Actual source code: lusol.c

  1: #define PETSCMAT_DLL

  3: /* 
  4:         Provides an interface to the LUSOL package of ....

  6: */
 7:  #include src/mat/impls/aij/seq/aij.h

  9: #if defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
 10: #define LU1FAC   lu1fac_
 11: #define LU6SOL   lu6sol_
 12: #define M1PAGE   m1page_
 13: #define M5SETX   m5setx_
 14: #define M6RDEL   m6rdel_
 15: #elif !defined(PETSC_HAVE_FORTRAN_CAPS)
 16: #define LU1FAC   lu1fac
 17: #define LU6SOL   lu6sol
 18: #define M1PAGE   m1page
 19: #define M5SETX   m5setx
 20: #define M6RDEL   m6rdel
 21: #endif

 24: /*
 25:     Dummy symbols that the MINOS files mi25bfac.f and mi15blas.f may require
 26: */
 27: void PETSC_STDCALL M1PAGE() {
 28:   ;
 29: }
 30: void PETSC_STDCALL M5SETX() {
 31:   ;
 32: }

 34: void PETSC_STDCALL M6RDEL() {
 35:   ;
 36: }

 39:                         double *parmlu, double *data, int *indc, int *indr,
 40:                         int *rowperm, int *colperm, int *collen, int *rowlen,
 41:                         int *colstart, int *rowstart, int *rploc, int *cploc,
 42:                         int *rpinv, int *cpinv, double *w, int *inform);

 45:                         int *size, int *luparm, double *parmlu, double *data,
 46:                         int *indc, int *indr, int *rowperm, int *colperm,
 47:                         int *collen, int *rowlen, int *colstart, int *rowstart,
 48:                         int *inform);

 51: EXTERN PetscErrorCode MatDuplicate_LUSOL(Mat,MatDuplicateOption,Mat*);

 53: typedef struct  {
 54:   double *data;
 55:   int *indc;
 56:   int *indr;

 58:   int *ip;
 59:   int *iq;
 60:   int *lenc;
 61:   int *lenr;
 62:   int *locc;
 63:   int *locr;
 64:   int *iploc;
 65:   int *iqloc;
 66:   int *ipinv;
 67:   int *iqinv;
 68:   double *mnsw;
 69:   double *mnsv;

 71:   double elbowroom;
 72:   double luroom;                /* Extra space allocated when factor fails   */
 73:   double parmlu[30];                /* Input/output to LUSOL                     */

 75:   int n;                        /* Number of rows/columns in matrix          */
 76:   int nz;                        /* Number of nonzeros                        */
 77:   int nnz;                        /* Number of nonzeros allocated for factors  */
 78:   int luparm[30];                /* Input/output to LUSOL                     */

 80:   PetscErrorCode (*MatDuplicate)(Mat,MatDuplicateOption,Mat*);
 81:   PetscErrorCode (*MatLUFactorSymbolic)(Mat,IS,IS,MatFactorInfo*,Mat*);
 82:   PetscErrorCode (*MatDestroy)(Mat);
 83:   PetscTruth CleanUpLUSOL;

 85: } Mat_LUSOL;

 87: /*  LUSOL input/Output Parameters (Description uses C-style indexes
 88:  *
 89:  *  Input parameters                                        Typical value
 90:  *
 91:  *  luparm(0) = nout     File number for printed messages.         6
 92:  *  luparm(1) = lprint   Print level.                              0
 93:  *                    < 0 suppresses output.
 94:  *                    = 0 gives error messages.
 95:  *                    = 1 gives debug output from some of the
 96:  *                        other routines in LUSOL.
 97:  *                   >= 2 gives the pivot row and column and the
 98:  *                        no. of rows and columns involved at
 99:  *                        each elimination step in lu1fac.
100:  *  luparm(2) = maxcol   lu1fac: maximum number of columns         5
101:  *                        searched allowed in a Markowitz-type
102:  *                        search for the next pivot element.
103:  *                        For some of the factorization, the
104:  *                        number of rows searched is
105:  *                        maxrow = maxcol - 1.
106:  *
107:  *
108:  *  Output parameters
109:  *
110:  *  luparm(9) = inform   Return code from last call to any LU routine.
111:  *  luparm(10) = nsing    No. of singularities marked in the
112:  *                        output array w(*).
113:  *  luparm(11) = jsing    Column index of last singularity.
114:  *  luparm(12) = minlen   Minimum recommended value for  lena.
115:  *  luparm(13) = maxlen   ?
116:  *  luparm(14) = nupdat   No. of updates performed by the lu8 routines.
117:  *  luparm(15) = nrank    No. of nonempty rows of U.
118:  *  luparm(16) = ndens1   No. of columns remaining when the density of
119:  *                        the matrix being factorized reached dens1.
120:  *  luparm(17) = ndens2   No. of columns remaining when the density of
121:  *                        the matrix being factorized reached dens2.
122:  *  luparm(18) = jumin    The column index associated with dumin.
123:  *  luparm(19) = numl0    No. of columns in initial  L.
124:  *  luparm(20) = lenl0    Size of initial  L  (no. of nonzeros).
125:  *  luparm(21) = lenu0    Size of initial  U.
126:  *  luparm(22) = lenl     Size of current  L.
127:  *  luparm(23) = lenu     Size of current  U.
128:  *  luparm(24) = lrow     Length of row file.
129:  *  luparm(25) = ncp      No. of compressions of LU data structures.
130:  *  luparm(26) = mersum   lu1fac: sum of Markowitz merit counts.
131:  *  luparm(27) = nutri    lu1fac: triangular rows in U.
132:  *  luparm(28) = nltri    lu1fac: triangular rows in L.
133:  *  luparm(29) =
134:  *
135:  *
136:  *  Input parameters                                        Typical value
137:  *
138:  *  parmlu(0) = elmax1   Max multiplier allowed in  L           10.0
139:  *                        during factor.
140:  *  parmlu(1) = elmax2   Max multiplier allowed in  L           10.0
141:  *                        during updates.
142:  *  parmlu(2) = small    Absolute tolerance for             eps**0.8
143:  *                        treating reals as zero.     IBM double: 3.0d-13
144:  *  parmlu(3) = utol1    Absolute tol for flagging          eps**0.66667
145:  *                        small diagonals of U.       IBM double: 3.7d-11
146:  *  parmlu(4) = utol2    Relative tol for flagging          eps**0.66667
147:  *                        small diagonals of U.       IBM double: 3.7d-11
148:  *  parmlu(5) = uspace   Factor limiting waste space in  U.      3.0
149:  *                        In lu1fac, the row or column lists
150:  *                        are compressed if their length
151:  *                        exceeds uspace times the length of
152:  *                        either file after the last compression.
153:  *  parmlu(6) = dens1    The density at which the Markowitz      0.3
154:  *                        strategy should search maxcol columns
155:  *                        and no rows.
156:  *  parmlu(7) = dens2    the density at which the Markowitz      0.6
157:  *                        strategy should search only 1 column
158:  *                        or (preferably) use a dense LU for
159:  *                        all the remaining rows and columns.
160:  *
161:  *
162:  *  Output parameters
163:  *
164:  *  parmlu(9) = amax     Maximum element in  A.
165:  *  parmlu(10) = elmax    Maximum multiplier in current  L.
166:  *  parmlu(11) = umax     Maximum element in current  U.
167:  *  parmlu(12) = dumax    Maximum diagonal in  U.
168:  *  parmlu(13) = dumin    Minimum diagonal in  U.
169:  *  parmlu(14) =
170:  *  parmlu(15) =
171:  *  parmlu(16) =
172:  *  parmlu(17) =
173:  *  parmlu(18) =
174:  *  parmlu(19) = resid    lu6sol: residual after solve with U or U'.
175:  *  ...
176:  *  parmlu(29) =
177:  */

179: #define Factorization_Tolerance       1e-1
180: #define Factorization_Pivot_Tolerance pow(2.2204460492503131E-16, 2.0 / 3.0) 
181: #define Factorization_Small_Tolerance 1e-15 /* pow(DBL_EPSILON, 0.8) */

186: PetscErrorCode PETSCMAT_DLLEXPORT MatConvert_LUSOL_SeqAIJ(Mat A,const MatType type,MatReuse reuse,Mat *newmat)
187: {
188:   /* This routine is only called to convert an unfactored PETSc-LUSOL matrix */
189:   /* to its base PETSc type, so we will ignore 'MatType type'. */
191:   Mat            B=*newmat;
192:   Mat_LUSOL      *lusol=(Mat_LUSOL *)A->spptr;

195:   if (reuse == MAT_INITIAL_MATRIX) {
196:     MatDuplicate(A,MAT_COPY_VALUES,&B);
197:   }
198:   B->ops->duplicate        = lusol->MatDuplicate;
199:   B->ops->lufactorsymbolic = lusol->MatLUFactorSymbolic;
200:   B->ops->destroy          = lusol->MatDestroy;
201: 
202:   PetscFree(lusol);

204:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_lusol_C","",PETSC_NULL);
205:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_lusol_seqaij_C","",PETSC_NULL);

207:   PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
208:   *newmat = B;
209:   return(0);
210: }

215: PetscErrorCode MatDestroy_LUSOL(Mat A)
216: {
218:   Mat_LUSOL *lusol=(Mat_LUSOL *)A->spptr;

221:   if (lusol->CleanUpLUSOL) {
222:     PetscFree(lusol->ip);
223:     PetscFree(lusol->iq);
224:     PetscFree(lusol->lenc);
225:     PetscFree(lusol->lenr);
226:     PetscFree(lusol->locc);
227:     PetscFree(lusol->locr);
228:     PetscFree(lusol->iploc);
229:     PetscFree(lusol->iqloc);
230:     PetscFree(lusol->ipinv);
231:     PetscFree(lusol->iqinv);
232:     PetscFree(lusol->mnsw);
233:     PetscFree(lusol->mnsv);
234:     PetscFree(lusol->indc);
235:   }

237:   MatConvert_LUSOL_SeqAIJ(A,MATSEQAIJ,MAT_REUSE_MATRIX,&A);
238:   (*A->ops->destroy)(A);
239:   return(0);
240: }

244: PetscErrorCode MatSolve_LUSOL(Mat A,Vec b,Vec x)
245: {
246:   Mat_LUSOL *lusol=(Mat_LUSOL*)A->spptr;
247:   double    *bb,*xx;
248:   int       mode=5;
250:   int       i,m,n,nnz,status;

253:   VecGetArray(x, &xx);
254:   VecGetArray(b, &bb);

256:   m = n = lusol->n;
257:   nnz = lusol->nnz;

259:   for (i = 0; i < m; i++)
260:     {
261:       lusol->mnsv[i] = bb[i];
262:     }

264:   LU6SOL(&mode, &m, &n, lusol->mnsv, xx, &nnz,
265:          lusol->luparm, lusol->parmlu, lusol->data,
266:          lusol->indc, lusol->indr, lusol->ip, lusol->iq,
267:          lusol->lenc, lusol->lenr, lusol->locc, lusol->locr, &status);

269:   if (status != 0)
270:     {
271:       SETERRQ(PETSC_ERR_ARG_SIZ,"solve failed");
272:     }

274:   VecRestoreArray(x, &xx);
275:   VecRestoreArray(b, &bb);
276:   return(0);
277: }

281: PetscErrorCode MatLUFactorNumeric_LUSOL(Mat A,MatFactorInfo *info,Mat *F)
282: {
283:   Mat_SeqAIJ     *a;
284:   Mat_LUSOL      *lusol = (Mat_LUSOL*)(*F)->spptr;
286:   int            m, n, nz, nnz, status;
287:   int            i, rs, re;
288:   int            factorizations;

291:   MatGetSize(A,&m,&n);
292:   a = (Mat_SeqAIJ *)A->data;

294:   if (m != lusol->n) {
295:     SETERRQ(PETSC_ERR_ARG_SIZ,"factorization struct inconsistent");
296:   }

298:   factorizations = 0;
299:   do
300:     {
301:       /*******************************************************************/
302:       /* Check the workspace allocation.                                 */
303:       /*******************************************************************/

305:       nz = a->nz;
306:       nnz = PetscMax(lusol->nnz, (int)(lusol->elbowroom*nz));
307:       nnz = PetscMax(nnz, 5*n);

309:       if (nnz < lusol->luparm[12]){
310:         nnz = (int)(lusol->luroom * lusol->luparm[12]);
311:       } else if ((factorizations > 0) && (lusol->luroom < 6)){
312:         lusol->luroom += 0.1;
313:       }

315:       nnz = PetscMax(nnz, (int)(lusol->luroom*(lusol->luparm[22] + lusol->luparm[23])));

317:       if (nnz > lusol->nnz){
318:         PetscFree(lusol->indc);
319:         PetscMalloc((sizeof(double)+2*sizeof(int))*nnz,&lusol->indc);
320:         lusol->indr = lusol->indc + nnz;
321:         lusol->data = (double *)(lusol->indr + nnz);
322:         lusol->nnz  = nnz;
323:       }

325:       /*******************************************************************/
326:       /* Fill in the data for the problem.      (1-based Fortran style)  */
327:       /*******************************************************************/

329:       nz = 0;
330:       for (i = 0; i < n; i++)
331:         {
332:           rs = a->i[i];
333:           re = a->i[i+1];

335:           while (rs < re)
336:             {
337:               if (a->a[rs] != 0.0)
338:                 {
339:                   lusol->indc[nz] = i + 1;
340:                   lusol->indr[nz] = a->j[rs] + 1;
341:                   lusol->data[nz] = a->a[rs];
342:                   nz++;
343:                 }
344:               rs++;
345:             }
346:         }

348:       /*******************************************************************/
349:       /* Do the factorization.                                           */
350:       /*******************************************************************/

352:       LU1FAC(&m, &n, &nz, &nnz,
353:              lusol->luparm, lusol->parmlu, lusol->data,
354:              lusol->indc, lusol->indr, lusol->ip, lusol->iq,
355:              lusol->lenc, lusol->lenr, lusol->locc, lusol->locr,
356:              lusol->iploc, lusol->iqloc, lusol->ipinv,
357:              lusol->iqinv, lusol->mnsw, &status);
358: 
359:       switch(status)
360:         {
361:         case 0:                /* factored */
362:           break;

364:         case 7:                /* insufficient memory */
365:           break;

367:         case 1:
368:         case -1:                /* singular */
369:           SETERRQ(PETSC_ERR_LIB,"Singular matrix");

371:         case 3:
372:         case 4:                /* error conditions */
373:           SETERRQ(PETSC_ERR_LIB,"matrix error");

375:         default:                /* unknown condition */
376:           SETERRQ(PETSC_ERR_LIB,"matrix unknown return code");
377:         }

379:       factorizations++;
380:     } while (status == 7);
381:   (*F)->assembled = PETSC_TRUE;
382:   return(0);
383: }

387: PetscErrorCode MatLUFactorSymbolic_LUSOL(Mat A, IS r, IS c,MatFactorInfo *info, Mat *F) {
388:   /************************************************************************/
389:   /* Input                                                                */
390:   /*     A  - matrix to factor                                            */
391:   /*     r  - row permutation (ignored)                                   */
392:   /*     c  - column permutation (ignored)                                */
393:   /*                                                                      */
394:   /* Output                                                               */
395:   /*     F  - matrix storing the factorization;                           */
396:   /************************************************************************/
397:   Mat       B;
398:   Mat_LUSOL *lusol;
400:   int        i, m, n, nz, nnz;

403: 
404:   /************************************************************************/
405:   /* Check the arguments.                                                 */
406:   /************************************************************************/

408:   MatGetSize(A, &m, &n);
409:   nz = ((Mat_SeqAIJ *)A->data)->nz;

411:   /************************************************************************/
412:   /* Create the factorization.                                            */
413:   /************************************************************************/

415:   MatCreate(A->comm,&B);
416:   MatSetSizes(B,PETSC_DECIDE,PETSC_DECIDE,m,n);
417:   MatSetType(B,A->type_name);
418:   MatSeqAIJSetPreallocation(B,0,PETSC_NULL);

420:   B->ops->lufactornumeric = MatLUFactorNumeric_LUSOL;
421:   B->ops->solve           = MatSolve_LUSOL;
422:   B->factor               = FACTOR_LU;
423:   lusol                   = (Mat_LUSOL*)(B->spptr);

425:   /************************************************************************/
426:   /* Initialize parameters                                                */
427:   /************************************************************************/

429:   for (i = 0; i < 30; i++)
430:     {
431:       lusol->luparm[i] = 0;
432:       lusol->parmlu[i] = 0;
433:     }

435:   lusol->luparm[1] = -1;
436:   lusol->luparm[2] = 5;
437:   lusol->luparm[7] = 1;

439:   lusol->parmlu[0] = 1 / Factorization_Tolerance;
440:   lusol->parmlu[1] = 1 / Factorization_Tolerance;
441:   lusol->parmlu[2] = Factorization_Small_Tolerance;
442:   lusol->parmlu[3] = Factorization_Pivot_Tolerance;
443:   lusol->parmlu[4] = Factorization_Pivot_Tolerance;
444:   lusol->parmlu[5] = 3.0;
445:   lusol->parmlu[6] = 0.3;
446:   lusol->parmlu[7] = 0.6;

448:   /************************************************************************/
449:   /* Allocate the workspace needed by LUSOL.                              */
450:   /************************************************************************/

452:   lusol->elbowroom = PetscMax(lusol->elbowroom, info->fill);
453:   nnz = PetscMax((int)(lusol->elbowroom*nz), 5*n);
454: 
455:   lusol->n = n;
456:   lusol->nz = nz;
457:   lusol->nnz = nnz;
458:   lusol->luroom = 1.75;

460:   PetscMalloc(sizeof(int)*n,&lusol->ip);
461:   PetscMalloc(sizeof(int)*n,&lusol->iq);
462:   PetscMalloc(sizeof(int)*n,&lusol->lenc);
463:   PetscMalloc(sizeof(int)*n,&lusol->lenr);
464:   PetscMalloc(sizeof(int)*n,&lusol->locc);
465:   PetscMalloc(sizeof(int)*n,&lusol->locr);
466:   PetscMalloc(sizeof(int)*n,&lusol->iploc);
467:   PetscMalloc(sizeof(int)*n,&lusol->iqloc);
468:   PetscMalloc(sizeof(int)*n,&lusol->ipinv);
469:   PetscMalloc(sizeof(int)*n,&lusol->iqinv);
470:   PetscMalloc(sizeof(double)*n,&lusol->mnsw);
471:   PetscMalloc(sizeof(double)*n,&lusol->mnsv);

473:   PetscMalloc((sizeof(double)+2*sizeof(int))*nnz,&lusol->indc);
474:   lusol->indr = lusol->indc + nnz;
475:   lusol->data = (double *)(lusol->indr + nnz);
476:   lusol->CleanUpLUSOL = PETSC_TRUE;
477:   *F = B;
478:   return(0);
479: }

484: PetscErrorCode PETSCMAT_DLLEXPORT MatConvert_SeqAIJ_LUSOL(Mat A,const MatType type,MatReuse reuse,Mat *newmat)
485: {
487:   PetscInt       m, n;
488:   Mat_LUSOL      *lusol;
489:   Mat            B=*newmat;

492:   MatGetSize(A, &m, &n);
493:   if (m != n) {
494:     SETERRQ(PETSC_ERR_ARG_SIZ,"matrix must be square");
495:   }
496:   if (reuse == MAT_INITIAL_MATRIX) {
497:     MatDuplicate(A,MAT_COPY_VALUES,&B);
498:   }
499: 
500:   PetscNew(Mat_LUSOL,&lusol);
501:   lusol->MatDuplicate        = A->ops->duplicate;
502:   lusol->MatLUFactorSymbolic = A->ops->lufactorsymbolic;
503:   lusol->MatDestroy          = A->ops->destroy;
504:   lusol->CleanUpLUSOL        = PETSC_FALSE;

506:   B->spptr                   = (void*)lusol;
507:   B->ops->duplicate          = MatDuplicate_LUSOL;
508:   B->ops->lufactorsymbolic   = MatLUFactorSymbolic_LUSOL;
509:   B->ops->destroy            = MatDestroy_LUSOL;

511:   PetscInfo(0,"Using LUSOL for LU factorization and solves.\n");
512:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_seqaij_lusol_C",
513:                                            "MatConvert_SeqAIJ_LUSOL",MatConvert_SeqAIJ_LUSOL);
514:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_lusol_seqaij_C",
515:                                            "MatConvert_LUSOL_SeqAIJ",MatConvert_LUSOL_SeqAIJ);
516:   PetscObjectChangeTypeName((PetscObject)B,type);
517:   *newmat = B;
518:   return(0);
519: }

524: PetscErrorCode MatDuplicate_LUSOL(Mat A, MatDuplicateOption op, Mat *M) {
526:   Mat_LUSOL *lu=(Mat_LUSOL *)A->spptr;
528:   (*lu->MatDuplicate)(A,op,M);
529:   PetscMemcpy((*M)->spptr,lu,sizeof(Mat_LUSOL));
530:   return(0);
531: }

533: /*MC
534:   MATLUSOL - MATLUSOL = "lusol" - A matrix type providing direct solvers (LU) for sequential matrices 
535:   via the external package LUSOL.

537:   If LUSOL is installed (see the manual for
538:   instructions on how to declare the existence of external packages),
539:   a matrix type can be constructed which invokes LUSOL solvers.
540:   After calling MatCreate(...,A), simply call MatSetType(A,MATLUSOL).
541:   This matrix type is only supported for double precision real.

543:   This matrix inherits from MATSEQAIJ.  As a result, MatSeqAIJSetPreallocation is 
544:   supported for this matrix type.  MatConvert can be called for a fast inplace conversion
545:   to and from the MATSEQAIJ matrix type.

547:   Options Database Keys:
548: . -mat_type lusol - sets the matrix type to "lusol" during a call to MatSetFromOptions()

550:    Level: beginner

552: .seealso: PCLU
553: M*/

558: PetscErrorCode PETSCMAT_DLLEXPORT MatCreate_LUSOL(Mat A)
559: {

563:   /* Change type name before calling MatSetType to force proper construction of SeqAIJ and LUSOL types */
564:   PetscObjectChangeTypeName((PetscObject)A,MATLUSOL);
565:   MatSetType(A,MATSEQAIJ);
566:   MatConvert_SeqAIJ_LUSOL(A,MATLUSOL,MAT_REUSE_MATRIX,&A);
567:   return(0);
568: }