1 2 /* 3 Provides an interface to the MUMPS sparse solver 4 */ 5 6 #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/ 7 #include <../src/mat/impls/sbaij/mpi/mpisbaij.h> 8 9 EXTERN_C_BEGIN 10 #if defined(PETSC_USE_COMPLEX) 11 #if defined(PETSC_USE_REAL_SINGLE) 12 #include <cmumps_c.h> 13 #else 14 #include <zmumps_c.h> 15 #endif 16 #else 17 #if defined(PETSC_USE_REAL_SINGLE) 18 #include <smumps_c.h> 19 #else 20 #include <dmumps_c.h> 21 #endif 22 #endif 23 EXTERN_C_END 24 #define JOB_INIT -1 25 #define JOB_FACTSYMBOLIC 1 26 #define JOB_FACTNUMERIC 2 27 #define JOB_SOLVE 3 28 #define JOB_END -2 29 30 /* calls to MUMPS */ 31 #if defined(PETSC_USE_COMPLEX) 32 #if defined(PETSC_USE_REAL_SINGLE) 33 #define PetscMUMPS_c cmumps_c 34 #else 35 #define PetscMUMPS_c zmumps_c 36 #endif 37 #else 38 #if defined(PETSC_USE_REAL_SINGLE) 39 #define PetscMUMPS_c smumps_c 40 #else 41 #define PetscMUMPS_c dmumps_c 42 #endif 43 #endif 44 45 /* declare MumpsScalar */ 46 #if defined(PETSC_USE_COMPLEX) 47 #if defined(PETSC_USE_REAL_SINGLE) 48 #define MumpsScalar mumps_complex 49 #else 50 #define MumpsScalar mumps_double_complex 51 #endif 52 #else 53 #define MumpsScalar PetscScalar 54 #endif 55 56 /* macros s.t. indices match MUMPS documentation */ 57 #define ICNTL(I) icntl[(I)-1] 58 #define CNTL(I) cntl[(I)-1] 59 #define INFOG(I) infog[(I)-1] 60 #define INFO(I) info[(I)-1] 61 #define RINFOG(I) rinfog[(I)-1] 62 #define RINFO(I) rinfo[(I)-1] 63 64 typedef struct { 65 #if defined(PETSC_USE_COMPLEX) 66 #if defined(PETSC_USE_REAL_SINGLE) 67 CMUMPS_STRUC_C id; 68 #else 69 ZMUMPS_STRUC_C id; 70 #endif 71 #else 72 #if defined(PETSC_USE_REAL_SINGLE) 73 SMUMPS_STRUC_C id; 74 #else 75 DMUMPS_STRUC_C id; 76 #endif 77 #endif 78 79 MatStructure matstruc; 80 PetscMPIInt myid,size; 81 PetscInt *irn,*jcn,nz,sym; 82 PetscScalar *val; 83 MPI_Comm comm_mumps; 84 PetscBool isAIJ; 85 PetscInt ICNTL9_pre; /* check if ICNTL(9) is changed from previous MatSolve */ 86 VecScatter scat_rhs, scat_sol; /* used by MatSolve() */ 87 Vec b_seq,x_seq; 88 PetscInt ninfo,*info; /* display INFO */ 89 PetscInt sizeredrhs; 90 PetscScalar *schur_sol; 91 PetscInt schur_sizesol; 92 93 PetscErrorCode (*ConvertToTriples)(Mat, int, MatReuse, int*, int**, int**, PetscScalar**); 94 } Mat_MUMPS; 95 96 extern PetscErrorCode MatDuplicate_MUMPS(Mat,MatDuplicateOption,Mat*); 97 98 static PetscErrorCode MatMumpsResetSchur_Private(Mat_MUMPS* mumps) 99 { 100 PetscErrorCode ierr; 101 102 PetscFunctionBegin; 103 ierr = PetscFree2(mumps->id.listvar_schur,mumps->id.schur);CHKERRQ(ierr); 104 ierr = PetscFree(mumps->id.redrhs);CHKERRQ(ierr); 105 ierr = PetscFree(mumps->schur_sol);CHKERRQ(ierr); 106 mumps->id.size_schur = 0; 107 mumps->id.schur_lld = 0; 108 mumps->id.ICNTL(19) = 0; 109 PetscFunctionReturn(0); 110 } 111 112 /* solve with rhs in mumps->id.redrhs and return in the same location */ 113 static PetscErrorCode MatMumpsSolveSchur_Private(Mat F) 114 { 115 Mat_MUMPS *mumps=(Mat_MUMPS*)F->data; 116 Mat S,B,X; 117 MatFactorSchurStatus schurstatus; 118 PetscInt sizesol; 119 PetscErrorCode ierr; 120 121 PetscFunctionBegin; 122 ierr = MatFactorFactorizeSchurComplement(F);CHKERRQ(ierr); 123 ierr = MatFactorGetSchurComplement(F,&S,&schurstatus);CHKERRQ(ierr); 124 ierr = MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,(PetscScalar*)mumps->id.redrhs,&B);CHKERRQ(ierr); 125 switch (schurstatus) { 126 case MAT_FACTOR_SCHUR_FACTORED: 127 ierr = MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,(PetscScalar*)mumps->id.redrhs,&X);CHKERRQ(ierr); 128 if (!mumps->id.ICNTL(9)) { /* transpose solve */ 129 ierr = MatMatSolveTranspose(S,B,X);CHKERRQ(ierr); 130 } else { 131 ierr = MatMatSolve(S,B,X);CHKERRQ(ierr); 132 } 133 break; 134 case MAT_FACTOR_SCHUR_INVERTED: 135 sizesol = mumps->id.nrhs*mumps->id.size_schur; 136 if (!mumps->schur_sol || sizesol > mumps->schur_sizesol) { 137 ierr = PetscFree(mumps->schur_sol);CHKERRQ(ierr); 138 ierr = PetscMalloc1(sizesol,&mumps->schur_sol);CHKERRQ(ierr); 139 mumps->schur_sizesol = sizesol; 140 } 141 ierr = MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,mumps->schur_sol,&X);CHKERRQ(ierr); 142 if (!mumps->id.ICNTL(9)) { /* transpose solve */ 143 ierr = MatTransposeMatMult(S,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&X);CHKERRQ(ierr); 144 } else { 145 ierr = MatMatMult(S,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&X);CHKERRQ(ierr); 146 } 147 ierr = MatCopy(X,B,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 148 break; 149 default: 150 SETERRQ1(PetscObjectComm((PetscObject)F),PETSC_ERR_SUP,"Unhandled MatFactorSchurStatus %D",F->schur_status); 151 break; 152 } 153 ierr = MatFactorRestoreSchurComplement(F,&S,schurstatus);CHKERRQ(ierr); 154 ierr = MatDestroy(&B);CHKERRQ(ierr); 155 ierr = MatDestroy(&X);CHKERRQ(ierr); 156 PetscFunctionReturn(0); 157 } 158 159 static PetscErrorCode MatMumpsHandleSchur_Private(Mat F, PetscBool expansion) 160 { 161 Mat_MUMPS *mumps=(Mat_MUMPS*)F->data; 162 PetscErrorCode ierr; 163 164 PetscFunctionBegin; 165 if (!mumps->id.ICNTL(19)) { /* do nothing when Schur complement has not been computed */ 166 PetscFunctionReturn(0); 167 } 168 if (!expansion) { /* prepare for the condensation step */ 169 PetscInt sizeredrhs = mumps->id.nrhs*mumps->id.size_schur; 170 /* allocate MUMPS internal array to store reduced right-hand sides */ 171 if (!mumps->id.redrhs || sizeredrhs > mumps->sizeredrhs) { 172 ierr = PetscFree(mumps->id.redrhs);CHKERRQ(ierr); 173 mumps->id.lredrhs = mumps->id.size_schur; 174 ierr = PetscMalloc1(mumps->id.nrhs*mumps->id.lredrhs,&mumps->id.redrhs);CHKERRQ(ierr); 175 mumps->sizeredrhs = mumps->id.nrhs*mumps->id.lredrhs; 176 } 177 mumps->id.ICNTL(26) = 1; /* condensation phase */ 178 } else { /* prepare for the expansion step */ 179 /* solve Schur complement (this has to be done by the MUMPS user, so basically us) */ 180 ierr = MatMumpsSolveSchur_Private(F);CHKERRQ(ierr); 181 mumps->id.ICNTL(26) = 2; /* expansion phase */ 182 PetscMUMPS_c(&mumps->id); 183 if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d\n",mumps->id.INFOG(1)); 184 /* restore defaults */ 185 mumps->id.ICNTL(26) = -1; 186 /* free MUMPS internal array for redrhs if we have solved for multiple rhs in order to save memory space */ 187 if (mumps->id.nrhs > 1) { 188 ierr = PetscFree(mumps->id.redrhs);CHKERRQ(ierr); 189 mumps->id.lredrhs = 0; 190 mumps->sizeredrhs = 0; 191 } 192 } 193 PetscFunctionReturn(0); 194 } 195 196 /* 197 MatConvertToTriples_A_B - convert Petsc matrix to triples: row[nz], col[nz], val[nz] 198 199 input: 200 A - matrix in aij,baij or sbaij (bs=1) format 201 shift - 0: C style output triple; 1: Fortran style output triple. 202 reuse - MAT_INITIAL_MATRIX: spaces are allocated and values are set for the triple 203 MAT_REUSE_MATRIX: only the values in v array are updated 204 output: 205 nnz - dim of r, c, and v (number of local nonzero entries of A) 206 r, c, v - row and col index, matrix values (matrix triples) 207 208 The returned values r, c, and sometimes v are obtained in a single PetscMalloc(). Then in MatDestroy_MUMPS() it is 209 freed with PetscFree((mumps->irn); This is not ideal code, the fact that v is ONLY sometimes part of mumps->irn means 210 that the PetscMalloc() cannot easily be replaced with a PetscMalloc3(). 211 212 */ 213 214 PetscErrorCode MatConvertToTriples_seqaij_seqaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 215 { 216 const PetscInt *ai,*aj,*ajj,M=A->rmap->n; 217 PetscInt nz,rnz,i,j; 218 PetscErrorCode ierr; 219 PetscInt *row,*col; 220 Mat_SeqAIJ *aa=(Mat_SeqAIJ*)A->data; 221 222 PetscFunctionBegin; 223 *v=aa->a; 224 if (reuse == MAT_INITIAL_MATRIX) { 225 nz = aa->nz; 226 ai = aa->i; 227 aj = aa->j; 228 *nnz = nz; 229 ierr = PetscMalloc1(2*nz, &row);CHKERRQ(ierr); 230 col = row + nz; 231 232 nz = 0; 233 for (i=0; i<M; i++) { 234 rnz = ai[i+1] - ai[i]; 235 ajj = aj + ai[i]; 236 for (j=0; j<rnz; j++) { 237 row[nz] = i+shift; col[nz++] = ajj[j] + shift; 238 } 239 } 240 *r = row; *c = col; 241 } 242 PetscFunctionReturn(0); 243 } 244 245 PetscErrorCode MatConvertToTriples_seqbaij_seqaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 246 { 247 Mat_SeqBAIJ *aa=(Mat_SeqBAIJ*)A->data; 248 const PetscInt *ai,*aj,*ajj,bs2 = aa->bs2; 249 PetscInt bs,M,nz,idx=0,rnz,i,j,k,m; 250 PetscErrorCode ierr; 251 PetscInt *row,*col; 252 253 PetscFunctionBegin; 254 ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr); 255 M = A->rmap->N/bs; 256 *v = aa->a; 257 if (reuse == MAT_INITIAL_MATRIX) { 258 ai = aa->i; aj = aa->j; 259 nz = bs2*aa->nz; 260 *nnz = nz; 261 ierr = PetscMalloc1(2*nz, &row);CHKERRQ(ierr); 262 col = row + nz; 263 264 for (i=0; i<M; i++) { 265 ajj = aj + ai[i]; 266 rnz = ai[i+1] - ai[i]; 267 for (k=0; k<rnz; k++) { 268 for (j=0; j<bs; j++) { 269 for (m=0; m<bs; m++) { 270 row[idx] = i*bs + m + shift; 271 col[idx++] = bs*(ajj[k]) + j + shift; 272 } 273 } 274 } 275 } 276 *r = row; *c = col; 277 } 278 PetscFunctionReturn(0); 279 } 280 281 PetscErrorCode MatConvertToTriples_seqsbaij_seqsbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 282 { 283 const PetscInt *ai, *aj,*ajj,M=A->rmap->n; 284 PetscInt nz,rnz,i,j; 285 PetscErrorCode ierr; 286 PetscInt *row,*col; 287 Mat_SeqSBAIJ *aa=(Mat_SeqSBAIJ*)A->data; 288 289 PetscFunctionBegin; 290 *v = aa->a; 291 if (reuse == MAT_INITIAL_MATRIX) { 292 nz = aa->nz; 293 ai = aa->i; 294 aj = aa->j; 295 *v = aa->a; 296 *nnz = nz; 297 ierr = PetscMalloc1(2*nz, &row);CHKERRQ(ierr); 298 col = row + nz; 299 300 nz = 0; 301 for (i=0; i<M; i++) { 302 rnz = ai[i+1] - ai[i]; 303 ajj = aj + ai[i]; 304 for (j=0; j<rnz; j++) { 305 row[nz] = i+shift; col[nz++] = ajj[j] + shift; 306 } 307 } 308 *r = row; *c = col; 309 } 310 PetscFunctionReturn(0); 311 } 312 313 PetscErrorCode MatConvertToTriples_seqaij_seqsbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 314 { 315 const PetscInt *ai,*aj,*ajj,*adiag,M=A->rmap->n; 316 PetscInt nz,rnz,i,j; 317 const PetscScalar *av,*v1; 318 PetscScalar *val; 319 PetscErrorCode ierr; 320 PetscInt *row,*col; 321 Mat_SeqAIJ *aa=(Mat_SeqAIJ*)A->data; 322 PetscBool missing; 323 324 PetscFunctionBegin; 325 ai =aa->i; aj=aa->j;av=aa->a; 326 adiag=aa->diag; 327 ierr = MatMissingDiagonal_SeqAIJ(A,&missing,&i);CHKERRQ(ierr); 328 if (reuse == MAT_INITIAL_MATRIX) { 329 /* count nz in the uppper triangular part of A */ 330 nz = 0; 331 if (missing) { 332 for (i=0; i<M; i++) { 333 if (PetscUnlikely(adiag[i] >= ai[i+1])) { 334 for (j=ai[i];j<ai[i+1];j++) { 335 if (aj[j] < i) continue; 336 nz++; 337 } 338 } else { 339 nz += ai[i+1] - adiag[i]; 340 } 341 } 342 } else { 343 for (i=0; i<M; i++) nz += ai[i+1] - adiag[i]; 344 } 345 *nnz = nz; 346 347 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 348 col = row + nz; 349 val = (PetscScalar*)(col + nz); 350 351 nz = 0; 352 if (missing) { 353 for (i=0; i<M; i++) { 354 if (PetscUnlikely(adiag[i] >= ai[i+1])) { 355 for (j=ai[i];j<ai[i+1];j++) { 356 if (aj[j] < i) continue; 357 row[nz] = i+shift; 358 col[nz] = aj[j]+shift; 359 val[nz] = av[j]; 360 nz++; 361 } 362 } else { 363 rnz = ai[i+1] - adiag[i]; 364 ajj = aj + adiag[i]; 365 v1 = av + adiag[i]; 366 for (j=0; j<rnz; j++) { 367 row[nz] = i+shift; col[nz] = ajj[j] + shift; val[nz++] = v1[j]; 368 } 369 } 370 } 371 } else { 372 for (i=0; i<M; i++) { 373 rnz = ai[i+1] - adiag[i]; 374 ajj = aj + adiag[i]; 375 v1 = av + adiag[i]; 376 for (j=0; j<rnz; j++) { 377 row[nz] = i+shift; col[nz] = ajj[j] + shift; val[nz++] = v1[j]; 378 } 379 } 380 } 381 *r = row; *c = col; *v = val; 382 } else { 383 nz = 0; val = *v; 384 if (missing) { 385 for (i=0; i <M; i++) { 386 if (PetscUnlikely(adiag[i] >= ai[i+1])) { 387 for (j=ai[i];j<ai[i+1];j++) { 388 if (aj[j] < i) continue; 389 val[nz++] = av[j]; 390 } 391 } else { 392 rnz = ai[i+1] - adiag[i]; 393 v1 = av + adiag[i]; 394 for (j=0; j<rnz; j++) { 395 val[nz++] = v1[j]; 396 } 397 } 398 } 399 } else { 400 for (i=0; i <M; i++) { 401 rnz = ai[i+1] - adiag[i]; 402 v1 = av + adiag[i]; 403 for (j=0; j<rnz; j++) { 404 val[nz++] = v1[j]; 405 } 406 } 407 } 408 } 409 PetscFunctionReturn(0); 410 } 411 412 PetscErrorCode MatConvertToTriples_mpisbaij_mpisbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 413 { 414 const PetscInt *ai, *aj, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj; 415 PetscErrorCode ierr; 416 PetscInt rstart,nz,i,j,jj,irow,countA,countB; 417 PetscInt *row,*col; 418 const PetscScalar *av, *bv,*v1,*v2; 419 PetscScalar *val; 420 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)A->data; 421 Mat_SeqSBAIJ *aa = (Mat_SeqSBAIJ*)(mat->A)->data; 422 Mat_SeqBAIJ *bb = (Mat_SeqBAIJ*)(mat->B)->data; 423 424 PetscFunctionBegin; 425 ai=aa->i; aj=aa->j; bi=bb->i; bj=bb->j; rstart= A->rmap->rstart; 426 av=aa->a; bv=bb->a; 427 428 garray = mat->garray; 429 430 if (reuse == MAT_INITIAL_MATRIX) { 431 nz = aa->nz + bb->nz; 432 *nnz = nz; 433 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 434 col = row + nz; 435 val = (PetscScalar*)(col + nz); 436 437 *r = row; *c = col; *v = val; 438 } else { 439 row = *r; col = *c; val = *v; 440 } 441 442 jj = 0; irow = rstart; 443 for (i=0; i<m; i++) { 444 ajj = aj + ai[i]; /* ptr to the beginning of this row */ 445 countA = ai[i+1] - ai[i]; 446 countB = bi[i+1] - bi[i]; 447 bjj = bj + bi[i]; 448 v1 = av + ai[i]; 449 v2 = bv + bi[i]; 450 451 /* A-part */ 452 for (j=0; j<countA; j++) { 453 if (reuse == MAT_INITIAL_MATRIX) { 454 row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift; 455 } 456 val[jj++] = v1[j]; 457 } 458 459 /* B-part */ 460 for (j=0; j < countB; j++) { 461 if (reuse == MAT_INITIAL_MATRIX) { 462 row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift; 463 } 464 val[jj++] = v2[j]; 465 } 466 irow++; 467 } 468 PetscFunctionReturn(0); 469 } 470 471 PetscErrorCode MatConvertToTriples_mpiaij_mpiaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 472 { 473 const PetscInt *ai, *aj, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj; 474 PetscErrorCode ierr; 475 PetscInt rstart,nz,i,j,jj,irow,countA,countB; 476 PetscInt *row,*col; 477 const PetscScalar *av, *bv,*v1,*v2; 478 PetscScalar *val; 479 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 480 Mat_SeqAIJ *aa = (Mat_SeqAIJ*)(mat->A)->data; 481 Mat_SeqAIJ *bb = (Mat_SeqAIJ*)(mat->B)->data; 482 483 PetscFunctionBegin; 484 ai=aa->i; aj=aa->j; bi=bb->i; bj=bb->j; rstart= A->rmap->rstart; 485 av=aa->a; bv=bb->a; 486 487 garray = mat->garray; 488 489 if (reuse == MAT_INITIAL_MATRIX) { 490 nz = aa->nz + bb->nz; 491 *nnz = nz; 492 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 493 col = row + nz; 494 val = (PetscScalar*)(col + nz); 495 496 *r = row; *c = col; *v = val; 497 } else { 498 row = *r; col = *c; val = *v; 499 } 500 501 jj = 0; irow = rstart; 502 for (i=0; i<m; i++) { 503 ajj = aj + ai[i]; /* ptr to the beginning of this row */ 504 countA = ai[i+1] - ai[i]; 505 countB = bi[i+1] - bi[i]; 506 bjj = bj + bi[i]; 507 v1 = av + ai[i]; 508 v2 = bv + bi[i]; 509 510 /* A-part */ 511 for (j=0; j<countA; j++) { 512 if (reuse == MAT_INITIAL_MATRIX) { 513 row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift; 514 } 515 val[jj++] = v1[j]; 516 } 517 518 /* B-part */ 519 for (j=0; j < countB; j++) { 520 if (reuse == MAT_INITIAL_MATRIX) { 521 row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift; 522 } 523 val[jj++] = v2[j]; 524 } 525 irow++; 526 } 527 PetscFunctionReturn(0); 528 } 529 530 PetscErrorCode MatConvertToTriples_mpibaij_mpiaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 531 { 532 Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)A->data; 533 Mat_SeqBAIJ *aa = (Mat_SeqBAIJ*)(mat->A)->data; 534 Mat_SeqBAIJ *bb = (Mat_SeqBAIJ*)(mat->B)->data; 535 const PetscInt *ai = aa->i, *bi = bb->i, *aj = aa->j, *bj = bb->j,*ajj, *bjj; 536 const PetscInt *garray = mat->garray,mbs=mat->mbs,rstart=A->rmap->rstart; 537 const PetscInt bs2=mat->bs2; 538 PetscErrorCode ierr; 539 PetscInt bs,nz,i,j,k,n,jj,irow,countA,countB,idx; 540 PetscInt *row,*col; 541 const PetscScalar *av=aa->a, *bv=bb->a,*v1,*v2; 542 PetscScalar *val; 543 544 PetscFunctionBegin; 545 ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr); 546 if (reuse == MAT_INITIAL_MATRIX) { 547 nz = bs2*(aa->nz + bb->nz); 548 *nnz = nz; 549 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 550 col = row + nz; 551 val = (PetscScalar*)(col + nz); 552 553 *r = row; *c = col; *v = val; 554 } else { 555 row = *r; col = *c; val = *v; 556 } 557 558 jj = 0; irow = rstart; 559 for (i=0; i<mbs; i++) { 560 countA = ai[i+1] - ai[i]; 561 countB = bi[i+1] - bi[i]; 562 ajj = aj + ai[i]; 563 bjj = bj + bi[i]; 564 v1 = av + bs2*ai[i]; 565 v2 = bv + bs2*bi[i]; 566 567 idx = 0; 568 /* A-part */ 569 for (k=0; k<countA; k++) { 570 for (j=0; j<bs; j++) { 571 for (n=0; n<bs; n++) { 572 if (reuse == MAT_INITIAL_MATRIX) { 573 row[jj] = irow + n + shift; 574 col[jj] = rstart + bs*ajj[k] + j + shift; 575 } 576 val[jj++] = v1[idx++]; 577 } 578 } 579 } 580 581 idx = 0; 582 /* B-part */ 583 for (k=0; k<countB; k++) { 584 for (j=0; j<bs; j++) { 585 for (n=0; n<bs; n++) { 586 if (reuse == MAT_INITIAL_MATRIX) { 587 row[jj] = irow + n + shift; 588 col[jj] = bs*garray[bjj[k]] + j + shift; 589 } 590 val[jj++] = v2[idx++]; 591 } 592 } 593 } 594 irow += bs; 595 } 596 PetscFunctionReturn(0); 597 } 598 599 PetscErrorCode MatConvertToTriples_mpiaij_mpisbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 600 { 601 const PetscInt *ai, *aj,*adiag, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj; 602 PetscErrorCode ierr; 603 PetscInt rstart,nz,nza,nzb,i,j,jj,irow,countA,countB; 604 PetscInt *row,*col; 605 const PetscScalar *av, *bv,*v1,*v2; 606 PetscScalar *val; 607 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 608 Mat_SeqAIJ *aa =(Mat_SeqAIJ*)(mat->A)->data; 609 Mat_SeqAIJ *bb =(Mat_SeqAIJ*)(mat->B)->data; 610 611 PetscFunctionBegin; 612 ai=aa->i; aj=aa->j; adiag=aa->diag; 613 bi=bb->i; bj=bb->j; garray = mat->garray; 614 av=aa->a; bv=bb->a; 615 616 rstart = A->rmap->rstart; 617 618 if (reuse == MAT_INITIAL_MATRIX) { 619 nza = 0; /* num of upper triangular entries in mat->A, including diagonals */ 620 nzb = 0; /* num of upper triangular entries in mat->B */ 621 for (i=0; i<m; i++) { 622 nza += (ai[i+1] - adiag[i]); 623 countB = bi[i+1] - bi[i]; 624 bjj = bj + bi[i]; 625 for (j=0; j<countB; j++) { 626 if (garray[bjj[j]] > rstart) nzb++; 627 } 628 } 629 630 nz = nza + nzb; /* total nz of upper triangular part of mat */ 631 *nnz = nz; 632 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 633 col = row + nz; 634 val = (PetscScalar*)(col + nz); 635 636 *r = row; *c = col; *v = val; 637 } else { 638 row = *r; col = *c; val = *v; 639 } 640 641 jj = 0; irow = rstart; 642 for (i=0; i<m; i++) { 643 ajj = aj + adiag[i]; /* ptr to the beginning of the diagonal of this row */ 644 v1 = av + adiag[i]; 645 countA = ai[i+1] - adiag[i]; 646 countB = bi[i+1] - bi[i]; 647 bjj = bj + bi[i]; 648 v2 = bv + bi[i]; 649 650 /* A-part */ 651 for (j=0; j<countA; j++) { 652 if (reuse == MAT_INITIAL_MATRIX) { 653 row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift; 654 } 655 val[jj++] = v1[j]; 656 } 657 658 /* B-part */ 659 for (j=0; j < countB; j++) { 660 if (garray[bjj[j]] > rstart) { 661 if (reuse == MAT_INITIAL_MATRIX) { 662 row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift; 663 } 664 val[jj++] = v2[j]; 665 } 666 } 667 irow++; 668 } 669 PetscFunctionReturn(0); 670 } 671 672 PetscErrorCode MatDestroy_MUMPS(Mat A) 673 { 674 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 675 PetscErrorCode ierr; 676 677 PetscFunctionBegin; 678 ierr = PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);CHKERRQ(ierr); 679 ierr = VecScatterDestroy(&mumps->scat_rhs);CHKERRQ(ierr); 680 ierr = VecScatterDestroy(&mumps->scat_sol);CHKERRQ(ierr); 681 ierr = VecDestroy(&mumps->b_seq);CHKERRQ(ierr); 682 ierr = VecDestroy(&mumps->x_seq);CHKERRQ(ierr); 683 ierr = PetscFree(mumps->id.perm_in);CHKERRQ(ierr); 684 ierr = PetscFree(mumps->irn);CHKERRQ(ierr); 685 ierr = PetscFree(mumps->info);CHKERRQ(ierr); 686 ierr = MatMumpsResetSchur_Private(mumps);CHKERRQ(ierr); 687 mumps->id.job = JOB_END; 688 PetscMUMPS_c(&mumps->id); 689 ierr = MPI_Comm_free(&mumps->comm_mumps);CHKERRQ(ierr); 690 ierr = PetscFree(A->data);CHKERRQ(ierr); 691 692 /* clear composed functions */ 693 ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverType_C",NULL);CHKERRQ(ierr); 694 ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorSetSchurIS_C",NULL);CHKERRQ(ierr); 695 ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorCreateSchurComplement_C",NULL);CHKERRQ(ierr); 696 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetIcntl_C",NULL);CHKERRQ(ierr); 697 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetIcntl_C",NULL);CHKERRQ(ierr); 698 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetCntl_C",NULL);CHKERRQ(ierr); 699 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetCntl_C",NULL);CHKERRQ(ierr); 700 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfo_C",NULL);CHKERRQ(ierr); 701 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfog_C",NULL);CHKERRQ(ierr); 702 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfo_C",NULL);CHKERRQ(ierr); 703 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfog_C",NULL);CHKERRQ(ierr); 704 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetMatInverse_C",NULL);CHKERRQ(ierr); 705 PetscFunctionReturn(0); 706 } 707 708 PetscErrorCode MatSolve_MUMPS(Mat A,Vec b,Vec x) 709 { 710 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 711 PetscScalar *array; 712 Vec b_seq; 713 IS is_iden,is_petsc; 714 PetscErrorCode ierr; 715 PetscInt i; 716 PetscBool second_solve = PETSC_FALSE; 717 static PetscBool cite1 = PETSC_FALSE,cite2 = PETSC_FALSE; 718 719 PetscFunctionBegin; 720 ierr = PetscCitationsRegister("@article{MUMPS01,\n author = {P.~R. Amestoy and I.~S. Duff and J.-Y. L'Excellent and J. Koster},\n title = {A fully asynchronous multifrontal solver using distributed dynamic scheduling},\n journal = {SIAM Journal on Matrix Analysis and Applications},\n volume = {23},\n number = {1},\n pages = {15--41},\n year = {2001}\n}\n",&cite1);CHKERRQ(ierr); 721 ierr = PetscCitationsRegister("@article{MUMPS02,\n author = {P.~R. Amestoy and A. Guermouche and J.-Y. L'Excellent and S. Pralet},\n title = {Hybrid scheduling for the parallel solution of linear systems},\n journal = {Parallel Computing},\n volume = {32},\n number = {2},\n pages = {136--156},\n year = {2006}\n}\n",&cite2);CHKERRQ(ierr); 722 723 if (A->factorerrortype) { 724 ierr = PetscInfo2(A,"MatSolve is called with singular matrix factor, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 725 ierr = VecSetInf(x);CHKERRQ(ierr); 726 PetscFunctionReturn(0); 727 } 728 729 mumps->id.nrhs = 1; 730 b_seq = mumps->b_seq; 731 if (mumps->size > 1) { 732 /* MUMPS only supports centralized rhs. Scatter b into a seqential rhs vector */ 733 ierr = VecScatterBegin(mumps->scat_rhs,b,b_seq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 734 ierr = VecScatterEnd(mumps->scat_rhs,b,b_seq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 735 if (!mumps->myid) {ierr = VecGetArray(b_seq,&array);CHKERRQ(ierr);} 736 } else { /* size == 1 */ 737 ierr = VecCopy(b,x);CHKERRQ(ierr); 738 ierr = VecGetArray(x,&array);CHKERRQ(ierr); 739 } 740 if (!mumps->myid) { /* define rhs on the host */ 741 mumps->id.nrhs = 1; 742 mumps->id.rhs = (MumpsScalar*)array; 743 } 744 745 /* 746 handle condensation step of Schur complement (if any) 747 We set by default ICNTL(26) == -1 when Schur indices have been provided by the user. 748 According to MUMPS (5.0.0) manual, any value should be harmful during the factorization phase 749 Unless the user provides a valid value for ICNTL(26), MatSolve and MatMatSolve routines solve the full system. 750 This requires an extra call to PetscMUMPS_c and the computation of the factors for S 751 */ 752 if (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2) { 753 if (mumps->size > 1) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Parallel Schur complements not yet supported from PETSc\n"); 754 second_solve = PETSC_TRUE; 755 ierr = MatMumpsHandleSchur_Private(A,PETSC_FALSE);CHKERRQ(ierr); 756 } 757 /* solve phase */ 758 /*-------------*/ 759 mumps->id.job = JOB_SOLVE; 760 PetscMUMPS_c(&mumps->id); 761 if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d\n",mumps->id.INFOG(1)); 762 763 /* handle expansion step of Schur complement (if any) */ 764 if (second_solve) { 765 ierr = MatMumpsHandleSchur_Private(A,PETSC_TRUE);CHKERRQ(ierr); 766 } 767 768 if (mumps->size > 1) { /* convert mumps distributed solution to petsc mpi x */ 769 if (mumps->scat_sol && mumps->ICNTL9_pre != mumps->id.ICNTL(9)) { 770 /* when id.ICNTL(9) changes, the contents of lsol_loc may change (not its size, lsol_loc), recreates scat_sol */ 771 ierr = VecScatterDestroy(&mumps->scat_sol);CHKERRQ(ierr); 772 } 773 if (!mumps->scat_sol) { /* create scatter scat_sol */ 774 ierr = ISCreateStride(PETSC_COMM_SELF,mumps->id.lsol_loc,0,1,&is_iden);CHKERRQ(ierr); /* from */ 775 for (i=0; i<mumps->id.lsol_loc; i++) { 776 mumps->id.isol_loc[i] -= 1; /* change Fortran style to C style */ 777 } 778 ierr = ISCreateGeneral(PETSC_COMM_SELF,mumps->id.lsol_loc,mumps->id.isol_loc,PETSC_COPY_VALUES,&is_petsc);CHKERRQ(ierr); /* to */ 779 ierr = VecScatterCreate(mumps->x_seq,is_iden,x,is_petsc,&mumps->scat_sol);CHKERRQ(ierr); 780 ierr = ISDestroy(&is_iden);CHKERRQ(ierr); 781 ierr = ISDestroy(&is_petsc);CHKERRQ(ierr); 782 783 mumps->ICNTL9_pre = mumps->id.ICNTL(9); /* save current value of id.ICNTL(9) */ 784 } 785 786 ierr = VecScatterBegin(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 787 ierr = VecScatterEnd(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 788 } 789 PetscFunctionReturn(0); 790 } 791 792 PetscErrorCode MatSolveTranspose_MUMPS(Mat A,Vec b,Vec x) 793 { 794 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 795 PetscErrorCode ierr; 796 797 PetscFunctionBegin; 798 mumps->id.ICNTL(9) = 0; 799 ierr = MatSolve_MUMPS(A,b,x);CHKERRQ(ierr); 800 mumps->id.ICNTL(9) = 1; 801 PetscFunctionReturn(0); 802 } 803 804 PetscErrorCode MatMatSolve_MUMPS(Mat A,Mat B,Mat X) 805 { 806 PetscErrorCode ierr; 807 Mat Bt = NULL; 808 PetscBool flg, flgT; 809 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 810 PetscInt i,nrhs,M; 811 PetscScalar *array,*bray; 812 813 PetscFunctionBegin; 814 ierr = PetscObjectTypeCompareAny((PetscObject)B,&flg,MATSEQDENSE,MATMPIDENSE,NULL);CHKERRQ(ierr); 815 ierr = PetscObjectTypeCompare((PetscObject)B,MATTRANSPOSEMAT,&flgT);CHKERRQ(ierr); 816 if (flgT) { 817 if (mumps->size > 1) SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_WRONG,"Matrix B must be MATDENSE matrix"); 818 ierr = MatTransposeGetMat(B,&Bt);CHKERRQ(ierr); 819 } else { 820 if (!flg) SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_WRONG,"Matrix B must be MATDENSE matrix"); 821 } 822 823 ierr = MatGetSize(B,&M,&nrhs);CHKERRQ(ierr); 824 mumps->id.nrhs = nrhs; 825 mumps->id.lrhs = M; 826 827 #if 0 828 if (mumps->id.ICNTL(30)) { 829 if (mumps->size == 1 && Bt) { 830 PetscBool done; 831 PetscScalar *aa; 832 PetscInt spnr,*ia,*ja; 833 834 ierr = MatSeqAIJGetArray(Bt,&aa);CHKERRQ(ierr); 835 ierr = MatGetRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&done);CHKERRQ(ierr); 836 if (!done) SETERRQ(PetscObjectComm((PetscObject)Bt),PETSC_ERR_ARG_WRONG,"Cannot get IJ structure"); 837 838 mumps->id.irhs_ptr = ia; 839 mumps->id.irhs_sparse = ja; 840 mumps->id.nz_rhs = ia[spnr] - 1; 841 mumps->id.rhs_sparse = (MumpsScalar*)aa; 842 mumps->id.ICNTL(20) = 1; /* rhs is sparse */ 843 844 /* solve phase */ 845 /*-------------*/ 846 mumps->id.job = JOB_SOLVE; 847 PetscMUMPS_c(&mumps->id); 848 if (mumps->id.INFOG(1) < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d INFOG(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFOG(2)); 849 } else { 850 SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"not done yet"); 851 } 852 PetscFunctionReturn(0); 853 } 854 #endif 855 ierr = PetscObjectTypeCompareAny((PetscObject)X,&flg,MATSEQDENSE,MATMPIDENSE,NULL);CHKERRQ(ierr); 856 if (!flg) SETERRQ(PetscObjectComm((PetscObject)X),PETSC_ERR_ARG_WRONG,"Matrix X must be MATDENSE matrix"); 857 858 if (B->rmap->n != X->rmap->n) SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_WRONG,"Matrix B and X must have same row distribution"); 859 860 if (mumps->size == 1) { 861 PetscScalar *aa; 862 PetscInt spnr,*ia,*ja; 863 PetscBool second_solve = PETSC_FALSE; 864 865 /* copy B to X */ 866 ierr = MatDenseGetArray(X,&array);CHKERRQ(ierr); 867 mumps->id.rhs = (MumpsScalar*)array; 868 if (!Bt) { 869 ierr = MatDenseGetArray(B,&bray);CHKERRQ(ierr); 870 ierr = PetscMemcpy(array,bray,M*nrhs*sizeof(PetscScalar));CHKERRQ(ierr); 871 ierr = MatDenseRestoreArray(B,&bray);CHKERRQ(ierr); 872 } else { 873 PetscBool done; 874 875 ierr = MatSeqAIJGetArray(Bt,&aa);CHKERRQ(ierr); 876 ierr = MatGetRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&done);CHKERRQ(ierr); 877 if (!done) SETERRQ(PetscObjectComm((PetscObject)Bt),PETSC_ERR_ARG_WRONG,"Cannot get IJ structure"); 878 mumps->id.irhs_ptr = ia; 879 mumps->id.irhs_sparse = ja; 880 mumps->id.nz_rhs = ia[spnr] - 1; 881 mumps->id.rhs_sparse = (MumpsScalar*)aa; 882 mumps->id.ICNTL(20) = 1; 883 } 884 /* handle condensation step of Schur complement (if any) */ 885 if (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2) { 886 second_solve = PETSC_TRUE; 887 ierr = MatMumpsHandleSchur_Private(A,PETSC_FALSE);CHKERRQ(ierr); 888 } 889 /* solve phase */ 890 /*-------------*/ 891 mumps->id.job = JOB_SOLVE; 892 PetscMUMPS_c(&mumps->id); 893 if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d\n",mumps->id.INFOG(1)); 894 895 /* handle expansion step of Schur complement (if any) */ 896 if (second_solve) { 897 ierr = MatMumpsHandleSchur_Private(A,PETSC_TRUE);CHKERRQ(ierr); 898 } 899 if (Bt) { 900 PetscBool done; 901 902 ierr = MatSeqAIJRestoreArray(Bt,&aa);CHKERRQ(ierr); 903 ierr = MatRestoreRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&done);CHKERRQ(ierr); 904 if (!done) SETERRQ(PetscObjectComm((PetscObject)Bt),PETSC_ERR_ARG_WRONG,"Cannot restore IJ structure"); 905 mumps->id.ICNTL(20) = 0; 906 } 907 ierr = MatDenseRestoreArray(X,&array);CHKERRQ(ierr); 908 } else { /*--------- parallel case --------*/ 909 PetscInt lsol_loc,nlsol_loc,*isol_loc,*idx,*iidx,*idxx,*isol_loc_save; 910 MumpsScalar *sol_loc,*sol_loc_save; 911 IS is_to,is_from; 912 PetscInt k,proc,j,m; 913 const PetscInt *rstart; 914 Vec v_mpi,b_seq,x_seq; 915 VecScatter scat_rhs,scat_sol; 916 917 if (mumps->size > 1 && mumps->id.ICNTL(19)) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Parallel Schur complements not yet supported from PETSc\n"); 918 919 /* create x_seq to hold local solution */ 920 isol_loc_save = mumps->id.isol_loc; /* save it for MatSovle() */ 921 sol_loc_save = mumps->id.sol_loc; 922 923 lsol_loc = mumps->id.INFO(23); 924 nlsol_loc = nrhs*lsol_loc; /* length of sol_loc */ 925 ierr = PetscMalloc2(nlsol_loc,&sol_loc,nlsol_loc,&isol_loc);CHKERRQ(ierr); 926 mumps->id.sol_loc = (MumpsScalar*)sol_loc; 927 mumps->id.isol_loc = isol_loc; 928 929 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,nlsol_loc,(PetscScalar*)sol_loc,&x_seq);CHKERRQ(ierr); 930 931 /* copy rhs matrix B into vector v_mpi */ 932 ierr = MatGetLocalSize(B,&m,NULL);CHKERRQ(ierr); 933 ierr = MatDenseGetArray(B,&bray);CHKERRQ(ierr); 934 ierr = VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),1,nrhs*m,nrhs*M,(const PetscScalar*)bray,&v_mpi);CHKERRQ(ierr); 935 ierr = MatDenseRestoreArray(B,&bray);CHKERRQ(ierr); 936 937 /* scatter v_mpi to b_seq because MUMPS only supports centralized rhs */ 938 /* idx: maps from k-th index of v_mpi to (i,j)-th global entry of B; 939 iidx: inverse of idx, will be used by scattering xx_seq -> X */ 940 ierr = PetscMalloc2(nrhs*M,&idx,nrhs*M,&iidx);CHKERRQ(ierr); 941 ierr = MatGetOwnershipRanges(B,&rstart);CHKERRQ(ierr); 942 k = 0; 943 for (proc=0; proc<mumps->size; proc++){ 944 for (j=0; j<nrhs; j++){ 945 for (i=rstart[proc]; i<rstart[proc+1]; i++){ 946 iidx[j*M + i] = k; 947 idx[k++] = j*M + i; 948 } 949 } 950 } 951 952 if (!mumps->myid) { 953 ierr = VecCreateSeq(PETSC_COMM_SELF,nrhs*M,&b_seq);CHKERRQ(ierr); 954 ierr = ISCreateGeneral(PETSC_COMM_SELF,nrhs*M,idx,PETSC_COPY_VALUES,&is_to);CHKERRQ(ierr); 955 ierr = ISCreateStride(PETSC_COMM_SELF,nrhs*M,0,1,&is_from);CHKERRQ(ierr); 956 } else { 957 ierr = VecCreateSeq(PETSC_COMM_SELF,0,&b_seq);CHKERRQ(ierr); 958 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_to);CHKERRQ(ierr); 959 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_from);CHKERRQ(ierr); 960 } 961 ierr = VecScatterCreate(v_mpi,is_from,b_seq,is_to,&scat_rhs);CHKERRQ(ierr); 962 ierr = VecScatterBegin(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 963 ierr = ISDestroy(&is_to);CHKERRQ(ierr); 964 ierr = ISDestroy(&is_from);CHKERRQ(ierr); 965 ierr = VecScatterEnd(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 966 967 if (!mumps->myid) { /* define rhs on the host */ 968 ierr = VecGetArray(b_seq,&bray);CHKERRQ(ierr); 969 mumps->id.rhs = (MumpsScalar*)bray; 970 ierr = VecRestoreArray(b_seq,&bray);CHKERRQ(ierr); 971 } 972 973 /* solve phase */ 974 /*-------------*/ 975 mumps->id.job = JOB_SOLVE; 976 PetscMUMPS_c(&mumps->id); 977 if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d\n",mumps->id.INFOG(1)); 978 979 /* scatter mumps distributed solution to petsc vector v_mpi, which shares local arrays with solution matrix X */ 980 ierr = MatDenseGetArray(X,&array);CHKERRQ(ierr); 981 ierr = VecPlaceArray(v_mpi,array);CHKERRQ(ierr); 982 983 /* create scatter scat_sol */ 984 ierr = PetscMalloc1(nlsol_loc,&idxx);CHKERRQ(ierr); 985 ierr = ISCreateStride(PETSC_COMM_SELF,nlsol_loc,0,1,&is_from);CHKERRQ(ierr); 986 for (i=0; i<lsol_loc; i++) { 987 isol_loc[i] -= 1; /* change Fortran style to C style */ 988 idxx[i] = iidx[isol_loc[i]]; 989 for (j=1; j<nrhs; j++){ 990 idxx[j*lsol_loc+i] = iidx[isol_loc[i]+j*M]; 991 } 992 } 993 ierr = ISCreateGeneral(PETSC_COMM_SELF,nlsol_loc,idxx,PETSC_COPY_VALUES,&is_to);CHKERRQ(ierr); 994 ierr = VecScatterCreate(x_seq,is_from,v_mpi,is_to,&scat_sol);CHKERRQ(ierr); 995 ierr = VecScatterBegin(scat_sol,x_seq,v_mpi,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 996 ierr = ISDestroy(&is_from);CHKERRQ(ierr); 997 ierr = ISDestroy(&is_to);CHKERRQ(ierr); 998 ierr = VecScatterEnd(scat_sol,x_seq,v_mpi,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 999 ierr = MatDenseRestoreArray(X,&array);CHKERRQ(ierr); 1000 1001 /* free spaces */ 1002 mumps->id.sol_loc = sol_loc_save; 1003 mumps->id.isol_loc = isol_loc_save; 1004 1005 ierr = PetscFree2(sol_loc,isol_loc);CHKERRQ(ierr); 1006 ierr = PetscFree2(idx,iidx);CHKERRQ(ierr); 1007 ierr = PetscFree(idxx);CHKERRQ(ierr); 1008 ierr = VecDestroy(&x_seq);CHKERRQ(ierr); 1009 ierr = VecDestroy(&v_mpi);CHKERRQ(ierr); 1010 ierr = VecDestroy(&b_seq);CHKERRQ(ierr); 1011 ierr = VecScatterDestroy(&scat_rhs);CHKERRQ(ierr); 1012 ierr = VecScatterDestroy(&scat_sol);CHKERRQ(ierr); 1013 } 1014 PetscFunctionReturn(0); 1015 } 1016 1017 #if !defined(PETSC_USE_COMPLEX) 1018 /* 1019 input: 1020 F: numeric factor 1021 output: 1022 nneg: total number of negative pivots 1023 nzero: total number of zero pivots 1024 npos: (global dimension of F) - nneg - nzero 1025 */ 1026 PetscErrorCode MatGetInertia_SBAIJMUMPS(Mat F,int *nneg,int *nzero,int *npos) 1027 { 1028 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1029 PetscErrorCode ierr; 1030 PetscMPIInt size; 1031 1032 PetscFunctionBegin; 1033 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)F),&size);CHKERRQ(ierr); 1034 /* MUMPS 4.3.1 calls ScaLAPACK when ICNTL(13)=0 (default), which does not offer the possibility to compute the inertia of a dense matrix. Set ICNTL(13)=1 to skip ScaLAPACK */ 1035 if (size > 1 && mumps->id.ICNTL(13) != 1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"ICNTL(13)=%d. -mat_mumps_icntl_13 must be set as 1 for correct global matrix inertia\n",mumps->id.INFOG(13)); 1036 1037 if (nneg) *nneg = mumps->id.INFOG(12); 1038 if (nzero || npos) { 1039 if (mumps->id.ICNTL(24) != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"-mat_mumps_icntl_24 must be set as 1 for null pivot row detection"); 1040 if (nzero) *nzero = mumps->id.INFOG(28); 1041 if (npos) *npos = F->rmap->N - (mumps->id.INFOG(12) + mumps->id.INFOG(28)); 1042 } 1043 PetscFunctionReturn(0); 1044 } 1045 #endif 1046 1047 PetscErrorCode MatFactorNumeric_MUMPS(Mat F,Mat A,const MatFactorInfo *info) 1048 { 1049 Mat_MUMPS *mumps =(Mat_MUMPS*)(F)->data; 1050 PetscErrorCode ierr; 1051 PetscBool isMPIAIJ; 1052 1053 PetscFunctionBegin; 1054 if (mumps->id.INFOG(1) < 0) { 1055 if (mumps->id.INFOG(1) == -6) { 1056 ierr = PetscInfo2(A,"MatFactorNumeric is called with singular matrix structure, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1057 } 1058 ierr = PetscInfo2(A,"MatFactorNumeric is called after analysis phase fails, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1059 PetscFunctionReturn(0); 1060 } 1061 1062 ierr = (*mumps->ConvertToTriples)(A, 1, MAT_REUSE_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);CHKERRQ(ierr); 1063 1064 /* numerical factorization phase */ 1065 /*-------------------------------*/ 1066 mumps->id.job = JOB_FACTNUMERIC; 1067 if (!mumps->id.ICNTL(18)) { /* A is centralized */ 1068 if (!mumps->myid) { 1069 mumps->id.a = (MumpsScalar*)mumps->val; 1070 } 1071 } else { 1072 mumps->id.a_loc = (MumpsScalar*)mumps->val; 1073 } 1074 PetscMUMPS_c(&mumps->id); 1075 if (mumps->id.INFOG(1) < 0) { 1076 if (A->erroriffailure) { 1077 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in numerical factorization phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2)); 1078 } else { 1079 if (mumps->id.INFOG(1) == -10) { /* numerically singular matrix */ 1080 ierr = PetscInfo2(F,"matrix is numerically singular, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1081 F->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1082 } else if (mumps->id.INFOG(1) == -13) { 1083 ierr = PetscInfo2(F,"MUMPS in numerical factorization phase: INFOG(1)=%d, cannot allocate required memory %d megabytes\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1084 F->factorerrortype = MAT_FACTOR_OUTMEMORY; 1085 } else if (mumps->id.INFOG(1) == -8 || mumps->id.INFOG(1) == -9 || (-16 < mumps->id.INFOG(1) && mumps->id.INFOG(1) < -10) ) { 1086 ierr = PetscInfo2(F,"MUMPS in numerical factorization phase: INFOG(1)=%d, INFO(2)=%d, problem with workarray \n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1087 F->factorerrortype = MAT_FACTOR_OUTMEMORY; 1088 } else { 1089 ierr = PetscInfo2(F,"MUMPS in numerical factorization phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1090 F->factorerrortype = MAT_FACTOR_OTHER; 1091 } 1092 } 1093 } 1094 if (!mumps->myid && mumps->id.ICNTL(16) > 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB," mumps->id.ICNTL(16):=%d\n",mumps->id.INFOG(16)); 1095 1096 F->assembled = PETSC_TRUE; 1097 mumps->matstruc = SAME_NONZERO_PATTERN; 1098 if (F->schur) { /* reset Schur status to unfactored */ 1099 if (mumps->id.ICNTL(19) == 1) { /* stored by rows */ 1100 mumps->id.ICNTL(19) = 2; 1101 ierr = MatTranspose(F->schur,MAT_INPLACE_MATRIX,&F->schur);CHKERRQ(ierr); 1102 } 1103 ierr = MatFactorRestoreSchurComplement(F,NULL,MAT_FACTOR_SCHUR_UNFACTORED);CHKERRQ(ierr); 1104 } 1105 1106 /* just to be sure that ICNTL(19) value returned by a call from MatMumpsGetIcntl is always consistent */ 1107 if (!mumps->sym && mumps->id.ICNTL(19) && mumps->id.ICNTL(19) != 1) mumps->id.ICNTL(19) = 3; 1108 1109 if (mumps->size > 1) { 1110 PetscInt lsol_loc; 1111 PetscScalar *sol_loc; 1112 1113 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&isMPIAIJ);CHKERRQ(ierr); 1114 1115 /* distributed solution; Create x_seq=sol_loc for repeated use */ 1116 if (mumps->x_seq) { 1117 ierr = VecScatterDestroy(&mumps->scat_sol);CHKERRQ(ierr); 1118 ierr = PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);CHKERRQ(ierr); 1119 ierr = VecDestroy(&mumps->x_seq);CHKERRQ(ierr); 1120 } 1121 lsol_loc = mumps->id.INFO(23); /* length of sol_loc */ 1122 ierr = PetscMalloc2(lsol_loc,&sol_loc,lsol_loc,&mumps->id.isol_loc);CHKERRQ(ierr); 1123 mumps->id.lsol_loc = lsol_loc; 1124 mumps->id.sol_loc = (MumpsScalar*)sol_loc; 1125 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lsol_loc,sol_loc,&mumps->x_seq);CHKERRQ(ierr); 1126 } 1127 PetscFunctionReturn(0); 1128 } 1129 1130 /* Sets MUMPS options from the options database */ 1131 PetscErrorCode PetscSetMUMPSFromOptions(Mat F, Mat A) 1132 { 1133 Mat_MUMPS *mumps = (Mat_MUMPS*)F->data; 1134 PetscErrorCode ierr; 1135 PetscInt icntl,info[40],i,ninfo=40; 1136 PetscBool flg; 1137 1138 PetscFunctionBegin; 1139 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"MUMPS Options","Mat");CHKERRQ(ierr); 1140 ierr = PetscOptionsInt("-mat_mumps_icntl_1","ICNTL(1): output stream for error messages","None",mumps->id.ICNTL(1),&icntl,&flg);CHKERRQ(ierr); 1141 if (flg) mumps->id.ICNTL(1) = icntl; 1142 ierr = PetscOptionsInt("-mat_mumps_icntl_2","ICNTL(2): output stream for diagnostic printing, statistics, and warning","None",mumps->id.ICNTL(2),&icntl,&flg);CHKERRQ(ierr); 1143 if (flg) mumps->id.ICNTL(2) = icntl; 1144 ierr = PetscOptionsInt("-mat_mumps_icntl_3","ICNTL(3): output stream for global information, collected on the host","None",mumps->id.ICNTL(3),&icntl,&flg);CHKERRQ(ierr); 1145 if (flg) mumps->id.ICNTL(3) = icntl; 1146 1147 ierr = PetscOptionsInt("-mat_mumps_icntl_4","ICNTL(4): level of printing (0 to 4)","None",mumps->id.ICNTL(4),&icntl,&flg);CHKERRQ(ierr); 1148 if (flg) mumps->id.ICNTL(4) = icntl; 1149 if (mumps->id.ICNTL(4) || PetscLogPrintInfo) mumps->id.ICNTL(3) = 6; /* resume MUMPS default id.ICNTL(3) = 6 */ 1150 1151 ierr = PetscOptionsInt("-mat_mumps_icntl_6","ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7)","None",mumps->id.ICNTL(6),&icntl,&flg);CHKERRQ(ierr); 1152 if (flg) mumps->id.ICNTL(6) = icntl; 1153 1154 ierr = PetscOptionsInt("-mat_mumps_icntl_7","ICNTL(7): computes a symmetric permutation in sequential analysis (0 to 7). 3=Scotch, 4=PORD, 5=Metis","None",mumps->id.ICNTL(7),&icntl,&flg);CHKERRQ(ierr); 1155 if (flg) { 1156 if (icntl== 1 && mumps->size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"pivot order be set by the user in PERM_IN -- not supported by the PETSc/MUMPS interface\n"); 1157 else mumps->id.ICNTL(7) = icntl; 1158 } 1159 1160 ierr = PetscOptionsInt("-mat_mumps_icntl_8","ICNTL(8): scaling strategy (-2 to 8 or 77)","None",mumps->id.ICNTL(8),&mumps->id.ICNTL(8),NULL);CHKERRQ(ierr); 1161 /* ierr = PetscOptionsInt("-mat_mumps_icntl_9","ICNTL(9): computes the solution using A or A^T","None",mumps->id.ICNTL(9),&mumps->id.ICNTL(9),NULL);CHKERRQ(ierr); handled by MatSolveTranspose_MUMPS() */ 1162 ierr = PetscOptionsInt("-mat_mumps_icntl_10","ICNTL(10): max num of refinements","None",mumps->id.ICNTL(10),&mumps->id.ICNTL(10),NULL);CHKERRQ(ierr); 1163 ierr = PetscOptionsInt("-mat_mumps_icntl_11","ICNTL(11): statistics related to an error analysis (via -ksp_view)","None",mumps->id.ICNTL(11),&mumps->id.ICNTL(11),NULL);CHKERRQ(ierr); 1164 ierr = PetscOptionsInt("-mat_mumps_icntl_12","ICNTL(12): an ordering strategy for symmetric matrices (0 to 3)","None",mumps->id.ICNTL(12),&mumps->id.ICNTL(12),NULL);CHKERRQ(ierr); 1165 ierr = PetscOptionsInt("-mat_mumps_icntl_13","ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting","None",mumps->id.ICNTL(13),&mumps->id.ICNTL(13),NULL);CHKERRQ(ierr); 1166 ierr = PetscOptionsInt("-mat_mumps_icntl_14","ICNTL(14): percentage increase in the estimated working space","None",mumps->id.ICNTL(14),&mumps->id.ICNTL(14),NULL);CHKERRQ(ierr); 1167 ierr = PetscOptionsInt("-mat_mumps_icntl_19","ICNTL(19): computes the Schur complement","None",mumps->id.ICNTL(19),&mumps->id.ICNTL(19),NULL);CHKERRQ(ierr); 1168 if (mumps->id.ICNTL(19) <= 0 || mumps->id.ICNTL(19) > 3) { /* reset any schur data (if any) */ 1169 ierr = MatDestroy(&F->schur);CHKERRQ(ierr); 1170 ierr = MatMumpsResetSchur_Private(mumps);CHKERRQ(ierr); 1171 } 1172 /* ierr = PetscOptionsInt("-mat_mumps_icntl_20","ICNTL(20): the format (dense or sparse) of the right-hand sides","None",mumps->id.ICNTL(20),&mumps->id.ICNTL(20),NULL);CHKERRQ(ierr); -- sparse rhs is not supported in PETSc API */ 1173 /* ierr = PetscOptionsInt("-mat_mumps_icntl_21","ICNTL(21): the distribution (centralized or distributed) of the solution vectors","None",mumps->id.ICNTL(21),&mumps->id.ICNTL(21),NULL);CHKERRQ(ierr); we only use distributed solution vector */ 1174 1175 ierr = PetscOptionsInt("-mat_mumps_icntl_22","ICNTL(22): in-core/out-of-core factorization and solve (0 or 1)","None",mumps->id.ICNTL(22),&mumps->id.ICNTL(22),NULL);CHKERRQ(ierr); 1176 ierr = PetscOptionsInt("-mat_mumps_icntl_23","ICNTL(23): max size of the working memory (MB) that can allocate per processor","None",mumps->id.ICNTL(23),&mumps->id.ICNTL(23),NULL);CHKERRQ(ierr); 1177 ierr = PetscOptionsInt("-mat_mumps_icntl_24","ICNTL(24): detection of null pivot rows (0 or 1)","None",mumps->id.ICNTL(24),&mumps->id.ICNTL(24),NULL);CHKERRQ(ierr); 1178 if (mumps->id.ICNTL(24)) { 1179 mumps->id.ICNTL(13) = 1; /* turn-off ScaLAPACK to help with the correct detection of null pivots */ 1180 } 1181 1182 ierr = PetscOptionsInt("-mat_mumps_icntl_25","ICNTL(25): computes a solution of a deficient matrix and a null space basis","None",mumps->id.ICNTL(25),&mumps->id.ICNTL(25),NULL);CHKERRQ(ierr); 1183 ierr = PetscOptionsInt("-mat_mumps_icntl_26","ICNTL(26): drives the solution phase if a Schur complement matrix","None",mumps->id.ICNTL(26),&mumps->id.ICNTL(26),NULL);CHKERRQ(ierr); 1184 ierr = PetscOptionsInt("-mat_mumps_icntl_27","ICNTL(27): the blocking size for multiple right-hand sides","None",mumps->id.ICNTL(27),&mumps->id.ICNTL(27),NULL);CHKERRQ(ierr); 1185 ierr = PetscOptionsInt("-mat_mumps_icntl_28","ICNTL(28): use 1 for sequential analysis and ictnl(7) ordering, or 2 for parallel analysis and ictnl(29) ordering","None",mumps->id.ICNTL(28),&mumps->id.ICNTL(28),NULL);CHKERRQ(ierr); 1186 ierr = PetscOptionsInt("-mat_mumps_icntl_29","ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis","None",mumps->id.ICNTL(29),&mumps->id.ICNTL(29),NULL);CHKERRQ(ierr); 1187 ierr = PetscOptionsInt("-mat_mumps_icntl_30","ICNTL(30): compute user-specified set of entries in inv(A)","None",mumps->id.ICNTL(30),&mumps->id.ICNTL(30),NULL);CHKERRQ(ierr); 1188 ierr = PetscOptionsInt("-mat_mumps_icntl_31","ICNTL(31): indicates which factors may be discarded during factorization","None",mumps->id.ICNTL(31),&mumps->id.ICNTL(31),NULL);CHKERRQ(ierr); 1189 /* ierr = PetscOptionsInt("-mat_mumps_icntl_32","ICNTL(32): performs the forward elemination of the right-hand sides during factorization","None",mumps->id.ICNTL(32),&mumps->id.ICNTL(32),NULL);CHKERRQ(ierr); -- not supported by PETSc API */ 1190 ierr = PetscOptionsInt("-mat_mumps_icntl_33","ICNTL(33): compute determinant","None",mumps->id.ICNTL(33),&mumps->id.ICNTL(33),NULL);CHKERRQ(ierr); 1191 ierr = PetscOptionsInt("-mat_mumps_icntl_35","ICNTL(35): activates Block Lock Rank (BLR) based factorization","None",mumps->id.ICNTL(35),&mumps->id.ICNTL(35),NULL);CHKERRQ(ierr); 1192 1193 ierr = PetscOptionsReal("-mat_mumps_cntl_1","CNTL(1): relative pivoting threshold","None",mumps->id.CNTL(1),&mumps->id.CNTL(1),NULL);CHKERRQ(ierr); 1194 ierr = PetscOptionsReal("-mat_mumps_cntl_2","CNTL(2): stopping criterion of refinement","None",mumps->id.CNTL(2),&mumps->id.CNTL(2),NULL);CHKERRQ(ierr); 1195 ierr = PetscOptionsReal("-mat_mumps_cntl_3","CNTL(3): absolute pivoting threshold","None",mumps->id.CNTL(3),&mumps->id.CNTL(3),NULL);CHKERRQ(ierr); 1196 ierr = PetscOptionsReal("-mat_mumps_cntl_4","CNTL(4): value for static pivoting","None",mumps->id.CNTL(4),&mumps->id.CNTL(4),NULL);CHKERRQ(ierr); 1197 ierr = PetscOptionsReal("-mat_mumps_cntl_5","CNTL(5): fixation for null pivots","None",mumps->id.CNTL(5),&mumps->id.CNTL(5),NULL);CHKERRQ(ierr); 1198 ierr = PetscOptionsReal("-mat_mumps_cntl_7","CNTL(7): dropping parameter used during BLR","None",mumps->id.CNTL(7),&mumps->id.CNTL(7),NULL);CHKERRQ(ierr); 1199 1200 ierr = PetscOptionsString("-mat_mumps_ooc_tmpdir", "out of core directory", "None", mumps->id.ooc_tmpdir, mumps->id.ooc_tmpdir, 256, NULL);CHKERRQ(ierr); 1201 1202 ierr = PetscOptionsIntArray("-mat_mumps_view_info","request INFO local to each processor","",info,&ninfo,NULL);CHKERRQ(ierr); 1203 if (ninfo) { 1204 if (ninfo > 40) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_USER,"number of INFO %d must <= 40\n",ninfo); 1205 ierr = PetscMalloc1(ninfo,&mumps->info);CHKERRQ(ierr); 1206 mumps->ninfo = ninfo; 1207 for (i=0; i<ninfo; i++) { 1208 if (info[i] < 0 || info[i]>40) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_USER,"index of INFO %d must between 1 and 40\n",ninfo); 1209 else mumps->info[i] = info[i]; 1210 } 1211 } 1212 1213 ierr = PetscOptionsEnd();CHKERRQ(ierr); 1214 PetscFunctionReturn(0); 1215 } 1216 1217 PetscErrorCode PetscInitializeMUMPS(Mat A,Mat_MUMPS *mumps) 1218 { 1219 PetscErrorCode ierr; 1220 1221 PetscFunctionBegin; 1222 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A), &mumps->myid);CHKERRQ(ierr); 1223 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)A),&mumps->size);CHKERRQ(ierr); 1224 ierr = MPI_Comm_dup(PetscObjectComm((PetscObject)A),&(mumps->comm_mumps));CHKERRQ(ierr); 1225 1226 mumps->id.comm_fortran = MPI_Comm_c2f(mumps->comm_mumps); 1227 1228 mumps->id.job = JOB_INIT; 1229 mumps->id.par = 1; /* host participates factorizaton and solve */ 1230 mumps->id.sym = mumps->sym; 1231 PetscMUMPS_c(&mumps->id); 1232 1233 mumps->scat_rhs = NULL; 1234 mumps->scat_sol = NULL; 1235 1236 /* set PETSc-MUMPS default options - override MUMPS default */ 1237 mumps->id.ICNTL(3) = 0; 1238 mumps->id.ICNTL(4) = 0; 1239 if (mumps->size == 1) { 1240 mumps->id.ICNTL(18) = 0; /* centralized assembled matrix input */ 1241 } else { 1242 mumps->id.ICNTL(18) = 3; /* distributed assembled matrix input */ 1243 mumps->id.ICNTL(20) = 0; /* rhs is in dense format */ 1244 mumps->id.ICNTL(21) = 1; /* distributed solution */ 1245 } 1246 1247 /* schur */ 1248 mumps->id.size_schur = 0; 1249 mumps->id.listvar_schur = NULL; 1250 mumps->id.schur = NULL; 1251 mumps->sizeredrhs = 0; 1252 mumps->schur_sol = NULL; 1253 mumps->schur_sizesol = 0; 1254 PetscFunctionReturn(0); 1255 } 1256 1257 PetscErrorCode MatFactorSymbolic_MUMPS_ReportIfError(Mat F,Mat A,const MatFactorInfo *info,Mat_MUMPS *mumps) 1258 { 1259 PetscErrorCode ierr; 1260 1261 PetscFunctionBegin; 1262 if (mumps->id.INFOG(1) < 0) { 1263 if (A->erroriffailure) { 1264 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in analysis phase: INFOG(1)=%d\n",mumps->id.INFOG(1)); 1265 } else { 1266 if (mumps->id.INFOG(1) == -6) { 1267 ierr = PetscInfo2(F,"matrix is singular in structure, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1268 F->factorerrortype = MAT_FACTOR_STRUCT_ZEROPIVOT; 1269 } else if (mumps->id.INFOG(1) == -5 || mumps->id.INFOG(1) == -7) { 1270 ierr = PetscInfo2(F,"problem of workspace, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1271 F->factorerrortype = MAT_FACTOR_OUTMEMORY; 1272 } else { 1273 ierr = PetscInfo2(F,"Error reported by MUMPS in analysis phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1274 F->factorerrortype = MAT_FACTOR_OTHER; 1275 } 1276 } 1277 } 1278 PetscFunctionReturn(0); 1279 } 1280 1281 /* Note Petsc r(=c) permutation is used when mumps->id.ICNTL(7)==1 with centralized assembled matrix input; otherwise r and c are ignored */ 1282 PetscErrorCode MatLUFactorSymbolic_AIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info) 1283 { 1284 Mat_MUMPS *mumps = (Mat_MUMPS*)F->data; 1285 PetscErrorCode ierr; 1286 Vec b; 1287 IS is_iden; 1288 const PetscInt M = A->rmap->N; 1289 1290 PetscFunctionBegin; 1291 mumps->matstruc = DIFFERENT_NONZERO_PATTERN; 1292 1293 /* Set MUMPS options from the options database */ 1294 ierr = PetscSetMUMPSFromOptions(F,A);CHKERRQ(ierr); 1295 1296 ierr = (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);CHKERRQ(ierr); 1297 1298 /* analysis phase */ 1299 /*----------------*/ 1300 mumps->id.job = JOB_FACTSYMBOLIC; 1301 mumps->id.n = M; 1302 switch (mumps->id.ICNTL(18)) { 1303 case 0: /* centralized assembled matrix input */ 1304 if (!mumps->myid) { 1305 mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn; 1306 if (mumps->id.ICNTL(6)>1) { 1307 mumps->id.a = (MumpsScalar*)mumps->val; 1308 } 1309 if (mumps->id.ICNTL(7) == 1) { /* use user-provide matrix ordering - assuming r = c ordering */ 1310 /* 1311 PetscBool flag; 1312 ierr = ISEqual(r,c,&flag);CHKERRQ(ierr); 1313 if (!flag) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"row_perm != col_perm"); 1314 ierr = ISView(r,PETSC_VIEWER_STDOUT_SELF); 1315 */ 1316 if (!mumps->myid) { 1317 const PetscInt *idx; 1318 PetscInt i,*perm_in; 1319 1320 ierr = PetscMalloc1(M,&perm_in);CHKERRQ(ierr); 1321 ierr = ISGetIndices(r,&idx);CHKERRQ(ierr); 1322 1323 mumps->id.perm_in = perm_in; 1324 for (i=0; i<M; i++) perm_in[i] = idx[i]+1; /* perm_in[]: start from 1, not 0! */ 1325 ierr = ISRestoreIndices(r,&idx);CHKERRQ(ierr); 1326 } 1327 } 1328 } 1329 break; 1330 case 3: /* distributed assembled matrix input (size>1) */ 1331 mumps->id.nz_loc = mumps->nz; 1332 mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn; 1333 if (mumps->id.ICNTL(6)>1) { 1334 mumps->id.a_loc = (MumpsScalar*)mumps->val; 1335 } 1336 /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */ 1337 if (!mumps->myid) { 1338 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->N,&mumps->b_seq);CHKERRQ(ierr); 1339 ierr = ISCreateStride(PETSC_COMM_SELF,A->rmap->N,0,1,&is_iden);CHKERRQ(ierr); 1340 } else { 1341 ierr = VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);CHKERRQ(ierr); 1342 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);CHKERRQ(ierr); 1343 } 1344 ierr = MatCreateVecs(A,NULL,&b);CHKERRQ(ierr); 1345 ierr = VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);CHKERRQ(ierr); 1346 ierr = ISDestroy(&is_iden);CHKERRQ(ierr); 1347 ierr = VecDestroy(&b);CHKERRQ(ierr); 1348 break; 1349 } 1350 PetscMUMPS_c(&mumps->id); 1351 ierr = MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);CHKERRQ(ierr); 1352 1353 F->ops->lufactornumeric = MatFactorNumeric_MUMPS; 1354 F->ops->solve = MatSolve_MUMPS; 1355 F->ops->solvetranspose = MatSolveTranspose_MUMPS; 1356 F->ops->matsolve = MatMatSolve_MUMPS; 1357 PetscFunctionReturn(0); 1358 } 1359 1360 /* Note the Petsc r and c permutations are ignored */ 1361 PetscErrorCode MatLUFactorSymbolic_BAIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info) 1362 { 1363 Mat_MUMPS *mumps = (Mat_MUMPS*)F->data; 1364 PetscErrorCode ierr; 1365 Vec b; 1366 IS is_iden; 1367 const PetscInt M = A->rmap->N; 1368 1369 PetscFunctionBegin; 1370 mumps->matstruc = DIFFERENT_NONZERO_PATTERN; 1371 1372 /* Set MUMPS options from the options database */ 1373 ierr = PetscSetMUMPSFromOptions(F,A);CHKERRQ(ierr); 1374 1375 ierr = (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);CHKERRQ(ierr); 1376 1377 /* analysis phase */ 1378 /*----------------*/ 1379 mumps->id.job = JOB_FACTSYMBOLIC; 1380 mumps->id.n = M; 1381 switch (mumps->id.ICNTL(18)) { 1382 case 0: /* centralized assembled matrix input */ 1383 if (!mumps->myid) { 1384 mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn; 1385 if (mumps->id.ICNTL(6)>1) { 1386 mumps->id.a = (MumpsScalar*)mumps->val; 1387 } 1388 } 1389 break; 1390 case 3: /* distributed assembled matrix input (size>1) */ 1391 mumps->id.nz_loc = mumps->nz; 1392 mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn; 1393 if (mumps->id.ICNTL(6)>1) { 1394 mumps->id.a_loc = (MumpsScalar*)mumps->val; 1395 } 1396 /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */ 1397 if (!mumps->myid) { 1398 ierr = VecCreateSeq(PETSC_COMM_SELF,A->cmap->N,&mumps->b_seq);CHKERRQ(ierr); 1399 ierr = ISCreateStride(PETSC_COMM_SELF,A->cmap->N,0,1,&is_iden);CHKERRQ(ierr); 1400 } else { 1401 ierr = VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);CHKERRQ(ierr); 1402 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);CHKERRQ(ierr); 1403 } 1404 ierr = MatCreateVecs(A,NULL,&b);CHKERRQ(ierr); 1405 ierr = VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);CHKERRQ(ierr); 1406 ierr = ISDestroy(&is_iden);CHKERRQ(ierr); 1407 ierr = VecDestroy(&b);CHKERRQ(ierr); 1408 break; 1409 } 1410 PetscMUMPS_c(&mumps->id); 1411 ierr = MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);CHKERRQ(ierr); 1412 1413 F->ops->lufactornumeric = MatFactorNumeric_MUMPS; 1414 F->ops->solve = MatSolve_MUMPS; 1415 F->ops->solvetranspose = MatSolveTranspose_MUMPS; 1416 PetscFunctionReturn(0); 1417 } 1418 1419 /* Note the Petsc r permutation and factor info are ignored */ 1420 PetscErrorCode MatCholeskyFactorSymbolic_MUMPS(Mat F,Mat A,IS r,const MatFactorInfo *info) 1421 { 1422 Mat_MUMPS *mumps = (Mat_MUMPS*)F->data; 1423 PetscErrorCode ierr; 1424 Vec b; 1425 IS is_iden; 1426 const PetscInt M = A->rmap->N; 1427 1428 PetscFunctionBegin; 1429 mumps->matstruc = DIFFERENT_NONZERO_PATTERN; 1430 1431 /* Set MUMPS options from the options database */ 1432 ierr = PetscSetMUMPSFromOptions(F,A);CHKERRQ(ierr); 1433 1434 ierr = (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);CHKERRQ(ierr); 1435 1436 /* analysis phase */ 1437 /*----------------*/ 1438 mumps->id.job = JOB_FACTSYMBOLIC; 1439 mumps->id.n = M; 1440 switch (mumps->id.ICNTL(18)) { 1441 case 0: /* centralized assembled matrix input */ 1442 if (!mumps->myid) { 1443 mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn; 1444 if (mumps->id.ICNTL(6)>1) { 1445 mumps->id.a = (MumpsScalar*)mumps->val; 1446 } 1447 } 1448 break; 1449 case 3: /* distributed assembled matrix input (size>1) */ 1450 mumps->id.nz_loc = mumps->nz; 1451 mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn; 1452 if (mumps->id.ICNTL(6)>1) { 1453 mumps->id.a_loc = (MumpsScalar*)mumps->val; 1454 } 1455 /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */ 1456 if (!mumps->myid) { 1457 ierr = VecCreateSeq(PETSC_COMM_SELF,A->cmap->N,&mumps->b_seq);CHKERRQ(ierr); 1458 ierr = ISCreateStride(PETSC_COMM_SELF,A->cmap->N,0,1,&is_iden);CHKERRQ(ierr); 1459 } else { 1460 ierr = VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);CHKERRQ(ierr); 1461 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);CHKERRQ(ierr); 1462 } 1463 ierr = MatCreateVecs(A,NULL,&b);CHKERRQ(ierr); 1464 ierr = VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);CHKERRQ(ierr); 1465 ierr = ISDestroy(&is_iden);CHKERRQ(ierr); 1466 ierr = VecDestroy(&b);CHKERRQ(ierr); 1467 break; 1468 } 1469 PetscMUMPS_c(&mumps->id); 1470 ierr = MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);CHKERRQ(ierr); 1471 1472 F->ops->choleskyfactornumeric = MatFactorNumeric_MUMPS; 1473 F->ops->solve = MatSolve_MUMPS; 1474 F->ops->solvetranspose = MatSolve_MUMPS; 1475 F->ops->matsolve = MatMatSolve_MUMPS; 1476 #if defined(PETSC_USE_COMPLEX) 1477 F->ops->getinertia = NULL; 1478 #else 1479 F->ops->getinertia = MatGetInertia_SBAIJMUMPS; 1480 #endif 1481 PetscFunctionReturn(0); 1482 } 1483 1484 PetscErrorCode MatView_MUMPS(Mat A,PetscViewer viewer) 1485 { 1486 PetscErrorCode ierr; 1487 PetscBool iascii; 1488 PetscViewerFormat format; 1489 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 1490 1491 PetscFunctionBegin; 1492 /* check if matrix is mumps type */ 1493 if (A->ops->solve != MatSolve_MUMPS) PetscFunctionReturn(0); 1494 1495 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1496 if (iascii) { 1497 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1498 if (format == PETSC_VIEWER_ASCII_INFO) { 1499 ierr = PetscViewerASCIIPrintf(viewer,"MUMPS run parameters:\n");CHKERRQ(ierr); 1500 ierr = PetscViewerASCIIPrintf(viewer," SYM (matrix type): %d \n",mumps->id.sym);CHKERRQ(ierr); 1501 ierr = PetscViewerASCIIPrintf(viewer," PAR (host participation): %d \n",mumps->id.par);CHKERRQ(ierr); 1502 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(1) (output for error): %d \n",mumps->id.ICNTL(1));CHKERRQ(ierr); 1503 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(2) (output of diagnostic msg): %d \n",mumps->id.ICNTL(2));CHKERRQ(ierr); 1504 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(3) (output for global info): %d \n",mumps->id.ICNTL(3));CHKERRQ(ierr); 1505 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(4) (level of printing): %d \n",mumps->id.ICNTL(4));CHKERRQ(ierr); 1506 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(5) (input mat struct): %d \n",mumps->id.ICNTL(5));CHKERRQ(ierr); 1507 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(6) (matrix prescaling): %d \n",mumps->id.ICNTL(6));CHKERRQ(ierr); 1508 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(7) (sequential matrix ordering):%d \n",mumps->id.ICNTL(7));CHKERRQ(ierr); 1509 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(8) (scaling strategy): %d \n",mumps->id.ICNTL(8));CHKERRQ(ierr); 1510 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(10) (max num of refinements): %d \n",mumps->id.ICNTL(10));CHKERRQ(ierr); 1511 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(11) (error analysis): %d \n",mumps->id.ICNTL(11));CHKERRQ(ierr); 1512 if (mumps->id.ICNTL(11)>0) { 1513 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(4) (inf norm of input mat): %g\n",mumps->id.RINFOG(4));CHKERRQ(ierr); 1514 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(5) (inf norm of solution): %g\n",mumps->id.RINFOG(5));CHKERRQ(ierr); 1515 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(6) (inf norm of residual): %g\n",mumps->id.RINFOG(6));CHKERRQ(ierr); 1516 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(7),RINFOG(8) (backward error est): %g, %g\n",mumps->id.RINFOG(7),mumps->id.RINFOG(8));CHKERRQ(ierr); 1517 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(9) (error estimate): %g \n",mumps->id.RINFOG(9));CHKERRQ(ierr); 1518 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(10),RINFOG(11)(condition numbers): %g, %g\n",mumps->id.RINFOG(10),mumps->id.RINFOG(11));CHKERRQ(ierr); 1519 } 1520 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(12) (efficiency control): %d \n",mumps->id.ICNTL(12));CHKERRQ(ierr); 1521 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(13) (efficiency control): %d \n",mumps->id.ICNTL(13));CHKERRQ(ierr); 1522 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(14) (percentage of estimated workspace increase): %d \n",mumps->id.ICNTL(14));CHKERRQ(ierr); 1523 /* ICNTL(15-17) not used */ 1524 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(18) (input mat struct): %d \n",mumps->id.ICNTL(18));CHKERRQ(ierr); 1525 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(19) (Schur complement info): %d \n",mumps->id.ICNTL(19));CHKERRQ(ierr); 1526 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(20) (rhs sparse pattern): %d \n",mumps->id.ICNTL(20));CHKERRQ(ierr); 1527 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(21) (solution struct): %d \n",mumps->id.ICNTL(21));CHKERRQ(ierr); 1528 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(22) (in-core/out-of-core facility): %d \n",mumps->id.ICNTL(22));CHKERRQ(ierr); 1529 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(23) (max size of memory can be allocated locally):%d \n",mumps->id.ICNTL(23));CHKERRQ(ierr); 1530 1531 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(24) (detection of null pivot rows): %d \n",mumps->id.ICNTL(24));CHKERRQ(ierr); 1532 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(25) (computation of a null space basis): %d \n",mumps->id.ICNTL(25));CHKERRQ(ierr); 1533 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(26) (Schur options for rhs or solution): %d \n",mumps->id.ICNTL(26));CHKERRQ(ierr); 1534 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(27) (experimental parameter): %d \n",mumps->id.ICNTL(27));CHKERRQ(ierr); 1535 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(28) (use parallel or sequential ordering): %d \n",mumps->id.ICNTL(28));CHKERRQ(ierr); 1536 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(29) (parallel ordering): %d \n",mumps->id.ICNTL(29));CHKERRQ(ierr); 1537 1538 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(30) (user-specified set of entries in inv(A)): %d \n",mumps->id.ICNTL(30));CHKERRQ(ierr); 1539 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(31) (factors is discarded in the solve phase): %d \n",mumps->id.ICNTL(31));CHKERRQ(ierr); 1540 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(33) (compute determinant): %d \n",mumps->id.ICNTL(33));CHKERRQ(ierr); 1541 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(35) (activate BLR based factorization): %d \n",mumps->id.ICNTL(35));CHKERRQ(ierr); 1542 1543 ierr = PetscViewerASCIIPrintf(viewer," CNTL(1) (relative pivoting threshold): %g \n",mumps->id.CNTL(1));CHKERRQ(ierr); 1544 ierr = PetscViewerASCIIPrintf(viewer," CNTL(2) (stopping criterion of refinement): %g \n",mumps->id.CNTL(2));CHKERRQ(ierr); 1545 ierr = PetscViewerASCIIPrintf(viewer," CNTL(3) (absolute pivoting threshold): %g \n",mumps->id.CNTL(3));CHKERRQ(ierr); 1546 ierr = PetscViewerASCIIPrintf(viewer," CNTL(4) (value of static pivoting): %g \n",mumps->id.CNTL(4));CHKERRQ(ierr); 1547 ierr = PetscViewerASCIIPrintf(viewer," CNTL(5) (fixation for null pivots): %g \n",mumps->id.CNTL(5));CHKERRQ(ierr); 1548 ierr = PetscViewerASCIIPrintf(viewer," CNTL(7) (dropping parameter for BLR): %g \n",mumps->id.CNTL(7));CHKERRQ(ierr); 1549 1550 /* infomation local to each processor */ 1551 ierr = PetscViewerASCIIPrintf(viewer, " RINFO(1) (local estimated flops for the elimination after analysis): \n");CHKERRQ(ierr); 1552 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1553 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %g \n",mumps->myid,mumps->id.RINFO(1));CHKERRQ(ierr); 1554 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1555 ierr = PetscViewerASCIIPrintf(viewer, " RINFO(2) (local estimated flops for the assembly after factorization): \n");CHKERRQ(ierr); 1556 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %g \n",mumps->myid,mumps->id.RINFO(2));CHKERRQ(ierr); 1557 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1558 ierr = PetscViewerASCIIPrintf(viewer, " RINFO(3) (local estimated flops for the elimination after factorization): \n");CHKERRQ(ierr); 1559 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %g \n",mumps->myid,mumps->id.RINFO(3));CHKERRQ(ierr); 1560 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1561 1562 ierr = PetscViewerASCIIPrintf(viewer, " INFO(15) (estimated size of (in MB) MUMPS internal data for running numerical factorization): \n");CHKERRQ(ierr); 1563 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(15));CHKERRQ(ierr); 1564 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1565 1566 ierr = PetscViewerASCIIPrintf(viewer, " INFO(16) (size of (in MB) MUMPS internal data used during numerical factorization): \n");CHKERRQ(ierr); 1567 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(16));CHKERRQ(ierr); 1568 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1569 1570 ierr = PetscViewerASCIIPrintf(viewer, " INFO(23) (num of pivots eliminated on this processor after factorization): \n");CHKERRQ(ierr); 1571 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(23));CHKERRQ(ierr); 1572 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1573 1574 if (mumps->ninfo && mumps->ninfo <= 40){ 1575 PetscInt i; 1576 for (i=0; i<mumps->ninfo; i++){ 1577 ierr = PetscViewerASCIIPrintf(viewer, " INFO(%d): \n",mumps->info[i]);CHKERRQ(ierr); 1578 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(mumps->info[i]));CHKERRQ(ierr); 1579 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1580 } 1581 } 1582 1583 1584 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1585 1586 if (!mumps->myid) { /* information from the host */ 1587 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(1) (global estimated flops for the elimination after analysis): %g \n",mumps->id.RINFOG(1));CHKERRQ(ierr); 1588 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(2) (global estimated flops for the assembly after factorization): %g \n",mumps->id.RINFOG(2));CHKERRQ(ierr); 1589 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(3) (global estimated flops for the elimination after factorization): %g \n",mumps->id.RINFOG(3));CHKERRQ(ierr); 1590 ierr = PetscViewerASCIIPrintf(viewer," (RINFOG(12) RINFOG(13))*2^INFOG(34) (determinant): (%g,%g)*(2^%d)\n",mumps->id.RINFOG(12),mumps->id.RINFOG(13),mumps->id.INFOG(34));CHKERRQ(ierr); 1591 1592 ierr = PetscViewerASCIIPrintf(viewer," INFOG(3) (estimated real workspace for factors on all processors after analysis): %d \n",mumps->id.INFOG(3));CHKERRQ(ierr); 1593 ierr = PetscViewerASCIIPrintf(viewer," INFOG(4) (estimated integer workspace for factors on all processors after analysis): %d \n",mumps->id.INFOG(4));CHKERRQ(ierr); 1594 ierr = PetscViewerASCIIPrintf(viewer," INFOG(5) (estimated maximum front size in the complete tree): %d \n",mumps->id.INFOG(5));CHKERRQ(ierr); 1595 ierr = PetscViewerASCIIPrintf(viewer," INFOG(6) (number of nodes in the complete tree): %d \n",mumps->id.INFOG(6));CHKERRQ(ierr); 1596 ierr = PetscViewerASCIIPrintf(viewer," INFOG(7) (ordering option effectively use after analysis): %d \n",mumps->id.INFOG(7));CHKERRQ(ierr); 1597 ierr = PetscViewerASCIIPrintf(viewer," INFOG(8) (structural symmetry in percent of the permuted matrix after analysis): %d \n",mumps->id.INFOG(8));CHKERRQ(ierr); 1598 ierr = PetscViewerASCIIPrintf(viewer," INFOG(9) (total real/complex workspace to store the matrix factors after factorization): %d \n",mumps->id.INFOG(9));CHKERRQ(ierr); 1599 ierr = PetscViewerASCIIPrintf(viewer," INFOG(10) (total integer space store the matrix factors after factorization): %d \n",mumps->id.INFOG(10));CHKERRQ(ierr); 1600 ierr = PetscViewerASCIIPrintf(viewer," INFOG(11) (order of largest frontal matrix after factorization): %d \n",mumps->id.INFOG(11));CHKERRQ(ierr); 1601 ierr = PetscViewerASCIIPrintf(viewer," INFOG(12) (number of off-diagonal pivots): %d \n",mumps->id.INFOG(12));CHKERRQ(ierr); 1602 ierr = PetscViewerASCIIPrintf(viewer," INFOG(13) (number of delayed pivots after factorization): %d \n",mumps->id.INFOG(13));CHKERRQ(ierr); 1603 ierr = PetscViewerASCIIPrintf(viewer," INFOG(14) (number of memory compress after factorization): %d \n",mumps->id.INFOG(14));CHKERRQ(ierr); 1604 ierr = PetscViewerASCIIPrintf(viewer," INFOG(15) (number of steps of iterative refinement after solution): %d \n",mumps->id.INFOG(15));CHKERRQ(ierr); 1605 ierr = PetscViewerASCIIPrintf(viewer," INFOG(16) (estimated size (in MB) of all MUMPS internal data for factorization after analysis: value on the most memory consuming processor): %d \n",mumps->id.INFOG(16));CHKERRQ(ierr); 1606 ierr = PetscViewerASCIIPrintf(viewer," INFOG(17) (estimated size of all MUMPS internal data for factorization after analysis: sum over all processors): %d \n",mumps->id.INFOG(17));CHKERRQ(ierr); 1607 ierr = PetscViewerASCIIPrintf(viewer," INFOG(18) (size of all MUMPS internal data allocated during factorization: value on the most memory consuming processor): %d \n",mumps->id.INFOG(18));CHKERRQ(ierr); 1608 ierr = PetscViewerASCIIPrintf(viewer," INFOG(19) (size of all MUMPS internal data allocated during factorization: sum over all processors): %d \n",mumps->id.INFOG(19));CHKERRQ(ierr); 1609 ierr = PetscViewerASCIIPrintf(viewer," INFOG(20) (estimated number of entries in the factors): %d \n",mumps->id.INFOG(20));CHKERRQ(ierr); 1610 ierr = PetscViewerASCIIPrintf(viewer," INFOG(21) (size in MB of memory effectively used during factorization - value on the most memory consuming processor): %d \n",mumps->id.INFOG(21));CHKERRQ(ierr); 1611 ierr = PetscViewerASCIIPrintf(viewer," INFOG(22) (size in MB of memory effectively used during factorization - sum over all processors): %d \n",mumps->id.INFOG(22));CHKERRQ(ierr); 1612 ierr = PetscViewerASCIIPrintf(viewer," INFOG(23) (after analysis: value of ICNTL(6) effectively used): %d \n",mumps->id.INFOG(23));CHKERRQ(ierr); 1613 ierr = PetscViewerASCIIPrintf(viewer," INFOG(24) (after analysis: value of ICNTL(12) effectively used): %d \n",mumps->id.INFOG(24));CHKERRQ(ierr); 1614 ierr = PetscViewerASCIIPrintf(viewer," INFOG(25) (after factorization: number of pivots modified by static pivoting): %d \n",mumps->id.INFOG(25));CHKERRQ(ierr); 1615 ierr = PetscViewerASCIIPrintf(viewer," INFOG(28) (after factorization: number of null pivots encountered): %d\n",mumps->id.INFOG(28));CHKERRQ(ierr); 1616 ierr = PetscViewerASCIIPrintf(viewer," INFOG(29) (after factorization: effective number of entries in the factors (sum over all processors)): %d\n",mumps->id.INFOG(29));CHKERRQ(ierr); 1617 ierr = PetscViewerASCIIPrintf(viewer," INFOG(30, 31) (after solution: size in Mbytes of memory used during solution phase): %d, %d\n",mumps->id.INFOG(30),mumps->id.INFOG(31));CHKERRQ(ierr); 1618 ierr = PetscViewerASCIIPrintf(viewer," INFOG(32) (after analysis: type of analysis done): %d\n",mumps->id.INFOG(32));CHKERRQ(ierr); 1619 ierr = PetscViewerASCIIPrintf(viewer," INFOG(33) (value used for ICNTL(8)): %d\n",mumps->id.INFOG(33));CHKERRQ(ierr); 1620 ierr = PetscViewerASCIIPrintf(viewer," INFOG(34) (exponent of the determinant if determinant is requested): %d\n",mumps->id.INFOG(34));CHKERRQ(ierr); 1621 } 1622 } 1623 } 1624 PetscFunctionReturn(0); 1625 } 1626 1627 PetscErrorCode MatGetInfo_MUMPS(Mat A,MatInfoType flag,MatInfo *info) 1628 { 1629 Mat_MUMPS *mumps =(Mat_MUMPS*)A->data; 1630 1631 PetscFunctionBegin; 1632 info->block_size = 1.0; 1633 info->nz_allocated = mumps->id.INFOG(20); 1634 info->nz_used = mumps->id.INFOG(20); 1635 info->nz_unneeded = 0.0; 1636 info->assemblies = 0.0; 1637 info->mallocs = 0.0; 1638 info->memory = 0.0; 1639 info->fill_ratio_given = 0; 1640 info->fill_ratio_needed = 0; 1641 info->factor_mallocs = 0; 1642 PetscFunctionReturn(0); 1643 } 1644 1645 /* -------------------------------------------------------------------------------------------*/ 1646 PetscErrorCode MatFactorSetSchurIS_MUMPS(Mat F, IS is) 1647 { 1648 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1649 const PetscInt *idxs; 1650 PetscInt size,i; 1651 PetscErrorCode ierr; 1652 1653 PetscFunctionBegin; 1654 ierr = ISGetLocalSize(is,&size);CHKERRQ(ierr); 1655 if (mumps->size > 1) { 1656 PetscBool ls,gs; 1657 1658 ls = mumps->myid ? (size ? PETSC_FALSE : PETSC_TRUE) : PETSC_TRUE; 1659 ierr = MPI_Allreduce(&ls,&gs,1,MPIU_BOOL,MPI_LAND,mumps->comm_mumps);CHKERRQ(ierr); 1660 if (!gs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MUMPS distributed parallel Schur complements not yet supported from PETSc\n"); 1661 } 1662 if (mumps->id.size_schur != size) { 1663 ierr = PetscFree2(mumps->id.listvar_schur,mumps->id.schur);CHKERRQ(ierr); 1664 mumps->id.size_schur = size; 1665 mumps->id.schur_lld = size; 1666 ierr = PetscMalloc2(size,&mumps->id.listvar_schur,size*size,&mumps->id.schur);CHKERRQ(ierr); 1667 } 1668 1669 /* Schur complement matrix */ 1670 ierr = MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.size_schur,(PetscScalar*)mumps->id.schur,&F->schur);CHKERRQ(ierr); 1671 if (mumps->sym == 1) { 1672 ierr = MatSetOption(F->schur,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 1673 } 1674 1675 /* MUMPS expects Fortran style indices */ 1676 ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 1677 ierr = PetscMemcpy(mumps->id.listvar_schur,idxs,size*sizeof(PetscInt));CHKERRQ(ierr); 1678 for (i=0;i<size;i++) mumps->id.listvar_schur[i]++; 1679 ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 1680 if (mumps->size > 1) { 1681 mumps->id.ICNTL(19) = 1; /* MUMPS returns Schur centralized on the host */ 1682 } else { 1683 if (F->factortype == MAT_FACTOR_LU) { 1684 mumps->id.ICNTL(19) = 3; /* MUMPS returns full matrix */ 1685 } else { 1686 mumps->id.ICNTL(19) = 2; /* MUMPS returns lower triangular part */ 1687 } 1688 } 1689 /* set a special value of ICNTL (not handled my MUMPS) to be used in the solve phase by PETSc */ 1690 mumps->id.ICNTL(26) = -1; 1691 PetscFunctionReturn(0); 1692 } 1693 1694 /* -------------------------------------------------------------------------------------------*/ 1695 PetscErrorCode MatFactorCreateSchurComplement_MUMPS(Mat F,Mat* S) 1696 { 1697 Mat St; 1698 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1699 PetscScalar *array; 1700 #if defined(PETSC_USE_COMPLEX) 1701 PetscScalar im = PetscSqrtScalar((PetscScalar)-1.0); 1702 #endif 1703 PetscErrorCode ierr; 1704 1705 PetscFunctionBegin; 1706 if (!mumps->id.ICNTL(19)) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ORDER,"Schur complement mode not selected! You should call MatFactorSetSchurIS to enable it"); 1707 ierr = MatCreate(PETSC_COMM_SELF,&St);CHKERRQ(ierr); 1708 ierr = MatSetSizes(St,PETSC_DECIDE,PETSC_DECIDE,mumps->id.size_schur,mumps->id.size_schur);CHKERRQ(ierr); 1709 ierr = MatSetType(St,MATDENSE);CHKERRQ(ierr); 1710 ierr = MatSetUp(St);CHKERRQ(ierr); 1711 ierr = MatDenseGetArray(St,&array);CHKERRQ(ierr); 1712 if (!mumps->sym) { /* MUMPS always return a full matrix */ 1713 if (mumps->id.ICNTL(19) == 1) { /* stored by rows */ 1714 PetscInt i,j,N=mumps->id.size_schur; 1715 for (i=0;i<N;i++) { 1716 for (j=0;j<N;j++) { 1717 #if !defined(PETSC_USE_COMPLEX) 1718 PetscScalar val = mumps->id.schur[i*N+j]; 1719 #else 1720 PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i; 1721 #endif 1722 array[j*N+i] = val; 1723 } 1724 } 1725 } else { /* stored by columns */ 1726 ierr = PetscMemcpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur*sizeof(PetscScalar));CHKERRQ(ierr); 1727 } 1728 } else { /* either full or lower-triangular (not packed) */ 1729 if (mumps->id.ICNTL(19) == 2) { /* lower triangular stored by columns */ 1730 PetscInt i,j,N=mumps->id.size_schur; 1731 for (i=0;i<N;i++) { 1732 for (j=i;j<N;j++) { 1733 #if !defined(PETSC_USE_COMPLEX) 1734 PetscScalar val = mumps->id.schur[i*N+j]; 1735 #else 1736 PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i; 1737 #endif 1738 array[i*N+j] = val; 1739 array[j*N+i] = val; 1740 } 1741 } 1742 } else if (mumps->id.ICNTL(19) == 3) { /* full matrix */ 1743 ierr = PetscMemcpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur*sizeof(PetscScalar));CHKERRQ(ierr); 1744 } else { /* ICNTL(19) == 1 lower triangular stored by rows */ 1745 PetscInt i,j,N=mumps->id.size_schur; 1746 for (i=0;i<N;i++) { 1747 for (j=0;j<i+1;j++) { 1748 #if !defined(PETSC_USE_COMPLEX) 1749 PetscScalar val = mumps->id.schur[i*N+j]; 1750 #else 1751 PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i; 1752 #endif 1753 array[i*N+j] = val; 1754 array[j*N+i] = val; 1755 } 1756 } 1757 } 1758 } 1759 ierr = MatDenseRestoreArray(St,&array);CHKERRQ(ierr); 1760 *S = St; 1761 PetscFunctionReturn(0); 1762 } 1763 1764 /* -------------------------------------------------------------------------------------------*/ 1765 PetscErrorCode MatMumpsSetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt ival) 1766 { 1767 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1768 1769 PetscFunctionBegin; 1770 mumps->id.ICNTL(icntl) = ival; 1771 PetscFunctionReturn(0); 1772 } 1773 1774 PetscErrorCode MatMumpsGetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt *ival) 1775 { 1776 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1777 1778 PetscFunctionBegin; 1779 *ival = mumps->id.ICNTL(icntl); 1780 PetscFunctionReturn(0); 1781 } 1782 1783 /*@ 1784 MatMumpsSetIcntl - Set MUMPS parameter ICNTL() 1785 1786 Logically Collective on Mat 1787 1788 Input Parameters: 1789 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 1790 . icntl - index of MUMPS parameter array ICNTL() 1791 - ival - value of MUMPS ICNTL(icntl) 1792 1793 Options Database: 1794 . -mat_mumps_icntl_<icntl> <ival> 1795 1796 Level: beginner 1797 1798 References: 1799 . MUMPS Users' Guide 1800 1801 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 1802 @*/ 1803 PetscErrorCode MatMumpsSetIcntl(Mat F,PetscInt icntl,PetscInt ival) 1804 { 1805 PetscErrorCode ierr; 1806 1807 PetscFunctionBegin; 1808 PetscValidType(F,1); 1809 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 1810 PetscValidLogicalCollectiveInt(F,icntl,2); 1811 PetscValidLogicalCollectiveInt(F,ival,3); 1812 ierr = PetscTryMethod(F,"MatMumpsSetIcntl_C",(Mat,PetscInt,PetscInt),(F,icntl,ival));CHKERRQ(ierr); 1813 PetscFunctionReturn(0); 1814 } 1815 1816 /*@ 1817 MatMumpsGetIcntl - Get MUMPS parameter ICNTL() 1818 1819 Logically Collective on Mat 1820 1821 Input Parameters: 1822 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 1823 - icntl - index of MUMPS parameter array ICNTL() 1824 1825 Output Parameter: 1826 . ival - value of MUMPS ICNTL(icntl) 1827 1828 Level: beginner 1829 1830 References: 1831 . MUMPS Users' Guide 1832 1833 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 1834 @*/ 1835 PetscErrorCode MatMumpsGetIcntl(Mat F,PetscInt icntl,PetscInt *ival) 1836 { 1837 PetscErrorCode ierr; 1838 1839 PetscFunctionBegin; 1840 PetscValidType(F,1); 1841 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 1842 PetscValidLogicalCollectiveInt(F,icntl,2); 1843 PetscValidIntPointer(ival,3); 1844 ierr = PetscUseMethod(F,"MatMumpsGetIcntl_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));CHKERRQ(ierr); 1845 PetscFunctionReturn(0); 1846 } 1847 1848 /* -------------------------------------------------------------------------------------------*/ 1849 PetscErrorCode MatMumpsSetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal val) 1850 { 1851 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1852 1853 PetscFunctionBegin; 1854 mumps->id.CNTL(icntl) = val; 1855 PetscFunctionReturn(0); 1856 } 1857 1858 PetscErrorCode MatMumpsGetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal *val) 1859 { 1860 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1861 1862 PetscFunctionBegin; 1863 *val = mumps->id.CNTL(icntl); 1864 PetscFunctionReturn(0); 1865 } 1866 1867 /*@ 1868 MatMumpsSetCntl - Set MUMPS parameter CNTL() 1869 1870 Logically Collective on Mat 1871 1872 Input Parameters: 1873 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 1874 . icntl - index of MUMPS parameter array CNTL() 1875 - val - value of MUMPS CNTL(icntl) 1876 1877 Options Database: 1878 . -mat_mumps_cntl_<icntl> <val> 1879 1880 Level: beginner 1881 1882 References: 1883 . MUMPS Users' Guide 1884 1885 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 1886 @*/ 1887 PetscErrorCode MatMumpsSetCntl(Mat F,PetscInt icntl,PetscReal val) 1888 { 1889 PetscErrorCode ierr; 1890 1891 PetscFunctionBegin; 1892 PetscValidType(F,1); 1893 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 1894 PetscValidLogicalCollectiveInt(F,icntl,2); 1895 PetscValidLogicalCollectiveReal(F,val,3); 1896 ierr = PetscTryMethod(F,"MatMumpsSetCntl_C",(Mat,PetscInt,PetscReal),(F,icntl,val));CHKERRQ(ierr); 1897 PetscFunctionReturn(0); 1898 } 1899 1900 /*@ 1901 MatMumpsGetCntl - Get MUMPS parameter CNTL() 1902 1903 Logically Collective on Mat 1904 1905 Input Parameters: 1906 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 1907 - icntl - index of MUMPS parameter array CNTL() 1908 1909 Output Parameter: 1910 . val - value of MUMPS CNTL(icntl) 1911 1912 Level: beginner 1913 1914 References: 1915 . MUMPS Users' Guide 1916 1917 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 1918 @*/ 1919 PetscErrorCode MatMumpsGetCntl(Mat F,PetscInt icntl,PetscReal *val) 1920 { 1921 PetscErrorCode ierr; 1922 1923 PetscFunctionBegin; 1924 PetscValidType(F,1); 1925 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 1926 PetscValidLogicalCollectiveInt(F,icntl,2); 1927 PetscValidRealPointer(val,3); 1928 ierr = PetscUseMethod(F,"MatMumpsGetCntl_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));CHKERRQ(ierr); 1929 PetscFunctionReturn(0); 1930 } 1931 1932 PetscErrorCode MatMumpsGetInfo_MUMPS(Mat F,PetscInt icntl,PetscInt *info) 1933 { 1934 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1935 1936 PetscFunctionBegin; 1937 *info = mumps->id.INFO(icntl); 1938 PetscFunctionReturn(0); 1939 } 1940 1941 PetscErrorCode MatMumpsGetInfog_MUMPS(Mat F,PetscInt icntl,PetscInt *infog) 1942 { 1943 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1944 1945 PetscFunctionBegin; 1946 *infog = mumps->id.INFOG(icntl); 1947 PetscFunctionReturn(0); 1948 } 1949 1950 PetscErrorCode MatMumpsGetRinfo_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfo) 1951 { 1952 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1953 1954 PetscFunctionBegin; 1955 *rinfo = mumps->id.RINFO(icntl); 1956 PetscFunctionReturn(0); 1957 } 1958 1959 PetscErrorCode MatMumpsGetRinfog_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfog) 1960 { 1961 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1962 1963 PetscFunctionBegin; 1964 *rinfog = mumps->id.RINFOG(icntl); 1965 PetscFunctionReturn(0); 1966 } 1967 1968 PetscErrorCode MatMumpsGetMatInverse_MUMPS(Mat F,Mat spRHS) 1969 { 1970 PetscErrorCode ierr; 1971 Mat Bt = NULL; 1972 PetscBool flgT; 1973 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1974 PetscBool done; 1975 PetscScalar *aa; 1976 PetscInt spnr,*ia,*ja; 1977 1978 PetscFunctionBegin; 1979 if (mumps->size > 1) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_SUP,"No support for parallel runs yet"); 1980 ierr = PetscObjectTypeCompare((PetscObject)spRHS,MATTRANSPOSEMAT,&flgT);CHKERRQ(ierr); 1981 if (flgT) { 1982 ierr = MatTransposeGetMat(spRHS,&Bt);CHKERRQ(ierr); 1983 } else { 1984 SETERRQ(PetscObjectComm((PetscObject)spRHS),PETSC_ERR_ARG_WRONG,"Matrix spRHS must be type MATTRANSPOSEMAT matrix"); 1985 } 1986 1987 ierr = MatMumpsSetIcntl(F,30,1);CHKERRQ(ierr); 1988 1989 ierr = MatSeqAIJGetArray(Bt,&aa);CHKERRQ(ierr); 1990 ierr = MatGetRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&done);CHKERRQ(ierr); 1991 if (!done) SETERRQ(PetscObjectComm((PetscObject)Bt),PETSC_ERR_ARG_WRONG,"Cannot get IJ structure"); 1992 1993 mumps->id.irhs_ptr = ia; 1994 mumps->id.irhs_sparse = ja; 1995 mumps->id.nz_rhs = ia[spnr] - 1; 1996 mumps->id.rhs_sparse = (MumpsScalar*)aa; 1997 mumps->id.ICNTL(20) = 1; /* rhs is sparse */ 1998 1999 /* solve phase */ 2000 /*-------------*/ 2001 mumps->id.job = JOB_SOLVE; 2002 PetscMUMPS_c(&mumps->id); 2003 if (mumps->id.INFOG(1) < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d INFOG(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFOG(2)); 2004 PetscFunctionReturn(0); 2005 } 2006 2007 /*@ 2008 MatMumpsGetMatInverse - Get user-specified set of entries in inverse of A 2009 2010 Logically Collective on Mat 2011 2012 Input Parameters: 2013 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2014 - spRHS - sequential sparse matrix in MATTRANSPOSEMAT format holding specified indices 2015 2016 Output Parameter: 2017 . - spRHS - requested entries of inverse of A 2018 2019 Level: beginner 2020 2021 References: 2022 . MUMPS Users' Guide 2023 2024 .seealso: MatGetFactor(), MatCreateTranspose() 2025 @*/ 2026 PetscErrorCode MatMumpsGetMatInverse(Mat F,Mat spRHS) 2027 { 2028 PetscErrorCode ierr; 2029 2030 PetscFunctionBegin; 2031 PetscValidType(F,1); 2032 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2033 PetscValidIntPointer(spRHS,2); 2034 ierr = PetscUseMethod(F,"MatMumpsGetMatInverse_C",(Mat,Mat),(F,spRHS));CHKERRQ(ierr); 2035 PetscFunctionReturn(0); 2036 } 2037 2038 /*@ 2039 MatMumpsGetInfo - Get MUMPS parameter INFO() 2040 2041 Logically Collective on Mat 2042 2043 Input Parameters: 2044 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2045 - icntl - index of MUMPS parameter array INFO() 2046 2047 Output Parameter: 2048 . ival - value of MUMPS INFO(icntl) 2049 2050 Level: beginner 2051 2052 References: 2053 . MUMPS Users' Guide 2054 2055 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 2056 @*/ 2057 PetscErrorCode MatMumpsGetInfo(Mat F,PetscInt icntl,PetscInt *ival) 2058 { 2059 PetscErrorCode ierr; 2060 2061 PetscFunctionBegin; 2062 PetscValidType(F,1); 2063 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2064 PetscValidIntPointer(ival,3); 2065 ierr = PetscUseMethod(F,"MatMumpsGetInfo_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));CHKERRQ(ierr); 2066 PetscFunctionReturn(0); 2067 } 2068 2069 /*@ 2070 MatMumpsGetInfog - Get MUMPS parameter INFOG() 2071 2072 Logically Collective on Mat 2073 2074 Input Parameters: 2075 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2076 - icntl - index of MUMPS parameter array INFOG() 2077 2078 Output Parameter: 2079 . ival - value of MUMPS INFOG(icntl) 2080 2081 Level: beginner 2082 2083 References: 2084 . MUMPS Users' Guide 2085 2086 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 2087 @*/ 2088 PetscErrorCode MatMumpsGetInfog(Mat F,PetscInt icntl,PetscInt *ival) 2089 { 2090 PetscErrorCode ierr; 2091 2092 PetscFunctionBegin; 2093 PetscValidType(F,1); 2094 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2095 PetscValidIntPointer(ival,3); 2096 ierr = PetscUseMethod(F,"MatMumpsGetInfog_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));CHKERRQ(ierr); 2097 PetscFunctionReturn(0); 2098 } 2099 2100 /*@ 2101 MatMumpsGetRinfo - Get MUMPS parameter RINFO() 2102 2103 Logically Collective on Mat 2104 2105 Input Parameters: 2106 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2107 - icntl - index of MUMPS parameter array RINFO() 2108 2109 Output Parameter: 2110 . val - value of MUMPS RINFO(icntl) 2111 2112 Level: beginner 2113 2114 References: 2115 . MUMPS Users' Guide 2116 2117 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 2118 @*/ 2119 PetscErrorCode MatMumpsGetRinfo(Mat F,PetscInt icntl,PetscReal *val) 2120 { 2121 PetscErrorCode ierr; 2122 2123 PetscFunctionBegin; 2124 PetscValidType(F,1); 2125 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2126 PetscValidRealPointer(val,3); 2127 ierr = PetscUseMethod(F,"MatMumpsGetRinfo_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));CHKERRQ(ierr); 2128 PetscFunctionReturn(0); 2129 } 2130 2131 /*@ 2132 MatMumpsGetRinfog - Get MUMPS parameter RINFOG() 2133 2134 Logically Collective on Mat 2135 2136 Input Parameters: 2137 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2138 - icntl - index of MUMPS parameter array RINFOG() 2139 2140 Output Parameter: 2141 . val - value of MUMPS RINFOG(icntl) 2142 2143 Level: beginner 2144 2145 References: 2146 . MUMPS Users' Guide 2147 2148 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 2149 @*/ 2150 PetscErrorCode MatMumpsGetRinfog(Mat F,PetscInt icntl,PetscReal *val) 2151 { 2152 PetscErrorCode ierr; 2153 2154 PetscFunctionBegin; 2155 PetscValidType(F,1); 2156 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2157 PetscValidRealPointer(val,3); 2158 ierr = PetscUseMethod(F,"MatMumpsGetRinfog_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));CHKERRQ(ierr); 2159 PetscFunctionReturn(0); 2160 } 2161 2162 /*MC 2163 MATSOLVERMUMPS - A matrix type providing direct solvers (LU and Cholesky) for 2164 distributed and sequential matrices via the external package MUMPS. 2165 2166 Works with MATAIJ and MATSBAIJ matrices 2167 2168 Use ./configure --download-mumps --download-scalapack --download-parmetis --download-metis --download-ptscotch to have PETSc installed with MUMPS 2169 2170 Use -pc_type cholesky or lu -pc_factor_mat_solver_type mumps to use this direct solver 2171 2172 Options Database Keys: 2173 + -mat_mumps_icntl_1 - ICNTL(1): output stream for error messages 2174 . -mat_mumps_icntl_2 - ICNTL(2): output stream for diagnostic printing, statistics, and warning 2175 . -mat_mumps_icntl_3 - ICNTL(3): output stream for global information, collected on the host 2176 . -mat_mumps_icntl_4 - ICNTL(4): level of printing (0 to 4) 2177 . -mat_mumps_icntl_6 - ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7) 2178 . -mat_mumps_icntl_7 - ICNTL(7): computes a symmetric permutation in sequential analysis (0 to 7). 3=Scotch, 4=PORD, 5=Metis 2179 . -mat_mumps_icntl_8 - ICNTL(8): scaling strategy (-2 to 8 or 77) 2180 . -mat_mumps_icntl_10 - ICNTL(10): max num of refinements 2181 . -mat_mumps_icntl_11 - ICNTL(11): statistics related to an error analysis (via -ksp_view) 2182 . -mat_mumps_icntl_12 - ICNTL(12): an ordering strategy for symmetric matrices (0 to 3) 2183 . -mat_mumps_icntl_13 - ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting 2184 . -mat_mumps_icntl_14 - ICNTL(14): percentage increase in the estimated working space 2185 . -mat_mumps_icntl_19 - ICNTL(19): computes the Schur complement 2186 . -mat_mumps_icntl_22 - ICNTL(22): in-core/out-of-core factorization and solve (0 or 1) 2187 . -mat_mumps_icntl_23 - ICNTL(23): max size of the working memory (MB) that can allocate per processor 2188 . -mat_mumps_icntl_24 - ICNTL(24): detection of null pivot rows (0 or 1) 2189 . -mat_mumps_icntl_25 - ICNTL(25): compute a solution of a deficient matrix and a null space basis 2190 . -mat_mumps_icntl_26 - ICNTL(26): drives the solution phase if a Schur complement matrix 2191 . -mat_mumps_icntl_28 - ICNTL(28): use 1 for sequential analysis and ictnl(7) ordering, or 2 for parallel analysis and ictnl(29) ordering 2192 . -mat_mumps_icntl_29 - ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis 2193 . -mat_mumps_icntl_30 - ICNTL(30): compute user-specified set of entries in inv(A) 2194 . -mat_mumps_icntl_31 - ICNTL(31): indicates which factors may be discarded during factorization 2195 . -mat_mumps_icntl_33 - ICNTL(33): compute determinant 2196 . -mat_mumps_cntl_1 - CNTL(1): relative pivoting threshold 2197 . -mat_mumps_cntl_2 - CNTL(2): stopping criterion of refinement 2198 . -mat_mumps_cntl_3 - CNTL(3): absolute pivoting threshold 2199 . -mat_mumps_cntl_4 - CNTL(4): value for static pivoting 2200 - -mat_mumps_cntl_5 - CNTL(5): fixation for null pivots 2201 2202 Level: beginner 2203 2204 Notes: When a MUMPS factorization fails inside a KSP solve, for example with a KSP_DIVERGED_PCSETUP_FAILED, one can find the MUMPS information about the failure by calling 2205 $ KSPGetPC(ksp,&pc); 2206 $ PCFactorGetMatrix(pc,&mat); 2207 $ MatMumpsGetInfo(mat,....); 2208 $ MatMumpsGetInfog(mat,....); etc. 2209 Or you can run with -ksp_error_if_not_converged and the program will be stopped and the information printed in the error message. 2210 2211 .seealso: PCFactorSetMatSolverType(), MatSolverType, MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog(), KSPGetPC(), PCGetFactor(), PCFactorGetMatrix() 2212 2213 M*/ 2214 2215 static PetscErrorCode MatFactorGetSolverType_mumps(Mat A,MatSolverType *type) 2216 { 2217 PetscFunctionBegin; 2218 *type = MATSOLVERMUMPS; 2219 PetscFunctionReturn(0); 2220 } 2221 2222 /* MatGetFactor for Seq and MPI AIJ matrices */ 2223 static PetscErrorCode MatGetFactor_aij_mumps(Mat A,MatFactorType ftype,Mat *F) 2224 { 2225 Mat B; 2226 PetscErrorCode ierr; 2227 Mat_MUMPS *mumps; 2228 PetscBool isSeqAIJ; 2229 2230 PetscFunctionBegin; 2231 /* Create the factorization matrix */ 2232 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isSeqAIJ);CHKERRQ(ierr); 2233 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2234 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 2235 ierr = PetscStrallocpy("mumps",&((PetscObject)B)->type_name);CHKERRQ(ierr); 2236 ierr = MatSetUp(B);CHKERRQ(ierr); 2237 2238 ierr = PetscNewLog(B,&mumps);CHKERRQ(ierr); 2239 2240 B->ops->view = MatView_MUMPS; 2241 B->ops->getinfo = MatGetInfo_MUMPS; 2242 2243 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);CHKERRQ(ierr); 2244 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);CHKERRQ(ierr); 2245 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);CHKERRQ(ierr); 2246 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);CHKERRQ(ierr); 2247 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);CHKERRQ(ierr); 2248 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);CHKERRQ(ierr); 2249 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);CHKERRQ(ierr); 2250 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);CHKERRQ(ierr); 2251 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);CHKERRQ(ierr); 2252 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);CHKERRQ(ierr); 2253 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);CHKERRQ(ierr); 2254 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetMatInverse_C",MatMumpsGetMatInverse_MUMPS);CHKERRQ(ierr); 2255 2256 if (ftype == MAT_FACTOR_LU) { 2257 B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS; 2258 B->factortype = MAT_FACTOR_LU; 2259 if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqaij; 2260 else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpiaij; 2261 mumps->sym = 0; 2262 } else { 2263 B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS; 2264 B->factortype = MAT_FACTOR_CHOLESKY; 2265 if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqsbaij; 2266 else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpisbaij; 2267 #if defined(PETSC_USE_COMPLEX) 2268 mumps->sym = 2; 2269 #else 2270 if (A->spd_set && A->spd) mumps->sym = 1; 2271 else mumps->sym = 2; 2272 #endif 2273 } 2274 2275 /* set solvertype */ 2276 ierr = PetscFree(B->solvertype);CHKERRQ(ierr); 2277 ierr = PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);CHKERRQ(ierr); 2278 2279 mumps->isAIJ = PETSC_TRUE; 2280 B->ops->destroy = MatDestroy_MUMPS; 2281 B->data = (void*)mumps; 2282 2283 ierr = PetscInitializeMUMPS(A,mumps);CHKERRQ(ierr); 2284 2285 *F = B; 2286 PetscFunctionReturn(0); 2287 } 2288 2289 /* MatGetFactor for Seq and MPI SBAIJ matrices */ 2290 static PetscErrorCode MatGetFactor_sbaij_mumps(Mat A,MatFactorType ftype,Mat *F) 2291 { 2292 Mat B; 2293 PetscErrorCode ierr; 2294 Mat_MUMPS *mumps; 2295 PetscBool isSeqSBAIJ; 2296 2297 PetscFunctionBegin; 2298 if (ftype != MAT_FACTOR_CHOLESKY) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Cannot use PETSc SBAIJ matrices with MUMPS LU, use AIJ matrix"); 2299 if (A->rmap->bs > 1) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Cannot use PETSc SBAIJ matrices with block size > 1 with MUMPS Cholesky, use AIJ matrix instead"); 2300 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQSBAIJ,&isSeqSBAIJ);CHKERRQ(ierr); 2301 /* Create the factorization matrix */ 2302 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2303 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 2304 ierr = PetscStrallocpy("mumps",&((PetscObject)B)->type_name);CHKERRQ(ierr); 2305 ierr = MatSetUp(B);CHKERRQ(ierr); 2306 2307 ierr = PetscNewLog(B,&mumps);CHKERRQ(ierr); 2308 if (isSeqSBAIJ) { 2309 mumps->ConvertToTriples = MatConvertToTriples_seqsbaij_seqsbaij; 2310 } else { 2311 mumps->ConvertToTriples = MatConvertToTriples_mpisbaij_mpisbaij; 2312 } 2313 2314 B->ops->getinfo = MatGetInfo_External; 2315 B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS; 2316 B->ops->view = MatView_MUMPS; 2317 2318 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);CHKERRQ(ierr); 2319 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);CHKERRQ(ierr); 2320 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);CHKERRQ(ierr); 2321 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);CHKERRQ(ierr); 2322 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);CHKERRQ(ierr); 2323 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);CHKERRQ(ierr); 2324 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);CHKERRQ(ierr); 2325 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);CHKERRQ(ierr); 2326 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);CHKERRQ(ierr); 2327 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);CHKERRQ(ierr); 2328 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);CHKERRQ(ierr); 2329 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetMatInverse_C",MatMumpsGetMatInverse_MUMPS);CHKERRQ(ierr); 2330 2331 B->factortype = MAT_FACTOR_CHOLESKY; 2332 #if defined(PETSC_USE_COMPLEX) 2333 mumps->sym = 2; 2334 #else 2335 if (A->spd_set && A->spd) mumps->sym = 1; 2336 else mumps->sym = 2; 2337 #endif 2338 2339 /* set solvertype */ 2340 ierr = PetscFree(B->solvertype);CHKERRQ(ierr); 2341 ierr = PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);CHKERRQ(ierr); 2342 2343 mumps->isAIJ = PETSC_FALSE; 2344 B->ops->destroy = MatDestroy_MUMPS; 2345 B->data = (void*)mumps; 2346 2347 ierr = PetscInitializeMUMPS(A,mumps);CHKERRQ(ierr); 2348 2349 *F = B; 2350 PetscFunctionReturn(0); 2351 } 2352 2353 static PetscErrorCode MatGetFactor_baij_mumps(Mat A,MatFactorType ftype,Mat *F) 2354 { 2355 Mat B; 2356 PetscErrorCode ierr; 2357 Mat_MUMPS *mumps; 2358 PetscBool isSeqBAIJ; 2359 2360 PetscFunctionBegin; 2361 /* Create the factorization matrix */ 2362 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQBAIJ,&isSeqBAIJ);CHKERRQ(ierr); 2363 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2364 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 2365 ierr = PetscStrallocpy("mumps",&((PetscObject)B)->type_name);CHKERRQ(ierr); 2366 ierr = MatSetUp(B);CHKERRQ(ierr); 2367 2368 ierr = PetscNewLog(B,&mumps);CHKERRQ(ierr); 2369 if (ftype == MAT_FACTOR_LU) { 2370 B->ops->lufactorsymbolic = MatLUFactorSymbolic_BAIJMUMPS; 2371 B->factortype = MAT_FACTOR_LU; 2372 if (isSeqBAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqbaij_seqaij; 2373 else mumps->ConvertToTriples = MatConvertToTriples_mpibaij_mpiaij; 2374 mumps->sym = 0; 2375 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot use PETSc BAIJ matrices with MUMPS Cholesky, use SBAIJ or AIJ matrix instead\n"); 2376 2377 B->ops->getinfo = MatGetInfo_External; 2378 B->ops->view = MatView_MUMPS; 2379 2380 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);CHKERRQ(ierr); 2381 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);CHKERRQ(ierr); 2382 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);CHKERRQ(ierr); 2383 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);CHKERRQ(ierr); 2384 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);CHKERRQ(ierr); 2385 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);CHKERRQ(ierr); 2386 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);CHKERRQ(ierr); 2387 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);CHKERRQ(ierr); 2388 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);CHKERRQ(ierr); 2389 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);CHKERRQ(ierr); 2390 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);CHKERRQ(ierr); 2391 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetMatInverse_C",MatMumpsGetMatInverse_MUMPS);CHKERRQ(ierr); 2392 2393 /* set solvertype */ 2394 ierr = PetscFree(B->solvertype);CHKERRQ(ierr); 2395 ierr = PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);CHKERRQ(ierr); 2396 2397 mumps->isAIJ = PETSC_TRUE; 2398 B->ops->destroy = MatDestroy_MUMPS; 2399 B->data = (void*)mumps; 2400 2401 ierr = PetscInitializeMUMPS(A,mumps);CHKERRQ(ierr); 2402 2403 *F = B; 2404 PetscFunctionReturn(0); 2405 } 2406 2407 PETSC_EXTERN PetscErrorCode MatSolverTypeRegister_MUMPS(void) 2408 { 2409 PetscErrorCode ierr; 2410 2411 PetscFunctionBegin; 2412 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);CHKERRQ(ierr); 2413 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);CHKERRQ(ierr); 2414 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);CHKERRQ(ierr); 2415 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);CHKERRQ(ierr); 2416 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPISBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);CHKERRQ(ierr); 2417 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);CHKERRQ(ierr); 2418 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);CHKERRQ(ierr); 2419 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);CHKERRQ(ierr); 2420 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);CHKERRQ(ierr); 2421 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQSBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);CHKERRQ(ierr); 2422 PetscFunctionReturn(0); 2423 } 2424 2425