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 #include <../src/mat/impls/sell/mpi/mpisell.h> 9 10 EXTERN_C_BEGIN 11 #if defined(PETSC_USE_COMPLEX) 12 #if defined(PETSC_USE_REAL_SINGLE) 13 #include <cmumps_c.h> 14 #else 15 #include <zmumps_c.h> 16 #endif 17 #else 18 #if defined(PETSC_USE_REAL_SINGLE) 19 #include <smumps_c.h> 20 #else 21 #include <dmumps_c.h> 22 #endif 23 #endif 24 EXTERN_C_END 25 #define JOB_INIT -1 26 #define JOB_FACTSYMBOLIC 1 27 #define JOB_FACTNUMERIC 2 28 #define JOB_SOLVE 3 29 #define JOB_END -2 30 31 /* calls to MUMPS */ 32 #if defined(PETSC_USE_COMPLEX) 33 #if defined(PETSC_USE_REAL_SINGLE) 34 #define PetscMUMPS_c cmumps_c 35 #else 36 #define PetscMUMPS_c zmumps_c 37 #endif 38 #else 39 #if defined(PETSC_USE_REAL_SINGLE) 40 #define PetscMUMPS_c smumps_c 41 #else 42 #define PetscMUMPS_c dmumps_c 43 #endif 44 #endif 45 46 /* declare MumpsScalar */ 47 #if defined(PETSC_USE_COMPLEX) 48 #if defined(PETSC_USE_REAL_SINGLE) 49 #define MumpsScalar mumps_complex 50 #else 51 #define MumpsScalar mumps_double_complex 52 #endif 53 #else 54 #define MumpsScalar PetscScalar 55 #endif 56 57 /* macros s.t. indices match MUMPS documentation */ 58 #define ICNTL(I) icntl[(I)-1] 59 #define CNTL(I) cntl[(I)-1] 60 #define INFOG(I) infog[(I)-1] 61 #define INFO(I) info[(I)-1] 62 #define RINFOG(I) rinfog[(I)-1] 63 #define RINFO(I) rinfo[(I)-1] 64 65 typedef struct { 66 #if defined(PETSC_USE_COMPLEX) 67 #if defined(PETSC_USE_REAL_SINGLE) 68 CMUMPS_STRUC_C id; 69 #else 70 ZMUMPS_STRUC_C id; 71 #endif 72 #else 73 #if defined(PETSC_USE_REAL_SINGLE) 74 SMUMPS_STRUC_C id; 75 #else 76 DMUMPS_STRUC_C id; 77 #endif 78 #endif 79 80 MatStructure matstruc; 81 PetscMPIInt myid,size; 82 PetscInt *irn,*jcn,nz,sym; 83 PetscScalar *val; 84 MPI_Comm comm_mumps; 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_seqsell_seqaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 246 { 247 Mat_SeqSELL *a=(Mat_SeqSELL*)A->data; 248 PetscInt *ptr; 249 250 PetscFunctionBegin; 251 *v = a->val; 252 if (reuse == MAT_INITIAL_MATRIX) { 253 PetscInt nz,i,j,row; 254 PetscErrorCode ierr; 255 256 nz = a->sliidx[a->totalslices]; 257 *nnz = nz; 258 ierr = PetscMalloc1(2*nz, &ptr);CHKERRQ(ierr); 259 *r = ptr; 260 *c = ptr + nz; 261 262 for (i=0; i<a->totalslices; i++) { 263 for (j=a->sliidx[i],row=0; j<a->sliidx[i+1]; j++,row=((row+1)&0x07)) { 264 *ptr++ = 8*i + row + shift; 265 } 266 } 267 for (i=0;i<nz;i++) *ptr++ = a->colidx[i] + shift; 268 } 269 PetscFunctionReturn(0); 270 } 271 272 PetscErrorCode MatConvertToTriples_seqbaij_seqaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 273 { 274 Mat_SeqBAIJ *aa=(Mat_SeqBAIJ*)A->data; 275 const PetscInt *ai,*aj,*ajj,bs2 = aa->bs2; 276 PetscInt bs,M,nz,idx=0,rnz,i,j,k,m; 277 PetscErrorCode ierr; 278 PetscInt *row,*col; 279 280 PetscFunctionBegin; 281 ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr); 282 M = A->rmap->N/bs; 283 *v = aa->a; 284 if (reuse == MAT_INITIAL_MATRIX) { 285 ai = aa->i; aj = aa->j; 286 nz = bs2*aa->nz; 287 *nnz = nz; 288 ierr = PetscMalloc1(2*nz, &row);CHKERRQ(ierr); 289 col = row + nz; 290 291 for (i=0; i<M; i++) { 292 ajj = aj + ai[i]; 293 rnz = ai[i+1] - ai[i]; 294 for (k=0; k<rnz; k++) { 295 for (j=0; j<bs; j++) { 296 for (m=0; m<bs; m++) { 297 row[idx] = i*bs + m + shift; 298 col[idx++] = bs*(ajj[k]) + j + shift; 299 } 300 } 301 } 302 } 303 *r = row; *c = col; 304 } 305 PetscFunctionReturn(0); 306 } 307 308 PetscErrorCode MatConvertToTriples_seqsbaij_seqsbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 309 { 310 const PetscInt *ai, *aj,*ajj,M=A->rmap->n; 311 PetscInt nz,rnz,i,j; 312 PetscErrorCode ierr; 313 PetscInt *row,*col; 314 Mat_SeqSBAIJ *aa=(Mat_SeqSBAIJ*)A->data; 315 316 PetscFunctionBegin; 317 *v = aa->a; 318 if (reuse == MAT_INITIAL_MATRIX) { 319 nz = aa->nz; 320 ai = aa->i; 321 aj = aa->j; 322 *v = aa->a; 323 *nnz = nz; 324 ierr = PetscMalloc1(2*nz, &row);CHKERRQ(ierr); 325 col = row + nz; 326 327 nz = 0; 328 for (i=0; i<M; i++) { 329 rnz = ai[i+1] - ai[i]; 330 ajj = aj + ai[i]; 331 for (j=0; j<rnz; j++) { 332 row[nz] = i+shift; col[nz++] = ajj[j] + shift; 333 } 334 } 335 *r = row; *c = col; 336 } 337 PetscFunctionReturn(0); 338 } 339 340 PetscErrorCode MatConvertToTriples_seqaij_seqsbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 341 { 342 const PetscInt *ai,*aj,*ajj,*adiag,M=A->rmap->n; 343 PetscInt nz,rnz,i,j; 344 const PetscScalar *av,*v1; 345 PetscScalar *val; 346 PetscErrorCode ierr; 347 PetscInt *row,*col; 348 Mat_SeqAIJ *aa=(Mat_SeqAIJ*)A->data; 349 PetscBool missing; 350 351 PetscFunctionBegin; 352 ai = aa->i; aj = aa->j; av = aa->a; 353 adiag = aa->diag; 354 ierr = MatMissingDiagonal_SeqAIJ(A,&missing,&i);CHKERRQ(ierr); 355 if (reuse == MAT_INITIAL_MATRIX) { 356 /* count nz in the upper triangular part of A */ 357 nz = 0; 358 if (missing) { 359 for (i=0; i<M; i++) { 360 if (PetscUnlikely(adiag[i] >= ai[i+1])) { 361 for (j=ai[i];j<ai[i+1];j++) { 362 if (aj[j] < i) continue; 363 nz++; 364 } 365 } else { 366 nz += ai[i+1] - adiag[i]; 367 } 368 } 369 } else { 370 for (i=0; i<M; i++) nz += ai[i+1] - adiag[i]; 371 } 372 *nnz = nz; 373 374 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 375 col = row + nz; 376 val = (PetscScalar*)(col + nz); 377 378 nz = 0; 379 if (missing) { 380 for (i=0; i<M; i++) { 381 if (PetscUnlikely(adiag[i] >= ai[i+1])) { 382 for (j=ai[i];j<ai[i+1];j++) { 383 if (aj[j] < i) continue; 384 row[nz] = i+shift; 385 col[nz] = aj[j]+shift; 386 val[nz] = av[j]; 387 nz++; 388 } 389 } else { 390 rnz = ai[i+1] - adiag[i]; 391 ajj = aj + adiag[i]; 392 v1 = av + adiag[i]; 393 for (j=0; j<rnz; j++) { 394 row[nz] = i+shift; col[nz] = ajj[j] + shift; val[nz++] = v1[j]; 395 } 396 } 397 } 398 } else { 399 for (i=0; i<M; i++) { 400 rnz = ai[i+1] - adiag[i]; 401 ajj = aj + adiag[i]; 402 v1 = av + adiag[i]; 403 for (j=0; j<rnz; j++) { 404 row[nz] = i+shift; col[nz] = ajj[j] + shift; val[nz++] = v1[j]; 405 } 406 } 407 } 408 *r = row; *c = col; *v = val; 409 } else { 410 nz = 0; val = *v; 411 if (missing) { 412 for (i=0; i <M; i++) { 413 if (PetscUnlikely(adiag[i] >= ai[i+1])) { 414 for (j=ai[i];j<ai[i+1];j++) { 415 if (aj[j] < i) continue; 416 val[nz++] = av[j]; 417 } 418 } else { 419 rnz = ai[i+1] - adiag[i]; 420 v1 = av + adiag[i]; 421 for (j=0; j<rnz; j++) { 422 val[nz++] = v1[j]; 423 } 424 } 425 } 426 } else { 427 for (i=0; i <M; i++) { 428 rnz = ai[i+1] - adiag[i]; 429 v1 = av + adiag[i]; 430 for (j=0; j<rnz; j++) { 431 val[nz++] = v1[j]; 432 } 433 } 434 } 435 } 436 PetscFunctionReturn(0); 437 } 438 439 PetscErrorCode MatConvertToTriples_mpisbaij_mpisbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 440 { 441 const PetscInt *ai, *aj, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj; 442 PetscErrorCode ierr; 443 PetscInt rstart,nz,i,j,jj,irow,countA,countB; 444 PetscInt *row,*col; 445 const PetscScalar *av, *bv,*v1,*v2; 446 PetscScalar *val; 447 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)A->data; 448 Mat_SeqSBAIJ *aa = (Mat_SeqSBAIJ*)(mat->A)->data; 449 Mat_SeqBAIJ *bb = (Mat_SeqBAIJ*)(mat->B)->data; 450 451 PetscFunctionBegin; 452 ai=aa->i; aj=aa->j; bi=bb->i; bj=bb->j; rstart= A->rmap->rstart; 453 av=aa->a; bv=bb->a; 454 455 garray = mat->garray; 456 457 if (reuse == MAT_INITIAL_MATRIX) { 458 nz = aa->nz + bb->nz; 459 *nnz = nz; 460 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 461 col = row + nz; 462 val = (PetscScalar*)(col + nz); 463 464 *r = row; *c = col; *v = val; 465 } else { 466 row = *r; col = *c; val = *v; 467 } 468 469 jj = 0; irow = rstart; 470 for (i=0; i<m; i++) { 471 ajj = aj + ai[i]; /* ptr to the beginning of this row */ 472 countA = ai[i+1] - ai[i]; 473 countB = bi[i+1] - bi[i]; 474 bjj = bj + bi[i]; 475 v1 = av + ai[i]; 476 v2 = bv + bi[i]; 477 478 /* A-part */ 479 for (j=0; j<countA; j++) { 480 if (reuse == MAT_INITIAL_MATRIX) { 481 row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift; 482 } 483 val[jj++] = v1[j]; 484 } 485 486 /* B-part */ 487 for (j=0; j < countB; j++) { 488 if (reuse == MAT_INITIAL_MATRIX) { 489 row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift; 490 } 491 val[jj++] = v2[j]; 492 } 493 irow++; 494 } 495 PetscFunctionReturn(0); 496 } 497 498 PetscErrorCode MatConvertToTriples_mpiaij_mpiaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 499 { 500 const PetscInt *ai, *aj, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj; 501 PetscErrorCode ierr; 502 PetscInt rstart,nz,i,j,jj,irow,countA,countB; 503 PetscInt *row,*col; 504 const PetscScalar *av, *bv,*v1,*v2; 505 PetscScalar *val; 506 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 507 Mat_SeqAIJ *aa = (Mat_SeqAIJ*)(mat->A)->data; 508 Mat_SeqAIJ *bb = (Mat_SeqAIJ*)(mat->B)->data; 509 510 PetscFunctionBegin; 511 ai=aa->i; aj=aa->j; bi=bb->i; bj=bb->j; rstart= A->rmap->rstart; 512 av=aa->a; bv=bb->a; 513 514 garray = mat->garray; 515 516 if (reuse == MAT_INITIAL_MATRIX) { 517 nz = aa->nz + bb->nz; 518 *nnz = nz; 519 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 520 col = row + nz; 521 val = (PetscScalar*)(col + nz); 522 523 *r = row; *c = col; *v = val; 524 } else { 525 row = *r; col = *c; val = *v; 526 } 527 528 jj = 0; irow = rstart; 529 for (i=0; i<m; i++) { 530 ajj = aj + ai[i]; /* ptr to the beginning of this row */ 531 countA = ai[i+1] - ai[i]; 532 countB = bi[i+1] - bi[i]; 533 bjj = bj + bi[i]; 534 v1 = av + ai[i]; 535 v2 = bv + bi[i]; 536 537 /* A-part */ 538 for (j=0; j<countA; j++) { 539 if (reuse == MAT_INITIAL_MATRIX) { 540 row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift; 541 } 542 val[jj++] = v1[j]; 543 } 544 545 /* B-part */ 546 for (j=0; j < countB; j++) { 547 if (reuse == MAT_INITIAL_MATRIX) { 548 row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift; 549 } 550 val[jj++] = v2[j]; 551 } 552 irow++; 553 } 554 PetscFunctionReturn(0); 555 } 556 557 PetscErrorCode MatConvertToTriples_mpibaij_mpiaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 558 { 559 Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)A->data; 560 Mat_SeqBAIJ *aa = (Mat_SeqBAIJ*)(mat->A)->data; 561 Mat_SeqBAIJ *bb = (Mat_SeqBAIJ*)(mat->B)->data; 562 const PetscInt *ai = aa->i, *bi = bb->i, *aj = aa->j, *bj = bb->j,*ajj, *bjj; 563 const PetscInt *garray = mat->garray,mbs=mat->mbs,rstart=A->rmap->rstart; 564 const PetscInt bs2=mat->bs2; 565 PetscErrorCode ierr; 566 PetscInt bs,nz,i,j,k,n,jj,irow,countA,countB,idx; 567 PetscInt *row,*col; 568 const PetscScalar *av=aa->a, *bv=bb->a,*v1,*v2; 569 PetscScalar *val; 570 571 PetscFunctionBegin; 572 ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr); 573 if (reuse == MAT_INITIAL_MATRIX) { 574 nz = bs2*(aa->nz + bb->nz); 575 *nnz = nz; 576 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 577 col = row + nz; 578 val = (PetscScalar*)(col + nz); 579 580 *r = row; *c = col; *v = val; 581 } else { 582 row = *r; col = *c; val = *v; 583 } 584 585 jj = 0; irow = rstart; 586 for (i=0; i<mbs; i++) { 587 countA = ai[i+1] - ai[i]; 588 countB = bi[i+1] - bi[i]; 589 ajj = aj + ai[i]; 590 bjj = bj + bi[i]; 591 v1 = av + bs2*ai[i]; 592 v2 = bv + bs2*bi[i]; 593 594 idx = 0; 595 /* A-part */ 596 for (k=0; k<countA; k++) { 597 for (j=0; j<bs; j++) { 598 for (n=0; n<bs; n++) { 599 if (reuse == MAT_INITIAL_MATRIX) { 600 row[jj] = irow + n + shift; 601 col[jj] = rstart + bs*ajj[k] + j + shift; 602 } 603 val[jj++] = v1[idx++]; 604 } 605 } 606 } 607 608 idx = 0; 609 /* B-part */ 610 for (k=0; k<countB; k++) { 611 for (j=0; j<bs; j++) { 612 for (n=0; n<bs; n++) { 613 if (reuse == MAT_INITIAL_MATRIX) { 614 row[jj] = irow + n + shift; 615 col[jj] = bs*garray[bjj[k]] + j + shift; 616 } 617 val[jj++] = v2[idx++]; 618 } 619 } 620 } 621 irow += bs; 622 } 623 PetscFunctionReturn(0); 624 } 625 626 PetscErrorCode MatConvertToTriples_mpiaij_mpisbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v) 627 { 628 const PetscInt *ai, *aj,*adiag, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj; 629 PetscErrorCode ierr; 630 PetscInt rstart,nz,nza,nzb,i,j,jj,irow,countA,countB; 631 PetscInt *row,*col; 632 const PetscScalar *av, *bv,*v1,*v2; 633 PetscScalar *val; 634 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 635 Mat_SeqAIJ *aa =(Mat_SeqAIJ*)(mat->A)->data; 636 Mat_SeqAIJ *bb =(Mat_SeqAIJ*)(mat->B)->data; 637 638 PetscFunctionBegin; 639 ai=aa->i; aj=aa->j; adiag=aa->diag; 640 bi=bb->i; bj=bb->j; garray = mat->garray; 641 av=aa->a; bv=bb->a; 642 643 rstart = A->rmap->rstart; 644 645 if (reuse == MAT_INITIAL_MATRIX) { 646 nza = 0; /* num of upper triangular entries in mat->A, including diagonals */ 647 nzb = 0; /* num of upper triangular entries in mat->B */ 648 for (i=0; i<m; i++) { 649 nza += (ai[i+1] - adiag[i]); 650 countB = bi[i+1] - bi[i]; 651 bjj = bj + bi[i]; 652 for (j=0; j<countB; j++) { 653 if (garray[bjj[j]] > rstart) nzb++; 654 } 655 } 656 657 nz = nza + nzb; /* total nz of upper triangular part of mat */ 658 *nnz = nz; 659 ierr = PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);CHKERRQ(ierr); 660 col = row + nz; 661 val = (PetscScalar*)(col + nz); 662 663 *r = row; *c = col; *v = val; 664 } else { 665 row = *r; col = *c; val = *v; 666 } 667 668 jj = 0; irow = rstart; 669 for (i=0; i<m; i++) { 670 ajj = aj + adiag[i]; /* ptr to the beginning of the diagonal of this row */ 671 v1 = av + adiag[i]; 672 countA = ai[i+1] - adiag[i]; 673 countB = bi[i+1] - bi[i]; 674 bjj = bj + bi[i]; 675 v2 = bv + bi[i]; 676 677 /* A-part */ 678 for (j=0; j<countA; j++) { 679 if (reuse == MAT_INITIAL_MATRIX) { 680 row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift; 681 } 682 val[jj++] = v1[j]; 683 } 684 685 /* B-part */ 686 for (j=0; j < countB; j++) { 687 if (garray[bjj[j]] > rstart) { 688 if (reuse == MAT_INITIAL_MATRIX) { 689 row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift; 690 } 691 val[jj++] = v2[j]; 692 } 693 } 694 irow++; 695 } 696 PetscFunctionReturn(0); 697 } 698 699 PetscErrorCode MatDestroy_MUMPS(Mat A) 700 { 701 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 702 PetscErrorCode ierr; 703 704 PetscFunctionBegin; 705 ierr = PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);CHKERRQ(ierr); 706 ierr = VecScatterDestroy(&mumps->scat_rhs);CHKERRQ(ierr); 707 ierr = VecScatterDestroy(&mumps->scat_sol);CHKERRQ(ierr); 708 ierr = VecDestroy(&mumps->b_seq);CHKERRQ(ierr); 709 ierr = VecDestroy(&mumps->x_seq);CHKERRQ(ierr); 710 ierr = PetscFree(mumps->id.perm_in);CHKERRQ(ierr); 711 ierr = PetscFree(mumps->irn);CHKERRQ(ierr); 712 ierr = PetscFree(mumps->info);CHKERRQ(ierr); 713 ierr = MatMumpsResetSchur_Private(mumps);CHKERRQ(ierr); 714 mumps->id.job = JOB_END; 715 PetscMUMPS_c(&mumps->id); 716 ierr = MPI_Comm_free(&mumps->comm_mumps);CHKERRQ(ierr); 717 ierr = PetscFree(A->data);CHKERRQ(ierr); 718 719 /* clear composed functions */ 720 ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverType_C",NULL);CHKERRQ(ierr); 721 ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorSetSchurIS_C",NULL);CHKERRQ(ierr); 722 ierr = PetscObjectComposeFunction((PetscObject)A,"MatFactorCreateSchurComplement_C",NULL);CHKERRQ(ierr); 723 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetIcntl_C",NULL);CHKERRQ(ierr); 724 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetIcntl_C",NULL);CHKERRQ(ierr); 725 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetCntl_C",NULL);CHKERRQ(ierr); 726 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetCntl_C",NULL);CHKERRQ(ierr); 727 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfo_C",NULL);CHKERRQ(ierr); 728 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfog_C",NULL);CHKERRQ(ierr); 729 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfo_C",NULL);CHKERRQ(ierr); 730 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfog_C",NULL);CHKERRQ(ierr); 731 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInverse_C",NULL);CHKERRQ(ierr); 732 ierr = PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInverseTranspose_C",NULL);CHKERRQ(ierr); 733 PetscFunctionReturn(0); 734 } 735 736 PetscErrorCode MatSolve_MUMPS(Mat A,Vec b,Vec x) 737 { 738 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 739 PetscScalar *array; 740 Vec b_seq; 741 IS is_iden,is_petsc; 742 PetscErrorCode ierr; 743 PetscInt i; 744 PetscBool second_solve = PETSC_FALSE; 745 static PetscBool cite1 = PETSC_FALSE,cite2 = PETSC_FALSE; 746 747 PetscFunctionBegin; 748 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); 749 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); 750 751 if (A->factorerrortype) { 752 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); 753 ierr = VecSetInf(x);CHKERRQ(ierr); 754 PetscFunctionReturn(0); 755 } 756 757 mumps->id.ICNTL(20)= 0; /* dense RHS */ 758 mumps->id.nrhs = 1; 759 b_seq = mumps->b_seq; 760 if (mumps->size > 1) { 761 /* MUMPS only supports centralized rhs. Scatter b into a seqential rhs vector */ 762 ierr = VecScatterBegin(mumps->scat_rhs,b,b_seq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 763 ierr = VecScatterEnd(mumps->scat_rhs,b,b_seq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 764 if (!mumps->myid) {ierr = VecGetArray(b_seq,&array);CHKERRQ(ierr);} 765 } else { /* size == 1 */ 766 ierr = VecCopy(b,x);CHKERRQ(ierr); 767 ierr = VecGetArray(x,&array);CHKERRQ(ierr); 768 } 769 if (!mumps->myid) { /* define rhs on the host */ 770 mumps->id.nrhs = 1; 771 mumps->id.rhs = (MumpsScalar*)array; 772 } 773 774 /* 775 handle condensation step of Schur complement (if any) 776 We set by default ICNTL(26) == -1 when Schur indices have been provided by the user. 777 According to MUMPS (5.0.0) manual, any value should be harmful during the factorization phase 778 Unless the user provides a valid value for ICNTL(26), MatSolve and MatMatSolve routines solve the full system. 779 This requires an extra call to PetscMUMPS_c and the computation of the factors for S 780 */ 781 if (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2) { 782 if (mumps->size > 1) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Parallel Schur complements not yet supported from PETSc\n"); 783 second_solve = PETSC_TRUE; 784 ierr = MatMumpsHandleSchur_Private(A,PETSC_FALSE);CHKERRQ(ierr); 785 } 786 /* solve phase */ 787 /*-------------*/ 788 mumps->id.job = JOB_SOLVE; 789 PetscMUMPS_c(&mumps->id); 790 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)); 791 792 /* handle expansion step of Schur complement (if any) */ 793 if (second_solve) { 794 ierr = MatMumpsHandleSchur_Private(A,PETSC_TRUE);CHKERRQ(ierr); 795 } 796 797 if (mumps->size > 1) { /* convert mumps distributed solution to petsc mpi x */ 798 if (mumps->scat_sol && mumps->ICNTL9_pre != mumps->id.ICNTL(9)) { 799 /* when id.ICNTL(9) changes, the contents of lsol_loc may change (not its size, lsol_loc), recreates scat_sol */ 800 ierr = VecScatterDestroy(&mumps->scat_sol);CHKERRQ(ierr); 801 } 802 if (!mumps->scat_sol) { /* create scatter scat_sol */ 803 ierr = ISCreateStride(PETSC_COMM_SELF,mumps->id.lsol_loc,0,1,&is_iden);CHKERRQ(ierr); /* from */ 804 for (i=0; i<mumps->id.lsol_loc; i++) { 805 mumps->id.isol_loc[i] -= 1; /* change Fortran style to C style */ 806 } 807 ierr = ISCreateGeneral(PETSC_COMM_SELF,mumps->id.lsol_loc,mumps->id.isol_loc,PETSC_COPY_VALUES,&is_petsc);CHKERRQ(ierr); /* to */ 808 ierr = VecScatterCreate(mumps->x_seq,is_iden,x,is_petsc,&mumps->scat_sol);CHKERRQ(ierr); 809 ierr = ISDestroy(&is_iden);CHKERRQ(ierr); 810 ierr = ISDestroy(&is_petsc);CHKERRQ(ierr); 811 812 mumps->ICNTL9_pre = mumps->id.ICNTL(9); /* save current value of id.ICNTL(9) */ 813 } 814 815 ierr = VecScatterBegin(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 816 ierr = VecScatterEnd(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 817 } 818 PetscFunctionReturn(0); 819 } 820 821 PetscErrorCode MatSolveTranspose_MUMPS(Mat A,Vec b,Vec x) 822 { 823 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 824 PetscErrorCode ierr; 825 826 PetscFunctionBegin; 827 mumps->id.ICNTL(9) = 0; 828 ierr = MatSolve_MUMPS(A,b,x);CHKERRQ(ierr); 829 mumps->id.ICNTL(9) = 1; 830 PetscFunctionReturn(0); 831 } 832 833 PetscErrorCode MatMatSolve_MUMPS(Mat A,Mat B,Mat X) 834 { 835 PetscErrorCode ierr; 836 Mat Bt = NULL; 837 PetscBool flg, flgT; 838 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 839 PetscInt i,nrhs,M; 840 PetscScalar *array,*bray; 841 PetscInt lsol_loc,nlsol_loc,*isol_loc,*idx,*iidx,*idxx,*isol_loc_save; 842 MumpsScalar *sol_loc,*sol_loc_save; 843 IS is_to,is_from; 844 PetscInt k,proc,j,m; 845 const PetscInt *rstart; 846 Vec v_mpi,b_seq,x_seq; 847 VecScatter scat_rhs,scat_sol; 848 PetscScalar *aa; 849 PetscInt spnr,*ia,*ja; 850 Mat_MPIAIJ *b = NULL; 851 852 PetscFunctionBegin; 853 ierr = PetscObjectTypeCompareAny((PetscObject)X,&flg,MATSEQDENSE,MATMPIDENSE,NULL);CHKERRQ(ierr); 854 if (!flg) SETERRQ(PetscObjectComm((PetscObject)X),PETSC_ERR_ARG_WRONG,"Matrix X must be MATDENSE matrix"); 855 856 ierr = PetscObjectTypeCompareAny((PetscObject)B,&flg,MATSEQDENSE,MATMPIDENSE,NULL);CHKERRQ(ierr); 857 if (flg) { /* dense B */ 858 if (B->rmap->n != X->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Matrix B and X must have same row distribution"); 859 mumps->id.ICNTL(20)= 0; /* dense RHS */ 860 } else { /* sparse B */ 861 ierr = PetscObjectTypeCompare((PetscObject)B,MATTRANSPOSEMAT,&flgT);CHKERRQ(ierr); 862 if (flgT) { /* input B is transpose of actural RHS matrix, 863 because mumps requires sparse compressed COLUMN storage! See MatMatTransposeSolve_MUMPS() */ 864 ierr = MatTransposeGetMat(B,&Bt);CHKERRQ(ierr); 865 } else SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_WRONG,"Matrix B must be MATDENSE matrix"); 866 mumps->id.ICNTL(20)= 1; /* sparse RHS */ 867 } 868 869 ierr = MatGetSize(B,&M,&nrhs);CHKERRQ(ierr); 870 mumps->id.nrhs = nrhs; 871 mumps->id.lrhs = M; 872 mumps->id.rhs = NULL; 873 874 if (mumps->size == 1) { 875 PetscScalar *aa; 876 PetscInt spnr,*ia,*ja; 877 PetscBool second_solve = PETSC_FALSE; 878 879 ierr = MatDenseGetArray(X,&array);CHKERRQ(ierr); 880 mumps->id.rhs = (MumpsScalar*)array; 881 882 if (!Bt) { /* dense B */ 883 /* copy B to X */ 884 ierr = MatDenseGetArray(B,&bray);CHKERRQ(ierr); 885 ierr = PetscMemcpy(array,bray,M*nrhs*sizeof(PetscScalar));CHKERRQ(ierr); 886 ierr = MatDenseRestoreArray(B,&bray);CHKERRQ(ierr); 887 } else { /* sparse B */ 888 ierr = MatSeqAIJGetArray(Bt,&aa);CHKERRQ(ierr); 889 ierr = MatGetRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);CHKERRQ(ierr); 890 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure"); 891 /* mumps requires ia and ja start at 1! */ 892 mumps->id.irhs_ptr = ia; 893 mumps->id.irhs_sparse = ja; 894 mumps->id.nz_rhs = ia[spnr] - 1; 895 mumps->id.rhs_sparse = (MumpsScalar*)aa; 896 } 897 /* handle condensation step of Schur complement (if any) */ 898 if (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2) { 899 second_solve = PETSC_TRUE; 900 ierr = MatMumpsHandleSchur_Private(A,PETSC_FALSE);CHKERRQ(ierr); 901 } 902 /* solve phase */ 903 /*-------------*/ 904 mumps->id.job = JOB_SOLVE; 905 PetscMUMPS_c(&mumps->id); 906 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)); 907 908 /* handle expansion step of Schur complement (if any) */ 909 if (second_solve) { 910 ierr = MatMumpsHandleSchur_Private(A,PETSC_TRUE);CHKERRQ(ierr); 911 } 912 if (Bt) { /* sparse B */ 913 ierr = MatSeqAIJRestoreArray(Bt,&aa);CHKERRQ(ierr); 914 ierr = MatRestoreRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);CHKERRQ(ierr); 915 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot restore IJ structure"); 916 } 917 ierr = MatDenseRestoreArray(X,&array);CHKERRQ(ierr); 918 PetscFunctionReturn(0); 919 } 920 921 /*--------- parallel case: MUMPS requires rhs B to be centralized on the host! --------*/ 922 if (mumps->size > 1 && mumps->id.ICNTL(19)) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Parallel Schur complements not yet supported from PETSc\n"); 923 924 /* create x_seq to hold mumps local solution */ 925 isol_loc_save = mumps->id.isol_loc; /* save it for MatSovle() */ 926 sol_loc_save = mumps->id.sol_loc; 927 928 lsol_loc = mumps->id.INFO(23); 929 nlsol_loc = nrhs*lsol_loc; /* length of sol_loc */ 930 ierr = PetscMalloc2(nlsol_loc,&sol_loc,nlsol_loc,&isol_loc);CHKERRQ(ierr); 931 mumps->id.sol_loc = (MumpsScalar*)sol_loc; 932 mumps->id.isol_loc = isol_loc; 933 934 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,nlsol_loc,(PetscScalar*)sol_loc,&x_seq);CHKERRQ(ierr); 935 936 /* scatter v_mpi to b_seq because MUMPS only supports centralized rhs */ 937 /* idx: maps from k-th index of v_mpi to (i,j)-th global entry of B; 938 iidx: inverse of idx, will be used by scattering mumps x_seq -> petsc X */ 939 ierr = PetscMalloc1(nrhs*M,&idx);CHKERRQ(ierr); 940 941 if (!Bt) { /* dense B */ 942 /* wrap dense rhs matrix B into a vector v_mpi */ 943 ierr = MatGetLocalSize(B,&m,NULL);CHKERRQ(ierr); 944 ierr = MatDenseGetArray(B,&bray);CHKERRQ(ierr); 945 ierr = VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),1,nrhs*m,nrhs*M,(const PetscScalar*)bray,&v_mpi);CHKERRQ(ierr); 946 ierr = MatDenseRestoreArray(B,&bray);CHKERRQ(ierr); 947 948 /* scatter v_mpi to b_seq in proc[0]. MUMPS requires rhs to be centralized on the host! */ 949 if (!mumps->myid) { 950 ierr = MatGetOwnershipRanges(B,&rstart);CHKERRQ(ierr); 951 k = 0; 952 for (proc=0; proc<mumps->size; proc++){ 953 for (j=0; j<nrhs; j++){ 954 for (i=rstart[proc]; i<rstart[proc+1]; i++){ 955 idx[k++] = j*M + i; 956 } 957 } 958 } 959 960 ierr = VecCreateSeq(PETSC_COMM_SELF,nrhs*M,&b_seq);CHKERRQ(ierr); 961 ierr = ISCreateGeneral(PETSC_COMM_SELF,nrhs*M,idx,PETSC_COPY_VALUES,&is_to);CHKERRQ(ierr); 962 ierr = ISCreateStride(PETSC_COMM_SELF,nrhs*M,0,1,&is_from);CHKERRQ(ierr); 963 } else { 964 ierr = VecCreateSeq(PETSC_COMM_SELF,0,&b_seq);CHKERRQ(ierr); 965 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_to);CHKERRQ(ierr); 966 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_from);CHKERRQ(ierr); 967 } 968 ierr = VecScatterCreate(v_mpi,is_from,b_seq,is_to,&scat_rhs);CHKERRQ(ierr); 969 ierr = VecScatterBegin(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 970 ierr = ISDestroy(&is_to);CHKERRQ(ierr); 971 ierr = ISDestroy(&is_from);CHKERRQ(ierr); 972 ierr = VecScatterEnd(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 973 974 if (!mumps->myid) { /* define rhs on the host */ 975 ierr = VecGetArray(b_seq,&bray);CHKERRQ(ierr); 976 mumps->id.rhs = (MumpsScalar*)bray; 977 ierr = VecRestoreArray(b_seq,&bray);CHKERRQ(ierr); 978 } 979 980 } else { /* sparse B */ 981 b = (Mat_MPIAIJ*)Bt->data; 982 983 /* wrap dense X into a vector v_mpi */ 984 ierr = MatGetLocalSize(X,&m,NULL);CHKERRQ(ierr); 985 ierr = MatDenseGetArray(X,&bray);CHKERRQ(ierr); 986 ierr = VecCreateMPIWithArray(PetscObjectComm((PetscObject)X),1,nrhs*m,nrhs*M,(const PetscScalar*)bray,&v_mpi);CHKERRQ(ierr); 987 ierr = MatDenseRestoreArray(X,&bray);CHKERRQ(ierr); 988 989 if (!mumps->myid) { 990 ierr = MatSeqAIJGetArray(b->A,&aa);CHKERRQ(ierr); 991 ierr = MatGetRowIJ(b->A,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);CHKERRQ(ierr); 992 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure"); 993 /* mumps requires ia and ja start at 1! */ 994 mumps->id.irhs_ptr = ia; 995 mumps->id.irhs_sparse = ja; 996 mumps->id.nz_rhs = ia[spnr] - 1; 997 mumps->id.rhs_sparse = (MumpsScalar*)aa; 998 } else { 999 mumps->id.irhs_ptr = NULL; 1000 mumps->id.irhs_sparse = NULL; 1001 mumps->id.nz_rhs = 0; 1002 mumps->id.rhs_sparse = NULL; 1003 } 1004 } 1005 1006 /* solve phase */ 1007 /*-------------*/ 1008 mumps->id.job = JOB_SOLVE; 1009 PetscMUMPS_c(&mumps->id); 1010 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)); 1011 1012 /* scatter mumps distributed solution to petsc vector v_mpi, which shares local arrays with solution matrix X */ 1013 ierr = MatDenseGetArray(X,&array);CHKERRQ(ierr); 1014 ierr = VecPlaceArray(v_mpi,array);CHKERRQ(ierr); 1015 1016 /* create scatter scat_sol */ 1017 ierr = MatGetOwnershipRanges(X,&rstart);CHKERRQ(ierr); 1018 /* iidx: inverse of idx computed above, used for scattering mumps x_seq to petsc X */ 1019 iidx = idx; 1020 k = 0; 1021 for (proc=0; proc<mumps->size; proc++){ 1022 for (j=0; j<nrhs; j++){ 1023 for (i=rstart[proc]; i<rstart[proc+1]; i++) iidx[j*M + i] = k++; 1024 } 1025 } 1026 1027 ierr = PetscMalloc1(nlsol_loc,&idxx);CHKERRQ(ierr); 1028 ierr = ISCreateStride(PETSC_COMM_SELF,nlsol_loc,0,1,&is_from);CHKERRQ(ierr); 1029 for (i=0; i<lsol_loc; i++) { 1030 isol_loc[i] -= 1; /* change Fortran style to C style */ 1031 idxx[i] = iidx[isol_loc[i]]; 1032 for (j=1; j<nrhs; j++){ 1033 idxx[j*lsol_loc+i] = iidx[isol_loc[i]+j*M]; 1034 } 1035 } 1036 ierr = ISCreateGeneral(PETSC_COMM_SELF,nlsol_loc,idxx,PETSC_COPY_VALUES,&is_to);CHKERRQ(ierr); 1037 ierr = VecScatterCreate(x_seq,is_from,v_mpi,is_to,&scat_sol);CHKERRQ(ierr); 1038 ierr = VecScatterBegin(scat_sol,x_seq,v_mpi,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1039 ierr = ISDestroy(&is_from);CHKERRQ(ierr); 1040 ierr = ISDestroy(&is_to);CHKERRQ(ierr); 1041 ierr = VecScatterEnd(scat_sol,x_seq,v_mpi,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1042 ierr = MatDenseRestoreArray(X,&array);CHKERRQ(ierr); 1043 1044 /* free spaces */ 1045 mumps->id.sol_loc = sol_loc_save; 1046 mumps->id.isol_loc = isol_loc_save; 1047 1048 ierr = PetscFree2(sol_loc,isol_loc);CHKERRQ(ierr); 1049 ierr = PetscFree(idx);CHKERRQ(ierr); 1050 ierr = PetscFree(idxx);CHKERRQ(ierr); 1051 ierr = VecDestroy(&x_seq);CHKERRQ(ierr); 1052 ierr = VecDestroy(&v_mpi);CHKERRQ(ierr); 1053 if (Bt) { 1054 if (!mumps->myid) { 1055 b = (Mat_MPIAIJ*)Bt->data; 1056 ierr = MatSeqAIJRestoreArray(b->A,&aa);CHKERRQ(ierr); 1057 ierr = MatRestoreRowIJ(b->A,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);CHKERRQ(ierr); 1058 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot restore IJ structure"); 1059 } 1060 } else { 1061 ierr = VecDestroy(&b_seq);CHKERRQ(ierr); 1062 ierr = VecScatterDestroy(&scat_rhs);CHKERRQ(ierr); 1063 } 1064 ierr = VecScatterDestroy(&scat_sol);CHKERRQ(ierr); 1065 PetscFunctionReturn(0); 1066 } 1067 1068 PetscErrorCode MatMatTransposeSolve_MUMPS(Mat A,Mat Bt,Mat X) 1069 { 1070 PetscErrorCode ierr; 1071 PetscBool flg; 1072 Mat B; 1073 1074 PetscFunctionBegin; 1075 ierr = PetscObjectTypeCompareAny((PetscObject)Bt,&flg,MATSEQAIJ,MATMPIAIJ,NULL);CHKERRQ(ierr); 1076 if (!flg) SETERRQ(PetscObjectComm((PetscObject)Bt),PETSC_ERR_ARG_WRONG,"Matrix Bt must be MATAIJ matrix"); 1077 1078 /* Create B=Bt^T that uses Bt's data structure */ 1079 ierr = MatCreateTranspose(Bt,&B);CHKERRQ(ierr); 1080 1081 ierr = MatMatSolve_MUMPS(A,B,X);CHKERRQ(ierr); 1082 ierr = MatDestroy(&B);CHKERRQ(ierr); 1083 PetscFunctionReturn(0); 1084 } 1085 1086 #if !defined(PETSC_USE_COMPLEX) 1087 /* 1088 input: 1089 F: numeric factor 1090 output: 1091 nneg: total number of negative pivots 1092 nzero: total number of zero pivots 1093 npos: (global dimension of F) - nneg - nzero 1094 */ 1095 PetscErrorCode MatGetInertia_SBAIJMUMPS(Mat F,int *nneg,int *nzero,int *npos) 1096 { 1097 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1098 PetscErrorCode ierr; 1099 PetscMPIInt size; 1100 1101 PetscFunctionBegin; 1102 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)F),&size);CHKERRQ(ierr); 1103 /* 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 */ 1104 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)); 1105 1106 if (nneg) *nneg = mumps->id.INFOG(12); 1107 if (nzero || npos) { 1108 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"); 1109 if (nzero) *nzero = mumps->id.INFOG(28); 1110 if (npos) *npos = F->rmap->N - (mumps->id.INFOG(12) + mumps->id.INFOG(28)); 1111 } 1112 PetscFunctionReturn(0); 1113 } 1114 #endif 1115 1116 PetscErrorCode MatFactorNumeric_MUMPS(Mat F,Mat A,const MatFactorInfo *info) 1117 { 1118 Mat_MUMPS *mumps =(Mat_MUMPS*)(F)->data; 1119 PetscErrorCode ierr; 1120 PetscBool isMPIAIJ; 1121 1122 PetscFunctionBegin; 1123 if (mumps->id.INFOG(1) < 0) { 1124 if (mumps->id.INFOG(1) == -6) { 1125 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); 1126 } 1127 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); 1128 PetscFunctionReturn(0); 1129 } 1130 1131 ierr = (*mumps->ConvertToTriples)(A, 1, MAT_REUSE_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);CHKERRQ(ierr); 1132 1133 /* numerical factorization phase */ 1134 /*-------------------------------*/ 1135 mumps->id.job = JOB_FACTNUMERIC; 1136 if (!mumps->id.ICNTL(18)) { /* A is centralized */ 1137 if (!mumps->myid) { 1138 mumps->id.a = (MumpsScalar*)mumps->val; 1139 } 1140 } else { 1141 mumps->id.a_loc = (MumpsScalar*)mumps->val; 1142 } 1143 PetscMUMPS_c(&mumps->id); 1144 if (mumps->id.INFOG(1) < 0) { 1145 if (A->erroriffailure) { 1146 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)); 1147 } else { 1148 if (mumps->id.INFOG(1) == -10) { /* numerically singular matrix */ 1149 ierr = PetscInfo2(F,"matrix is numerically singular, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1150 F->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1151 } else if (mumps->id.INFOG(1) == -13) { 1152 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); 1153 F->factorerrortype = MAT_FACTOR_OUTMEMORY; 1154 } else if (mumps->id.INFOG(1) == -8 || mumps->id.INFOG(1) == -9 || (-16 < mumps->id.INFOG(1) && mumps->id.INFOG(1) < -10) ) { 1155 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); 1156 F->factorerrortype = MAT_FACTOR_OUTMEMORY; 1157 } else { 1158 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); 1159 F->factorerrortype = MAT_FACTOR_OTHER; 1160 } 1161 } 1162 } 1163 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)); 1164 1165 F->assembled = PETSC_TRUE; 1166 mumps->matstruc = SAME_NONZERO_PATTERN; 1167 if (F->schur) { /* reset Schur status to unfactored */ 1168 if (mumps->id.ICNTL(19) == 1) { /* stored by rows */ 1169 mumps->id.ICNTL(19) = 2; 1170 ierr = MatTranspose(F->schur,MAT_INPLACE_MATRIX,&F->schur);CHKERRQ(ierr); 1171 } 1172 ierr = MatFactorRestoreSchurComplement(F,NULL,MAT_FACTOR_SCHUR_UNFACTORED);CHKERRQ(ierr); 1173 } 1174 1175 /* just to be sure that ICNTL(19) value returned by a call from MatMumpsGetIcntl is always consistent */ 1176 if (!mumps->sym && mumps->id.ICNTL(19) && mumps->id.ICNTL(19) != 1) mumps->id.ICNTL(19) = 3; 1177 1178 if (mumps->size > 1) { 1179 PetscInt lsol_loc; 1180 PetscScalar *sol_loc; 1181 1182 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&isMPIAIJ);CHKERRQ(ierr); 1183 1184 /* distributed solution; Create x_seq=sol_loc for repeated use */ 1185 if (mumps->x_seq) { 1186 ierr = VecScatterDestroy(&mumps->scat_sol);CHKERRQ(ierr); 1187 ierr = PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);CHKERRQ(ierr); 1188 ierr = VecDestroy(&mumps->x_seq);CHKERRQ(ierr); 1189 } 1190 lsol_loc = mumps->id.INFO(23); /* length of sol_loc */ 1191 ierr = PetscMalloc2(lsol_loc,&sol_loc,lsol_loc,&mumps->id.isol_loc);CHKERRQ(ierr); 1192 mumps->id.lsol_loc = lsol_loc; 1193 mumps->id.sol_loc = (MumpsScalar*)sol_loc; 1194 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lsol_loc,sol_loc,&mumps->x_seq);CHKERRQ(ierr); 1195 } 1196 PetscFunctionReturn(0); 1197 } 1198 1199 /* Sets MUMPS options from the options database */ 1200 PetscErrorCode PetscSetMUMPSFromOptions(Mat F, Mat A) 1201 { 1202 Mat_MUMPS *mumps = (Mat_MUMPS*)F->data; 1203 PetscErrorCode ierr; 1204 PetscInt icntl,info[40],i,ninfo=40; 1205 PetscBool flg; 1206 1207 PetscFunctionBegin; 1208 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"MUMPS Options","Mat");CHKERRQ(ierr); 1209 ierr = PetscOptionsInt("-mat_mumps_icntl_1","ICNTL(1): output stream for error messages","None",mumps->id.ICNTL(1),&icntl,&flg);CHKERRQ(ierr); 1210 if (flg) mumps->id.ICNTL(1) = icntl; 1211 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); 1212 if (flg) mumps->id.ICNTL(2) = icntl; 1213 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); 1214 if (flg) mumps->id.ICNTL(3) = icntl; 1215 1216 ierr = PetscOptionsInt("-mat_mumps_icntl_4","ICNTL(4): level of printing (0 to 4)","None",mumps->id.ICNTL(4),&icntl,&flg);CHKERRQ(ierr); 1217 if (flg) mumps->id.ICNTL(4) = icntl; 1218 if (mumps->id.ICNTL(4) || PetscLogPrintInfo) mumps->id.ICNTL(3) = 6; /* resume MUMPS default id.ICNTL(3) = 6 */ 1219 1220 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); 1221 if (flg) mumps->id.ICNTL(6) = icntl; 1222 1223 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); 1224 if (flg) { 1225 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"); 1226 else mumps->id.ICNTL(7) = icntl; 1227 } 1228 1229 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); 1230 /* 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() */ 1231 ierr = PetscOptionsInt("-mat_mumps_icntl_10","ICNTL(10): max num of refinements","None",mumps->id.ICNTL(10),&mumps->id.ICNTL(10),NULL);CHKERRQ(ierr); 1232 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); 1233 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); 1234 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); 1235 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); 1236 ierr = PetscOptionsInt("-mat_mumps_icntl_19","ICNTL(19): computes the Schur complement","None",mumps->id.ICNTL(19),&mumps->id.ICNTL(19),NULL);CHKERRQ(ierr); 1237 if (mumps->id.ICNTL(19) <= 0 || mumps->id.ICNTL(19) > 3) { /* reset any schur data (if any) */ 1238 ierr = MatDestroy(&F->schur);CHKERRQ(ierr); 1239 ierr = MatMumpsResetSchur_Private(mumps);CHKERRQ(ierr); 1240 } 1241 /* 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 */ 1242 /* 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 */ 1243 1244 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); 1245 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); 1246 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); 1247 if (mumps->id.ICNTL(24)) { 1248 mumps->id.ICNTL(13) = 1; /* turn-off ScaLAPACK to help with the correct detection of null pivots */ 1249 } 1250 1251 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); 1252 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); 1253 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); 1254 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); 1255 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); 1256 /* 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); */ /* call MatMumpsGetInverse() directly */ 1257 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); 1258 /* 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 */ 1259 ierr = PetscOptionsInt("-mat_mumps_icntl_33","ICNTL(33): compute determinant","None",mumps->id.ICNTL(33),&mumps->id.ICNTL(33),NULL);CHKERRQ(ierr); 1260 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); 1261 1262 ierr = PetscOptionsReal("-mat_mumps_cntl_1","CNTL(1): relative pivoting threshold","None",mumps->id.CNTL(1),&mumps->id.CNTL(1),NULL);CHKERRQ(ierr); 1263 ierr = PetscOptionsReal("-mat_mumps_cntl_2","CNTL(2): stopping criterion of refinement","None",mumps->id.CNTL(2),&mumps->id.CNTL(2),NULL);CHKERRQ(ierr); 1264 ierr = PetscOptionsReal("-mat_mumps_cntl_3","CNTL(3): absolute pivoting threshold","None",mumps->id.CNTL(3),&mumps->id.CNTL(3),NULL);CHKERRQ(ierr); 1265 ierr = PetscOptionsReal("-mat_mumps_cntl_4","CNTL(4): value for static pivoting","None",mumps->id.CNTL(4),&mumps->id.CNTL(4),NULL);CHKERRQ(ierr); 1266 ierr = PetscOptionsReal("-mat_mumps_cntl_5","CNTL(5): fixation for null pivots","None",mumps->id.CNTL(5),&mumps->id.CNTL(5),NULL);CHKERRQ(ierr); 1267 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); 1268 1269 ierr = PetscOptionsString("-mat_mumps_ooc_tmpdir", "out of core directory", "None", mumps->id.ooc_tmpdir, mumps->id.ooc_tmpdir, 256, NULL);CHKERRQ(ierr); 1270 1271 ierr = PetscOptionsIntArray("-mat_mumps_view_info","request INFO local to each processor","",info,&ninfo,NULL);CHKERRQ(ierr); 1272 if (ninfo) { 1273 if (ninfo > 40) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_USER,"number of INFO %d must <= 40\n",ninfo); 1274 ierr = PetscMalloc1(ninfo,&mumps->info);CHKERRQ(ierr); 1275 mumps->ninfo = ninfo; 1276 for (i=0; i<ninfo; i++) { 1277 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); 1278 else mumps->info[i] = info[i]; 1279 } 1280 } 1281 1282 ierr = PetscOptionsEnd();CHKERRQ(ierr); 1283 PetscFunctionReturn(0); 1284 } 1285 1286 PetscErrorCode PetscInitializeMUMPS(Mat A,Mat_MUMPS *mumps) 1287 { 1288 PetscErrorCode ierr; 1289 1290 PetscFunctionBegin; 1291 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A), &mumps->myid);CHKERRQ(ierr); 1292 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)A),&mumps->size);CHKERRQ(ierr); 1293 ierr = MPI_Comm_dup(PetscObjectComm((PetscObject)A),&(mumps->comm_mumps));CHKERRQ(ierr); 1294 1295 mumps->id.comm_fortran = MPI_Comm_c2f(mumps->comm_mumps); 1296 1297 mumps->id.job = JOB_INIT; 1298 mumps->id.par = 1; /* host participates factorizaton and solve */ 1299 mumps->id.sym = mumps->sym; 1300 PetscMUMPS_c(&mumps->id); 1301 1302 mumps->scat_rhs = NULL; 1303 mumps->scat_sol = NULL; 1304 1305 /* set PETSc-MUMPS default options - override MUMPS default */ 1306 mumps->id.ICNTL(3) = 0; 1307 mumps->id.ICNTL(4) = 0; 1308 if (mumps->size == 1) { 1309 mumps->id.ICNTL(18) = 0; /* centralized assembled matrix input */ 1310 } else { 1311 mumps->id.ICNTL(18) = 3; /* distributed assembled matrix input */ 1312 mumps->id.ICNTL(20) = 0; /* rhs is in dense format */ 1313 mumps->id.ICNTL(21) = 1; /* distributed solution */ 1314 } 1315 1316 /* schur */ 1317 mumps->id.size_schur = 0; 1318 mumps->id.listvar_schur = NULL; 1319 mumps->id.schur = NULL; 1320 mumps->sizeredrhs = 0; 1321 mumps->schur_sol = NULL; 1322 mumps->schur_sizesol = 0; 1323 PetscFunctionReturn(0); 1324 } 1325 1326 PetscErrorCode MatFactorSymbolic_MUMPS_ReportIfError(Mat F,Mat A,const MatFactorInfo *info,Mat_MUMPS *mumps) 1327 { 1328 PetscErrorCode ierr; 1329 1330 PetscFunctionBegin; 1331 if (mumps->id.INFOG(1) < 0) { 1332 if (A->erroriffailure) { 1333 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in analysis phase: INFOG(1)=%d\n",mumps->id.INFOG(1)); 1334 } else { 1335 if (mumps->id.INFOG(1) == -6) { 1336 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); 1337 F->factorerrortype = MAT_FACTOR_STRUCT_ZEROPIVOT; 1338 } else if (mumps->id.INFOG(1) == -5 || mumps->id.INFOG(1) == -7) { 1339 ierr = PetscInfo2(F,"problem of workspace, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));CHKERRQ(ierr); 1340 F->factorerrortype = MAT_FACTOR_OUTMEMORY; 1341 } else { 1342 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); 1343 F->factorerrortype = MAT_FACTOR_OTHER; 1344 } 1345 } 1346 } 1347 PetscFunctionReturn(0); 1348 } 1349 1350 /* Note Petsc r(=c) permutation is used when mumps->id.ICNTL(7)==1 with centralized assembled matrix input; otherwise r and c are ignored */ 1351 PetscErrorCode MatLUFactorSymbolic_AIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info) 1352 { 1353 Mat_MUMPS *mumps = (Mat_MUMPS*)F->data; 1354 PetscErrorCode ierr; 1355 Vec b; 1356 IS is_iden; 1357 const PetscInt M = A->rmap->N; 1358 1359 PetscFunctionBegin; 1360 mumps->matstruc = DIFFERENT_NONZERO_PATTERN; 1361 1362 /* Set MUMPS options from the options database */ 1363 ierr = PetscSetMUMPSFromOptions(F,A);CHKERRQ(ierr); 1364 1365 ierr = (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);CHKERRQ(ierr); 1366 1367 /* analysis phase */ 1368 /*----------------*/ 1369 mumps->id.job = JOB_FACTSYMBOLIC; 1370 mumps->id.n = M; 1371 switch (mumps->id.ICNTL(18)) { 1372 case 0: /* centralized assembled matrix input */ 1373 if (!mumps->myid) { 1374 mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn; 1375 if (mumps->id.ICNTL(6)>1) { 1376 mumps->id.a = (MumpsScalar*)mumps->val; 1377 } 1378 if (mumps->id.ICNTL(7) == 1) { /* use user-provide matrix ordering - assuming r = c ordering */ 1379 /* 1380 PetscBool flag; 1381 ierr = ISEqual(r,c,&flag);CHKERRQ(ierr); 1382 if (!flag) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"row_perm != col_perm"); 1383 ierr = ISView(r,PETSC_VIEWER_STDOUT_SELF); 1384 */ 1385 if (!mumps->myid) { 1386 const PetscInt *idx; 1387 PetscInt i,*perm_in; 1388 1389 ierr = PetscMalloc1(M,&perm_in);CHKERRQ(ierr); 1390 ierr = ISGetIndices(r,&idx);CHKERRQ(ierr); 1391 1392 mumps->id.perm_in = perm_in; 1393 for (i=0; i<M; i++) perm_in[i] = idx[i]+1; /* perm_in[]: start from 1, not 0! */ 1394 ierr = ISRestoreIndices(r,&idx);CHKERRQ(ierr); 1395 } 1396 } 1397 } 1398 break; 1399 case 3: /* distributed assembled matrix input (size>1) */ 1400 mumps->id.nz_loc = mumps->nz; 1401 mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn; 1402 if (mumps->id.ICNTL(6)>1) { 1403 mumps->id.a_loc = (MumpsScalar*)mumps->val; 1404 } 1405 /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */ 1406 if (!mumps->myid) { 1407 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->N,&mumps->b_seq);CHKERRQ(ierr); 1408 ierr = ISCreateStride(PETSC_COMM_SELF,A->rmap->N,0,1,&is_iden);CHKERRQ(ierr); 1409 } else { 1410 ierr = VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);CHKERRQ(ierr); 1411 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);CHKERRQ(ierr); 1412 } 1413 ierr = MatCreateVecs(A,NULL,&b);CHKERRQ(ierr); 1414 ierr = VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);CHKERRQ(ierr); 1415 ierr = ISDestroy(&is_iden);CHKERRQ(ierr); 1416 ierr = VecDestroy(&b);CHKERRQ(ierr); 1417 break; 1418 } 1419 PetscMUMPS_c(&mumps->id); 1420 ierr = MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);CHKERRQ(ierr); 1421 1422 F->ops->lufactornumeric = MatFactorNumeric_MUMPS; 1423 F->ops->solve = MatSolve_MUMPS; 1424 F->ops->solvetranspose = MatSolveTranspose_MUMPS; 1425 F->ops->matsolve = MatMatSolve_MUMPS; 1426 F->ops->mattransposesolve = MatMatTransposeSolve_MUMPS; 1427 PetscFunctionReturn(0); 1428 } 1429 1430 /* Note the Petsc r and c permutations are ignored */ 1431 PetscErrorCode MatLUFactorSymbolic_BAIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info) 1432 { 1433 Mat_MUMPS *mumps = (Mat_MUMPS*)F->data; 1434 PetscErrorCode ierr; 1435 Vec b; 1436 IS is_iden; 1437 const PetscInt M = A->rmap->N; 1438 1439 PetscFunctionBegin; 1440 mumps->matstruc = DIFFERENT_NONZERO_PATTERN; 1441 1442 /* Set MUMPS options from the options database */ 1443 ierr = PetscSetMUMPSFromOptions(F,A);CHKERRQ(ierr); 1444 1445 ierr = (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);CHKERRQ(ierr); 1446 1447 /* analysis phase */ 1448 /*----------------*/ 1449 mumps->id.job = JOB_FACTSYMBOLIC; 1450 mumps->id.n = M; 1451 switch (mumps->id.ICNTL(18)) { 1452 case 0: /* centralized assembled matrix input */ 1453 if (!mumps->myid) { 1454 mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn; 1455 if (mumps->id.ICNTL(6)>1) { 1456 mumps->id.a = (MumpsScalar*)mumps->val; 1457 } 1458 } 1459 break; 1460 case 3: /* distributed assembled matrix input (size>1) */ 1461 mumps->id.nz_loc = mumps->nz; 1462 mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn; 1463 if (mumps->id.ICNTL(6)>1) { 1464 mumps->id.a_loc = (MumpsScalar*)mumps->val; 1465 } 1466 /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */ 1467 if (!mumps->myid) { 1468 ierr = VecCreateSeq(PETSC_COMM_SELF,A->cmap->N,&mumps->b_seq);CHKERRQ(ierr); 1469 ierr = ISCreateStride(PETSC_COMM_SELF,A->cmap->N,0,1,&is_iden);CHKERRQ(ierr); 1470 } else { 1471 ierr = VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);CHKERRQ(ierr); 1472 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);CHKERRQ(ierr); 1473 } 1474 ierr = MatCreateVecs(A,NULL,&b);CHKERRQ(ierr); 1475 ierr = VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);CHKERRQ(ierr); 1476 ierr = ISDestroy(&is_iden);CHKERRQ(ierr); 1477 ierr = VecDestroy(&b);CHKERRQ(ierr); 1478 break; 1479 } 1480 PetscMUMPS_c(&mumps->id); 1481 ierr = MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);CHKERRQ(ierr); 1482 1483 F->ops->lufactornumeric = MatFactorNumeric_MUMPS; 1484 F->ops->solve = MatSolve_MUMPS; 1485 F->ops->solvetranspose = MatSolveTranspose_MUMPS; 1486 PetscFunctionReturn(0); 1487 } 1488 1489 /* Note the Petsc r permutation and factor info are ignored */ 1490 PetscErrorCode MatCholeskyFactorSymbolic_MUMPS(Mat F,Mat A,IS r,const MatFactorInfo *info) 1491 { 1492 Mat_MUMPS *mumps = (Mat_MUMPS*)F->data; 1493 PetscErrorCode ierr; 1494 Vec b; 1495 IS is_iden; 1496 const PetscInt M = A->rmap->N; 1497 1498 PetscFunctionBegin; 1499 mumps->matstruc = DIFFERENT_NONZERO_PATTERN; 1500 1501 /* Set MUMPS options from the options database */ 1502 ierr = PetscSetMUMPSFromOptions(F,A);CHKERRQ(ierr); 1503 1504 ierr = (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);CHKERRQ(ierr); 1505 1506 /* analysis phase */ 1507 /*----------------*/ 1508 mumps->id.job = JOB_FACTSYMBOLIC; 1509 mumps->id.n = M; 1510 switch (mumps->id.ICNTL(18)) { 1511 case 0: /* centralized assembled matrix input */ 1512 if (!mumps->myid) { 1513 mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn; 1514 if (mumps->id.ICNTL(6)>1) { 1515 mumps->id.a = (MumpsScalar*)mumps->val; 1516 } 1517 } 1518 break; 1519 case 3: /* distributed assembled matrix input (size>1) */ 1520 mumps->id.nz_loc = mumps->nz; 1521 mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn; 1522 if (mumps->id.ICNTL(6)>1) { 1523 mumps->id.a_loc = (MumpsScalar*)mumps->val; 1524 } 1525 /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */ 1526 if (!mumps->myid) { 1527 ierr = VecCreateSeq(PETSC_COMM_SELF,A->cmap->N,&mumps->b_seq);CHKERRQ(ierr); 1528 ierr = ISCreateStride(PETSC_COMM_SELF,A->cmap->N,0,1,&is_iden);CHKERRQ(ierr); 1529 } else { 1530 ierr = VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);CHKERRQ(ierr); 1531 ierr = ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);CHKERRQ(ierr); 1532 } 1533 ierr = MatCreateVecs(A,NULL,&b);CHKERRQ(ierr); 1534 ierr = VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);CHKERRQ(ierr); 1535 ierr = ISDestroy(&is_iden);CHKERRQ(ierr); 1536 ierr = VecDestroy(&b);CHKERRQ(ierr); 1537 break; 1538 } 1539 PetscMUMPS_c(&mumps->id); 1540 ierr = MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);CHKERRQ(ierr); 1541 1542 F->ops->choleskyfactornumeric = MatFactorNumeric_MUMPS; 1543 F->ops->solve = MatSolve_MUMPS; 1544 F->ops->solvetranspose = MatSolve_MUMPS; 1545 F->ops->matsolve = MatMatSolve_MUMPS; 1546 #if defined(PETSC_USE_COMPLEX) 1547 F->ops->getinertia = NULL; 1548 #else 1549 F->ops->getinertia = MatGetInertia_SBAIJMUMPS; 1550 #endif 1551 PetscFunctionReturn(0); 1552 } 1553 1554 PetscErrorCode MatView_MUMPS(Mat A,PetscViewer viewer) 1555 { 1556 PetscErrorCode ierr; 1557 PetscBool iascii; 1558 PetscViewerFormat format; 1559 Mat_MUMPS *mumps=(Mat_MUMPS*)A->data; 1560 1561 PetscFunctionBegin; 1562 /* check if matrix is mumps type */ 1563 if (A->ops->solve != MatSolve_MUMPS) PetscFunctionReturn(0); 1564 1565 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1566 if (iascii) { 1567 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1568 if (format == PETSC_VIEWER_ASCII_INFO) { 1569 ierr = PetscViewerASCIIPrintf(viewer,"MUMPS run parameters:\n");CHKERRQ(ierr); 1570 ierr = PetscViewerASCIIPrintf(viewer," SYM (matrix type): %d \n",mumps->id.sym);CHKERRQ(ierr); 1571 ierr = PetscViewerASCIIPrintf(viewer," PAR (host participation): %d \n",mumps->id.par);CHKERRQ(ierr); 1572 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(1) (output for error): %d \n",mumps->id.ICNTL(1));CHKERRQ(ierr); 1573 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(2) (output of diagnostic msg): %d \n",mumps->id.ICNTL(2));CHKERRQ(ierr); 1574 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(3) (output for global info): %d \n",mumps->id.ICNTL(3));CHKERRQ(ierr); 1575 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(4) (level of printing): %d \n",mumps->id.ICNTL(4));CHKERRQ(ierr); 1576 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(5) (input mat struct): %d \n",mumps->id.ICNTL(5));CHKERRQ(ierr); 1577 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(6) (matrix prescaling): %d \n",mumps->id.ICNTL(6));CHKERRQ(ierr); 1578 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(7) (sequential matrix ordering):%d \n",mumps->id.ICNTL(7));CHKERRQ(ierr); 1579 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(8) (scaling strategy): %d \n",mumps->id.ICNTL(8));CHKERRQ(ierr); 1580 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(10) (max num of refinements): %d \n",mumps->id.ICNTL(10));CHKERRQ(ierr); 1581 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(11) (error analysis): %d \n",mumps->id.ICNTL(11));CHKERRQ(ierr); 1582 if (mumps->id.ICNTL(11)>0) { 1583 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(4) (inf norm of input mat): %g\n",mumps->id.RINFOG(4));CHKERRQ(ierr); 1584 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(5) (inf norm of solution): %g\n",mumps->id.RINFOG(5));CHKERRQ(ierr); 1585 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(6) (inf norm of residual): %g\n",mumps->id.RINFOG(6));CHKERRQ(ierr); 1586 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(7),RINFOG(8) (backward error est): %g, %g\n",mumps->id.RINFOG(7),mumps->id.RINFOG(8));CHKERRQ(ierr); 1587 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(9) (error estimate): %g \n",mumps->id.RINFOG(9));CHKERRQ(ierr); 1588 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(10),RINFOG(11)(condition numbers): %g, %g\n",mumps->id.RINFOG(10),mumps->id.RINFOG(11));CHKERRQ(ierr); 1589 } 1590 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(12) (efficiency control): %d \n",mumps->id.ICNTL(12));CHKERRQ(ierr); 1591 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(13) (efficiency control): %d \n",mumps->id.ICNTL(13));CHKERRQ(ierr); 1592 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(14) (percentage of estimated workspace increase): %d \n",mumps->id.ICNTL(14));CHKERRQ(ierr); 1593 /* ICNTL(15-17) not used */ 1594 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(18) (input mat struct): %d \n",mumps->id.ICNTL(18));CHKERRQ(ierr); 1595 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(19) (Schur complement info): %d \n",mumps->id.ICNTL(19));CHKERRQ(ierr); 1596 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(20) (rhs sparse pattern): %d \n",mumps->id.ICNTL(20));CHKERRQ(ierr); 1597 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(21) (solution struct): %d \n",mumps->id.ICNTL(21));CHKERRQ(ierr); 1598 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(22) (in-core/out-of-core facility): %d \n",mumps->id.ICNTL(22));CHKERRQ(ierr); 1599 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(23) (max size of memory can be allocated locally):%d \n",mumps->id.ICNTL(23));CHKERRQ(ierr); 1600 1601 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(24) (detection of null pivot rows): %d \n",mumps->id.ICNTL(24));CHKERRQ(ierr); 1602 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(25) (computation of a null space basis): %d \n",mumps->id.ICNTL(25));CHKERRQ(ierr); 1603 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(26) (Schur options for rhs or solution): %d \n",mumps->id.ICNTL(26));CHKERRQ(ierr); 1604 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(27) (experimental parameter): %d \n",mumps->id.ICNTL(27));CHKERRQ(ierr); 1605 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(28) (use parallel or sequential ordering): %d \n",mumps->id.ICNTL(28));CHKERRQ(ierr); 1606 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(29) (parallel ordering): %d \n",mumps->id.ICNTL(29));CHKERRQ(ierr); 1607 1608 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(30) (user-specified set of entries in inv(A)): %d \n",mumps->id.ICNTL(30));CHKERRQ(ierr); 1609 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(31) (factors is discarded in the solve phase): %d \n",mumps->id.ICNTL(31));CHKERRQ(ierr); 1610 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(33) (compute determinant): %d \n",mumps->id.ICNTL(33));CHKERRQ(ierr); 1611 ierr = PetscViewerASCIIPrintf(viewer," ICNTL(35) (activate BLR based factorization): %d \n",mumps->id.ICNTL(35));CHKERRQ(ierr); 1612 1613 ierr = PetscViewerASCIIPrintf(viewer," CNTL(1) (relative pivoting threshold): %g \n",mumps->id.CNTL(1));CHKERRQ(ierr); 1614 ierr = PetscViewerASCIIPrintf(viewer," CNTL(2) (stopping criterion of refinement): %g \n",mumps->id.CNTL(2));CHKERRQ(ierr); 1615 ierr = PetscViewerASCIIPrintf(viewer," CNTL(3) (absolute pivoting threshold): %g \n",mumps->id.CNTL(3));CHKERRQ(ierr); 1616 ierr = PetscViewerASCIIPrintf(viewer," CNTL(4) (value of static pivoting): %g \n",mumps->id.CNTL(4));CHKERRQ(ierr); 1617 ierr = PetscViewerASCIIPrintf(viewer," CNTL(5) (fixation for null pivots): %g \n",mumps->id.CNTL(5));CHKERRQ(ierr); 1618 ierr = PetscViewerASCIIPrintf(viewer," CNTL(7) (dropping parameter for BLR): %g \n",mumps->id.CNTL(7));CHKERRQ(ierr); 1619 1620 /* infomation local to each processor */ 1621 ierr = PetscViewerASCIIPrintf(viewer, " RINFO(1) (local estimated flops for the elimination after analysis): \n");CHKERRQ(ierr); 1622 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1623 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %g \n",mumps->myid,mumps->id.RINFO(1));CHKERRQ(ierr); 1624 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1625 ierr = PetscViewerASCIIPrintf(viewer, " RINFO(2) (local estimated flops for the assembly after factorization): \n");CHKERRQ(ierr); 1626 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %g \n",mumps->myid,mumps->id.RINFO(2));CHKERRQ(ierr); 1627 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1628 ierr = PetscViewerASCIIPrintf(viewer, " RINFO(3) (local estimated flops for the elimination after factorization): \n");CHKERRQ(ierr); 1629 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %g \n",mumps->myid,mumps->id.RINFO(3));CHKERRQ(ierr); 1630 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1631 1632 ierr = PetscViewerASCIIPrintf(viewer, " INFO(15) (estimated size of (in MB) MUMPS internal data for running numerical factorization): \n");CHKERRQ(ierr); 1633 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(15));CHKERRQ(ierr); 1634 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1635 1636 ierr = PetscViewerASCIIPrintf(viewer, " INFO(16) (size of (in MB) MUMPS internal data used during numerical factorization): \n");CHKERRQ(ierr); 1637 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(16));CHKERRQ(ierr); 1638 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1639 1640 ierr = PetscViewerASCIIPrintf(viewer, " INFO(23) (num of pivots eliminated on this processor after factorization): \n");CHKERRQ(ierr); 1641 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(23));CHKERRQ(ierr); 1642 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1643 1644 if (mumps->ninfo && mumps->ninfo <= 40){ 1645 PetscInt i; 1646 for (i=0; i<mumps->ninfo; i++){ 1647 ierr = PetscViewerASCIIPrintf(viewer, " INFO(%d): \n",mumps->info[i]);CHKERRQ(ierr); 1648 ierr = PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(mumps->info[i]));CHKERRQ(ierr); 1649 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1650 } 1651 } 1652 1653 1654 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1655 1656 if (!mumps->myid) { /* information from the host */ 1657 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(1) (global estimated flops for the elimination after analysis): %g \n",mumps->id.RINFOG(1));CHKERRQ(ierr); 1658 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(2) (global estimated flops for the assembly after factorization): %g \n",mumps->id.RINFOG(2));CHKERRQ(ierr); 1659 ierr = PetscViewerASCIIPrintf(viewer," RINFOG(3) (global estimated flops for the elimination after factorization): %g \n",mumps->id.RINFOG(3));CHKERRQ(ierr); 1660 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); 1661 1662 ierr = PetscViewerASCIIPrintf(viewer," INFOG(3) (estimated real workspace for factors on all processors after analysis): %d \n",mumps->id.INFOG(3));CHKERRQ(ierr); 1663 ierr = PetscViewerASCIIPrintf(viewer," INFOG(4) (estimated integer workspace for factors on all processors after analysis): %d \n",mumps->id.INFOG(4));CHKERRQ(ierr); 1664 ierr = PetscViewerASCIIPrintf(viewer," INFOG(5) (estimated maximum front size in the complete tree): %d \n",mumps->id.INFOG(5));CHKERRQ(ierr); 1665 ierr = PetscViewerASCIIPrintf(viewer," INFOG(6) (number of nodes in the complete tree): %d \n",mumps->id.INFOG(6));CHKERRQ(ierr); 1666 ierr = PetscViewerASCIIPrintf(viewer," INFOG(7) (ordering option effectively use after analysis): %d \n",mumps->id.INFOG(7));CHKERRQ(ierr); 1667 ierr = PetscViewerASCIIPrintf(viewer," INFOG(8) (structural symmetry in percent of the permuted matrix after analysis): %d \n",mumps->id.INFOG(8));CHKERRQ(ierr); 1668 ierr = PetscViewerASCIIPrintf(viewer," INFOG(9) (total real/complex workspace to store the matrix factors after factorization): %d \n",mumps->id.INFOG(9));CHKERRQ(ierr); 1669 ierr = PetscViewerASCIIPrintf(viewer," INFOG(10) (total integer space store the matrix factors after factorization): %d \n",mumps->id.INFOG(10));CHKERRQ(ierr); 1670 ierr = PetscViewerASCIIPrintf(viewer," INFOG(11) (order of largest frontal matrix after factorization): %d \n",mumps->id.INFOG(11));CHKERRQ(ierr); 1671 ierr = PetscViewerASCIIPrintf(viewer," INFOG(12) (number of off-diagonal pivots): %d \n",mumps->id.INFOG(12));CHKERRQ(ierr); 1672 ierr = PetscViewerASCIIPrintf(viewer," INFOG(13) (number of delayed pivots after factorization): %d \n",mumps->id.INFOG(13));CHKERRQ(ierr); 1673 ierr = PetscViewerASCIIPrintf(viewer," INFOG(14) (number of memory compress after factorization): %d \n",mumps->id.INFOG(14));CHKERRQ(ierr); 1674 ierr = PetscViewerASCIIPrintf(viewer," INFOG(15) (number of steps of iterative refinement after solution): %d \n",mumps->id.INFOG(15));CHKERRQ(ierr); 1675 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); 1676 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); 1677 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); 1678 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); 1679 ierr = PetscViewerASCIIPrintf(viewer," INFOG(20) (estimated number of entries in the factors): %d \n",mumps->id.INFOG(20));CHKERRQ(ierr); 1680 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); 1681 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); 1682 ierr = PetscViewerASCIIPrintf(viewer," INFOG(23) (after analysis: value of ICNTL(6) effectively used): %d \n",mumps->id.INFOG(23));CHKERRQ(ierr); 1683 ierr = PetscViewerASCIIPrintf(viewer," INFOG(24) (after analysis: value of ICNTL(12) effectively used): %d \n",mumps->id.INFOG(24));CHKERRQ(ierr); 1684 ierr = PetscViewerASCIIPrintf(viewer," INFOG(25) (after factorization: number of pivots modified by static pivoting): %d \n",mumps->id.INFOG(25));CHKERRQ(ierr); 1685 ierr = PetscViewerASCIIPrintf(viewer," INFOG(28) (after factorization: number of null pivots encountered): %d\n",mumps->id.INFOG(28));CHKERRQ(ierr); 1686 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); 1687 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); 1688 ierr = PetscViewerASCIIPrintf(viewer," INFOG(32) (after analysis: type of analysis done): %d\n",mumps->id.INFOG(32));CHKERRQ(ierr); 1689 ierr = PetscViewerASCIIPrintf(viewer," INFOG(33) (value used for ICNTL(8)): %d\n",mumps->id.INFOG(33));CHKERRQ(ierr); 1690 ierr = PetscViewerASCIIPrintf(viewer," INFOG(34) (exponent of the determinant if determinant is requested): %d\n",mumps->id.INFOG(34));CHKERRQ(ierr); 1691 } 1692 } 1693 } 1694 PetscFunctionReturn(0); 1695 } 1696 1697 PetscErrorCode MatGetInfo_MUMPS(Mat A,MatInfoType flag,MatInfo *info) 1698 { 1699 Mat_MUMPS *mumps =(Mat_MUMPS*)A->data; 1700 1701 PetscFunctionBegin; 1702 info->block_size = 1.0; 1703 info->nz_allocated = mumps->id.INFOG(20); 1704 info->nz_used = mumps->id.INFOG(20); 1705 info->nz_unneeded = 0.0; 1706 info->assemblies = 0.0; 1707 info->mallocs = 0.0; 1708 info->memory = 0.0; 1709 info->fill_ratio_given = 0; 1710 info->fill_ratio_needed = 0; 1711 info->factor_mallocs = 0; 1712 PetscFunctionReturn(0); 1713 } 1714 1715 /* -------------------------------------------------------------------------------------------*/ 1716 PetscErrorCode MatFactorSetSchurIS_MUMPS(Mat F, IS is) 1717 { 1718 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1719 const PetscInt *idxs; 1720 PetscInt size,i; 1721 PetscErrorCode ierr; 1722 1723 PetscFunctionBegin; 1724 ierr = ISGetLocalSize(is,&size);CHKERRQ(ierr); 1725 if (mumps->size > 1) { 1726 PetscBool ls,gs; 1727 1728 ls = mumps->myid ? (size ? PETSC_FALSE : PETSC_TRUE) : PETSC_TRUE; 1729 ierr = MPI_Allreduce(&ls,&gs,1,MPIU_BOOL,MPI_LAND,mumps->comm_mumps);CHKERRQ(ierr); 1730 if (!gs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MUMPS distributed parallel Schur complements not yet supported from PETSc\n"); 1731 } 1732 if (mumps->id.size_schur != size) { 1733 ierr = PetscFree2(mumps->id.listvar_schur,mumps->id.schur);CHKERRQ(ierr); 1734 mumps->id.size_schur = size; 1735 mumps->id.schur_lld = size; 1736 ierr = PetscMalloc2(size,&mumps->id.listvar_schur,size*size,&mumps->id.schur);CHKERRQ(ierr); 1737 } 1738 1739 /* Schur complement matrix */ 1740 ierr = MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.size_schur,(PetscScalar*)mumps->id.schur,&F->schur);CHKERRQ(ierr); 1741 if (mumps->sym == 1) { 1742 ierr = MatSetOption(F->schur,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 1743 } 1744 1745 /* MUMPS expects Fortran style indices */ 1746 ierr = ISGetIndices(is,&idxs);CHKERRQ(ierr); 1747 ierr = PetscMemcpy(mumps->id.listvar_schur,idxs,size*sizeof(PetscInt));CHKERRQ(ierr); 1748 for (i=0;i<size;i++) mumps->id.listvar_schur[i]++; 1749 ierr = ISRestoreIndices(is,&idxs);CHKERRQ(ierr); 1750 if (mumps->size > 1) { 1751 mumps->id.ICNTL(19) = 1; /* MUMPS returns Schur centralized on the host */ 1752 } else { 1753 if (F->factortype == MAT_FACTOR_LU) { 1754 mumps->id.ICNTL(19) = 3; /* MUMPS returns full matrix */ 1755 } else { 1756 mumps->id.ICNTL(19) = 2; /* MUMPS returns lower triangular part */ 1757 } 1758 } 1759 /* set a special value of ICNTL (not handled my MUMPS) to be used in the solve phase by PETSc */ 1760 mumps->id.ICNTL(26) = -1; 1761 PetscFunctionReturn(0); 1762 } 1763 1764 /* -------------------------------------------------------------------------------------------*/ 1765 PetscErrorCode MatFactorCreateSchurComplement_MUMPS(Mat F,Mat* S) 1766 { 1767 Mat St; 1768 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1769 PetscScalar *array; 1770 #if defined(PETSC_USE_COMPLEX) 1771 PetscScalar im = PetscSqrtScalar((PetscScalar)-1.0); 1772 #endif 1773 PetscErrorCode ierr; 1774 1775 PetscFunctionBegin; 1776 if (!mumps->id.ICNTL(19)) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ORDER,"Schur complement mode not selected! You should call MatFactorSetSchurIS to enable it"); 1777 ierr = MatCreate(PETSC_COMM_SELF,&St);CHKERRQ(ierr); 1778 ierr = MatSetSizes(St,PETSC_DECIDE,PETSC_DECIDE,mumps->id.size_schur,mumps->id.size_schur);CHKERRQ(ierr); 1779 ierr = MatSetType(St,MATDENSE);CHKERRQ(ierr); 1780 ierr = MatSetUp(St);CHKERRQ(ierr); 1781 ierr = MatDenseGetArray(St,&array);CHKERRQ(ierr); 1782 if (!mumps->sym) { /* MUMPS always return a full matrix */ 1783 if (mumps->id.ICNTL(19) == 1) { /* stored by rows */ 1784 PetscInt i,j,N=mumps->id.size_schur; 1785 for (i=0;i<N;i++) { 1786 for (j=0;j<N;j++) { 1787 #if !defined(PETSC_USE_COMPLEX) 1788 PetscScalar val = mumps->id.schur[i*N+j]; 1789 #else 1790 PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i; 1791 #endif 1792 array[j*N+i] = val; 1793 } 1794 } 1795 } else { /* stored by columns */ 1796 ierr = PetscMemcpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur*sizeof(PetscScalar));CHKERRQ(ierr); 1797 } 1798 } else { /* either full or lower-triangular (not packed) */ 1799 if (mumps->id.ICNTL(19) == 2) { /* lower triangular stored by columns */ 1800 PetscInt i,j,N=mumps->id.size_schur; 1801 for (i=0;i<N;i++) { 1802 for (j=i;j<N;j++) { 1803 #if !defined(PETSC_USE_COMPLEX) 1804 PetscScalar val = mumps->id.schur[i*N+j]; 1805 #else 1806 PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i; 1807 #endif 1808 array[i*N+j] = val; 1809 array[j*N+i] = val; 1810 } 1811 } 1812 } else if (mumps->id.ICNTL(19) == 3) { /* full matrix */ 1813 ierr = PetscMemcpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur*sizeof(PetscScalar));CHKERRQ(ierr); 1814 } else { /* ICNTL(19) == 1 lower triangular stored by rows */ 1815 PetscInt i,j,N=mumps->id.size_schur; 1816 for (i=0;i<N;i++) { 1817 for (j=0;j<i+1;j++) { 1818 #if !defined(PETSC_USE_COMPLEX) 1819 PetscScalar val = mumps->id.schur[i*N+j]; 1820 #else 1821 PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i; 1822 #endif 1823 array[i*N+j] = val; 1824 array[j*N+i] = val; 1825 } 1826 } 1827 } 1828 } 1829 ierr = MatDenseRestoreArray(St,&array);CHKERRQ(ierr); 1830 *S = St; 1831 PetscFunctionReturn(0); 1832 } 1833 1834 /* -------------------------------------------------------------------------------------------*/ 1835 PetscErrorCode MatMumpsSetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt ival) 1836 { 1837 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1838 1839 PetscFunctionBegin; 1840 mumps->id.ICNTL(icntl) = ival; 1841 PetscFunctionReturn(0); 1842 } 1843 1844 PetscErrorCode MatMumpsGetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt *ival) 1845 { 1846 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1847 1848 PetscFunctionBegin; 1849 *ival = mumps->id.ICNTL(icntl); 1850 PetscFunctionReturn(0); 1851 } 1852 1853 /*@ 1854 MatMumpsSetIcntl - Set MUMPS parameter ICNTL() 1855 1856 Logically Collective on Mat 1857 1858 Input Parameters: 1859 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 1860 . icntl - index of MUMPS parameter array ICNTL() 1861 - ival - value of MUMPS ICNTL(icntl) 1862 1863 Options Database: 1864 . -mat_mumps_icntl_<icntl> <ival> 1865 1866 Level: beginner 1867 1868 References: 1869 . MUMPS Users' Guide 1870 1871 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 1872 @*/ 1873 PetscErrorCode MatMumpsSetIcntl(Mat F,PetscInt icntl,PetscInt ival) 1874 { 1875 PetscErrorCode ierr; 1876 1877 PetscFunctionBegin; 1878 PetscValidType(F,1); 1879 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 1880 PetscValidLogicalCollectiveInt(F,icntl,2); 1881 PetscValidLogicalCollectiveInt(F,ival,3); 1882 ierr = PetscTryMethod(F,"MatMumpsSetIcntl_C",(Mat,PetscInt,PetscInt),(F,icntl,ival));CHKERRQ(ierr); 1883 PetscFunctionReturn(0); 1884 } 1885 1886 /*@ 1887 MatMumpsGetIcntl - Get MUMPS parameter ICNTL() 1888 1889 Logically Collective on Mat 1890 1891 Input Parameters: 1892 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 1893 - icntl - index of MUMPS parameter array ICNTL() 1894 1895 Output Parameter: 1896 . ival - value of MUMPS ICNTL(icntl) 1897 1898 Level: beginner 1899 1900 References: 1901 . MUMPS Users' Guide 1902 1903 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 1904 @*/ 1905 PetscErrorCode MatMumpsGetIcntl(Mat F,PetscInt icntl,PetscInt *ival) 1906 { 1907 PetscErrorCode ierr; 1908 1909 PetscFunctionBegin; 1910 PetscValidType(F,1); 1911 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 1912 PetscValidLogicalCollectiveInt(F,icntl,2); 1913 PetscValidIntPointer(ival,3); 1914 ierr = PetscUseMethod(F,"MatMumpsGetIcntl_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));CHKERRQ(ierr); 1915 PetscFunctionReturn(0); 1916 } 1917 1918 /* -------------------------------------------------------------------------------------------*/ 1919 PetscErrorCode MatMumpsSetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal val) 1920 { 1921 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1922 1923 PetscFunctionBegin; 1924 mumps->id.CNTL(icntl) = val; 1925 PetscFunctionReturn(0); 1926 } 1927 1928 PetscErrorCode MatMumpsGetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal *val) 1929 { 1930 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 1931 1932 PetscFunctionBegin; 1933 *val = mumps->id.CNTL(icntl); 1934 PetscFunctionReturn(0); 1935 } 1936 1937 /*@ 1938 MatMumpsSetCntl - Set MUMPS parameter CNTL() 1939 1940 Logically Collective on Mat 1941 1942 Input Parameters: 1943 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 1944 . icntl - index of MUMPS parameter array CNTL() 1945 - val - value of MUMPS CNTL(icntl) 1946 1947 Options Database: 1948 . -mat_mumps_cntl_<icntl> <val> 1949 1950 Level: beginner 1951 1952 References: 1953 . MUMPS Users' Guide 1954 1955 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 1956 @*/ 1957 PetscErrorCode MatMumpsSetCntl(Mat F,PetscInt icntl,PetscReal val) 1958 { 1959 PetscErrorCode ierr; 1960 1961 PetscFunctionBegin; 1962 PetscValidType(F,1); 1963 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 1964 PetscValidLogicalCollectiveInt(F,icntl,2); 1965 PetscValidLogicalCollectiveReal(F,val,3); 1966 ierr = PetscTryMethod(F,"MatMumpsSetCntl_C",(Mat,PetscInt,PetscReal),(F,icntl,val));CHKERRQ(ierr); 1967 PetscFunctionReturn(0); 1968 } 1969 1970 /*@ 1971 MatMumpsGetCntl - Get MUMPS parameter CNTL() 1972 1973 Logically Collective on Mat 1974 1975 Input Parameters: 1976 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 1977 - icntl - index of MUMPS parameter array CNTL() 1978 1979 Output Parameter: 1980 . val - value of MUMPS CNTL(icntl) 1981 1982 Level: beginner 1983 1984 References: 1985 . MUMPS Users' Guide 1986 1987 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 1988 @*/ 1989 PetscErrorCode MatMumpsGetCntl(Mat F,PetscInt icntl,PetscReal *val) 1990 { 1991 PetscErrorCode ierr; 1992 1993 PetscFunctionBegin; 1994 PetscValidType(F,1); 1995 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 1996 PetscValidLogicalCollectiveInt(F,icntl,2); 1997 PetscValidRealPointer(val,3); 1998 ierr = PetscUseMethod(F,"MatMumpsGetCntl_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));CHKERRQ(ierr); 1999 PetscFunctionReturn(0); 2000 } 2001 2002 PetscErrorCode MatMumpsGetInfo_MUMPS(Mat F,PetscInt icntl,PetscInt *info) 2003 { 2004 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 2005 2006 PetscFunctionBegin; 2007 *info = mumps->id.INFO(icntl); 2008 PetscFunctionReturn(0); 2009 } 2010 2011 PetscErrorCode MatMumpsGetInfog_MUMPS(Mat F,PetscInt icntl,PetscInt *infog) 2012 { 2013 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 2014 2015 PetscFunctionBegin; 2016 *infog = mumps->id.INFOG(icntl); 2017 PetscFunctionReturn(0); 2018 } 2019 2020 PetscErrorCode MatMumpsGetRinfo_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfo) 2021 { 2022 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 2023 2024 PetscFunctionBegin; 2025 *rinfo = mumps->id.RINFO(icntl); 2026 PetscFunctionReturn(0); 2027 } 2028 2029 PetscErrorCode MatMumpsGetRinfog_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfog) 2030 { 2031 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 2032 2033 PetscFunctionBegin; 2034 *rinfog = mumps->id.RINFOG(icntl); 2035 PetscFunctionReturn(0); 2036 } 2037 2038 PetscErrorCode MatMumpsGetInverse_MUMPS(Mat F,Mat spRHS) 2039 { 2040 PetscErrorCode ierr; 2041 Mat Bt = NULL,Btseq = NULL; 2042 PetscBool flg; 2043 Mat_MUMPS *mumps =(Mat_MUMPS*)F->data; 2044 PetscScalar *aa; 2045 PetscInt spnr,*ia,*ja; 2046 2047 PetscFunctionBegin; 2048 PetscValidIntPointer(spRHS,2); 2049 ierr = PetscObjectTypeCompare((PetscObject)spRHS,MATTRANSPOSEMAT,&flg);CHKERRQ(ierr); 2050 if (flg) { 2051 ierr = MatTransposeGetMat(spRHS,&Bt);CHKERRQ(ierr); 2052 } else SETERRQ(PetscObjectComm((PetscObject)spRHS),PETSC_ERR_ARG_WRONG,"Matrix spRHS must be type MATTRANSPOSEMAT matrix"); 2053 2054 ierr = MatMumpsSetIcntl(F,30,1);CHKERRQ(ierr); 2055 2056 if (mumps->size > 1) { 2057 Mat_MPIAIJ *b = (Mat_MPIAIJ*)Bt->data; 2058 Btseq = b->A; 2059 } else { 2060 Btseq = Bt; 2061 } 2062 2063 if (!mumps->myid) { 2064 ierr = MatSeqAIJGetArray(Btseq,&aa);CHKERRQ(ierr); 2065 ierr = MatGetRowIJ(Btseq,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);CHKERRQ(ierr); 2066 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure"); 2067 2068 mumps->id.irhs_ptr = ia; 2069 mumps->id.irhs_sparse = ja; 2070 mumps->id.nz_rhs = ia[spnr] - 1; 2071 mumps->id.rhs_sparse = (MumpsScalar*)aa; 2072 } else { 2073 mumps->id.irhs_ptr = NULL; 2074 mumps->id.irhs_sparse = NULL; 2075 mumps->id.nz_rhs = 0; 2076 mumps->id.rhs_sparse = NULL; 2077 } 2078 mumps->id.ICNTL(20) = 1; /* rhs is sparse */ 2079 mumps->id.ICNTL(21) = 0; /* solution is in assembled centralized format */ 2080 2081 /* solve phase */ 2082 /*-------------*/ 2083 mumps->id.job = JOB_SOLVE; 2084 PetscMUMPS_c(&mumps->id); 2085 if (mumps->id.INFOG(1) < 0) 2086 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2)); 2087 2088 if (!mumps->myid) { 2089 ierr = MatSeqAIJRestoreArray(Btseq,&aa);CHKERRQ(ierr); 2090 ierr = MatRestoreRowIJ(Btseq,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);CHKERRQ(ierr); 2091 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure"); 2092 } 2093 PetscFunctionReturn(0); 2094 } 2095 2096 /*@ 2097 MatMumpsGetInverse - Get user-specified set of entries in inverse of A 2098 2099 Logically Collective on Mat 2100 2101 Input Parameters: 2102 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2103 - spRHS - sequential sparse matrix in MATTRANSPOSEMAT format holding specified indices in processor[0] 2104 2105 Output Parameter: 2106 . spRHS - requested entries of inverse of A 2107 2108 Level: beginner 2109 2110 References: 2111 . MUMPS Users' Guide 2112 2113 .seealso: MatGetFactor(), MatCreateTranspose() 2114 @*/ 2115 PetscErrorCode MatMumpsGetInverse(Mat F,Mat spRHS) 2116 { 2117 PetscErrorCode ierr; 2118 2119 PetscFunctionBegin; 2120 PetscValidType(F,1); 2121 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2122 ierr = PetscUseMethod(F,"MatMumpsGetInverse_C",(Mat,Mat),(F,spRHS));CHKERRQ(ierr); 2123 PetscFunctionReturn(0); 2124 } 2125 2126 PetscErrorCode MatMumpsGetInverseTranspose_MUMPS(Mat F,Mat spRHST) 2127 { 2128 PetscErrorCode ierr; 2129 Mat spRHS; 2130 2131 PetscFunctionBegin; 2132 ierr = MatCreateTranspose(spRHST,&spRHS);CHKERRQ(ierr); 2133 ierr = MatMumpsGetInverse_MUMPS(F,spRHS);CHKERRQ(ierr); 2134 ierr = MatDestroy(&spRHS);CHKERRQ(ierr); 2135 PetscFunctionReturn(0); 2136 } 2137 2138 /*@ 2139 MatMumpsGetInverseTranspose - Get user-specified set of entries in inverse of matrix A^T 2140 2141 Logically Collective on Mat 2142 2143 Input Parameters: 2144 + F - the factored matrix of A obtained by calling MatGetFactor() from PETSc-MUMPS interface 2145 - spRHST - sequential sparse matrix in MATAIJ format holding specified indices of A^T in processor[0] 2146 2147 Output Parameter: 2148 . spRHST - requested entries of inverse of A^T 2149 2150 Level: beginner 2151 2152 References: 2153 . MUMPS Users' Guide 2154 2155 .seealso: MatGetFactor(), MatCreateTranspose(), MatMumpsGetInverse() 2156 @*/ 2157 PetscErrorCode MatMumpsGetInverseTranspose(Mat F,Mat spRHST) 2158 { 2159 PetscErrorCode ierr; 2160 PetscBool flg; 2161 2162 PetscFunctionBegin; 2163 PetscValidType(F,1); 2164 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2165 ierr = PetscObjectTypeCompareAny((PetscObject)spRHST,&flg,MATSEQAIJ,MATMPIAIJ,NULL);CHKERRQ(ierr); 2166 if (!flg) SETERRQ(PetscObjectComm((PetscObject)spRHST),PETSC_ERR_ARG_WRONG,"Matrix spRHST must be MATAIJ matrix"); 2167 2168 ierr = PetscUseMethod(F,"MatMumpsGetInverseTranspose_C",(Mat,Mat),(F,spRHST));CHKERRQ(ierr); 2169 PetscFunctionReturn(0); 2170 } 2171 2172 /*@ 2173 MatMumpsGetInfo - Get MUMPS parameter INFO() 2174 2175 Logically Collective on Mat 2176 2177 Input Parameters: 2178 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2179 - icntl - index of MUMPS parameter array INFO() 2180 2181 Output Parameter: 2182 . ival - value of MUMPS INFO(icntl) 2183 2184 Level: beginner 2185 2186 References: 2187 . MUMPS Users' Guide 2188 2189 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 2190 @*/ 2191 PetscErrorCode MatMumpsGetInfo(Mat F,PetscInt icntl,PetscInt *ival) 2192 { 2193 PetscErrorCode ierr; 2194 2195 PetscFunctionBegin; 2196 PetscValidType(F,1); 2197 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2198 PetscValidIntPointer(ival,3); 2199 ierr = PetscUseMethod(F,"MatMumpsGetInfo_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));CHKERRQ(ierr); 2200 PetscFunctionReturn(0); 2201 } 2202 2203 /*@ 2204 MatMumpsGetInfog - Get MUMPS parameter INFOG() 2205 2206 Logically Collective on Mat 2207 2208 Input Parameters: 2209 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2210 - icntl - index of MUMPS parameter array INFOG() 2211 2212 Output Parameter: 2213 . ival - value of MUMPS INFOG(icntl) 2214 2215 Level: beginner 2216 2217 References: 2218 . MUMPS Users' Guide 2219 2220 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 2221 @*/ 2222 PetscErrorCode MatMumpsGetInfog(Mat F,PetscInt icntl,PetscInt *ival) 2223 { 2224 PetscErrorCode ierr; 2225 2226 PetscFunctionBegin; 2227 PetscValidType(F,1); 2228 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2229 PetscValidIntPointer(ival,3); 2230 ierr = PetscUseMethod(F,"MatMumpsGetInfog_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));CHKERRQ(ierr); 2231 PetscFunctionReturn(0); 2232 } 2233 2234 /*@ 2235 MatMumpsGetRinfo - Get MUMPS parameter RINFO() 2236 2237 Logically Collective on Mat 2238 2239 Input Parameters: 2240 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2241 - icntl - index of MUMPS parameter array RINFO() 2242 2243 Output Parameter: 2244 . val - value of MUMPS RINFO(icntl) 2245 2246 Level: beginner 2247 2248 References: 2249 . MUMPS Users' Guide 2250 2251 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 2252 @*/ 2253 PetscErrorCode MatMumpsGetRinfo(Mat F,PetscInt icntl,PetscReal *val) 2254 { 2255 PetscErrorCode ierr; 2256 2257 PetscFunctionBegin; 2258 PetscValidType(F,1); 2259 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2260 PetscValidRealPointer(val,3); 2261 ierr = PetscUseMethod(F,"MatMumpsGetRinfo_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));CHKERRQ(ierr); 2262 PetscFunctionReturn(0); 2263 } 2264 2265 /*@ 2266 MatMumpsGetRinfog - Get MUMPS parameter RINFOG() 2267 2268 Logically Collective on Mat 2269 2270 Input Parameters: 2271 + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface 2272 - icntl - index of MUMPS parameter array RINFOG() 2273 2274 Output Parameter: 2275 . val - value of MUMPS RINFOG(icntl) 2276 2277 Level: beginner 2278 2279 References: 2280 . MUMPS Users' Guide 2281 2282 .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog() 2283 @*/ 2284 PetscErrorCode MatMumpsGetRinfog(Mat F,PetscInt icntl,PetscReal *val) 2285 { 2286 PetscErrorCode ierr; 2287 2288 PetscFunctionBegin; 2289 PetscValidType(F,1); 2290 if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix"); 2291 PetscValidRealPointer(val,3); 2292 ierr = PetscUseMethod(F,"MatMumpsGetRinfog_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));CHKERRQ(ierr); 2293 PetscFunctionReturn(0); 2294 } 2295 2296 /*MC 2297 MATSOLVERMUMPS - A matrix type providing direct solvers (LU and Cholesky) for 2298 distributed and sequential matrices via the external package MUMPS. 2299 2300 Works with MATAIJ and MATSBAIJ matrices 2301 2302 Use ./configure --download-mumps --download-scalapack --download-parmetis --download-metis --download-ptscotch to have PETSc installed with MUMPS 2303 2304 Use -pc_type cholesky or lu -pc_factor_mat_solver_type mumps to use this direct solver 2305 2306 Options Database Keys: 2307 + -mat_mumps_icntl_1 - ICNTL(1): output stream for error messages 2308 . -mat_mumps_icntl_2 - ICNTL(2): output stream for diagnostic printing, statistics, and warning 2309 . -mat_mumps_icntl_3 - ICNTL(3): output stream for global information, collected on the host 2310 . -mat_mumps_icntl_4 - ICNTL(4): level of printing (0 to 4) 2311 . -mat_mumps_icntl_6 - ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7) 2312 . -mat_mumps_icntl_7 - ICNTL(7): computes a symmetric permutation in sequential analysis (0 to 7). 3=Scotch, 4=PORD, 5=Metis 2313 . -mat_mumps_icntl_8 - ICNTL(8): scaling strategy (-2 to 8 or 77) 2314 . -mat_mumps_icntl_10 - ICNTL(10): max num of refinements 2315 . -mat_mumps_icntl_11 - ICNTL(11): statistics related to an error analysis (via -ksp_view) 2316 . -mat_mumps_icntl_12 - ICNTL(12): an ordering strategy for symmetric matrices (0 to 3) 2317 . -mat_mumps_icntl_13 - ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting 2318 . -mat_mumps_icntl_14 - ICNTL(14): percentage increase in the estimated working space 2319 . -mat_mumps_icntl_19 - ICNTL(19): computes the Schur complement 2320 . -mat_mumps_icntl_22 - ICNTL(22): in-core/out-of-core factorization and solve (0 or 1) 2321 . -mat_mumps_icntl_23 - ICNTL(23): max size of the working memory (MB) that can allocate per processor 2322 . -mat_mumps_icntl_24 - ICNTL(24): detection of null pivot rows (0 or 1) 2323 . -mat_mumps_icntl_25 - ICNTL(25): compute a solution of a deficient matrix and a null space basis 2324 . -mat_mumps_icntl_26 - ICNTL(26): drives the solution phase if a Schur complement matrix 2325 . -mat_mumps_icntl_28 - ICNTL(28): use 1 for sequential analysis and ictnl(7) ordering, or 2 for parallel analysis and ictnl(29) ordering 2326 . -mat_mumps_icntl_29 - ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis 2327 . -mat_mumps_icntl_30 - ICNTL(30): compute user-specified set of entries in inv(A) 2328 . -mat_mumps_icntl_31 - ICNTL(31): indicates which factors may be discarded during factorization 2329 . -mat_mumps_icntl_33 - ICNTL(33): compute determinant 2330 . -mat_mumps_cntl_1 - CNTL(1): relative pivoting threshold 2331 . -mat_mumps_cntl_2 - CNTL(2): stopping criterion of refinement 2332 . -mat_mumps_cntl_3 - CNTL(3): absolute pivoting threshold 2333 . -mat_mumps_cntl_4 - CNTL(4): value for static pivoting 2334 - -mat_mumps_cntl_5 - CNTL(5): fixation for null pivots 2335 2336 Level: beginner 2337 2338 Notes: 2339 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 2340 $ KSPGetPC(ksp,&pc); 2341 $ PCFactorGetMatrix(pc,&mat); 2342 $ MatMumpsGetInfo(mat,....); 2343 $ MatMumpsGetInfog(mat,....); etc. 2344 Or you can run with -ksp_error_if_not_converged and the program will be stopped and the information printed in the error message. 2345 2346 .seealso: PCFactorSetMatSolverType(), MatSolverType, MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog(), KSPGetPC(), PCGetFactor(), PCFactorGetMatrix() 2347 2348 M*/ 2349 2350 static PetscErrorCode MatFactorGetSolverType_mumps(Mat A,MatSolverType *type) 2351 { 2352 PetscFunctionBegin; 2353 *type = MATSOLVERMUMPS; 2354 PetscFunctionReturn(0); 2355 } 2356 2357 /* MatGetFactor for Seq and MPI AIJ matrices */ 2358 static PetscErrorCode MatGetFactor_aij_mumps(Mat A,MatFactorType ftype,Mat *F) 2359 { 2360 Mat B; 2361 PetscErrorCode ierr; 2362 Mat_MUMPS *mumps; 2363 PetscBool isSeqAIJ; 2364 2365 PetscFunctionBegin; 2366 /* Create the factorization matrix */ 2367 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isSeqAIJ);CHKERRQ(ierr); 2368 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2369 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 2370 ierr = PetscStrallocpy("mumps",&((PetscObject)B)->type_name);CHKERRQ(ierr); 2371 ierr = MatSetUp(B);CHKERRQ(ierr); 2372 2373 ierr = PetscNewLog(B,&mumps);CHKERRQ(ierr); 2374 2375 B->ops->view = MatView_MUMPS; 2376 B->ops->getinfo = MatGetInfo_MUMPS; 2377 2378 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);CHKERRQ(ierr); 2379 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);CHKERRQ(ierr); 2380 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);CHKERRQ(ierr); 2381 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);CHKERRQ(ierr); 2382 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);CHKERRQ(ierr); 2383 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);CHKERRQ(ierr); 2384 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);CHKERRQ(ierr); 2385 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);CHKERRQ(ierr); 2386 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);CHKERRQ(ierr); 2387 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);CHKERRQ(ierr); 2388 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);CHKERRQ(ierr); 2389 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverse_C",MatMumpsGetInverse_MUMPS);CHKERRQ(ierr); 2390 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverseTranspose_C",MatMumpsGetInverseTranspose_MUMPS);CHKERRQ(ierr); 2391 2392 if (ftype == MAT_FACTOR_LU) { 2393 B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS; 2394 B->factortype = MAT_FACTOR_LU; 2395 if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqaij; 2396 else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpiaij; 2397 mumps->sym = 0; 2398 } else { 2399 B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS; 2400 B->factortype = MAT_FACTOR_CHOLESKY; 2401 if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqsbaij; 2402 else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpisbaij; 2403 #if defined(PETSC_USE_COMPLEX) 2404 mumps->sym = 2; 2405 #else 2406 if (A->spd_set && A->spd) mumps->sym = 1; 2407 else mumps->sym = 2; 2408 #endif 2409 } 2410 2411 /* set solvertype */ 2412 ierr = PetscFree(B->solvertype);CHKERRQ(ierr); 2413 ierr = PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);CHKERRQ(ierr); 2414 2415 B->ops->destroy = MatDestroy_MUMPS; 2416 B->data = (void*)mumps; 2417 2418 ierr = PetscInitializeMUMPS(A,mumps);CHKERRQ(ierr); 2419 2420 *F = B; 2421 PetscFunctionReturn(0); 2422 } 2423 2424 /* MatGetFactor for Seq and MPI SBAIJ matrices */ 2425 static PetscErrorCode MatGetFactor_sbaij_mumps(Mat A,MatFactorType ftype,Mat *F) 2426 { 2427 Mat B; 2428 PetscErrorCode ierr; 2429 Mat_MUMPS *mumps; 2430 PetscBool isSeqSBAIJ; 2431 2432 PetscFunctionBegin; 2433 if (ftype != MAT_FACTOR_CHOLESKY) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Cannot use PETSc SBAIJ matrices with MUMPS LU, use AIJ matrix"); 2434 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"); 2435 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQSBAIJ,&isSeqSBAIJ);CHKERRQ(ierr); 2436 /* Create the factorization matrix */ 2437 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2438 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 2439 ierr = PetscStrallocpy("mumps",&((PetscObject)B)->type_name);CHKERRQ(ierr); 2440 ierr = MatSetUp(B);CHKERRQ(ierr); 2441 2442 ierr = PetscNewLog(B,&mumps);CHKERRQ(ierr); 2443 if (isSeqSBAIJ) { 2444 mumps->ConvertToTriples = MatConvertToTriples_seqsbaij_seqsbaij; 2445 } else { 2446 mumps->ConvertToTriples = MatConvertToTriples_mpisbaij_mpisbaij; 2447 } 2448 2449 B->ops->getinfo = MatGetInfo_External; 2450 B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS; 2451 B->ops->view = MatView_MUMPS; 2452 2453 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);CHKERRQ(ierr); 2454 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);CHKERRQ(ierr); 2455 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);CHKERRQ(ierr); 2456 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);CHKERRQ(ierr); 2457 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);CHKERRQ(ierr); 2458 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);CHKERRQ(ierr); 2459 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);CHKERRQ(ierr); 2460 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);CHKERRQ(ierr); 2461 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);CHKERRQ(ierr); 2462 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);CHKERRQ(ierr); 2463 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);CHKERRQ(ierr); 2464 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverse_C",MatMumpsGetInverse_MUMPS);CHKERRQ(ierr); 2465 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverseTranspose_C",MatMumpsGetInverseTranspose_MUMPS);CHKERRQ(ierr); 2466 2467 B->factortype = MAT_FACTOR_CHOLESKY; 2468 #if defined(PETSC_USE_COMPLEX) 2469 mumps->sym = 2; 2470 #else 2471 if (A->spd_set && A->spd) mumps->sym = 1; 2472 else mumps->sym = 2; 2473 #endif 2474 2475 /* set solvertype */ 2476 ierr = PetscFree(B->solvertype);CHKERRQ(ierr); 2477 ierr = PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);CHKERRQ(ierr); 2478 2479 B->ops->destroy = MatDestroy_MUMPS; 2480 B->data = (void*)mumps; 2481 2482 ierr = PetscInitializeMUMPS(A,mumps);CHKERRQ(ierr); 2483 2484 *F = B; 2485 PetscFunctionReturn(0); 2486 } 2487 2488 static PetscErrorCode MatGetFactor_baij_mumps(Mat A,MatFactorType ftype,Mat *F) 2489 { 2490 Mat B; 2491 PetscErrorCode ierr; 2492 Mat_MUMPS *mumps; 2493 PetscBool isSeqBAIJ; 2494 2495 PetscFunctionBegin; 2496 /* Create the factorization matrix */ 2497 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQBAIJ,&isSeqBAIJ);CHKERRQ(ierr); 2498 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2499 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 2500 ierr = PetscStrallocpy("mumps",&((PetscObject)B)->type_name);CHKERRQ(ierr); 2501 ierr = MatSetUp(B);CHKERRQ(ierr); 2502 2503 ierr = PetscNewLog(B,&mumps);CHKERRQ(ierr); 2504 if (ftype == MAT_FACTOR_LU) { 2505 B->ops->lufactorsymbolic = MatLUFactorSymbolic_BAIJMUMPS; 2506 B->factortype = MAT_FACTOR_LU; 2507 if (isSeqBAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqbaij_seqaij; 2508 else mumps->ConvertToTriples = MatConvertToTriples_mpibaij_mpiaij; 2509 mumps->sym = 0; 2510 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot use PETSc BAIJ matrices with MUMPS Cholesky, use SBAIJ or AIJ matrix instead\n"); 2511 2512 B->ops->getinfo = MatGetInfo_External; 2513 B->ops->view = MatView_MUMPS; 2514 2515 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);CHKERRQ(ierr); 2516 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);CHKERRQ(ierr); 2517 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);CHKERRQ(ierr); 2518 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);CHKERRQ(ierr); 2519 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);CHKERRQ(ierr); 2520 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);CHKERRQ(ierr); 2521 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);CHKERRQ(ierr); 2522 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);CHKERRQ(ierr); 2523 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);CHKERRQ(ierr); 2524 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);CHKERRQ(ierr); 2525 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);CHKERRQ(ierr); 2526 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverse_C",MatMumpsGetInverse_MUMPS);CHKERRQ(ierr); 2527 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverseTranspose_C",MatMumpsGetInverseTranspose_MUMPS);CHKERRQ(ierr); 2528 2529 /* set solvertype */ 2530 ierr = PetscFree(B->solvertype);CHKERRQ(ierr); 2531 ierr = PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);CHKERRQ(ierr); 2532 2533 B->ops->destroy = MatDestroy_MUMPS; 2534 B->data = (void*)mumps; 2535 2536 ierr = PetscInitializeMUMPS(A,mumps);CHKERRQ(ierr); 2537 2538 *F = B; 2539 PetscFunctionReturn(0); 2540 } 2541 2542 /* MatGetFactor for Seq and MPI SELL matrices */ 2543 static PetscErrorCode MatGetFactor_sell_mumps(Mat A,MatFactorType ftype,Mat *F) 2544 { 2545 Mat B; 2546 PetscErrorCode ierr; 2547 Mat_MUMPS *mumps; 2548 PetscBool isSeqSELL; 2549 2550 PetscFunctionBegin; 2551 /* Create the factorization matrix */ 2552 ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQSELL,&isSeqSELL);CHKERRQ(ierr); 2553 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2554 ierr = MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);CHKERRQ(ierr); 2555 ierr = PetscStrallocpy("mumps",&((PetscObject)B)->type_name);CHKERRQ(ierr); 2556 ierr = MatSetUp(B);CHKERRQ(ierr); 2557 2558 ierr = PetscNewLog(B,&mumps);CHKERRQ(ierr); 2559 2560 B->ops->view = MatView_MUMPS; 2561 B->ops->getinfo = MatGetInfo_MUMPS; 2562 2563 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);CHKERRQ(ierr); 2564 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);CHKERRQ(ierr); 2565 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);CHKERRQ(ierr); 2566 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);CHKERRQ(ierr); 2567 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);CHKERRQ(ierr); 2568 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);CHKERRQ(ierr); 2569 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);CHKERRQ(ierr); 2570 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);CHKERRQ(ierr); 2571 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);CHKERRQ(ierr); 2572 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);CHKERRQ(ierr); 2573 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);CHKERRQ(ierr); 2574 2575 if (ftype == MAT_FACTOR_LU) { 2576 B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS; 2577 B->factortype = MAT_FACTOR_LU; 2578 if (isSeqSELL) mumps->ConvertToTriples = MatConvertToTriples_seqsell_seqaij; 2579 else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"To be implemented"); 2580 mumps->sym = 0; 2581 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"To be implemented"); 2582 2583 /* set solvertype */ 2584 ierr = PetscFree(B->solvertype);CHKERRQ(ierr); 2585 ierr = PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);CHKERRQ(ierr); 2586 2587 B->ops->destroy = MatDestroy_MUMPS; 2588 B->data = (void*)mumps; 2589 2590 ierr = PetscInitializeMUMPS(A,mumps);CHKERRQ(ierr); 2591 2592 *F = B; 2593 PetscFunctionReturn(0); 2594 } 2595 2596 PETSC_EXTERN PetscErrorCode MatSolverTypeRegister_MUMPS(void) 2597 { 2598 PetscErrorCode ierr; 2599 2600 PetscFunctionBegin; 2601 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);CHKERRQ(ierr); 2602 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);CHKERRQ(ierr); 2603 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);CHKERRQ(ierr); 2604 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);CHKERRQ(ierr); 2605 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATMPISBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);CHKERRQ(ierr); 2606 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);CHKERRQ(ierr); 2607 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);CHKERRQ(ierr); 2608 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);CHKERRQ(ierr); 2609 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);CHKERRQ(ierr); 2610 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQSBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);CHKERRQ(ierr); 2611 ierr = MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQSELL,MAT_FACTOR_LU,MatGetFactor_sell_mumps);CHKERRQ(ierr); 2612 PetscFunctionReturn(0); 2613 } 2614 2615