Actual source code: superlu_dist.c

petsc-master 2015-03-28
Report Typos and Errors
  2: /*
  3:         Provides an interface to the SuperLU_DIST_2.2 sparse solver
  4: */

  6: #include <../src/mat/impls/aij/seq/aij.h>
  7: #include <../src/mat/impls/aij/mpi/mpiaij.h>
  8: #if defined(PETSC_HAVE_STDLIB_H) /* This is to get around weird problem with SuperLU on cray */
  9: #include <stdlib.h>
 10: #endif

 12: #if defined(PETSC_USE_64BIT_INDICES)
 13: /* ugly SuperLU_Dist variable telling it to use long long int */
 14: #define _LONGINT
 15: #endif

 17: EXTERN_C_BEGIN
 18: #if defined(PETSC_USE_COMPLEX)
 19: #include <superlu_zdefs.h>
 20: #else
 21: #include <superlu_ddefs.h>
 22: #endif
 23: EXTERN_C_END

 25: /*
 26:     GLOBAL - The sparse matrix and right hand side are all stored initially on process 0. Should be called centralized
 27:     DISTRIBUTED - The sparse matrix and right hand size are initially stored across the entire MPI communicator.
 28: */
 29: typedef enum {GLOBAL,DISTRIBUTED} SuperLU_MatInputMode;
 30: const char *SuperLU_MatInputModes[] = {"GLOBAL","DISTRIBUTED","SuperLU_MatInputMode","PETSC_",0};

 32: typedef struct {
 33:   int_t                nprow,npcol,*row,*col;
 34:   gridinfo_t           grid;
 35:   superlu_options_t    options;
 36:   SuperMatrix          A_sup;
 37:   ScalePermstruct_t    ScalePermstruct;
 38:   LUstruct_t           LUstruct;
 39:   int                  StatPrint;
 40:   SuperLU_MatInputMode MatInputMode;
 41:   SOLVEstruct_t        SOLVEstruct;
 42:   fact_t               FactPattern;
 43:   MPI_Comm             comm_superlu;
 44: #if defined(PETSC_USE_COMPLEX)
 45:   doublecomplex        *val;
 46: #else
 47:   double               *val;
 48: #endif
 49:   PetscBool            matsolve_iscalled,matmatsolve_iscalled;
 50:   PetscBool            CleanUpSuperLU_Dist;  /* Flag to clean up (non-global) SuperLU objects during Destroy */
 51: } Mat_SuperLU_DIST;

 53: extern PetscErrorCode MatFactorInfo_SuperLU_DIST(Mat,PetscViewer);
 54: extern PetscErrorCode MatLUFactorNumeric_SuperLU_DIST(Mat,Mat,const MatFactorInfo*);
 55: extern PetscErrorCode MatDestroy_SuperLU_DIST(Mat);
 56: extern PetscErrorCode MatView_SuperLU_DIST(Mat,PetscViewer);
 57: extern PetscErrorCode MatSolve_SuperLU_DIST(Mat,Vec,Vec);
 58: extern PetscErrorCode MatLUFactorSymbolic_SuperLU_DIST(Mat,Mat,IS,IS,const MatFactorInfo*);
 59: extern PetscErrorCode MatDestroy_MPIAIJ(Mat);

 63: PetscErrorCode MatGetDiagonal_SuperLU_DIST(Mat A,Vec v)
 64: {
 66:   SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Mat type: SuperLU_DIST factor");
 67:   return(0);
 68: }

 72: PetscErrorCode MatDestroy_SuperLU_DIST(Mat A)
 73: {
 74:   PetscErrorCode   ierr;
 75:   Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)A->spptr;
 76:   PetscBool        flg;

 79:   if (lu && lu->CleanUpSuperLU_Dist) {
 80:     /* Deallocate SuperLU_DIST storage */
 81:     if (lu->MatInputMode == GLOBAL) {
 82:       PetscStackCall("SuperLU_DIST:Destroy_CompCol_Matrix_dist",Destroy_CompCol_Matrix_dist(&lu->A_sup));
 83:     } else {
 84:       PetscStackCall("SuperLU_DIST:Destroy_CompRowLoc_Matrix_dist",Destroy_CompRowLoc_Matrix_dist(&lu->A_sup));
 85:       if (lu->options.SolveInitialized) {
 86: #if defined(PETSC_USE_COMPLEX)
 87:         PetscStackCall("SuperLU_DIST:zSolveFinalize",zSolveFinalize(&lu->options, &lu->SOLVEstruct));
 88: #else
 89:         PetscStackCall("SuperLU_DIST:dSolveFinalize",dSolveFinalize(&lu->options, &lu->SOLVEstruct));
 90: #endif
 91:       }
 92:     }
 93:     PetscStackCall("SuperLU_DIST:Destroy_LU",Destroy_LU(A->cmap->N, &lu->grid, &lu->LUstruct));
 94:     PetscStackCall("SuperLU_DIST:ScalePermstructFree",ScalePermstructFree(&lu->ScalePermstruct));
 95:     PetscStackCall("SuperLU_DIST:LUstructFree",LUstructFree(&lu->LUstruct));

 97:     /* Release the SuperLU_DIST process grid. */
 98:     PetscStackCall("SuperLU_DIST:superlu_gridexit",superlu_gridexit(&lu->grid));
 99:     MPI_Comm_free(&(lu->comm_superlu));
100:   }
101:   PetscFree(A->spptr);

103:   PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&flg);
104:   if (flg) {
105:     MatDestroy_SeqAIJ(A);
106:   } else {
107:     MatDestroy_MPIAIJ(A);
108:   }
109:   return(0);
110: }

114: PetscErrorCode MatSolve_SuperLU_DIST(Mat A,Vec b_mpi,Vec x)
115: {
116:   Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)A->spptr;
117:   PetscErrorCode   ierr;
118:   PetscMPIInt      size;
119:   PetscInt         m=A->rmap->n,M=A->rmap->N,N=A->cmap->N;
120:   SuperLUStat_t    stat;
121:   double           berr[1];
122:   PetscScalar      *bptr;
123:   PetscInt         nrhs=1;
124:   Vec              x_seq;
125:   IS               iden;
126:   VecScatter       scat;
127:   int              info; /* SuperLU_Dist info code is ALWAYS an int, even with long long indices */
128:   static PetscBool cite = PETSC_FALSE;

131:   PetscCitationsRegister("@article{lidemmel03,\n  author = {Xiaoye S. Li and James W. Demmel},\n  title = {{SuperLU_DIST}: A Scalable Distributed-Memory Sparse Direct\n           Solver for Unsymmetric Linear Systems},\n  journal = {ACM Trans. Mathematical Software},\n  volume = {29},\n  number = {2},\n  pages = {110-140},\n  year = 2003\n}\n",&cite);
132:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
133:   if (size > 1 && lu->MatInputMode == GLOBAL) {
134:     /* global mat input, convert b to x_seq */
135:     VecCreateSeq(PETSC_COMM_SELF,N,&x_seq);
136:     ISCreateStride(PETSC_COMM_SELF,N,0,1,&iden);
137:     VecScatterCreate(b_mpi,iden,x_seq,iden,&scat);
138:     ISDestroy(&iden);

140:     VecScatterBegin(scat,b_mpi,x_seq,INSERT_VALUES,SCATTER_FORWARD);
141:     VecScatterEnd(scat,b_mpi,x_seq,INSERT_VALUES,SCATTER_FORWARD);
142:     VecGetArray(x_seq,&bptr);
143:   } else { /* size==1 || distributed mat input */
144:     if (lu->options.SolveInitialized && !lu->matsolve_iscalled) {
145:       /* see comments in MatMatSolve() */
146: #if defined(PETSC_USE_COMPLEX)
147:       PetscStackCall("SuperLU_DIST:zSolveFinalize",zSolveFinalize(&lu->options, &lu->SOLVEstruct));
148: #else
149:       PetscStackCall("SuperLU_DIST:dSolveFinalize",dSolveFinalize(&lu->options, &lu->SOLVEstruct));
150: #endif
151:       lu->options.SolveInitialized = NO;
152:     }
153:     VecCopy(b_mpi,x);
154:     VecGetArray(x,&bptr);
155:   }

157:   if (lu->options.Fact != FACTORED) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"SuperLU_DIST options.Fact mush equal FACTORED");

159:   PetscStackCall("SuperLU_DIST:PStatInit",PStatInit(&stat));        /* Initialize the statistics variables. */
160:   if (lu->MatInputMode == GLOBAL) {
161: #if defined(PETSC_USE_COMPLEX)
162:     PetscStackCall("SuperLU_DIST:pzgssvx_ABglobal",pzgssvx_ABglobal(&lu->options,&lu->A_sup,&lu->ScalePermstruct,(doublecomplex*)bptr,M,nrhs,&lu->grid,&lu->LUstruct,berr,&stat,&info));
163: #else
164:     PetscStackCall("SuperLU_DIST:pdgssvx_ABglobal",pdgssvx_ABglobal(&lu->options, &lu->A_sup, &lu->ScalePermstruct,bptr,M,nrhs,&lu->grid,&lu->LUstruct,berr,&stat,&info));
165: #endif
166:   } else { /* distributed mat input */
167: #if defined(PETSC_USE_COMPLEX)
168:     PetscStackCall("SuperLU_DIST:pzgssvx",pzgssvx(&lu->options,&lu->A_sup,&lu->ScalePermstruct,(doublecomplex*)bptr,m,nrhs,&lu->grid,&lu->LUstruct,&lu->SOLVEstruct,berr,&stat,&info));
169: #else
170:     PetscStackCall("SuperLU_DIST:pdgssvx",pdgssvx(&lu->options,&lu->A_sup,&lu->ScalePermstruct,bptr,m,nrhs,&lu->grid,&lu->LUstruct,&lu->SOLVEstruct,berr,&stat,&info));
171: #endif
172:   }
173:   if (info) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"pdgssvx fails, info: %d\n",info);

175:   if (lu->options.PrintStat) PStatPrint(&lu->options, &stat, &lu->grid);      /* Print the statistics. */
176:   PetscStackCall("SuperLU_DIST:PStatFree",PStatFree(&stat));

178:   if (size > 1 && lu->MatInputMode == GLOBAL) {
179:     /* convert seq x to mpi x */
180:     VecRestoreArray(x_seq,&bptr);
181:     VecScatterBegin(scat,x_seq,x,INSERT_VALUES,SCATTER_REVERSE);
182:     VecScatterEnd(scat,x_seq,x,INSERT_VALUES,SCATTER_REVERSE);
183:     VecScatterDestroy(&scat);
184:     VecDestroy(&x_seq);
185:   } else {
186:     VecRestoreArray(x,&bptr);

188:     lu->matsolve_iscalled    = PETSC_TRUE;
189:     lu->matmatsolve_iscalled = PETSC_FALSE;
190:   }
191:   return(0);
192: }

196: PetscErrorCode MatMatSolve_SuperLU_DIST(Mat A,Mat B_mpi,Mat X)
197: {
198:   Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)A->spptr;
199:   PetscErrorCode   ierr;
200:   PetscMPIInt      size;
201:   PetscInt         M=A->rmap->N,m=A->rmap->n,nrhs;
202:   SuperLUStat_t    stat;
203:   double           berr[1];
204:   PetscScalar      *bptr;
205:   int              info; /* SuperLU_Dist info code is ALWAYS an int, even with long long indices */
206:   PetscBool        flg;

209:   if (lu->options.Fact != FACTORED) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"SuperLU_DIST options.Fact mush equal FACTORED");
210:   PetscObjectTypeCompareAny((PetscObject)B_mpi,&flg,MATSEQDENSE,MATMPIDENSE,NULL);
211:   if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Matrix B must be MATDENSE matrix");
212:   PetscObjectTypeCompareAny((PetscObject)X,&flg,MATSEQDENSE,MATMPIDENSE,NULL);
213:   if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Matrix X must be MATDENSE matrix");

215:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
216:   if (size > 1 && lu->MatInputMode == GLOBAL) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatInputMode=GLOBAL for nproc %d>1 is not supported",size);
217:   /* size==1 or distributed mat input */
218:   if (lu->options.SolveInitialized && !lu->matmatsolve_iscalled) {
219:     /* communication pattern of SOLVEstruct is unlikely created for matmatsolve,
220:        thus destroy it and create a new SOLVEstruct.
221:        Otherwise it may result in memory corruption or incorrect solution
222:        See src/mat/examples/tests/ex125.c */
223: #if defined(PETSC_USE_COMPLEX)
224:     PetscStackCall("SuperLU_DIST:zSolveFinalize",zSolveFinalize(&lu->options, &lu->SOLVEstruct));
225: #else
226:     PetscStackCall("SuperLU_DIST:dSolveFinalize",dSolveFinalize(&lu->options, &lu->SOLVEstruct));
227: #endif
228:     lu->options.SolveInitialized = NO;
229:   }
230:   MatCopy(B_mpi,X,SAME_NONZERO_PATTERN);

232:   MatGetSize(B_mpi,NULL,&nrhs);

234:   PetscStackCall("SuperLU_DIST:PStatInit",PStatInit(&stat));        /* Initialize the statistics variables. */
235:   MatDenseGetArray(X,&bptr);
236:   if (lu->MatInputMode == GLOBAL) { /* size == 1 */
237: #if defined(PETSC_USE_COMPLEX)
238:     PetscStackCall("SuperLU_DIST:pzgssvx_ABglobal",pzgssvx_ABglobal(&lu->options, &lu->A_sup, &lu->ScalePermstruct,(doublecomplex*)bptr, M, nrhs,&lu->grid, &lu->LUstruct, berr, &stat, &info));
239: #else
240:     PetscStackCall("SuperLU_DIST:pdgssvx_ABglobal",pdgssvx_ABglobal(&lu->options, &lu->A_sup, &lu->ScalePermstruct,bptr, M, nrhs, &lu->grid, &lu->LUstruct, berr, &stat, &info));
241: #endif
242:   } else { /* distributed mat input */
243: #if defined(PETSC_USE_COMPLEX)
244:     PetscStackCall("SuperLU_DIST:pzgssvx",pzgssvx(&lu->options,&lu->A_sup,&lu->ScalePermstruct,(doublecomplex*)bptr,m,nrhs,&lu->grid, &lu->LUstruct,&lu->SOLVEstruct,berr,&stat,&info));
245: #else
246:     PetscStackCall("SuperLU_DIST:pdgssvx",pdgssvx(&lu->options,&lu->A_sup,&lu->ScalePermstruct,bptr,m,nrhs,&lu->grid,&lu->LUstruct,&lu->SOLVEstruct,berr,&stat,&info));
247: #endif
248:   }
249:   if (info) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"pdgssvx fails, info: %d\n",info);
250:   MatDenseRestoreArray(X,&bptr);

252:   if (lu->options.PrintStat) PStatPrint(&lu->options, &stat, &lu->grid); /* Print the statistics. */
253:   PetscStackCall("SuperLU_DIST:PStatFree",PStatFree(&stat));
254:   lu->matsolve_iscalled    = PETSC_FALSE;
255:   lu->matmatsolve_iscalled = PETSC_TRUE;
256:   return(0);
257: }


262: PetscErrorCode MatLUFactorNumeric_SuperLU_DIST(Mat F,Mat A,const MatFactorInfo *info)
263: {
264:   Mat              *tseq,A_seq = NULL;
265:   Mat_SeqAIJ       *aa,*bb;
266:   Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)(F)->spptr;
267:   PetscErrorCode   ierr;
268:   PetscInt         M=A->rmap->N,N=A->cmap->N,i,*ai,*aj,*bi,*bj,nz,rstart,*garray,
269:                    m=A->rmap->n, colA_start,j,jcol,jB,countA,countB,*bjj,*ajj;
270:   int              sinfo;   /* SuperLU_Dist info flag is always an int even with long long indices */
271:   PetscMPIInt      size;
272:   SuperLUStat_t    stat;
273:   double           *berr=0;
274:   IS               isrow;
275:   Mat              F_diag=NULL;
276: #if defined(PETSC_USE_COMPLEX)
277:   doublecomplex    *av, *bv;
278: #else
279:   double           *av, *bv;
280: #endif

283:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);

285:   if (lu->MatInputMode == GLOBAL) { /* global mat input */
286:     if (size > 1) { /* convert mpi A to seq mat A */
287:       ISCreateStride(PETSC_COMM_SELF,M,0,1,&isrow);
288:       MatGetSubMatrices(A,1,&isrow,&isrow,MAT_INITIAL_MATRIX,&tseq);
289:       ISDestroy(&isrow);

291:       A_seq = *tseq;
292:       PetscFree(tseq);
293:       aa    = (Mat_SeqAIJ*)A_seq->data;
294:     } else {
295:       PetscBool flg;
296:       PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);
297:       if (flg) {
298:         Mat_MPIAIJ *At = (Mat_MPIAIJ*)A->data;
299:         A = At->A;
300:       }
301:       aa =  (Mat_SeqAIJ*)A->data;
302:     }

304:     /* Convert Petsc NR matrix to SuperLU_DIST NC.
305:        Note: memories of lu->val, col and row are allocated by CompRow_to_CompCol_dist()! */
306:     if (lu->options.Fact != DOFACT) {/* successive numeric factorization, sparsity pattern is reused. */
307:       PetscStackCall("SuperLU_DIST:Destroy_CompCol_Matrix_dist",Destroy_CompCol_Matrix_dist(&lu->A_sup));
308:       if (lu->FactPattern == SamePattern_SameRowPerm) {
309:         lu->options.Fact = SamePattern_SameRowPerm; /* matrix has similar numerical values */
310:       } else { /* lu->FactPattern == SamePattern */
311:         PetscStackCall("SuperLU_DIST:Destroy_LU",Destroy_LU(N, &lu->grid, &lu->LUstruct));
312:         lu->options.Fact = SamePattern;
313:       }
314:     }
315: #if defined(PETSC_USE_COMPLEX)
316:     PetscStackCall("SuperLU_DIST:zCompRow_to_CompCol_dist",zCompRow_to_CompCol_dist(M,N,aa->nz,(doublecomplex*)aa->a,(int_t*)aa->j,(int_t*)aa->i,&lu->val,&lu->col, &lu->row));
317: #else
318:     PetscStackCall("SuperLU_DIST:dCompRow_to_CompCol_dist",dCompRow_to_CompCol_dist(M,N,aa->nz,aa->a,(int_t*)aa->j,(int_t*)aa->i,&lu->val, &lu->col, &lu->row));
319: #endif

321:     /* Create compressed column matrix A_sup. */
322: #if defined(PETSC_USE_COMPLEX)
323:     PetscStackCall("SuperLU_DIST:zCreate_CompCol_Matrix_dist",zCreate_CompCol_Matrix_dist(&lu->A_sup, M, N, aa->nz, lu->val, lu->col, lu->row, SLU_NC, SLU_Z, SLU_GE));
324: #else
325:     PetscStackCall("SuperLU_DIST:dCreate_CompCol_Matrix_dist",dCreate_CompCol_Matrix_dist(&lu->A_sup, M, N, aa->nz, lu->val, lu->col, lu->row, SLU_NC, SLU_D, SLU_GE));
326: #endif
327:   } else { /* distributed mat input */
328:     Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
329:     aa=(Mat_SeqAIJ*)(mat->A)->data;
330:     bb=(Mat_SeqAIJ*)(mat->B)->data;
331:     ai=aa->i; aj=aa->j;
332:     bi=bb->i; bj=bb->j;
333: #if defined(PETSC_USE_COMPLEX)
334:     av=(doublecomplex*)aa->a;
335:     bv=(doublecomplex*)bb->a;
336: #else
337:     av=aa->a;
338:     bv=bb->a;
339: #endif
340:     rstart = A->rmap->rstart;
341:     nz     = aa->nz + bb->nz;
342:     garray = mat->garray;

344:     if (lu->options.Fact == DOFACT) { /* first numeric factorization */
345: #if defined(PETSC_USE_COMPLEX)
346:       PetscStackCall("SuperLU_DIST:zallocateA_dist",zallocateA_dist(m, nz, &lu->val, &lu->col, &lu->row));
347: #else
348:       PetscStackCall("SuperLU_DIST:dallocateA_dist",dallocateA_dist(m, nz, &lu->val, &lu->col, &lu->row));
349: #endif
350:     } else { /* successive numeric factorization, sparsity pattern and perm_c are reused. */
351:       /* Destroy_CompRowLoc_Matrix_dist(&lu->A_sup); */ /* this leads to crash! However, see SuperLU_DIST_2.5/EXAMPLE/pzdrive2.c */
352:       if (lu->FactPattern == SamePattern_SameRowPerm) {
353:         lu->options.Fact = SamePattern_SameRowPerm; /* matrix has similar numerical values */
354:       } else {
355:         PetscStackCall("SuperLU_DIST:Destroy_LU",Destroy_LU(N, &lu->grid, &lu->LUstruct)); /* Deallocate storage associated with the L and U matrices. */
356:         lu->options.Fact = SamePattern;
357:       }
358:     }
359:     nz = 0;
360:     for (i=0; i<m; i++) {
361:       lu->row[i] = nz;
362:       countA     = ai[i+1] - ai[i];
363:       countB     = bi[i+1] - bi[i];
364:       ajj        = aj + ai[i]; /* ptr to the beginning of this row */
365:       bjj        = bj + bi[i];

367:       /* B part, smaller col index */
368:       colA_start = rstart + ajj[0]; /* the smallest global col index of A */
369:       jB         = 0;
370:       for (j=0; j<countB; j++) {
371:         jcol = garray[bjj[j]];
372:         if (jcol > colA_start) {
373:           jB = j;
374:           break;
375:         }
376:         lu->col[nz]   = jcol;
377:         lu->val[nz++] = *bv++;
378:         if (j==countB-1) jB = countB;
379:       }

381:       /* A part */
382:       for (j=0; j<countA; j++) {
383:         lu->col[nz]   = rstart + ajj[j];
384:         lu->val[nz++] = *av++;
385:       }

387:       /* B part, larger col index */
388:       for (j=jB; j<countB; j++) {
389:         lu->col[nz]   = garray[bjj[j]];
390:         lu->val[nz++] = *bv++;
391:       }
392:     }
393:     lu->row[m] = nz;
394: #if defined(PETSC_USE_COMPLEX)
395:     PetscStackCall("SuperLU_DIST:zCreate_CompRowLoc_Matrix_dist",zCreate_CompRowLoc_Matrix_dist(&lu->A_sup, M, N, nz, m, rstart,lu->val, lu->col, lu->row, SLU_NR_loc, SLU_Z, SLU_GE));
396: #else
397:     PetscStackCall("SuperLU_DIST:dCreate_CompRowLoc_Matrix_dist",dCreate_CompRowLoc_Matrix_dist(&lu->A_sup, M, N, nz, m, rstart,lu->val, lu->col, lu->row, SLU_NR_loc, SLU_D, SLU_GE));
398: #endif
399:   }

401:   /* Factor the matrix. */
402:   PetscStackCall("SuperLU_DIST:PStatInit",PStatInit(&stat));   /* Initialize the statistics variables. */
403:   if (lu->MatInputMode == GLOBAL) { /* global mat input */
404: #if defined(PETSC_USE_COMPLEX)
405:     PetscStackCall("SuperLU_DIST:pzgssvx_ABglobal",pzgssvx_ABglobal(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, M, 0,&lu->grid, &lu->LUstruct, berr, &stat, &sinfo));
406: #else
407:     PetscStackCall("SuperLU_DIST:pdgssvx_ABglobal",pdgssvx_ABglobal(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, M, 0,&lu->grid, &lu->LUstruct, berr, &stat, &sinfo));
408: #endif
409:   } else { /* distributed mat input */
410: #if defined(PETSC_USE_COMPLEX)
411:     PetscStackCall("SuperLU_DIST:pzgssvx",pzgssvx(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, m, 0, &lu->grid,&lu->LUstruct, &lu->SOLVEstruct, berr, &stat, &sinfo));
412:     if (sinfo) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"pzgssvx fails, info: %d\n",sinfo);
413: #else
414:     PetscStackCall("SuperLU_DIST:pdgssvx",pdgssvx(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, m, 0, &lu->grid,&lu->LUstruct, &lu->SOLVEstruct, berr, &stat, &sinfo));
415:     if (sinfo) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"pdgssvx fails, info: %d\n",sinfo);
416: #endif
417:   }

419:   if (lu->MatInputMode == GLOBAL && size > 1) {
420:     MatDestroy(&A_seq);
421:   }

423:   if (lu->options.PrintStat) {
424:     PStatPrint(&lu->options, &stat, &lu->grid);  /* Print the statistics. */
425:   }
426:   PStatFree(&stat);
427:   if (size > 1) {
428:     F_diag            = ((Mat_MPIAIJ*)(F)->data)->A;
429:     F_diag->assembled = PETSC_TRUE;
430:   }
431:   (F)->assembled    = PETSC_TRUE;
432:   (F)->preallocated = PETSC_TRUE;
433:   lu->options.Fact  = FACTORED; /* The factored form of A is supplied. Local option used by this func. only */
434:   return(0);
435: }

437: /* Note the Petsc r and c permutations are ignored */
440: PetscErrorCode MatLUFactorSymbolic_SuperLU_DIST(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
441: {
442:   Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)F->spptr;
443:   PetscInt         M   = A->rmap->N,N=A->cmap->N;

446:   /* Initialize the SuperLU process grid. */
447:   PetscStackCall("SuperLU_DIST:superlu_gridinit",superlu_gridinit(lu->comm_superlu, lu->nprow, lu->npcol, &lu->grid));

449:   /* Initialize ScalePermstruct and LUstruct. */
450:   PetscStackCall("SuperLU_DIST:ScalePermstructInit",ScalePermstructInit(M, N, &lu->ScalePermstruct));
451:   PetscStackCall("SuperLU_DIST:LUstructInit",LUstructInit(N, &lu->LUstruct));
452:   F->ops->lufactornumeric = MatLUFactorNumeric_SuperLU_DIST;
453:   F->ops->solve           = MatSolve_SuperLU_DIST;
454:   F->ops->matsolve        = MatMatSolve_SuperLU_DIST;
455:   lu->CleanUpSuperLU_Dist = PETSC_TRUE;
456:   return(0);
457: }

461: PetscErrorCode MatFactorGetSolverPackage_aij_superlu_dist(Mat A,const MatSolverPackage *type)
462: {
464:   *type = MATSOLVERSUPERLU_DIST;
465:   return(0);
466: }

470: PETSC_EXTERN PetscErrorCode MatGetFactor_aij_superlu_dist(Mat A,MatFactorType ftype,Mat *F)
471: {
472:   Mat               B;
473:   Mat_SuperLU_DIST  *lu;
474:   PetscErrorCode    ierr;
475:   PetscInt          M=A->rmap->N,N=A->cmap->N,indx;
476:   PetscMPIInt       size;
477:   superlu_options_t options;
478:   PetscBool         flg;
479:   const char        *colperm[]     = {"NATURAL","MMD_AT_PLUS_A","MMD_ATA","METIS_AT_PLUS_A","PARMETIS"};
480:   const char        *rowperm[]     = {"LargeDiag","NATURAL"};
481:   const char        *factPattern[] = {"SamePattern","SamePattern_SameRowPerm"};
482:   PetscBool         set;

485:   /* Create the factorization matrix */
486:   MatCreate(PetscObjectComm((PetscObject)A),&B);
487:   MatSetSizes(B,A->rmap->n,A->cmap->n,M,N);
488:   MatSetType(B,((PetscObject)A)->type_name);
489:   MatSeqAIJSetPreallocation(B,0,NULL);
490:   MatMPIAIJSetPreallocation(B,0,NULL,0,NULL);

492:   B->ops->lufactorsymbolic = MatLUFactorSymbolic_SuperLU_DIST;
493:   B->ops->view             = MatView_SuperLU_DIST;
494:   B->ops->destroy          = MatDestroy_SuperLU_DIST;
495:   B->ops->getdiagonal      = MatGetDiagonal_SuperLU_DIST;

497:   PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_aij_superlu_dist);

499:   B->factortype = MAT_FACTOR_LU;

501:   PetscNewLog(B,&lu);
502:   B->spptr = lu;

504:   /* Set the default input options:
505:      options.Fact              = DOFACT;
506:      options.Equil             = YES;
507:      options.ParSymbFact       = NO;
508:      options.ColPerm           = METIS_AT_PLUS_A;
509:      options.RowPerm           = LargeDiag;
510:      options.ReplaceTinyPivot  = YES;
511:      options.IterRefine        = DOUBLE;
512:      options.Trans             = NOTRANS;
513:      options.SolveInitialized  = NO; -hold the communication pattern used MatSolve() and MatMatSolve()
514:      options.RefineInitialized = NO;
515:      options.PrintStat         = YES;
516:   */
517:   set_default_options_dist(&options);

519:   MPI_Comm_dup(PetscObjectComm((PetscObject)A),&(lu->comm_superlu));
520:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
521:   /* Default num of process columns and rows */
522:   lu->npcol = (int_t) (0.5 + PetscSqrtReal((PetscReal)size));
523:   if (!lu->npcol) lu->npcol = 1;
524:   while (lu->npcol > 0) {
525:     lu->nprow = (int_t) (size/lu->npcol);
526:     if (size == lu->nprow * lu->npcol) break;
527:     lu->npcol--;
528:   }

530:   PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"SuperLU_Dist Options","Mat");
531:   PetscOptionsInt("-mat_superlu_dist_r","Number rows in processor partition","None",lu->nprow,(PetscInt*)&lu->nprow,NULL);
532:   PetscOptionsInt("-mat_superlu_dist_c","Number columns in processor partition","None",lu->npcol,(PetscInt*)&lu->npcol,NULL);
533:   if (size != lu->nprow * lu->npcol) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number of processes %d must equal to nprow %d * npcol %d",size,lu->nprow,lu->npcol);

535:   lu->MatInputMode = DISTRIBUTED;

537:   PetscOptionsEnum("-mat_superlu_dist_matinput","Matrix input mode (global or distributed)","None",SuperLU_MatInputModes,(PetscEnum)lu->MatInputMode,(PetscEnum*)&lu->MatInputMode,NULL);
538:   if (lu->MatInputMode == DISTRIBUTED && size == 1) lu->MatInputMode = GLOBAL;

540:   PetscOptionsBool("-mat_superlu_dist_equil","Equilibrate matrix","None",options.Equil ? PETSC_TRUE : PETSC_FALSE,&flg,&set);
541:   if (set && !flg) options.Equil = NO;

543:   PetscOptionsEList("-mat_superlu_dist_rowperm","Row permutation","None",rowperm,2,rowperm[0],&indx,&flg);
544:   if (flg) {
545:     switch (indx) {
546:     case 0:
547:       options.RowPerm = LargeDiag;
548:       break;
549:     case 1:
550:       options.RowPerm = NOROWPERM;
551:       break;
552:     }
553:   }

555:   PetscOptionsEList("-mat_superlu_dist_colperm","Column permutation","None",colperm,5,colperm[3],&indx,&flg);
556:   if (flg) {
557:     switch (indx) {
558:     case 0:
559:       options.ColPerm = NATURAL;
560:       break;
561:     case 1:
562:       options.ColPerm = MMD_AT_PLUS_A;
563:       break;
564:     case 2:
565:       options.ColPerm = MMD_ATA;
566:       break;
567:     case 3:
568:       options.ColPerm = METIS_AT_PLUS_A;
569:       break;
570:     case 4:
571:       options.ColPerm = PARMETIS;   /* only works for np>1 */
572:       break;
573:     default:
574:       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown column permutation");
575:     }
576:   }

578:   PetscOptionsBool("-mat_superlu_dist_replacetinypivot","Replace tiny pivots","None",options.ReplaceTinyPivot ? PETSC_TRUE : PETSC_FALSE,&flg,&set);
579:   if (set && !flg) options.ReplaceTinyPivot = NO;

581:   options.ParSymbFact = NO;

583:   PetscOptionsBool("-mat_superlu_dist_parsymbfact","Parallel symbolic factorization","None",PETSC_FALSE,&flg,&set);
584:   if (set && flg) {
585: #if defined(PETSC_HAVE_PARMETIS)
586:     options.ParSymbFact = YES;
587:     options.ColPerm     = PARMETIS;   /* in v2.2, PARMETIS is forced for ParSymbFact regardless of user ordering setting */
588: #else
589:     printf("parsymbfact needs PARMETIS");
590: #endif
591:   }

593:   lu->FactPattern = SamePattern_SameRowPerm;

595:   PetscOptionsEList("-mat_superlu_dist_fact","Sparsity pattern for repeated matrix factorization","None",factPattern,2,factPattern[1],&indx,&flg);
596:   if (flg) {
597:     switch (indx) {
598:     case 0:
599:       lu->FactPattern = SamePattern;
600:       break;
601:     case 1:
602:       lu->FactPattern = SamePattern_SameRowPerm;
603:       break;
604:     }
605:   }

607:   options.IterRefine = NOREFINE;
608:   PetscOptionsBool("-mat_superlu_dist_iterrefine","Use iterative refinement","None",options.IterRefine == NOREFINE ? PETSC_FALSE : PETSC_TRUE ,&flg,&set);
609:   if (set) {
610:     if (flg) options.IterRefine = SLU_DOUBLE;
611:     else options.IterRefine = NOREFINE;
612:   }

614:   if (PetscLogPrintInfo) options.PrintStat = YES;
615:   else options.PrintStat = NO;
616:   PetscOptionsBool("-mat_superlu_dist_statprint","Print factorization information","None",(PetscBool)options.PrintStat,(PetscBool*)&options.PrintStat,NULL);
617:   PetscOptionsEnd();

619:   lu->options              = options;
620:   lu->options.Fact         = DOFACT;
621:   lu->matsolve_iscalled    = PETSC_FALSE;
622:   lu->matmatsolve_iscalled = PETSC_FALSE;

624:   *F = B;
625:   return(0);
626: }

630: PETSC_EXTERN PetscErrorCode MatSolverPackageRegister_SuperLU_DIST(void)
631: {
634:   MatSolverPackageRegister(MATSOLVERSUPERLU_DIST,MATMPIAIJ,  MAT_FACTOR_LU,MatGetFactor_aij_superlu_dist);
635:   MatSolverPackageRegister(MATSOLVERSUPERLU_DIST,MATSEQAIJ,  MAT_FACTOR_LU,MatGetFactor_aij_superlu_dist);
636:   return(0);
637: }

641: PetscErrorCode MatFactorInfo_SuperLU_DIST(Mat A,PetscViewer viewer)
642: {
643:   Mat_SuperLU_DIST  *lu=(Mat_SuperLU_DIST*)A->spptr;
644:   superlu_options_t options;
645:   PetscErrorCode    ierr;

648:   /* check if matrix is superlu_dist type */
649:   if (A->ops->solve != MatSolve_SuperLU_DIST) return(0);

651:   options = lu->options;
652:   PetscViewerASCIIPrintf(viewer,"SuperLU_DIST run parameters:\n");
653:   PetscViewerASCIIPrintf(viewer,"  Process grid nprow %D x npcol %D \n",lu->nprow,lu->npcol);
654:   PetscViewerASCIIPrintf(viewer,"  Equilibrate matrix %s \n",PetscBools[options.Equil != NO]);
655:   PetscViewerASCIIPrintf(viewer,"  Matrix input mode %d \n",lu->MatInputMode);
656:   PetscViewerASCIIPrintf(viewer,"  Replace tiny pivots %s \n",PetscBools[options.ReplaceTinyPivot != NO]);
657:   PetscViewerASCIIPrintf(viewer,"  Use iterative refinement %s \n",PetscBools[options.IterRefine == SLU_DOUBLE]);
658:   PetscViewerASCIIPrintf(viewer,"  Processors in row %d col partition %d \n",lu->nprow,lu->npcol);
659:   PetscViewerASCIIPrintf(viewer,"  Row permutation %s \n",(options.RowPerm == NOROWPERM) ? "NATURAL" : "LargeDiag");

661:   switch (options.ColPerm) {
662:   case NATURAL:
663:     PetscViewerASCIIPrintf(viewer,"  Column permutation NATURAL\n");
664:     break;
665:   case MMD_AT_PLUS_A:
666:     PetscViewerASCIIPrintf(viewer,"  Column permutation MMD_AT_PLUS_A\n");
667:     break;
668:   case MMD_ATA:
669:     PetscViewerASCIIPrintf(viewer,"  Column permutation MMD_ATA\n");
670:     break;
671:   case METIS_AT_PLUS_A:
672:     PetscViewerASCIIPrintf(viewer,"  Column permutation METIS_AT_PLUS_A\n");
673:     break;
674:   case PARMETIS:
675:     PetscViewerASCIIPrintf(viewer,"  Column permutation PARMETIS\n");
676:     break;
677:   default:
678:     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown column permutation");
679:   }

681:   PetscViewerASCIIPrintf(viewer,"  Parallel symbolic factorization %s \n",PetscBools[options.ParSymbFact != NO]);

683:   if (lu->FactPattern == SamePattern) {
684:     PetscViewerASCIIPrintf(viewer,"  Repeated factorization SamePattern\n");
685:   } else {
686:     PetscViewerASCIIPrintf(viewer,"  Repeated factorization SamePattern_SameRowPerm\n");
687:   }
688:   return(0);
689: }

693: PetscErrorCode MatView_SuperLU_DIST(Mat A,PetscViewer viewer)
694: {
695:   PetscErrorCode    ierr;
696:   PetscBool         iascii;
697:   PetscViewerFormat format;

700:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
701:   if (iascii) {
702:     PetscViewerGetFormat(viewer,&format);
703:     if (format == PETSC_VIEWER_ASCII_INFO) {
704:       MatFactorInfo_SuperLU_DIST(A,viewer);
705:     }
706:   }
707:   return(0);
708: }


711: /*MC
712:   MATSOLVERSUPERLU_DIST - Parallel direct solver package for LU factorization

714:    Works with AIJ matrices

716:   Options Database Keys:
717: + -mat_superlu_dist_r <n> - number of rows in processor partition
718: . -mat_superlu_dist_c <n> - number of columns in processor partition
719: . -mat_superlu_dist_matinput <0,1> - matrix input mode; 0=global, 1=distributed
720: . -mat_superlu_dist_equil - equilibrate the matrix
721: . -mat_superlu_dist_rowperm <LargeDiag,NATURAL> - row permutation
722: . -mat_superlu_dist_colperm <MMD_AT_PLUS_A,MMD_ATA,NATURAL> - column permutation
723: . -mat_superlu_dist_replacetinypivot - replace tiny pivots
724: . -mat_superlu_dist_fact <SamePattern> - (choose one of) SamePattern SamePattern_SameRowPerm
725: . -mat_superlu_dist_iterrefine - use iterative refinement
726: - -mat_superlu_dist_statprint - print factorization information

728:    Level: beginner

730: .seealso: PCLU

732: .seealso: PCFactorSetMatSolverPackage(), MatSolverPackage

734: M*/