| 1 |
/*! \file SparseRenumbering.cpp |
| 2 |
\brief tridiagonal factorization algorithm with Cuthill-McKee |
| 3 |
\author François-Xavier Roux, ONERA, Laboratoire Jacques-Louis Lions |
| 4 |
\author Atsushi Suzuki, Laboratoire Jacques-Louis Lions |
| 5 |
\date Jul. 12th 2015 |
| 6 |
\date Nov. 30th 2016 |
| 7 |
*/ |
| 8 |
|
| 9 |
// This file is part of Dissection |
| 10 |
// |
| 11 |
// Dissection is free software: you can redistribute it and/or modify |
| 12 |
// it 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 |
// Linking Dissection statically or dynamically with other modules is making |
| 17 |
// a combined work based on Disssection. Thus, the terms and conditions of |
| 18 |
// the GNU General Public License cover the whole combination. |
| 19 |
// |
| 20 |
// As a special exception, the copyright holders of Dissection give you |
| 21 |
// permission to combine Dissection program with free software programs or |
| 22 |
// libraries that are released under the GNU LGPL and with independent modules |
| 23 |
// that communicate with Dissection solely through the Dissection-fortran |
| 24 |
// interface. You may copy and distribute such a system following the terms of |
| 25 |
// the GNU GPL for Dissection and the licenses of the other code concerned, |
| 26 |
// provided that you include the source code of that other code when and as |
| 27 |
// the GNU GPL requires distribution of source code and provided that you do |
| 28 |
// not modify the Dissection-fortran interface. |
| 29 |
// |
| 30 |
// Note that people who make modified versions of Dissection are not obligated |
| 31 |
// to grant this special exception for their modified versions; it is their |
| 32 |
// choice whether to do so. The GNU General Public License gives permission to |
| 33 |
// release a modified version without this exception; this exception also makes |
| 34 |
// it possible to release a modified version which carries forward this |
| 35 |
// exception. If you modify the Dissection-fortran interface, this exception |
| 36 |
// does not apply to your modified version of Dissection, and you must remove |
| 37 |
// this exception when you distribute your modified version. |
| 38 |
// |
| 39 |
// This exception is an additional permission under section 7 of the GNU |
| 40 |
// General Public License, version 3 ("GPLv3") |
| 41 |
// |
| 42 |
// Dissection is distributed in the hope that it will be useful, |
| 43 |
// but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 44 |
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 45 |
// GNU General Public License for more details. |
| 46 |
// |
| 47 |
// You should have received a copy of the GNU General Public License |
| 48 |
// along with Dissection. If not, see <http://www.gnu.org/licenses/>. |
| 49 |
// |
| 50 |
|
| 51 |
#include <cstdio> |
| 52 |
#include <cstdlib> |
| 53 |
#include <list> |
| 54 |
#include <algorithm> |
| 55 |
#include "Algebra/SparseRenumbering.hpp" |
| 56 |
#include "Driver/DissectionDefault.hpp" |
| 57 |
#include "Compiler/arithmetic.hpp" |
| 58 |
#include "Compiler/DissectionIO.hpp" |
| 59 |
|
| 60 |
using std::list; |
| 61 |
|
| 62 |
void CMK_number(const int dim, const int *ptrows, const int *indcols, |
| 63 |
vector<int> &new2old, const bool verbose, FILE *fp) |
| 64 |
{ |
| 65 |
vector<int> list, indic, connect; |
| 66 |
vector<bool> mask; |
| 67 |
int i1_prev, i2_prev, i0, i1; |
| 68 |
double profile, min_profile; |
| 69 |
|
| 70 |
connect.resize(dim); |
| 71 |
list.resize(dim); |
| 72 |
indic.resize(dim); |
| 73 |
mask.resize(dim); |
| 74 |
|
| 75 |
// compute initial profile |
| 76 |
i1_prev = 0; |
| 77 |
i2_prev = indcols[ptrows[0]]; |
| 78 |
for (int k = (ptrows[0] + 1); k < ptrows[1]; k++) { |
| 79 |
i2_prev = indcols[k] > i2_prev ? indcols[k] : i2_prev; |
| 80 |
} |
| 81 |
min_profile = ((double)(i2_prev - i1_prev + 2) |
| 82 |
* (double)(i2_prev - i1_prev + 1)) * 0.5; |
| 83 |
while (i2_prev < (dim - 1)) { |
| 84 |
int i1 = i2_prev + 1; |
| 85 |
int i2 = indcols[ptrows[i1_prev]]; |
| 86 |
for (int k = (ptrows[i1_prev] + 1); k < ptrows[i2_prev + 1]; k++) { |
| 87 |
i2 = indcols[k] > i2 ? indcols[k] : i2; |
| 88 |
} |
| 89 |
min_profile += (double)(i2 - i1 + 2) * (double)(i2 - i1 + 1) * 0.5; |
| 90 |
i1_prev = i1; |
| 91 |
i2_prev = i2; |
| 92 |
} |
| 93 |
diss_printf(verbose, fp, |
| 94 |
"%s %d : Cuthill McKee renumbering : init block : %g :: ", |
| 95 |
__FILE__, __LINE__, min_profile); |
| 96 |
|
| 97 |
// initialize frontal renumbering from an excentric point |
| 98 |
for (int i = 0; i < dim; i++) { |
| 99 |
connect[i] = ptrows[i + 1] - ptrows[i]; |
| 100 |
mask[i] = false; |
| 101 |
} |
| 102 |
i0 = 0; // initialization : critical when connent[] == dim |
| 103 |
{ |
| 104 |
int imin = dim; // find index that attains minimun of connect[] |
| 105 |
for (int i = 0; i < dim; i++) { |
| 106 |
if (imin > connect[i]) { |
| 107 |
imin = connect[i]; |
| 108 |
i0 = i; |
| 109 |
} |
| 110 |
} |
| 111 |
} |
| 112 |
mask[i0] = true; |
| 113 |
profile = frontal_numb(dim, ptrows, indcols, i0, list, indic, connect); |
| 114 |
i1 = list[dim - 1]; |
| 115 |
|
| 116 |
diss_printf(verbose, fp, "- > node %d : %g :: ", i0, profile); |
| 117 |
if (profile < min_profile) { |
| 118 |
min_profile = profile; |
| 119 |
} |
| 120 |
else { |
| 121 |
for (int i = 0; i < dim; i++) { |
| 122 |
list[i] = i; |
| 123 |
} |
| 124 |
} |
| 125 |
profile = min_profile; |
| 126 |
min_profile = min_profile + 1.0; |
| 127 |
|
| 128 |
// repeat frontal numbering as long as profile improves |
| 129 |
while (profile < min_profile) { |
| 130 |
min_profile = profile; |
| 131 |
for (int i = 0; i < dim; i++) { |
| 132 |
new2old[i] = list[i]; |
| 133 |
} |
| 134 |
for (int i = 0; i < dim; i++) { |
| 135 |
connect[i] = ptrows[i + 1] - ptrows[i]; |
| 136 |
} |
| 137 |
{ |
| 138 |
i0 = 0; // initialization : critical when connent[] == dim |
| 139 |
int imin = dim; |
| 140 |
for (int k = ptrows[i1]; k < ptrows[i1 + 1]; k++) { |
| 141 |
const int ii = indcols[k]; |
| 142 |
if (imin > connect[ii]) { |
| 143 |
imin = connect[ii]; |
| 144 |
i0 = ii; |
| 145 |
} |
| 146 |
} |
| 147 |
if (mask[i0]) { |
| 148 |
for (int i = (dim - 1); i >=0; i--) { |
| 149 |
if (!mask[list[i]]) { |
| 150 |
i0 = list[i]; |
| 151 |
break; |
| 152 |
} |
| 153 |
} |
| 154 |
} |
| 155 |
} |
| 156 |
mask[i0] = true; |
| 157 |
profile = frontal_numb(dim, ptrows, indcols, i0, list, indic, connect); |
| 158 |
i1 = list[dim - 1]; |
| 159 |
diss_printf(verbose, fp, "- > node: %d : %g :: ", i0, profile); |
| 160 |
} // while |
| 161 |
|
| 162 |
diss_printf(verbose, fp, "-> optimazied : %g\n", min_profile); |
| 163 |
} |
| 164 |
|
| 165 |
double frontal_numb(const int dim, const int *ptrows, const int *indcols, |
| 166 |
const int i0, |
| 167 |
vector<int> &list, |
| 168 |
vector<int> &indic, vector<int> &connect) |
| 169 |
{ |
| 170 |
int n, nn, i1, i2, i1_prev, i2_prev, k1, k2; |
| 171 |
double profile; |
| 172 |
// compute nodal connectivity degree |
| 173 |
for (int i = 0; i < dim; i++) { |
| 174 |
connect[i] = ptrows[i + 1] - ptrows[i]; |
| 175 |
} |
| 176 |
#if 1 |
| 177 |
for (int i = 0; i < dim; i++) { |
| 178 |
indic[i] = 0; |
| 179 |
} |
| 180 |
n = 0; |
| 181 |
nn = (-1); |
| 182 |
list[0] = i0; |
| 183 |
indic[i0] = 1; |
| 184 |
while (n < (dim - 1)) { |
| 185 |
nn++; |
| 186 |
int j = list[nn]; |
| 187 |
// renumber un-numbered neigbours of j, in increasing connectivity order |
| 188 |
int k = 0; |
| 189 |
for (int l = ptrows[j]; l < ptrows[j + 1]; l++) { |
| 190 |
const int i = indcols[l]; |
| 191 |
if (indic[i] == 0) { |
| 192 |
int ki = 1; |
| 193 |
for ( ; ki <= k; ki++) { |
| 194 |
if (connect[i] < connect[list[n + ki]]) { |
| 195 |
break; |
| 196 |
} |
| 197 |
} |
| 198 |
k++; |
| 199 |
for (int kk = k; kk >= (ki + 1); kk--) { |
| 200 |
list[n + kk] = list[n + kk - 1]; |
| 201 |
} |
| 202 |
list[n + ki] = i; |
| 203 |
} // if (indic[i] == 0) |
| 204 |
} // update connectivity degree of un-numbered nodes |
| 205 |
for (int kk = (n + 1); kk <= (n + k); kk++) { |
| 206 |
const int ii = list[kk]; |
| 207 |
for (int m = ptrows[ii]; m < ptrows[ii + 1]; m++) { |
| 208 |
connect[indcols[m]]--; |
| 209 |
} |
| 210 |
} |
| 211 |
for (int kk = (n + 1); kk <= (n + k); kk++) { |
| 212 |
indic[list[kk]] = 1; |
| 213 |
} |
| 214 |
n += k; |
| 215 |
} // while (n < (dim - 1)) |
| 216 |
// build old to new correspondance |
| 217 |
for (int i = 0; i < dim; i++) { |
| 218 |
indic[list[i]] = i; |
| 219 |
} |
| 220 |
#else |
| 221 |
std::list<int> idexlist; |
| 222 |
vector<bool> mask; |
| 223 |
mask.resize(dim, false); |
| 224 |
idexlist.push_back(i0); |
| 225 |
std::list<int>::iterator it = idexlist.begin(); |
| 226 |
mask[(*it)] = true; |
| 227 |
n = 0; |
| 228 |
std::list<int>::iterator nt = idexlist.begin(); |
| 229 |
while (n < (dim - 1)) { |
| 230 |
int k = 0; |
| 231 |
for (int l = ptrows[(*it)]; l < ptrows[(*it) + 1]; l++) { |
| 232 |
const int ii = indcols[l]; |
| 233 |
if (!mask[ii]) { |
| 234 |
bool flag = false; |
| 235 |
std::list<int>::iterator kt = nt; |
| 236 |
++kt; |
| 237 |
for (int kk = 0 ; kk < k; kk++, ++kt) { |
| 238 |
if (connect[ii] < connect[(*kt)]) { |
| 239 |
idexlist.insert(kt, ii); |
| 240 |
flag = true; |
| 241 |
break; |
| 242 |
} |
| 243 |
} |
| 244 |
if (!flag) { |
| 245 |
idexlist.push_back(ii); |
| 246 |
} |
| 247 |
k++; |
| 248 |
} |
| 249 |
} |
| 250 |
++it; |
| 251 |
std::list<int>::iterator kt = nt; |
| 252 |
int kk = 0; |
| 253 |
for ( ; kk < k; ++kt, kk++) { |
| 254 |
for (int m = ptrows[(*kt)]; m < ptrows[(*kt) + 1]; m++) { |
| 255 |
connect[indcols[m]]--; |
| 256 |
} |
| 257 |
mask[(*kt)] = true; |
| 258 |
++nt; |
| 259 |
} |
| 260 |
n += k; |
| 261 |
} |
| 262 |
{ |
| 263 |
int i = 0; |
| 264 |
std::list<int>::iterator it = idexlist.begin(); |
| 265 |
for (; it != idexlist.end(); ++it, i++) { |
| 266 |
list[i] = (*it); |
| 267 |
indic[(*it)] = i; |
| 268 |
} |
| 269 |
} |
| 270 |
idexlist.clear(); |
| 271 |
#endif |
| 272 |
|
| 273 |
// compute profile |
| 274 |
i1_prev = 0; |
| 275 |
k1 = ptrows[list[i1_prev]]; |
| 276 |
k2 = ptrows[list[i1_prev] + 1]; |
| 277 |
i2_prev = indic[indcols[k1]]; |
| 278 |
for (int k = (k1 + 1); k < k2; k++) { |
| 279 |
i2_prev = indic[indcols[k]] > i2_prev ? indic[indcols[k]] : i2_prev; |
| 280 |
} |
| 281 |
profile = ((double)(i2_prev - i1_prev + 1) * |
| 282 |
(double)(i2_prev - i1_prev + 2) * 0.5); |
| 283 |
while (i2_prev < (dim - 1)) { |
| 284 |
i1 = i2_prev + 1; |
| 285 |
i2 = i1; |
| 286 |
for (int i = i1_prev; i <= i2_prev; i++) { |
| 287 |
k1 = ptrows[list[i]]; |
| 288 |
k2 = ptrows[list[i] + 1]; |
| 289 |
i2 = std::max<int>((int)i2, (int)indic[indcols[k1]]); |
| 290 |
for (int k = (k1 + 1); k < k2; k++) { |
| 291 |
i2 = indic[indcols[k]] > i2 ? indic[indcols[k]] : i2; |
| 292 |
} |
| 293 |
} |
| 294 |
profile += ((double)(i2 - i1 + 1) * (double)(i2 - i1 + 2) * 0.5); |
| 295 |
i1_prev = i1; |
| 296 |
i2_prev = i2; |
| 297 |
} |
| 298 |
return profile; |
| 299 |
} |
| 300 |
|
| 301 |
int point_front(const int dim, const int *ptrows, const int *indcols, |
| 302 |
vector<int> &new2old, vector<int> &p_front, const bool verbose, |
| 303 |
FILE *fp) |
| 304 |
{ |
| 305 |
int nfront, ilast; |
| 306 |
vector<int> old2new, p_front1; |
| 307 |
vector<bool> mask; |
| 308 |
const int size_thrs = 24; |
| 309 |
old2new.resize(dim); |
| 310 |
for (int i = 0; i < dim; i++) { |
| 311 |
old2new[new2old[i]] = i; |
| 312 |
} |
| 313 |
p_front.resize(dim + 1); // for safty |
| 314 |
|
| 315 |
nfront = 0; |
| 316 |
p_front[0] = 0; |
| 317 |
ilast = (-1); |
| 318 |
while (ilast < (dim - 1)) { // loop ilast |
| 319 |
ilast++; |
| 320 |
bool flag_cont = false; |
| 321 |
bool flag_break = false; |
| 322 |
const int ii = new2old[ilast]; |
| 323 |
nfront++; |
| 324 |
int isup = (-1); |
| 325 |
for (int k = ptrows[ii]; k < ptrows[ii + 1]; k++) { |
| 326 |
const int jj = old2new[indcols[k]]; |
| 327 |
isup = jj > isup ? jj : isup; |
| 328 |
} |
| 329 |
p_front[nfront] = isup + 1; |
| 330 |
while (isup > ilast) { |
| 331 |
ilast = isup; |
| 332 |
for (int i = p_front[nfront - 1]; i < p_front[nfront]; i++) { |
| 333 |
const int ii = new2old[i]; |
| 334 |
for (int k = ptrows[ii]; k < ptrows[ii + 1]; k++) { |
| 335 |
const int jj = old2new[indcols[k]]; |
| 336 |
isup = jj > isup ? jj : isup; |
| 337 |
} |
| 338 |
} // loop : i |
| 339 |
if (isup == ilast) { |
| 340 |
flag_cont = true; |
| 341 |
break; // continue loop_ilast |
| 342 |
} |
| 343 |
nfront++; |
| 344 |
p_front[nfront] = isup + 1; |
| 345 |
if (isup == (dim - 1)) { |
| 346 |
flag_break = true; |
| 347 |
break; // exit loop_ilast |
| 348 |
} |
| 349 |
} |
| 350 |
if (flag_cont) { |
| 351 |
continue; |
| 352 |
} |
| 353 |
if (flag_break) { |
| 354 |
break; |
| 355 |
} |
| 356 |
} // loop ilast |
| 357 |
old2new.clear(); |
| 358 |
mask.resize(nfront + 1); |
| 359 |
p_front1.resize(nfront + 1); |
| 360 |
for (int k = 0; k < nfront; k++) { |
| 361 |
const int itmp = p_front[k + 1] - p_front[k]; |
| 362 |
if (itmp >= size_thrs) { |
| 363 |
mask[k] = true; |
| 364 |
} |
| 365 |
else { |
| 366 |
mask[k] = false; |
| 367 |
} |
| 368 |
} |
| 369 |
mask[nfront] = true; |
| 370 |
int n = 0; |
| 371 |
{ |
| 372 |
int k = 0; |
| 373 |
p_front1[0] = p_front[0]; |
| 374 |
int k0 = k; |
| 375 |
while (k < nfront) { |
| 376 |
int flag = 0; |
| 377 |
if (!mask[k0]) { |
| 378 |
if (!mask[k + 1]) { |
| 379 |
if ((p_front[k + 1] - p_front[k0]) < size_thrs) { |
| 380 |
flag = 2; |
| 381 |
} |
| 382 |
else { |
| 383 |
flag = (-2); |
| 384 |
} |
| 385 |
} |
| 386 |
else { |
| 387 |
if ((k == 0) && mask[1]) { |
| 388 |
flag = 2; |
| 389 |
} |
| 390 |
else { |
| 391 |
if ((k0 == k) || (k == (nfront - 1))) { |
| 392 |
flag = 0; |
| 393 |
} |
| 394 |
else { |
| 395 |
flag = (-1); |
| 396 |
} |
| 397 |
} |
| 398 |
} // if (!mask[k + 1]) |
| 399 |
} // if (!mask[k0]) |
| 400 |
else { |
| 401 |
if (k == (nfront - 2) && !mask[nfront - 1]) { |
| 402 |
flag = 1; |
| 403 |
} |
| 404 |
} |
| 405 |
switch (flag) { |
| 406 |
case 1: |
| 407 |
p_front1[n + 1] = p_front[k + 2]; |
| 408 |
n++; |
| 409 |
k += 2; |
| 410 |
k0 = k; |
| 411 |
break; |
| 412 |
case (-2): |
| 413 |
p_front1[n + 1] = p_front[k]; |
| 414 |
n++; |
| 415 |
k0 = k; |
| 416 |
break; |
| 417 |
case (-1): |
| 418 |
p_front1[n + 1] = p_front[k]; |
| 419 |
p_front1[n + 2] = p_front[k + 1]; |
| 420 |
n += 2; |
| 421 |
k++; |
| 422 |
k0 = k; |
| 423 |
break; |
| 424 |
case (0): |
| 425 |
p_front1[n + 1] = p_front[k + 1]; |
| 426 |
n++; |
| 427 |
k++; |
| 428 |
k0 = k; |
| 429 |
break; |
| 430 |
case 2: |
| 431 |
k++; |
| 432 |
break; |
| 433 |
} |
| 434 |
} // while (k < nfront) |
| 435 |
} // scope for k, n etc. |
| 436 |
if ((p_front1[0] != 0)|| (p_front1[n] != p_front[nfront])) { |
| 437 |
diss_printf(verbose, fp, "%s %d : error %d -> %d\n", __FILE__, __LINE__, |
| 438 |
p_front[nfront], p_front1[n]); |
| 439 |
exit(-1); |
| 440 |
} |
| 441 |
nfront = n; |
| 442 |
p_front.resize(nfront + 1); |
| 443 |
for (int k = 0; k <= n; k++) { |
| 444 |
p_front[k] = p_front1[k]; |
| 445 |
} |
| 446 |
return nfront; |
| 447 |
} |
| 448 |
|
| 449 |
int getColorMaskCSR(int *color_mask, const CSR_indirect *csr, |
| 450 |
const bool verbose, FILE *fp) |
| 451 |
{ |
| 452 |
const int dim = csr->n; |
| 453 |
int *prow = csr->ptRows; |
| 454 |
int *indcols = csr->indCols; |
| 455 |
|
| 456 |
for (int i = 0; i < dim; i++) { |
| 457 |
color_mask[i] = (-1); |
| 458 |
} |
| 459 |
|
| 460 |
list<int> graph_start; |
| 461 |
int num_isolated = 0; |
| 462 |
for (int i = 0 ; i < dim; i++) { |
| 463 |
if ((prow[i] + 1) == prow[i + 1]) { |
| 464 |
num_isolated++; |
| 465 |
color_mask[i] = 0; |
| 466 |
} |
| 467 |
} |
| 468 |
if (num_isolated > 0) { |
| 469 |
diss_printf(verbose, fp, |
| 470 |
"%s %d : isolated = %d\n", __FILE__, __LINE__, num_isolated); |
| 471 |
} |
| 472 |
int color = 0; |
| 473 |
for (int i = 0; i < dim; i++) { |
| 474 |
if (color_mask[i] == (-1)) { |
| 475 |
color++; |
| 476 |
color_mask[i] = color; |
| 477 |
graph_start.clear(); |
| 478 |
graph_start.push_back(i); |
| 479 |
for (list<int>::iterator it = graph_start.begin(); |
| 480 |
it != graph_start.end(); ++it) { |
| 481 |
for (int k = prow[(*it)]; k < prow[(*it) + 1]; k++) { |
| 482 |
if (color_mask[indcols[k]] == (-1)) { |
| 483 |
color_mask[indcols[k]] = color; |
| 484 |
graph_start.push_back(indcols[k]); |
| 485 |
} |
| 486 |
} |
| 487 |
} // loop : it |
| 488 |
} |
| 489 |
} // loop : i |
| 490 |
// post process merge color with small size into previous color |
| 491 |
int reduced = 0; |
| 492 |
diss_printf(verbose, fp, "%s %d : before color = %d\n", |
| 493 |
__FILE__, __LINE__, color); |
| 494 |
int m = 1; |
| 495 |
while (m <= color) { |
| 496 |
int count = 0; |
| 497 |
for (int i = 0; i < dim; i++) { |
| 498 |
if (color_mask[i] == m) { |
| 499 |
count++; |
| 500 |
} |
| 501 |
} |
| 502 |
if ((count <= DIM_AUG_KERN) && (color > 1)){ // to treat very small matrix |
| 503 |
reduced++; |
| 504 |
color--; |
| 505 |
for (int i = 0; i < dim; i++) { |
| 506 |
if (color_mask[i] == m) { |
| 507 |
color_mask[i] = (-1); |
| 508 |
} |
| 509 |
else if (color_mask[i] > m) { |
| 510 |
color_mask[i]--; |
| 511 |
} |
| 512 |
} |
| 513 |
} |
| 514 |
else { |
| 515 |
m++; |
| 516 |
} |
| 517 |
} |
| 518 |
diss_printf(verbose, fp, "%s %d : after fused = %d color = %d\n", |
| 519 |
__FILE__, __LINE__, reduced, color); |
| 520 |
return color; |
| 521 |
} |
| 522 |
|