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_newdatastruct; 120 (*B)->ops->lufactorsymbolic = MatLUFactorSymbolic_SeqAIJ_newdatastruct; 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_newdatastruct; 126 (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJ_newdatastruct; 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_inplace" 135 PetscErrorCode MatLUFactorSymbolic_SeqAIJ_inplace(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; 148 149 PetscFunctionBegin; 150 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 151 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 152 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 153 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 154 155 /* get new row pointers */ 156 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 157 bi[0] = 0; 158 159 /* bdiag is location of diagonal in factor */ 160 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 161 bdiag[0] = 0; 162 163 /* linked list for storing column indices of the active row */ 164 nlnk = n + 1; 165 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 166 167 ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr); 168 169 /* initial FreeSpace size is f*(ai[n]+1) */ 170 f = info->fill; 171 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 172 current_space = free_space; 173 174 for (i=0; i<n; i++) { 175 /* copy previous fill into linked list */ 176 nzi = 0; 177 nnz = ai[r[i]+1] - ai[r[i]]; 178 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 179 ajtmp = aj + ai[r[i]]; 180 ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 181 nzi += nlnk; 182 183 /* add pivot rows into linked list */ 184 row = lnk[n]; 185 while (row < i) { 186 nzbd = bdiag[row] - bi[row] + 1; /* num of entries in the row with column index <= row */ 187 ajtmp = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */ 188 ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr); 189 nzi += nlnk; 190 row = lnk[row]; 191 } 192 bi[i+1] = bi[i] + nzi; 193 im[i] = nzi; 194 195 /* mark bdiag */ 196 nzbd = 0; 197 nnz = nzi; 198 k = lnk[n]; 199 while (nnz-- && k < i){ 200 nzbd++; 201 k = lnk[k]; 202 } 203 bdiag[i] = bi[i] + nzbd; 204 205 /* if free space is not available, make more free space */ 206 if (current_space->local_remaining<nzi) { 207 nnz = (n - i)*nzi; /* estimated and max additional space needed */ 208 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 209 reallocs++; 210 } 211 212 /* copy data into free space, then initialize lnk */ 213 ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 214 bi_ptr[i] = current_space->array; 215 current_space->array += nzi; 216 current_space->local_used += nzi; 217 current_space->local_remaining -= nzi; 218 } 219 #if defined(PETSC_USE_INFO) 220 if (ai[n] != 0) { 221 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 222 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 223 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 224 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr); 225 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 226 } else { 227 ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr); 228 } 229 #endif 230 231 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 232 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 233 234 /* destroy list of free space and other temporary array(s) */ 235 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 236 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 237 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 238 ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr); 239 240 /* put together the new matrix */ 241 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 242 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 243 b = (Mat_SeqAIJ*)(B)->data; 244 b->free_a = PETSC_TRUE; 245 b->free_ij = PETSC_TRUE; 246 b->singlemalloc = PETSC_FALSE; 247 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 248 b->j = bj; 249 b->i = bi; 250 b->diag = bdiag; 251 b->ilen = 0; 252 b->imax = 0; 253 b->row = isrow; 254 b->col = iscol; 255 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 256 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 257 b->icol = isicol; 258 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 259 260 /* In b structure: Free imax, ilen, old a, old j. Allocate solve_work, new a, new j */ 261 ierr = PetscLogObjectMemory(B,(bi[n]-n)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 262 b->maxnz = b->nz = bi[n] ; 263 264 (B)->factor = MAT_FACTOR_LU; 265 (B)->info.factor_mallocs = reallocs; 266 (B)->info.fill_ratio_given = f; 267 268 if (ai[n]) { 269 (B)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]); 270 } else { 271 (B)->info.fill_ratio_needed = 0.0; 272 } 273 (B)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_inplace; 274 (B)->ops->solve = MatSolve_SeqAIJ_inplace; 275 (B)->ops->solvetranspose = MatSolveTranspose_SeqAIJ_inplace; 276 /* switch to inodes if appropriate */ 277 ierr = MatLUFactorSymbolic_SeqAIJ_Inode(B,A,isrow,iscol,info);CHKERRQ(ierr); 278 PetscFunctionReturn(0); 279 } 280 281 #undef __FUNCT__ 282 #define __FUNCT__ "MatLUFactorSymbolic_SeqAIJ_newdatastruct" 283 PetscErrorCode MatLUFactorSymbolic_SeqAIJ_newdatastruct(Mat B,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 284 { 285 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 286 IS isicol; 287 PetscErrorCode ierr; 288 const PetscInt *r,*ic; 289 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j; 290 PetscInt *bi,*bj,*ajtmp; 291 PetscInt *bdiag,row,nnz,nzi,reallocs=0,nzbd,*im; 292 PetscReal f; 293 PetscInt nlnk,*lnk,k,**bi_ptr; 294 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 295 PetscBT lnkbt; 296 PetscTruth newdatastruct= PETSC_FALSE; 297 298 PetscFunctionBegin; 299 ierr = PetscOptionsGetTruth(PETSC_NULL,"-lu_old",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 300 if(newdatastruct){ 301 ierr = MatLUFactorSymbolic_SeqAIJ_inplace(B,A,isrow,iscol,info);CHKERRQ(ierr); 302 PetscFunctionReturn(0); 303 } 304 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 305 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 306 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 307 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 308 309 /* get new row and diagonal pointers, must be allocated separately because they will be given to the Mat_SeqAIJ and freed separately */ 310 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 311 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 312 bi[0] = bdiag[0] = 0; 313 314 /* linked list for storing column indices of the active row */ 315 nlnk = n + 1; 316 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 317 318 ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr); 319 320 /* initial FreeSpace size is f*(ai[n]+1) */ 321 f = info->fill; 322 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 323 current_space = free_space; 324 325 for (i=0; i<n; i++) { 326 /* copy previous fill into linked list */ 327 nzi = 0; 328 nnz = ai[r[i]+1] - ai[r[i]]; 329 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 330 ajtmp = aj + ai[r[i]]; 331 ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 332 nzi += nlnk; 333 334 /* add pivot rows into linked list */ 335 row = lnk[n]; 336 while (row < i){ 337 nzbd = bdiag[row] + 1; /* num of entries in the row with column index <= row */ 338 ajtmp = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */ 339 ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr); 340 nzi += nlnk; 341 row = lnk[row]; 342 } 343 bi[i+1] = bi[i] + nzi; 344 im[i] = nzi; 345 346 /* mark bdiag */ 347 nzbd = 0; 348 nnz = nzi; 349 k = lnk[n]; 350 while (nnz-- && k < i){ 351 nzbd++; 352 k = lnk[k]; 353 } 354 bdiag[i] = nzbd; /* note: bdiag[i] = nnzL as input for PetscFreeSpaceContiguous_LU() */ 355 356 /* if free space is not available, make more free space */ 357 if (current_space->local_remaining<nzi) { 358 nnz = 2*(n - i)*nzi; /* estimated and max additional space needed */ 359 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 360 reallocs++; 361 } 362 363 /* copy data into free space, then initialize lnk */ 364 ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 365 bi_ptr[i] = current_space->array; 366 current_space->array += nzi; 367 current_space->local_used += nzi; 368 current_space->local_remaining -= nzi; 369 } 370 #if defined(PETSC_USE_INFO) 371 if (ai[n] != 0) { 372 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 373 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 374 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 375 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr); 376 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 377 } else { 378 ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr); 379 } 380 #endif 381 382 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 383 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 384 385 /* destroy list of free space and other temporary array(s) */ 386 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 387 ierr = PetscFreeSpaceContiguous_LU_v2(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 388 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 389 ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr); 390 391 /* put together the new matrix */ 392 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 393 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 394 b = (Mat_SeqAIJ*)(B)->data; 395 b->free_a = PETSC_TRUE; 396 b->free_ij = PETSC_TRUE; 397 b->singlemalloc = PETSC_FALSE; 398 ierr = PetscMalloc((bdiag[0]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 399 b->j = bj; 400 b->i = bi; 401 b->diag = bdiag; 402 b->ilen = 0; 403 b->imax = 0; 404 b->row = isrow; 405 b->col = iscol; 406 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 407 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 408 b->icol = isicol; 409 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 410 411 /* In b structure: Free imax, ilen, old a, old j. Allocate solve_work, new a, new j */ 412 ierr = PetscLogObjectMemory(B,(bdiag[0]+1)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 413 b->maxnz = b->nz = bdiag[0]+1; 414 B->factor = MAT_FACTOR_LU; 415 B->info.factor_mallocs = reallocs; 416 B->info.fill_ratio_given = f; 417 418 if (ai[n]) { 419 B->info.fill_ratio_needed = ((PetscReal)(bdiag[0]+1))/((PetscReal)ai[n]); 420 } else { 421 B->info.fill_ratio_needed = 0.0; 422 } 423 B->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 424 /* switch to inodes if appropriate */ 425 ierr = Mat_CheckInode_FactorLU(B,PETSC_FALSE);CHKERRQ(ierr); 426 PetscFunctionReturn(0); 427 } 428 429 /* 430 Trouble in factorization, should we dump the original matrix? 431 */ 432 #undef __FUNCT__ 433 #define __FUNCT__ "MatFactorDumpMatrix" 434 PetscErrorCode MatFactorDumpMatrix(Mat A) 435 { 436 PetscErrorCode ierr; 437 PetscTruth flg = PETSC_FALSE; 438 439 PetscFunctionBegin; 440 ierr = PetscOptionsGetTruth(PETSC_NULL,"-mat_factor_dump_on_error",&flg,PETSC_NULL);CHKERRQ(ierr); 441 if (flg) { 442 PetscViewer viewer; 443 char filename[PETSC_MAX_PATH_LEN]; 444 445 ierr = PetscSNPrintf(filename,PETSC_MAX_PATH_LEN,"matrix_factor_error.%d",PetscGlobalRank);CHKERRQ(ierr); 446 ierr = PetscViewerBinaryOpen(((PetscObject)A)->comm,filename,FILE_MODE_WRITE,&viewer);CHKERRQ(ierr); 447 ierr = MatView(A,viewer);CHKERRQ(ierr); 448 ierr = PetscViewerDestroy(viewer);CHKERRQ(ierr); 449 } 450 PetscFunctionReturn(0); 451 } 452 453 #undef __FUNCT__ 454 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_newdatastruct" 455 PetscErrorCode MatLUFactorNumeric_SeqAIJ_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 456 { 457 Mat C=B; 458 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 459 IS isrow = b->row,isicol = b->icol; 460 PetscErrorCode ierr; 461 const PetscInt *r,*ic,*ics; 462 PetscInt i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 463 PetscInt *ajtmp,*bjtmp,nz,nzL,row,*bdiag=b->diag,*pj; 464 MatScalar *rtmp,*pc,multiplier,*v,*pv,*aa=a->a; 465 PetscTruth row_identity,col_identity; 466 467 LUShift_Ctx sctx; 468 PetscInt *ddiag,newshift; 469 PetscReal rs; 470 MatScalar d; 471 472 PetscFunctionBegin; 473 /* MatPivotSetUp(): initialize shift context sctx */ 474 ierr = PetscMemzero(&sctx,sizeof(LUShift_Ctx));CHKERRQ(ierr); 475 476 /* if both shift schemes are chosen by user, only use info->shiftpd */ 477 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 478 ddiag = a->diag; 479 sctx.shift_top = info->zeropivot; 480 for (i=0; i<n; i++) { 481 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 482 d = (aa)[ddiag[i]]; 483 rs = -PetscAbsScalar(d) - PetscRealPart(d); 484 v = aa+ai[i]; 485 nz = ai[i+1] - ai[i]; 486 for (j=0; j<nz; j++) 487 rs += PetscAbsScalar(v[j]); 488 if (rs>sctx.shift_top) sctx.shift_top = rs; 489 } 490 sctx.shift_top *= 1.1; 491 sctx.nshift_max = 5; 492 sctx.shift_lo = 0.; 493 sctx.shift_hi = 1.; 494 } 495 496 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 497 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 498 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 499 ics = ic; 500 501 do { 502 sctx.lushift = PETSC_FALSE; 503 for (i=0; i<n; i++){ 504 /* zero rtmp */ 505 /* L part */ 506 nz = bi[i+1] - bi[i]; 507 bjtmp = bj + bi[i]; 508 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 509 510 /* U part */ 511 nz = bdiag[i]-bdiag[i+1]; 512 bjtmp = bj + bdiag[i+1]+1; 513 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 514 515 /* load in initial (unfactored row) */ 516 nz = ai[r[i]+1] - ai[r[i]]; 517 ajtmp = aj + ai[r[i]]; 518 v = aa + ai[r[i]]; 519 for (j=0; j<nz; j++) { 520 rtmp[ics[ajtmp[j]]] = v[j]; 521 } 522 /* ZeropivotApply() */ 523 rtmp[i] += sctx.shift_amount; /* shift the diagonal of the matrix */ 524 525 /* elimination */ 526 bjtmp = bj + bi[i]; 527 row = *bjtmp++; 528 nzL = bi[i+1] - bi[i]; 529 for(k=0; k < nzL;k++) { 530 pc = rtmp + row; 531 if (*pc != 0.0) { 532 pv = b->a + bdiag[row]; 533 multiplier = *pc * (*pv); 534 *pc = multiplier; 535 pj = b->j + bdiag[row+1]+1; /* beginning of U(row,:) */ 536 pv = b->a + bdiag[row+1]+1; 537 nz = bdiag[row]-bdiag[row+1]-1; /* num of entries in U(row,:) excluding diag */ 538 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 539 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 540 } 541 row = *bjtmp++; 542 } 543 544 /* finished row so stick it into b->a */ 545 rs = 0.0; 546 /* L part */ 547 pv = b->a + bi[i] ; 548 pj = b->j + bi[i] ; 549 nz = bi[i+1] - bi[i]; 550 for (j=0; j<nz; j++) { 551 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 552 } 553 554 /* U part */ 555 pv = b->a + bdiag[i+1]+1; 556 pj = b->j + bdiag[i+1]+1; 557 nz = bdiag[i] - bdiag[i+1]-1; 558 for (j=0; j<nz; j++) { 559 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 560 } 561 562 /* MatPivotCheck() */ 563 sctx.rs = rs; 564 sctx.pv = rtmp[i]; 565 if (info->shiftnz){ 566 ierr = MatPivotCheck_nz(info,sctx,i,newshift);CHKERRQ(ierr); 567 } else if (info->shiftpd){ 568 ierr = MatPivotCheck_pd(info,sctx,i,newshift);CHKERRQ(ierr); 569 } else if (info->shiftinblocks){ 570 ierr = MatPivotCheck_inblocks(info,sctx,i,newshift);CHKERRQ(ierr); 571 } else { 572 ierr = MatPivotCheck_none(info,sctx,i,newshift);CHKERRQ(ierr); 573 } 574 rtmp[i] = sctx.pv; 575 if (newshift == 1) break; 576 577 /* Mark diagonal and invert diagonal for simplier triangular solves */ 578 pv = b->a + bdiag[i]; 579 *pv = 1.0/rtmp[i]; 580 581 } /* endof for (i=0; i<n; i++){ */ 582 583 /* MatPivotRefine() */ 584 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max){ 585 /* 586 * if no shift in this attempt & shifting & started shifting & can refine, 587 * then try lower shift 588 */ 589 sctx.shift_hi = sctx.shift_fraction; 590 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 591 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 592 sctx.lushift = PETSC_TRUE; 593 sctx.nshift++; 594 } 595 } while (sctx.lushift); 596 597 ierr = PetscFree(rtmp);CHKERRQ(ierr); 598 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 599 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 600 if (b->inode.use){ 601 C->ops->solve = MatSolve_SeqAIJ_Inode_newdatastruct; 602 } else { 603 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 604 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 605 if (row_identity && col_identity) { 606 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 607 } else { 608 C->ops->solve = MatSolve_SeqAIJ_newdatastruct; 609 } 610 } 611 C->ops->solveadd = MatSolveAdd_SeqAIJ_newdatastruct; 612 C->ops->solvetranspose = MatSolveTranspose_SeqAIJ_newdatastruct; 613 C->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ_newdatastruct; 614 C->ops->matsolve = MatMatSolve_SeqAIJ_newdatastruct; 615 C->assembled = PETSC_TRUE; 616 C->preallocated = PETSC_TRUE; 617 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 618 619 /* MatPivotView() */ 620 if (sctx.nshift){ 621 if (info->shiftpd) { 622 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); 623 } else if (info->shiftnz) { 624 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 625 } else if (info->shiftinblocks){ 626 ierr = PetscInfo2(A,"number of shift_inblocks applied %D, each shift_amount %G\n",sctx.nshift,info->shiftinblocks);CHKERRQ(ierr); 627 } 628 } 629 PetscFunctionReturn(0); 630 } 631 632 #undef __FUNCT__ 633 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_inplace" 634 PetscErrorCode MatLUFactorNumeric_SeqAIJ_inplace(Mat B,Mat A,const MatFactorInfo *info) 635 { 636 Mat C=B; 637 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 638 IS isrow = b->row,isicol = b->icol; 639 PetscErrorCode ierr; 640 const PetscInt *r,*ic,*ics; 641 PetscInt nz,row,i,j,n=A->rmap->n,diag; 642 const PetscInt *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 643 const PetscInt *ajtmp,*bjtmp,*diag_offset = b->diag,*pj; 644 MatScalar *pv,*rtmp,*pc,multiplier,d; 645 const MatScalar *v,*aa=a->a; 646 PetscReal rs=0.0; 647 LUShift_Ctx sctx; 648 PetscInt newshift,*ddiag; 649 650 PetscFunctionBegin; 651 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 652 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 653 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 654 ics = ic; 655 656 /* initialize shift context sctx */ 657 sctx.nshift = 0; 658 sctx.nshift_max = 0; 659 sctx.shift_top = 0.0; 660 sctx.shift_lo = 0.0; 661 sctx.shift_hi = 0.0; 662 sctx.shift_fraction = 0.0; 663 sctx.shift_amount = 0.0; 664 665 /* if both shift schemes are chosen by user, only use info->shiftpd */ 666 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 667 ddiag = a->diag; 668 sctx.shift_top = info->zeropivot; 669 for (i=0; i<n; i++) { 670 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 671 d = (aa)[ddiag[i]]; 672 rs = -PetscAbsScalar(d) - PetscRealPart(d); 673 v = aa+ai[i]; 674 nz = ai[i+1] - ai[i]; 675 for (j=0; j<nz; j++) 676 rs += PetscAbsScalar(v[j]); 677 if (rs>sctx.shift_top) sctx.shift_top = rs; 678 } 679 sctx.shift_top *= 1.1; 680 sctx.nshift_max = 5; 681 sctx.shift_lo = 0.; 682 sctx.shift_hi = 1.; 683 } 684 685 do { 686 sctx.lushift = PETSC_FALSE; 687 for (i=0; i<n; i++){ 688 nz = bi[i+1] - bi[i]; 689 bjtmp = bj + bi[i]; 690 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 691 692 /* load in initial (unfactored row) */ 693 nz = ai[r[i]+1] - ai[r[i]]; 694 ajtmp = aj + ai[r[i]]; 695 v = aa + ai[r[i]]; 696 for (j=0; j<nz; j++) { 697 rtmp[ics[ajtmp[j]]] = v[j]; 698 } 699 rtmp[ics[r[i]]] += sctx.shift_amount; /* shift the diagonal of the matrix */ 700 /* if (sctx.shift_amount > 0.0) printf("row %d, shift %g\n",i,sctx.shift_amount); */ 701 702 row = *bjtmp++; 703 while (row < i) { 704 pc = rtmp + row; 705 if (*pc != 0.0) { 706 pv = b->a + diag_offset[row]; 707 pj = b->j + diag_offset[row] + 1; 708 multiplier = *pc / *pv++; 709 *pc = multiplier; 710 nz = bi[row+1] - diag_offset[row] - 1; 711 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 712 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 713 } 714 row = *bjtmp++; 715 } 716 /* finished row so stick it into b->a */ 717 pv = b->a + bi[i] ; 718 pj = b->j + bi[i] ; 719 nz = bi[i+1] - bi[i]; 720 diag = diag_offset[i] - bi[i]; 721 rs = 0.0; 722 for (j=0; j<nz; j++) { 723 pv[j] = rtmp[pj[j]]; 724 rs += PetscAbsScalar(pv[j]); 725 } 726 rs -= PetscAbsScalar(pv[diag]); 727 728 /* 9/13/02 Victor Eijkhout suggested scaling zeropivot by rs for matrices with funny scalings */ 729 sctx.rs = rs; 730 sctx.pv = pv[diag]; 731 ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); 732 if (newshift == 1) break; 733 } 734 735 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max) { 736 /* 737 * if no shift in this attempt & shifting & started shifting & can refine, 738 * then try lower shift 739 */ 740 sctx.shift_hi = sctx.shift_fraction; 741 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 742 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 743 sctx.lushift = PETSC_TRUE; 744 sctx.nshift++; 745 } 746 } while (sctx.lushift); 747 748 /* invert diagonal entries for simplier triangular solves */ 749 for (i=0; i<n; i++) { 750 b->a[diag_offset[i]] = 1.0/b->a[diag_offset[i]]; 751 } 752 ierr = PetscFree(rtmp);CHKERRQ(ierr); 753 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 754 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 755 if (b->inode.use) { 756 C->ops->solve = MatSolve_SeqAIJ_Inode_inplace; 757 } else { 758 PetscTruth row_identity, col_identity; 759 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 760 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 761 if (row_identity && col_identity) { 762 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_inplace; 763 } else { 764 C->ops->solve = MatSolve_SeqAIJ_inplace; 765 } 766 } 767 C->ops->solveadd = MatSolveAdd_SeqAIJ_inplace; 768 C->ops->solvetranspose = MatSolveTranspose_SeqAIJ_inplace; 769 C->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ_inplace; 770 C->ops->matsolve = MatMatSolve_SeqAIJ_inplace; 771 C->assembled = PETSC_TRUE; 772 C->preallocated = PETSC_TRUE; 773 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 774 if (sctx.nshift){ 775 if (info->shiftpd) { 776 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); 777 } else if (info->shiftnz) { 778 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 779 } 780 } 781 PetscFunctionReturn(0); 782 } 783 784 /* 785 This routine implements inplace ILU(0) with row or/and column permutations. 786 Input: 787 A - original matrix 788 Output; 789 A - a->i (rowptr) is same as original rowptr, but factored i-the row is stored in rowperm[i] 790 a->j (col index) is permuted by the inverse of colperm, then sorted 791 a->a reordered accordingly with a->j 792 a->diag (ptr to diagonal elements) is updated. 793 */ 794 #undef __FUNCT__ 795 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_InplaceWithPerm" 796 PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat B,Mat A,const MatFactorInfo *info) 797 { 798 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 799 IS isrow = a->row,isicol = a->icol; 800 PetscErrorCode ierr; 801 const PetscInt *r,*ic,*ics; 802 PetscInt i,j,n=A->rmap->n,*ai=a->i,*aj=a->j; 803 PetscInt *ajtmp,nz,row; 804 PetscInt *diag = a->diag,nbdiag,*pj; 805 PetscScalar *rtmp,*pc,multiplier,d; 806 MatScalar *v,*pv; 807 PetscReal rs; 808 LUShift_Ctx sctx; 809 PetscInt newshift; 810 811 PetscFunctionBegin; 812 if (A != B) SETERRQ(PETSC_ERR_ARG_INCOMP,"input and output matrix must have same address"); 813 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 814 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 815 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&rtmp);CHKERRQ(ierr); 816 ierr = PetscMemzero(rtmp,(n+1)*sizeof(PetscScalar));CHKERRQ(ierr); 817 ics = ic; 818 819 sctx.shift_top = 0.; 820 sctx.nshift_max = 0; 821 sctx.shift_lo = 0.; 822 sctx.shift_hi = 0.; 823 sctx.shift_fraction = 0.; 824 825 /* if both shift schemes are chosen by user, only use info->shiftpd */ 826 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 827 sctx.shift_top = 0.; 828 for (i=0; i<n; i++) { 829 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 830 d = (a->a)[diag[i]]; 831 rs = -PetscAbsScalar(d) - PetscRealPart(d); 832 v = a->a+ai[i]; 833 nz = ai[i+1] - ai[i]; 834 for (j=0; j<nz; j++) 835 rs += PetscAbsScalar(v[j]); 836 if (rs>sctx.shift_top) sctx.shift_top = rs; 837 } 838 if (sctx.shift_top < info->zeropivot) sctx.shift_top = info->zeropivot; 839 sctx.shift_top *= 1.1; 840 sctx.nshift_max = 5; 841 sctx.shift_lo = 0.; 842 sctx.shift_hi = 1.; 843 } 844 845 sctx.shift_amount = 0.; 846 sctx.nshift = 0; 847 do { 848 sctx.lushift = PETSC_FALSE; 849 for (i=0; i<n; i++){ 850 /* load in initial unfactored row */ 851 nz = ai[r[i]+1] - ai[r[i]]; 852 ajtmp = aj + ai[r[i]]; 853 v = a->a + ai[r[i]]; 854 /* sort permuted ajtmp and values v accordingly */ 855 for (j=0; j<nz; j++) ajtmp[j] = ics[ajtmp[j]]; 856 ierr = PetscSortIntWithScalarArray(nz,ajtmp,v);CHKERRQ(ierr); 857 858 diag[r[i]] = ai[r[i]]; 859 for (j=0; j<nz; j++) { 860 rtmp[ajtmp[j]] = v[j]; 861 if (ajtmp[j] < i) diag[r[i]]++; /* update a->diag */ 862 } 863 rtmp[r[i]] += sctx.shift_amount; /* shift the diagonal of the matrix */ 864 865 row = *ajtmp++; 866 while (row < i) { 867 pc = rtmp + row; 868 if (*pc != 0.0) { 869 pv = a->a + diag[r[row]]; 870 pj = aj + diag[r[row]] + 1; 871 872 multiplier = *pc / *pv++; 873 *pc = multiplier; 874 nz = ai[r[row]+1] - diag[r[row]] - 1; 875 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 876 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 877 } 878 row = *ajtmp++; 879 } 880 /* finished row so overwrite it onto a->a */ 881 pv = a->a + ai[r[i]] ; 882 pj = aj + ai[r[i]] ; 883 nz = ai[r[i]+1] - ai[r[i]]; 884 nbdiag = diag[r[i]] - ai[r[i]]; /* num of entries before the diagonal */ 885 886 rs = 0.0; 887 for (j=0; j<nz; j++) { 888 pv[j] = rtmp[pj[j]]; 889 if (j != nbdiag) rs += PetscAbsScalar(pv[j]); 890 } 891 892 /* 9/13/02 Victor Eijkhout suggested scaling zeropivot by rs for matrices with funny scalings */ 893 sctx.rs = rs; 894 sctx.pv = pv[nbdiag]; 895 ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); 896 if (newshift == 1) break; 897 } 898 899 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max) { 900 /* 901 * if no shift in this attempt & shifting & started shifting & can refine, 902 * then try lower shift 903 */ 904 sctx.shift_hi = sctx.shift_fraction; 905 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 906 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 907 sctx.lushift = PETSC_TRUE; 908 sctx.nshift++; 909 } 910 } while (sctx.lushift); 911 912 /* invert diagonal entries for simplier triangular solves */ 913 for (i=0; i<n; i++) { 914 a->a[diag[r[i]]] = 1.0/a->a[diag[r[i]]]; 915 } 916 917 ierr = PetscFree(rtmp);CHKERRQ(ierr); 918 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 919 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 920 A->ops->solve = MatSolve_SeqAIJ_InplaceWithPerm; 921 A->ops->solveadd = MatSolveAdd_SeqAIJ_inplace; 922 A->ops->solvetranspose = MatSolveTranspose_SeqAIJ_inplace; 923 A->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ_inplace; 924 A->assembled = PETSC_TRUE; 925 A->preallocated = PETSC_TRUE; 926 ierr = PetscLogFlops(A->cmap->n);CHKERRQ(ierr); 927 if (sctx.nshift){ 928 if (info->shiftpd) { 929 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); 930 } else if (info->shiftnz) { 931 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 932 } 933 } 934 PetscFunctionReturn(0); 935 } 936 937 /* ----------------------------------------------------------- */ 938 #undef __FUNCT__ 939 #define __FUNCT__ "MatLUFactor_SeqAIJ" 940 PetscErrorCode MatLUFactor_SeqAIJ(Mat A,IS row,IS col,const MatFactorInfo *info) 941 { 942 PetscErrorCode ierr; 943 Mat C; 944 945 PetscFunctionBegin; 946 ierr = MatGetFactor(A,MAT_SOLVER_PETSC,MAT_FACTOR_LU,&C);CHKERRQ(ierr); 947 ierr = MatLUFactorSymbolic(C,A,row,col,info);CHKERRQ(ierr); 948 ierr = MatLUFactorNumeric(C,A,info);CHKERRQ(ierr); 949 A->ops->solve = C->ops->solve; 950 A->ops->solvetranspose = C->ops->solvetranspose; 951 ierr = MatHeaderCopy(A,C);CHKERRQ(ierr); 952 ierr = PetscLogObjectParent(A,((Mat_SeqAIJ*)(A->data))->icol);CHKERRQ(ierr); 953 PetscFunctionReturn(0); 954 } 955 /* ----------------------------------------------------------- */ 956 957 958 #undef __FUNCT__ 959 #define __FUNCT__ "MatSolve_SeqAIJ_inplace" 960 PetscErrorCode MatSolve_SeqAIJ_inplace(Mat A,Vec bb,Vec xx) 961 { 962 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 963 IS iscol = a->col,isrow = a->row; 964 PetscErrorCode ierr; 965 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 966 PetscInt nz; 967 const PetscInt *rout,*cout,*r,*c; 968 PetscScalar *x,*tmp,*tmps,sum; 969 const PetscScalar *b; 970 const MatScalar *aa = a->a,*v; 971 972 PetscFunctionBegin; 973 if (!n) PetscFunctionReturn(0); 974 975 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 976 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 977 tmp = a->solve_work; 978 979 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 980 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 981 982 /* forward solve the lower triangular */ 983 tmp[0] = b[*r++]; 984 tmps = tmp; 985 for (i=1; i<n; i++) { 986 v = aa + ai[i] ; 987 vi = aj + ai[i] ; 988 nz = a->diag[i] - ai[i]; 989 sum = b[*r++]; 990 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 991 tmp[i] = sum; 992 } 993 994 /* backward solve the upper triangular */ 995 for (i=n-1; i>=0; i--){ 996 v = aa + a->diag[i] + 1; 997 vi = aj + a->diag[i] + 1; 998 nz = ai[i+1] - a->diag[i] - 1; 999 sum = tmp[i]; 1000 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1001 x[*c--] = tmp[i] = sum*aa[a->diag[i]]; 1002 } 1003 1004 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1005 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1006 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1007 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1008 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1009 PetscFunctionReturn(0); 1010 } 1011 1012 #undef __FUNCT__ 1013 #define __FUNCT__ "MatMatSolve_SeqAIJ_inplace" 1014 PetscErrorCode MatMatSolve_SeqAIJ_inplace(Mat A,Mat B,Mat X) 1015 { 1016 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1017 IS iscol = a->col,isrow = a->row; 1018 PetscErrorCode ierr; 1019 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 1020 PetscInt nz,neq; 1021 const PetscInt *rout,*cout,*r,*c; 1022 PetscScalar *x,*b,*tmp,*tmps,sum; 1023 const MatScalar *aa = a->a,*v; 1024 PetscTruth bisdense,xisdense; 1025 1026 PetscFunctionBegin; 1027 if (!n) PetscFunctionReturn(0); 1028 1029 ierr = PetscTypeCompare((PetscObject)B,MATSEQDENSE,&bisdense);CHKERRQ(ierr); 1030 if (!bisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"B matrix must be a SeqDense matrix"); 1031 ierr = PetscTypeCompare((PetscObject)X,MATSEQDENSE,&xisdense);CHKERRQ(ierr); 1032 if (!xisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"X matrix must be a SeqDense matrix"); 1033 1034 ierr = MatGetArray(B,&b);CHKERRQ(ierr); 1035 ierr = MatGetArray(X,&x);CHKERRQ(ierr); 1036 1037 tmp = a->solve_work; 1038 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1039 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1040 1041 for (neq=0; neq<B->cmap->n; neq++){ 1042 /* forward solve the lower triangular */ 1043 tmp[0] = b[r[0]]; 1044 tmps = tmp; 1045 for (i=1; i<n; i++) { 1046 v = aa + ai[i] ; 1047 vi = aj + ai[i] ; 1048 nz = a->diag[i] - ai[i]; 1049 sum = b[r[i]]; 1050 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1051 tmp[i] = sum; 1052 } 1053 /* backward solve the upper triangular */ 1054 for (i=n-1; i>=0; i--){ 1055 v = aa + a->diag[i] + 1; 1056 vi = aj + a->diag[i] + 1; 1057 nz = ai[i+1] - a->diag[i] - 1; 1058 sum = tmp[i]; 1059 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1060 x[c[i]] = tmp[i] = sum*aa[a->diag[i]]; 1061 } 1062 1063 b += n; 1064 x += n; 1065 } 1066 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1067 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1068 ierr = MatRestoreArray(B,&b);CHKERRQ(ierr); 1069 ierr = MatRestoreArray(X,&x);CHKERRQ(ierr); 1070 ierr = PetscLogFlops(B->cmap->n*(2.0*a->nz - n));CHKERRQ(ierr); 1071 PetscFunctionReturn(0); 1072 } 1073 1074 #undef __FUNCT__ 1075 #define __FUNCT__ "MatMatSolve_SeqAIJ_newdatastruct" 1076 PetscErrorCode MatMatSolve_SeqAIJ_newdatastruct(Mat A,Mat B,Mat X) 1077 { 1078 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1079 IS iscol = a->col,isrow = a->row; 1080 PetscErrorCode ierr; 1081 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j,*adiag = a->diag; 1082 PetscInt nz,neq; 1083 const PetscInt *rout,*cout,*r,*c; 1084 PetscScalar *x,*b,*tmp,sum; 1085 const MatScalar *aa = a->a,*v; 1086 PetscTruth bisdense,xisdense; 1087 1088 PetscFunctionBegin; 1089 if (!n) PetscFunctionReturn(0); 1090 1091 ierr = PetscTypeCompare((PetscObject)B,MATSEQDENSE,&bisdense);CHKERRQ(ierr); 1092 if (!bisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"B matrix must be a SeqDense matrix"); 1093 ierr = PetscTypeCompare((PetscObject)X,MATSEQDENSE,&xisdense);CHKERRQ(ierr); 1094 if (!xisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"X matrix must be a SeqDense matrix"); 1095 1096 ierr = MatGetArray(B,&b);CHKERRQ(ierr); 1097 ierr = MatGetArray(X,&x);CHKERRQ(ierr); 1098 1099 tmp = a->solve_work; 1100 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1101 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1102 1103 for (neq=0; neq<B->cmap->n; neq++){ 1104 /* forward solve the lower triangular */ 1105 tmp[0] = b[r[0]]; 1106 v = aa; 1107 vi = aj; 1108 for (i=1; i<n; i++) { 1109 nz = ai[i+1] - ai[i]; 1110 sum = b[r[i]]; 1111 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 1112 tmp[i] = sum; 1113 v += nz; vi += nz; 1114 } 1115 1116 /* backward solve the upper triangular */ 1117 for (i=n-1; i>=0; i--){ 1118 v = aa + adiag[i+1]+1; 1119 vi = aj + adiag[i+1]+1; 1120 nz = adiag[i]-adiag[i+1]-1; 1121 sum = tmp[i]; 1122 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 1123 x[c[i]] = tmp[i] = sum*v[nz]; /* v[nz] = aa[adiag[i]] */ 1124 } 1125 1126 b += n; 1127 x += n; 1128 } 1129 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1130 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1131 ierr = MatRestoreArray(B,&b);CHKERRQ(ierr); 1132 ierr = MatRestoreArray(X,&x);CHKERRQ(ierr); 1133 ierr = PetscLogFlops(B->cmap->n*(2.0*a->nz - n));CHKERRQ(ierr); 1134 PetscFunctionReturn(0); 1135 } 1136 1137 #undef __FUNCT__ 1138 #define __FUNCT__ "MatSolve_SeqAIJ_InplaceWithPerm" 1139 PetscErrorCode MatSolve_SeqAIJ_InplaceWithPerm(Mat A,Vec bb,Vec xx) 1140 { 1141 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1142 IS iscol = a->col,isrow = a->row; 1143 PetscErrorCode ierr; 1144 const PetscInt *r,*c,*rout,*cout; 1145 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 1146 PetscInt nz,row; 1147 PetscScalar *x,*b,*tmp,*tmps,sum; 1148 const MatScalar *aa = a->a,*v; 1149 1150 PetscFunctionBegin; 1151 if (!n) PetscFunctionReturn(0); 1152 1153 ierr = VecGetArray(bb,&b);CHKERRQ(ierr); 1154 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1155 tmp = a->solve_work; 1156 1157 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1158 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1159 1160 /* forward solve the lower triangular */ 1161 tmp[0] = b[*r++]; 1162 tmps = tmp; 1163 for (row=1; row<n; row++) { 1164 i = rout[row]; /* permuted row */ 1165 v = aa + ai[i] ; 1166 vi = aj + ai[i] ; 1167 nz = a->diag[i] - ai[i]; 1168 sum = b[*r++]; 1169 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1170 tmp[row] = sum; 1171 } 1172 1173 /* backward solve the upper triangular */ 1174 for (row=n-1; row>=0; row--){ 1175 i = rout[row]; /* permuted row */ 1176 v = aa + a->diag[i] + 1; 1177 vi = aj + a->diag[i] + 1; 1178 nz = ai[i+1] - a->diag[i] - 1; 1179 sum = tmp[row]; 1180 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1181 x[*c--] = tmp[row] = sum*aa[a->diag[i]]; 1182 } 1183 1184 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1185 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1186 ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr); 1187 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1188 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1189 PetscFunctionReturn(0); 1190 } 1191 1192 /* ----------------------------------------------------------- */ 1193 #include "../src/mat/impls/aij/seq/ftn-kernels/fsolve.h" 1194 #undef __FUNCT__ 1195 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering_inplace" 1196 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_inplace(Mat A,Vec bb,Vec xx) 1197 { 1198 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1199 PetscErrorCode ierr; 1200 PetscInt n = A->rmap->n; 1201 const PetscInt *ai = a->i,*aj = a->j,*adiag = a->diag; 1202 PetscScalar *x; 1203 const PetscScalar *b; 1204 const MatScalar *aa = a->a; 1205 #if !defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 1206 PetscInt adiag_i,i,nz,ai_i; 1207 const PetscInt *vi; 1208 const MatScalar *v; 1209 PetscScalar sum; 1210 #endif 1211 1212 PetscFunctionBegin; 1213 if (!n) PetscFunctionReturn(0); 1214 1215 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1216 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1217 1218 #if defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 1219 fortransolveaij_(&n,x,ai,aj,adiag,aa,b); 1220 #else 1221 /* forward solve the lower triangular */ 1222 x[0] = b[0]; 1223 for (i=1; i<n; i++) { 1224 ai_i = ai[i]; 1225 v = aa + ai_i; 1226 vi = aj + ai_i; 1227 nz = adiag[i] - ai_i; 1228 sum = b[i]; 1229 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 1230 x[i] = sum; 1231 } 1232 1233 /* backward solve the upper triangular */ 1234 for (i=n-1; i>=0; i--){ 1235 adiag_i = adiag[i]; 1236 v = aa + adiag_i + 1; 1237 vi = aj + adiag_i + 1; 1238 nz = ai[i+1] - adiag_i - 1; 1239 sum = x[i]; 1240 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 1241 x[i] = sum*aa[adiag_i]; 1242 } 1243 #endif 1244 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1245 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1246 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1247 PetscFunctionReturn(0); 1248 } 1249 1250 #undef __FUNCT__ 1251 #define __FUNCT__ "MatSolveAdd_SeqAIJ_inplace" 1252 PetscErrorCode MatSolveAdd_SeqAIJ_inplace(Mat A,Vec bb,Vec yy,Vec xx) 1253 { 1254 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1255 IS iscol = a->col,isrow = a->row; 1256 PetscErrorCode ierr; 1257 PetscInt i, n = A->rmap->n,j; 1258 PetscInt nz; 1259 const PetscInt *rout,*cout,*r,*c,*vi,*ai = a->i,*aj = a->j; 1260 PetscScalar *x,*tmp,sum; 1261 const PetscScalar *b; 1262 const MatScalar *aa = a->a,*v; 1263 1264 PetscFunctionBegin; 1265 if (yy != xx) {ierr = VecCopy(yy,xx);CHKERRQ(ierr);} 1266 1267 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1268 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1269 tmp = a->solve_work; 1270 1271 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1272 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1273 1274 /* forward solve the lower triangular */ 1275 tmp[0] = b[*r++]; 1276 for (i=1; i<n; i++) { 1277 v = aa + ai[i] ; 1278 vi = aj + ai[i] ; 1279 nz = a->diag[i] - ai[i]; 1280 sum = b[*r++]; 1281 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1282 tmp[i] = sum; 1283 } 1284 1285 /* backward solve the upper triangular */ 1286 for (i=n-1; i>=0; i--){ 1287 v = aa + a->diag[i] + 1; 1288 vi = aj + a->diag[i] + 1; 1289 nz = ai[i+1] - a->diag[i] - 1; 1290 sum = tmp[i]; 1291 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1292 tmp[i] = sum*aa[a->diag[i]]; 1293 x[*c--] += tmp[i]; 1294 } 1295 1296 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1297 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1298 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1299 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1300 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1301 1302 PetscFunctionReturn(0); 1303 } 1304 1305 #undef __FUNCT__ 1306 #define __FUNCT__ "MatSolveAdd_SeqAIJ_newdatastruct" 1307 PetscErrorCode MatSolveAdd_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec yy,Vec xx) 1308 { 1309 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1310 IS iscol = a->col,isrow = a->row; 1311 PetscErrorCode ierr; 1312 PetscInt i, n = A->rmap->n,j; 1313 PetscInt nz; 1314 const PetscInt *rout,*cout,*r,*c,*vi,*ai = a->i,*aj = a->j,*adiag = a->diag; 1315 PetscScalar *x,*tmp,sum; 1316 const PetscScalar *b; 1317 const MatScalar *aa = a->a,*v; 1318 1319 PetscFunctionBegin; 1320 if (yy != xx) {ierr = VecCopy(yy,xx);CHKERRQ(ierr);} 1321 1322 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1323 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1324 tmp = a->solve_work; 1325 1326 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1327 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1328 1329 /* forward solve the lower triangular */ 1330 tmp[0] = b[r[0]]; 1331 v = aa; 1332 vi = aj; 1333 for (i=1; i<n; i++) { 1334 nz = ai[i+1] - ai[i]; 1335 sum = b[r[i]]; 1336 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1337 tmp[i] = sum; 1338 v += nz; vi += nz; 1339 } 1340 1341 /* backward solve the upper triangular */ 1342 v = aa + adiag[n-1]; 1343 vi = aj + adiag[n-1]; 1344 for (i=n-1; i>=0; i--){ 1345 nz = adiag[i] - adiag[i+1] - 1; 1346 sum = tmp[i]; 1347 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1348 tmp[i] = sum*v[nz]; 1349 x[c[i]] += tmp[i]; 1350 v += nz+1; vi += nz+1; 1351 } 1352 1353 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1354 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1355 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1356 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1357 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1358 1359 PetscFunctionReturn(0); 1360 } 1361 1362 #undef __FUNCT__ 1363 #define __FUNCT__ "MatSolveTranspose_SeqAIJ_inplace" 1364 PetscErrorCode MatSolveTranspose_SeqAIJ_inplace(Mat A,Vec bb,Vec xx) 1365 { 1366 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1367 IS iscol = a->col,isrow = a->row; 1368 PetscErrorCode ierr; 1369 const PetscInt *rout,*cout,*r,*c,*diag = a->diag,*ai = a->i,*aj = a->j,*vi; 1370 PetscInt i,n = A->rmap->n,j; 1371 PetscInt nz; 1372 PetscScalar *x,*tmp,s1; 1373 const MatScalar *aa = a->a,*v; 1374 const PetscScalar *b; 1375 1376 PetscFunctionBegin; 1377 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1378 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1379 tmp = a->solve_work; 1380 1381 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1382 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1383 1384 /* copy the b into temp work space according to permutation */ 1385 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1386 1387 /* forward solve the U^T */ 1388 for (i=0; i<n; i++) { 1389 v = aa + diag[i] ; 1390 vi = aj + diag[i] + 1; 1391 nz = ai[i+1] - diag[i] - 1; 1392 s1 = tmp[i]; 1393 s1 *= (*v++); /* multiply by inverse of diagonal entry */ 1394 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1395 tmp[i] = s1; 1396 } 1397 1398 /* backward solve the L^T */ 1399 for (i=n-1; i>=0; i--){ 1400 v = aa + diag[i] - 1 ; 1401 vi = aj + diag[i] - 1 ; 1402 nz = diag[i] - ai[i]; 1403 s1 = tmp[i]; 1404 for (j=0; j>-nz; j--) tmp[vi[j]] -= s1*v[j]; 1405 } 1406 1407 /* copy tmp into x according to permutation */ 1408 for (i=0; i<n; i++) x[r[i]] = tmp[i]; 1409 1410 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1411 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1412 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1413 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1414 1415 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1416 PetscFunctionReturn(0); 1417 } 1418 1419 #undef __FUNCT__ 1420 #define __FUNCT__ "MatSolveTranspose_SeqAIJ_newdatastruct" 1421 PetscErrorCode MatSolveTranspose_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec xx) 1422 { 1423 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1424 IS iscol = a->col,isrow = a->row; 1425 PetscErrorCode ierr; 1426 const PetscInt *rout,*cout,*r,*c,*adiag = a->diag,*ai = a->i,*aj = a->j,*vi; 1427 PetscInt i,n = A->rmap->n,j; 1428 PetscInt nz; 1429 PetscScalar *x,*tmp,s1; 1430 const MatScalar *aa = a->a,*v; 1431 const PetscScalar *b; 1432 1433 PetscFunctionBegin; 1434 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1435 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1436 tmp = a->solve_work; 1437 1438 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1439 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1440 1441 /* copy the b into temp work space according to permutation */ 1442 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1443 1444 /* forward solve the U^T */ 1445 for (i=0; i<n; i++) { 1446 v = aa + adiag[i+1] + 1; 1447 vi = aj + adiag[i+1] + 1; 1448 nz = adiag[i] - adiag[i+1] - 1; 1449 s1 = tmp[i]; 1450 s1 *= v[nz]; /* multiply by inverse of diagonal entry */ 1451 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1452 tmp[i] = s1; 1453 } 1454 1455 /* backward solve the L^T */ 1456 for (i=n-1; i>=0; i--){ 1457 v = aa + ai[i]; 1458 vi = aj + ai[i]; 1459 nz = ai[i+1] - ai[i]; 1460 s1 = tmp[i]; 1461 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1462 } 1463 1464 /* copy tmp into x according to permutation */ 1465 for (i=0; i<n; i++) x[r[i]] = tmp[i]; 1466 1467 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1468 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1469 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1470 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1471 1472 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1473 PetscFunctionReturn(0); 1474 } 1475 1476 #undef __FUNCT__ 1477 #define __FUNCT__ "MatSolveTransposeAdd_SeqAIJ_inplace" 1478 PetscErrorCode MatSolveTransposeAdd_SeqAIJ_inplace(Mat A,Vec bb,Vec zz,Vec xx) 1479 { 1480 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1481 IS iscol = a->col,isrow = a->row; 1482 PetscErrorCode ierr; 1483 const PetscInt *rout,*cout,*r,*c,*diag = a->diag,*ai = a->i,*aj = a->j,*vi; 1484 PetscInt i,n = A->rmap->n,j; 1485 PetscInt nz; 1486 PetscScalar *x,*tmp,s1; 1487 const MatScalar *aa = a->a,*v; 1488 const PetscScalar *b; 1489 1490 PetscFunctionBegin; 1491 if (zz != xx) {ierr = VecCopy(zz,xx);CHKERRQ(ierr);} 1492 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1493 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1494 tmp = a->solve_work; 1495 1496 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1497 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1498 1499 /* copy the b into temp work space according to permutation */ 1500 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1501 1502 /* forward solve the U^T */ 1503 for (i=0; i<n; i++) { 1504 v = aa + diag[i] ; 1505 vi = aj + diag[i] + 1; 1506 nz = ai[i+1] - diag[i] - 1; 1507 s1 = tmp[i]; 1508 s1 *= (*v++); /* multiply by inverse of diagonal entry */ 1509 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1510 tmp[i] = s1; 1511 } 1512 1513 /* backward solve the L^T */ 1514 for (i=n-1; i>=0; i--){ 1515 v = aa + diag[i] - 1 ; 1516 vi = aj + diag[i] - 1 ; 1517 nz = diag[i] - ai[i]; 1518 s1 = tmp[i]; 1519 for (j=0; j>-nz; j--) tmp[vi[j]] -= s1*v[j]; 1520 } 1521 1522 /* copy tmp into x according to permutation */ 1523 for (i=0; i<n; i++) x[r[i]] += tmp[i]; 1524 1525 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1526 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1527 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1528 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1529 1530 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1531 PetscFunctionReturn(0); 1532 } 1533 1534 #undef __FUNCT__ 1535 #define __FUNCT__ "MatSolveTransposeAdd_SeqAIJ_newdatastruct" 1536 PetscErrorCode MatSolveTransposeAdd_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec zz,Vec xx) 1537 { 1538 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1539 IS iscol = a->col,isrow = a->row; 1540 PetscErrorCode ierr; 1541 const PetscInt *rout,*cout,*r,*c,*adiag = a->diag,*ai = a->i,*aj = a->j,*vi; 1542 PetscInt i,n = A->rmap->n,j; 1543 PetscInt nz; 1544 PetscScalar *x,*tmp,s1; 1545 const MatScalar *aa = a->a,*v; 1546 const PetscScalar *b; 1547 1548 PetscFunctionBegin; 1549 if (zz != xx) {ierr = VecCopy(zz,xx);CHKERRQ(ierr);} 1550 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1551 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1552 tmp = a->solve_work; 1553 1554 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1555 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1556 1557 /* copy the b into temp work space according to permutation */ 1558 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1559 1560 /* forward solve the U^T */ 1561 for (i=0; i<n; i++) { 1562 v = aa + adiag[i+1] + 1; 1563 vi = aj + adiag[i+1] + 1; 1564 nz = adiag[i] - adiag[i+1] - 1; 1565 s1 = tmp[i]; 1566 s1 *= v[nz]; /* multiply by inverse of diagonal entry */ 1567 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1568 tmp[i] = s1; 1569 } 1570 1571 1572 /* backward solve the L^T */ 1573 for (i=n-1; i>=0; i--){ 1574 v = aa + ai[i] ; 1575 vi = aj + ai[i]; 1576 nz = ai[i+1] - ai[i]; 1577 s1 = tmp[i]; 1578 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1579 } 1580 1581 /* copy tmp into x according to permutation */ 1582 for (i=0; i<n; i++) x[r[i]] += tmp[i]; 1583 1584 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1585 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1586 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1587 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1588 1589 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1590 PetscFunctionReturn(0); 1591 } 1592 1593 /* ----------------------------------------------------------------*/ 1594 1595 EXTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat,Mat,MatDuplicateOption,PetscTruth); 1596 1597 /* 1598 ilu() under revised new data structure. 1599 Factored arrays bj and ba are stored as 1600 L(0,:), L(1,:), ...,L(n-1,:), U(n-1,:),...,U(i,:),U(i-1,:),...,U(0,:) 1601 1602 bi=fact->i is an array of size n+1, in which 1603 bi+ 1604 bi[i]: points to 1st entry of L(i,:),i=0,...,n-1 1605 bi[n]: points to L(n-1,n-1)+1 1606 1607 bdiag=fact->diag is an array of size n+1,in which 1608 bdiag[i]: points to diagonal of U(i,:), i=0,...,n-1 1609 bdiag[n]: points to entry of U(n-1,0)-1 1610 1611 U(i,:) contains bdiag[i] as its last entry, i.e., 1612 U(i,:) = (u[i,i+1],...,u[i,n-1],diag[i]) 1613 */ 1614 #undef __FUNCT__ 1615 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct" 1616 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1617 { 1618 1619 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1620 PetscErrorCode ierr; 1621 PetscInt n=A->rmap->n,*ai=a->i,*aj,*adiag=a->diag; 1622 PetscInt i,j,nz,*bi,*bj,*bdiag; 1623 PetscTruth missing; 1624 IS isicol; 1625 1626 PetscFunctionBegin; 1627 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); 1628 ierr = MatMissingDiagonal(A,&missing,&i);CHKERRQ(ierr); 1629 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",i); 1630 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1631 1632 ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES,PETSC_FALSE);CHKERRQ(ierr); 1633 b = (Mat_SeqAIJ*)(fact)->data; 1634 1635 /* allocate matrix arrays for new data structure */ 1636 ierr = PetscMalloc3(ai[n]+1,PetscScalar,&b->a,ai[n]+1,PetscInt,&b->j,n+1,PetscInt,&b->i);CHKERRQ(ierr); 1637 ierr = PetscLogObjectMemory(fact,ai[n]*(sizeof(PetscScalar)+sizeof(PetscInt))+(n+1)*sizeof(PetscInt));CHKERRQ(ierr); 1638 b->singlemalloc = PETSC_TRUE; 1639 if (!b->diag){ 1640 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&b->diag);CHKERRQ(ierr); 1641 ierr = PetscLogObjectMemory(fact,(n+1)*sizeof(PetscInt));CHKERRQ(ierr); 1642 } 1643 bdiag = b->diag; 1644 1645 if (n > 0) { 1646 ierr = PetscMemzero(b->a,(ai[n])*sizeof(MatScalar));CHKERRQ(ierr); 1647 } 1648 1649 /* set bi and bj with new data structure */ 1650 bi = b->i; 1651 bj = b->j; 1652 1653 /* L part */ 1654 bi[0] = 0; 1655 for (i=0; i<n; i++){ 1656 nz = adiag[i] - ai[i]; 1657 bi[i+1] = bi[i] + nz; 1658 aj = a->j + ai[i]; 1659 for (j=0; j<nz; j++){ 1660 *bj = aj[j]; bj++; 1661 } 1662 } 1663 1664 /* U part */ 1665 bdiag[n] = bi[n]-1; 1666 for (i=n-1; i>=0; i--){ 1667 nz = ai[i+1] - adiag[i] - 1; 1668 aj = a->j + adiag[i] + 1; 1669 for (j=0; j<nz; j++){ 1670 *bj = aj[j]; bj++; 1671 } 1672 /* diag[i] */ 1673 *bj = i; bj++; 1674 bdiag[i] = bdiag[i+1] + nz + 1; 1675 } 1676 1677 fact->factor = MAT_FACTOR_ILU; 1678 fact->info.factor_mallocs = 0; 1679 fact->info.fill_ratio_given = info->fill; 1680 fact->info.fill_ratio_needed = 1.0; 1681 fact->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 1682 1683 b = (Mat_SeqAIJ*)(fact)->data; 1684 b->row = isrow; 1685 b->col = iscol; 1686 b->icol = isicol; 1687 ierr = PetscMalloc((fact->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1688 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1689 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1690 PetscFunctionReturn(0); 1691 } 1692 1693 #undef __FUNCT__ 1694 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_newdatastruct" 1695 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1696 { 1697 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1698 IS isicol; 1699 PetscErrorCode ierr; 1700 const PetscInt *r,*ic; 1701 PetscInt n=A->rmap->n,*ai=a->i,*aj=a->j; 1702 PetscInt *bi,*cols,nnz,*cols_lvl; 1703 PetscInt *bdiag,prow,fm,nzbd,reallocs=0,dcount=0; 1704 PetscInt i,levels,diagonal_fill; 1705 PetscTruth col_identity,row_identity; 1706 PetscReal f; 1707 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL; 1708 PetscBT lnkbt; 1709 PetscInt nzi,*bj,**bj_ptr,**bjlvl_ptr; 1710 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1711 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 1712 PetscTruth newdatastruct=PETSC_FALSE; 1713 1714 PetscFunctionBegin; 1715 ierr = PetscOptionsGetTruth(PETSC_NULL,"-ilu_old",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 1716 if(newdatastruct){ 1717 ierr = MatILUFactorSymbolic_SeqAIJ_inplace(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1718 PetscFunctionReturn(0); 1719 } 1720 1721 levels = (PetscInt)info->levels; 1722 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 1723 ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 1724 1725 if (!levels && row_identity && col_identity) { 1726 /* special case: ilu(0) with natural ordering */ 1727 ierr = MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1728 ierr = Mat_CheckInode_FactorLU(fact,PETSC_FALSE);CHKERRQ(ierr); 1729 PetscFunctionReturn(0); 1730 } 1731 1732 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); 1733 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1734 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1735 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 1736 1737 /* get new row and diagonal pointers, must be allocated separately because they will be given to the Mat_SeqAIJ and freed separately */ 1738 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1739 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 1740 bi[0] = bdiag[0] = 0; 1741 1742 ierr = PetscMalloc2(n,PetscInt*,&bj_ptr,n,PetscInt*,&bjlvl_ptr);CHKERRQ(ierr); 1743 1744 /* create a linked list for storing column indices of the active row */ 1745 nlnk = n + 1; 1746 ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1747 1748 /* initial FreeSpace size is f*(ai[n]+1) */ 1749 f = info->fill; 1750 diagonal_fill = (PetscInt)info->diagonal_fill; 1751 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 1752 current_space = free_space; 1753 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr); 1754 current_space_lvl = free_space_lvl; 1755 1756 for (i=0; i<n; i++) { 1757 nzi = 0; 1758 /* copy current row into linked list */ 1759 nnz = ai[r[i]+1] - ai[r[i]]; 1760 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 1761 cols = aj + ai[r[i]]; 1762 lnk[i] = -1; /* marker to indicate if diagonal exists */ 1763 ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1764 nzi += nlnk; 1765 1766 /* make sure diagonal entry is included */ 1767 if (diagonal_fill && lnk[i] == -1) { 1768 fm = n; 1769 while (lnk[fm] < i) fm = lnk[fm]; 1770 lnk[i] = lnk[fm]; /* insert diagonal into linked list */ 1771 lnk[fm] = i; 1772 lnk_lvl[i] = 0; 1773 nzi++; dcount++; 1774 } 1775 1776 /* add pivot rows into the active row */ 1777 nzbd = 0; 1778 prow = lnk[n]; 1779 while (prow < i) { 1780 nnz = bdiag[prow]; 1781 cols = bj_ptr[prow] + nnz + 1; 1782 cols_lvl = bjlvl_ptr[prow] + nnz + 1; 1783 nnz = bi[prow+1] - bi[prow] - nnz - 1; 1784 ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr); 1785 nzi += nlnk; 1786 prow = lnk[prow]; 1787 nzbd++; 1788 } 1789 bdiag[i] = nzbd; 1790 bi[i+1] = bi[i] + nzi; 1791 1792 /* if free space is not available, make more free space */ 1793 if (current_space->local_remaining<nzi) { 1794 nnz = 2*nzi*(n - i); /* estimated and max additional space needed */ 1795 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 1796 ierr = PetscFreeSpaceGet(nnz,¤t_space_lvl);CHKERRQ(ierr); 1797 reallocs++; 1798 } 1799 1800 /* copy data into free_space and free_space_lvl, then initialize lnk */ 1801 ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 1802 bj_ptr[i] = current_space->array; 1803 bjlvl_ptr[i] = current_space_lvl->array; 1804 1805 /* make sure the active row i has diagonal entry */ 1806 if (*(bj_ptr[i]+bdiag[i]) != i) { 1807 SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\ 1808 try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i); 1809 } 1810 1811 current_space->array += nzi; 1812 current_space->local_used += nzi; 1813 current_space->local_remaining -= nzi; 1814 current_space_lvl->array += nzi; 1815 current_space_lvl->local_used += nzi; 1816 current_space_lvl->local_remaining -= nzi; 1817 } 1818 1819 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 1820 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 1821 1822 /* destroy list of free space and other temporary arrays */ 1823 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1824 1825 /* copy free_space into bj and free free_space; set bi, bj, bdiag in new datastructure; */ 1826 ierr = PetscFreeSpaceContiguous_LU_v2(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 1827 1828 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1829 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 1830 ierr = PetscFree2(bj_ptr,bjlvl_ptr);CHKERRQ(ierr); 1831 1832 #if defined(PETSC_USE_INFO) 1833 { 1834 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 1835 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 1836 ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 1837 ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr); 1838 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 1839 if (diagonal_fill) { 1840 ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr); 1841 } 1842 } 1843 #endif 1844 1845 /* put together the new matrix */ 1846 ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 1847 ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr); 1848 b = (Mat_SeqAIJ*)(fact)->data; 1849 b->free_a = PETSC_TRUE; 1850 b->free_ij = PETSC_TRUE; 1851 b->singlemalloc = PETSC_FALSE; 1852 ierr = PetscMalloc((bdiag[0]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 1853 b->j = bj; 1854 b->i = bi; 1855 b->diag = bdiag; 1856 b->ilen = 0; 1857 b->imax = 0; 1858 b->row = isrow; 1859 b->col = iscol; 1860 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1861 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1862 b->icol = isicol; 1863 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1864 /* In b structure: Free imax, ilen, old a, old j. 1865 Allocate bdiag, solve_work, new a, new j */ 1866 ierr = PetscLogObjectMemory(fact,(bdiag[0]+1)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 1867 b->maxnz = b->nz = bdiag[0]+1; 1868 (fact)->info.factor_mallocs = reallocs; 1869 (fact)->info.fill_ratio_given = f; 1870 (fact)->info.fill_ratio_needed = ((PetscReal)(bdiag[0]+1))/((PetscReal)ai[n]); 1871 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 1872 ierr = Mat_CheckInode_FactorLU(fact,PETSC_FALSE);CHKERRQ(ierr); 1873 PetscFunctionReturn(0); 1874 } 1875 1876 #undef __FUNCT__ 1877 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_inplace" 1878 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_inplace(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1879 { 1880 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1881 IS isicol; 1882 PetscErrorCode ierr; 1883 const PetscInt *r,*ic; 1884 PetscInt n=A->rmap->n,*ai=a->i,*aj=a->j,d; 1885 PetscInt *bi,*cols,nnz,*cols_lvl; 1886 PetscInt *bdiag,prow,fm,nzbd,reallocs=0,dcount=0; 1887 PetscInt i,levels,diagonal_fill; 1888 PetscTruth col_identity,row_identity; 1889 PetscReal f; 1890 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL; 1891 PetscBT lnkbt; 1892 PetscInt nzi,*bj,**bj_ptr,**bjlvl_ptr; 1893 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1894 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 1895 PetscTruth missing; 1896 1897 PetscFunctionBegin; 1898 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); 1899 f = info->fill; 1900 levels = (PetscInt)info->levels; 1901 diagonal_fill = (PetscInt)info->diagonal_fill; 1902 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1903 1904 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 1905 ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 1906 if (!levels && row_identity && col_identity) { /* special case: ilu(0) with natural ordering */ 1907 ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES,PETSC_TRUE);CHKERRQ(ierr); 1908 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_inplace; 1909 1910 fact->factor = MAT_FACTOR_ILU; 1911 (fact)->info.factor_mallocs = 0; 1912 (fact)->info.fill_ratio_given = info->fill; 1913 (fact)->info.fill_ratio_needed = 1.0; 1914 b = (Mat_SeqAIJ*)(fact)->data; 1915 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 1916 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 1917 b->row = isrow; 1918 b->col = iscol; 1919 b->icol = isicol; 1920 ierr = PetscMalloc(((fact)->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1921 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1922 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1923 ierr = MatILUFactorSymbolic_SeqAIJ_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1924 PetscFunctionReturn(0); 1925 } 1926 1927 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1928 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 1929 1930 /* get new row and diagonal pointers, must be allocated separately because they will be given to the Mat_SeqAIJ and freed separately */ 1931 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1932 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 1933 bi[0] = bdiag[0] = 0; 1934 1935 ierr = PetscMalloc2(n,PetscInt*,&bj_ptr,n,PetscInt*,&bjlvl_ptr);CHKERRQ(ierr); 1936 1937 /* create a linked list for storing column indices of the active row */ 1938 nlnk = n + 1; 1939 ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1940 1941 /* initial FreeSpace size is f*(ai[n]+1) */ 1942 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 1943 current_space = free_space; 1944 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr); 1945 current_space_lvl = free_space_lvl; 1946 1947 for (i=0; i<n; i++) { 1948 nzi = 0; 1949 /* copy current row into linked list */ 1950 nnz = ai[r[i]+1] - ai[r[i]]; 1951 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 1952 cols = aj + ai[r[i]]; 1953 lnk[i] = -1; /* marker to indicate if diagonal exists */ 1954 ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1955 nzi += nlnk; 1956 1957 /* make sure diagonal entry is included */ 1958 if (diagonal_fill && lnk[i] == -1) { 1959 fm = n; 1960 while (lnk[fm] < i) fm = lnk[fm]; 1961 lnk[i] = lnk[fm]; /* insert diagonal into linked list */ 1962 lnk[fm] = i; 1963 lnk_lvl[i] = 0; 1964 nzi++; dcount++; 1965 } 1966 1967 /* add pivot rows into the active row */ 1968 nzbd = 0; 1969 prow = lnk[n]; 1970 while (prow < i) { 1971 nnz = bdiag[prow]; 1972 cols = bj_ptr[prow] + nnz + 1; 1973 cols_lvl = bjlvl_ptr[prow] + nnz + 1; 1974 nnz = bi[prow+1] - bi[prow] - nnz - 1; 1975 ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr); 1976 nzi += nlnk; 1977 prow = lnk[prow]; 1978 nzbd++; 1979 } 1980 bdiag[i] = nzbd; 1981 bi[i+1] = bi[i] + nzi; 1982 1983 /* if free space is not available, make more free space */ 1984 if (current_space->local_remaining<nzi) { 1985 nnz = nzi*(n - i); /* estimated and max additional space needed */ 1986 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 1987 ierr = PetscFreeSpaceGet(nnz,¤t_space_lvl);CHKERRQ(ierr); 1988 reallocs++; 1989 } 1990 1991 /* copy data into free_space and free_space_lvl, then initialize lnk */ 1992 ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 1993 bj_ptr[i] = current_space->array; 1994 bjlvl_ptr[i] = current_space_lvl->array; 1995 1996 /* make sure the active row i has diagonal entry */ 1997 if (*(bj_ptr[i]+bdiag[i]) != i) { 1998 SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\ 1999 try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i); 2000 } 2001 2002 current_space->array += nzi; 2003 current_space->local_used += nzi; 2004 current_space->local_remaining -= nzi; 2005 current_space_lvl->array += nzi; 2006 current_space_lvl->local_used += nzi; 2007 current_space_lvl->local_remaining -= nzi; 2008 } 2009 2010 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 2011 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 2012 2013 /* destroy list of free space and other temporary arrays */ 2014 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 2015 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); /* copy free_space -> bj */ 2016 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2017 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2018 ierr = PetscFree2(bj_ptr,bjlvl_ptr);CHKERRQ(ierr); 2019 2020 #if defined(PETSC_USE_INFO) 2021 { 2022 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 2023 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 2024 ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2025 ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr); 2026 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 2027 if (diagonal_fill) { 2028 ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr); 2029 } 2030 } 2031 #endif 2032 2033 /* put together the new matrix */ 2034 ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 2035 ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr); 2036 b = (Mat_SeqAIJ*)(fact)->data; 2037 b->free_a = PETSC_TRUE; 2038 b->free_ij = PETSC_TRUE; 2039 b->singlemalloc = PETSC_FALSE; 2040 ierr = PetscMalloc(bi[n]*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 2041 b->j = bj; 2042 b->i = bi; 2043 for (i=0; i<n; i++) bdiag[i] += bi[i]; 2044 b->diag = bdiag; 2045 b->ilen = 0; 2046 b->imax = 0; 2047 b->row = isrow; 2048 b->col = iscol; 2049 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2050 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2051 b->icol = isicol; 2052 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2053 /* In b structure: Free imax, ilen, old a, old j. 2054 Allocate bdiag, solve_work, new a, new j */ 2055 ierr = PetscLogObjectMemory(fact,(bi[n]-n) * (sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 2056 b->maxnz = b->nz = bi[n] ; 2057 (fact)->info.factor_mallocs = reallocs; 2058 (fact)->info.fill_ratio_given = f; 2059 (fact)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]); 2060 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_inplace; 2061 ierr = MatILUFactorSymbolic_SeqAIJ_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); 2062 PetscFunctionReturn(0); 2063 } 2064 2065 #undef __FUNCT__ 2066 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ_newdatastruct" 2067 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 2068 { 2069 Mat C = B; 2070 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 2071 Mat_SeqSBAIJ *b=(Mat_SeqSBAIJ*)C->data; 2072 IS ip=b->row,iip = b->icol; 2073 PetscErrorCode ierr; 2074 const PetscInt *rip,*riip; 2075 PetscInt i,j,mbs=A->rmap->n,*bi=b->i,*bj=b->j,*bdiag=b->diag,*bjtmp; 2076 PetscInt *ai=a->i,*aj=a->j; 2077 PetscInt k,jmin,jmax,*c2r,*il,col,nexti,ili,nz; 2078 MatScalar *rtmp,*ba=b->a,*bval,*aa=a->a,dk,uikdi; 2079 PetscTruth perm_identity; 2080 2081 LUShift_Ctx sctx; 2082 PetscInt newshift; 2083 PetscReal rs; 2084 MatScalar d,*v; 2085 2086 PetscFunctionBegin; 2087 /* MatPivotSetUp(): initialize shift context sctx */ 2088 ierr = PetscMemzero(&sctx,sizeof(LUShift_Ctx));CHKERRQ(ierr); 2089 2090 /* if both shift schemes are chosen by user, only use info->shiftpd */ 2091 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 2092 sctx.shift_top = info->zeropivot; 2093 for (i=0; i<mbs; i++) { 2094 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 2095 d = (aa)[a->diag[i]]; 2096 rs = -PetscAbsScalar(d) - PetscRealPart(d); 2097 v = aa+ai[i]; 2098 nz = ai[i+1] - ai[i]; 2099 for (j=0; j<nz; j++) 2100 rs += PetscAbsScalar(v[j]); 2101 if (rs>sctx.shift_top) sctx.shift_top = rs; 2102 } 2103 sctx.shift_top *= 1.1; 2104 sctx.nshift_max = 5; 2105 sctx.shift_lo = 0.; 2106 sctx.shift_hi = 1.; 2107 } 2108 2109 ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 2110 ierr = ISGetIndices(iip,&riip);CHKERRQ(ierr); 2111 2112 /* allocate working arrays 2113 c2r: linked list, keep track of pivot rows for a given column. c2r[col]: head of the list for a given col 2114 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 2115 */ 2116 ierr = PetscMalloc3(mbs,MatScalar,&rtmp,mbs,PetscInt,&il,mbs,PetscInt,&c2r);CHKERRQ(ierr); 2117 2118 do { 2119 sctx.lushift = PETSC_FALSE; 2120 2121 for (i=0; i<mbs; i++) c2r[i] = mbs; 2122 il[0] = 0; 2123 2124 for (k = 0; k<mbs; k++){ 2125 /* zero rtmp */ 2126 nz = bi[k+1] - bi[k]; 2127 bjtmp = bj + bi[k]; 2128 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 2129 2130 /* load in initial unfactored row */ 2131 bval = ba + bi[k]; 2132 jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 2133 for (j = jmin; j < jmax; j++){ 2134 col = riip[aj[j]]; 2135 if (col >= k){ /* only take upper triangular entry */ 2136 rtmp[col] = aa[j]; 2137 *bval++ = 0.0; /* for in-place factorization */ 2138 } 2139 } 2140 /* shift the diagonal of the matrix: ZeropivotApply() */ 2141 rtmp[k] += sctx.shift_amount; /* shift the diagonal of the matrix */ 2142 2143 /* modify k-th row by adding in those rows i with U(i,k)!=0 */ 2144 dk = rtmp[k]; 2145 i = c2r[k]; /* first row to be added to k_th row */ 2146 2147 while (i < k){ 2148 nexti = c2r[i]; /* next row to be added to k_th row */ 2149 2150 /* compute multiplier, update diag(k) and U(i,k) */ 2151 ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2152 uikdi = - ba[ili]*ba[bdiag[i]]; /* diagonal(k) */ 2153 dk += uikdi*ba[ili]; /* update diag[k] */ 2154 ba[ili] = uikdi; /* -U(i,k) */ 2155 2156 /* add multiple of row i to k-th row */ 2157 jmin = ili + 1; jmax = bi[i+1]; 2158 if (jmin < jmax){ 2159 for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 2160 /* update il and c2r for row i */ 2161 il[i] = jmin; 2162 j = bj[jmin]; c2r[i] = c2r[j]; c2r[j] = i; 2163 } 2164 i = nexti; 2165 } 2166 2167 /* copy data into U(k,:) */ 2168 rs = 0.0; 2169 jmin = bi[k]; jmax = bi[k+1]-1; 2170 if (jmin < jmax) { 2171 for (j=jmin; j<jmax; j++){ 2172 col = bj[j]; ba[j] = rtmp[col]; rs += PetscAbsScalar(ba[j]); 2173 } 2174 /* add the k-th row into il and c2r */ 2175 il[k] = jmin; 2176 i = bj[jmin]; c2r[k] = c2r[i]; c2r[i] = k; 2177 } 2178 2179 /* MatPivotCheck() */ 2180 sctx.rs = rs; 2181 sctx.pv = dk; 2182 if (info->shiftnz){ 2183 ierr = MatPivotCheck_nz(info,sctx,k,newshift);CHKERRQ(ierr); 2184 } else if (info->shiftpd){ 2185 ierr = MatPivotCheck_pd(info,sctx,k,newshift);CHKERRQ(ierr); 2186 } else if (info->shiftinblocks){ 2187 ierr = MatPivotCheck_inblocks(info,sctx,k,newshift);CHKERRQ(ierr); 2188 } else { 2189 ierr = MatPivotCheck_none(info,sctx,k,newshift);CHKERRQ(ierr); 2190 } 2191 dk = sctx.pv; 2192 if (newshift == 1) break; 2193 2194 ba[bdiag[k]] = 1.0/dk; /* U(k,k) */ 2195 } 2196 } while (sctx.lushift); 2197 2198 ierr = PetscFree3(rtmp,il,c2r);CHKERRQ(ierr); 2199 ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 2200 ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr); 2201 2202 ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr); 2203 if (perm_identity){ 2204 B->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering_newdatastruct; 2205 B->ops->solvetranspose = MatSolve_SeqSBAIJ_1_NaturalOrdering_newdatastruct; 2206 B->ops->forwardsolve = 0; 2207 B->ops->backwardsolve = 0; 2208 } else { 2209 B->ops->solve = MatSolve_SeqSBAIJ_1_newdatastruct; 2210 B->ops->solvetranspose = MatSolve_SeqSBAIJ_1_newdatastruct; 2211 B->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1_newdatastruct; 2212 B->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1_newdatastruct; 2213 } 2214 2215 C->assembled = PETSC_TRUE; 2216 C->preallocated = PETSC_TRUE; 2217 ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr); 2218 2219 /* MatPivotView() */ 2220 if (sctx.nshift){ 2221 if (info->shiftpd) { 2222 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); 2223 } else if (info->shiftnz) { 2224 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2225 } else if (info->shiftinblocks){ 2226 ierr = PetscInfo2(A,"number of shift_inblocks applied %D, each shift_amount %G\n",sctx.nshift,info->shiftinblocks);CHKERRQ(ierr); 2227 } 2228 } 2229 PetscFunctionReturn(0); 2230 } 2231 2232 #undef __FUNCT__ 2233 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ_inplace" 2234 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_inplace(Mat B,Mat A,const MatFactorInfo *info) 2235 { 2236 Mat C = B; 2237 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 2238 Mat_SeqSBAIJ *b=(Mat_SeqSBAIJ*)C->data; 2239 IS ip=b->row,iip = b->icol; 2240 PetscErrorCode ierr; 2241 const PetscInt *rip,*riip; 2242 PetscInt i,j,mbs=A->rmap->n,*bi=b->i,*bj=b->j,*bcol,*bjtmp; 2243 PetscInt *ai=a->i,*aj=a->j; 2244 PetscInt k,jmin,jmax,*jl,*il,col,nexti,ili,nz; 2245 MatScalar *rtmp,*ba=b->a,*bval,*aa=a->a,dk,uikdi; 2246 PetscReal zeropivot,rs,shiftnz; 2247 PetscReal shiftpd; 2248 ChShift_Ctx sctx; 2249 PetscInt newshift; 2250 PetscTruth perm_identity; 2251 2252 PetscFunctionBegin; 2253 shiftnz = info->shiftnz; 2254 shiftpd = info->shiftpd; 2255 zeropivot = info->zeropivot; 2256 2257 ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 2258 ierr = ISGetIndices(iip,&riip);CHKERRQ(ierr); 2259 2260 /* initialization */ 2261 ierr = PetscMalloc3(mbs,MatScalar,&rtmp,mbs,PetscInt,&il,mbs,PetscInt,&jl);CHKERRQ(ierr); 2262 sctx.shift_amount = 0; 2263 sctx.nshift = 0; 2264 do { 2265 sctx.chshift = PETSC_FALSE; 2266 for (i=0; i<mbs; i++) jl[i] = mbs; 2267 il[0] = 0; 2268 2269 for (k = 0; k<mbs; k++){ 2270 /* zero rtmp */ 2271 nz = bi[k+1] - bi[k]; 2272 bjtmp = bj + bi[k]; 2273 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 2274 2275 bval = ba + bi[k]; 2276 /* initialize k-th row by the perm[k]-th row of A */ 2277 jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 2278 for (j = jmin; j < jmax; j++){ 2279 col = riip[aj[j]]; 2280 if (col >= k){ /* only take upper triangular entry */ 2281 rtmp[col] = aa[j]; 2282 *bval++ = 0.0; /* for in-place factorization */ 2283 } 2284 } 2285 /* shift the diagonal of the matrix */ 2286 if (sctx.nshift) rtmp[k] += sctx.shift_amount; 2287 2288 /* modify k-th row by adding in those rows i with U(i,k)!=0 */ 2289 dk = rtmp[k]; 2290 i = jl[k]; /* first row to be added to k_th row */ 2291 2292 while (i < k){ 2293 nexti = jl[i]; /* next row to be added to k_th row */ 2294 2295 /* compute multiplier, update diag(k) and U(i,k) */ 2296 ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2297 uikdi = - ba[ili]*ba[bi[i]]; /* diagonal(k) */ 2298 dk += uikdi*ba[ili]; 2299 ba[ili] = uikdi; /* -U(i,k) */ 2300 2301 /* add multiple of row i to k-th row */ 2302 jmin = ili + 1; jmax = bi[i+1]; 2303 if (jmin < jmax){ 2304 for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 2305 /* update il and jl for row i */ 2306 il[i] = jmin; 2307 j = bj[jmin]; jl[i] = jl[j]; jl[j] = i; 2308 } 2309 i = nexti; 2310 } 2311 2312 /* shift the diagonals when zero pivot is detected */ 2313 /* compute rs=sum of abs(off-diagonal) */ 2314 rs = 0.0; 2315 jmin = bi[k]+1; 2316 nz = bi[k+1] - jmin; 2317 bcol = bj + jmin; 2318 for (j=0; j<nz; j++) { 2319 rs += PetscAbsScalar(rtmp[bcol[j]]); 2320 } 2321 2322 sctx.rs = rs; 2323 sctx.pv = dk; 2324 ierr = MatCholeskyCheckShift_inline(info,sctx,k,newshift);CHKERRQ(ierr); 2325 2326 if (newshift == 1) { 2327 if (!sctx.shift_amount) { 2328 sctx.shift_amount = 1e-5; 2329 } 2330 break; 2331 } 2332 2333 /* copy data into U(k,:) */ 2334 ba[bi[k]] = 1.0/dk; /* U(k,k) */ 2335 jmin = bi[k]+1; jmax = bi[k+1]; 2336 if (jmin < jmax) { 2337 for (j=jmin; j<jmax; j++){ 2338 col = bj[j]; ba[j] = rtmp[col]; 2339 } 2340 /* add the k-th row into il and jl */ 2341 il[k] = jmin; 2342 i = bj[jmin]; jl[k] = jl[i]; jl[i] = k; 2343 } 2344 } 2345 } while (sctx.chshift); 2346 ierr = PetscFree3(rtmp,il,jl);CHKERRQ(ierr); 2347 ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 2348 ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr); 2349 2350 ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr); 2351 if (perm_identity){ 2352 B->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering_inplace; 2353 B->ops->solvetranspose = MatSolve_SeqSBAIJ_1_NaturalOrdering_inplace; 2354 B->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1_NaturalOrdering_inplace; 2355 B->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1_NaturalOrdering_inplace; 2356 } else { 2357 B->ops->solve = MatSolve_SeqSBAIJ_1_inplace; 2358 B->ops->solvetranspose = MatSolve_SeqSBAIJ_1_inplace; 2359 B->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1_inplace; 2360 B->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1_inplace; 2361 } 2362 2363 C->assembled = PETSC_TRUE; 2364 C->preallocated = PETSC_TRUE; 2365 ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr); 2366 if (sctx.nshift){ 2367 if (shiftnz) { 2368 ierr = PetscInfo2(A,"number of shiftnz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2369 } else if (shiftpd) { 2370 ierr = PetscInfo2(A,"number of shiftpd tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2371 } 2372 } 2373 PetscFunctionReturn(0); 2374 } 2375 2376 /* 2377 icc() under revised new data structure. 2378 Factored arrays bj and ba are stored as 2379 U(0,:),...,U(i,:),U(n-1,:) 2380 2381 ui=fact->i is an array of size n+1, in which 2382 ui+ 2383 ui[i]: points to 1st entry of U(i,:),i=0,...,n-1 2384 ui[n]: points to U(n-1,n-1)+1 2385 2386 udiag=fact->diag is an array of size n,in which 2387 udiag[i]: points to diagonal of U(i,:), i=0,...,n-1 2388 2389 U(i,:) contains udiag[i] as its last entry, i.e., 2390 U(i,:) = (u[i,i+1],...,u[i,n-1],diag[i]) 2391 */ 2392 2393 #undef __FUNCT__ 2394 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ_newdatastruct" 2395 PetscErrorCode MatICCFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2396 { 2397 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2398 Mat_SeqSBAIJ *b; 2399 PetscErrorCode ierr; 2400 PetscTruth perm_identity,missing; 2401 PetscInt reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui,*udiag; 2402 const PetscInt *rip,*riip; 2403 PetscInt jmin,jmax,nzk,k,j,*jl,prow,*il,nextprow; 2404 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL,d; 2405 PetscInt ncols,ncols_upper,*cols,*ajtmp,*uj,**uj_ptr,**uj_lvl_ptr; 2406 PetscReal fill=info->fill,levels=info->levels; 2407 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2408 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2409 PetscBT lnkbt; 2410 IS iperm; 2411 PetscTruth newdatastruct=PETSC_FALSE; 2412 2413 PetscFunctionBegin; 2414 ierr = PetscOptionsGetTruth(PETSC_NULL,"-icc_old",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 2415 if(newdatastruct){ 2416 ierr = MatICCFactorSymbolic_SeqAIJ_inplace(fact,A,perm,info);CHKERRQ(ierr); 2417 PetscFunctionReturn(0); 2418 } 2419 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); 2420 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2421 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2422 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2423 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2424 2425 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2426 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2427 ui[0] = 0; 2428 2429 /* ICC(0) without matrix ordering: simply rearrange column indices */ 2430 if (!levels && perm_identity) { 2431 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2432 cols = uj; 2433 for (i=0; i<am; i++) { 2434 ncols = ai[i+1] - a->diag[i]; 2435 ui[i+1] = ui[i] + ncols; 2436 udiag[i] = ui[i+1] - 1; /* points to the last entry of U(i,:) */ 2437 2438 aj = a->j + a->diag[i] + 1; /* 1st entry of U(i,:) without diagonal */ 2439 ncols--; /* exclude diagonal */ 2440 for (j=0; j<ncols; j++) *cols++ = aj[j]; 2441 *cols++ = i; /* diagoanl is located as the last entry of U(i,:) */ 2442 } 2443 } else { /* case: levels>0 || (levels=0 && !perm_identity) */ 2444 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2445 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2446 2447 /* initialization */ 2448 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ajtmp);CHKERRQ(ierr); 2449 2450 /* jl: linked list for storing indices of the pivot rows 2451 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2452 ierr = PetscMalloc4(am,PetscInt*,&uj_ptr,am,PetscInt*,&uj_lvl_ptr,am,PetscInt,&jl,am,PetscInt,&il);CHKERRQ(ierr); 2453 for (i=0; i<am; i++){ 2454 jl[i] = am; il[i] = 0; 2455 } 2456 2457 /* create and initialize a linked list for storing column indices of the active row k */ 2458 nlnk = am + 1; 2459 ierr = PetscIncompleteLLCreate(am,am,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2460 2461 /* initial FreeSpace size is fill*(ai[am]+1) */ 2462 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2463 current_space = free_space; 2464 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space_lvl);CHKERRQ(ierr); 2465 current_space_lvl = free_space_lvl; 2466 2467 for (k=0; k<am; k++){ /* for each active row k */ 2468 /* initialize lnk by the column indices of row rip[k] of A */ 2469 nzk = 0; 2470 ncols = ai[rip[k]+1] - ai[rip[k]]; 2471 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2472 ncols_upper = 0; 2473 for (j=0; j<ncols; j++){ 2474 i = *(aj + ai[rip[k]] + j); /* unpermuted column index */ 2475 if (riip[i] >= k){ /* only take upper triangular entry */ 2476 ajtmp[ncols_upper] = i; 2477 ncols_upper++; 2478 } 2479 } 2480 ierr = PetscIncompleteLLInit(ncols_upper,ajtmp,am,riip,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2481 nzk += nlnk; 2482 2483 /* update lnk by computing fill-in for each pivot row to be merged in */ 2484 prow = jl[k]; /* 1st pivot row */ 2485 2486 while (prow < k){ 2487 nextprow = jl[prow]; 2488 2489 /* merge prow into k-th row */ 2490 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2491 jmax = ui[prow+1]; 2492 ncols = jmax-jmin; 2493 i = jmin - ui[prow]; 2494 cols = uj_ptr[prow] + i; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2495 uj = uj_lvl_ptr[prow] + i; /* levels of cols */ 2496 j = *(uj - 1); 2497 ierr = PetscICCLLAddSorted(ncols,cols,levels,uj,am,nlnk,lnk,lnk_lvl,lnkbt,j);CHKERRQ(ierr); 2498 nzk += nlnk; 2499 2500 /* update il and jl for prow */ 2501 if (jmin < jmax){ 2502 il[prow] = jmin; 2503 j = *cols; jl[prow] = jl[j]; jl[j] = prow; 2504 } 2505 prow = nextprow; 2506 } 2507 2508 /* if free space is not available, make more free space */ 2509 if (current_space->local_remaining<nzk) { 2510 i = am - k + 1; /* num of unfactored rows */ 2511 i *= PetscMin(nzk, i-1); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2512 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2513 ierr = PetscFreeSpaceGet(i,¤t_space_lvl);CHKERRQ(ierr); 2514 reallocs++; 2515 } 2516 2517 /* copy data into free_space and free_space_lvl, then initialize lnk */ 2518 if (nzk == 0) SETERRQ1(PETSC_ERR_ARG_WRONG,"Empty row %D in ICC matrix factor",k); 2519 ierr = PetscIncompleteLLClean(am,am,nzk,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 2520 2521 /* add the k-th row into il and jl */ 2522 if (nzk > 1){ 2523 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2524 jl[k] = jl[i]; jl[i] = k; 2525 il[k] = ui[k] + 1; 2526 } 2527 uj_ptr[k] = current_space->array; 2528 uj_lvl_ptr[k] = current_space_lvl->array; 2529 2530 current_space->array += nzk; 2531 current_space->local_used += nzk; 2532 current_space->local_remaining -= nzk; 2533 2534 current_space_lvl->array += nzk; 2535 current_space_lvl->local_used += nzk; 2536 current_space_lvl->local_remaining -= nzk; 2537 2538 ui[k+1] = ui[k] + nzk; 2539 } 2540 2541 #if defined(PETSC_USE_INFO) 2542 if (ai[am] != 0) { 2543 PetscReal af = (PetscReal)ui[am]/((PetscReal)ai[am]); 2544 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2545 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2546 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2547 } else { 2548 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2549 } 2550 #endif 2551 2552 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2553 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2554 ierr = PetscFree4(uj_ptr,uj_lvl_ptr,jl,il);CHKERRQ(ierr); 2555 ierr = PetscFree(ajtmp);CHKERRQ(ierr); 2556 2557 /* destroy list of free space and other temporary array(s) */ 2558 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2559 ierr = PetscFreeSpaceContiguous_Cholesky(&free_space,uj,am,ui,udiag);CHKERRQ(ierr); /* store matrix factor in newdatastruct */ 2560 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2561 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2562 2563 } /* end of case: levels>0 || (levels=0 && !perm_identity) */ 2564 2565 /* put together the new matrix in MATSEQSBAIJ format */ 2566 2567 b = (Mat_SeqSBAIJ*)(fact)->data; 2568 b->singlemalloc = PETSC_FALSE; 2569 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2570 b->j = uj; 2571 b->i = ui; 2572 b->diag = udiag; 2573 b->free_diag = PETSC_TRUE; 2574 b->ilen = 0; 2575 b->imax = 0; 2576 b->row = perm; 2577 b->col = perm; 2578 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2579 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2580 b->icol = iperm; 2581 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2582 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2583 ierr = PetscLogObjectMemory((fact),(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2584 b->maxnz = b->nz = ui[am]; 2585 b->free_a = PETSC_TRUE; 2586 b->free_ij = PETSC_TRUE; 2587 2588 fact->info.factor_mallocs = reallocs; 2589 fact->info.fill_ratio_given = fill; 2590 if (ai[am] != 0) { 2591 fact->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2592 } else { 2593 fact->info.fill_ratio_needed = 0.0; 2594 } 2595 fact->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_newdatastruct; 2596 PetscFunctionReturn(0); 2597 } 2598 2599 #undef __FUNCT__ 2600 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ_inplace" 2601 PetscErrorCode MatICCFactorSymbolic_SeqAIJ_inplace(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2602 { 2603 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2604 Mat_SeqSBAIJ *b; 2605 PetscErrorCode ierr; 2606 PetscTruth perm_identity,missing; 2607 PetscInt reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui,*udiag; 2608 const PetscInt *rip,*riip; 2609 PetscInt jmin,jmax,nzk,k,j,*jl,prow,*il,nextprow; 2610 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL,d; 2611 PetscInt ncols,ncols_upper,*cols,*ajtmp,*uj,**uj_ptr,**uj_lvl_ptr; 2612 PetscReal fill=info->fill,levels=info->levels; 2613 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2614 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2615 PetscBT lnkbt; 2616 IS iperm; 2617 2618 PetscFunctionBegin; 2619 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); 2620 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2621 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2622 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2623 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2624 2625 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2626 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2627 ui[0] = 0; 2628 2629 /* ICC(0) without matrix ordering: simply copies fill pattern */ 2630 if (!levels && perm_identity) { 2631 2632 for (i=0; i<am; i++) { 2633 ui[i+1] = ui[i] + ai[i+1] - a->diag[i]; 2634 udiag[i] = ui[i]; 2635 } 2636 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2637 cols = uj; 2638 for (i=0; i<am; i++) { 2639 aj = a->j + a->diag[i]; 2640 ncols = ui[i+1] - ui[i]; 2641 for (j=0; j<ncols; j++) *cols++ = *aj++; 2642 } 2643 } else { /* case: levels>0 || (levels=0 && !perm_identity) */ 2644 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2645 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2646 2647 /* initialization */ 2648 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ajtmp);CHKERRQ(ierr); 2649 2650 /* jl: linked list for storing indices of the pivot rows 2651 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2652 ierr = PetscMalloc4(am,PetscInt*,&uj_ptr,am,PetscInt*,&uj_lvl_ptr,am,PetscInt,&jl,am,PetscInt,&il);CHKERRQ(ierr); 2653 for (i=0; i<am; i++){ 2654 jl[i] = am; il[i] = 0; 2655 } 2656 2657 /* create and initialize a linked list for storing column indices of the active row k */ 2658 nlnk = am + 1; 2659 ierr = PetscIncompleteLLCreate(am,am,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2660 2661 /* initial FreeSpace size is fill*(ai[am]+1) */ 2662 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2663 current_space = free_space; 2664 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space_lvl);CHKERRQ(ierr); 2665 current_space_lvl = free_space_lvl; 2666 2667 for (k=0; k<am; k++){ /* for each active row k */ 2668 /* initialize lnk by the column indices of row rip[k] of A */ 2669 nzk = 0; 2670 ncols = ai[rip[k]+1] - ai[rip[k]]; 2671 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2672 ncols_upper = 0; 2673 for (j=0; j<ncols; j++){ 2674 i = *(aj + ai[rip[k]] + j); /* unpermuted column index */ 2675 if (riip[i] >= k){ /* only take upper triangular entry */ 2676 ajtmp[ncols_upper] = i; 2677 ncols_upper++; 2678 } 2679 } 2680 ierr = PetscIncompleteLLInit(ncols_upper,ajtmp,am,riip,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2681 nzk += nlnk; 2682 2683 /* update lnk by computing fill-in for each pivot row to be merged in */ 2684 prow = jl[k]; /* 1st pivot row */ 2685 2686 while (prow < k){ 2687 nextprow = jl[prow]; 2688 2689 /* merge prow into k-th row */ 2690 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2691 jmax = ui[prow+1]; 2692 ncols = jmax-jmin; 2693 i = jmin - ui[prow]; 2694 cols = uj_ptr[prow] + i; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2695 uj = uj_lvl_ptr[prow] + i; /* levels of cols */ 2696 j = *(uj - 1); 2697 ierr = PetscICCLLAddSorted(ncols,cols,levels,uj,am,nlnk,lnk,lnk_lvl,lnkbt,j);CHKERRQ(ierr); 2698 nzk += nlnk; 2699 2700 /* update il and jl for prow */ 2701 if (jmin < jmax){ 2702 il[prow] = jmin; 2703 j = *cols; jl[prow] = jl[j]; jl[j] = prow; 2704 } 2705 prow = nextprow; 2706 } 2707 2708 /* if free space is not available, make more free space */ 2709 if (current_space->local_remaining<nzk) { 2710 i = am - k + 1; /* num of unfactored rows */ 2711 i *= PetscMin(nzk, (i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2712 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2713 ierr = PetscFreeSpaceGet(i,¤t_space_lvl);CHKERRQ(ierr); 2714 reallocs++; 2715 } 2716 2717 /* copy data into free_space and free_space_lvl, then initialize lnk */ 2718 if (nzk == 0) SETERRQ1(PETSC_ERR_ARG_WRONG,"Empty row %D in ICC matrix factor",k); 2719 ierr = PetscIncompleteLLClean(am,am,nzk,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 2720 2721 /* add the k-th row into il and jl */ 2722 if (nzk > 1){ 2723 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2724 jl[k] = jl[i]; jl[i] = k; 2725 il[k] = ui[k] + 1; 2726 } 2727 uj_ptr[k] = current_space->array; 2728 uj_lvl_ptr[k] = current_space_lvl->array; 2729 2730 current_space->array += nzk; 2731 current_space->local_used += nzk; 2732 current_space->local_remaining -= nzk; 2733 2734 current_space_lvl->array += nzk; 2735 current_space_lvl->local_used += nzk; 2736 current_space_lvl->local_remaining -= nzk; 2737 2738 ui[k+1] = ui[k] + nzk; 2739 } 2740 2741 #if defined(PETSC_USE_INFO) 2742 if (ai[am] != 0) { 2743 PetscReal af = (PetscReal)ui[am]/((PetscReal)ai[am]); 2744 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2745 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2746 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2747 } else { 2748 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2749 } 2750 #endif 2751 2752 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2753 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2754 ierr = PetscFree4(uj_ptr,uj_lvl_ptr,jl,il);CHKERRQ(ierr); 2755 ierr = PetscFree(ajtmp);CHKERRQ(ierr); 2756 2757 /* destroy list of free space and other temporary array(s) */ 2758 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2759 ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr); 2760 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2761 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2762 2763 } /* end of case: levels>0 || (levels=0 && !perm_identity) */ 2764 2765 /* put together the new matrix in MATSEQSBAIJ format */ 2766 2767 b = (Mat_SeqSBAIJ*)fact->data; 2768 b->singlemalloc = PETSC_FALSE; 2769 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2770 b->j = uj; 2771 b->i = ui; 2772 b->diag = udiag; 2773 b->free_diag = PETSC_TRUE; 2774 b->ilen = 0; 2775 b->imax = 0; 2776 b->row = perm; 2777 b->col = perm; 2778 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2779 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2780 b->icol = iperm; 2781 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2782 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2783 ierr = PetscLogObjectMemory(fact,(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2784 b->maxnz = b->nz = ui[am]; 2785 b->free_a = PETSC_TRUE; 2786 b->free_ij = PETSC_TRUE; 2787 2788 fact->info.factor_mallocs = reallocs; 2789 fact->info.fill_ratio_given = fill; 2790 if (ai[am] != 0) { 2791 fact->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2792 } else { 2793 fact->info.fill_ratio_needed = 0.0; 2794 } 2795 fact->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_inplace; 2796 PetscFunctionReturn(0); 2797 } 2798 2799 PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2800 { 2801 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2802 Mat_SeqSBAIJ *b; 2803 PetscErrorCode ierr; 2804 PetscTruth perm_identity; 2805 PetscReal fill = info->fill; 2806 const PetscInt *rip,*riip; 2807 PetscInt i,am=A->rmap->n,*ai=a->i,*aj=a->j,reallocs=0,prow; 2808 PetscInt *jl,jmin,jmax,nzk,*ui,k,j,*il,nextprow; 2809 PetscInt nlnk,*lnk,ncols,ncols_upper,*cols,*uj,**ui_ptr,*uj_ptr,*udiag; 2810 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2811 PetscBT lnkbt; 2812 IS iperm; 2813 PetscTruth newdatastruct=PETSC_FALSE; 2814 2815 PetscFunctionBegin; 2816 ierr = PetscOptionsGetTruth(PETSC_NULL,"-cholesky_old",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 2817 if(newdatastruct){ 2818 ierr = MatCholeskyFactorSymbolic_SeqAIJ_inplace(fact,A,perm,info);CHKERRQ(ierr); 2819 PetscFunctionReturn(0); 2820 } 2821 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); 2822 /* check whether perm is the identity mapping */ 2823 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2824 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2825 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2826 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2827 2828 /* initialization */ 2829 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2830 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2831 ui[0] = 0; 2832 2833 /* jl: linked list for storing indices of the pivot rows 2834 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2835 ierr = PetscMalloc4(am,PetscInt*,&ui_ptr,am,PetscInt,&jl,am,PetscInt,&il,am,PetscInt,&cols);CHKERRQ(ierr); 2836 for (i=0; i<am; i++){ 2837 jl[i] = am; il[i] = 0; 2838 } 2839 2840 /* create and initialize a linked list for storing column indices of the active row k */ 2841 nlnk = am + 1; 2842 ierr = PetscLLCreate(am,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2843 2844 /* initial FreeSpace size is fill*(ai[am]+1) */ 2845 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2846 current_space = free_space; 2847 2848 for (k=0; k<am; k++){ /* for each active row k */ 2849 /* initialize lnk by the column indices of row rip[k] of A */ 2850 nzk = 0; 2851 ncols = ai[rip[k]+1] - ai[rip[k]]; 2852 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2853 ncols_upper = 0; 2854 for (j=0; j<ncols; j++){ 2855 i = riip[*(aj + ai[rip[k]] + j)]; 2856 if (i >= k){ /* only take upper triangular entry */ 2857 cols[ncols_upper] = i; 2858 ncols_upper++; 2859 } 2860 } 2861 ierr = PetscLLAdd(ncols_upper,cols,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2862 nzk += nlnk; 2863 2864 /* update lnk by computing fill-in for each pivot row to be merged in */ 2865 prow = jl[k]; /* 1st pivot row */ 2866 2867 while (prow < k){ 2868 nextprow = jl[prow]; 2869 /* merge prow into k-th row */ 2870 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2871 jmax = ui[prow+1]; 2872 ncols = jmax-jmin; 2873 uj_ptr = ui_ptr[prow] + jmin - ui[prow]; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2874 ierr = PetscLLAddSorted(ncols,uj_ptr,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2875 nzk += nlnk; 2876 2877 /* update il and jl for prow */ 2878 if (jmin < jmax){ 2879 il[prow] = jmin; 2880 j = *uj_ptr; jl[prow] = jl[j]; jl[j] = prow; 2881 } 2882 prow = nextprow; 2883 } 2884 2885 /* if free space is not available, make more free space */ 2886 if (current_space->local_remaining<nzk) { 2887 i = am - k + 1; /* num of unfactored rows */ 2888 i *= PetscMin(nzk,i-1); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2889 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2890 reallocs++; 2891 } 2892 2893 /* copy data into free space, then initialize lnk */ 2894 ierr = PetscLLClean(am,am,nzk,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 2895 2896 /* add the k-th row into il and jl */ 2897 if (nzk-1 > 0){ 2898 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2899 jl[k] = jl[i]; jl[i] = k; 2900 il[k] = ui[k] + 1; 2901 } 2902 ui_ptr[k] = current_space->array; 2903 current_space->array += nzk; 2904 current_space->local_used += nzk; 2905 current_space->local_remaining -= nzk; 2906 2907 ui[k+1] = ui[k] + nzk; 2908 } 2909 2910 #if defined(PETSC_USE_INFO) 2911 if (ai[am] != 0) { 2912 PetscReal af = (PetscReal)(ui[am])/((PetscReal)ai[am]); 2913 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2914 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2915 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2916 } else { 2917 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2918 } 2919 #endif 2920 2921 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2922 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2923 ierr = PetscFree4(ui_ptr,jl,il,cols);CHKERRQ(ierr); 2924 2925 /* destroy list of free space and other temporary array(s) */ 2926 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2927 ierr = PetscFreeSpaceContiguous_Cholesky(&free_space,uj,am,ui,udiag);CHKERRQ(ierr); /* store matrix factor in newdatastruct */ 2928 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2929 2930 /* put together the new matrix in MATSEQSBAIJ format */ 2931 2932 b = (Mat_SeqSBAIJ*)fact->data; 2933 b->singlemalloc = PETSC_FALSE; 2934 b->free_a = PETSC_TRUE; 2935 b->free_ij = PETSC_TRUE; 2936 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2937 b->j = uj; 2938 b->i = ui; 2939 b->diag = udiag; 2940 b->free_diag = PETSC_TRUE; 2941 b->ilen = 0; 2942 b->imax = 0; 2943 b->row = perm; 2944 b->col = perm; 2945 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2946 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2947 b->icol = iperm; 2948 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2949 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2950 ierr = PetscLogObjectMemory(fact,(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2951 b->maxnz = b->nz = ui[am]; 2952 2953 fact->info.factor_mallocs = reallocs; 2954 fact->info.fill_ratio_given = fill; 2955 if (ai[am] != 0) { 2956 fact->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2957 } else { 2958 fact->info.fill_ratio_needed = 0.0; 2959 } 2960 fact->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_newdatastruct; 2961 PetscFunctionReturn(0); 2962 } 2963 2964 #undef __FUNCT__ 2965 #define __FUNCT__ "MatCholeskyFactorSymbolic_SeqAIJ_inplace" 2966 PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ_inplace(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2967 { 2968 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2969 Mat_SeqSBAIJ *b; 2970 PetscErrorCode ierr; 2971 PetscTruth perm_identity; 2972 PetscReal fill = info->fill; 2973 const PetscInt *rip,*riip; 2974 PetscInt i,am=A->rmap->n,*ai=a->i,*aj=a->j,reallocs=0,prow; 2975 PetscInt *jl,jmin,jmax,nzk,*ui,k,j,*il,nextprow; 2976 PetscInt nlnk,*lnk,ncols,ncols_upper,*cols,*uj,**ui_ptr,*uj_ptr; 2977 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2978 PetscBT lnkbt; 2979 IS iperm; 2980 2981 PetscFunctionBegin; 2982 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); 2983 /* check whether perm is the identity mapping */ 2984 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2985 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2986 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2987 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2988 2989 /* initialization */ 2990 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2991 ui[0] = 0; 2992 2993 /* jl: linked list for storing indices of the pivot rows 2994 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2995 ierr = PetscMalloc4(am,PetscInt*,&ui_ptr,am,PetscInt,&jl,am,PetscInt,&il,am,PetscInt,&cols);CHKERRQ(ierr); 2996 for (i=0; i<am; i++){ 2997 jl[i] = am; il[i] = 0; 2998 } 2999 3000 /* create and initialize a linked list for storing column indices of the active row k */ 3001 nlnk = am + 1; 3002 ierr = PetscLLCreate(am,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3003 3004 /* initial FreeSpace size is fill*(ai[am]+1) */ 3005 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 3006 current_space = free_space; 3007 3008 for (k=0; k<am; k++){ /* for each active row k */ 3009 /* initialize lnk by the column indices of row rip[k] of A */ 3010 nzk = 0; 3011 ncols = ai[rip[k]+1] - ai[rip[k]]; 3012 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 3013 ncols_upper = 0; 3014 for (j=0; j<ncols; j++){ 3015 i = riip[*(aj + ai[rip[k]] + j)]; 3016 if (i >= k){ /* only take upper triangular entry */ 3017 cols[ncols_upper] = i; 3018 ncols_upper++; 3019 } 3020 } 3021 ierr = PetscLLAdd(ncols_upper,cols,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3022 nzk += nlnk; 3023 3024 /* update lnk by computing fill-in for each pivot row to be merged in */ 3025 prow = jl[k]; /* 1st pivot row */ 3026 3027 while (prow < k){ 3028 nextprow = jl[prow]; 3029 /* merge prow into k-th row */ 3030 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 3031 jmax = ui[prow+1]; 3032 ncols = jmax-jmin; 3033 uj_ptr = ui_ptr[prow] + jmin - ui[prow]; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 3034 ierr = PetscLLAddSorted(ncols,uj_ptr,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3035 nzk += nlnk; 3036 3037 /* update il and jl for prow */ 3038 if (jmin < jmax){ 3039 il[prow] = jmin; 3040 j = *uj_ptr; jl[prow] = jl[j]; jl[j] = prow; 3041 } 3042 prow = nextprow; 3043 } 3044 3045 /* if free space is not available, make more free space */ 3046 if (current_space->local_remaining<nzk) { 3047 i = am - k + 1; /* num of unfactored rows */ 3048 i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 3049 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 3050 reallocs++; 3051 } 3052 3053 /* copy data into free space, then initialize lnk */ 3054 ierr = PetscLLClean(am,am,nzk,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 3055 3056 /* add the k-th row into il and jl */ 3057 if (nzk-1 > 0){ 3058 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 3059 jl[k] = jl[i]; jl[i] = k; 3060 il[k] = ui[k] + 1; 3061 } 3062 ui_ptr[k] = current_space->array; 3063 current_space->array += nzk; 3064 current_space->local_used += nzk; 3065 current_space->local_remaining -= nzk; 3066 3067 ui[k+1] = ui[k] + nzk; 3068 } 3069 3070 #if defined(PETSC_USE_INFO) 3071 if (ai[am] != 0) { 3072 PetscReal af = (PetscReal)(ui[am])/((PetscReal)ai[am]); 3073 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 3074 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 3075 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 3076 } else { 3077 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 3078 } 3079 #endif 3080 3081 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 3082 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 3083 ierr = PetscFree4(ui_ptr,jl,il,cols);CHKERRQ(ierr); 3084 3085 /* destroy list of free space and other temporary array(s) */ 3086 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 3087 ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr); 3088 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 3089 3090 /* put together the new matrix in MATSEQSBAIJ format */ 3091 3092 b = (Mat_SeqSBAIJ*)fact->data; 3093 b->singlemalloc = PETSC_FALSE; 3094 b->free_a = PETSC_TRUE; 3095 b->free_ij = PETSC_TRUE; 3096 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 3097 b->j = uj; 3098 b->i = ui; 3099 b->diag = 0; 3100 b->ilen = 0; 3101 b->imax = 0; 3102 b->row = perm; 3103 b->col = perm; 3104 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 3105 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 3106 b->icol = iperm; 3107 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 3108 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 3109 ierr = PetscLogObjectMemory(fact,(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 3110 b->maxnz = b->nz = ui[am]; 3111 3112 fact->info.factor_mallocs = reallocs; 3113 fact->info.fill_ratio_given = fill; 3114 if (ai[am] != 0) { 3115 fact->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 3116 } else { 3117 fact->info.fill_ratio_needed = 0.0; 3118 } 3119 fact->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_inplace; 3120 PetscFunctionReturn(0); 3121 } 3122 3123 #undef __FUNCT__ 3124 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering_newdatastruct" 3125 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_newdatastruct(Mat A,Vec bb,Vec xx) 3126 { 3127 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3128 PetscErrorCode ierr; 3129 PetscInt n = A->rmap->n; 3130 const PetscInt *ai = a->i,*aj = a->j,*adiag = a->diag,*vi; 3131 PetscScalar *x,sum; 3132 const PetscScalar *b; 3133 const MatScalar *aa = a->a,*v; 3134 PetscInt i,nz; 3135 3136 PetscFunctionBegin; 3137 if (!n) PetscFunctionReturn(0); 3138 3139 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3140 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 3141 3142 /* forward solve the lower triangular */ 3143 x[0] = b[0]; 3144 v = aa; 3145 vi = aj; 3146 for (i=1; i<n; i++) { 3147 nz = ai[i+1] - ai[i]; 3148 sum = b[i]; 3149 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 3150 v += nz; 3151 vi += nz; 3152 x[i] = sum; 3153 } 3154 3155 /* backward solve the upper triangular */ 3156 for (i=n-1; i>=0; i--){ 3157 v = aa + adiag[i+1] + 1; 3158 vi = aj + adiag[i+1] + 1; 3159 nz = adiag[i] - adiag[i+1]-1; 3160 sum = x[i]; 3161 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 3162 x[i] = sum*v[nz]; /* x[i]=aa[adiag[i]]*sum; v++; */ 3163 } 3164 3165 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 3166 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3167 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 3168 PetscFunctionReturn(0); 3169 } 3170 3171 #undef __FUNCT__ 3172 #define __FUNCT__ "MatSolve_SeqAIJ_newdatastruct" 3173 PetscErrorCode MatSolve_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec xx) 3174 { 3175 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3176 IS iscol = a->col,isrow = a->row; 3177 PetscErrorCode ierr; 3178 PetscInt i,n=A->rmap->n,*vi,*ai=a->i,*aj=a->j,*adiag = a->diag,nz; 3179 const PetscInt *rout,*cout,*r,*c; 3180 PetscScalar *x,*tmp,sum; 3181 const PetscScalar *b; 3182 const MatScalar *aa = a->a,*v; 3183 3184 PetscFunctionBegin; 3185 if (!n) PetscFunctionReturn(0); 3186 3187 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3188 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 3189 tmp = a->solve_work; 3190 3191 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 3192 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 3193 3194 /* forward solve the lower triangular */ 3195 tmp[0] = b[r[0]]; 3196 v = aa; 3197 vi = aj; 3198 for (i=1; i<n; i++) { 3199 nz = ai[i+1] - ai[i]; 3200 sum = b[r[i]]; 3201 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 3202 tmp[i] = sum; 3203 v += nz; vi += nz; 3204 } 3205 3206 /* backward solve the upper triangular */ 3207 for (i=n-1; i>=0; i--){ 3208 v = aa + adiag[i+1]+1; 3209 vi = aj + adiag[i+1]+1; 3210 nz = adiag[i]-adiag[i+1]-1; 3211 sum = tmp[i]; 3212 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 3213 x[c[i]] = tmp[i] = sum*v[nz]; /* v[nz] = aa[adiag[i]] */ 3214 } 3215 3216 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 3217 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 3218 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3219 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 3220 ierr = PetscLogFlops(2*a->nz - A->cmap->n);CHKERRQ(ierr); 3221 PetscFunctionReturn(0); 3222 } 3223 3224 #undef __FUNCT__ 3225 #define __FUNCT__ "MatILUDTFactor_SeqAIJ" 3226 PetscErrorCode MatILUDTFactor_SeqAIJ(Mat A,IS isrow,IS iscol,const MatFactorInfo *info,Mat *fact) 3227 { 3228 Mat B = *fact; 3229 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b; 3230 IS isicol; 3231 PetscErrorCode ierr; 3232 const PetscInt *r,*ic; 3233 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j,*ajtmp,*adiag; 3234 PetscInt *bi,*bj,*bdiag,*bdiag_rev; 3235 PetscInt row,nzi,nzi_bl,nzi_bu,*im,nzi_al,nzi_au; 3236 PetscInt nlnk,*lnk; 3237 PetscBT lnkbt; 3238 PetscTruth row_identity,icol_identity,both_identity; 3239 MatScalar *aatmp,*pv,*batmp,*ba,*rtmp,*pc,multiplier,*vtmp,diag_tmp; 3240 const PetscInt *ics; 3241 PetscInt j,nz,*pj,*bjtmp,k,ncut,*jtmp; 3242 PetscReal dt=info->dt,dtcol=info->dtcol,shift=info->shiftinblocks; 3243 PetscInt dtcount=(PetscInt)info->dtcount,nnz_max; 3244 PetscTruth missing; 3245 3246 PetscFunctionBegin; 3247 3248 if (dt == PETSC_DEFAULT) dt = 0.005; 3249 if (dtcol == PETSC_DEFAULT) dtcol = 0.01; /* XXX unused! */ 3250 if (dtcount == PETSC_DEFAULT) dtcount = (PetscInt)(1.5*a->rmax); 3251 3252 /* ------- symbolic factorization, can be reused ---------*/ 3253 ierr = MatMissingDiagonal(A,&missing,&i);CHKERRQ(ierr); 3254 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",i); 3255 adiag=a->diag; 3256 3257 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 3258 3259 /* bdiag is location of diagonal in factor */ 3260 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); /* becomes b->diag */ 3261 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag_rev);CHKERRQ(ierr); /* temporary */ 3262 3263 /* allocate row pointers bi */ 3264 ierr = PetscMalloc((2*n+2)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 3265 3266 /* allocate bj and ba; max num of nonzero entries is (ai[n]+2*n*dtcount+2) */ 3267 if (dtcount > n-1) dtcount = n-1; /* diagonal is excluded */ 3268 nnz_max = ai[n]+2*n*dtcount+2; 3269 3270 ierr = PetscMalloc((nnz_max+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 3271 ierr = PetscMalloc((nnz_max+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 3272 3273 /* put together the new matrix */ 3274 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 3275 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 3276 b = (Mat_SeqAIJ*)B->data; 3277 b->free_a = PETSC_TRUE; 3278 b->free_ij = PETSC_TRUE; 3279 b->singlemalloc = PETSC_FALSE; 3280 b->a = ba; 3281 b->j = bj; 3282 b->i = bi; 3283 b->diag = bdiag; 3284 b->ilen = 0; 3285 b->imax = 0; 3286 b->row = isrow; 3287 b->col = iscol; 3288 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 3289 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 3290 b->icol = isicol; 3291 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 3292 3293 ierr = PetscLogObjectMemory(B,nnz_max*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 3294 b->maxnz = nnz_max; 3295 3296 B->factor = MAT_FACTOR_ILUDT; 3297 B->info.factor_mallocs = 0; 3298 B->info.fill_ratio_given = ((PetscReal)nnz_max)/((PetscReal)ai[n]); 3299 CHKMEMQ; 3300 /* ------- end of symbolic factorization ---------*/ 3301 3302 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 3303 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 3304 ics = ic; 3305 3306 /* linked list for storing column indices of the active row */ 3307 nlnk = n + 1; 3308 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3309 3310 /* im: used by PetscLLAddSortedLU(); jtmp: working array for column indices of active row */ 3311 ierr = PetscMalloc2(n,PetscInt,&im,n,PetscInt,&jtmp);CHKERRQ(ierr); 3312 /* rtmp, vtmp: working arrays for sparse and contiguous row entries of active row */ 3313 ierr = PetscMalloc2(n,MatScalar,&rtmp,n,MatScalar,&vtmp);CHKERRQ(ierr); 3314 ierr = PetscMemzero(rtmp,n*sizeof(MatScalar));CHKERRQ(ierr); 3315 3316 bi[0] = 0; 3317 bdiag[0] = nnz_max-1; /* location of diag[0] in factor B */ 3318 bdiag_rev[n] = bdiag[0]; 3319 bi[2*n+1] = bdiag[0]+1; /* endof bj and ba array */ 3320 for (i=0; i<n; i++) { 3321 /* copy initial fill into linked list */ 3322 nzi = 0; /* nonzeros for active row i */ 3323 nzi = ai[r[i]+1] - ai[r[i]]; 3324 if (!nzi) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 3325 nzi_al = adiag[r[i]] - ai[r[i]]; 3326 nzi_au = ai[r[i]+1] - adiag[r[i]] -1; 3327 ajtmp = aj + ai[r[i]]; 3328 ierr = PetscLLAddPerm(nzi,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3329 3330 /* load in initial (unfactored row) */ 3331 aatmp = a->a + ai[r[i]]; 3332 for (j=0; j<nzi; j++) { 3333 rtmp[ics[*ajtmp++]] = *aatmp++; 3334 } 3335 3336 /* add pivot rows into linked list */ 3337 row = lnk[n]; 3338 while (row < i ) { 3339 nzi_bl = bi[row+1] - bi[row] + 1; 3340 bjtmp = bj + bdiag[row+1]+1; /* points to 1st column next to the diagonal in U */ 3341 ierr = PetscLLAddSortedLU(bjtmp,row,nlnk,lnk,lnkbt,i,nzi_bl,im);CHKERRQ(ierr); 3342 nzi += nlnk; 3343 row = lnk[row]; 3344 } 3345 3346 /* copy data from lnk into jtmp, then initialize lnk */ 3347 ierr = PetscLLClean(n,n,nzi,lnk,jtmp,lnkbt);CHKERRQ(ierr); 3348 3349 /* numerical factorization */ 3350 bjtmp = jtmp; 3351 row = *bjtmp++; /* 1st pivot row */ 3352 while ( row < i ) { 3353 pc = rtmp + row; 3354 pv = ba + bdiag[row]; /* 1./(diag of the pivot row) */ 3355 multiplier = (*pc) * (*pv); 3356 *pc = multiplier; 3357 if (PetscAbsScalar(*pc) > dt){ /* apply tolerance dropping rule */ 3358 pj = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */ 3359 pv = ba + bdiag[row+1] + 1; 3360 /* if (multiplier < -1.0 or multiplier >1.0) printf("row/prow %d, %d, multiplier %g\n",i,row,multiplier); */ 3361 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */ 3362 for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++); 3363 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 3364 } 3365 row = *bjtmp++; 3366 } 3367 3368 /* copy sparse rtmp into contiguous vtmp; separate L and U part */ 3369 diag_tmp = rtmp[i]; /* save diagonal value - may not needed?? */ 3370 nzi_bl = 0; j = 0; 3371 while (jtmp[j] < i){ /* Note: jtmp is sorted */ 3372 vtmp[j] = rtmp[jtmp[j]]; rtmp[jtmp[j]]=0.0; 3373 nzi_bl++; j++; 3374 } 3375 nzi_bu = nzi - nzi_bl -1; 3376 while (j < nzi){ 3377 vtmp[j] = rtmp[jtmp[j]]; rtmp[jtmp[j]]=0.0; 3378 j++; 3379 } 3380 3381 bjtmp = bj + bi[i]; 3382 batmp = ba + bi[i]; 3383 /* apply level dropping rule to L part */ 3384 ncut = nzi_al + dtcount; 3385 if (ncut < nzi_bl){ 3386 ierr = PetscSortSplit(ncut,nzi_bl,vtmp,jtmp);CHKERRQ(ierr); 3387 ierr = PetscSortIntWithScalarArray(ncut,jtmp,vtmp);CHKERRQ(ierr); 3388 } else { 3389 ncut = nzi_bl; 3390 } 3391 for (j=0; j<ncut; j++){ 3392 bjtmp[j] = jtmp[j]; 3393 batmp[j] = vtmp[j]; 3394 /* printf(" (%d,%g),",bjtmp[j],batmp[j]); */ 3395 } 3396 bi[i+1] = bi[i] + ncut; 3397 nzi = ncut + 1; 3398 3399 /* apply level dropping rule to U part */ 3400 ncut = nzi_au + dtcount; 3401 if (ncut < nzi_bu){ 3402 ierr = PetscSortSplit(ncut,nzi_bu,vtmp+nzi_bl+1,jtmp+nzi_bl+1);CHKERRQ(ierr); 3403 ierr = PetscSortIntWithScalarArray(ncut,jtmp+nzi_bl+1,vtmp+nzi_bl+1);CHKERRQ(ierr); 3404 } else { 3405 ncut = nzi_bu; 3406 } 3407 nzi += ncut; 3408 3409 /* mark bdiagonal */ 3410 bdiag[i+1] = bdiag[i] - (ncut + 1); 3411 bdiag_rev[n-i-1] = bdiag[i+1]; 3412 bi[2*n - i] = bi[2*n - i +1] - (ncut + 1); 3413 bjtmp = bj + bdiag[i]; 3414 batmp = ba + bdiag[i]; 3415 *bjtmp = i; 3416 *batmp = diag_tmp; /* rtmp[i]; */ 3417 if (*batmp == 0.0) { 3418 *batmp = dt+shift; 3419 /* printf(" row %d add shift %g\n",i,shift); */ 3420 } 3421 *batmp = 1.0/(*batmp); /* invert diagonal entries for simplier triangular solves */ 3422 /* printf(" (%d,%g),",*bjtmp,*batmp); */ 3423 3424 bjtmp = bj + bdiag[i+1]+1; 3425 batmp = ba + bdiag[i+1]+1; 3426 for (k=0; k<ncut; k++){ 3427 bjtmp[k] = jtmp[nzi_bl+1+k]; 3428 batmp[k] = vtmp[nzi_bl+1+k]; 3429 /* printf(" (%d,%g),",bjtmp[k],batmp[k]); */ 3430 } 3431 /* printf("\n"); */ 3432 3433 im[i] = nzi; /* used by PetscLLAddSortedLU() */ 3434 /* 3435 printf("row %d: bi %d, bdiag %d\n",i,bi[i],bdiag[i]); 3436 printf(" ----------------------------\n"); 3437 */ 3438 } /* for (i=0; i<n; i++) */ 3439 /* printf("end of L %d, beginning of U %d\n",bi[n],bdiag[n]); */ 3440 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]); 3441 3442 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 3443 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 3444 3445 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 3446 ierr = PetscFree2(im,jtmp);CHKERRQ(ierr); 3447 ierr = PetscFree2(rtmp,vtmp);CHKERRQ(ierr); 3448 ierr = PetscFree(bdiag_rev);CHKERRQ(ierr); 3449 3450 ierr = PetscLogFlops(B->cmap->n);CHKERRQ(ierr); 3451 b->maxnz = b->nz = bi[n] + bdiag[0] - bdiag[n]; 3452 3453 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 3454 ierr = ISIdentity(isicol,&icol_identity);CHKERRQ(ierr); 3455 both_identity = (PetscTruth) (row_identity && icol_identity); 3456 if (row_identity && icol_identity) { 3457 B->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 3458 } else { 3459 B->ops->solve = MatSolve_SeqAIJ_newdatastruct; 3460 } 3461 3462 B->ops->lufactorsymbolic = MatILUDTFactorSymbolic_SeqAIJ; 3463 B->ops->lufactornumeric = MatILUDTFactorNumeric_SeqAIJ; 3464 B->ops->solveadd = 0; 3465 B->ops->solvetranspose = 0; 3466 B->ops->solvetransposeadd = 0; 3467 B->ops->matsolve = 0; 3468 B->assembled = PETSC_TRUE; 3469 B->preallocated = PETSC_TRUE; 3470 PetscFunctionReturn(0); 3471 } 3472 3473 /* a wraper of MatILUDTFactor_SeqAIJ() */ 3474 #undef __FUNCT__ 3475 #define __FUNCT__ "MatILUDTFactorSymbolic_SeqAIJ" 3476 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS row,IS col,const MatFactorInfo *info) 3477 { 3478 PetscErrorCode ierr; 3479 3480 PetscFunctionBegin; 3481 ierr = MatILUDTFactor_SeqAIJ(A,row,col,info,&fact);CHKERRQ(ierr); 3482 3483 fact->ops->lufactornumeric = MatILUDTFactorNumeric_SeqAIJ; 3484 PetscFunctionReturn(0); 3485 } 3486 3487 /* 3488 same as MatLUFactorNumeric_SeqAIJ(), except using contiguous array matrix factors 3489 - intend to replace existing MatLUFactorNumeric_SeqAIJ() 3490 */ 3491 #undef __FUNCT__ 3492 #define __FUNCT__ "MatILUDTFactorNumeric_SeqAIJ" 3493 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorNumeric_SeqAIJ(Mat fact,Mat A,const MatFactorInfo *info) 3494 { 3495 Mat C=fact; 3496 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 3497 IS isrow = b->row,isicol = b->icol; 3498 PetscErrorCode ierr; 3499 const PetscInt *r,*ic,*ics; 3500 PetscInt i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 3501 PetscInt *ajtmp,*bjtmp,nz,nzl,nzu,row,*bdiag = b->diag,*pj; 3502 MatScalar *rtmp,*pc,multiplier,*v,*pv,*aa=a->a; 3503 PetscReal dt=info->dt,shift=info->shiftinblocks; 3504 PetscTruth row_identity, col_identity; 3505 3506 PetscFunctionBegin; 3507 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 3508 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 3509 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 3510 ics = ic; 3511 3512 for (i=0; i<n; i++){ 3513 /* initialize rtmp array */ 3514 nzl = bi[i+1] - bi[i]; /* num of nozeros in L(i,:) */ 3515 bjtmp = bj + bi[i]; 3516 for (j=0; j<nzl; j++) rtmp[*bjtmp++] = 0.0; 3517 rtmp[i] = 0.0; 3518 nzu = bdiag[i] - bdiag[i+1]; /* num of nozeros in U(i,:) */ 3519 bjtmp = bj + bdiag[i+1] + 1; 3520 for (j=0; j<nzu; j++) rtmp[*bjtmp++] = 0.0; 3521 3522 /* load in initial unfactored row of A */ 3523 /* printf("row %d\n",i); */ 3524 nz = ai[r[i]+1] - ai[r[i]]; 3525 ajtmp = aj + ai[r[i]]; 3526 v = aa + ai[r[i]]; 3527 for (j=0; j<nz; j++) { 3528 rtmp[ics[*ajtmp++]] = v[j]; 3529 /* printf(" (%d,%g),",ics[ajtmp[j]],rtmp[ics[ajtmp[j]]]); */ 3530 } 3531 /* printf("\n"); */ 3532 3533 /* numerical factorization */ 3534 bjtmp = bj + bi[i]; /* point to 1st entry of L(i,:) */ 3535 nzl = bi[i+1] - bi[i]; /* num of entries in L(i,:) */ 3536 k = 0; 3537 while (k < nzl){ 3538 row = *bjtmp++; 3539 /* printf(" prow %d\n",row); */ 3540 pc = rtmp + row; 3541 pv = b->a + bdiag[row]; /* 1./(diag of the pivot row) */ 3542 multiplier = (*pc) * (*pv); 3543 *pc = multiplier; 3544 if (PetscAbsScalar(multiplier) > dt){ 3545 pj = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */ 3546 pv = b->a + bdiag[row+1] + 1; 3547 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */ 3548 for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++); 3549 /* ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); */ 3550 } 3551 k++; 3552 } 3553 3554 /* finished row so stick it into b->a */ 3555 /* L-part */ 3556 pv = b->a + bi[i] ; 3557 pj = bj + bi[i] ; 3558 nzl = bi[i+1] - bi[i]; 3559 for (j=0; j<nzl; j++) { 3560 pv[j] = rtmp[pj[j]]; 3561 /* printf(" (%d,%g),",pj[j],pv[j]); */ 3562 } 3563 3564 /* diagonal: invert diagonal entries for simplier triangular solves */ 3565 if (rtmp[i] == 0.0) rtmp[i] = dt+shift; 3566 b->a[bdiag[i]] = 1.0/rtmp[i]; 3567 /* printf(" (%d,%g),",i,b->a[bdiag[i]]); */ 3568 3569 /* U-part */ 3570 pv = b->a + bdiag[i+1] + 1; 3571 pj = bj + bdiag[i+1] + 1; 3572 nzu = bdiag[i] - bdiag[i+1] - 1; 3573 for (j=0; j<nzu; j++) { 3574 pv[j] = rtmp[pj[j]]; 3575 /* printf(" (%d,%g),",pj[j],pv[j]); */ 3576 } 3577 /* printf("\n"); */ 3578 } 3579 3580 ierr = PetscFree(rtmp);CHKERRQ(ierr); 3581 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 3582 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 3583 3584 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 3585 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 3586 if (row_identity && col_identity) { 3587 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 3588 } else { 3589 C->ops->solve = MatSolve_SeqAIJ_newdatastruct; 3590 } 3591 C->ops->solveadd = 0; 3592 C->ops->solvetranspose = 0; 3593 C->ops->solvetransposeadd = 0; 3594 C->ops->matsolve = 0; 3595 C->assembled = PETSC_TRUE; 3596 C->preallocated = PETSC_TRUE; 3597 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 3598 PetscFunctionReturn(0); 3599 } 3600