_flag (wflg, wbig, W, n) ;
|
alpar@9
|
1296 /* at this point, W [0..n-1] < wflg holds */
|
alpar@9
|
1297
|
alpar@9
|
1298 /* ========================================================================= */
|
alpar@9
|
1299 /* SUPERVARIABLE DETECTION */
|
alpar@9
|
1300 /* ========================================================================= */
|
alpar@9
|
1301
|
alpar@9
|
1302 AMD_DEBUG1 (("Detecting supervariables:\n")) ;
|
alpar@9
|
1303 for (pme = pme1 ; pme <= pme2 ; pme++)
|
alpar@9
|
1304 {
|
alpar@9
|
1305 i = Iw [pme] ;
|
alpar@9
|
1306 ASSERT (i >= 0 && i < n) ;
|
alpar@9
|
1307 AMD_DEBUG2 (("Consider i "ID" nv "ID"\n", i, Nv [i])) ;
|
alpar@9
|
1308 if (Nv [i] < 0)
|
alpar@9
|
1309 {
|
alpar@9
|
1310 /* i is a principal variable in Lme */
|
alpar@9
|
1311
|
alpar@9
|
1312 /* ---------------------------------------------------------
|
alpar@9
|
1313 * examine all hash buckets with 2 or more variables. We do
|
alpar@9
|
1314 * this by examing all unique hash keys for supervariables in
|
alpar@9
|
1315 * the pattern Lme of the current element, me
|
alpar@9
|
1316 * --------------------------------------------------------- */
|
alpar@9
|
1317
|
alpar@9
|
1318 /* let i = head of hash bucket, and empty the hash bucket */
|
alpar@9
|
1319 ASSERT (Last [i] >= 0 && Last [i] < n) ;
|
alpar@9
|
1320 hash = Last [i] ;
|
alpar@9
|
1321
|
alpar@9
|
1322 /* if Hhead array is not used: */
|
alpar@9
|
1323 j = Head [hash] ;
|
alpar@9
|
1324 if (j == EMPTY)
|
alpar@9
|
1325 {
|
alpar@9
|
1326 /* hash bucket and degree list are both empty */
|
alpar@9
|
1327 i = EMPTY ;
|
alpar@9
|
1328 }
|
alpar@9
|
1329 else if (j < EMPTY)
|
alpar@9
|
1330 {
|
alpar@9
|
1331 /* degree list is empty */
|
alpar@9
|
1332 i = FLIP (j) ;
|
alpar@9
|
1333 Head [hash] = EMPTY ;
|
alpar@9
|
1334 }
|
alpar@9
|
1335 else
|
alpar@9
|
1336 {
|
alpar@9
|
1337 /* degree list is not empty, restore Last [j] of head j */
|
alpar@9
|
1338 i = Last [j] ;
|
alpar@9
|
1339 Last [j] = EMPTY ;
|
alpar@9
|
1340 }
|
alpar@9
|
1341
|
alpar@9
|
1342 /* if separate Hhead array is used: *
|
alpar@9
|
1343 i = Hhead [hash] ;
|
alpar@9
|
1344 Hhead [hash] = EMPTY ;
|
alpar@9
|
1345 */
|
alpar@9
|
1346
|
alpar@9
|
1347 ASSERT (i >= EMPTY && i < n) ;
|
alpar@9
|
1348 AMD_DEBUG2 (("----i "ID" hash "ID"\n", i, hash)) ;
|
alpar@9
|
1349
|
alpar@9
|
1350 while (i != EMPTY && Next [i] != EMPTY)
|
alpar@9
|
1351 {
|
alpar@9
|
1352
|
alpar@9
|
1353 /* -----------------------------------------------------
|
alpar@9
|
1354 * this bucket has one or more variables following i.
|
alpar@9
|
1355 * scan all of them to see if i can absorb any entries
|
alpar@9
|
1356 * that follow i in hash bucket. Scatter i into w.
|
alpar@9
|
1357 * ----------------------------------------------------- */
|
alpar@9
|
1358
|
alpar@9
|
1359 ln = Len [i] ;
|
alpar@9
|
1360 eln = Elen [i] ;
|
alpar@9
|
1361 ASSERT (ln >= 0 && eln >= 0) ;
|
alpar@9
|
1362 ASSERT (Pe [i] >= 0 && Pe [i] < iwlen) ;
|
alpar@9
|
1363 /* do not flag the first element in the list (me) */
|
alpar@9
|
1364 for (p = Pe [i] + 1 ; p <= Pe [i] + ln - 1 ; p++)
|
alpar@9
|
1365 {
|
alpar@9
|
1366 ASSERT (Iw [p] >= 0 && Iw [p] < n) ;
|
alpar@9
|
1367 W [Iw [p]] = wflg ;
|
alpar@9
|
1368 }
|
alpar@9
|
1369
|
alpar@9
|
1370 /* ----------------------------------------------------- */
|
alpar@9
|
1371 /* scan every other entry j following i in bucket */
|
alpar@9
|
1372 /* ----------------------------------------------------- */
|
alpar@9
|
1373
|
alpar@9
|
1374 jlast = i ;
|
alpar@9
|
1375 j = Next [i] ;
|
alpar@9
|
1376 ASSERT (j >= EMPTY && j < n) ;
|
alpar@9
|
1377
|
alpar@9
|
1378 while (j != EMPTY)
|
alpar@9
|
1379 {
|
alpar@9
|
1380 /* ------------------------------------------------- */
|
alpar@9
|
1381 /* check if j and i have identical nonzero pattern */
|
alpar@9
|
1382 /* ------------------------------------------------- */
|
alpar@9
|
1383
|
alpar@9
|
1384 AMD_DEBUG3 (("compare i "ID" and j "ID"\n", i,j)) ;
|
alpar@9
|
1385
|
alpar@9
|
1386 /* check if i and j have the same Len and Elen */
|
alpar@9
|
1387 ASSERT (Len [j] >= 0 && Elen [j] >= 0) ;
|
alpar@9
|
1388 ASSERT (Pe [j] >= 0 && Pe [j] < iwlen) ;
|
alpar@9
|
1389 ok = (Len [j] == ln) && (Elen [j] == eln) ;
|
alpar@9
|
1390 /* skip the first element in the list (me) */
|
alpar@9
|
1391 for (p = Pe [j] + 1 ; ok && p <= Pe [j] + ln - 1 ; p++)
|
alpar@9
|
1392 {
|
alpar@9
|
1393 ASSERT (Iw [p] >= 0 && Iw [p] < n) ;
|
alpar@9
|
1394 if (W [Iw [p]] != wflg) ok = 0 ;
|
alpar@9
|
1395 }
|
alpar@9
|
1396 if (ok)
|
alpar@9
|
1397 {
|
alpar@9
|
1398 /* --------------------------------------------- */
|
alpar@9
|
1399 /* found it! j can be absorbed into i */
|
alpar@9
|
1400 /* --------------------------------------------- */
|
alpar@9
|
1401
|
alpar@9
|
1402 AMD_DEBUG1 (("found it! j "ID" => i "ID"\n", j,i));
|
alpar@9
|
1403 Pe [j] = FLIP (i) ;
|
alpar@9
|
1404 /* both Nv [i] and Nv [j] are negated since they */
|
alpar@9
|
1405 /* are in Lme, and the absolute values of each */
|
alpar@9
|
1406 /* are the number of variables in i and j: */
|
alpar@9
|
1407 Nv [i] += Nv [j] ;
|
alpar@9
|
1408 Nv [j] = 0 ;
|
alpar@9
|
1409 Elen [j] = EMPTY ;
|
alpar@9
|
1410 /* delete j from hash bucket */
|
alpar@9
|
1411 ASSERT (j != Next [j]) ;
|
alpar@9
|
1412 j = Next [j] ;
|
alpar@9
|
1413 Next [jlast] = j ;
|
alpar@9
|
1414
|
alpar@9
|
1415 }
|
alpar@9
|
1416 else
|
alpar@9
|
1417 {
|
alpar@9
|
1418 /* j cannot be absorbed into i */
|
alpar@9
|
1419 jlast = j ;
|
alpar@9
|
1420 ASSERT (j != Next [j]) ;
|
alpar@9
|
1421 j = Next [j] ;
|
alpar@9
|
1422 }
|
alpar@9
|
1423 ASSERT (j >= EMPTY && j < n) ;
|
alpar@9
|
1424 }
|
alpar@9
|
1425
|
alpar@9
|
1426 /* -----------------------------------------------------
|
alpar@9
|
1427 * no more variables can be absorbed into i
|
alpar@9
|
1428 * go to next i in bucket and clear flag array
|
alpar@9
|
1429 * ----------------------------------------------------- */
|
alpar@9
|
1430
|
alpar@9
|
1431 wflg++ ;
|
alpar@9
|
1432 i = Next [i] ;
|
alpar@9
|
1433 ASSERT (i >= EMPTY && i < n) ;
|
alpar@9
|
1434
|
alpar@9
|
1435 }
|
alpar@9
|
1436 }
|
alpar@9
|
1437 }
|
alpar@9
|
1438 AMD_DEBUG2 (("detect done\n")) ;
|
alpar@9
|
1439
|
alpar@9
|
1440 /* ========================================================================= */
|
alpar@9
|
1441 /* RESTORE DEGREE LISTS AND REMOVE NONPRINCIPAL SUPERVARIABLES FROM ELEMENT */
|
alpar@9
|
1442 /* ========================================================================= */
|
alpar@9
|
1443
|
alpar@9
|
1444 p = pme1 ;
|
alpar@9
|
1445 nleft = n - nel ;
|
alpar@9
|
1446 for (pme = pme1 ; pme <= pme2 ; pme++)
|
alpar@9
|
1447 {
|
alpar@9
|
1448 i = Iw [pme] ;
|
alpar@9
|
1449 ASSERT (i >= 0 && i < n) ;
|
alpar@9
|
1450 nvi = -Nv [i] ;
|
alpar@9
|
1451 AMD_DEBUG3 (("Restore i "ID" "ID"\n", i, nvi)) ;
|
alpar@9
|
1452 if (nvi > 0)
|
alpar@9
|
1453 {
|
alpar@9
|
1454 /* i is a principal variable in Lme */
|
alpar@9
|
1455 /* restore Nv [i] to signify that i is principal */
|
alpar@9
|
1456 Nv [i] = nvi ;
|
alpar@9
|
1457
|
alpar@9
|
1458 /* --------------------------------------------------------- */
|
alpar@9
|
1459 /* compute the external degree (add size of current element) */
|
alpar@9
|
1460 /* --------------------------------------------------------- */
|
alpar@9
|
1461
|
alpar@9
|
1462 deg = Degree [i] + degme - nvi ;
|
alpar@9
|
1463 deg = MIN (deg, nleft - nvi) ;
|
alpar@9
|
1464 ASSERT (IMPLIES (aggressive, deg > 0) && deg >= 0 && deg < n) ;
|
alpar@9
|
1465
|
alpar@9
|
1466 /* --------------------------------------------------------- */
|
alpar@9
|
1467 /* place the supervariable at the head of the degree list */
|
alpar@9
|
1468 /* --------------------------------------------------------- */
|
alpar@9
|
1469
|
alpar@9
|
1470 inext = Head [deg] ;
|
alpar@9
|
1471 ASSERT (inext >= EMPTY && inext < n) ;
|
alpar@9
|
1472 if (inext != EMPTY) Last [inext] = i ;
|
alpar@9
|
1473 Next [i] = inext ;
|
alpar@9
|
1474 Last [i] = EMPTY ;
|
alpar@9
|
1475 Head [deg] = i ;
|
alpar@9
|
1476
|
alpar@9
|
1477 /* --------------------------------------------------------- */
|
alpar@9
|
1478 /* save the new degree, and find the minimum degree */
|
alpar@9
|
1479 /* --------------------------------------------------------- */
|
alpar@9
|
1480
|
alpar@9
|
1481 mindeg = MIN (mindeg, deg) ;
|
alpar@9
|
1482 Degree [i] = deg ;
|
alpar@9
|
1483
|
alpar@9
|
1484 /* --------------------------------------------------------- */
|
alpar@9
|
1485 /* place the supervariable in the element pattern */
|
alpar@9
|
1486 /* --------------------------------------------------------- */
|
alpar@9
|
1487
|
alpar@9
|
1488 Iw [p++] = i ;
|
alpar@9
|
1489
|
alpar@9
|
1490 }
|
alpar@9
|
1491 }
|
alpar@9
|
1492 AMD_DEBUG2 (("restore done\n")) ;
|
alpar@9
|
1493
|
alpar@9
|
1494 /* ========================================================================= */
|
alpar@9
|
1495 /* FINALIZE THE NEW ELEMENT */
|
alpar@9
|
1496 /* ========================================================================= */
|
alpar@9
|
1497
|
alpar@9
|
1498 AMD_DEBUG2 (("ME = "ID" DONE\n", me)) ;
|
alpar@9
|
1499 Nv [me] = nvpiv ;
|
alpar@9
|
1500 /* save the length of the list for the new element me */
|
alpar@9
|
1501 Len [me] = p - pme1 ;
|
alpar@9
|
1502 if (Len [me] == 0)
|
alpar@9
|
1503 {
|
alpar@9
|
1504 /* there is nothing left of the current pivot element */
|
alpar@9
|
1505 /* it is a root of the assembly tree */
|
alpar@9
|
1506 Pe [me] = EMPTY ;
|
alpar@9
|
1507 W [me] = 0 ;
|
alpar@9
|
1508 }
|
alpar@9
|
1509 if (elenme != 0)
|
alpar@9
|
1510 {
|
alpar@9
|
1511 /* element was not constructed in place: deallocate part of */
|
alpar@9
|
1512 /* it since newly nonprincipal variables may have been removed */
|
alpar@9
|
1513 pfree = p ;
|
alpar@9
|
1514 }
|
alpar@9
|
1515
|
alpar@9
|
1516 /* The new element has nvpiv pivots and the size of the contribution
|
alpar@9
|
1517 * block for a multifrontal method is degme-by-degme, not including
|
alpar@9
|
1518 * the "dense" rows/columns. If the "dense" rows/columns are included,
|
alpar@9
|
1519 * the frontal matrix is no larger than
|
alpar@9
|
1520 * (degme+ndense)-by-(degme+ndense).
|
alpar@9
|
1521 */
|
alpar@9
|
1522
|
alpar@9
|
1523 if (Info != (double *) NULL)
|
alpar@9
|
1524 {
|
alpar@9
|
1525 f = nvpiv ;
|
alpar@9
|
1526 r = degme + ndense ;
|
alpar@9
|
1527 dmax = MAX (dmax, f + r) ;
|
alpar@9
|
1528
|
alpar@9
|
1529 /* number of nonzeros in L (excluding the diagonal) */
|
alpar@9
|
1530 lnzme = f*r + (f-1)*f/2 ;
|
alpar@9
|
1531 lnz += lnzme ;
|
alpar@9
|
1532
|
alpar@9
|
1533 /* number of divide operations for LDL' and for LU */
|
alpar@9
|
1534 ndiv += lnzme ;
|
alpar@9
|
1535
|
alpar@9
|
1536 /* number of multiply-subtract pairs for LU */
|
alpar@9
|
1537 s = f*r*r + r*(f-1)*f + (f-1)*f*(2*f-1)/6 ;
|
alpar@9
|
1538 nms_lu += s ;
|
alpar@9
|
1539
|
alpar@9
|
1540 /* number of multiply-subtract pairs for LDL' */
|
alpar@9
|
1541 nms_ldl += (s + lnzme)/2 ;
|
alpar@9
|
1542 }
|
alpar@9
|
1543
|
alpar@9
|
1544 #ifndef NDEBUG
|
alpar@9
|
1545 AMD_DEBUG2 (("finalize done nel "ID" n "ID"\n ::::\n", nel, n)) ;
|
alpar@9
|
1546 for (pme = Pe [me] ; pme <= Pe [me] + Len [me] - 1 ; pme++)
|
alpar@9
|
1547 {
|
alpar@9
|
1548 AMD_DEBUG3 ((" "ID"", Iw [pme])) ;
|
alpar@9
|
1549 }
|
alpar@9
|
1550 AMD_DEBUG3 (("\n")) ;
|
alpar@9
|
1551 #endif
|
alpar@9
|
1552
|
alpar@9
|
1553 }
|
alpar@9
|
1554
|
alpar@9
|
1555 /* ========================================================================= */
|
alpar@9
|
1556 /* DONE SELECTING PIVOTS */
|
alpar@9
|
1557 /* ========================================================================= */
|
alpar@9
|
1558
|
alpar@9
|
1559 if (Info != (double *) NULL)
|
alpar@9
|
1560 {
|
alpar@9
|
1561
|
alpar@9
|
1562 /* count the work to factorize the ndense-by-ndense submatrix */
|
alpar@9
|
1563 f = ndense ;
|
alpar@9
|
1564 dmax = MAX (dmax, (double) ndense) ;
|
alpar@9
|
1565
|
alpar@9
|
1566 /* number of nonzeros in L (excluding the diagonal) */
|
alpar@9
|
1567 lnzme = (f-1)*f/2 ;
|
alpar@9
|
1568 lnz += lnzme ;
|
alpar@9
|
1569
|
alpar@9
|
1570 /* number of divide operations for LDL' and for LU */
|
alpar@9
|
1571 ndiv += lnzme ;
|
alpar@9
|
1572
|
alpar@9
|
1573 /* number of multiply-subtract pairs for LU */
|
alpar@9
|
1574 s = (f-1)*f*(2*f-1)/6 ;
|
alpar@9
|
1575 nms_lu += s ;
|
alpar@9
|
1576
|
alpar@9
|
1577 /* number of multiply-subtract pairs for LDL' */
|
alpar@9
|
1578 nms_ldl += (s + lnzme)/2 ;
|
alpar@9
|
1579
|
alpar@9
|
1580 /* number of nz's in L (excl. diagonal) */
|
alpar@9
|
1581 Info [AMD_LNZ] = lnz ;
|
alpar@9
|
1582
|
alpar@9
|
1583 /* number of divide ops for LU and LDL' */
|
alpar@9
|
1584 Info [AMD_NDIV] = ndiv ;
|
alpar@9
|
1585
|
alpar@9
|
1586 /* number of multiply-subtract pairs for LDL' */
|
alpar@9
|
1587 Info [AMD_NMULTSUBS_LDL] = nms_ldl ;
|
alpar@9
|
1588
|
alpar@9
|
1589 /* number of multiply-subtract pairs for LU */
|
alpar@9
|
1590 Info [AMD_NMULTSUBS_LU] = nms_lu ;
|
alpar@9
|
1591
|
alpar@9
|
1592 /* number of "dense" rows/columns */
|
alpar@9
|
1593 Info [AMD_NDENSE] = ndense ;
|
alpar@9
|
1594
|
alpar@9
|
1595 /* largest front is dmax-by-dmax */
|
alpar@9
|
1596 Info [AMD_DMAX] = dmax ;
|
alpar@9
|
1597
|
alpar@9
|
1598 /* number of garbage collections in AMD */
|
alpar@9
|
1599 Info [AMD_NCMPA] = ncmpa ;
|
alpar@9
|
1600
|
alpar@9
|
1601 /* successful ordering */
|
alpar@9
|
1602 Info [AMD_STATUS] = AMD_OK ;
|
alpar@9
|
1603 }
|
alpar@9
|
1604
|
alpar@9
|
1605 /* ========================================================================= */
|
alpar@9
|
1606 /* POST-ORDERING */
|
alpar@9
|
1607 /* ========================================================================= */
|
alpar@9
|
1608
|
alpar@9
|
1609 /* -------------------------------------------------------------------------
|
alpar@9
|
1610 * Variables at this point:
|
alpar@9
|
1611 *
|
alpar@9
|
1612 * Pe: holds the elimination tree. The parent of j is FLIP (Pe [j]),
|
alpar@9
|
1613 * or EMPTY if j is a root. The tree holds both elements and
|
alpar@9
|
1614 * non-principal (unordered) variables absorbed into them.
|
alpar@9
|
1615 * Dense variables are non-principal and unordered.
|
alpar@9
|
1616 *
|
alpar@9
|
1617 * Elen: holds the size of each element, including the diagonal part.
|
alpar@9
|
1618 * FLIP (Elen [e]) > 0 if e is an element. For unordered
|
alpar@9
|
1619 * variables i, Elen [i] is EMPTY.
|
alpar@9
|
1620 *
|
alpar@9
|
1621 * Nv: Nv [e] > 0 is the number of pivots represented by the element e.
|
alpar@9
|
1622 * For unordered variables i, Nv [i] is zero.
|
alpar@9
|
1623 *
|
alpar@9
|
1624 * Contents no longer needed:
|
alpar@9
|
1625 * W, Iw, Len, Degree, Head, Next, Last.
|
alpar@9
|
1626 *
|
alpar@9
|
1627 * The matrix itself has been destroyed.
|
alpar@9
|
1628 *
|
alpar@9
|
1629 * n: the size of the matrix.
|
alpar@9
|
1630 * No other scalars needed (pfree, iwlen, etc.)
|
alpar@9
|
1631 * ------------------------------------------------------------------------- */
|
alpar@9
|
1632
|
alpar@9
|
1633 /* restore Pe */
|
alpar@9
|
1634 for (i = 0 ; i < n ; i++)
|
alpar@9
|
1635 {
|
alpar@9
|
1636 Pe [i] = FLIP (Pe [i]) ;
|
alpar@9
|
1637 }
|
alpar@9
|
1638
|
alpar@9
|
1639 /* restore Elen, for output information, and for postordering */
|
alpar@9
|
1640 for (i = 0 ; i < n ; i++)
|
alpar@9
|
1641 {
|
alpar@9
|
1642 Elen [i] = FLIP (Elen [i]) ;
|
alpar@9
|
1643 }
|
alpar@9
|
1644
|
alpar@9
|
1645 /* Now the parent of j is Pe [j], or EMPTY if j is a root. Elen [e] > 0
|
alpar@9
|
1646 * is the size of element e. Elen [i] is EMPTY for unordered variable i. */
|
alpar@9
|
1647
|
alpar@9
|
1648 #ifndef NDEBUG
|
alpar@9
|
1649 AMD_DEBUG2 (("\nTree:\n")) ;
|
alpar@9
|
1650 for (i = 0 ; i < n ; i++)
|
alpar@9
|
1651 {
|
alpar@9
|
1652 AMD_DEBUG2 ((" "ID" parent: "ID" ", i, Pe [i])) ;
|
alpar@9
|
1653 ASSERT (Pe [i] >= EMPTY && Pe [i] < n) ;
|
alpar@9
|
1654 if (Nv [i] > 0)
|
alpar@9
|
1655 {
|
alpar@9
|
1656 /* this is an element */
|
alpar@9
|
1657 e = i ;
|
alpar@9
|
1658 AMD_DEBUG2 ((" element, size is "ID"\n", Elen [i])) ;
|
alpar@9
|
1659 ASSERT (Elen [e] > 0) ;
|
alpar@9
|
1660 }
|
alpar@9
|
1661 AMD_DEBUG2 (("\n")) ;
|
alpar@9
|
1662 }
|
alpar@9
|
1663 AMD_DEBUG2 (("\nelements:\n")) ;
|
alpar@9
|
1664 for (e = 0 ; e < n ; e++)
|
alpar@9
|
1665 {
|
alpar@9
|
1666 if (Nv [e] > 0)
|
alpar@9
|
1667 {
|
alpar@9
|
1668 AMD_DEBUG3 (("Element e= "ID" size "ID" nv "ID" \n", e,
|
alpar@9
|
1669 Elen [e], Nv [e])) ;
|
alpar@9
|
1670 }
|
alpar@9
|
1671 }
|
alpar@9
|
1672 AMD_DEBUG2 (("\nvariables:\n")) ;
|
alpar@9
|
1673 for (i = 0 ; i < n ; i++)
|
alpar@9
|
1674 {
|
alpar@9
|
1675 Int cnt ;
|
alpar@9
|
1676 if (Nv [i] == 0)
|
alpar@9
|
1677 {
|
alpar@9
|
1678 AMD_DEBUG3 (("i unordered: "ID"\n", i)) ;
|
alpar@9
|
1679 j = Pe [i] ;
|
alpar@9
|
1680 cnt = 0 ;
|
alpar@9
|
1681 AMD_DEBUG3 ((" j: "ID"\n", j)) ;
|
alpar@9
|
1682 if (j == EMPTY)
|
alpar@9
|
1683 {
|
alpar@9
|
1684 AMD_DEBUG3 ((" i is a dense variable\n")) ;
|
alpar@9
|
1685 }
|
alpar@9
|
1686 else
|
alpar@9
|
1687 {
|
alpar@9
|
1688 ASSERT (j >= 0 && j < n) ;
|
alpar@9
|
1689 while (Nv [j] == 0)
|
alpar@9
|
1690 {
|
alpar@9
|
1691 AMD_DEBUG3 ((" j : "ID"\n", j)) ;
|
alpar@9
|
1692 j = Pe [j] ;
|
alpar@9
|
1693 AMD_DEBUG3 ((" j:: "ID"\n", j)) ;
|
alpar@9
|
1694 cnt++ ;
|
alpar@9
|
1695 if (cnt > n) break ;
|
alpar@9
|
1696 }
|
alpar@9
|
1697 e = j ;
|
alpar@9
|
1698 AMD_DEBUG3 ((" got to e: "ID"\n", e)) ;
|
alpar@9
|
1699 }
|
alpar@9
|
1700 }
|
alpar@9
|
1701 }
|
alpar@9
|
1702 #endif
|
alpar@9
|
1703
|
alpar@9
|
1704 /* ========================================================================= */
|
alpar@9
|
1705 /* compress the paths of the variables */
|
alpar@9
|
1706 /* ========================================================================= */
|
alpar@9
|
1707
|
alpar@9
|
1708 for (i = 0 ; i < n ; i++)
|
alpar@9
|
1709 {
|
alpar@9
|
1710 if (Nv [i] == 0)
|
alpar@9
|
1711 {
|
alpar@9
|
1712
|
alpar@9
|
1713 /* -------------------------------------------------------------
|
alpar@9
|
1714 * i is an un-ordered row. Traverse the tree from i until
|
alpar@9
|
1715 * reaching an element, e. The element, e, was the principal
|
alpar@9
|
1716 * supervariable of i and all nodes in the path from i to when e
|
alpar@9
|
1717 * was selected as pivot.
|
alpar@9
|
1718 * ------------------------------------------------------------- */
|
alpar@9
|
1719
|
alpar@9
|
1720 AMD_DEBUG1 (("Path compression, i unordered: "ID"\n", i)) ;
|
alpar@9
|
1721 j = Pe [i] ;
|
alpar@9
|
1722 ASSERT (j >= EMPTY && j < n) ;
|
alpar@9
|
1723 AMD_DEBUG3 ((" j: "ID"\n", j)) ;
|
alpar@9
|
1724 if (j == EMPTY)
|
alpar@9
|
1725 {
|
alpar@9
|
1726 /* Skip a dense variable. It has no parent. */
|
alpar@9
|
1727 AMD_DEBUG3 ((" i is a dense variable\n")) ;
|
alpar@9
|
1728 continue ;
|
alpar@9
|
1729 }
|
alpar@9
|
1730
|
alpar@9
|
1731 /* while (j is a variable) */
|
alpar@9
|
1732 while (Nv [j] == 0)
|
alpar@9
|
1733 {
|
alpar@9
|
1734 AMD_DEBUG3 ((" j : "ID"\n", j)) ;
|
alpar@9
|
1735 j = Pe [j] ;
|
alpar@9
|
1736 AMD_DEBUG3 ((" j:: "ID"\n", j)) ;
|
alpar@9
|
1737 ASSERT (j >= 0 && j < n) ;
|
alpar@9
|
1738 }
|
alpar@9
|
1739 /* got to an element e */
|
alpar@9
|
1740 e = j ;
|
alpar@9
|
1741 AMD_DEBUG3 (("got to e: "ID"\n", e)) ;
|
alpar@9
|
1742
|
alpar@9
|
1743 /* -------------------------------------------------------------
|
alpar@9
|
1744 * traverse the path again from i to e, and compress the path
|
alpar@9
|
1745 * (all nodes point to e). Path compression allows this code to
|
alpar@9
|
1746 * compute in O(n) time.
|
alpar@9
|
1747 * ------------------------------------------------------------- */
|
alpar@9
|
1748
|
alpar@9
|
1749 j = i ;
|
alpar@9
|
1750 /* while (j is a variable) */
|
alpar@9
|
1751 while (Nv [j] == 0)
|
alpar@9
|
1752 {
|
alpar@9
|
1753 jnext = Pe [j] ;
|
alpar@9
|
1754 AMD_DEBUG3 (("j "ID" jnext "ID"\n", j, jnext)) ;
|
alpar@9
|
1755 Pe [j] = e ;
|
alpar@9
|
1756 j = jnext ;
|
alpar@9
|
1757 ASSERT (j >= 0 && j < n) ;
|
alpar@9
|
1758 }
|
alpar@9
|
1759 }
|
alpar@9
|
1760 }
|
alpar@9
|
1761
|
alpar@9
|
1762 /* ========================================================================= */
|
alpar@9
|
1763 /* postorder the assembly tree */
|
alpar@9
|
1764 /* ========================================================================= */
|
alpar@9
|
1765
|
alpar@9
|
1766 AMD_postorder (n, Pe, Nv, Elen,
|
alpar@9
|
1767 W, /* output order */
|
alpar@9
|
1768 Head, Next, Last) ; /* workspace */
|
alpar@9
|
1769
|
alpar@9
|
1770 /* ========================================================================= */
|
alpar@9
|
1771 /* compute output permutation and inverse permutation */
|
alpar@9
|
1772 /* ========================================================================= */
|
alpar@9
|
1773
|
alpar@9
|
1774 /* W [e] = k means that element e is the kth element in the new
|
alpar@9
|
1775 * order. e is in the range 0 to n-1, and k is in the range 0 to
|
alpar@9
|
1776 * the number of elements. Use Head for inverse order. */
|
alpar@9
|
1777
|
alpar@9
|
1778 for (k = 0 ; k < n ; k++)
|
alpar@9
|
1779 {
|
alpar@9
|
1780 Head [k] = EMPTY ;
|
alpar@9
|
1781 Next [k] = EMPTY ;
|
alpar@9
|
1782 }
|
alpar@9
|
1783 for (e = 0 ; e < n ; e++)
|
alpar@9
|
1784 {
|
alpar@9
|
1785 k = W [e] ;
|
alpar@9
|
1786 ASSERT ((k == EMPTY) == (Nv [e] == 0)) ;
|
alpar@9
|
1787 if (k != EMPTY)
|
alpar@9
|
1788 {
|
alpar@9
|
1789 ASSERT (k >= 0 && k < n) ;
|
alpar@9
|
1790 Head [k] = e ;
|
alpar@9
|
1791 }
|
alpar@9
|
1792 }
|
alpar@9
|
1793
|
alpar@9
|
1794 /* construct output inverse permutation in Next,
|
alpar@9
|
1795 * and permutation in Last */
|
alpar@9
|
1796 nel = 0 ;
|
alpar@9
|
1797 for (k = 0 ; k < n ; k++)
|
alpar@9
|
1798 {
|
alpar@9
|
1799 e = Head [k] ;
|
alpar@9
|
1800 if (e == EMPTY) break ;
|
alpar@9
|
1801 ASSERT (e >= 0 && e < n && Nv [e] > 0) ;
|
alpar@9
|
1802 Next [e] = nel ;
|
alpar@9
|
1803 nel += Nv [e] ;
|
alpar@9
|
1804 }
|
alpar@9
|
1805 ASSERT (nel == n - ndense) ;
|
alpar@9
|
1806
|
alpar@9
|
1807 /* order non-principal variables (dense, & those merged into supervar's) */
|
alpar@9
|
1808 for (i = 0 ; i < n ; i++)
|
alpar@9
|
1809 {
|
alpar@9
|
1810 if (Nv [i] == 0)
|
alpar@9
|
1811 {
|
alpar@9
|
1812 e = Pe [i] ;
|
alpar@9
|
1813 ASSERT (e >= EMPTY && e < n) ;
|
alpar@9
|
1814 if (e != EMPTY)
|
alpar@9
|
1815 {
|
alpar@9
|
1816 /* This is an unordered variable that was merged
|
alpar@9
|
1817 * into element e via supernode detection or mass
|
alpar@9
|
1818 * elimination of i when e became the pivot element.
|
alpar@9
|
1819 * Place i in order just before e. */
|
alpar@9
|
1820 ASSERT (Next [i] == EMPTY && Nv [e] > 0) ;
|
alpar@9
|
1821 Next [i] = Next [e] ;
|
alpar@9
|
1822 Next [e]++ ;
|
alpar@9
|
1823 }
|
alpar@9
|
1824 else
|
alpar@9
|
1825 {
|
alpar@9
|
1826 /* This is a dense unordered variable, with no parent.
|
alpar@9
|
1827 * Place it last in the output order. */
|
alpar@9
|
1828 Next [i] = nel++ ;
|
alpar@9
|
1829 }
|
alpar@9
|
1830 }
|
alpar@9
|
1831 }
|
alpar@9
|
1832 ASSERT (nel == n) ;
|
alpar@9
|
1833
|
alpar@9
|
1834 AMD_DEBUG2 (("\n\nPerm:\n")) ;
|
alpar@9
|
1835 for (i = 0 ; i < n ; i++)
|
alpar@9
|
1836 {
|
alpar@9
|
1837 k = Next [i] ;
|
alpar@9
|
1838 ASSERT (k >= 0 && k < n) ;
|
alpar@9
|
1839 Last [k] = i ;
|
alpar@9
|
1840 AMD_DEBUG2 ((" perm ["ID"] = "ID"\n", k, i)) ;
|
alpar@9
|
1841 }
|
alpar@9
|
1842 }
|