1 #define PETSCMAT_DLL 2 3 4 #include "../src/mat/impls/aij/seq/aij.h" 5 #include "../src/mat/impls/sbaij/seq/sbaij.h" 6 #include "petscbt.h" 7 #include "../src/mat/utils/freespace.h" 8 9 EXTERN_C_BEGIN 10 #undef __FUNCT__ 11 #define __FUNCT__ "MatOrdering_Flow_SeqAIJ" 12 /* 13 Computes an ordering to get most of the large numerical values in the lower triangular part of the matrix 14 */ 15 PetscErrorCode MatOrdering_Flow_SeqAIJ(Mat mat,const MatOrderingType type,IS *irow,IS *icol) 16 { 17 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->data; 18 PetscErrorCode ierr; 19 PetscInt i,j,jj,k, kk,n = mat->rmap->n, current = 0, newcurrent = 0,*order; 20 const PetscInt *ai = a->i, *aj = a->j; 21 const PetscScalar *aa = a->a; 22 PetscTruth *done; 23 PetscReal best,past = 0,future; 24 25 PetscFunctionBegin; 26 /* pick initial row */ 27 best = -1; 28 for (i=0; i<n; i++) { 29 future = 0.0; 30 for (j=ai[i]; j<ai[i+1]; j++) { 31 if (aj[j] != i) future += PetscAbsScalar(aa[j]); else past = PetscAbsScalar(aa[j]); 32 } 33 if (!future) future = 1.e-10; /* if there is zero in the upper diagonal part want to rank this row high */ 34 if (past/future > best) { 35 best = past/future; 36 current = i; 37 } 38 } 39 40 ierr = PetscMalloc(n*sizeof(PetscTruth),&done);CHKERRQ(ierr); 41 ierr = PetscMemzero(done,n*sizeof(PetscTruth));CHKERRQ(ierr); 42 ierr = PetscMalloc(n*sizeof(PetscInt),&order);CHKERRQ(ierr); 43 order[0] = current; 44 for (i=0; i<n-1; i++) { 45 done[current] = PETSC_TRUE; 46 best = -1; 47 /* loop over all neighbors of current pivot */ 48 for (j=ai[current]; j<ai[current+1]; j++) { 49 jj = aj[j]; 50 if (done[jj]) continue; 51 /* loop over columns of potential next row computing weights for below and above diagonal */ 52 past = future = 0.0; 53 for (k=ai[jj]; k<ai[jj+1]; k++) { 54 kk = aj[k]; 55 if (done[kk]) past += PetscAbsScalar(aa[k]); 56 else if (kk != jj) future += PetscAbsScalar(aa[k]); 57 } 58 if (!future) future = 1.e-10; /* if there is zero in the upper diagonal part want to rank this row high */ 59 if (past/future > best) { 60 best = past/future; 61 newcurrent = jj; 62 } 63 } 64 if (best == -1) { /* no neighbors to select from so select best of all that remain */ 65 best = -1; 66 for (k=0; k<n; k++) { 67 if (done[k]) continue; 68 future = 0.0; 69 past = 0.0; 70 for (j=ai[k]; j<ai[k+1]; j++) { 71 kk = aj[j]; 72 if (done[kk]) past += PetscAbsScalar(aa[j]); 73 else if (kk != k) future += PetscAbsScalar(aa[j]); 74 } 75 if (!future) future = 1.e-10; /* if there is zero in the upper diagonal part want to rank this row high */ 76 if (past/future > best) { 77 best = past/future; 78 newcurrent = k; 79 } 80 } 81 } 82 if (current == newcurrent) SETERRQ(PETSC_ERR_PLIB,"newcurrent cannot be current"); 83 current = newcurrent; 84 order[i+1] = current; 85 } 86 ierr = ISCreateGeneral(PETSC_COMM_SELF,n,order,irow);CHKERRQ(ierr); 87 *icol = *irow; 88 ierr = PetscObjectReference((PetscObject)*irow);CHKERRQ(ierr); 89 ierr = PetscFree(done);CHKERRQ(ierr); 90 ierr = PetscFree(order);CHKERRQ(ierr); 91 PetscFunctionReturn(0); 92 } 93 EXTERN_C_END 94 95 EXTERN_C_BEGIN 96 #undef __FUNCT__ 97 #define __FUNCT__ "MatGetFactorAvailable_seqaij_petsc" 98 PetscErrorCode MatGetFactorAvailable_seqaij_petsc(Mat A,MatFactorType ftype,PetscTruth *flg) 99 { 100 PetscFunctionBegin; 101 *flg = PETSC_TRUE; 102 PetscFunctionReturn(0); 103 } 104 EXTERN_C_END 105 106 EXTERN_C_BEGIN 107 #undef __FUNCT__ 108 #define __FUNCT__ "MatGetFactor_seqaij_petsc" 109 PetscErrorCode MatGetFactor_seqaij_petsc(Mat A,MatFactorType ftype,Mat *B) 110 { 111 PetscInt n = A->rmap->n; 112 PetscErrorCode ierr; 113 114 PetscFunctionBegin; 115 ierr = MatCreate(((PetscObject)A)->comm,B);CHKERRQ(ierr); 116 ierr = MatSetSizes(*B,n,n,n,n);CHKERRQ(ierr); 117 if (ftype == MAT_FACTOR_LU || ftype == MAT_FACTOR_ILU || ftype == MAT_FACTOR_ILUDT){ 118 ierr = MatSetType(*B,MATSEQAIJ);CHKERRQ(ierr); 119 (*B)->ops->ilufactorsymbolic = MatILUFactorSymbolic_SeqAIJ; 120 (*B)->ops->lufactorsymbolic = MatLUFactorSymbolic_SeqAIJ; 121 (*B)->ops->iludtfactor = MatILUDTFactor_SeqAIJ; 122 } else if (ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC) { 123 ierr = MatSetType(*B,MATSEQSBAIJ);CHKERRQ(ierr); 124 ierr = MatSeqSBAIJSetPreallocation(*B,1,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 125 (*B)->ops->iccfactorsymbolic = MatICCFactorSymbolic_SeqAIJ; 126 (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJ; 127 } else SETERRQ(PETSC_ERR_SUP,"Factor type not supported"); 128 (*B)->factor = ftype; 129 PetscFunctionReturn(0); 130 } 131 EXTERN_C_END 132 133 #undef __FUNCT__ 134 #define __FUNCT__ "MatLUFactorSymbolic_SeqAIJ" 135 PetscErrorCode MatLUFactorSymbolic_SeqAIJ(Mat B,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 136 { 137 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 138 IS isicol; 139 PetscErrorCode ierr; 140 const PetscInt *r,*ic; 141 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j; 142 PetscInt *bi,*bj,*ajtmp; 143 PetscInt *bdiag,row,nnz,nzi,reallocs=0,nzbd,*im; 144 PetscReal f; 145 PetscInt nlnk,*lnk,k,**bi_ptr; 146 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 147 PetscBT lnkbt; 148 PetscTruth newdatastruct= PETSC_FALSE; 149 150 PetscFunctionBegin; 151 ierr = PetscOptionsGetTruth(PETSC_NULL,"-lu_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 152 if(newdatastruct){ 153 ierr = MatLUFactorSymbolic_SeqAIJ_newdatastruct(B,A,isrow,iscol,info);CHKERRQ(ierr); 154 PetscFunctionReturn(0); 155 } 156 157 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 158 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 159 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 160 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 161 162 /* get new row pointers */ 163 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 164 bi[0] = 0; 165 166 /* bdiag is location of diagonal in factor */ 167 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 168 bdiag[0] = 0; 169 170 /* linked list for storing column indices of the active row */ 171 nlnk = n + 1; 172 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 173 174 ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr); 175 176 /* initial FreeSpace size is f*(ai[n]+1) */ 177 f = info->fill; 178 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 179 current_space = free_space; 180 181 for (i=0; i<n; i++) { 182 /* copy previous fill into linked list */ 183 nzi = 0; 184 nnz = ai[r[i]+1] - ai[r[i]]; 185 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 186 ajtmp = aj + ai[r[i]]; 187 ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 188 nzi += nlnk; 189 190 /* add pivot rows into linked list */ 191 row = lnk[n]; 192 while (row < i) { 193 nzbd = bdiag[row] - bi[row] + 1; /* num of entries in the row with column index <= row */ 194 ajtmp = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */ 195 ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr); 196 nzi += nlnk; 197 row = lnk[row]; 198 } 199 bi[i+1] = bi[i] + nzi; 200 im[i] = nzi; 201 202 /* mark bdiag */ 203 nzbd = 0; 204 nnz = nzi; 205 k = lnk[n]; 206 while (nnz-- && k < i){ 207 nzbd++; 208 k = lnk[k]; 209 } 210 bdiag[i] = bi[i] + nzbd; 211 212 /* if free space is not available, make more free space */ 213 if (current_space->local_remaining<nzi) { 214 nnz = (n - i)*nzi; /* estimated and max additional space needed */ 215 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 216 reallocs++; 217 } 218 219 /* copy data into free space, then initialize lnk */ 220 ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 221 bi_ptr[i] = current_space->array; 222 current_space->array += nzi; 223 current_space->local_used += nzi; 224 current_space->local_remaining -= nzi; 225 } 226 #if defined(PETSC_USE_INFO) 227 if (ai[n] != 0) { 228 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 229 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 230 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 231 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr); 232 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 233 } else { 234 ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr); 235 } 236 #endif 237 238 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 239 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 240 241 /* destroy list of free space and other temporary array(s) */ 242 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 243 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 244 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 245 ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr); 246 247 /* put together the new matrix */ 248 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 249 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 250 b = (Mat_SeqAIJ*)(B)->data; 251 b->free_a = PETSC_TRUE; 252 b->free_ij = PETSC_TRUE; 253 b->singlemalloc = PETSC_FALSE; 254 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 255 b->j = bj; 256 b->i = bi; 257 b->diag = bdiag; 258 b->ilen = 0; 259 b->imax = 0; 260 b->row = isrow; 261 b->col = iscol; 262 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 263 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 264 b->icol = isicol; 265 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 266 267 /* In b structure: Free imax, ilen, old a, old j. Allocate solve_work, new a, new j */ 268 ierr = PetscLogObjectMemory(B,(bi[n]-n)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 269 b->maxnz = b->nz = bi[n] ; 270 271 (B)->factor = MAT_FACTOR_LU; 272 (B)->info.factor_mallocs = reallocs; 273 (B)->info.fill_ratio_given = f; 274 275 if (ai[n]) { 276 (B)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]); 277 } else { 278 (B)->info.fill_ratio_needed = 0.0; 279 } 280 (B)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ; 281 (B)->ops->solve = MatSolve_SeqAIJ; 282 (B)->ops->solvetranspose = MatSolveTranspose_SeqAIJ; 283 /* switch to inodes if appropriate */ 284 ierr = MatLUFactorSymbolic_SeqAIJ_Inode(B,A,isrow,iscol,info);CHKERRQ(ierr); 285 PetscFunctionReturn(0); 286 } 287 288 #undef __FUNCT__ 289 #define __FUNCT__ "MatLUFactorSymbolic_SeqAIJ_newdatastruct" 290 PetscErrorCode MatLUFactorSymbolic_SeqAIJ_newdatastruct(Mat B,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 291 { 292 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 293 IS isicol; 294 PetscErrorCode ierr; 295 const PetscInt *r,*ic; 296 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j; 297 PetscInt *bi,*bj,*ajtmp; 298 PetscInt *bdiag,row,nnz,nzi,reallocs=0,nzbd,*im; 299 PetscReal f; 300 PetscInt nlnk,*lnk,k,**bi_ptr; 301 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 302 PetscBT lnkbt; 303 304 PetscFunctionBegin; 305 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 306 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 307 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 308 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 309 310 /* get new row and diagonal pointers, must be allocated separately because they will be given to the Mat_SeqAIJ and freed separately */ 311 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 312 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 313 bi[0] = bdiag[0] = 0; 314 315 /* linked list for storing column indices of the active row */ 316 nlnk = n + 1; 317 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 318 319 ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr); 320 321 /* initial FreeSpace size is f*(ai[n]+1) */ 322 f = info->fill; 323 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 324 current_space = free_space; 325 326 for (i=0; i<n; i++) { 327 /* copy previous fill into linked list */ 328 nzi = 0; 329 nnz = ai[r[i]+1] - ai[r[i]]; 330 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 331 ajtmp = aj + ai[r[i]]; 332 ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 333 nzi += nlnk; 334 335 /* add pivot rows into linked list */ 336 row = lnk[n]; 337 while (row < i){ 338 nzbd = bdiag[row] + 1; /* num of entries in the row with column index <= row */ 339 ajtmp = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */ 340 ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr); 341 nzi += nlnk; 342 row = lnk[row]; 343 } 344 bi[i+1] = bi[i] + nzi; 345 im[i] = nzi; 346 347 /* mark bdiag */ 348 nzbd = 0; 349 nnz = nzi; 350 k = lnk[n]; 351 while (nnz-- && k < i){ 352 nzbd++; 353 k = lnk[k]; 354 } 355 bdiag[i] = nzbd; /* note: bdiag[i] = nnzL as input for PetscFreeSpaceContiguous_LU() */ 356 357 /* if free space is not available, make more free space */ 358 if (current_space->local_remaining<nzi) { 359 nnz = 2*(n - i)*nzi; /* estimated and max additional space needed */ 360 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 361 reallocs++; 362 } 363 364 /* copy data into free space, then initialize lnk */ 365 ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 366 bi_ptr[i] = current_space->array; 367 current_space->array += nzi; 368 current_space->local_used += nzi; 369 current_space->local_remaining -= nzi; 370 } 371 #if defined(PETSC_USE_INFO) 372 if (ai[n] != 0) { 373 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 374 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 375 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 376 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr); 377 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 378 } else { 379 ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr); 380 } 381 #endif 382 383 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 384 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 385 386 /* destroy list of free space and other temporary array(s) */ 387 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 388 ierr = PetscFreeSpaceContiguous_LU_v2(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 389 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 390 ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr); 391 392 /* put together the new matrix */ 393 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 394 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 395 b = (Mat_SeqAIJ*)(B)->data; 396 b->free_a = PETSC_TRUE; 397 b->free_ij = PETSC_TRUE; 398 b->singlemalloc = PETSC_FALSE; 399 ierr = PetscMalloc((bdiag[0]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 400 b->j = bj; 401 b->i = bi; 402 b->diag = bdiag; 403 b->ilen = 0; 404 b->imax = 0; 405 b->row = isrow; 406 b->col = iscol; 407 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 408 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 409 b->icol = isicol; 410 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 411 412 /* In b structure: Free imax, ilen, old a, old j. Allocate solve_work, new a, new j */ 413 ierr = PetscLogObjectMemory(B,(bdiag[0]+1)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 414 b->maxnz = b->nz = bdiag[0]+1; 415 (B)->factor = MAT_FACTOR_LU; 416 (B)->info.factor_mallocs = reallocs; 417 (B)->info.fill_ratio_given = f; 418 419 if (ai[n]) { 420 (B)->info.fill_ratio_needed = ((PetscReal)(bdiag[0]+1))/((PetscReal)ai[n]); 421 } else { 422 (B)->info.fill_ratio_needed = 0.0; 423 } 424 (B)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 425 /* switch to inodes if appropriate */ 426 ierr = Mat_CheckInode_FactorLU(B,PETSC_FALSE);CHKERRQ(ierr); 427 PetscFunctionReturn(0); 428 } 429 430 /* 431 Trouble in factorization, should we dump the original matrix? 432 */ 433 #undef __FUNCT__ 434 #define __FUNCT__ "MatFactorDumpMatrix" 435 PetscErrorCode MatFactorDumpMatrix(Mat A) 436 { 437 PetscErrorCode ierr; 438 PetscTruth flg = PETSC_FALSE; 439 440 PetscFunctionBegin; 441 ierr = PetscOptionsGetTruth(PETSC_NULL,"-mat_factor_dump_on_error",&flg,PETSC_NULL);CHKERRQ(ierr); 442 if (flg) { 443 PetscViewer viewer; 444 char filename[PETSC_MAX_PATH_LEN]; 445 446 ierr = PetscSNPrintf(filename,PETSC_MAX_PATH_LEN,"matrix_factor_error.%d",PetscGlobalRank);CHKERRQ(ierr); 447 ierr = PetscViewerBinaryOpen(((PetscObject)A)->comm,filename,FILE_MODE_WRITE,&viewer);CHKERRQ(ierr); 448 ierr = MatView(A,viewer);CHKERRQ(ierr); 449 ierr = PetscViewerDestroy(viewer);CHKERRQ(ierr); 450 } 451 PetscFunctionReturn(0); 452 } 453 454 #undef __FUNCT__ 455 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_newdatastruct" 456 PetscErrorCode MatLUFactorNumeric_SeqAIJ_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 457 { 458 Mat C=B; 459 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 460 IS isrow = b->row,isicol = b->icol; 461 PetscErrorCode ierr; 462 const PetscInt *r,*ic,*ics; 463 PetscInt i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 464 PetscInt *ajtmp,*bjtmp,nz,nzL,row,*bdiag=b->diag,*pj; 465 MatScalar *rtmp,*pc,multiplier,*v,*pv,*aa=a->a; 466 PetscTruth row_identity,col_identity; 467 468 LUShift_Ctx sctx; 469 PetscInt *ddiag,newshift; 470 PetscReal rs; 471 MatScalar d; 472 473 PetscFunctionBegin; 474 /* MatPivotSetUp(): initialize shift context sctx */ 475 ierr = PetscMemzero(&sctx,sizeof(LUShift_Ctx));CHKERRQ(ierr); 476 477 /* if both shift schemes are chosen by user, only use info->shiftpd */ 478 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 479 ddiag = a->diag; 480 sctx.shift_top = info->zeropivot; 481 for (i=0; i<n; i++) { 482 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 483 d = (aa)[ddiag[i]]; 484 rs = -PetscAbsScalar(d) - PetscRealPart(d); 485 v = aa+ai[i]; 486 nz = ai[i+1] - ai[i]; 487 for (j=0; j<nz; j++) 488 rs += PetscAbsScalar(v[j]); 489 if (rs>sctx.shift_top) sctx.shift_top = rs; 490 } 491 sctx.shift_top *= 1.1; 492 sctx.nshift_max = 5; 493 sctx.shift_lo = 0.; 494 sctx.shift_hi = 1.; 495 } 496 497 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 498 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 499 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 500 ics = ic; 501 502 do { 503 sctx.lushift = PETSC_FALSE; 504 for (i=0; i<n; i++){ 505 /* zero rtmp */ 506 /* L part */ 507 nz = bi[i+1] - bi[i]; 508 bjtmp = bj + bi[i]; 509 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 510 511 /* U part */ 512 nz = bdiag[i]-bdiag[i+1]; 513 bjtmp = bj + bdiag[i+1]+1; 514 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 515 516 /* load in initial (unfactored row) */ 517 nz = ai[r[i]+1] - ai[r[i]]; 518 ajtmp = aj + ai[r[i]]; 519 v = aa + ai[r[i]]; 520 for (j=0; j<nz; j++) { 521 rtmp[ics[ajtmp[j]]] = v[j]; 522 } 523 /* ZeropivotApply() */ 524 rtmp[i] += sctx.shift_amount; /* shift the diagonal of the matrix */ 525 526 /* elimination */ 527 bjtmp = bj + bi[i]; 528 row = *bjtmp++; 529 nzL = bi[i+1] - bi[i]; 530 for(k=0; k < nzL;k++) { 531 pc = rtmp + row; 532 if (*pc != 0.0) { 533 pv = b->a + bdiag[row]; 534 multiplier = *pc * (*pv); 535 *pc = multiplier; 536 pj = b->j + bdiag[row+1]+1; /* beginning of U(row,:) */ 537 pv = b->a + bdiag[row+1]+1; 538 nz = bdiag[row]-bdiag[row+1]-1; /* num of entries in U(row,:) excluding diag */ 539 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 540 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 541 } 542 row = *bjtmp++; 543 } 544 545 /* finished row so stick it into b->a */ 546 rs = 0.0; 547 /* L part */ 548 pv = b->a + bi[i] ; 549 pj = b->j + bi[i] ; 550 nz = bi[i+1] - bi[i]; 551 for (j=0; j<nz; j++) { 552 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 553 } 554 555 /* U part */ 556 pv = b->a + bdiag[i+1]+1; 557 pj = b->j + bdiag[i+1]+1; 558 nz = bdiag[i] - bdiag[i+1]-1; 559 for (j=0; j<nz; j++) { 560 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 561 } 562 563 /* MatPivotCheck() */ 564 sctx.rs = rs; 565 sctx.pv = rtmp[i]; 566 /* ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); */ 567 if (info->shiftnz){ 568 ierr = MatPivotCheck_nz(info,sctx,i,newshift);CHKERRQ(ierr); 569 } else if (info->shiftpd){ 570 ierr = MatPivotCheck_pd(info,sctx,i,newshift);CHKERRQ(ierr); 571 } else if (info->shiftinblocks){ 572 ierr = MatPivotCheck_inblocks(info,sctx,i,newshift);CHKERRQ(ierr); 573 } else { 574 ierr = MatPivotCheck_none(info,sctx,i,newshift);CHKERRQ(ierr); 575 } 576 rtmp[i] = sctx.pv; 577 if (newshift == 1) break; 578 579 /* Mark diagonal and invert diagonal for simplier triangular solves */ 580 pv = b->a + bdiag[i]; 581 *pv = 1.0/rtmp[i]; 582 583 } /* endof for (i=0; i<n; i++){ */ 584 585 /* MatPivotRefine() */ 586 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max){ 587 /* 588 * if no shift in this attempt & shifting & started shifting & can refine, 589 * then try lower shift 590 */ 591 sctx.shift_hi = sctx.shift_fraction; 592 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 593 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 594 sctx.lushift = PETSC_TRUE; 595 sctx.nshift++; 596 } 597 } while (sctx.lushift); 598 599 ierr = PetscFree(rtmp);CHKERRQ(ierr); 600 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 601 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 602 if (b->inode.use){ 603 C->ops->solve = MatSolve_SeqAIJ_Inode_newdatastruct; 604 } else { 605 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 606 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 607 if (row_identity && col_identity) { 608 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 609 } else { 610 C->ops->solve = MatSolve_SeqAIJ_newdatastruct; 611 } 612 } 613 C->ops->solveadd = MatSolveAdd_SeqAIJ_newdatastruct; 614 C->ops->solvetranspose = MatSolveTranspose_SeqAIJ_newdatastruct; 615 C->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ_newdatastruct; 616 C->ops->matsolve = MatMatSolve_SeqAIJ_newdatastruct; 617 C->assembled = PETSC_TRUE; 618 C->preallocated = PETSC_TRUE; 619 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 620 621 /* MatPivotView() */ 622 if (sctx.nshift){ 623 if (info->shiftpd) { 624 ierr = PetscInfo4(A,"number of shift_pd tries %D, shift_amount %G, diagonal shifted up by %e fraction top_value %e\n",sctx.nshift,sctx.shift_amount,sctx.shift_fraction,sctx.shift_top);CHKERRQ(ierr); 625 } else if (info->shiftnz) { 626 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 627 } else if (info->shiftinblocks){ 628 ierr = PetscInfo2(A,"number of shift_inblocks applied %D, each shift_amount %G\n",sctx.nshift,info->shiftinblocks);CHKERRQ(ierr); 629 } 630 } 631 PetscFunctionReturn(0); 632 } 633 634 #undef __FUNCT__ 635 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ" 636 PetscErrorCode MatLUFactorNumeric_SeqAIJ(Mat B,Mat A,const MatFactorInfo *info) 637 { 638 Mat C=B; 639 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 640 IS isrow = b->row,isicol = b->icol; 641 PetscErrorCode ierr; 642 const PetscInt *r,*ic,*ics; 643 PetscInt nz,row,i,j,n=A->rmap->n,diag; 644 const PetscInt *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 645 const PetscInt *ajtmp,*bjtmp,*diag_offset = b->diag,*pj; 646 MatScalar *pv,*rtmp,*pc,multiplier,d; 647 const MatScalar *v,*aa=a->a; 648 PetscReal rs=0.0; 649 LUShift_Ctx sctx; 650 PetscInt newshift,*ddiag; 651 652 PetscFunctionBegin; 653 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 654 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 655 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 656 ics = ic; 657 658 /* initialize shift context sctx */ 659 sctx.nshift = 0; 660 sctx.nshift_max = 0; 661 sctx.shift_top = 0.0; 662 sctx.shift_lo = 0.0; 663 sctx.shift_hi = 0.0; 664 sctx.shift_fraction = 0.0; 665 sctx.shift_amount = 0.0; 666 667 /* if both shift schemes are chosen by user, only use info->shiftpd */ 668 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 669 ddiag = a->diag; 670 sctx.shift_top = info->zeropivot; 671 for (i=0; i<n; i++) { 672 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 673 d = (aa)[ddiag[i]]; 674 rs = -PetscAbsScalar(d) - PetscRealPart(d); 675 v = aa+ai[i]; 676 nz = ai[i+1] - ai[i]; 677 for (j=0; j<nz; j++) 678 rs += PetscAbsScalar(v[j]); 679 if (rs>sctx.shift_top) sctx.shift_top = rs; 680 } 681 sctx.shift_top *= 1.1; 682 sctx.nshift_max = 5; 683 sctx.shift_lo = 0.; 684 sctx.shift_hi = 1.; 685 } 686 687 do { 688 sctx.lushift = PETSC_FALSE; 689 for (i=0; i<n; i++){ 690 nz = bi[i+1] - bi[i]; 691 bjtmp = bj + bi[i]; 692 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 693 694 /* load in initial (unfactored row) */ 695 nz = ai[r[i]+1] - ai[r[i]]; 696 ajtmp = aj + ai[r[i]]; 697 v = aa + ai[r[i]]; 698 for (j=0; j<nz; j++) { 699 rtmp[ics[ajtmp[j]]] = v[j]; 700 } 701 rtmp[ics[r[i]]] += sctx.shift_amount; /* shift the diagonal of the matrix */ 702 /* if (sctx.shift_amount > 0.0) printf("row %d, shift %g\n",i,sctx.shift_amount); */ 703 704 row = *bjtmp++; 705 while (row < i) { 706 pc = rtmp + row; 707 if (*pc != 0.0) { 708 pv = b->a + diag_offset[row]; 709 pj = b->j + diag_offset[row] + 1; 710 multiplier = *pc / *pv++; 711 *pc = multiplier; 712 nz = bi[row+1] - diag_offset[row] - 1; 713 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 714 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 715 } 716 row = *bjtmp++; 717 } 718 /* finished row so stick it into b->a */ 719 pv = b->a + bi[i] ; 720 pj = b->j + bi[i] ; 721 nz = bi[i+1] - bi[i]; 722 diag = diag_offset[i] - bi[i]; 723 rs = 0.0; 724 for (j=0; j<nz; j++) { 725 pv[j] = rtmp[pj[j]]; 726 rs += PetscAbsScalar(pv[j]); 727 } 728 rs -= PetscAbsScalar(pv[diag]); 729 730 /* 9/13/02 Victor Eijkhout suggested scaling zeropivot by rs for matrices with funny scalings */ 731 sctx.rs = rs; 732 sctx.pv = pv[diag]; 733 ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); 734 if (newshift == 1) break; 735 } 736 737 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max) { 738 /* 739 * if no shift in this attempt & shifting & started shifting & can refine, 740 * then try lower shift 741 */ 742 sctx.shift_hi = sctx.shift_fraction; 743 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 744 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 745 sctx.lushift = PETSC_TRUE; 746 sctx.nshift++; 747 } 748 } while (sctx.lushift); 749 750 /* invert diagonal entries for simplier triangular solves */ 751 for (i=0; i<n; i++) { 752 b->a[diag_offset[i]] = 1.0/b->a[diag_offset[i]]; 753 } 754 ierr = PetscFree(rtmp);CHKERRQ(ierr); 755 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 756 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 757 if (b->inode.use) { 758 C->ops->solve = MatSolve_SeqAIJ_Inode; 759 } else { 760 PetscTruth row_identity, col_identity; 761 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 762 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 763 if (row_identity && col_identity) { 764 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering; 765 } else { 766 C->ops->solve = MatSolve_SeqAIJ; 767 } 768 } 769 C->ops->solveadd = MatSolveAdd_SeqAIJ; 770 C->ops->solvetranspose = MatSolveTranspose_SeqAIJ; 771 C->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ; 772 C->ops->matsolve = MatMatSolve_SeqAIJ; 773 C->assembled = PETSC_TRUE; 774 C->preallocated = PETSC_TRUE; 775 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 776 if (sctx.nshift){ 777 if (info->shiftpd) { 778 ierr = PetscInfo4(A,"number of shift_pd tries %D, shift_amount %G, diagonal shifted up by %e fraction top_value %e\n",sctx.nshift,sctx.shift_amount,sctx.shift_fraction,sctx.shift_top);CHKERRQ(ierr); 779 } else if (info->shiftnz) { 780 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 781 } 782 } 783 PetscFunctionReturn(0); 784 } 785 786 /* 787 This routine implements inplace ILU(0) with row or/and column permutations. 788 Input: 789 A - original matrix 790 Output; 791 A - a->i (rowptr) is same as original rowptr, but factored i-the row is stored in rowperm[i] 792 a->j (col index) is permuted by the inverse of colperm, then sorted 793 a->a reordered accordingly with a->j 794 a->diag (ptr to diagonal elements) is updated. 795 */ 796 #undef __FUNCT__ 797 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_InplaceWithPerm" 798 PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat B,Mat A,const MatFactorInfo *info) 799 { 800 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 801 IS isrow = a->row,isicol = a->icol; 802 PetscErrorCode ierr; 803 const PetscInt *r,*ic,*ics; 804 PetscInt i,j,n=A->rmap->n,*ai=a->i,*aj=a->j; 805 PetscInt *ajtmp,nz,row; 806 PetscInt *diag = a->diag,nbdiag,*pj; 807 PetscScalar *rtmp,*pc,multiplier,d; 808 MatScalar *v,*pv; 809 PetscReal rs; 810 LUShift_Ctx sctx; 811 PetscInt newshift; 812 813 PetscFunctionBegin; 814 if (A != B) SETERRQ(PETSC_ERR_ARG_INCOMP,"input and output matrix must have same address"); 815 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 816 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 817 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&rtmp);CHKERRQ(ierr); 818 ierr = PetscMemzero(rtmp,(n+1)*sizeof(PetscScalar));CHKERRQ(ierr); 819 ics = ic; 820 821 sctx.shift_top = 0.; 822 sctx.nshift_max = 0; 823 sctx.shift_lo = 0.; 824 sctx.shift_hi = 0.; 825 sctx.shift_fraction = 0.; 826 827 /* if both shift schemes are chosen by user, only use info->shiftpd */ 828 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 829 sctx.shift_top = 0.; 830 for (i=0; i<n; i++) { 831 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 832 d = (a->a)[diag[i]]; 833 rs = -PetscAbsScalar(d) - PetscRealPart(d); 834 v = a->a+ai[i]; 835 nz = ai[i+1] - ai[i]; 836 for (j=0; j<nz; j++) 837 rs += PetscAbsScalar(v[j]); 838 if (rs>sctx.shift_top) sctx.shift_top = rs; 839 } 840 if (sctx.shift_top < info->zeropivot) sctx.shift_top = info->zeropivot; 841 sctx.shift_top *= 1.1; 842 sctx.nshift_max = 5; 843 sctx.shift_lo = 0.; 844 sctx.shift_hi = 1.; 845 } 846 847 sctx.shift_amount = 0.; 848 sctx.nshift = 0; 849 do { 850 sctx.lushift = PETSC_FALSE; 851 for (i=0; i<n; i++){ 852 /* load in initial unfactored row */ 853 nz = ai[r[i]+1] - ai[r[i]]; 854 ajtmp = aj + ai[r[i]]; 855 v = a->a + ai[r[i]]; 856 /* sort permuted ajtmp and values v accordingly */ 857 for (j=0; j<nz; j++) ajtmp[j] = ics[ajtmp[j]]; 858 ierr = PetscSortIntWithScalarArray(nz,ajtmp,v);CHKERRQ(ierr); 859 860 diag[r[i]] = ai[r[i]]; 861 for (j=0; j<nz; j++) { 862 rtmp[ajtmp[j]] = v[j]; 863 if (ajtmp[j] < i) diag[r[i]]++; /* update a->diag */ 864 } 865 rtmp[r[i]] += sctx.shift_amount; /* shift the diagonal of the matrix */ 866 867 row = *ajtmp++; 868 while (row < i) { 869 pc = rtmp + row; 870 if (*pc != 0.0) { 871 pv = a->a + diag[r[row]]; 872 pj = aj + diag[r[row]] + 1; 873 874 multiplier = *pc / *pv++; 875 *pc = multiplier; 876 nz = ai[r[row]+1] - diag[r[row]] - 1; 877 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 878 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 879 } 880 row = *ajtmp++; 881 } 882 /* finished row so overwrite it onto a->a */ 883 pv = a->a + ai[r[i]] ; 884 pj = aj + ai[r[i]] ; 885 nz = ai[r[i]+1] - ai[r[i]]; 886 nbdiag = diag[r[i]] - ai[r[i]]; /* num of entries before the diagonal */ 887 888 rs = 0.0; 889 for (j=0; j<nz; j++) { 890 pv[j] = rtmp[pj[j]]; 891 if (j != nbdiag) rs += PetscAbsScalar(pv[j]); 892 } 893 894 /* 9/13/02 Victor Eijkhout suggested scaling zeropivot by rs for matrices with funny scalings */ 895 sctx.rs = rs; 896 sctx.pv = pv[nbdiag]; 897 ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); 898 if (newshift == 1) break; 899 } 900 901 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max) { 902 /* 903 * if no shift in this attempt & shifting & started shifting & can refine, 904 * then try lower shift 905 */ 906 sctx.shift_hi = sctx.shift_fraction; 907 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 908 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 909 sctx.lushift = PETSC_TRUE; 910 sctx.nshift++; 911 } 912 } while (sctx.lushift); 913 914 /* invert diagonal entries for simplier triangular solves */ 915 for (i=0; i<n; i++) { 916 a->a[diag[r[i]]] = 1.0/a->a[diag[r[i]]]; 917 } 918 919 ierr = PetscFree(rtmp);CHKERRQ(ierr); 920 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 921 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 922 A->ops->solve = MatSolve_SeqAIJ_InplaceWithPerm; 923 A->ops->solveadd = MatSolveAdd_SeqAIJ; 924 A->ops->solvetranspose = MatSolveTranspose_SeqAIJ; 925 A->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ; 926 A->assembled = PETSC_TRUE; 927 A->preallocated = PETSC_TRUE; 928 ierr = PetscLogFlops(A->cmap->n);CHKERRQ(ierr); 929 if (sctx.nshift){ 930 if (info->shiftpd) { 931 ierr = PetscInfo4(A,"number of shift_pd tries %D, shift_amount %G, diagonal shifted up by %e fraction top_value %e\n",sctx.nshift,sctx.shift_amount,sctx.shift_fraction,sctx.shift_top);CHKERRQ(ierr); 932 } else if (info->shiftnz) { 933 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 934 } 935 } 936 PetscFunctionReturn(0); 937 } 938 939 /* ----------------------------------------------------------- */ 940 #undef __FUNCT__ 941 #define __FUNCT__ "MatLUFactor_SeqAIJ" 942 PetscErrorCode MatLUFactor_SeqAIJ(Mat A,IS row,IS col,const MatFactorInfo *info) 943 { 944 PetscErrorCode ierr; 945 Mat C; 946 947 PetscFunctionBegin; 948 ierr = MatGetFactor(A,MAT_SOLVER_PETSC,MAT_FACTOR_LU,&C);CHKERRQ(ierr); 949 ierr = MatLUFactorSymbolic(C,A,row,col,info);CHKERRQ(ierr); 950 ierr = MatLUFactorNumeric(C,A,info);CHKERRQ(ierr); 951 A->ops->solve = C->ops->solve; 952 A->ops->solvetranspose = C->ops->solvetranspose; 953 ierr = MatHeaderCopy(A,C);CHKERRQ(ierr); 954 ierr = PetscLogObjectParent(A,((Mat_SeqAIJ*)(A->data))->icol);CHKERRQ(ierr); 955 PetscFunctionReturn(0); 956 } 957 /* ----------------------------------------------------------- */ 958 959 960 #undef __FUNCT__ 961 #define __FUNCT__ "MatSolve_SeqAIJ" 962 PetscErrorCode MatSolve_SeqAIJ(Mat A,Vec bb,Vec xx) 963 { 964 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 965 IS iscol = a->col,isrow = a->row; 966 PetscErrorCode ierr; 967 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 968 PetscInt nz; 969 const PetscInt *rout,*cout,*r,*c; 970 PetscScalar *x,*tmp,*tmps,sum; 971 const PetscScalar *b; 972 const MatScalar *aa = a->a,*v; 973 974 PetscFunctionBegin; 975 if (!n) PetscFunctionReturn(0); 976 977 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 978 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 979 tmp = a->solve_work; 980 981 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 982 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 983 984 /* forward solve the lower triangular */ 985 tmp[0] = b[*r++]; 986 tmps = tmp; 987 for (i=1; i<n; i++) { 988 v = aa + ai[i] ; 989 vi = aj + ai[i] ; 990 nz = a->diag[i] - ai[i]; 991 sum = b[*r++]; 992 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 993 tmp[i] = sum; 994 } 995 996 /* backward solve the upper triangular */ 997 for (i=n-1; i>=0; i--){ 998 v = aa + a->diag[i] + 1; 999 vi = aj + a->diag[i] + 1; 1000 nz = ai[i+1] - a->diag[i] - 1; 1001 sum = tmp[i]; 1002 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1003 x[*c--] = tmp[i] = sum*aa[a->diag[i]]; 1004 } 1005 1006 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1007 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1008 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1009 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1010 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1011 PetscFunctionReturn(0); 1012 } 1013 1014 #undef __FUNCT__ 1015 #define __FUNCT__ "MatMatSolve_SeqAIJ" 1016 PetscErrorCode MatMatSolve_SeqAIJ(Mat A,Mat B,Mat X) 1017 { 1018 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1019 IS iscol = a->col,isrow = a->row; 1020 PetscErrorCode ierr; 1021 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 1022 PetscInt nz,neq; 1023 const PetscInt *rout,*cout,*r,*c; 1024 PetscScalar *x,*b,*tmp,*tmps,sum; 1025 const MatScalar *aa = a->a,*v; 1026 PetscTruth bisdense,xisdense; 1027 1028 PetscFunctionBegin; 1029 if (!n) PetscFunctionReturn(0); 1030 1031 ierr = PetscTypeCompare((PetscObject)B,MATSEQDENSE,&bisdense);CHKERRQ(ierr); 1032 if (!bisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"B matrix must be a SeqDense matrix"); 1033 ierr = PetscTypeCompare((PetscObject)X,MATSEQDENSE,&xisdense);CHKERRQ(ierr); 1034 if (!xisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"X matrix must be a SeqDense matrix"); 1035 1036 ierr = MatGetArray(B,&b);CHKERRQ(ierr); 1037 ierr = MatGetArray(X,&x);CHKERRQ(ierr); 1038 1039 tmp = a->solve_work; 1040 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1041 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1042 1043 for (neq=0; neq<B->cmap->n; neq++){ 1044 /* forward solve the lower triangular */ 1045 tmp[0] = b[r[0]]; 1046 tmps = tmp; 1047 for (i=1; i<n; i++) { 1048 v = aa + ai[i] ; 1049 vi = aj + ai[i] ; 1050 nz = a->diag[i] - ai[i]; 1051 sum = b[r[i]]; 1052 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1053 tmp[i] = sum; 1054 } 1055 /* backward solve the upper triangular */ 1056 for (i=n-1; i>=0; i--){ 1057 v = aa + a->diag[i] + 1; 1058 vi = aj + a->diag[i] + 1; 1059 nz = ai[i+1] - a->diag[i] - 1; 1060 sum = tmp[i]; 1061 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1062 x[c[i]] = tmp[i] = sum*aa[a->diag[i]]; 1063 } 1064 1065 b += n; 1066 x += n; 1067 } 1068 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1069 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1070 ierr = MatRestoreArray(B,&b);CHKERRQ(ierr); 1071 ierr = MatRestoreArray(X,&x);CHKERRQ(ierr); 1072 ierr = PetscLogFlops(B->cmap->n*(2.0*a->nz - n));CHKERRQ(ierr); 1073 PetscFunctionReturn(0); 1074 } 1075 1076 #undef __FUNCT__ 1077 #define __FUNCT__ "MatMatSolve_SeqAIJ_newdatastruct" 1078 PetscErrorCode MatMatSolve_SeqAIJ_newdatastruct(Mat A,Mat B,Mat X) 1079 { 1080 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1081 IS iscol = a->col,isrow = a->row; 1082 PetscErrorCode ierr; 1083 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j,*adiag = a->diag; 1084 PetscInt nz,neq; 1085 const PetscInt *rout,*cout,*r,*c; 1086 PetscScalar *x,*b,*tmp,sum; 1087 const MatScalar *aa = a->a,*v; 1088 PetscTruth bisdense,xisdense; 1089 1090 PetscFunctionBegin; 1091 if (!n) PetscFunctionReturn(0); 1092 1093 ierr = PetscTypeCompare((PetscObject)B,MATSEQDENSE,&bisdense);CHKERRQ(ierr); 1094 if (!bisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"B matrix must be a SeqDense matrix"); 1095 ierr = PetscTypeCompare((PetscObject)X,MATSEQDENSE,&xisdense);CHKERRQ(ierr); 1096 if (!xisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"X matrix must be a SeqDense matrix"); 1097 1098 ierr = MatGetArray(B,&b);CHKERRQ(ierr); 1099 ierr = MatGetArray(X,&x);CHKERRQ(ierr); 1100 1101 tmp = a->solve_work; 1102 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1103 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1104 1105 for (neq=0; neq<B->cmap->n; neq++){ 1106 /* forward solve the lower triangular */ 1107 tmp[0] = b[r[0]]; 1108 v = aa; 1109 vi = aj; 1110 for (i=1; i<n; i++) { 1111 nz = ai[i+1] - ai[i]; 1112 sum = b[r[i]]; 1113 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 1114 tmp[i] = sum; 1115 v += nz; vi += nz; 1116 } 1117 1118 /* backward solve the upper triangular */ 1119 for (i=n-1; i>=0; i--){ 1120 v = aa + adiag[i+1]+1; 1121 vi = aj + adiag[i+1]+1; 1122 nz = adiag[i]-adiag[i+1]-1; 1123 sum = tmp[i]; 1124 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 1125 x[c[i]] = tmp[i] = sum*v[nz]; /* v[nz] = aa[adiag[i]] */ 1126 } 1127 1128 b += n; 1129 x += n; 1130 } 1131 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1132 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1133 ierr = MatRestoreArray(B,&b);CHKERRQ(ierr); 1134 ierr = MatRestoreArray(X,&x);CHKERRQ(ierr); 1135 ierr = PetscLogFlops(B->cmap->n*(2.0*a->nz - n));CHKERRQ(ierr); 1136 PetscFunctionReturn(0); 1137 } 1138 1139 #undef __FUNCT__ 1140 #define __FUNCT__ "MatSolve_SeqAIJ_InplaceWithPerm" 1141 PetscErrorCode MatSolve_SeqAIJ_InplaceWithPerm(Mat A,Vec bb,Vec xx) 1142 { 1143 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1144 IS iscol = a->col,isrow = a->row; 1145 PetscErrorCode ierr; 1146 const PetscInt *r,*c,*rout,*cout; 1147 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 1148 PetscInt nz,row; 1149 PetscScalar *x,*b,*tmp,*tmps,sum; 1150 const MatScalar *aa = a->a,*v; 1151 1152 PetscFunctionBegin; 1153 if (!n) PetscFunctionReturn(0); 1154 1155 ierr = VecGetArray(bb,&b);CHKERRQ(ierr); 1156 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1157 tmp = a->solve_work; 1158 1159 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1160 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1161 1162 /* forward solve the lower triangular */ 1163 tmp[0] = b[*r++]; 1164 tmps = tmp; 1165 for (row=1; row<n; row++) { 1166 i = rout[row]; /* permuted row */ 1167 v = aa + ai[i] ; 1168 vi = aj + ai[i] ; 1169 nz = a->diag[i] - ai[i]; 1170 sum = b[*r++]; 1171 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1172 tmp[row] = sum; 1173 } 1174 1175 /* backward solve the upper triangular */ 1176 for (row=n-1; row>=0; row--){ 1177 i = rout[row]; /* permuted row */ 1178 v = aa + a->diag[i] + 1; 1179 vi = aj + a->diag[i] + 1; 1180 nz = ai[i+1] - a->diag[i] - 1; 1181 sum = tmp[row]; 1182 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1183 x[*c--] = tmp[row] = sum*aa[a->diag[i]]; 1184 } 1185 1186 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1187 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1188 ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr); 1189 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1190 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1191 PetscFunctionReturn(0); 1192 } 1193 1194 /* ----------------------------------------------------------- */ 1195 #include "../src/mat/impls/aij/seq/ftn-kernels/fsolve.h" 1196 #undef __FUNCT__ 1197 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering" 1198 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat A,Vec bb,Vec xx) 1199 { 1200 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1201 PetscErrorCode ierr; 1202 PetscInt n = A->rmap->n; 1203 const PetscInt *ai = a->i,*aj = a->j,*adiag = a->diag; 1204 PetscScalar *x; 1205 const PetscScalar *b; 1206 const MatScalar *aa = a->a; 1207 #if !defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 1208 PetscInt adiag_i,i,nz,ai_i; 1209 const PetscInt *vi; 1210 const MatScalar *v; 1211 PetscScalar sum; 1212 #endif 1213 1214 PetscFunctionBegin; 1215 if (!n) PetscFunctionReturn(0); 1216 1217 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1218 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1219 1220 #if defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 1221 fortransolveaij_(&n,x,ai,aj,adiag,aa,b); 1222 #else 1223 /* forward solve the lower triangular */ 1224 x[0] = b[0]; 1225 for (i=1; i<n; i++) { 1226 ai_i = ai[i]; 1227 v = aa + ai_i; 1228 vi = aj + ai_i; 1229 nz = adiag[i] - ai_i; 1230 sum = b[i]; 1231 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 1232 x[i] = sum; 1233 } 1234 1235 /* backward solve the upper triangular */ 1236 for (i=n-1; i>=0; i--){ 1237 adiag_i = adiag[i]; 1238 v = aa + adiag_i + 1; 1239 vi = aj + adiag_i + 1; 1240 nz = ai[i+1] - adiag_i - 1; 1241 sum = x[i]; 1242 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 1243 x[i] = sum*aa[adiag_i]; 1244 } 1245 #endif 1246 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1247 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1248 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1249 PetscFunctionReturn(0); 1250 } 1251 1252 #undef __FUNCT__ 1253 #define __FUNCT__ "MatSolveAdd_SeqAIJ" 1254 PetscErrorCode MatSolveAdd_SeqAIJ(Mat A,Vec bb,Vec yy,Vec xx) 1255 { 1256 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1257 IS iscol = a->col,isrow = a->row; 1258 PetscErrorCode ierr; 1259 PetscInt i, n = A->rmap->n,j; 1260 PetscInt nz; 1261 const PetscInt *rout,*cout,*r,*c,*vi,*ai = a->i,*aj = a->j; 1262 PetscScalar *x,*tmp,sum; 1263 const PetscScalar *b; 1264 const MatScalar *aa = a->a,*v; 1265 1266 PetscFunctionBegin; 1267 if (yy != xx) {ierr = VecCopy(yy,xx);CHKERRQ(ierr);} 1268 1269 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1270 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1271 tmp = a->solve_work; 1272 1273 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1274 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1275 1276 /* forward solve the lower triangular */ 1277 tmp[0] = b[*r++]; 1278 for (i=1; i<n; i++) { 1279 v = aa + ai[i] ; 1280 vi = aj + ai[i] ; 1281 nz = a->diag[i] - ai[i]; 1282 sum = b[*r++]; 1283 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1284 tmp[i] = sum; 1285 } 1286 1287 /* backward solve the upper triangular */ 1288 for (i=n-1; i>=0; i--){ 1289 v = aa + a->diag[i] + 1; 1290 vi = aj + a->diag[i] + 1; 1291 nz = ai[i+1] - a->diag[i] - 1; 1292 sum = tmp[i]; 1293 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1294 tmp[i] = sum*aa[a->diag[i]]; 1295 x[*c--] += tmp[i]; 1296 } 1297 1298 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1299 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1300 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1301 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1302 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1303 1304 PetscFunctionReturn(0); 1305 } 1306 1307 #undef __FUNCT__ 1308 #define __FUNCT__ "MatSolveAdd_SeqAIJ_newdatastruct" 1309 PetscErrorCode MatSolveAdd_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec yy,Vec xx) 1310 { 1311 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1312 IS iscol = a->col,isrow = a->row; 1313 PetscErrorCode ierr; 1314 PetscInt i, n = A->rmap->n,j; 1315 PetscInt nz; 1316 const PetscInt *rout,*cout,*r,*c,*vi,*ai = a->i,*aj = a->j,*adiag = a->diag; 1317 PetscScalar *x,*tmp,sum; 1318 const PetscScalar *b; 1319 const MatScalar *aa = a->a,*v; 1320 1321 PetscFunctionBegin; 1322 if (yy != xx) {ierr = VecCopy(yy,xx);CHKERRQ(ierr);} 1323 1324 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1325 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1326 tmp = a->solve_work; 1327 1328 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1329 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1330 1331 /* forward solve the lower triangular */ 1332 tmp[0] = b[r[0]]; 1333 v = aa; 1334 vi = aj; 1335 for (i=1; i<n; i++) { 1336 nz = ai[i+1] - ai[i]; 1337 sum = b[r[i]]; 1338 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1339 tmp[i] = sum; 1340 v += nz; vi += nz; 1341 } 1342 1343 /* backward solve the upper triangular */ 1344 v = aa + adiag[n-1]; 1345 vi = aj + adiag[n-1]; 1346 for (i=n-1; i>=0; i--){ 1347 nz = adiag[i] - adiag[i+1] - 1; 1348 sum = tmp[i]; 1349 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1350 tmp[i] = sum*v[nz]; 1351 x[c[i]] += tmp[i]; 1352 v += nz+1; vi += nz+1; 1353 } 1354 1355 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1356 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1357 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1358 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1359 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1360 1361 PetscFunctionReturn(0); 1362 } 1363 1364 #undef __FUNCT__ 1365 #define __FUNCT__ "MatSolveTranspose_SeqAIJ" 1366 PetscErrorCode MatSolveTranspose_SeqAIJ(Mat A,Vec bb,Vec xx) 1367 { 1368 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1369 IS iscol = a->col,isrow = a->row; 1370 PetscErrorCode ierr; 1371 const PetscInt *rout,*cout,*r,*c,*diag = a->diag,*ai = a->i,*aj = a->j,*vi; 1372 PetscInt i,n = A->rmap->n,j; 1373 PetscInt nz; 1374 PetscScalar *x,*tmp,s1; 1375 const MatScalar *aa = a->a,*v; 1376 const PetscScalar *b; 1377 1378 PetscFunctionBegin; 1379 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1380 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1381 tmp = a->solve_work; 1382 1383 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1384 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1385 1386 /* copy the b into temp work space according to permutation */ 1387 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1388 1389 /* forward solve the U^T */ 1390 for (i=0; i<n; i++) { 1391 v = aa + diag[i] ; 1392 vi = aj + diag[i] + 1; 1393 nz = ai[i+1] - diag[i] - 1; 1394 s1 = tmp[i]; 1395 s1 *= (*v++); /* multiply by inverse of diagonal entry */ 1396 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1397 tmp[i] = s1; 1398 } 1399 1400 /* backward solve the L^T */ 1401 for (i=n-1; i>=0; i--){ 1402 v = aa + diag[i] - 1 ; 1403 vi = aj + diag[i] - 1 ; 1404 nz = diag[i] - ai[i]; 1405 s1 = tmp[i]; 1406 for (j=0; j>-nz; j--) tmp[vi[j]] -= s1*v[j]; 1407 } 1408 1409 /* copy tmp into x according to permutation */ 1410 for (i=0; i<n; i++) x[r[i]] = tmp[i]; 1411 1412 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1413 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1414 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1415 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1416 1417 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1418 PetscFunctionReturn(0); 1419 } 1420 1421 #undef __FUNCT__ 1422 #define __FUNCT__ "MatSolveTranspose_SeqAIJ_newdatastruct" 1423 PetscErrorCode MatSolveTranspose_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec xx) 1424 { 1425 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1426 IS iscol = a->col,isrow = a->row; 1427 PetscErrorCode ierr; 1428 const PetscInt *rout,*cout,*r,*c,*adiag = a->diag,*ai = a->i,*aj = a->j,*vi; 1429 PetscInt i,n = A->rmap->n,j; 1430 PetscInt nz; 1431 PetscScalar *x,*tmp,s1; 1432 const MatScalar *aa = a->a,*v; 1433 const PetscScalar *b; 1434 1435 PetscFunctionBegin; 1436 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1437 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1438 tmp = a->solve_work; 1439 1440 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1441 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1442 1443 /* copy the b into temp work space according to permutation */ 1444 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1445 1446 /* forward solve the U^T */ 1447 for (i=0; i<n; i++) { 1448 v = aa + adiag[i+1] + 1; 1449 vi = aj + adiag[i+1] + 1; 1450 nz = adiag[i] - adiag[i+1] - 1; 1451 s1 = tmp[i]; 1452 s1 *= v[nz]; /* multiply by inverse of diagonal entry */ 1453 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1454 tmp[i] = s1; 1455 } 1456 1457 /* backward solve the L^T */ 1458 for (i=n-1; i>=0; i--){ 1459 v = aa + ai[i]; 1460 vi = aj + ai[i]; 1461 nz = ai[i+1] - ai[i]; 1462 s1 = tmp[i]; 1463 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1464 } 1465 1466 /* copy tmp into x according to permutation */ 1467 for (i=0; i<n; i++) x[r[i]] = tmp[i]; 1468 1469 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1470 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1471 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1472 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1473 1474 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1475 PetscFunctionReturn(0); 1476 } 1477 1478 #undef __FUNCT__ 1479 #define __FUNCT__ "MatSolveTransposeAdd_SeqAIJ" 1480 PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat A,Vec bb,Vec zz,Vec xx) 1481 { 1482 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1483 IS iscol = a->col,isrow = a->row; 1484 PetscErrorCode ierr; 1485 const PetscInt *rout,*cout,*r,*c,*diag = a->diag,*ai = a->i,*aj = a->j,*vi; 1486 PetscInt i,n = A->rmap->n,j; 1487 PetscInt nz; 1488 PetscScalar *x,*tmp,s1; 1489 const MatScalar *aa = a->a,*v; 1490 const PetscScalar *b; 1491 1492 PetscFunctionBegin; 1493 if (zz != xx) {ierr = VecCopy(zz,xx);CHKERRQ(ierr);} 1494 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1495 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1496 tmp = a->solve_work; 1497 1498 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1499 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1500 1501 /* copy the b into temp work space according to permutation */ 1502 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1503 1504 /* forward solve the U^T */ 1505 for (i=0; i<n; i++) { 1506 v = aa + diag[i] ; 1507 vi = aj + diag[i] + 1; 1508 nz = ai[i+1] - diag[i] - 1; 1509 s1 = tmp[i]; 1510 s1 *= (*v++); /* multiply by inverse of diagonal entry */ 1511 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1512 tmp[i] = s1; 1513 } 1514 1515 /* backward solve the L^T */ 1516 for (i=n-1; i>=0; i--){ 1517 v = aa + diag[i] - 1 ; 1518 vi = aj + diag[i] - 1 ; 1519 nz = diag[i] - ai[i]; 1520 s1 = tmp[i]; 1521 for (j=0; j>-nz; j--) tmp[vi[j]] -= s1*v[j]; 1522 } 1523 1524 /* copy tmp into x according to permutation */ 1525 for (i=0; i<n; i++) x[r[i]] += tmp[i]; 1526 1527 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1528 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1529 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1530 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1531 1532 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1533 PetscFunctionReturn(0); 1534 } 1535 1536 #undef __FUNCT__ 1537 #define __FUNCT__ "MatSolveTransposeAdd_SeqAIJ_newdatastruct" 1538 PetscErrorCode MatSolveTransposeAdd_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec zz,Vec xx) 1539 { 1540 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1541 IS iscol = a->col,isrow = a->row; 1542 PetscErrorCode ierr; 1543 const PetscInt *rout,*cout,*r,*c,*adiag = a->diag,*ai = a->i,*aj = a->j,*vi; 1544 PetscInt i,n = A->rmap->n,j; 1545 PetscInt nz; 1546 PetscScalar *x,*tmp,s1; 1547 const MatScalar *aa = a->a,*v; 1548 const PetscScalar *b; 1549 1550 PetscFunctionBegin; 1551 if (zz != xx) {ierr = VecCopy(zz,xx);CHKERRQ(ierr);} 1552 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1553 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1554 tmp = a->solve_work; 1555 1556 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1557 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1558 1559 /* copy the b into temp work space according to permutation */ 1560 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1561 1562 /* forward solve the U^T */ 1563 for (i=0; i<n; i++) { 1564 v = aa + adiag[i+1] + 1; 1565 vi = aj + adiag[i+1] + 1; 1566 nz = adiag[i] - adiag[i+1] - 1; 1567 s1 = tmp[i]; 1568 s1 *= v[nz]; /* multiply by inverse of diagonal entry */ 1569 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1570 tmp[i] = s1; 1571 } 1572 1573 1574 /* backward solve the L^T */ 1575 for (i=n-1; i>=0; i--){ 1576 v = aa + ai[i] ; 1577 vi = aj + ai[i]; 1578 nz = ai[i+1] - ai[i]; 1579 s1 = tmp[i]; 1580 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1581 } 1582 1583 /* copy tmp into x according to permutation */ 1584 for (i=0; i<n; i++) x[r[i]] += tmp[i]; 1585 1586 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1587 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1588 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1589 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1590 1591 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1592 PetscFunctionReturn(0); 1593 } 1594 1595 /* ----------------------------------------------------------------*/ 1596 1597 EXTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat,Mat,MatDuplicateOption,PetscTruth); 1598 1599 /* 1600 ilu() under revised new data structure. 1601 Factored arrays bj and ba are stored as 1602 L(0,:), L(1,:), ...,L(n-1,:), U(n-1,:),...,U(i,:),U(i-1,:),...,U(0,:) 1603 1604 bi=fact->i is an array of size n+1, in which 1605 bi+ 1606 bi[i]: points to 1st entry of L(i,:),i=0,...,n-1 1607 bi[n]: points to L(n-1,n-1)+1 1608 1609 bdiag=fact->diag is an array of size n+1,in which 1610 bdiag[i]: points to diagonal of U(i,:), i=0,...,n-1 1611 bdiag[n]: points to entry of U(n-1,0)-1 1612 1613 U(i,:) contains bdiag[i] as its last entry, i.e., 1614 U(i,:) = (u[i,i+1],...,u[i,n-1],diag[i]) 1615 */ 1616 #undef __FUNCT__ 1617 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct" 1618 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1619 { 1620 1621 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1622 PetscErrorCode ierr; 1623 PetscInt n=A->rmap->n,*ai=a->i,*aj,*adiag=a->diag; 1624 PetscInt i,j,nz,*bi,*bj,*bdiag; 1625 PetscTruth missing; 1626 IS isicol; 1627 1628 PetscFunctionBegin; 1629 if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n); 1630 ierr = MatMissingDiagonal(A,&missing,&i);CHKERRQ(ierr); 1631 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",i); 1632 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1633 1634 ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES,PETSC_FALSE);CHKERRQ(ierr); 1635 b = (Mat_SeqAIJ*)(fact)->data; 1636 1637 /* allocate matrix arrays for new data structure */ 1638 ierr = PetscMalloc3(ai[n]+1,PetscScalar,&b->a,ai[n]+1,PetscInt,&b->j,n+1,PetscInt,&b->i);CHKERRQ(ierr); 1639 ierr = PetscLogObjectMemory(fact,ai[n]*(sizeof(PetscScalar)+sizeof(PetscInt))+(n+1)*sizeof(PetscInt));CHKERRQ(ierr); 1640 b->singlemalloc = PETSC_TRUE; 1641 if (!b->diag){ 1642 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&b->diag);CHKERRQ(ierr); 1643 ierr = PetscLogObjectMemory(fact,(n+1)*sizeof(PetscInt));CHKERRQ(ierr); 1644 } 1645 bdiag = b->diag; 1646 1647 if (n > 0) { 1648 ierr = PetscMemzero(b->a,(ai[n])*sizeof(MatScalar));CHKERRQ(ierr); 1649 } 1650 1651 /* set bi and bj with new data structure */ 1652 bi = b->i; 1653 bj = b->j; 1654 1655 /* L part */ 1656 bi[0] = 0; 1657 for (i=0; i<n; i++){ 1658 nz = adiag[i] - ai[i]; 1659 bi[i+1] = bi[i] + nz; 1660 aj = a->j + ai[i]; 1661 for (j=0; j<nz; j++){ 1662 *bj = aj[j]; bj++; 1663 } 1664 } 1665 1666 /* U part */ 1667 bdiag[n] = bi[n]-1; 1668 for (i=n-1; i>=0; i--){ 1669 nz = ai[i+1] - adiag[i] - 1; 1670 aj = a->j + adiag[i] + 1; 1671 for (j=0; j<nz; j++){ 1672 *bj = aj[j]; bj++; 1673 } 1674 /* diag[i] */ 1675 *bj = i; bj++; 1676 bdiag[i] = bdiag[i+1] + nz + 1; 1677 } 1678 1679 fact->factor = MAT_FACTOR_ILU; 1680 fact->info.factor_mallocs = 0; 1681 fact->info.fill_ratio_given = info->fill; 1682 fact->info.fill_ratio_needed = 1.0; 1683 fact->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 1684 1685 b = (Mat_SeqAIJ*)(fact)->data; 1686 b->row = isrow; 1687 b->col = iscol; 1688 b->icol = isicol; 1689 ierr = PetscMalloc((fact->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1690 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1691 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1692 PetscFunctionReturn(0); 1693 } 1694 1695 #undef __FUNCT__ 1696 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_newdatastruct" 1697 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1698 { 1699 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1700 IS isicol; 1701 PetscErrorCode ierr; 1702 const PetscInt *r,*ic; 1703 PetscInt n=A->rmap->n,*ai=a->i,*aj=a->j; 1704 PetscInt *bi,*cols,nnz,*cols_lvl; 1705 PetscInt *bdiag,prow,fm,nzbd,reallocs=0,dcount=0; 1706 PetscInt i,levels,diagonal_fill; 1707 PetscTruth col_identity,row_identity; 1708 PetscReal f; 1709 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL; 1710 PetscBT lnkbt; 1711 PetscInt nzi,*bj,**bj_ptr,**bjlvl_ptr; 1712 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1713 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 1714 1715 PetscFunctionBegin; 1716 /* printf("MatILUFactorSymbolic_SeqAIJ_newdatastruct ...\n"); */ 1717 levels = (PetscInt)info->levels; 1718 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 1719 ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 1720 1721 if (!levels && row_identity && col_identity) { 1722 /* special case: ilu(0) with natural ordering */ 1723 ierr = MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1724 ierr = Mat_CheckInode_FactorLU(fact,PETSC_FALSE);CHKERRQ(ierr); 1725 PetscFunctionReturn(0); 1726 } 1727 1728 if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n); 1729 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1730 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1731 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 1732 1733 /* get new row and diagonal pointers, must be allocated separately because they will be given to the Mat_SeqAIJ and freed separately */ 1734 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1735 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 1736 bi[0] = bdiag[0] = 0; 1737 1738 ierr = PetscMalloc2(n,PetscInt*,&bj_ptr,n,PetscInt*,&bjlvl_ptr);CHKERRQ(ierr); 1739 1740 /* create a linked list for storing column indices of the active row */ 1741 nlnk = n + 1; 1742 ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1743 1744 /* initial FreeSpace size is f*(ai[n]+1) */ 1745 f = info->fill; 1746 diagonal_fill = (PetscInt)info->diagonal_fill; 1747 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 1748 current_space = free_space; 1749 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr); 1750 current_space_lvl = free_space_lvl; 1751 1752 for (i=0; i<n; i++) { 1753 nzi = 0; 1754 /* copy current row into linked list */ 1755 nnz = ai[r[i]+1] - ai[r[i]]; 1756 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 1757 cols = aj + ai[r[i]]; 1758 lnk[i] = -1; /* marker to indicate if diagonal exists */ 1759 ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1760 nzi += nlnk; 1761 1762 /* make sure diagonal entry is included */ 1763 if (diagonal_fill && lnk[i] == -1) { 1764 fm = n; 1765 while (lnk[fm] < i) fm = lnk[fm]; 1766 lnk[i] = lnk[fm]; /* insert diagonal into linked list */ 1767 lnk[fm] = i; 1768 lnk_lvl[i] = 0; 1769 nzi++; dcount++; 1770 } 1771 1772 /* add pivot rows into the active row */ 1773 nzbd = 0; 1774 prow = lnk[n]; 1775 while (prow < i) { 1776 nnz = bdiag[prow]; 1777 cols = bj_ptr[prow] + nnz + 1; 1778 cols_lvl = bjlvl_ptr[prow] + nnz + 1; 1779 nnz = bi[prow+1] - bi[prow] - nnz - 1; 1780 ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr); 1781 nzi += nlnk; 1782 prow = lnk[prow]; 1783 nzbd++; 1784 } 1785 bdiag[i] = nzbd; 1786 bi[i+1] = bi[i] + nzi; 1787 1788 /* if free space is not available, make more free space */ 1789 if (current_space->local_remaining<nzi) { 1790 nnz = 2*nzi*(n - i); /* estimated and max additional space needed */ 1791 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 1792 ierr = PetscFreeSpaceGet(nnz,¤t_space_lvl);CHKERRQ(ierr); 1793 reallocs++; 1794 } 1795 1796 /* copy data into free_space and free_space_lvl, then initialize lnk */ 1797 ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 1798 bj_ptr[i] = current_space->array; 1799 bjlvl_ptr[i] = current_space_lvl->array; 1800 1801 /* make sure the active row i has diagonal entry */ 1802 if (*(bj_ptr[i]+bdiag[i]) != i) { 1803 SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\ 1804 try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i); 1805 } 1806 1807 current_space->array += nzi; 1808 current_space->local_used += nzi; 1809 current_space->local_remaining -= nzi; 1810 current_space_lvl->array += nzi; 1811 current_space_lvl->local_used += nzi; 1812 current_space_lvl->local_remaining -= nzi; 1813 } 1814 1815 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 1816 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 1817 1818 /* destroy list of free space and other temporary arrays */ 1819 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1820 1821 /* copy free_space into bj and free free_space; set bi, bj, bdiag in new datastructure; */ 1822 ierr = PetscFreeSpaceContiguous_LU_v2(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 1823 1824 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1825 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 1826 ierr = PetscFree2(bj_ptr,bjlvl_ptr);CHKERRQ(ierr); 1827 1828 #if defined(PETSC_USE_INFO) 1829 { 1830 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 1831 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 1832 ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 1833 ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr); 1834 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 1835 if (diagonal_fill) { 1836 ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr); 1837 } 1838 } 1839 #endif 1840 1841 /* put together the new matrix */ 1842 ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 1843 ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr); 1844 b = (Mat_SeqAIJ*)(fact)->data; 1845 b->free_a = PETSC_TRUE; 1846 b->free_ij = PETSC_TRUE; 1847 b->singlemalloc = PETSC_FALSE; 1848 ierr = PetscMalloc((bdiag[0]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 1849 b->j = bj; 1850 b->i = bi; 1851 b->diag = bdiag; 1852 b->ilen = 0; 1853 b->imax = 0; 1854 b->row = isrow; 1855 b->col = iscol; 1856 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1857 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1858 b->icol = isicol; 1859 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1860 /* In b structure: Free imax, ilen, old a, old j. 1861 Allocate bdiag, solve_work, new a, new j */ 1862 ierr = PetscLogObjectMemory(fact,(bdiag[0]+1)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 1863 b->maxnz = b->nz = bdiag[0]+1; 1864 (fact)->info.factor_mallocs = reallocs; 1865 (fact)->info.fill_ratio_given = f; 1866 (fact)->info.fill_ratio_needed = ((PetscReal)(bdiag[0]+1))/((PetscReal)ai[n]); 1867 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 1868 ierr = Mat_CheckInode_FactorLU(fact,PETSC_FALSE);CHKERRQ(ierr); 1869 PetscFunctionReturn(0); 1870 } 1871 1872 #undef __FUNCT__ 1873 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ" 1874 PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1875 { 1876 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1877 IS isicol; 1878 PetscErrorCode ierr; 1879 const PetscInt *r,*ic; 1880 PetscInt n=A->rmap->n,*ai=a->i,*aj=a->j,d; 1881 PetscInt *bi,*cols,nnz,*cols_lvl; 1882 PetscInt *bdiag,prow,fm,nzbd,reallocs=0,dcount=0; 1883 PetscInt i,levels,diagonal_fill; 1884 PetscTruth col_identity,row_identity; 1885 PetscReal f; 1886 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL; 1887 PetscBT lnkbt; 1888 PetscInt nzi,*bj,**bj_ptr,**bjlvl_ptr; 1889 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1890 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 1891 PetscTruth missing; 1892 PetscTruth newdatastruct=PETSC_FALSE; 1893 1894 PetscFunctionBegin; 1895 ierr = PetscOptionsGetTruth(PETSC_NULL,"-ilu_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 1896 if(newdatastruct){ 1897 ierr = MatILUFactorSymbolic_SeqAIJ_newdatastruct(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1898 PetscFunctionReturn(0); 1899 } 1900 1901 if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n); 1902 f = info->fill; 1903 levels = (PetscInt)info->levels; 1904 diagonal_fill = (PetscInt)info->diagonal_fill; 1905 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1906 1907 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 1908 ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 1909 if (!levels && row_identity && col_identity) { /* special case: ilu(0) with natural ordering */ 1910 ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES,PETSC_TRUE);CHKERRQ(ierr); 1911 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ; 1912 1913 fact->factor = MAT_FACTOR_ILU; 1914 (fact)->info.factor_mallocs = 0; 1915 (fact)->info.fill_ratio_given = info->fill; 1916 (fact)->info.fill_ratio_needed = 1.0; 1917 b = (Mat_SeqAIJ*)(fact)->data; 1918 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 1919 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 1920 b->row = isrow; 1921 b->col = iscol; 1922 b->icol = isicol; 1923 ierr = PetscMalloc(((fact)->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1924 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1925 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1926 ierr = MatILUFactorSymbolic_SeqAIJ_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1927 PetscFunctionReturn(0); 1928 } 1929 1930 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1931 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 1932 1933 /* get new row and diagonal pointers, must be allocated separately because they will be given to the Mat_SeqAIJ and freed separately */ 1934 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1935 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 1936 bi[0] = bdiag[0] = 0; 1937 1938 ierr = PetscMalloc2(n,PetscInt*,&bj_ptr,n,PetscInt*,&bjlvl_ptr);CHKERRQ(ierr); 1939 1940 /* create a linked list for storing column indices of the active row */ 1941 nlnk = n + 1; 1942 ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1943 1944 /* initial FreeSpace size is f*(ai[n]+1) */ 1945 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 1946 current_space = free_space; 1947 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr); 1948 current_space_lvl = free_space_lvl; 1949 1950 for (i=0; i<n; i++) { 1951 nzi = 0; 1952 /* copy current row into linked list */ 1953 nnz = ai[r[i]+1] - ai[r[i]]; 1954 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 1955 cols = aj + ai[r[i]]; 1956 lnk[i] = -1; /* marker to indicate if diagonal exists */ 1957 ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1958 nzi += nlnk; 1959 1960 /* make sure diagonal entry is included */ 1961 if (diagonal_fill && lnk[i] == -1) { 1962 fm = n; 1963 while (lnk[fm] < i) fm = lnk[fm]; 1964 lnk[i] = lnk[fm]; /* insert diagonal into linked list */ 1965 lnk[fm] = i; 1966 lnk_lvl[i] = 0; 1967 nzi++; dcount++; 1968 } 1969 1970 /* add pivot rows into the active row */ 1971 nzbd = 0; 1972 prow = lnk[n]; 1973 while (prow < i) { 1974 nnz = bdiag[prow]; 1975 cols = bj_ptr[prow] + nnz + 1; 1976 cols_lvl = bjlvl_ptr[prow] + nnz + 1; 1977 nnz = bi[prow+1] - bi[prow] - nnz - 1; 1978 ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr); 1979 nzi += nlnk; 1980 prow = lnk[prow]; 1981 nzbd++; 1982 } 1983 bdiag[i] = nzbd; 1984 bi[i+1] = bi[i] + nzi; 1985 1986 /* if free space is not available, make more free space */ 1987 if (current_space->local_remaining<nzi) { 1988 nnz = nzi*(n - i); /* estimated and max additional space needed */ 1989 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 1990 ierr = PetscFreeSpaceGet(nnz,¤t_space_lvl);CHKERRQ(ierr); 1991 reallocs++; 1992 } 1993 1994 /* copy data into free_space and free_space_lvl, then initialize lnk */ 1995 ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 1996 bj_ptr[i] = current_space->array; 1997 bjlvl_ptr[i] = current_space_lvl->array; 1998 1999 /* make sure the active row i has diagonal entry */ 2000 if (*(bj_ptr[i]+bdiag[i]) != i) { 2001 SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\ 2002 try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i); 2003 } 2004 2005 current_space->array += nzi; 2006 current_space->local_used += nzi; 2007 current_space->local_remaining -= nzi; 2008 current_space_lvl->array += nzi; 2009 current_space_lvl->local_used += nzi; 2010 current_space_lvl->local_remaining -= nzi; 2011 } 2012 2013 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 2014 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 2015 2016 /* destroy list of free space and other temporary arrays */ 2017 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 2018 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); /* copy free_space -> bj */ 2019 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2020 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2021 ierr = PetscFree2(bj_ptr,bjlvl_ptr);CHKERRQ(ierr); 2022 2023 #if defined(PETSC_USE_INFO) 2024 { 2025 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 2026 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 2027 ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2028 ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr); 2029 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 2030 if (diagonal_fill) { 2031 ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr); 2032 } 2033 } 2034 #endif 2035 2036 /* put together the new matrix */ 2037 ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 2038 ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr); 2039 b = (Mat_SeqAIJ*)(fact)->data; 2040 b->free_a = PETSC_TRUE; 2041 b->free_ij = PETSC_TRUE; 2042 b->singlemalloc = PETSC_FALSE; 2043 ierr = PetscMalloc(bi[n]*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 2044 b->j = bj; 2045 b->i = bi; 2046 for (i=0; i<n; i++) bdiag[i] += bi[i]; 2047 b->diag = bdiag; 2048 b->ilen = 0; 2049 b->imax = 0; 2050 b->row = isrow; 2051 b->col = iscol; 2052 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2053 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2054 b->icol = isicol; 2055 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2056 /* In b structure: Free imax, ilen, old a, old j. 2057 Allocate bdiag, solve_work, new a, new j */ 2058 ierr = PetscLogObjectMemory(fact,(bi[n]-n) * (sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 2059 b->maxnz = b->nz = bi[n] ; 2060 (fact)->info.factor_mallocs = reallocs; 2061 (fact)->info.fill_ratio_given = f; 2062 (fact)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]); 2063 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ; 2064 ierr = MatILUFactorSymbolic_SeqAIJ_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); 2065 PetscFunctionReturn(0); 2066 } 2067 2068 #undef __FUNCT__ 2069 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ_newdatastruct" 2070 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 2071 { 2072 Mat C = B; 2073 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 2074 Mat_SeqSBAIJ *b=(Mat_SeqSBAIJ*)C->data; 2075 IS ip=b->row,iip = b->icol; 2076 PetscErrorCode ierr; 2077 const PetscInt *rip,*riip; 2078 PetscInt i,j,mbs=A->rmap->n,*bi=b->i,*bj=b->j,*bdiag=b->diag,*bjtmp; 2079 PetscInt *ai=a->i,*aj=a->j; 2080 PetscInt k,jmin,jmax,*c2r,*il,col,nexti,ili,nz; 2081 MatScalar *rtmp,*ba=b->a,*bval,*aa=a->a,dk,uikdi; 2082 PetscTruth perm_identity; 2083 2084 LUShift_Ctx sctx; 2085 PetscInt newshift; 2086 PetscReal rs; 2087 MatScalar d,*v; 2088 2089 PetscFunctionBegin; 2090 /* MatPivotSetUp(): initialize shift context sctx */ 2091 ierr = PetscMemzero(&sctx,sizeof(LUShift_Ctx));CHKERRQ(ierr); 2092 2093 /* if both shift schemes are chosen by user, only use info->shiftpd */ 2094 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 2095 sctx.shift_top = info->zeropivot; 2096 for (i=0; i<mbs; i++) { 2097 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 2098 d = (aa)[a->diag[i]]; 2099 rs = -PetscAbsScalar(d) - PetscRealPart(d); 2100 v = aa+ai[i]; 2101 nz = ai[i+1] - ai[i]; 2102 for (j=0; j<nz; j++) 2103 rs += PetscAbsScalar(v[j]); 2104 if (rs>sctx.shift_top) sctx.shift_top = rs; 2105 } 2106 sctx.shift_top *= 1.1; 2107 sctx.nshift_max = 5; 2108 sctx.shift_lo = 0.; 2109 sctx.shift_hi = 1.; 2110 } 2111 2112 ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 2113 ierr = ISGetIndices(iip,&riip);CHKERRQ(ierr); 2114 2115 /* allocate working arrays 2116 c2r: linked list, keep track of pivot rows for a given column. c2r[col]: head of the list for a given col 2117 il: for active k row, il[i] gives the index of the 1st nonzero entry in U[i,k:n-1] in bj and ba arrays 2118 */ 2119 ierr = PetscMalloc3(mbs,MatScalar,&rtmp,mbs,PetscInt,&il,mbs,PetscInt,&c2r);CHKERRQ(ierr); 2120 2121 do { 2122 sctx.lushift = PETSC_FALSE; 2123 2124 for (i=0; i<mbs; i++) c2r[i] = mbs; 2125 il[0] = 0; 2126 2127 for (k = 0; k<mbs; k++){ 2128 /* zero rtmp */ 2129 nz = bi[k+1] - bi[k]; 2130 bjtmp = bj + bi[k]; 2131 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 2132 2133 /* load in initial unfactored row */ 2134 bval = ba + bi[k]; 2135 jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 2136 for (j = jmin; j < jmax; j++){ 2137 col = riip[aj[j]]; 2138 if (col >= k){ /* only take upper triangular entry */ 2139 rtmp[col] = aa[j]; 2140 *bval++ = 0.0; /* for in-place factorization */ 2141 } 2142 } 2143 /* shift the diagonal of the matrix: ZeropivotApply() */ 2144 rtmp[k] += sctx.shift_amount; /* shift the diagonal of the matrix */ 2145 2146 /* modify k-th row by adding in those rows i with U(i,k)!=0 */ 2147 dk = rtmp[k]; 2148 i = c2r[k]; /* first row to be added to k_th row */ 2149 2150 while (i < k){ 2151 nexti = c2r[i]; /* next row to be added to k_th row */ 2152 2153 /* compute multiplier, update diag(k) and U(i,k) */ 2154 ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2155 uikdi = - ba[ili]*ba[bdiag[i]]; /* diagonal(k) */ 2156 dk += uikdi*ba[ili]; /* update diag[k] */ 2157 ba[ili] = uikdi; /* -U(i,k) */ 2158 2159 /* add multiple of row i to k-th row */ 2160 jmin = ili + 1; jmax = bi[i+1]; 2161 if (jmin < jmax){ 2162 for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 2163 /* update il and c2r for row i */ 2164 il[i] = jmin; 2165 j = bj[jmin]; c2r[i] = c2r[j]; c2r[j] = i; 2166 } 2167 i = nexti; 2168 } 2169 2170 /* copy data into U(k,:) */ 2171 rs = 0.0; 2172 jmin = bi[k]; jmax = bi[k+1]-1; 2173 if (jmin < jmax) { 2174 for (j=jmin; j<jmax; j++){ 2175 col = bj[j]; ba[j] = rtmp[col]; rs += PetscAbsScalar(ba[j]); 2176 } 2177 /* add the k-th row into il and c2r */ 2178 il[k] = jmin; 2179 i = bj[jmin]; c2r[k] = c2r[i]; c2r[i] = k; 2180 } 2181 2182 /* MatPivotCheck() */ 2183 sctx.rs = rs; 2184 sctx.pv = dk; 2185 if (info->shiftnz){ 2186 ierr = MatPivotCheck_nz(info,sctx,k,newshift);CHKERRQ(ierr); 2187 } else if (info->shiftpd){ 2188 ierr = MatPivotCheck_pd(info,sctx,k,newshift);CHKERRQ(ierr); 2189 } else if (info->shiftinblocks){ 2190 ierr = MatPivotCheck_inblocks(info,sctx,k,newshift);CHKERRQ(ierr); 2191 } else { 2192 ierr = MatPivotCheck_none(info,sctx,k,newshift);CHKERRQ(ierr); 2193 } 2194 dk = sctx.pv; 2195 if (newshift == 1) break; 2196 2197 ba[bdiag[k]] = 1.0/dk; /* U(k,k) */ 2198 } 2199 } while (sctx.lushift); 2200 2201 ierr = PetscFree3(rtmp,il,c2r);CHKERRQ(ierr); 2202 ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 2203 ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr); 2204 2205 ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr); 2206 if (perm_identity){ 2207 (B)->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering_newdatastruct; 2208 (B)->ops->solvetranspose = 0; 2209 (B)->ops->forwardsolve = 0; 2210 (B)->ops->backwardsolve = 0; 2211 } else { 2212 (B)->ops->solve = MatSolve_SeqSBAIJ_1_newdatastruct; 2213 (B)->ops->solvetranspose = 0; 2214 (B)->ops->forwardsolve = 0; 2215 (B)->ops->backwardsolve = 0; 2216 } 2217 2218 C->assembled = PETSC_TRUE; 2219 C->preallocated = PETSC_TRUE; 2220 ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr); 2221 2222 /* MatPivotView() */ 2223 if (sctx.nshift){ 2224 if (info->shiftpd) { 2225 ierr = PetscInfo4(A,"number of shift_pd tries %D, shift_amount %G, diagonal shifted up by %e fraction top_value %e\n",sctx.nshift,sctx.shift_amount,sctx.shift_fraction,sctx.shift_top);CHKERRQ(ierr); 2226 } else if (info->shiftnz) { 2227 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2228 } else if (info->shiftinblocks){ 2229 ierr = PetscInfo2(A,"number of shift_inblocks applied %D, each shift_amount %G\n",sctx.nshift,info->shiftinblocks);CHKERRQ(ierr); 2230 } 2231 } 2232 PetscFunctionReturn(0); 2233 } 2234 2235 #undef __FUNCT__ 2236 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ" 2237 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat B,Mat A,const MatFactorInfo *info) 2238 { 2239 Mat C = B; 2240 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 2241 Mat_SeqSBAIJ *b=(Mat_SeqSBAIJ*)C->data; 2242 IS ip=b->row,iip = b->icol; 2243 PetscErrorCode ierr; 2244 const PetscInt *rip,*riip; 2245 PetscInt i,j,mbs=A->rmap->n,*bi=b->i,*bj=b->j,*bcol,*bjtmp; 2246 PetscInt *ai=a->i,*aj=a->j; 2247 PetscInt k,jmin,jmax,*jl,*il,col,nexti,ili,nz; 2248 MatScalar *rtmp,*ba=b->a,*bval,*aa=a->a,dk,uikdi; 2249 PetscReal zeropivot,rs,shiftnz; 2250 PetscReal shiftpd; 2251 ChShift_Ctx sctx; 2252 PetscInt newshift; 2253 PetscTruth perm_identity; 2254 2255 PetscFunctionBegin; 2256 shiftnz = info->shiftnz; 2257 shiftpd = info->shiftpd; 2258 zeropivot = info->zeropivot; 2259 2260 ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 2261 ierr = ISGetIndices(iip,&riip);CHKERRQ(ierr); 2262 2263 /* initialization */ 2264 ierr = PetscMalloc3(mbs,MatScalar,&rtmp,mbs,PetscInt,&il,mbs,PetscInt,&jl);CHKERRQ(ierr); 2265 sctx.shift_amount = 0; 2266 sctx.nshift = 0; 2267 do { 2268 sctx.chshift = PETSC_FALSE; 2269 for (i=0; i<mbs; i++) jl[i] = mbs; 2270 il[0] = 0; 2271 2272 for (k = 0; k<mbs; k++){ 2273 /* zero rtmp */ 2274 nz = bi[k+1] - bi[k]; 2275 bjtmp = bj + bi[k]; 2276 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 2277 2278 bval = ba + bi[k]; 2279 /* initialize k-th row by the perm[k]-th row of A */ 2280 jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 2281 for (j = jmin; j < jmax; j++){ 2282 col = riip[aj[j]]; 2283 if (col >= k){ /* only take upper triangular entry */ 2284 rtmp[col] = aa[j]; 2285 *bval++ = 0.0; /* for in-place factorization */ 2286 } 2287 } 2288 /* shift the diagonal of the matrix */ 2289 if (sctx.nshift) rtmp[k] += sctx.shift_amount; 2290 2291 /* modify k-th row by adding in those rows i with U(i,k)!=0 */ 2292 dk = rtmp[k]; 2293 i = jl[k]; /* first row to be added to k_th row */ 2294 2295 while (i < k){ 2296 nexti = jl[i]; /* next row to be added to k_th row */ 2297 2298 /* compute multiplier, update diag(k) and U(i,k) */ 2299 ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2300 uikdi = - ba[ili]*ba[bi[i]]; /* diagonal(k) */ 2301 dk += uikdi*ba[ili]; 2302 ba[ili] = uikdi; /* -U(i,k) */ 2303 2304 /* add multiple of row i to k-th row */ 2305 jmin = ili + 1; jmax = bi[i+1]; 2306 if (jmin < jmax){ 2307 for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 2308 /* update il and jl for row i */ 2309 il[i] = jmin; 2310 j = bj[jmin]; jl[i] = jl[j]; jl[j] = i; 2311 } 2312 i = nexti; 2313 } 2314 2315 /* shift the diagonals when zero pivot is detected */ 2316 /* compute rs=sum of abs(off-diagonal) */ 2317 rs = 0.0; 2318 jmin = bi[k]+1; 2319 nz = bi[k+1] - jmin; 2320 bcol = bj + jmin; 2321 for (j=0; j<nz; j++) { 2322 rs += PetscAbsScalar(rtmp[bcol[j]]); 2323 } 2324 2325 sctx.rs = rs; 2326 sctx.pv = dk; 2327 ierr = MatCholeskyCheckShift_inline(info,sctx,k,newshift);CHKERRQ(ierr); 2328 2329 if (newshift == 1) { 2330 if (!sctx.shift_amount) { 2331 sctx.shift_amount = 1e-5; 2332 } 2333 break; 2334 } 2335 2336 /* copy data into U(k,:) */ 2337 ba[bi[k]] = 1.0/dk; /* U(k,k) */ 2338 jmin = bi[k]+1; jmax = bi[k+1]; 2339 if (jmin < jmax) { 2340 for (j=jmin; j<jmax; j++){ 2341 col = bj[j]; ba[j] = rtmp[col]; 2342 } 2343 /* add the k-th row into il and jl */ 2344 il[k] = jmin; 2345 i = bj[jmin]; jl[k] = jl[i]; jl[i] = k; 2346 } 2347 } 2348 } while (sctx.chshift); 2349 ierr = PetscFree3(rtmp,il,jl);CHKERRQ(ierr); 2350 ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 2351 ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr); 2352 2353 ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr); 2354 if (perm_identity){ 2355 (B)->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering; 2356 (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1_NaturalOrdering; 2357 (B)->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1_NaturalOrdering; 2358 (B)->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1_NaturalOrdering; 2359 } else { 2360 (B)->ops->solve = MatSolve_SeqSBAIJ_1; 2361 (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1; 2362 (B)->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1; 2363 (B)->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1; 2364 } 2365 2366 C->assembled = PETSC_TRUE; 2367 C->preallocated = PETSC_TRUE; 2368 ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr); 2369 if (sctx.nshift){ 2370 if (shiftnz) { 2371 ierr = PetscInfo2(A,"number of shiftnz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2372 } else if (shiftpd) { 2373 ierr = PetscInfo2(A,"number of shiftpd tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2374 } 2375 } 2376 PetscFunctionReturn(0); 2377 } 2378 2379 /* 2380 icc() under revised new data structure. 2381 Factored arrays bj and ba are stored as 2382 U(0,:),...,U(i,:),U(n-1,:) 2383 2384 ui=fact->i is an array of size n+1, in which 2385 ui+ 2386 ui[i]: points to 1st entry of U(i,:),i=0,...,n-1 2387 ui[n]: points to U(n-1,n-1)+1 2388 2389 udiag=fact->diag is an array of size n,in which 2390 udiag[i]: points to diagonal of U(i,:), i=0,...,n-1 2391 2392 U(i,:) contains udiag[i] as its last entry, i.e., 2393 U(i,:) = (u[i,i+1],...,u[i,n-1],diag[i]) 2394 */ 2395 2396 #undef __FUNCT__ 2397 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ_newdatastruct" 2398 PetscErrorCode MatICCFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2399 { 2400 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2401 Mat_SeqSBAIJ *b; 2402 PetscErrorCode ierr; 2403 PetscTruth perm_identity,missing; 2404 PetscInt reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui,*udiag; 2405 const PetscInt *rip,*riip; 2406 PetscInt jmin,jmax,nzk,k,j,*jl,prow,*il,nextprow; 2407 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL,d; 2408 PetscInt ncols,ncols_upper,*cols,*ajtmp,*uj,**uj_ptr,**uj_lvl_ptr; 2409 PetscReal fill=info->fill,levels=info->levels; 2410 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2411 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2412 PetscBT lnkbt; 2413 IS iperm; 2414 2415 PetscFunctionBegin; 2416 if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n); 2417 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2418 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2419 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2420 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2421 2422 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2423 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2424 ui[0] = 0; 2425 2426 /* ICC(0) without matrix ordering: simply rearrange column indices */ 2427 if (!levels && perm_identity) { 2428 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2429 cols = uj; 2430 for (i=0; i<am; i++) { 2431 ncols = ai[i+1] - a->diag[i]; 2432 ui[i+1] = ui[i] + ncols; 2433 udiag[i] = ui[i+1] - 1; /* points to the last entry of U(i,:) */ 2434 2435 aj = a->j + a->diag[i] + 1; /* 1st entry of U(i,:) without diagonal */ 2436 ncols--; /* exclude diagonal */ 2437 for (j=0; j<ncols; j++) *cols++ = aj[j]; 2438 *cols++ = i; /* diagoanl is located as the last entry of U(i,:) */ 2439 } 2440 } else { /* case: levels>0 || (levels=0 && !perm_identity) */ 2441 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2442 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2443 2444 /* initialization */ 2445 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ajtmp);CHKERRQ(ierr); 2446 2447 /* jl: linked list for storing indices of the pivot rows 2448 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2449 ierr = PetscMalloc4(am,PetscInt*,&uj_ptr,am,PetscInt*,&uj_lvl_ptr,am,PetscInt,&jl,am,PetscInt,&il);CHKERRQ(ierr); 2450 for (i=0; i<am; i++){ 2451 jl[i] = am; il[i] = 0; 2452 } 2453 2454 /* create and initialize a linked list for storing column indices of the active row k */ 2455 nlnk = am + 1; 2456 ierr = PetscIncompleteLLCreate(am,am,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2457 2458 /* initial FreeSpace size is fill*(ai[am]+1) */ 2459 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2460 current_space = free_space; 2461 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space_lvl);CHKERRQ(ierr); 2462 current_space_lvl = free_space_lvl; 2463 2464 for (k=0; k<am; k++){ /* for each active row k */ 2465 /* initialize lnk by the column indices of row rip[k] of A */ 2466 nzk = 0; 2467 ncols = ai[rip[k]+1] - ai[rip[k]]; 2468 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2469 ncols_upper = 0; 2470 for (j=0; j<ncols; j++){ 2471 i = *(aj + ai[rip[k]] + j); /* unpermuted column index */ 2472 if (riip[i] >= k){ /* only take upper triangular entry */ 2473 ajtmp[ncols_upper] = i; 2474 ncols_upper++; 2475 } 2476 } 2477 ierr = PetscIncompleteLLInit(ncols_upper,ajtmp,am,riip,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2478 nzk += nlnk; 2479 2480 /* update lnk by computing fill-in for each pivot row to be merged in */ 2481 prow = jl[k]; /* 1st pivot row */ 2482 2483 while (prow < k){ 2484 nextprow = jl[prow]; 2485 2486 /* merge prow into k-th row */ 2487 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2488 jmax = ui[prow+1]; 2489 ncols = jmax-jmin; 2490 i = jmin - ui[prow]; 2491 cols = uj_ptr[prow] + i; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2492 uj = uj_lvl_ptr[prow] + i; /* levels of cols */ 2493 j = *(uj - 1); 2494 ierr = PetscICCLLAddSorted(ncols,cols,levels,uj,am,nlnk,lnk,lnk_lvl,lnkbt,j);CHKERRQ(ierr); 2495 nzk += nlnk; 2496 2497 /* update il and jl for prow */ 2498 if (jmin < jmax){ 2499 il[prow] = jmin; 2500 j = *cols; jl[prow] = jl[j]; jl[j] = prow; 2501 } 2502 prow = nextprow; 2503 } 2504 2505 /* if free space is not available, make more free space */ 2506 if (current_space->local_remaining<nzk) { 2507 i = am - k + 1; /* num of unfactored rows */ 2508 i *= PetscMin(nzk, i-1); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2509 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2510 ierr = PetscFreeSpaceGet(i,¤t_space_lvl);CHKERRQ(ierr); 2511 reallocs++; 2512 } 2513 2514 /* copy data into free_space and free_space_lvl, then initialize lnk */ 2515 if (nzk == 0) SETERRQ1(PETSC_ERR_ARG_WRONG,"Empty row %D in ICC matrix factor",k); 2516 ierr = PetscIncompleteLLClean(am,am,nzk,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 2517 2518 /* add the k-th row into il and jl */ 2519 if (nzk > 1){ 2520 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2521 jl[k] = jl[i]; jl[i] = k; 2522 il[k] = ui[k] + 1; 2523 } 2524 uj_ptr[k] = current_space->array; 2525 uj_lvl_ptr[k] = current_space_lvl->array; 2526 2527 current_space->array += nzk; 2528 current_space->local_used += nzk; 2529 current_space->local_remaining -= nzk; 2530 2531 current_space_lvl->array += nzk; 2532 current_space_lvl->local_used += nzk; 2533 current_space_lvl->local_remaining -= nzk; 2534 2535 ui[k+1] = ui[k] + nzk; 2536 } 2537 2538 #if defined(PETSC_USE_INFO) 2539 if (ai[am] != 0) { 2540 PetscReal af = (PetscReal)ui[am]/((PetscReal)ai[am]); 2541 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2542 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2543 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2544 } else { 2545 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2546 } 2547 #endif 2548 2549 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2550 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2551 ierr = PetscFree4(uj_ptr,uj_lvl_ptr,jl,il);CHKERRQ(ierr); 2552 ierr = PetscFree(ajtmp);CHKERRQ(ierr); 2553 2554 /* destroy list of free space and other temporary array(s) */ 2555 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2556 ierr = PetscFreeSpaceContiguous_Cholesky(&free_space,uj,am,ui,udiag);CHKERRQ(ierr); /* store matrix factor in newdatastruct */ 2557 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2558 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2559 2560 } /* end of case: levels>0 || (levels=0 && !perm_identity) */ 2561 2562 /* put together the new matrix in MATSEQSBAIJ format */ 2563 2564 b = (Mat_SeqSBAIJ*)(fact)->data; 2565 b->singlemalloc = PETSC_FALSE; 2566 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2567 b->j = uj; 2568 b->i = ui; 2569 b->diag = udiag; 2570 b->free_diag = PETSC_TRUE; 2571 b->ilen = 0; 2572 b->imax = 0; 2573 b->row = perm; 2574 b->col = perm; 2575 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2576 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2577 b->icol = iperm; 2578 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2579 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2580 ierr = PetscLogObjectMemory((fact),(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2581 b->maxnz = b->nz = ui[am]; 2582 b->free_a = PETSC_TRUE; 2583 b->free_ij = PETSC_TRUE; 2584 2585 (fact)->info.factor_mallocs = reallocs; 2586 (fact)->info.fill_ratio_given = fill; 2587 if (ai[am] != 0) { 2588 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2589 } else { 2590 (fact)->info.fill_ratio_needed = 0.0; 2591 } 2592 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_newdatastruct; 2593 PetscFunctionReturn(0); 2594 } 2595 2596 #undef __FUNCT__ 2597 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ" 2598 PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2599 { 2600 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2601 Mat_SeqSBAIJ *b; 2602 PetscErrorCode ierr; 2603 PetscTruth perm_identity,missing; 2604 PetscInt reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui,*udiag; 2605 const PetscInt *rip,*riip; 2606 PetscInt jmin,jmax,nzk,k,j,*jl,prow,*il,nextprow; 2607 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL,d; 2608 PetscInt ncols,ncols_upper,*cols,*ajtmp,*uj,**uj_ptr,**uj_lvl_ptr; 2609 PetscReal fill=info->fill,levels=info->levels; 2610 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2611 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2612 PetscBT lnkbt; 2613 IS iperm; 2614 PetscTruth newdatastruct=PETSC_FALSE; 2615 2616 PetscFunctionBegin; 2617 ierr = PetscOptionsGetTruth(PETSC_NULL,"-icc_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 2618 if(newdatastruct){ 2619 ierr = MatICCFactorSymbolic_SeqAIJ_newdatastruct(fact,A,perm,info);CHKERRQ(ierr); 2620 PetscFunctionReturn(0); 2621 } 2622 2623 if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n); 2624 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2625 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2626 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2627 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2628 2629 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2630 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2631 ui[0] = 0; 2632 2633 /* ICC(0) without matrix ordering: simply copies fill pattern */ 2634 if (!levels && perm_identity) { 2635 2636 for (i=0; i<am; i++) { 2637 ui[i+1] = ui[i] + ai[i+1] - a->diag[i]; 2638 udiag[i] = ui[i]; 2639 } 2640 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2641 cols = uj; 2642 for (i=0; i<am; i++) { 2643 aj = a->j + a->diag[i]; 2644 ncols = ui[i+1] - ui[i]; 2645 for (j=0; j<ncols; j++) *cols++ = *aj++; 2646 } 2647 } else { /* case: levels>0 || (levels=0 && !perm_identity) */ 2648 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2649 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2650 2651 /* initialization */ 2652 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ajtmp);CHKERRQ(ierr); 2653 2654 /* jl: linked list for storing indices of the pivot rows 2655 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2656 ierr = PetscMalloc4(am,PetscInt*,&uj_ptr,am,PetscInt*,&uj_lvl_ptr,am,PetscInt,&jl,am,PetscInt,&il);CHKERRQ(ierr); 2657 for (i=0; i<am; i++){ 2658 jl[i] = am; il[i] = 0; 2659 } 2660 2661 /* create and initialize a linked list for storing column indices of the active row k */ 2662 nlnk = am + 1; 2663 ierr = PetscIncompleteLLCreate(am,am,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2664 2665 /* initial FreeSpace size is fill*(ai[am]+1) */ 2666 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2667 current_space = free_space; 2668 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space_lvl);CHKERRQ(ierr); 2669 current_space_lvl = free_space_lvl; 2670 2671 for (k=0; k<am; k++){ /* for each active row k */ 2672 /* initialize lnk by the column indices of row rip[k] of A */ 2673 nzk = 0; 2674 ncols = ai[rip[k]+1] - ai[rip[k]]; 2675 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2676 ncols_upper = 0; 2677 for (j=0; j<ncols; j++){ 2678 i = *(aj + ai[rip[k]] + j); /* unpermuted column index */ 2679 if (riip[i] >= k){ /* only take upper triangular entry */ 2680 ajtmp[ncols_upper] = i; 2681 ncols_upper++; 2682 } 2683 } 2684 ierr = PetscIncompleteLLInit(ncols_upper,ajtmp,am,riip,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2685 nzk += nlnk; 2686 2687 /* update lnk by computing fill-in for each pivot row to be merged in */ 2688 prow = jl[k]; /* 1st pivot row */ 2689 2690 while (prow < k){ 2691 nextprow = jl[prow]; 2692 2693 /* merge prow into k-th row */ 2694 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2695 jmax = ui[prow+1]; 2696 ncols = jmax-jmin; 2697 i = jmin - ui[prow]; 2698 cols = uj_ptr[prow] + i; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2699 uj = uj_lvl_ptr[prow] + i; /* levels of cols */ 2700 j = *(uj - 1); 2701 ierr = PetscICCLLAddSorted(ncols,cols,levels,uj,am,nlnk,lnk,lnk_lvl,lnkbt,j);CHKERRQ(ierr); 2702 nzk += nlnk; 2703 2704 /* update il and jl for prow */ 2705 if (jmin < jmax){ 2706 il[prow] = jmin; 2707 j = *cols; jl[prow] = jl[j]; jl[j] = prow; 2708 } 2709 prow = nextprow; 2710 } 2711 2712 /* if free space is not available, make more free space */ 2713 if (current_space->local_remaining<nzk) { 2714 i = am - k + 1; /* num of unfactored rows */ 2715 i *= PetscMin(nzk, (i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2716 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2717 ierr = PetscFreeSpaceGet(i,¤t_space_lvl);CHKERRQ(ierr); 2718 reallocs++; 2719 } 2720 2721 /* copy data into free_space and free_space_lvl, then initialize lnk */ 2722 if (nzk == 0) SETERRQ1(PETSC_ERR_ARG_WRONG,"Empty row %D in ICC matrix factor",k); 2723 ierr = PetscIncompleteLLClean(am,am,nzk,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 2724 2725 /* add the k-th row into il and jl */ 2726 if (nzk > 1){ 2727 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2728 jl[k] = jl[i]; jl[i] = k; 2729 il[k] = ui[k] + 1; 2730 } 2731 uj_ptr[k] = current_space->array; 2732 uj_lvl_ptr[k] = current_space_lvl->array; 2733 2734 current_space->array += nzk; 2735 current_space->local_used += nzk; 2736 current_space->local_remaining -= nzk; 2737 2738 current_space_lvl->array += nzk; 2739 current_space_lvl->local_used += nzk; 2740 current_space_lvl->local_remaining -= nzk; 2741 2742 ui[k+1] = ui[k] + nzk; 2743 } 2744 2745 #if defined(PETSC_USE_INFO) 2746 if (ai[am] != 0) { 2747 PetscReal af = (PetscReal)ui[am]/((PetscReal)ai[am]); 2748 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2749 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2750 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2751 } else { 2752 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2753 } 2754 #endif 2755 2756 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2757 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2758 ierr = PetscFree4(uj_ptr,uj_lvl_ptr,jl,il);CHKERRQ(ierr); 2759 ierr = PetscFree(ajtmp);CHKERRQ(ierr); 2760 2761 /* destroy list of free space and other temporary array(s) */ 2762 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2763 ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr); 2764 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2765 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2766 2767 } /* end of case: levels>0 || (levels=0 && !perm_identity) */ 2768 2769 /* put together the new matrix in MATSEQSBAIJ format */ 2770 2771 b = (Mat_SeqSBAIJ*)(fact)->data; 2772 b->singlemalloc = PETSC_FALSE; 2773 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2774 b->j = uj; 2775 b->i = ui; 2776 b->diag = udiag; 2777 b->free_diag = PETSC_TRUE; 2778 b->ilen = 0; 2779 b->imax = 0; 2780 b->row = perm; 2781 b->col = perm; 2782 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2783 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2784 b->icol = iperm; 2785 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2786 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2787 ierr = PetscLogObjectMemory((fact),(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2788 b->maxnz = b->nz = ui[am]; 2789 b->free_a = PETSC_TRUE; 2790 b->free_ij = PETSC_TRUE; 2791 2792 (fact)->info.factor_mallocs = reallocs; 2793 (fact)->info.fill_ratio_given = fill; 2794 if (ai[am] != 0) { 2795 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2796 } else { 2797 (fact)->info.fill_ratio_needed = 0.0; 2798 } 2799 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ; 2800 PetscFunctionReturn(0); 2801 } 2802 2803 PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2804 { 2805 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2806 Mat_SeqSBAIJ *b; 2807 PetscErrorCode ierr; 2808 PetscTruth perm_identity; 2809 PetscReal fill = info->fill; 2810 const PetscInt *rip,*riip; 2811 PetscInt i,am=A->rmap->n,*ai=a->i,*aj=a->j,reallocs=0,prow; 2812 PetscInt *jl,jmin,jmax,nzk,*ui,k,j,*il,nextprow; 2813 PetscInt nlnk,*lnk,ncols,ncols_upper,*cols,*uj,**ui_ptr,*uj_ptr,*udiag; 2814 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2815 PetscBT lnkbt; 2816 IS iperm; 2817 2818 PetscFunctionBegin; 2819 if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n); 2820 /* check whether perm is the identity mapping */ 2821 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2822 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2823 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2824 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2825 2826 /* initialization */ 2827 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2828 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2829 ui[0] = 0; 2830 2831 /* jl: linked list for storing indices of the pivot rows 2832 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2833 ierr = PetscMalloc4(am,PetscInt*,&ui_ptr,am,PetscInt,&jl,am,PetscInt,&il,am,PetscInt,&cols);CHKERRQ(ierr); 2834 for (i=0; i<am; i++){ 2835 jl[i] = am; il[i] = 0; 2836 } 2837 2838 /* create and initialize a linked list for storing column indices of the active row k */ 2839 nlnk = am + 1; 2840 ierr = PetscLLCreate(am,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2841 2842 /* initial FreeSpace size is fill*(ai[am]+1) */ 2843 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2844 current_space = free_space; 2845 2846 for (k=0; k<am; k++){ /* for each active row k */ 2847 /* initialize lnk by the column indices of row rip[k] of A */ 2848 nzk = 0; 2849 ncols = ai[rip[k]+1] - ai[rip[k]]; 2850 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2851 ncols_upper = 0; 2852 for (j=0; j<ncols; j++){ 2853 i = riip[*(aj + ai[rip[k]] + j)]; 2854 if (i >= k){ /* only take upper triangular entry */ 2855 cols[ncols_upper] = i; 2856 ncols_upper++; 2857 } 2858 } 2859 ierr = PetscLLAdd(ncols_upper,cols,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2860 nzk += nlnk; 2861 2862 /* update lnk by computing fill-in for each pivot row to be merged in */ 2863 prow = jl[k]; /* 1st pivot row */ 2864 2865 while (prow < k){ 2866 nextprow = jl[prow]; 2867 /* merge prow into k-th row */ 2868 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2869 jmax = ui[prow+1]; 2870 ncols = jmax-jmin; 2871 uj_ptr = ui_ptr[prow] + jmin - ui[prow]; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2872 ierr = PetscLLAddSorted(ncols,uj_ptr,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2873 nzk += nlnk; 2874 2875 /* update il and jl for prow */ 2876 if (jmin < jmax){ 2877 il[prow] = jmin; 2878 j = *uj_ptr; jl[prow] = jl[j]; jl[j] = prow; 2879 } 2880 prow = nextprow; 2881 } 2882 2883 /* if free space is not available, make more free space */ 2884 if (current_space->local_remaining<nzk) { 2885 i = am - k + 1; /* num of unfactored rows */ 2886 i *= PetscMin(nzk,i-1); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2887 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2888 reallocs++; 2889 } 2890 2891 /* copy data into free space, then initialize lnk */ 2892 ierr = PetscLLClean(am,am,nzk,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 2893 2894 /* add the k-th row into il and jl */ 2895 if (nzk-1 > 0){ 2896 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2897 jl[k] = jl[i]; jl[i] = k; 2898 il[k] = ui[k] + 1; 2899 } 2900 ui_ptr[k] = current_space->array; 2901 current_space->array += nzk; 2902 current_space->local_used += nzk; 2903 current_space->local_remaining -= nzk; 2904 2905 ui[k+1] = ui[k] + nzk; 2906 } 2907 2908 #if defined(PETSC_USE_INFO) 2909 if (ai[am] != 0) { 2910 PetscReal af = (PetscReal)(ui[am])/((PetscReal)ai[am]); 2911 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2912 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2913 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2914 } else { 2915 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2916 } 2917 #endif 2918 2919 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2920 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2921 ierr = PetscFree4(ui_ptr,jl,il,cols);CHKERRQ(ierr); 2922 2923 /* destroy list of free space and other temporary array(s) */ 2924 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2925 ierr = PetscFreeSpaceContiguous_Cholesky(&free_space,uj,am,ui,udiag);CHKERRQ(ierr); /* store matrix factor in newdatastruct */ 2926 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2927 2928 /* put together the new matrix in MATSEQSBAIJ format */ 2929 2930 b = (Mat_SeqSBAIJ*)(fact)->data; 2931 b->singlemalloc = PETSC_FALSE; 2932 b->free_a = PETSC_TRUE; 2933 b->free_ij = PETSC_TRUE; 2934 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2935 b->j = uj; 2936 b->i = ui; 2937 b->diag = udiag; 2938 b->free_diag = PETSC_TRUE; 2939 b->ilen = 0; 2940 b->imax = 0; 2941 b->row = perm; 2942 b->col = perm; 2943 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2944 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2945 b->icol = iperm; 2946 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2947 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2948 ierr = PetscLogObjectMemory(fact,(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2949 b->maxnz = b->nz = ui[am]; 2950 2951 (fact)->info.factor_mallocs = reallocs; 2952 (fact)->info.fill_ratio_given = fill; 2953 if (ai[am] != 0) { 2954 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2955 } else { 2956 (fact)->info.fill_ratio_needed = 0.0; 2957 } 2958 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_newdatastruct; 2959 PetscFunctionReturn(0); 2960 } 2961 2962 #undef __FUNCT__ 2963 #define __FUNCT__ "MatCholeskyFactorSymbolic_SeqAIJ" 2964 PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2965 { 2966 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2967 Mat_SeqSBAIJ *b; 2968 PetscErrorCode ierr; 2969 PetscTruth perm_identity; 2970 PetscReal fill = info->fill; 2971 const PetscInt *rip,*riip; 2972 PetscInt i,am=A->rmap->n,*ai=a->i,*aj=a->j,reallocs=0,prow; 2973 PetscInt *jl,jmin,jmax,nzk,*ui,k,j,*il,nextprow; 2974 PetscInt nlnk,*lnk,ncols,ncols_upper,*cols,*uj,**ui_ptr,*uj_ptr; 2975 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2976 PetscBT lnkbt; 2977 IS iperm; 2978 PetscTruth newdatastruct=PETSC_FALSE; 2979 2980 PetscFunctionBegin; 2981 ierr = PetscOptionsGetTruth(PETSC_NULL,"-cholesky_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 2982 if(newdatastruct){ 2983 ierr = MatCholeskyFactorSymbolic_SeqAIJ_newdatastruct(fact,A,perm,info);CHKERRQ(ierr); 2984 PetscFunctionReturn(0); 2985 } 2986 2987 if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n); 2988 /* check whether perm is the identity mapping */ 2989 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2990 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2991 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2992 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2993 2994 /* initialization */ 2995 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2996 ui[0] = 0; 2997 2998 /* jl: linked list for storing indices of the pivot rows 2999 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 3000 ierr = PetscMalloc4(am,PetscInt*,&ui_ptr,am,PetscInt,&jl,am,PetscInt,&il,am,PetscInt,&cols);CHKERRQ(ierr); 3001 for (i=0; i<am; i++){ 3002 jl[i] = am; il[i] = 0; 3003 } 3004 3005 /* create and initialize a linked list for storing column indices of the active row k */ 3006 nlnk = am + 1; 3007 ierr = PetscLLCreate(am,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3008 3009 /* initial FreeSpace size is fill*(ai[am]+1) */ 3010 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 3011 current_space = free_space; 3012 3013 for (k=0; k<am; k++){ /* for each active row k */ 3014 /* initialize lnk by the column indices of row rip[k] of A */ 3015 nzk = 0; 3016 ncols = ai[rip[k]+1] - ai[rip[k]]; 3017 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 3018 ncols_upper = 0; 3019 for (j=0; j<ncols; j++){ 3020 i = riip[*(aj + ai[rip[k]] + j)]; 3021 if (i >= k){ /* only take upper triangular entry */ 3022 cols[ncols_upper] = i; 3023 ncols_upper++; 3024 } 3025 } 3026 ierr = PetscLLAdd(ncols_upper,cols,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3027 nzk += nlnk; 3028 3029 /* update lnk by computing fill-in for each pivot row to be merged in */ 3030 prow = jl[k]; /* 1st pivot row */ 3031 3032 while (prow < k){ 3033 nextprow = jl[prow]; 3034 /* merge prow into k-th row */ 3035 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 3036 jmax = ui[prow+1]; 3037 ncols = jmax-jmin; 3038 uj_ptr = ui_ptr[prow] + jmin - ui[prow]; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 3039 ierr = PetscLLAddSorted(ncols,uj_ptr,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3040 nzk += nlnk; 3041 3042 /* update il and jl for prow */ 3043 if (jmin < jmax){ 3044 il[prow] = jmin; 3045 j = *uj_ptr; jl[prow] = jl[j]; jl[j] = prow; 3046 } 3047 prow = nextprow; 3048 } 3049 3050 /* if free space is not available, make more free space */ 3051 if (current_space->local_remaining<nzk) { 3052 i = am - k + 1; /* num of unfactored rows */ 3053 i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 3054 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 3055 reallocs++; 3056 } 3057 3058 /* copy data into free space, then initialize lnk */ 3059 ierr = PetscLLClean(am,am,nzk,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 3060 3061 /* add the k-th row into il and jl */ 3062 if (nzk-1 > 0){ 3063 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 3064 jl[k] = jl[i]; jl[i] = k; 3065 il[k] = ui[k] + 1; 3066 } 3067 ui_ptr[k] = current_space->array; 3068 current_space->array += nzk; 3069 current_space->local_used += nzk; 3070 current_space->local_remaining -= nzk; 3071 3072 ui[k+1] = ui[k] + nzk; 3073 } 3074 3075 #if defined(PETSC_USE_INFO) 3076 if (ai[am] != 0) { 3077 PetscReal af = (PetscReal)(ui[am])/((PetscReal)ai[am]); 3078 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 3079 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 3080 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 3081 } else { 3082 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 3083 } 3084 #endif 3085 3086 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 3087 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 3088 ierr = PetscFree4(ui_ptr,jl,il,cols);CHKERRQ(ierr); 3089 3090 /* destroy list of free space and other temporary array(s) */ 3091 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 3092 ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr); 3093 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 3094 3095 /* put together the new matrix in MATSEQSBAIJ format */ 3096 3097 b = (Mat_SeqSBAIJ*)(fact)->data; 3098 b->singlemalloc = PETSC_FALSE; 3099 b->free_a = PETSC_TRUE; 3100 b->free_ij = PETSC_TRUE; 3101 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 3102 b->j = uj; 3103 b->i = ui; 3104 b->diag = 0; 3105 b->ilen = 0; 3106 b->imax = 0; 3107 b->row = perm; 3108 b->col = perm; 3109 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 3110 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 3111 b->icol = iperm; 3112 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 3113 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 3114 ierr = PetscLogObjectMemory(fact,(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 3115 b->maxnz = b->nz = ui[am]; 3116 3117 (fact)->info.factor_mallocs = reallocs; 3118 (fact)->info.fill_ratio_given = fill; 3119 if (ai[am] != 0) { 3120 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 3121 } else { 3122 (fact)->info.fill_ratio_needed = 0.0; 3123 } 3124 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ; 3125 PetscFunctionReturn(0); 3126 } 3127 3128 #undef __FUNCT__ 3129 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering_newdatastruct" 3130 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_newdatastruct(Mat A,Vec bb,Vec xx) 3131 { 3132 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3133 PetscErrorCode ierr; 3134 PetscInt n = A->rmap->n; 3135 const PetscInt *ai = a->i,*aj = a->j,*adiag = a->diag,*vi; 3136 PetscScalar *x,sum; 3137 const PetscScalar *b; 3138 const MatScalar *aa = a->a,*v; 3139 PetscInt i,nz; 3140 3141 PetscFunctionBegin; 3142 if (!n) PetscFunctionReturn(0); 3143 3144 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3145 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 3146 3147 /* forward solve the lower triangular */ 3148 x[0] = b[0]; 3149 v = aa; 3150 vi = aj; 3151 for (i=1; i<n; i++) { 3152 nz = ai[i+1] - ai[i]; 3153 sum = b[i]; 3154 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 3155 v += nz; 3156 vi += nz; 3157 x[i] = sum; 3158 } 3159 3160 /* backward solve the upper triangular */ 3161 for (i=n-1; i>=0; i--){ 3162 v = aa + adiag[i+1] + 1; 3163 vi = aj + adiag[i+1] + 1; 3164 nz = adiag[i] - adiag[i+1]-1; 3165 sum = x[i]; 3166 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 3167 x[i] = sum*v[nz]; /* x[i]=aa[adiag[i]]*sum; v++; */ 3168 } 3169 3170 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 3171 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3172 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 3173 PetscFunctionReturn(0); 3174 } 3175 3176 #undef __FUNCT__ 3177 #define __FUNCT__ "MatSolve_SeqAIJ_newdatastruct" 3178 PetscErrorCode MatSolve_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec xx) 3179 { 3180 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3181 IS iscol = a->col,isrow = a->row; 3182 PetscErrorCode ierr; 3183 PetscInt i,n=A->rmap->n,*vi,*ai=a->i,*aj=a->j,*adiag = a->diag,nz; 3184 const PetscInt *rout,*cout,*r,*c; 3185 PetscScalar *x,*tmp,sum; 3186 const PetscScalar *b; 3187 const MatScalar *aa = a->a,*v; 3188 3189 PetscFunctionBegin; 3190 if (!n) PetscFunctionReturn(0); 3191 3192 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3193 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 3194 tmp = a->solve_work; 3195 3196 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 3197 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 3198 3199 /* forward solve the lower triangular */ 3200 tmp[0] = b[r[0]]; 3201 v = aa; 3202 vi = aj; 3203 for (i=1; i<n; i++) { 3204 nz = ai[i+1] - ai[i]; 3205 sum = b[r[i]]; 3206 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 3207 tmp[i] = sum; 3208 v += nz; vi += nz; 3209 } 3210 3211 /* backward solve the upper triangular */ 3212 for (i=n-1; i>=0; i--){ 3213 v = aa + adiag[i+1]+1; 3214 vi = aj + adiag[i+1]+1; 3215 nz = adiag[i]-adiag[i+1]-1; 3216 sum = tmp[i]; 3217 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 3218 x[c[i]] = tmp[i] = sum*v[nz]; /* v[nz] = aa[adiag[i]] */ 3219 } 3220 3221 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 3222 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 3223 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3224 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 3225 ierr = PetscLogFlops(2*a->nz - A->cmap->n);CHKERRQ(ierr); 3226 PetscFunctionReturn(0); 3227 } 3228 3229 #undef __FUNCT__ 3230 #define __FUNCT__ "MatILUDTFactor_SeqAIJ" 3231 PetscErrorCode MatILUDTFactor_SeqAIJ(Mat A,IS isrow,IS iscol,const MatFactorInfo *info,Mat *fact) 3232 { 3233 Mat B = *fact; 3234 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b; 3235 IS isicol; 3236 PetscErrorCode ierr; 3237 const PetscInt *r,*ic; 3238 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j,*ajtmp,*adiag; 3239 PetscInt *bi,*bj,*bdiag,*bdiag_rev; 3240 PetscInt row,nzi,nzi_bl,nzi_bu,*im,nzi_al,nzi_au; 3241 PetscInt nlnk,*lnk; 3242 PetscBT lnkbt; 3243 PetscTruth row_identity,icol_identity,both_identity; 3244 MatScalar *aatmp,*pv,*batmp,*ba,*rtmp,*pc,multiplier,*vtmp,diag_tmp; 3245 const PetscInt *ics; 3246 PetscInt j,nz,*pj,*bjtmp,k,ncut,*jtmp; 3247 PetscReal dt=info->dt,dtcol=info->dtcol,shift=info->shiftinblocks; 3248 PetscInt dtcount=(PetscInt)info->dtcount,nnz_max; 3249 PetscTruth missing; 3250 3251 PetscFunctionBegin; 3252 3253 if (dt == PETSC_DEFAULT) dt = 0.005; 3254 if (dtcol == PETSC_DEFAULT) dtcol = 0.01; /* XXX unused! */ 3255 if (dtcount == PETSC_DEFAULT) dtcount = (PetscInt)(1.5*a->rmax); 3256 3257 /* ------- symbolic factorization, can be reused ---------*/ 3258 ierr = MatMissingDiagonal(A,&missing,&i);CHKERRQ(ierr); 3259 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",i); 3260 adiag=a->diag; 3261 3262 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 3263 3264 /* bdiag is location of diagonal in factor */ 3265 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); /* becomes b->diag */ 3266 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag_rev);CHKERRQ(ierr); /* temporary */ 3267 3268 /* allocate row pointers bi */ 3269 ierr = PetscMalloc((2*n+2)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 3270 3271 /* allocate bj and ba; max num of nonzero entries is (ai[n]+2*n*dtcount+2) */ 3272 if (dtcount > n-1) dtcount = n-1; /* diagonal is excluded */ 3273 nnz_max = ai[n]+2*n*dtcount+2; 3274 3275 ierr = PetscMalloc((nnz_max+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 3276 ierr = PetscMalloc((nnz_max+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 3277 3278 /* put together the new matrix */ 3279 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 3280 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 3281 b = (Mat_SeqAIJ*)(B)->data; 3282 b->free_a = PETSC_TRUE; 3283 b->free_ij = PETSC_TRUE; 3284 b->singlemalloc = PETSC_FALSE; 3285 b->a = ba; 3286 b->j = bj; 3287 b->i = bi; 3288 b->diag = bdiag; 3289 b->ilen = 0; 3290 b->imax = 0; 3291 b->row = isrow; 3292 b->col = iscol; 3293 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 3294 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 3295 b->icol = isicol; 3296 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 3297 3298 ierr = PetscLogObjectMemory(B,nnz_max*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 3299 b->maxnz = nnz_max; 3300 3301 (B)->factor = MAT_FACTOR_ILUDT; 3302 (B)->info.factor_mallocs = 0; 3303 (B)->info.fill_ratio_given = ((PetscReal)nnz_max)/((PetscReal)ai[n]); 3304 CHKMEMQ; 3305 /* ------- end of symbolic factorization ---------*/ 3306 3307 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 3308 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 3309 ics = ic; 3310 3311 /* linked list for storing column indices of the active row */ 3312 nlnk = n + 1; 3313 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3314 3315 /* im: used by PetscLLAddSortedLU(); jtmp: working array for column indices of active row */ 3316 ierr = PetscMalloc2(n,PetscInt,&im,n,PetscInt,&jtmp);CHKERRQ(ierr); 3317 /* rtmp, vtmp: working arrays for sparse and contiguous row entries of active row */ 3318 ierr = PetscMalloc2(n,MatScalar,&rtmp,n,MatScalar,&vtmp);CHKERRQ(ierr); 3319 ierr = PetscMemzero(rtmp,n*sizeof(MatScalar));CHKERRQ(ierr); 3320 3321 bi[0] = 0; 3322 bdiag[0] = nnz_max-1; /* location of diag[0] in factor B */ 3323 bdiag_rev[n] = bdiag[0]; 3324 bi[2*n+1] = bdiag[0]+1; /* endof bj and ba array */ 3325 for (i=0; i<n; i++) { 3326 /* copy initial fill into linked list */ 3327 nzi = 0; /* nonzeros for active row i */ 3328 nzi = ai[r[i]+1] - ai[r[i]]; 3329 if (!nzi) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 3330 nzi_al = adiag[r[i]] - ai[r[i]]; 3331 nzi_au = ai[r[i]+1] - adiag[r[i]] -1; 3332 ajtmp = aj + ai[r[i]]; 3333 ierr = PetscLLAddPerm(nzi,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3334 3335 /* load in initial (unfactored row) */ 3336 aatmp = a->a + ai[r[i]]; 3337 for (j=0; j<nzi; j++) { 3338 rtmp[ics[*ajtmp++]] = *aatmp++; 3339 } 3340 3341 /* add pivot rows into linked list */ 3342 row = lnk[n]; 3343 while (row < i ) { 3344 nzi_bl = bi[row+1] - bi[row] + 1; 3345 bjtmp = bj + bdiag[row+1]+1; /* points to 1st column next to the diagonal in U */ 3346 ierr = PetscLLAddSortedLU(bjtmp,row,nlnk,lnk,lnkbt,i,nzi_bl,im);CHKERRQ(ierr); 3347 nzi += nlnk; 3348 row = lnk[row]; 3349 } 3350 3351 /* copy data from lnk into jtmp, then initialize lnk */ 3352 ierr = PetscLLClean(n,n,nzi,lnk,jtmp,lnkbt);CHKERRQ(ierr); 3353 3354 /* numerical factorization */ 3355 bjtmp = jtmp; 3356 row = *bjtmp++; /* 1st pivot row */ 3357 while ( row < i ) { 3358 pc = rtmp + row; 3359 pv = ba + bdiag[row]; /* 1./(diag of the pivot row) */ 3360 multiplier = (*pc) * (*pv); 3361 *pc = multiplier; 3362 if (PetscAbsScalar(*pc) > dt){ /* apply tolerance dropping rule */ 3363 pj = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */ 3364 pv = ba + bdiag[row+1] + 1; 3365 /* if (multiplier < -1.0 or multiplier >1.0) printf("row/prow %d, %d, multiplier %g\n",i,row,multiplier); */ 3366 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */ 3367 for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++); 3368 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 3369 } 3370 row = *bjtmp++; 3371 } 3372 3373 /* copy sparse rtmp into contiguous vtmp; separate L and U part */ 3374 diag_tmp = rtmp[i]; /* save diagonal value - may not needed?? */ 3375 nzi_bl = 0; j = 0; 3376 while (jtmp[j] < i){ /* Note: jtmp is sorted */ 3377 vtmp[j] = rtmp[jtmp[j]]; rtmp[jtmp[j]]=0.0; 3378 nzi_bl++; j++; 3379 } 3380 nzi_bu = nzi - nzi_bl -1; 3381 while (j < nzi){ 3382 vtmp[j] = rtmp[jtmp[j]]; rtmp[jtmp[j]]=0.0; 3383 j++; 3384 } 3385 3386 bjtmp = bj + bi[i]; 3387 batmp = ba + bi[i]; 3388 /* apply level dropping rule to L part */ 3389 ncut = nzi_al + dtcount; 3390 if (ncut < nzi_bl){ 3391 ierr = PetscSortSplit(ncut,nzi_bl,vtmp,jtmp);CHKERRQ(ierr); 3392 ierr = PetscSortIntWithScalarArray(ncut,jtmp,vtmp);CHKERRQ(ierr); 3393 } else { 3394 ncut = nzi_bl; 3395 } 3396 for (j=0; j<ncut; j++){ 3397 bjtmp[j] = jtmp[j]; 3398 batmp[j] = vtmp[j]; 3399 /* printf(" (%d,%g),",bjtmp[j],batmp[j]); */ 3400 } 3401 bi[i+1] = bi[i] + ncut; 3402 nzi = ncut + 1; 3403 3404 /* apply level dropping rule to U part */ 3405 ncut = nzi_au + dtcount; 3406 if (ncut < nzi_bu){ 3407 ierr = PetscSortSplit(ncut,nzi_bu,vtmp+nzi_bl+1,jtmp+nzi_bl+1);CHKERRQ(ierr); 3408 ierr = PetscSortIntWithScalarArray(ncut,jtmp+nzi_bl+1,vtmp+nzi_bl+1);CHKERRQ(ierr); 3409 } else { 3410 ncut = nzi_bu; 3411 } 3412 nzi += ncut; 3413 3414 /* mark bdiagonal */ 3415 bdiag[i+1] = bdiag[i] - (ncut + 1); 3416 bdiag_rev[n-i-1] = bdiag[i+1]; 3417 bi[2*n - i] = bi[2*n - i +1] - (ncut + 1); 3418 bjtmp = bj + bdiag[i]; 3419 batmp = ba + bdiag[i]; 3420 *bjtmp = i; 3421 *batmp = diag_tmp; /* rtmp[i]; */ 3422 if (*batmp == 0.0) { 3423 *batmp = dt+shift; 3424 /* printf(" row %d add shift %g\n",i,shift); */ 3425 } 3426 *batmp = 1.0/(*batmp); /* invert diagonal entries for simplier triangular solves */ 3427 /* printf(" (%d,%g),",*bjtmp,*batmp); */ 3428 3429 bjtmp = bj + bdiag[i+1]+1; 3430 batmp = ba + bdiag[i+1]+1; 3431 for (k=0; k<ncut; k++){ 3432 bjtmp[k] = jtmp[nzi_bl+1+k]; 3433 batmp[k] = vtmp[nzi_bl+1+k]; 3434 /* printf(" (%d,%g),",bjtmp[k],batmp[k]); */ 3435 } 3436 /* printf("\n"); */ 3437 3438 im[i] = nzi; /* used by PetscLLAddSortedLU() */ 3439 /* 3440 printf("row %d: bi %d, bdiag %d\n",i,bi[i],bdiag[i]); 3441 printf(" ----------------------------\n"); 3442 */ 3443 } /* for (i=0; i<n; i++) */ 3444 /* printf("end of L %d, beginning of U %d\n",bi[n],bdiag[n]); */ 3445 if (bi[n] >= bdiag[n]) SETERRQ2(PETSC_ERR_ARG_SIZ,"end of L array %d cannot >= the beginning of U array %d",bi[n],bdiag[n]); 3446 3447 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 3448 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 3449 3450 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 3451 ierr = PetscFree2(im,jtmp);CHKERRQ(ierr); 3452 ierr = PetscFree2(rtmp,vtmp);CHKERRQ(ierr); 3453 ierr = PetscFree(bdiag_rev);CHKERRQ(ierr); 3454 3455 ierr = PetscLogFlops(B->cmap->n);CHKERRQ(ierr); 3456 b->maxnz = b->nz = bi[n] + bdiag[0] - bdiag[n]; 3457 3458 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 3459 ierr = ISIdentity(isicol,&icol_identity);CHKERRQ(ierr); 3460 both_identity = (PetscTruth) (row_identity && icol_identity); 3461 if (row_identity && icol_identity) { 3462 B->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 3463 } else { 3464 B->ops->solve = MatSolve_SeqAIJ_newdatastruct; 3465 } 3466 3467 B->ops->lufactorsymbolic = MatILUDTFactorSymbolic_SeqAIJ; 3468 B->ops->lufactornumeric = MatILUDTFactorNumeric_SeqAIJ; 3469 B->ops->solveadd = 0; 3470 B->ops->solvetranspose = 0; 3471 B->ops->solvetransposeadd = 0; 3472 B->ops->matsolve = 0; 3473 B->assembled = PETSC_TRUE; 3474 B->preallocated = PETSC_TRUE; 3475 PetscFunctionReturn(0); 3476 } 3477 3478 /* a wraper of MatILUDTFactor_SeqAIJ() */ 3479 #undef __FUNCT__ 3480 #define __FUNCT__ "MatILUDTFactorSymbolic_SeqAIJ" 3481 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS row,IS col,const MatFactorInfo *info) 3482 { 3483 PetscErrorCode ierr; 3484 3485 PetscFunctionBegin; 3486 ierr = MatILUDTFactor_SeqAIJ(A,row,col,info,&fact);CHKERRQ(ierr); 3487 3488 fact->ops->lufactornumeric = MatILUDTFactorNumeric_SeqAIJ; 3489 PetscFunctionReturn(0); 3490 } 3491 3492 /* 3493 same as MatLUFactorNumeric_SeqAIJ(), except using contiguous array matrix factors 3494 - intend to replace existing MatLUFactorNumeric_SeqAIJ() 3495 */ 3496 #undef __FUNCT__ 3497 #define __FUNCT__ "MatILUDTFactorNumeric_SeqAIJ" 3498 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorNumeric_SeqAIJ(Mat fact,Mat A,const MatFactorInfo *info) 3499 { 3500 Mat C=fact; 3501 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 3502 IS isrow = b->row,isicol = b->icol; 3503 PetscErrorCode ierr; 3504 const PetscInt *r,*ic,*ics; 3505 PetscInt i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 3506 PetscInt *ajtmp,*bjtmp,nz,nzl,nzu,row,*bdiag = b->diag,*pj; 3507 MatScalar *rtmp,*pc,multiplier,*v,*pv,*aa=a->a; 3508 PetscReal dt=info->dt,shift=info->shiftinblocks; 3509 PetscTruth row_identity, col_identity; 3510 3511 PetscFunctionBegin; 3512 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 3513 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 3514 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 3515 ics = ic; 3516 3517 for (i=0; i<n; i++){ 3518 /* initialize rtmp array */ 3519 nzl = bi[i+1] - bi[i]; /* num of nozeros in L(i,:) */ 3520 bjtmp = bj + bi[i]; 3521 for (j=0; j<nzl; j++) rtmp[*bjtmp++] = 0.0; 3522 rtmp[i] = 0.0; 3523 nzu = bdiag[i] - bdiag[i+1]; /* num of nozeros in U(i,:) */ 3524 bjtmp = bj + bdiag[i+1] + 1; 3525 for (j=0; j<nzu; j++) rtmp[*bjtmp++] = 0.0; 3526 3527 /* load in initial unfactored row of A */ 3528 /* printf("row %d\n",i); */ 3529 nz = ai[r[i]+1] - ai[r[i]]; 3530 ajtmp = aj + ai[r[i]]; 3531 v = aa + ai[r[i]]; 3532 for (j=0; j<nz; j++) { 3533 rtmp[ics[*ajtmp++]] = v[j]; 3534 /* printf(" (%d,%g),",ics[ajtmp[j]],rtmp[ics[ajtmp[j]]]); */ 3535 } 3536 /* printf("\n"); */ 3537 3538 /* numerical factorization */ 3539 bjtmp = bj + bi[i]; /* point to 1st entry of L(i,:) */ 3540 nzl = bi[i+1] - bi[i]; /* num of entries in L(i,:) */ 3541 k = 0; 3542 while (k < nzl){ 3543 row = *bjtmp++; 3544 /* printf(" prow %d\n",row); */ 3545 pc = rtmp + row; 3546 pv = b->a + bdiag[row]; /* 1./(diag of the pivot row) */ 3547 multiplier = (*pc) * (*pv); 3548 *pc = multiplier; 3549 if (PetscAbsScalar(multiplier) > dt){ 3550 pj = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */ 3551 pv = b->a + bdiag[row+1] + 1; 3552 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */ 3553 for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++); 3554 /* ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); */ 3555 } 3556 k++; 3557 } 3558 3559 /* finished row so stick it into b->a */ 3560 /* L-part */ 3561 pv = b->a + bi[i] ; 3562 pj = bj + bi[i] ; 3563 nzl = bi[i+1] - bi[i]; 3564 for (j=0; j<nzl; j++) { 3565 pv[j] = rtmp[pj[j]]; 3566 /* printf(" (%d,%g),",pj[j],pv[j]); */ 3567 } 3568 3569 /* diagonal: invert diagonal entries for simplier triangular solves */ 3570 if (rtmp[i] == 0.0) rtmp[i] = dt+shift; 3571 b->a[bdiag[i]] = 1.0/rtmp[i]; 3572 /* printf(" (%d,%g),",i,b->a[bdiag[i]]); */ 3573 3574 /* U-part */ 3575 pv = b->a + bdiag[i+1] + 1; 3576 pj = bj + bdiag[i+1] + 1; 3577 nzu = bdiag[i] - bdiag[i+1] - 1; 3578 for (j=0; j<nzu; j++) { 3579 pv[j] = rtmp[pj[j]]; 3580 /* printf(" (%d,%g),",pj[j],pv[j]); */ 3581 } 3582 /* printf("\n"); */ 3583 } 3584 3585 ierr = PetscFree(rtmp);CHKERRQ(ierr); 3586 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 3587 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 3588 3589 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 3590 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 3591 if (row_identity && col_identity) { 3592 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 3593 } else { 3594 C->ops->solve = MatSolve_SeqAIJ_newdatastruct; 3595 } 3596 C->ops->solveadd = 0; 3597 C->ops->solvetranspose = 0; 3598 C->ops->solvetransposeadd = 0; 3599 C->ops->matsolve = 0; 3600 C->assembled = PETSC_TRUE; 3601 C->preallocated = PETSC_TRUE; 3602 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 3603 PetscFunctionReturn(0); 3604 } 3605