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; 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; 69 past = 0; 70 for (j=ai[k]; j<ai[k+1]; j++) { 71 kk = aj[j]; 72 if (done[kk]) past += PetscAbsScalar(aa[j]); 73 else if (kk != k) future += PetscAbsScalar(aa[j]); 74 } 75 if (!future) future = 1.e-10; /* if there is zero in the upper diagonal part want to rank this row high */ 76 if (past/future > best) { 77 best = past/future; 78 newcurrent = k; 79 } 80 } 81 } 82 if (current == newcurrent) SETERRQ(PETSC_ERR_PLIB,"newcurrent cannot be current"); 83 current = newcurrent; 84 order[i+1] = current; 85 } 86 ierr = ISCreateGeneral(PETSC_COMM_SELF,n,order,irow);CHKERRQ(ierr); 87 *icol = *irow; 88 ierr = PetscObjectReference((PetscObject)*irow);CHKERRQ(ierr); 89 ierr = PetscFree(done);CHKERRQ(ierr); 90 ierr = PetscFree(order);CHKERRQ(ierr); 91 PetscFunctionReturn(0); 92 } 93 EXTERN_C_END 94 95 EXTERN_C_BEGIN 96 #undef __FUNCT__ 97 #define __FUNCT__ "MatGetFactorAvailable_seqaij_petsc" 98 PetscErrorCode MatGetFactorAvailable_seqaij_petsc(Mat A,MatFactorType ftype,PetscTruth *flg) 99 { 100 PetscFunctionBegin; 101 *flg = PETSC_TRUE; 102 PetscFunctionReturn(0); 103 } 104 EXTERN_C_END 105 106 EXTERN_C_BEGIN 107 #undef __FUNCT__ 108 #define __FUNCT__ "MatGetFactor_seqaij_petsc" 109 PetscErrorCode MatGetFactor_seqaij_petsc(Mat A,MatFactorType ftype,Mat *B) 110 { 111 PetscInt n = A->rmap->n; 112 PetscErrorCode ierr; 113 114 PetscFunctionBegin; 115 ierr = MatCreate(((PetscObject)A)->comm,B);CHKERRQ(ierr); 116 ierr = MatSetSizes(*B,n,n,n,n);CHKERRQ(ierr); 117 if (ftype == MAT_FACTOR_LU || ftype == MAT_FACTOR_ILU || ftype == MAT_FACTOR_ILUDT){ 118 ierr = MatSetType(*B,MATSEQAIJ);CHKERRQ(ierr); 119 (*B)->ops->ilufactorsymbolic = MatILUFactorSymbolic_SeqAIJ; 120 (*B)->ops->lufactorsymbolic = MatLUFactorSymbolic_SeqAIJ; 121 (*B)->ops->iludtfactor = MatILUDTFactor_SeqAIJ; 122 } else if (ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC) { 123 ierr = MatSetType(*B,MATSEQSBAIJ);CHKERRQ(ierr); 124 ierr = MatSeqSBAIJSetPreallocation(*B,1,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 125 (*B)->ops->iccfactorsymbolic = MatICCFactorSymbolic_SeqAIJ; 126 (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJ; 127 } else SETERRQ(PETSC_ERR_SUP,"Factor type not supported"); 128 (*B)->factor = ftype; 129 PetscFunctionReturn(0); 130 } 131 EXTERN_C_END 132 133 #undef __FUNCT__ 134 #define __FUNCT__ "MatLUFactorSymbolic_SeqAIJ" 135 PetscErrorCode MatLUFactorSymbolic_SeqAIJ(Mat B,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 136 { 137 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 138 IS isicol; 139 PetscErrorCode ierr; 140 const PetscInt *r,*ic; 141 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j; 142 PetscInt *bi,*bj,*ajtmp; 143 PetscInt *bdiag,row,nnz,nzi,reallocs=0,nzbd,*im; 144 PetscReal f; 145 PetscInt nlnk,*lnk,k,**bi_ptr; 146 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 147 PetscBT lnkbt; 148 PetscTruth newdatastruct= PETSC_FALSE; 149 150 PetscFunctionBegin; 151 ierr = PetscOptionsGetTruth(PETSC_NULL,"-lu_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 152 if(newdatastruct){ 153 ierr = MatLUFactorSymbolic_SeqAIJ_newdatastruct(B,A,isrow,iscol,info);CHKERRQ(ierr); 154 PetscFunctionReturn(0); 155 } 156 157 158 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 159 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 160 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 161 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 162 163 /* get new row pointers */ 164 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 165 bi[0] = 0; 166 167 /* bdiag is location of diagonal in factor */ 168 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 169 bdiag[0] = 0; 170 171 /* linked list for storing column indices of the active row */ 172 nlnk = n + 1; 173 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 174 175 ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr); 176 177 /* initial FreeSpace size is f*(ai[n]+1) */ 178 f = info->fill; 179 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 180 current_space = free_space; 181 182 for (i=0; i<n; i++) { 183 /* copy previous fill into linked list */ 184 nzi = 0; 185 nnz = ai[r[i]+1] - ai[r[i]]; 186 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 187 ajtmp = aj + ai[r[i]]; 188 ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 189 nzi += nlnk; 190 191 /* add pivot rows into linked list */ 192 row = lnk[n]; 193 while (row < i) { 194 nzbd = bdiag[row] - bi[row] + 1; /* num of entries in the row with column index <= row */ 195 ajtmp = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */ 196 ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr); 197 nzi += nlnk; 198 row = lnk[row]; 199 } 200 bi[i+1] = bi[i] + nzi; 201 im[i] = nzi; 202 203 /* mark bdiag */ 204 nzbd = 0; 205 nnz = nzi; 206 k = lnk[n]; 207 while (nnz-- && k < i){ 208 nzbd++; 209 k = lnk[k]; 210 } 211 bdiag[i] = bi[i] + nzbd; 212 213 /* if free space is not available, make more free space */ 214 if (current_space->local_remaining<nzi) { 215 nnz = (n - i)*nzi; /* estimated and max additional space needed */ 216 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 217 reallocs++; 218 } 219 220 /* copy data into free space, then initialize lnk */ 221 ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 222 bi_ptr[i] = current_space->array; 223 current_space->array += nzi; 224 current_space->local_used += nzi; 225 current_space->local_remaining -= nzi; 226 } 227 #if defined(PETSC_USE_INFO) 228 if (ai[n] != 0) { 229 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 230 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 231 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 232 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr); 233 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 234 } else { 235 ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr); 236 } 237 #endif 238 239 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 240 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 241 242 /* destroy list of free space and other temporary array(s) */ 243 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 244 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 245 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 246 ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr); 247 248 /* put together the new matrix */ 249 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 250 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 251 b = (Mat_SeqAIJ*)(B)->data; 252 b->free_a = PETSC_TRUE; 253 b->free_ij = PETSC_TRUE; 254 b->singlemalloc = PETSC_FALSE; 255 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 256 b->j = bj; 257 b->i = bi; 258 b->diag = bdiag; 259 b->ilen = 0; 260 b->imax = 0; 261 b->row = isrow; 262 b->col = iscol; 263 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 264 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 265 b->icol = isicol; 266 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 267 268 /* In b structure: Free imax, ilen, old a, old j. Allocate solve_work, new a, new j */ 269 ierr = PetscLogObjectMemory(B,(bi[n]-n)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 270 b->maxnz = b->nz = bi[n] ; 271 272 (B)->factor = MAT_FACTOR_LU; 273 (B)->info.factor_mallocs = reallocs; 274 (B)->info.fill_ratio_given = f; 275 276 if (ai[n]) { 277 (B)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]); 278 } else { 279 (B)->info.fill_ratio_needed = 0.0; 280 } 281 (B)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ; 282 (B)->ops->solve = MatSolve_SeqAIJ; 283 (B)->ops->solvetranspose = MatSolveTranspose_SeqAIJ; 284 /* switch to inodes if appropriate */ 285 ierr = MatLUFactorSymbolic_SeqAIJ_Inode(B,A,isrow,iscol,info);CHKERRQ(ierr); 286 PetscFunctionReturn(0); 287 } 288 289 #undef __FUNCT__ 290 #define __FUNCT__ "MatLUFactorSymbolic_SeqAIJ_newdatastruct" 291 PetscErrorCode MatLUFactorSymbolic_SeqAIJ_newdatastruct(Mat B,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 292 { 293 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 294 IS isicol; 295 PetscErrorCode ierr; 296 const PetscInt *r,*ic; 297 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j; 298 PetscInt *bi,*bj,*ajtmp; 299 PetscInt *bdiag,row,nnz,nzi,reallocs=0,nzbd,*im; 300 PetscReal f; 301 PetscInt nlnk,*lnk,k,**bi_ptr; 302 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 303 PetscBT lnkbt; 304 305 PetscFunctionBegin; 306 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 307 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 308 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 309 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 310 311 /* get new row and diagonal pointers */ 312 ierr = PetscMalloc2(n+1,PetscInt,&bi,n+1,PetscInt,&bdiag);CHKERRQ(ierr); 313 bi[0] = bdiag[0] = 0; 314 315 /* linked list for storing column indices of the active row */ 316 nlnk = n + 1; 317 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 318 319 ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr); 320 321 /* initial FreeSpace size is f*(ai[n]+1) */ 322 f = info->fill; 323 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 324 current_space = free_space; 325 326 for (i=0; i<n; i++) { 327 /* copy previous fill into linked list */ 328 nzi = 0; 329 nnz = ai[r[i]+1] - ai[r[i]]; 330 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 331 ajtmp = aj + ai[r[i]]; 332 ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 333 nzi += nlnk; 334 335 /* add pivot rows into linked list */ 336 row = lnk[n]; 337 while (row < i){ 338 nzbd = bdiag[row] + 1; /* num of entries in the row with column index <= row */ 339 ajtmp = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */ 340 ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr); 341 nzi += nlnk; 342 row = lnk[row]; 343 } 344 bi[i+1] = bi[i] + nzi; 345 im[i] = nzi; 346 347 /* mark bdiag */ 348 nzbd = 0; 349 nnz = nzi; 350 k = lnk[n]; 351 while (nnz-- && k < i){ 352 nzbd++; 353 k = lnk[k]; 354 } 355 bdiag[i] = nzbd; /* note: bdiag[i] = nnzL as input for PetscFreeSpaceContiguous_LU() */ 356 357 /* if free space is not available, make more free space */ 358 if (current_space->local_remaining<nzi) { 359 nnz = 2*(n - i)*nzi; /* estimated and max additional space needed */ 360 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 361 reallocs++; 362 } 363 364 /* copy data into free space, then initialize lnk */ 365 ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 366 bi_ptr[i] = current_space->array; 367 current_space->array += nzi; 368 current_space->local_used += nzi; 369 current_space->local_remaining -= nzi; 370 } 371 #if defined(PETSC_USE_INFO) 372 if (ai[n] != 0) { 373 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 374 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 375 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 376 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr); 377 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 378 } else { 379 ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr); 380 } 381 #endif 382 383 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 384 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 385 386 /* destroy list of free space and other temporary array(s) */ 387 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 388 ierr = PetscFreeSpaceContiguous_LU_v2(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 389 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 390 ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr); 391 392 /* put together the new matrix */ 393 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 394 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 395 b = (Mat_SeqAIJ*)(B)->data; 396 b->free_a = PETSC_TRUE; 397 b->free_ij = PETSC_TRUE; 398 b->singlemalloc = PETSC_FALSE; 399 ierr = PetscMalloc((bdiag[0]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 400 b->j = bj; 401 b->i = bi; 402 b->diag = bdiag; 403 b->ilen = 0; 404 b->imax = 0; 405 b->row = isrow; 406 b->col = iscol; 407 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 408 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 409 b->icol = isicol; 410 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 411 412 /* In b structure: Free imax, ilen, old a, old j. Allocate solve_work, new a, new j */ 413 ierr = PetscLogObjectMemory(B,(bdiag[0]+1)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 414 b->maxnz = b->nz = bdiag[0]+1; 415 (B)->factor = MAT_FACTOR_LU; 416 (B)->info.factor_mallocs = reallocs; 417 (B)->info.fill_ratio_given = f; 418 419 if (ai[n]) { 420 (B)->info.fill_ratio_needed = ((PetscReal)(bdiag[0]+1))/((PetscReal)ai[n]); 421 } else { 422 (B)->info.fill_ratio_needed = 0.0; 423 } 424 (B)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 425 PetscFunctionReturn(0); 426 } 427 428 /* 429 Trouble in factorization, should we dump the original matrix? 430 */ 431 #undef __FUNCT__ 432 #define __FUNCT__ "MatFactorDumpMatrix" 433 PetscErrorCode MatFactorDumpMatrix(Mat A) 434 { 435 PetscErrorCode ierr; 436 PetscTruth flg = PETSC_FALSE; 437 438 PetscFunctionBegin; 439 ierr = PetscOptionsGetTruth(PETSC_NULL,"-mat_factor_dump_on_error",&flg,PETSC_NULL);CHKERRQ(ierr); 440 if (flg) { 441 PetscViewer viewer; 442 char filename[PETSC_MAX_PATH_LEN]; 443 444 ierr = PetscSNPrintf(filename,PETSC_MAX_PATH_LEN,"matrix_factor_error.%d",PetscGlobalRank);CHKERRQ(ierr); 445 ierr = PetscViewerBinaryOpen(((PetscObject)A)->comm,filename,FILE_MODE_WRITE,&viewer);CHKERRQ(ierr); 446 ierr = MatView(A,viewer);CHKERRQ(ierr); 447 ierr = PetscViewerDestroy(viewer);CHKERRQ(ierr); 448 } 449 PetscFunctionReturn(0); 450 } 451 452 extern PetscErrorCode MatSolve_SeqAIJ_Inode(Mat,Vec,Vec); 453 454 #undef __FUNCT__ 455 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_newdatastruct" 456 PetscErrorCode MatLUFactorNumeric_SeqAIJ_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 457 { 458 Mat C=B; 459 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 460 IS isrow = b->row,isicol = b->icol; 461 PetscErrorCode ierr; 462 const PetscInt *r,*ic,*ics; 463 PetscInt i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 464 PetscInt *ajtmp,*bjtmp,nz,nzL,row,*bdiag=b->diag,*pj; 465 MatScalar *rtmp,*pc,multiplier,*v,*pv,*aa=a->a; 466 PetscTruth row_identity,col_identity; 467 468 LUShift_Ctx sctx; 469 PetscInt *ddiag,newshift; 470 PetscReal rs; 471 MatScalar d; 472 473 PetscFunctionBegin; 474 /* MatPivotSetUp(): initialize shift context sctx */ 475 ierr = PetscMemzero(&sctx,sizeof(LUShift_Ctx));CHKERRQ(ierr); 476 477 /* if both shift schemes are chosen by user, only use info->shiftpd */ 478 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 479 ddiag = a->diag; 480 sctx.shift_top = info->zeropivot; 481 for (i=0; i<n; i++) { 482 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 483 d = (aa)[ddiag[i]]; 484 rs = -PetscAbsScalar(d) - PetscRealPart(d); 485 v = aa+ai[i]; 486 nz = ai[i+1] - ai[i]; 487 for (j=0; j<nz; j++) 488 rs += PetscAbsScalar(v[j]); 489 if (rs>sctx.shift_top) sctx.shift_top = rs; 490 } 491 sctx.shift_top *= 1.1; 492 sctx.nshift_max = 5; 493 sctx.shift_lo = 0.; 494 sctx.shift_hi = 1.; 495 } 496 497 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 498 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 499 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 500 ics = ic; 501 502 do { 503 sctx.lushift = PETSC_FALSE; 504 for (i=0; i<n; i++){ 505 /* zero rtmp */ 506 /* L part */ 507 nz = bi[i+1] - bi[i]; 508 bjtmp = bj + bi[i]; 509 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 510 511 /* U part */ 512 nz = bdiag[i]-bdiag[i+1]; 513 bjtmp = bj + bdiag[i+1]+1; 514 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 515 516 /* load in initial (unfactored row) */ 517 nz = ai[r[i]+1] - ai[r[i]]; 518 ajtmp = aj + ai[r[i]]; 519 v = aa + ai[r[i]]; 520 for (j=0; j<nz; j++) { 521 rtmp[ics[ajtmp[j]]] = v[j]; 522 } 523 /* ZeropivotApply() */ 524 rtmp[ics[r[i]]] += sctx.shift_amount; /* shift the diagonal of the matrix */ 525 526 /* elimination */ 527 bjtmp = bj + bi[i]; 528 row = *bjtmp++; 529 nzL = bi[i+1] - bi[i]; 530 for(k=0; k < nzL;k++) { 531 pc = rtmp + row; 532 if (*pc != 0.0) { 533 pv = b->a + bdiag[row]; 534 multiplier = *pc * (*pv); 535 *pc = multiplier; 536 pj = b->j + bdiag[row+1]+1; /* beginning of U(row,:) */ 537 pv = b->a + bdiag[row+1]+1; 538 nz = bdiag[row]-bdiag[row+1]-1; /* num of entries in U(row,:) excluding diag */ 539 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 540 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 541 } 542 row = *bjtmp++; 543 } 544 545 /* finished row so stick it into b->a */ 546 rs = 0.0; 547 /* L part */ 548 pv = b->a + bi[i] ; 549 pj = b->j + bi[i] ; 550 nz = bi[i+1] - bi[i]; 551 for (j=0; j<nz; j++) { 552 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 553 } 554 555 /* U part */ 556 pv = b->a + bdiag[i+1]+1; 557 pj = b->j + bdiag[i+1]+1; 558 nz = bdiag[i] - bdiag[i+1]-1; 559 for (j=0; j<nz; j++) { 560 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 561 } 562 563 /* MatPivotCheck() */ 564 sctx.rs = rs; 565 sctx.pv = rtmp[i]; 566 /* ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); */ 567 if (info->shiftnz){ 568 ierr = MatPivotCheck_nz(info,sctx,i,newshift);CHKERRQ(ierr); 569 } else if (info->shiftpd){ 570 ierr = MatPivotCheck_pd(info,sctx,i,newshift);CHKERRQ(ierr); 571 } else if (info->shiftinblocks){ 572 ierr = MatPivotCheck_inblocks(info,sctx,i,newshift);CHKERRQ(ierr); 573 } else { 574 ierr = MatPivotCheck_none(info,sctx,i,newshift);CHKERRQ(ierr); 575 } 576 rtmp[i] = sctx.pv; 577 if (newshift == 1) break; 578 579 /* Mark diagonal and invert diagonal for simplier triangular solves */ 580 pv = b->a + bdiag[i]; 581 *pv = 1.0/rtmp[i]; 582 583 } /* endof for (i=0; i<n; i++){ */ 584 585 /* MatPivotRefine() */ 586 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max){ 587 /* 588 * if no shift in this attempt & shifting & started shifting & can refine, 589 * then try lower shift 590 */ 591 sctx.shift_hi = sctx.shift_fraction; 592 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 593 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 594 sctx.lushift = PETSC_TRUE; 595 sctx.nshift++; 596 } 597 } while (sctx.lushift); 598 599 ierr = PetscFree(rtmp);CHKERRQ(ierr); 600 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 601 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 602 603 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_v2; 607 } else { 608 C->ops->solve = MatSolve_SeqAIJ_newdatastruct_v2; 609 } 610 611 C->ops->solveadd = 0; 612 C->ops->solvetranspose = 0; 613 C->ops->solvetransposeadd = 0; 614 C->ops->matsolve = 0; 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" 634 PetscErrorCode MatLUFactorNumeric_SeqAIJ(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; 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; 763 } else { 764 C->ops->solve = MatSolve_SeqAIJ; 765 } 766 } 767 C->ops->solveadd = MatSolveAdd_SeqAIJ; 768 C->ops->solvetranspose = MatSolveTranspose_SeqAIJ; 769 C->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ; 770 C->ops->matsolve = MatMatSolve_SeqAIJ; 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; 922 A->ops->solvetranspose = MatSolveTranspose_SeqAIJ; 923 A->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ; 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" 960 PetscErrorCode MatSolve_SeqAIJ(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" 1014 PetscErrorCode MatMatSolve_SeqAIJ(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__ "MatSolve_SeqAIJ_InplaceWithPerm" 1076 PetscErrorCode MatSolve_SeqAIJ_InplaceWithPerm(Mat A,Vec bb,Vec xx) 1077 { 1078 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1079 IS iscol = a->col,isrow = a->row; 1080 PetscErrorCode ierr; 1081 const PetscInt *r,*c,*rout,*cout; 1082 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 1083 PetscInt nz,row; 1084 PetscScalar *x,*b,*tmp,*tmps,sum; 1085 const MatScalar *aa = a->a,*v; 1086 1087 PetscFunctionBegin; 1088 if (!n) PetscFunctionReturn(0); 1089 1090 ierr = VecGetArray(bb,&b);CHKERRQ(ierr); 1091 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1092 tmp = a->solve_work; 1093 1094 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1095 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1096 1097 /* forward solve the lower triangular */ 1098 tmp[0] = b[*r++]; 1099 tmps = tmp; 1100 for (row=1; row<n; row++) { 1101 i = rout[row]; /* permuted row */ 1102 v = aa + ai[i] ; 1103 vi = aj + ai[i] ; 1104 nz = a->diag[i] - ai[i]; 1105 sum = b[*r++]; 1106 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1107 tmp[row] = sum; 1108 } 1109 1110 /* backward solve the upper triangular */ 1111 for (row=n-1; row>=0; row--){ 1112 i = rout[row]; /* permuted row */ 1113 v = aa + a->diag[i] + 1; 1114 vi = aj + a->diag[i] + 1; 1115 nz = ai[i+1] - a->diag[i] - 1; 1116 sum = tmp[row]; 1117 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1118 x[*c--] = tmp[row] = sum*aa[a->diag[i]]; 1119 } 1120 1121 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1122 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1123 ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr); 1124 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1125 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1126 PetscFunctionReturn(0); 1127 } 1128 1129 /* ----------------------------------------------------------- */ 1130 #include "../src/mat/impls/aij/seq/ftn-kernels/fsolve.h" 1131 #undef __FUNCT__ 1132 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering" 1133 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat A,Vec bb,Vec xx) 1134 { 1135 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1136 PetscErrorCode ierr; 1137 PetscInt n = A->rmap->n; 1138 const PetscInt *ai = a->i,*aj = a->j,*adiag = a->diag; 1139 PetscScalar *x; 1140 const PetscScalar *b; 1141 const MatScalar *aa = a->a; 1142 #if !defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 1143 PetscInt adiag_i,i,nz,ai_i; 1144 const PetscInt *vi; 1145 const MatScalar *v; 1146 PetscScalar sum; 1147 #endif 1148 1149 PetscFunctionBegin; 1150 if (!n) PetscFunctionReturn(0); 1151 1152 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1153 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1154 1155 #if defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 1156 fortransolveaij_(&n,x,ai,aj,adiag,aa,b); 1157 #else 1158 /* forward solve the lower triangular */ 1159 x[0] = b[0]; 1160 for (i=1; i<n; i++) { 1161 ai_i = ai[i]; 1162 v = aa + ai_i; 1163 vi = aj + ai_i; 1164 nz = adiag[i] - ai_i; 1165 sum = b[i]; 1166 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 1167 x[i] = sum; 1168 } 1169 1170 /* backward solve the upper triangular */ 1171 for (i=n-1; i>=0; i--){ 1172 adiag_i = adiag[i]; 1173 v = aa + adiag_i + 1; 1174 vi = aj + adiag_i + 1; 1175 nz = ai[i+1] - adiag_i - 1; 1176 sum = x[i]; 1177 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 1178 x[i] = sum*aa[adiag_i]; 1179 } 1180 #endif 1181 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1182 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1183 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1184 PetscFunctionReturn(0); 1185 } 1186 1187 #undef __FUNCT__ 1188 #define __FUNCT__ "MatSolveAdd_SeqAIJ" 1189 PetscErrorCode MatSolveAdd_SeqAIJ(Mat A,Vec bb,Vec yy,Vec xx) 1190 { 1191 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1192 IS iscol = a->col,isrow = a->row; 1193 PetscErrorCode ierr; 1194 PetscInt i, n = A->rmap->n,j; 1195 PetscInt nz; 1196 const PetscInt *rout,*cout,*r,*c,*vi,*ai = a->i,*aj = a->j; 1197 PetscScalar *x,*tmp,sum; 1198 const PetscScalar *b; 1199 const MatScalar *aa = a->a,*v; 1200 1201 PetscFunctionBegin; 1202 if (yy != xx) {ierr = VecCopy(yy,xx);CHKERRQ(ierr);} 1203 1204 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1205 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1206 tmp = a->solve_work; 1207 1208 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1209 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1210 1211 /* forward solve the lower triangular */ 1212 tmp[0] = b[*r++]; 1213 for (i=1; i<n; i++) { 1214 v = aa + ai[i] ; 1215 vi = aj + ai[i] ; 1216 nz = a->diag[i] - ai[i]; 1217 sum = b[*r++]; 1218 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1219 tmp[i] = sum; 1220 } 1221 1222 /* backward solve the upper triangular */ 1223 for (i=n-1; i>=0; i--){ 1224 v = aa + a->diag[i] + 1; 1225 vi = aj + a->diag[i] + 1; 1226 nz = ai[i+1] - a->diag[i] - 1; 1227 sum = tmp[i]; 1228 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1229 tmp[i] = sum*aa[a->diag[i]]; 1230 x[*c--] += tmp[i]; 1231 } 1232 1233 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1234 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1235 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1236 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1237 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1238 1239 PetscFunctionReturn(0); 1240 } 1241 1242 #undef __FUNCT__ 1243 #define __FUNCT__ "MatSolveTranspose_SeqAIJ" 1244 PetscErrorCode MatSolveTranspose_SeqAIJ(Mat A,Vec bb,Vec xx) 1245 { 1246 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1247 IS iscol = a->col,isrow = a->row; 1248 PetscErrorCode ierr; 1249 const PetscInt *rout,*cout,*r,*c,*diag = a->diag,*ai = a->i,*aj = a->j,*vi; 1250 PetscInt i,n = A->rmap->n,j; 1251 PetscInt nz; 1252 PetscScalar *x,*tmp,s1; 1253 const MatScalar *aa = a->a,*v; 1254 const PetscScalar *b; 1255 1256 PetscFunctionBegin; 1257 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1258 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1259 tmp = a->solve_work; 1260 1261 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1262 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1263 1264 /* copy the b into temp work space according to permutation */ 1265 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1266 1267 /* forward solve the U^T */ 1268 for (i=0; i<n; i++) { 1269 v = aa + diag[i] ; 1270 vi = aj + diag[i] + 1; 1271 nz = ai[i+1] - diag[i] - 1; 1272 s1 = tmp[i]; 1273 s1 *= (*v++); /* multiply by inverse of diagonal entry */ 1274 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1275 tmp[i] = s1; 1276 } 1277 1278 /* backward solve the L^T */ 1279 for (i=n-1; i>=0; i--){ 1280 v = aa + diag[i] - 1 ; 1281 vi = aj + diag[i] - 1 ; 1282 nz = diag[i] - ai[i]; 1283 s1 = tmp[i]; 1284 for (j=0; j>-nz; j--) tmp[vi[j]] -= s1*v[j]; 1285 } 1286 1287 /* copy tmp into x according to permutation */ 1288 for (i=0; i<n; i++) x[r[i]] = tmp[i]; 1289 1290 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1291 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1292 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1293 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1294 1295 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1296 PetscFunctionReturn(0); 1297 } 1298 1299 #undef __FUNCT__ 1300 #define __FUNCT__ "MatSolveTransposeAdd_SeqAIJ" 1301 PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat A,Vec bb,Vec zz,Vec xx) 1302 { 1303 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1304 IS iscol = a->col,isrow = a->row; 1305 PetscErrorCode ierr; 1306 const PetscInt *rout,*cout,*r,*c,*diag = a->diag,*ai = a->i,*aj = a->j,*vi; 1307 PetscInt i,n = A->rmap->n,j; 1308 PetscInt nz; 1309 PetscScalar *x,*tmp,s1; 1310 const MatScalar *aa = a->a,*v; 1311 const PetscScalar *b; 1312 1313 PetscFunctionBegin; 1314 if (zz != xx) {ierr = VecCopy(zz,xx);CHKERRQ(ierr);} 1315 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1316 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1317 tmp = a->solve_work; 1318 1319 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1320 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1321 1322 /* copy the b into temp work space according to permutation */ 1323 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1324 1325 /* forward solve the U^T */ 1326 for (i=0; i<n; i++) { 1327 v = aa + diag[i] ; 1328 vi = aj + diag[i] + 1; 1329 nz = ai[i+1] - diag[i] - 1; 1330 s1 = tmp[i]; 1331 s1 *= (*v++); /* multiply by inverse of diagonal entry */ 1332 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1333 tmp[i] = s1; 1334 } 1335 1336 /* backward solve the L^T */ 1337 for (i=n-1; i>=0; i--){ 1338 v = aa + diag[i] - 1 ; 1339 vi = aj + diag[i] - 1 ; 1340 nz = diag[i] - ai[i]; 1341 s1 = tmp[i]; 1342 for (j=0; j>-nz; j--) tmp[vi[j]] -= s1*v[j]; 1343 } 1344 1345 /* copy tmp into x according to permutation */ 1346 for (i=0; i<n; i++) x[r[i]] += tmp[i]; 1347 1348 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1349 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1350 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1351 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1352 1353 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1354 PetscFunctionReturn(0); 1355 } 1356 1357 /* ----------------------------------------------------------------*/ 1358 EXTERN PetscErrorCode Mat_CheckInode(Mat,PetscTruth); 1359 EXTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat,Mat,MatDuplicateOption,PetscTruth); 1360 1361 /* 1362 ilu() under revised new data structure. 1363 Factored arrays bj and ba are stored as 1364 L(0,:), L(1,:), ...,L(n-1,:), U(n-1,:),...,U(i,:),U(i-1,:),...,U(0,:) 1365 1366 bi=fact->i is an array of size n+1, in which 1367 bi+ 1368 bi[i]: points to 1st entry of L(i,:),i=0,...,n-1 1369 bi[n]: points to L(n-1,n-1)+1 1370 1371 bdiag=fact->diag is an array of size n+1,in which 1372 bdiag[i]: points to diagonal of U(i,:), i=0,...,n-1 1373 bdiag[n]: points to diagonal of U(n-1,0)-1 1374 1375 U(i,:) contains bdiag[i] as its last entry, i.e., 1376 U(i,:) = (u[i,i+1],...,u[i,n-1],diag[i]) 1377 */ 1378 #undef __FUNCT__ 1379 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct" 1380 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1381 { 1382 1383 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1384 PetscErrorCode ierr; 1385 PetscInt n=A->rmap->n,*ai=a->i,*aj,*adiag=a->diag; 1386 PetscInt i,j,nz,*bi,*bj,*bdiag,bi_temp; 1387 PetscTruth missing; 1388 IS isicol; 1389 1390 PetscFunctionBegin; 1391 /* printf("MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct ...\n"); */ 1392 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); 1393 ierr = MatMissingDiagonal(A,&missing,&i);CHKERRQ(ierr); 1394 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",i); 1395 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1396 1397 ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES,PETSC_FALSE);CHKERRQ(ierr); 1398 b = (Mat_SeqAIJ*)(fact)->data; 1399 1400 /* allocate matrix arrays for new data structure */ 1401 ierr = PetscMalloc3(ai[n]+1,PetscScalar,&b->a,ai[n]+1,PetscInt,&b->j,n+1,PetscInt,&b->i);CHKERRQ(ierr); 1402 ierr = PetscLogObjectMemory(fact,ai[n]*(sizeof(PetscScalar)+sizeof(PetscInt))+(n+1)*sizeof(PetscInt));CHKERRQ(ierr); 1403 b->singlemalloc = PETSC_TRUE; 1404 if (!b->diag){ 1405 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&b->diag);CHKERRQ(ierr); 1406 ierr = PetscLogObjectMemory(fact,(n+1)*sizeof(PetscInt));CHKERRQ(ierr); 1407 } 1408 bdiag = b->diag; 1409 1410 if (n > 0) { 1411 ierr = PetscMemzero(b->a,(ai[n])*sizeof(MatScalar));CHKERRQ(ierr); 1412 } 1413 1414 /* set bi and bj with new data structure */ 1415 bi = b->i; 1416 bj = b->j; 1417 1418 /* L part */ 1419 bi[0] = 0; 1420 for (i=0; i<n; i++){ 1421 nz = adiag[i] - ai[i]; 1422 bi[i+1] = bi[i] + nz; 1423 aj = a->j + ai[i]; 1424 for (j=0; j<nz; j++){ 1425 *bj = aj[j]; bj++; 1426 } 1427 } 1428 1429 /* U part */ 1430 bi_temp = bi[n]; 1431 bdiag[n] = bi[n]-1; 1432 for (i=n-1; i>=0; i--){ 1433 nz = ai[i+1] - adiag[i] - 1; 1434 bi_temp = bi_temp + nz + 1; 1435 aj = a->j + adiag[i] + 1; 1436 for (j=0; j<nz; j++){ 1437 *bj = aj[j]; bj++; 1438 } 1439 /* diag[i] */ 1440 *bj = i; bj++; 1441 bdiag[i] = bi_temp - 1; 1442 } 1443 1444 fact->factor = MAT_FACTOR_ILU; 1445 fact->info.factor_mallocs = 0; 1446 fact->info.fill_ratio_given = info->fill; 1447 fact->info.fill_ratio_needed = 1.0; 1448 fact->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 1449 1450 b = (Mat_SeqAIJ*)(fact)->data; 1451 b->row = isrow; 1452 b->col = iscol; 1453 b->icol = isicol; 1454 ierr = PetscMalloc((fact->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1455 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1456 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1457 PetscFunctionReturn(0); 1458 } 1459 1460 #undef __FUNCT__ 1461 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_newdatastruct" 1462 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1463 { 1464 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1465 IS isicol; 1466 PetscErrorCode ierr; 1467 const PetscInt *r,*ic; 1468 PetscInt n=A->rmap->n,*ai=a->i,*aj=a->j; 1469 PetscInt *bi,*cols,nnz,*cols_lvl; 1470 PetscInt *bdiag,prow,fm,nzbd,reallocs=0,dcount=0; 1471 PetscInt i,levels,diagonal_fill; 1472 PetscTruth col_identity,row_identity; 1473 PetscReal f; 1474 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL; 1475 PetscBT lnkbt; 1476 PetscInt nzi,*bj,**bj_ptr,**bjlvl_ptr; 1477 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1478 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 1479 1480 PetscFunctionBegin; 1481 /* printf("MatILUFactorSymbolic_SeqAIJ_newdatastruct ...\n"); */ 1482 levels = (PetscInt)info->levels; 1483 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 1484 ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 1485 1486 if (!levels && row_identity && col_identity) { 1487 /* special case: ilu(0) with natural ordering */ 1488 ierr = MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1489 PetscFunctionReturn(0); 1490 } 1491 1492 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); 1493 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1494 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1495 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 1496 1497 /* get new row and diagonal pointers */ 1498 ierr = PetscMalloc2(n+1,PetscInt,&bi,n+1,PetscInt,&bdiag);CHKERRQ(ierr); 1499 bi[0] = bdiag[0] = 0; 1500 1501 ierr = PetscMalloc2(n,PetscInt*,&bj_ptr,n,PetscInt*,&bjlvl_ptr);CHKERRQ(ierr); 1502 1503 /* create a linked list for storing column indices of the active row */ 1504 nlnk = n + 1; 1505 ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1506 1507 /* initial FreeSpace size is f*(ai[n]+1) */ 1508 f = info->fill; 1509 diagonal_fill = (PetscInt)info->diagonal_fill; 1510 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 1511 current_space = free_space; 1512 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr); 1513 current_space_lvl = free_space_lvl; 1514 1515 for (i=0; i<n; i++) { 1516 nzi = 0; 1517 /* copy current row into linked list */ 1518 nnz = ai[r[i]+1] - ai[r[i]]; 1519 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 1520 cols = aj + ai[r[i]]; 1521 lnk[i] = -1; /* marker to indicate if diagonal exists */ 1522 ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1523 nzi += nlnk; 1524 1525 /* make sure diagonal entry is included */ 1526 if (diagonal_fill && lnk[i] == -1) { 1527 fm = n; 1528 while (lnk[fm] < i) fm = lnk[fm]; 1529 lnk[i] = lnk[fm]; /* insert diagonal into linked list */ 1530 lnk[fm] = i; 1531 lnk_lvl[i] = 0; 1532 nzi++; dcount++; 1533 } 1534 1535 /* add pivot rows into the active row */ 1536 nzbd = 0; 1537 prow = lnk[n]; 1538 while (prow < i) { 1539 nnz = bdiag[prow]; 1540 cols = bj_ptr[prow] + nnz + 1; 1541 cols_lvl = bjlvl_ptr[prow] + nnz + 1; 1542 nnz = bi[prow+1] - bi[prow] - nnz - 1; 1543 ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr); 1544 nzi += nlnk; 1545 prow = lnk[prow]; 1546 nzbd++; 1547 } 1548 bdiag[i] = nzbd; 1549 bi[i+1] = bi[i] + nzi; 1550 1551 /* if free space is not available, make more free space */ 1552 if (current_space->local_remaining<nzi) { 1553 nnz = 2*nzi*(n - i); /* estimated and max additional space needed */ 1554 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 1555 ierr = PetscFreeSpaceGet(nnz,¤t_space_lvl);CHKERRQ(ierr); 1556 reallocs++; 1557 } 1558 1559 /* copy data into free_space and free_space_lvl, then initialize lnk */ 1560 ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 1561 bj_ptr[i] = current_space->array; 1562 bjlvl_ptr[i] = current_space_lvl->array; 1563 1564 /* make sure the active row i has diagonal entry */ 1565 if (*(bj_ptr[i]+bdiag[i]) != i) { 1566 SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\ 1567 try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i); 1568 } 1569 1570 current_space->array += nzi; 1571 current_space->local_used += nzi; 1572 current_space->local_remaining -= nzi; 1573 current_space_lvl->array += nzi; 1574 current_space_lvl->local_used += nzi; 1575 current_space_lvl->local_remaining -= nzi; 1576 } 1577 1578 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 1579 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 1580 1581 /* destroy list of free space and other temporary arrays */ 1582 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1583 1584 /* copy free_space into bj and free free_space; set bi, bj, bdiag in new datastructure; */ 1585 ierr = PetscFreeSpaceContiguous_LU_v2(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 1586 1587 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1588 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 1589 ierr = PetscFree2(bj_ptr,bjlvl_ptr);CHKERRQ(ierr); 1590 1591 #if defined(PETSC_USE_INFO) 1592 { 1593 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 1594 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 1595 ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 1596 ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr); 1597 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 1598 if (diagonal_fill) { 1599 ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr); 1600 } 1601 } 1602 #endif 1603 1604 /* put together the new matrix */ 1605 ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 1606 ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr); 1607 b = (Mat_SeqAIJ*)(fact)->data; 1608 b->free_a = PETSC_TRUE; 1609 b->free_ij = PETSC_TRUE; 1610 b->singlemalloc = PETSC_FALSE; 1611 ierr = PetscMalloc((bdiag[0]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 1612 b->j = bj; 1613 b->i = bi; 1614 b->diag = bdiag; 1615 b->ilen = 0; 1616 b->imax = 0; 1617 b->row = isrow; 1618 b->col = iscol; 1619 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1620 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1621 b->icol = isicol; 1622 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1623 /* In b structure: Free imax, ilen, old a, old j. 1624 Allocate bdiag, solve_work, new a, new j */ 1625 ierr = PetscLogObjectMemory(fact,(bdiag[0]+1)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 1626 b->maxnz = b->nz = bdiag[0]+1; 1627 (fact)->info.factor_mallocs = reallocs; 1628 (fact)->info.fill_ratio_given = f; 1629 (fact)->info.fill_ratio_needed = ((PetscReal)(bdiag[0]+1))/((PetscReal)ai[n]); 1630 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 1631 /* ierr = MatILUFactorSymbolic_SeqAIJ_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); */ 1632 PetscFunctionReturn(0); 1633 } 1634 1635 #undef __FUNCT__ 1636 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ" 1637 PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1638 { 1639 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1640 IS isicol; 1641 PetscErrorCode ierr; 1642 const PetscInt *r,*ic; 1643 PetscInt n=A->rmap->n,*ai=a->i,*aj=a->j,d; 1644 PetscInt *bi,*cols,nnz,*cols_lvl; 1645 PetscInt *bdiag,prow,fm,nzbd,reallocs=0,dcount=0; 1646 PetscInt i,levels,diagonal_fill; 1647 PetscTruth col_identity,row_identity; 1648 PetscReal f; 1649 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL; 1650 PetscBT lnkbt; 1651 PetscInt nzi,*bj,**bj_ptr,**bjlvl_ptr; 1652 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1653 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 1654 PetscTruth missing; 1655 PetscTruth newdatastruct=PETSC_FALSE; 1656 1657 PetscFunctionBegin; 1658 ierr = PetscOptionsGetTruth(PETSC_NULL,"-ilu_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 1659 if(newdatastruct){ 1660 ierr = MatILUFactorSymbolic_SeqAIJ_newdatastruct(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1661 PetscFunctionReturn(0); 1662 } 1663 1664 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); 1665 f = info->fill; 1666 levels = (PetscInt)info->levels; 1667 diagonal_fill = (PetscInt)info->diagonal_fill; 1668 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1669 1670 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 1671 ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 1672 if (!levels && row_identity && col_identity) { /* special case: ilu(0) with natural ordering */ 1673 ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES,PETSC_TRUE);CHKERRQ(ierr); 1674 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ; 1675 1676 fact->factor = MAT_FACTOR_ILU; 1677 (fact)->info.factor_mallocs = 0; 1678 (fact)->info.fill_ratio_given = info->fill; 1679 (fact)->info.fill_ratio_needed = 1.0; 1680 b = (Mat_SeqAIJ*)(fact)->data; 1681 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 1682 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 1683 b->row = isrow; 1684 b->col = iscol; 1685 b->icol = isicol; 1686 ierr = PetscMalloc(((fact)->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1687 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1688 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1689 ierr = MatILUFactorSymbolic_SeqAIJ_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1690 PetscFunctionReturn(0); 1691 } 1692 1693 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1694 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 1695 1696 /* get new row and diagonal pointers */ 1697 ierr = PetscMalloc2(n+1,PetscInt,&bi,n+1,PetscInt,&bdiag);CHKERRQ(ierr); 1698 bi[0] = bdiag[0] = 0; 1699 1700 ierr = PetscMalloc2(n,PetscInt*,&bj_ptr,n,PetscInt*,&bjlvl_ptr);CHKERRQ(ierr); 1701 1702 /* create a linked list for storing column indices of the active row */ 1703 nlnk = n + 1; 1704 ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1705 1706 /* initial FreeSpace size is f*(ai[n]+1) */ 1707 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 1708 current_space = free_space; 1709 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr); 1710 current_space_lvl = free_space_lvl; 1711 1712 for (i=0; i<n; i++) { 1713 nzi = 0; 1714 /* copy current row into linked list */ 1715 nnz = ai[r[i]+1] - ai[r[i]]; 1716 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 1717 cols = aj + ai[r[i]]; 1718 lnk[i] = -1; /* marker to indicate if diagonal exists */ 1719 ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1720 nzi += nlnk; 1721 1722 /* make sure diagonal entry is included */ 1723 if (diagonal_fill && lnk[i] == -1) { 1724 fm = n; 1725 while (lnk[fm] < i) fm = lnk[fm]; 1726 lnk[i] = lnk[fm]; /* insert diagonal into linked list */ 1727 lnk[fm] = i; 1728 lnk_lvl[i] = 0; 1729 nzi++; dcount++; 1730 } 1731 1732 /* add pivot rows into the active row */ 1733 nzbd = 0; 1734 prow = lnk[n]; 1735 while (prow < i) { 1736 nnz = bdiag[prow]; 1737 cols = bj_ptr[prow] + nnz + 1; 1738 cols_lvl = bjlvl_ptr[prow] + nnz + 1; 1739 nnz = bi[prow+1] - bi[prow] - nnz - 1; 1740 ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr); 1741 nzi += nlnk; 1742 prow = lnk[prow]; 1743 nzbd++; 1744 } 1745 bdiag[i] = nzbd; 1746 bi[i+1] = bi[i] + nzi; 1747 1748 /* if free space is not available, make more free space */ 1749 if (current_space->local_remaining<nzi) { 1750 nnz = nzi*(n - i); /* estimated and max additional space needed */ 1751 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 1752 ierr = PetscFreeSpaceGet(nnz,¤t_space_lvl);CHKERRQ(ierr); 1753 reallocs++; 1754 } 1755 1756 /* copy data into free_space and free_space_lvl, then initialize lnk */ 1757 ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 1758 bj_ptr[i] = current_space->array; 1759 bjlvl_ptr[i] = current_space_lvl->array; 1760 1761 /* make sure the active row i has diagonal entry */ 1762 if (*(bj_ptr[i]+bdiag[i]) != i) { 1763 SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\ 1764 try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i); 1765 } 1766 1767 current_space->array += nzi; 1768 current_space->local_used += nzi; 1769 current_space->local_remaining -= nzi; 1770 current_space_lvl->array += nzi; 1771 current_space_lvl->local_used += nzi; 1772 current_space_lvl->local_remaining -= nzi; 1773 } 1774 1775 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 1776 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 1777 1778 /* destroy list of free space and other temporary arrays */ 1779 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1780 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); /* copy free_space -> bj */ 1781 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1782 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 1783 ierr = PetscFree2(bj_ptr,bjlvl_ptr);CHKERRQ(ierr); 1784 1785 #if defined(PETSC_USE_INFO) 1786 { 1787 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 1788 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 1789 ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 1790 ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr); 1791 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 1792 if (diagonal_fill) { 1793 ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr); 1794 } 1795 } 1796 #endif 1797 1798 /* put together the new matrix */ 1799 ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 1800 ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr); 1801 b = (Mat_SeqAIJ*)(fact)->data; 1802 b->free_a = PETSC_TRUE; 1803 b->free_ij = PETSC_TRUE; 1804 b->singlemalloc = PETSC_FALSE; 1805 ierr = PetscMalloc(bi[n]*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 1806 b->j = bj; 1807 b->i = bi; 1808 for (i=0; i<n; i++) bdiag[i] += bi[i]; 1809 b->diag = bdiag; 1810 b->ilen = 0; 1811 b->imax = 0; 1812 b->row = isrow; 1813 b->col = iscol; 1814 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1815 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1816 b->icol = isicol; 1817 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1818 /* In b structure: Free imax, ilen, old a, old j. 1819 Allocate bdiag, solve_work, new a, new j */ 1820 ierr = PetscLogObjectMemory(fact,(bi[n]-n) * (sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 1821 b->maxnz = b->nz = bi[n] ; 1822 (fact)->info.factor_mallocs = reallocs; 1823 (fact)->info.fill_ratio_given = f; 1824 (fact)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]); 1825 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ; 1826 ierr = MatILUFactorSymbolic_SeqAIJ_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1827 PetscFunctionReturn(0); 1828 } 1829 1830 #undef __FUNCT__ 1831 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ_newdatastruct" 1832 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 1833 { 1834 Mat C = B; 1835 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 1836 Mat_SeqSBAIJ *b=(Mat_SeqSBAIJ*)C->data; 1837 IS ip=b->row,iip = b->icol; 1838 PetscErrorCode ierr; 1839 const PetscInt *rip,*riip; 1840 PetscInt i,j,mbs=A->rmap->n,*bi=b->i,*bj=b->j,*bdiag=b->diag,*bjtmp; 1841 PetscInt *ai=a->i,*aj=a->j; 1842 PetscInt k,jmin,jmax,*c2r,*il,col,nexti,ili,nz; 1843 MatScalar *rtmp,*ba=b->a,*bval,*aa=a->a,dk,uikdi; 1844 PetscReal zeropivot,shiftnz; 1845 PetscReal shiftpd; 1846 PetscTruth perm_identity; 1847 1848 PetscFunctionBegin; 1849 1850 shiftnz = info->shiftnz; 1851 shiftpd = info->shiftpd; 1852 zeropivot = info->zeropivot; 1853 1854 ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 1855 ierr = ISGetIndices(iip,&riip);CHKERRQ(ierr); 1856 1857 /* allocate working arrays 1858 c2r: linked list, keep track of pivot rows for a given column. c2r[col]: head of the list for a given col 1859 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 1860 */ 1861 ierr = PetscMalloc3(mbs,MatScalar,&rtmp,mbs,PetscInt,&il,mbs,PetscInt,&c2r);CHKERRQ(ierr); 1862 1863 for (i=0; i<mbs; i++) { 1864 c2r[i] = mbs; 1865 } 1866 il[0] = 0; 1867 1868 for (k = 0; k<mbs; k++){ 1869 /* zero rtmp */ 1870 nz = bi[k+1] - bi[k]; 1871 bjtmp = bj + bi[k]; 1872 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 1873 1874 /* load in initial unfactored row */ 1875 bval = ba + bi[k]; 1876 jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 1877 for (j = jmin; j < jmax; j++){ 1878 col = riip[aj[j]]; 1879 if (col >= k){ /* only take upper triangular entry */ 1880 rtmp[col] = aa[j]; 1881 *bval++ = 0.0; /* for in-place factorization */ 1882 } 1883 } 1884 /* shift the diagonal of the matrix */ 1885 1886 /* modify k-th row by adding in those rows i with U(i,k)!=0 */ 1887 dk = rtmp[k]; 1888 i = c2r[k]; /* first row to be added to k_th row */ 1889 1890 while (i < k){ 1891 nexti = c2r[i]; /* next row to be added to k_th row */ 1892 1893 /* compute multiplier, update diag(k) and U(i,k) */ 1894 ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 1895 uikdi = - ba[ili]*ba[bdiag[i]]; /* diagonal(k) */ 1896 dk += uikdi*ba[ili]; /* update diag[k] */ 1897 ba[ili] = uikdi; /* -U(i,k) */ 1898 1899 /* add multiple of row i to k-th row */ 1900 jmin = ili + 1; jmax = bi[i+1]; 1901 if (jmin < jmax){ 1902 for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 1903 /* update il and c2r for row i */ 1904 il[i] = jmin; 1905 j = bj[jmin]; c2r[i] = c2r[j]; c2r[j] = i; 1906 } 1907 i = nexti; 1908 } 1909 1910 /* copy data into U(k,:) */ 1911 ba[bdiag[k]] = 1.0/dk; /* U(k,k) */ 1912 jmin = bi[k]; jmax = bi[k+1]-1; 1913 if (jmin < jmax) { 1914 for (j=jmin; j<jmax; j++){ 1915 col = bj[j]; ba[j] = rtmp[col]; 1916 } 1917 /* add the k-th row into il and c2r */ 1918 il[k] = jmin; 1919 i = bj[jmin]; c2r[k] = c2r[i]; c2r[i] = k; 1920 } 1921 } 1922 1923 ierr = PetscFree3(rtmp,il,c2r);CHKERRQ(ierr); 1924 ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 1925 ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr); 1926 1927 ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr); 1928 if (perm_identity){ 1929 (B)->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering_newdatastruct; 1930 (B)->ops->solvetranspose = 0; 1931 (B)->ops->forwardsolve = 0; 1932 (B)->ops->backwardsolve = 0; 1933 } else { 1934 (B)->ops->solve = MatSolve_SeqSBAIJ_1_newdatastruct; 1935 (B)->ops->solvetranspose = 0; 1936 (B)->ops->forwardsolve = 0; 1937 (B)->ops->backwardsolve = 0; 1938 } 1939 1940 C->assembled = PETSC_TRUE; 1941 C->preallocated = PETSC_TRUE; 1942 ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr); 1943 PetscFunctionReturn(0); 1944 } 1945 1946 #undef __FUNCT__ 1947 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ" 1948 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat B,Mat A,const MatFactorInfo *info) 1949 { 1950 Mat C = B; 1951 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 1952 Mat_SeqSBAIJ *b=(Mat_SeqSBAIJ*)C->data; 1953 IS ip=b->row,iip = b->icol; 1954 PetscErrorCode ierr; 1955 const PetscInt *rip,*riip; 1956 PetscInt i,j,mbs=A->rmap->n,*bi=b->i,*bj=b->j,*bcol; 1957 PetscInt *ai=a->i,*aj=a->j; 1958 PetscInt k,jmin,jmax,*jl,*il,col,nexti,ili,nz; 1959 MatScalar *rtmp,*ba=b->a,*bval,*aa=a->a,dk,uikdi; 1960 PetscReal zeropivot,rs,shiftnz; 1961 PetscReal shiftpd; 1962 ChShift_Ctx sctx; 1963 PetscInt newshift; 1964 PetscTruth perm_identity; 1965 1966 PetscFunctionBegin; 1967 shiftnz = info->shiftnz; 1968 shiftpd = info->shiftpd; 1969 zeropivot = info->zeropivot; 1970 1971 ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 1972 ierr = ISGetIndices(iip,&riip);CHKERRQ(ierr); 1973 1974 /* initialization */ 1975 ierr = PetscMalloc3(mbs,MatScalar,&rtmp,mbs,PetscInt,&il,mbs,PetscInt,&jl);CHKERRQ(ierr); 1976 sctx.shift_amount = 0; 1977 sctx.nshift = 0; 1978 do { 1979 sctx.chshift = PETSC_FALSE; 1980 for (i=0; i<mbs; i++) { 1981 rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0; 1982 } 1983 1984 for (k = 0; k<mbs; k++){ 1985 bval = ba + bi[k]; 1986 /* initialize k-th row by the perm[k]-th row of A */ 1987 jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 1988 for (j = jmin; j < jmax; j++){ 1989 col = riip[aj[j]]; 1990 if (col >= k){ /* only take upper triangular entry */ 1991 rtmp[col] = aa[j]; 1992 *bval++ = 0.0; /* for in-place factorization */ 1993 } 1994 } 1995 /* shift the diagonal of the matrix */ 1996 if (sctx.nshift) rtmp[k] += sctx.shift_amount; 1997 1998 /* modify k-th row by adding in those rows i with U(i,k)!=0 */ 1999 dk = rtmp[k]; 2000 i = jl[k]; /* first row to be added to k_th row */ 2001 2002 while (i < k){ 2003 nexti = jl[i]; /* next row to be added to k_th row */ 2004 2005 /* compute multiplier, update diag(k) and U(i,k) */ 2006 ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2007 uikdi = - ba[ili]*ba[bi[i]]; /* diagonal(k) */ 2008 dk += uikdi*ba[ili]; 2009 ba[ili] = uikdi; /* -U(i,k) */ 2010 2011 /* add multiple of row i to k-th row */ 2012 jmin = ili + 1; jmax = bi[i+1]; 2013 if (jmin < jmax){ 2014 for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 2015 /* update il and jl for row i */ 2016 il[i] = jmin; 2017 j = bj[jmin]; jl[i] = jl[j]; jl[j] = i; 2018 } 2019 i = nexti; 2020 } 2021 2022 /* shift the diagonals when zero pivot is detected */ 2023 /* compute rs=sum of abs(off-diagonal) */ 2024 rs = 0.0; 2025 jmin = bi[k]+1; 2026 nz = bi[k+1] - jmin; 2027 bcol = bj + jmin; 2028 for (j=0; j<nz; j++) { 2029 rs += PetscAbsScalar(rtmp[bcol[j]]); 2030 } 2031 2032 sctx.rs = rs; 2033 sctx.pv = dk; 2034 ierr = MatCholeskyCheckShift_inline(info,sctx,k,newshift);CHKERRQ(ierr); 2035 2036 if (newshift == 1) { 2037 if (!sctx.shift_amount) { 2038 sctx.shift_amount = 1e-5; 2039 } 2040 break; 2041 } 2042 2043 /* copy data into U(k,:) */ 2044 ba[bi[k]] = 1.0/dk; /* U(k,k) */ 2045 jmin = bi[k]+1; jmax = bi[k+1]; 2046 if (jmin < jmax) { 2047 for (j=jmin; j<jmax; j++){ 2048 col = bj[j]; ba[j] = rtmp[col]; rtmp[col] = 0.0; 2049 } 2050 /* add the k-th row into il and jl */ 2051 il[k] = jmin; 2052 i = bj[jmin]; jl[k] = jl[i]; jl[i] = k; 2053 } 2054 } 2055 } while (sctx.chshift); 2056 ierr = PetscFree3(rtmp,il,jl);CHKERRQ(ierr); 2057 ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 2058 ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr); 2059 2060 ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr); 2061 if (perm_identity){ 2062 (B)->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering; 2063 (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1_NaturalOrdering; 2064 (B)->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1_NaturalOrdering; 2065 (B)->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1_NaturalOrdering; 2066 } else { 2067 (B)->ops->solve = MatSolve_SeqSBAIJ_1; 2068 (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1; 2069 (B)->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1; 2070 (B)->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1; 2071 } 2072 2073 C->assembled = PETSC_TRUE; 2074 C->preallocated = PETSC_TRUE; 2075 ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr); 2076 if (sctx.nshift){ 2077 if (shiftnz) { 2078 ierr = PetscInfo2(A,"number of shiftnz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2079 } else if (shiftpd) { 2080 ierr = PetscInfo2(A,"number of shiftpd tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2081 } 2082 } 2083 PetscFunctionReturn(0); 2084 } 2085 2086 /* 2087 icc() under revised new data structure. 2088 Factored arrays bj and ba are stored as 2089 U(0,:),...,U(i,:),U(n-1,:) 2090 2091 ui=fact->i is an array of size n+1, in which 2092 ui+ 2093 ui[i]: points to 1st entry of U(i,:),i=0,...,n-1 2094 ui[n]: points to U(n-1,n-1)+1 2095 2096 udiag=fact->diag is an array of size n,in which 2097 udiag[i]: points to diagonal of U(i,:), i=0,...,n-1 2098 2099 U(i,:) contains udiag[i] as its last entry, i.e., 2100 U(i,:) = (u[i,i+1],...,u[i,n-1],diag[i]) 2101 */ 2102 2103 #undef __FUNCT__ 2104 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ_newdatastruct" 2105 PetscErrorCode MatICCFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2106 { 2107 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2108 Mat_SeqSBAIJ *b; 2109 PetscErrorCode ierr; 2110 PetscTruth perm_identity,missing; 2111 PetscInt reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui,*udiag; 2112 const PetscInt *rip,*riip; 2113 PetscInt jmin,jmax,nzk,k,j,*jl,prow,*il,nextprow; 2114 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL,d; 2115 PetscInt ncols,ncols_upper,*cols,*ajtmp,*uj,**uj_ptr,**uj_lvl_ptr; 2116 PetscReal fill=info->fill,levels=info->levels; 2117 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2118 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2119 PetscBT lnkbt; 2120 IS iperm; 2121 2122 PetscFunctionBegin; 2123 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); 2124 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2125 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2126 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2127 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2128 2129 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2130 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2131 ui[0] = 0; 2132 2133 /* ICC(0) without matrix ordering: simply rearrange column indices */ 2134 if (!levels && perm_identity) { 2135 2136 for (i=0; i<am; i++) { 2137 ui[i+1] = ui[i] + ai[i+1] - a->diag[i]; 2138 udiag[i] = ui[i+1] - 1; /* points to the last entry of U(i,:) */ 2139 } 2140 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2141 cols = uj; 2142 for (i=0; i<am; i++) { 2143 aj = a->j + a->diag[i] + 1; /* 1st entry of U(i,:) without diagonal */ 2144 ncols = ui[i+1] - ui[i] - 1; 2145 for (j=0; j<ncols; j++) *cols++ = aj[j]; 2146 *cols++ = i; /* diagoanl is located as the last entry of U(i,:) */ 2147 } 2148 } else { /* case: levels>0 || (levels=0 && !perm_identity) */ 2149 if (levels) SETERRQ(PETSC_ERR_ARG_WRONG,"not done yet"); 2150 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2151 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2152 2153 /* initialization */ 2154 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ajtmp);CHKERRQ(ierr); 2155 2156 /* jl: linked list for storing indices of the pivot rows 2157 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2158 ierr = PetscMalloc4(am,PetscInt*,&uj_ptr,am,PetscInt*,&uj_lvl_ptr,am,PetscInt,&jl,am,PetscInt,&il);CHKERRQ(ierr); 2159 for (i=0; i<am; i++){ 2160 jl[i] = am; il[i] = 0; 2161 } 2162 2163 /* create and initialize a linked list for storing column indices of the active row k */ 2164 nlnk = am + 1; 2165 ierr = PetscIncompleteLLCreate(am,am,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2166 2167 /* initial FreeSpace size is fill*(ai[am]+1) */ 2168 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2169 current_space = free_space; 2170 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space_lvl);CHKERRQ(ierr); 2171 current_space_lvl = free_space_lvl; 2172 2173 for (k=0; k<am; k++){ /* for each active row k */ 2174 /* initialize lnk by the column indices of row rip[k] of A */ 2175 nzk = 0; 2176 ncols = ai[rip[k]+1] - ai[rip[k]]; 2177 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2178 ncols_upper = 0; 2179 for (j=0; j<ncols; j++){ 2180 i = *(aj + ai[rip[k]] + j); /* unpermuted column index */ 2181 if (riip[i] >= k){ /* only take upper triangular entry */ 2182 ajtmp[ncols_upper] = i; 2183 ncols_upper++; 2184 } 2185 } 2186 ierr = PetscIncompleteLLInit(ncols_upper,ajtmp,am,riip,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2187 nzk += nlnk; 2188 2189 /* update lnk by computing fill-in for each pivot row to be merged in */ 2190 prow = jl[k]; /* 1st pivot row */ 2191 2192 while (prow < k){ 2193 nextprow = jl[prow]; 2194 2195 /* merge prow into k-th row */ 2196 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2197 jmax = ui[prow+1]; 2198 ncols = jmax-jmin; 2199 i = jmin - ui[prow]; 2200 cols = uj_ptr[prow] + i; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2201 uj = uj_lvl_ptr[prow] + i; /* levels of cols */ 2202 j = *(uj - 1); 2203 ierr = PetscICCLLAddSorted(ncols,cols,levels,uj,am,nlnk,lnk,lnk_lvl,lnkbt,j);CHKERRQ(ierr); 2204 nzk += nlnk; 2205 2206 /* update il and jl for prow */ 2207 if (jmin < jmax){ 2208 il[prow] = jmin; 2209 j = *cols; jl[prow] = jl[j]; jl[j] = prow; 2210 } 2211 prow = nextprow; 2212 } 2213 2214 /* if free space is not available, make more free space */ 2215 if (current_space->local_remaining<nzk) { 2216 i = am - k + 1; /* num of unfactored rows */ 2217 i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2218 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2219 ierr = PetscFreeSpaceGet(i,¤t_space_lvl);CHKERRQ(ierr); 2220 reallocs++; 2221 } 2222 2223 /* copy data into free_space and free_space_lvl, then initialize lnk */ 2224 if (nzk == 0) SETERRQ1(PETSC_ERR_ARG_WRONG,"Empty row %D in ICC matrix factor",k); 2225 ierr = PetscIncompleteLLClean(am,am,nzk,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 2226 2227 /* add the k-th row into il and jl */ 2228 if (nzk > 1){ 2229 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2230 jl[k] = jl[i]; jl[i] = k; 2231 il[k] = ui[k] + 1; 2232 } 2233 uj_ptr[k] = current_space->array; 2234 uj_lvl_ptr[k] = current_space_lvl->array; 2235 2236 current_space->array += nzk; 2237 current_space->local_used += nzk; 2238 current_space->local_remaining -= nzk; 2239 2240 current_space_lvl->array += nzk; 2241 current_space_lvl->local_used += nzk; 2242 current_space_lvl->local_remaining -= nzk; 2243 2244 ui[k+1] = ui[k] + nzk; 2245 } 2246 2247 #if defined(PETSC_USE_INFO) 2248 if (ai[am] != 0) { 2249 PetscReal af = (PetscReal)ui[am]/((PetscReal)ai[am]); 2250 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2251 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2252 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2253 } else { 2254 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2255 } 2256 #endif 2257 2258 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2259 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2260 ierr = PetscFree4(uj_ptr,uj_lvl_ptr,jl,il);CHKERRQ(ierr); 2261 ierr = PetscFree(ajtmp);CHKERRQ(ierr); 2262 2263 /* destroy list of free space and other temporary array(s) */ 2264 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2265 ierr = PetscFreeSpaceContiguous_Cholesky(&free_space,uj,am,ui,udiag);CHKERRQ(ierr); 2266 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2267 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2268 2269 } /* end of case: levels>0 || (levels=0 && !perm_identity) */ 2270 2271 /* put together the new matrix in MATSEQSBAIJ format */ 2272 2273 b = (Mat_SeqSBAIJ*)(fact)->data; 2274 b->singlemalloc = PETSC_FALSE; 2275 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2276 b->j = uj; 2277 b->i = ui; 2278 b->diag = udiag; 2279 b->free_diag = PETSC_TRUE; 2280 b->ilen = 0; 2281 b->imax = 0; 2282 b->row = perm; 2283 b->col = perm; 2284 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2285 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2286 b->icol = iperm; 2287 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2288 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2289 ierr = PetscLogObjectMemory((fact),(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2290 b->maxnz = b->nz = ui[am]; 2291 b->free_a = PETSC_TRUE; 2292 b->free_ij = PETSC_TRUE; 2293 2294 (fact)->info.factor_mallocs = reallocs; 2295 (fact)->info.fill_ratio_given = fill; 2296 if (ai[am] != 0) { 2297 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2298 } else { 2299 (fact)->info.fill_ratio_needed = 0.0; 2300 } 2301 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_newdatastruct; 2302 PetscFunctionReturn(0); 2303 } 2304 2305 #undef __FUNCT__ 2306 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ" 2307 PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2308 { 2309 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2310 Mat_SeqSBAIJ *b; 2311 PetscErrorCode ierr; 2312 PetscTruth perm_identity,missing; 2313 PetscInt reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui,*udiag; 2314 const PetscInt *rip,*riip; 2315 PetscInt jmin,jmax,nzk,k,j,*jl,prow,*il,nextprow; 2316 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL,d; 2317 PetscInt ncols,ncols_upper,*cols,*ajtmp,*uj,**uj_ptr,**uj_lvl_ptr; 2318 PetscReal fill=info->fill,levels=info->levels; 2319 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2320 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2321 PetscBT lnkbt; 2322 IS iperm; 2323 PetscTruth newdatastruct=PETSC_FALSE; 2324 2325 PetscFunctionBegin; 2326 ierr = PetscOptionsGetTruth(PETSC_NULL,"-icc_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 2327 if(newdatastruct){ 2328 ierr = MatICCFactorSymbolic_SeqAIJ_newdatastruct(fact,A,perm,info);CHKERRQ(ierr); 2329 PetscFunctionReturn(0); 2330 } 2331 2332 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); 2333 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2334 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2335 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2336 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2337 2338 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2339 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2340 ui[0] = 0; 2341 2342 /* ICC(0) without matrix ordering: simply copies fill pattern */ 2343 if (!levels && perm_identity) { 2344 2345 for (i=0; i<am; i++) { 2346 ui[i+1] = ui[i] + ai[i+1] - a->diag[i]; 2347 udiag[i] = ui[i]; 2348 } 2349 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2350 cols = uj; 2351 for (i=0; i<am; i++) { 2352 aj = a->j + a->diag[i]; 2353 ncols = ui[i+1] - ui[i]; 2354 for (j=0; j<ncols; j++) *cols++ = *aj++; 2355 } 2356 } else { /* case: levels>0 || (levels=0 && !perm_identity) */ 2357 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2358 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2359 2360 /* initialization */ 2361 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ajtmp);CHKERRQ(ierr); 2362 2363 /* jl: linked list for storing indices of the pivot rows 2364 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2365 ierr = PetscMalloc4(am,PetscInt*,&uj_ptr,am,PetscInt*,&uj_lvl_ptr,am,PetscInt,&jl,am,PetscInt,&il);CHKERRQ(ierr); 2366 for (i=0; i<am; i++){ 2367 jl[i] = am; il[i] = 0; 2368 } 2369 2370 /* create and initialize a linked list for storing column indices of the active row k */ 2371 nlnk = am + 1; 2372 ierr = PetscIncompleteLLCreate(am,am,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2373 2374 /* initial FreeSpace size is fill*(ai[am]+1) */ 2375 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2376 current_space = free_space; 2377 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space_lvl);CHKERRQ(ierr); 2378 current_space_lvl = free_space_lvl; 2379 2380 for (k=0; k<am; k++){ /* for each active row k */ 2381 /* initialize lnk by the column indices of row rip[k] of A */ 2382 nzk = 0; 2383 ncols = ai[rip[k]+1] - ai[rip[k]]; 2384 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2385 ncols_upper = 0; 2386 for (j=0; j<ncols; j++){ 2387 i = *(aj + ai[rip[k]] + j); /* unpermuted column index */ 2388 if (riip[i] >= k){ /* only take upper triangular entry */ 2389 ajtmp[ncols_upper] = i; 2390 ncols_upper++; 2391 } 2392 } 2393 ierr = PetscIncompleteLLInit(ncols_upper,ajtmp,am,riip,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2394 nzk += nlnk; 2395 2396 /* update lnk by computing fill-in for each pivot row to be merged in */ 2397 prow = jl[k]; /* 1st pivot row */ 2398 2399 while (prow < k){ 2400 nextprow = jl[prow]; 2401 2402 /* merge prow into k-th row */ 2403 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2404 jmax = ui[prow+1]; 2405 ncols = jmax-jmin; 2406 i = jmin - ui[prow]; 2407 cols = uj_ptr[prow] + i; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2408 uj = uj_lvl_ptr[prow] + i; /* levels of cols */ 2409 j = *(uj - 1); 2410 ierr = PetscICCLLAddSorted(ncols,cols,levels,uj,am,nlnk,lnk,lnk_lvl,lnkbt,j);CHKERRQ(ierr); 2411 nzk += nlnk; 2412 2413 /* update il and jl for prow */ 2414 if (jmin < jmax){ 2415 il[prow] = jmin; 2416 j = *cols; jl[prow] = jl[j]; jl[j] = prow; 2417 } 2418 prow = nextprow; 2419 } 2420 2421 /* if free space is not available, make more free space */ 2422 if (current_space->local_remaining<nzk) { 2423 i = am - k + 1; /* num of unfactored rows */ 2424 i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2425 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2426 ierr = PetscFreeSpaceGet(i,¤t_space_lvl);CHKERRQ(ierr); 2427 reallocs++; 2428 } 2429 2430 /* copy data into free_space and free_space_lvl, then initialize lnk */ 2431 if (nzk == 0) SETERRQ1(PETSC_ERR_ARG_WRONG,"Empty row %D in ICC matrix factor",k); 2432 ierr = PetscIncompleteLLClean(am,am,nzk,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 2433 2434 /* add the k-th row into il and jl */ 2435 if (nzk > 1){ 2436 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2437 jl[k] = jl[i]; jl[i] = k; 2438 il[k] = ui[k] + 1; 2439 } 2440 uj_ptr[k] = current_space->array; 2441 uj_lvl_ptr[k] = current_space_lvl->array; 2442 2443 current_space->array += nzk; 2444 current_space->local_used += nzk; 2445 current_space->local_remaining -= nzk; 2446 2447 current_space_lvl->array += nzk; 2448 current_space_lvl->local_used += nzk; 2449 current_space_lvl->local_remaining -= nzk; 2450 2451 ui[k+1] = ui[k] + nzk; 2452 } 2453 2454 #if defined(PETSC_USE_INFO) 2455 if (ai[am] != 0) { 2456 PetscReal af = (PetscReal)ui[am]/((PetscReal)ai[am]); 2457 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2458 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2459 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2460 } else { 2461 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2462 } 2463 #endif 2464 2465 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2466 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2467 ierr = PetscFree4(uj_ptr,uj_lvl_ptr,jl,il);CHKERRQ(ierr); 2468 ierr = PetscFree(ajtmp);CHKERRQ(ierr); 2469 2470 /* destroy list of free space and other temporary array(s) */ 2471 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2472 ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr); 2473 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2474 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2475 2476 } /* end of case: levels>0 || (levels=0 && !perm_identity) */ 2477 2478 /* put together the new matrix in MATSEQSBAIJ format */ 2479 2480 b = (Mat_SeqSBAIJ*)(fact)->data; 2481 b->singlemalloc = PETSC_FALSE; 2482 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2483 b->j = uj; 2484 b->i = ui; 2485 b->diag = udiag; 2486 b->free_diag = PETSC_TRUE; 2487 b->ilen = 0; 2488 b->imax = 0; 2489 b->row = perm; 2490 b->col = perm; 2491 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2492 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2493 b->icol = iperm; 2494 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2495 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2496 ierr = PetscLogObjectMemory((fact),(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2497 b->maxnz = b->nz = ui[am]; 2498 b->free_a = PETSC_TRUE; 2499 b->free_ij = PETSC_TRUE; 2500 2501 (fact)->info.factor_mallocs = reallocs; 2502 (fact)->info.fill_ratio_given = fill; 2503 if (ai[am] != 0) { 2504 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2505 } else { 2506 (fact)->info.fill_ratio_needed = 0.0; 2507 } 2508 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ; 2509 PetscFunctionReturn(0); 2510 } 2511 2512 #undef __FUNCT__ 2513 #define __FUNCT__ "MatCholeskyFactorSymbolic_SeqAIJ" 2514 PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2515 { 2516 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2517 Mat_SeqSBAIJ *b; 2518 PetscErrorCode ierr; 2519 PetscTruth perm_identity; 2520 PetscReal fill = info->fill; 2521 const PetscInt *rip,*riip; 2522 PetscInt i,am=A->rmap->n,*ai=a->i,*aj=a->j,reallocs=0,prow; 2523 PetscInt *jl,jmin,jmax,nzk,*ui,k,j,*il,nextprow; 2524 PetscInt nlnk,*lnk,ncols,ncols_upper,*cols,*uj,**ui_ptr,*uj_ptr; 2525 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2526 PetscBT lnkbt; 2527 IS iperm; 2528 2529 PetscFunctionBegin; 2530 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); 2531 /* check whether perm is the identity mapping */ 2532 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2533 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2534 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2535 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2536 2537 /* initialization */ 2538 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2539 ui[0] = 0; 2540 2541 /* jl: linked list for storing indices of the pivot rows 2542 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2543 ierr = PetscMalloc4(am,PetscInt*,&ui_ptr,am,PetscInt,&jl,am,PetscInt,&il,am,PetscInt,&cols);CHKERRQ(ierr); 2544 for (i=0; i<am; i++){ 2545 jl[i] = am; il[i] = 0; 2546 } 2547 2548 /* create and initialize a linked list for storing column indices of the active row k */ 2549 nlnk = am + 1; 2550 ierr = PetscLLCreate(am,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2551 2552 /* initial FreeSpace size is fill*(ai[am]+1) */ 2553 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2554 current_space = free_space; 2555 2556 for (k=0; k<am; k++){ /* for each active row k */ 2557 /* initialize lnk by the column indices of row rip[k] of A */ 2558 nzk = 0; 2559 ncols = ai[rip[k]+1] - ai[rip[k]]; 2560 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2561 ncols_upper = 0; 2562 for (j=0; j<ncols; j++){ 2563 i = riip[*(aj + ai[rip[k]] + j)]; 2564 if (i >= k){ /* only take upper triangular entry */ 2565 cols[ncols_upper] = i; 2566 ncols_upper++; 2567 } 2568 } 2569 ierr = PetscLLAdd(ncols_upper,cols,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2570 nzk += nlnk; 2571 2572 /* update lnk by computing fill-in for each pivot row to be merged in */ 2573 prow = jl[k]; /* 1st pivot row */ 2574 2575 while (prow < k){ 2576 nextprow = jl[prow]; 2577 /* merge prow into k-th row */ 2578 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2579 jmax = ui[prow+1]; 2580 ncols = jmax-jmin; 2581 uj_ptr = ui_ptr[prow] + jmin - ui[prow]; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2582 ierr = PetscLLAddSorted(ncols,uj_ptr,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2583 nzk += nlnk; 2584 2585 /* update il and jl for prow */ 2586 if (jmin < jmax){ 2587 il[prow] = jmin; 2588 j = *uj_ptr; jl[prow] = jl[j]; jl[j] = prow; 2589 } 2590 prow = nextprow; 2591 } 2592 2593 /* if free space is not available, make more free space */ 2594 if (current_space->local_remaining<nzk) { 2595 i = am - k + 1; /* num of unfactored rows */ 2596 i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2597 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2598 reallocs++; 2599 } 2600 2601 /* copy data into free space, then initialize lnk */ 2602 ierr = PetscLLClean(am,am,nzk,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 2603 2604 /* add the k-th row into il and jl */ 2605 if (nzk-1 > 0){ 2606 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2607 jl[k] = jl[i]; jl[i] = k; 2608 il[k] = ui[k] + 1; 2609 } 2610 ui_ptr[k] = current_space->array; 2611 current_space->array += nzk; 2612 current_space->local_used += nzk; 2613 current_space->local_remaining -= nzk; 2614 2615 ui[k+1] = ui[k] + nzk; 2616 } 2617 2618 #if defined(PETSC_USE_INFO) 2619 if (ai[am] != 0) { 2620 PetscReal af = (PetscReal)(ui[am])/((PetscReal)ai[am]); 2621 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2622 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2623 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2624 } else { 2625 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2626 } 2627 #endif 2628 2629 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2630 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2631 ierr = PetscFree4(ui_ptr,jl,il,cols);CHKERRQ(ierr); 2632 2633 /* destroy list of free space and other temporary array(s) */ 2634 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2635 ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr); 2636 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2637 2638 /* put together the new matrix in MATSEQSBAIJ format */ 2639 2640 b = (Mat_SeqSBAIJ*)(fact)->data; 2641 b->singlemalloc = PETSC_FALSE; 2642 b->free_a = PETSC_TRUE; 2643 b->free_ij = PETSC_TRUE; 2644 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2645 b->j = uj; 2646 b->i = ui; 2647 b->diag = 0; 2648 b->ilen = 0; 2649 b->imax = 0; 2650 b->row = perm; 2651 b->col = perm; 2652 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2653 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2654 b->icol = iperm; 2655 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2656 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2657 ierr = PetscLogObjectMemory(fact,(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2658 b->maxnz = b->nz = ui[am]; 2659 2660 (fact)->info.factor_mallocs = reallocs; 2661 (fact)->info.fill_ratio_given = fill; 2662 if (ai[am] != 0) { 2663 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2664 } else { 2665 (fact)->info.fill_ratio_needed = 0.0; 2666 } 2667 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ; 2668 PetscFunctionReturn(0); 2669 } 2670 2671 #undef __FUNCT__ 2672 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering_newdatastruct" 2673 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_newdatastruct(Mat A,Vec bb,Vec xx) 2674 { 2675 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2676 PetscErrorCode ierr; 2677 PetscInt n = A->rmap->n; 2678 const PetscInt *ai = a->i,*aj = a->j,*vi; 2679 PetscScalar *x,sum; 2680 const PetscScalar *b; 2681 const MatScalar *aa = a->a,*v; 2682 PetscInt i,nz; 2683 2684 PetscFunctionBegin; 2685 if (!n) PetscFunctionReturn(0); 2686 2687 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 2688 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 2689 2690 /* forward solve the lower triangular */ 2691 x[0] = b[0]; 2692 v = aa; 2693 vi = aj; 2694 for (i=1; i<n; i++) { 2695 nz = ai[i+1] - ai[i]; 2696 sum = b[i]; 2697 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 2698 v += nz; 2699 vi += nz; 2700 x[i] = sum; 2701 } 2702 2703 /* backward solve the upper triangular */ 2704 v = aa + ai[n+1]; 2705 vi = aj + ai[n+1]; 2706 for (i=n-1; i>=0; i--){ 2707 nz = ai[2*n-i +1] - ai[2*n-i]-1; 2708 sum = x[i]; 2709 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 2710 v += nz; 2711 vi += nz; vi++; 2712 x[i] = *v++ *sum; /* x[i]=aa[adiag[i]]*sum; v++; */ 2713 } 2714 2715 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 2716 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 2717 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 2718 PetscFunctionReturn(0); 2719 } 2720 2721 #undef __FUNCT__ 2722 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering_newdatastruct_v2" 2723 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_newdatastruct_v2(Mat A,Vec bb,Vec xx) 2724 { 2725 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2726 PetscErrorCode ierr; 2727 PetscInt n = A->rmap->n; 2728 const PetscInt *ai = a->i,*aj = a->j,*adiag = a->diag,*vi; 2729 PetscScalar *x,sum; 2730 const PetscScalar *b; 2731 const MatScalar *aa = a->a,*v; 2732 PetscInt i,nz; 2733 2734 PetscFunctionBegin; 2735 if (!n) PetscFunctionReturn(0); 2736 2737 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 2738 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 2739 2740 /* forward solve the lower triangular */ 2741 x[0] = b[0]; 2742 v = aa; 2743 vi = aj; 2744 for (i=1; i<n; i++) { 2745 nz = ai[i+1] - ai[i]; 2746 sum = b[i]; 2747 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 2748 v += nz; 2749 vi += nz; 2750 x[i] = sum; 2751 } 2752 2753 /* backward solve the upper triangular */ 2754 /* v = aa + ai[n+1]; 2755 vi = aj + ai[n+1]; */ 2756 v = aa + adiag[n-1]; 2757 vi = aj + adiag[n-1]; 2758 for (i=n-1; i>=0; i--){ 2759 nz = adiag[i] - adiag[i+1]-1; 2760 sum = x[i]; 2761 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 2762 v += nz; 2763 vi += nz; vi++; 2764 x[i] = *v++ *sum; /* x[i]=aa[adiag[i]]*sum; v++; */ 2765 } 2766 2767 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 2768 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 2769 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 2770 PetscFunctionReturn(0); 2771 } 2772 2773 #undef __FUNCT__ 2774 #define __FUNCT__ "MatSolve_SeqAIJ_newdatastruct" 2775 PetscErrorCode MatSolve_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec xx) 2776 { 2777 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2778 IS iscol = a->col,isrow = a->row; 2779 PetscErrorCode ierr; 2780 PetscInt i,n=A->rmap->n,*vi,*ai=a->i,*aj=a->j,nz,k; 2781 const PetscInt *rout,*cout,*r,*c; 2782 PetscScalar *x,*tmp,*tmps,sum; 2783 const PetscScalar *b; 2784 const MatScalar *aa = a->a,*v; 2785 2786 PetscFunctionBegin; 2787 if (!n) PetscFunctionReturn(0); 2788 2789 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 2790 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 2791 tmp = a->solve_work; 2792 2793 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 2794 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 2795 2796 /* forward solve the lower triangular */ 2797 tmp[0] = b[*r++]; 2798 tmps = tmp; 2799 v = aa; 2800 vi = aj; 2801 for (i=1; i<n; i++) { 2802 nz = ai[i+1] - ai[i]; 2803 sum = b[*r++]; 2804 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 2805 tmp[i] = sum; 2806 v += nz; vi += nz; 2807 } 2808 2809 /* backward solve the upper triangular */ 2810 k = n+1; 2811 v = aa + ai[k]; /* 1st entry of U(n-1,:) */ 2812 vi = aj + ai[k]; 2813 for (i=n-1; i>=0; i--){ 2814 k = 2*n-i; 2815 nz = ai[k +1] - ai[k] - 1; 2816 sum = tmp[i]; 2817 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 2818 x[c[i]] = tmp[i] = sum*v[nz]; /* v[nz] = aa[adiag[i]] */ 2819 v += nz+1; vi += nz+1; 2820 } 2821 2822 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 2823 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 2824 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 2825 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 2826 ierr = PetscLogFlops(2*a->nz - A->cmap->n);CHKERRQ(ierr); 2827 PetscFunctionReturn(0); 2828 } 2829 2830 #undef __FUNCT__ 2831 #define __FUNCT__ "MatSolve_SeqAIJ_newdatastruct_v2" 2832 PetscErrorCode MatSolve_SeqAIJ_newdatastruct_v2(Mat A,Vec bb,Vec xx) 2833 { 2834 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2835 IS iscol = a->col,isrow = a->row; 2836 PetscErrorCode ierr; 2837 PetscInt i,n=A->rmap->n,*vi,*ai=a->i,*aj=a->j,*adiag = a->diag,nz; 2838 const PetscInt *rout,*cout,*r,*c; 2839 PetscScalar *x,*tmp,sum; 2840 const PetscScalar *b; 2841 const MatScalar *aa = a->a,*v; 2842 2843 PetscFunctionBegin; 2844 if (!n) PetscFunctionReturn(0); 2845 2846 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 2847 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 2848 tmp = a->solve_work; 2849 2850 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 2851 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 2852 2853 /* forward solve the lower triangular */ 2854 /* tmp[0] = b[*r++]; */ 2855 tmp[0] = b[r[0]]; 2856 v = aa; 2857 vi = aj; 2858 for (i=1; i<n; i++) { 2859 nz = ai[i+1] - ai[i]; 2860 sum = b[r[i]]; 2861 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 2862 tmp[i] = sum; 2863 v += nz; vi += nz; 2864 } 2865 2866 /* backward solve the upper triangular */ 2867 /* v = aa + ai[k]; *//* 1st entry of U(n-1,:) */ 2868 /* vi = aj + ai[k]; */ 2869 v = aa + adiag[n-1]; 2870 vi = aj + adiag[n-1]; 2871 for (i=n-1; i>=0; i--){ 2872 /* nz = ai[k +1] - ai[k] - 1;*/ 2873 nz = adiag[i]-adiag[i+1]-1; 2874 sum = tmp[i]; 2875 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 2876 x[c[i]] = tmp[i] = sum*v[nz]; /* v[nz] = aa[adiag[i]] */ 2877 v += nz+1; vi += nz+1; 2878 } 2879 2880 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 2881 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 2882 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 2883 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 2884 ierr = PetscLogFlops(2*a->nz - A->cmap->n);CHKERRQ(ierr); 2885 PetscFunctionReturn(0); 2886 } 2887 2888 #undef __FUNCT__ 2889 #define __FUNCT__ "MatILUDTFactor_SeqAIJ" 2890 PetscErrorCode MatILUDTFactor_SeqAIJ(Mat A,IS isrow,IS iscol,const MatFactorInfo *info,Mat *fact) 2891 { 2892 Mat B = *fact; 2893 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b; 2894 IS isicol; 2895 PetscErrorCode ierr; 2896 const PetscInt *r,*ic; 2897 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j,*ajtmp,*adiag; 2898 PetscInt *bi,*bj,*bdiag,*bdiag_rev; 2899 PetscInt row,nzi,nzi_bl,nzi_bu,*im,nzi_al,nzi_au; 2900 PetscInt nlnk,*lnk; 2901 PetscBT lnkbt; 2902 PetscTruth row_identity,icol_identity,both_identity; 2903 MatScalar *aatmp,*pv,*batmp,*ba,*rtmp,*pc,multiplier,*vtmp,diag_tmp; 2904 const PetscInt *ics; 2905 PetscInt j,nz,*pj,*bjtmp,k,ncut,*jtmp; 2906 PetscReal dt=info->dt,dtcol=info->dtcol,shift=info->shiftinblocks; 2907 PetscInt dtcount=(PetscInt)info->dtcount,nnz_max; 2908 PetscTruth missing; 2909 2910 PetscFunctionBegin; 2911 2912 if (dt == PETSC_DEFAULT) dt = 0.005; 2913 if (dtcol == PETSC_DEFAULT) dtcol = 0.01; /* XXX unused! */ 2914 if (dtcount == PETSC_DEFAULT) dtcount = (PetscInt)(1.5*a->rmax); 2915 2916 /* ------- symbolic factorization, can be reused ---------*/ 2917 ierr = MatMissingDiagonal(A,&missing,&i);CHKERRQ(ierr); 2918 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",i); 2919 adiag=a->diag; 2920 2921 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 2922 2923 /* bdiag is location of diagonal in factor */ 2924 ierr = PetscMalloc2(n+1,PetscInt,&bdiag,n+1,PetscInt,&bdiag_rev);CHKERRQ(ierr); 2925 2926 /* allocate row pointers bi */ 2927 ierr = PetscMalloc((2*n+2)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 2928 2929 /* allocate bj and ba; max num of nonzero entries is (ai[n]+2*n*dtcount+2) */ 2930 if (dtcount > n-1) dtcount = n-1; /* diagonal is excluded */ 2931 nnz_max = ai[n]+2*n*dtcount+2; 2932 2933 ierr = PetscMalloc((nnz_max+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 2934 ierr = PetscMalloc((nnz_max+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 2935 2936 /* put together the new matrix */ 2937 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 2938 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 2939 b = (Mat_SeqAIJ*)(B)->data; 2940 b->free_a = PETSC_TRUE; 2941 b->free_ij = PETSC_TRUE; 2942 b->singlemalloc = PETSC_FALSE; 2943 b->a = ba; 2944 b->j = bj; 2945 b->i = bi; 2946 b->diag = bdiag; 2947 b->ilen = 0; 2948 b->imax = 0; 2949 b->row = isrow; 2950 b->col = iscol; 2951 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2952 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2953 b->icol = isicol; 2954 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2955 2956 ierr = PetscLogObjectMemory(B,nnz_max*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2957 b->maxnz = nnz_max; 2958 2959 (B)->factor = MAT_FACTOR_ILUDT; 2960 (B)->info.factor_mallocs = 0; 2961 (B)->info.fill_ratio_given = ((PetscReal)nnz_max)/((PetscReal)ai[n]); 2962 CHKMEMQ; 2963 /* ------- end of symbolic factorization ---------*/ 2964 2965 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 2966 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 2967 ics = ic; 2968 2969 /* linked list for storing column indices of the active row */ 2970 nlnk = n + 1; 2971 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2972 2973 /* im: used by PetscLLAddSortedLU(); jtmp: working array for column indices of active row */ 2974 ierr = PetscMalloc2(n,PetscInt,&im,n,PetscInt,&jtmp);CHKERRQ(ierr); 2975 /* rtmp, vtmp: working arrays for sparse and contiguous row entries of active row */ 2976 ierr = PetscMalloc2(n,MatScalar,&rtmp,n,MatScalar,&vtmp);CHKERRQ(ierr); 2977 ierr = PetscMemzero(rtmp,n*sizeof(MatScalar));CHKERRQ(ierr); 2978 2979 bi[0] = 0; 2980 bdiag[0] = nnz_max-1; /* location of diag[0] in factor B */ 2981 bdiag_rev[n] = bdiag[0]; 2982 bi[2*n+1] = bdiag[0]+1; /* endof bj and ba array */ 2983 for (i=0; i<n; i++) { 2984 /* copy initial fill into linked list */ 2985 nzi = 0; /* nonzeros for active row i */ 2986 nzi = ai[r[i]+1] - ai[r[i]]; 2987 if (!nzi) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 2988 nzi_al = adiag[r[i]] - ai[r[i]]; 2989 nzi_au = ai[r[i]+1] - adiag[r[i]] -1; 2990 ajtmp = aj + ai[r[i]]; 2991 ierr = PetscLLAddPerm(nzi,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 2992 2993 /* load in initial (unfactored row) */ 2994 aatmp = a->a + ai[r[i]]; 2995 for (j=0; j<nzi; j++) { 2996 rtmp[ics[*ajtmp++]] = *aatmp++; 2997 } 2998 2999 /* add pivot rows into linked list */ 3000 row = lnk[n]; 3001 while (row < i ) { 3002 nzi_bl = bi[row+1] - bi[row] + 1; 3003 bjtmp = bj + bdiag[row+1]+1; /* points to 1st column next to the diagonal in U */ 3004 ierr = PetscLLAddSortedLU(bjtmp,row,nlnk,lnk,lnkbt,i,nzi_bl,im);CHKERRQ(ierr); 3005 nzi += nlnk; 3006 row = lnk[row]; 3007 } 3008 3009 /* copy data from lnk into jtmp, then initialize lnk */ 3010 ierr = PetscLLClean(n,n,nzi,lnk,jtmp,lnkbt);CHKERRQ(ierr); 3011 3012 /* numerical factorization */ 3013 bjtmp = jtmp; 3014 row = *bjtmp++; /* 1st pivot row */ 3015 while ( row < i ) { 3016 pc = rtmp + row; 3017 pv = ba + bdiag[row]; /* 1./(diag of the pivot row) */ 3018 multiplier = (*pc) * (*pv); 3019 *pc = multiplier; 3020 if (PetscAbsScalar(*pc) > dt){ /* apply tolerance dropping rule */ 3021 pj = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */ 3022 pv = ba + bdiag[row+1] + 1; 3023 /* if (multiplier < -1.0 or multiplier >1.0) printf("row/prow %d, %d, multiplier %g\n",i,row,multiplier); */ 3024 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */ 3025 for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++); 3026 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 3027 } 3028 row = *bjtmp++; 3029 } 3030 3031 /* copy sparse rtmp into contiguous vtmp; separate L and U part */ 3032 diag_tmp = rtmp[i]; /* save diagonal value - may not needed?? */ 3033 nzi_bl = 0; j = 0; 3034 while (jtmp[j] < i){ /* Note: jtmp is sorted */ 3035 vtmp[j] = rtmp[jtmp[j]]; rtmp[jtmp[j]]=0.0; 3036 nzi_bl++; j++; 3037 } 3038 nzi_bu = nzi - nzi_bl -1; 3039 while (j < nzi){ 3040 vtmp[j] = rtmp[jtmp[j]]; rtmp[jtmp[j]]=0.0; 3041 j++; 3042 } 3043 3044 bjtmp = bj + bi[i]; 3045 batmp = ba + bi[i]; 3046 /* apply level dropping rule to L part */ 3047 ncut = nzi_al + dtcount; 3048 if (ncut < nzi_bl){ 3049 ierr = PetscSortSplit(ncut,nzi_bl,vtmp,jtmp);CHKERRQ(ierr); 3050 ierr = PetscSortIntWithScalarArray(ncut,jtmp,vtmp);CHKERRQ(ierr); 3051 } else { 3052 ncut = nzi_bl; 3053 } 3054 for (j=0; j<ncut; j++){ 3055 bjtmp[j] = jtmp[j]; 3056 batmp[j] = vtmp[j]; 3057 /* printf(" (%d,%g),",bjtmp[j],batmp[j]); */ 3058 } 3059 bi[i+1] = bi[i] + ncut; 3060 nzi = ncut + 1; 3061 3062 /* apply level dropping rule to U part */ 3063 ncut = nzi_au + dtcount; 3064 if (ncut < nzi_bu){ 3065 ierr = PetscSortSplit(ncut,nzi_bu,vtmp+nzi_bl+1,jtmp+nzi_bl+1);CHKERRQ(ierr); 3066 ierr = PetscSortIntWithScalarArray(ncut,jtmp+nzi_bl+1,vtmp+nzi_bl+1);CHKERRQ(ierr); 3067 } else { 3068 ncut = nzi_bu; 3069 } 3070 nzi += ncut; 3071 3072 /* mark bdiagonal */ 3073 bdiag[i+1] = bdiag[i] - (ncut + 1); 3074 bdiag_rev[n-i-1] = bdiag[i+1]; 3075 bi[2*n - i] = bi[2*n - i +1] - (ncut + 1); 3076 bjtmp = bj + bdiag[i]; 3077 batmp = ba + bdiag[i]; 3078 *bjtmp = i; 3079 *batmp = diag_tmp; /* rtmp[i]; */ 3080 if (*batmp == 0.0) { 3081 *batmp = dt+shift; 3082 /* printf(" row %d add shift %g\n",i,shift); */ 3083 } 3084 *batmp = 1.0/(*batmp); /* invert diagonal entries for simplier triangular solves */ 3085 /* printf(" (%d,%g),",*bjtmp,*batmp); */ 3086 3087 bjtmp = bj + bdiag[i+1]+1; 3088 batmp = ba + bdiag[i+1]+1; 3089 for (k=0; k<ncut; k++){ 3090 bjtmp[k] = jtmp[nzi_bl+1+k]; 3091 batmp[k] = vtmp[nzi_bl+1+k]; 3092 /* printf(" (%d,%g),",bjtmp[k],batmp[k]); */ 3093 } 3094 /* printf("\n"); */ 3095 3096 im[i] = nzi; /* used by PetscLLAddSortedLU() */ 3097 /* 3098 printf("row %d: bi %d, bdiag %d\n",i,bi[i],bdiag[i]); 3099 printf(" ----------------------------\n"); 3100 */ 3101 } /* for (i=0; i<n; i++) */ 3102 /* printf("end of L %d, beginning of U %d\n",bi[n],bdiag[n]); */ 3103 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]); 3104 3105 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 3106 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 3107 3108 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 3109 ierr = PetscFree2(im,jtmp);CHKERRQ(ierr); 3110 ierr = PetscFree2(rtmp,vtmp);CHKERRQ(ierr); 3111 3112 ierr = PetscLogFlops(B->cmap->n);CHKERRQ(ierr); 3113 b->maxnz = b->nz = bi[n] + bdiag[0] - bdiag[n]; 3114 3115 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 3116 ierr = ISIdentity(isicol,&icol_identity);CHKERRQ(ierr); 3117 both_identity = (PetscTruth) (row_identity && icol_identity); 3118 if (row_identity && icol_identity) { 3119 B->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 3120 } else { 3121 B->ops->solve = MatSolve_SeqAIJ_newdatastruct; 3122 } 3123 3124 B->ops->lufactorsymbolic = MatILUDTFactorSymbolic_SeqAIJ; 3125 B->ops->lufactornumeric = MatILUDTFactorNumeric_SeqAIJ; 3126 B->ops->solveadd = 0; 3127 B->ops->solvetranspose = 0; 3128 B->ops->solvetransposeadd = 0; 3129 B->ops->matsolve = 0; 3130 B->assembled = PETSC_TRUE; 3131 B->preallocated = PETSC_TRUE; 3132 PetscFunctionReturn(0); 3133 } 3134 3135 /* a wraper of MatILUDTFactor_SeqAIJ() */ 3136 #undef __FUNCT__ 3137 #define __FUNCT__ "MatILUDTFactorSymbolic_SeqAIJ" 3138 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS row,IS col,const MatFactorInfo *info) 3139 { 3140 PetscErrorCode ierr; 3141 3142 PetscFunctionBegin; 3143 ierr = MatILUDTFactor_SeqAIJ(A,row,col,info,&fact);CHKERRQ(ierr); 3144 3145 fact->ops->lufactornumeric = MatILUDTFactorNumeric_SeqAIJ; 3146 PetscFunctionReturn(0); 3147 } 3148 3149 /* 3150 same as MatLUFactorNumeric_SeqAIJ(), except using contiguous array matrix factors 3151 - intend to replace existing MatLUFactorNumeric_SeqAIJ() 3152 */ 3153 #undef __FUNCT__ 3154 #define __FUNCT__ "MatILUDTFactorNumeric_SeqAIJ" 3155 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorNumeric_SeqAIJ(Mat fact,Mat A,const MatFactorInfo *info) 3156 { 3157 Mat C=fact; 3158 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 3159 IS isrow = b->row,isicol = b->icol; 3160 PetscErrorCode ierr; 3161 const PetscInt *r,*ic,*ics; 3162 PetscInt i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 3163 PetscInt *ajtmp,*bjtmp,nz,nzl,nzu,row,*bdiag = b->diag,*pj; 3164 MatScalar *rtmp,*pc,multiplier,*v,*pv,*aa=a->a; 3165 PetscReal dt=info->dt,shift=info->shiftinblocks; 3166 PetscTruth row_identity, col_identity; 3167 3168 PetscFunctionBegin; 3169 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 3170 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 3171 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 3172 ics = ic; 3173 3174 for (i=0; i<n; i++){ 3175 /* initialize rtmp array */ 3176 nzl = bi[i+1] - bi[i]; /* num of nozeros in L(i,:) */ 3177 bjtmp = bj + bi[i]; 3178 for (j=0; j<nzl; j++) rtmp[*bjtmp++] = 0.0; 3179 rtmp[i] = 0.0; 3180 nzu = bdiag[i] - bdiag[i+1]; /* num of nozeros in U(i,:) */ 3181 bjtmp = bj + bdiag[i+1] + 1; 3182 for (j=0; j<nzu; j++) rtmp[*bjtmp++] = 0.0; 3183 3184 /* load in initial unfactored row of A */ 3185 /* printf("row %d\n",i); */ 3186 nz = ai[r[i]+1] - ai[r[i]]; 3187 ajtmp = aj + ai[r[i]]; 3188 v = aa + ai[r[i]]; 3189 for (j=0; j<nz; j++) { 3190 rtmp[ics[*ajtmp++]] = v[j]; 3191 /* printf(" (%d,%g),",ics[ajtmp[j]],rtmp[ics[ajtmp[j]]]); */ 3192 } 3193 /* printf("\n"); */ 3194 3195 /* numerical factorization */ 3196 bjtmp = bj + bi[i]; /* point to 1st entry of L(i,:) */ 3197 nzl = bi[i+1] - bi[i]; /* num of entries in L(i,:) */ 3198 k = 0; 3199 while (k < nzl){ 3200 row = *bjtmp++; 3201 /* printf(" prow %d\n",row); */ 3202 pc = rtmp + row; 3203 pv = b->a + bdiag[row]; /* 1./(diag of the pivot row) */ 3204 multiplier = (*pc) * (*pv); 3205 *pc = multiplier; 3206 if (PetscAbsScalar(multiplier) > dt){ 3207 pj = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */ 3208 pv = b->a + bdiag[row+1] + 1; 3209 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */ 3210 for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++); 3211 /* ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); */ 3212 } 3213 k++; 3214 } 3215 3216 /* finished row so stick it into b->a */ 3217 /* L-part */ 3218 pv = b->a + bi[i] ; 3219 pj = bj + bi[i] ; 3220 nzl = bi[i+1] - bi[i]; 3221 for (j=0; j<nzl; j++) { 3222 pv[j] = rtmp[pj[j]]; 3223 /* printf(" (%d,%g),",pj[j],pv[j]); */ 3224 } 3225 3226 /* diagonal: invert diagonal entries for simplier triangular solves */ 3227 if (rtmp[i] == 0.0) rtmp[i] = dt+shift; 3228 b->a[bdiag[i]] = 1.0/rtmp[i]; 3229 /* printf(" (%d,%g),",i,b->a[bdiag[i]]); */ 3230 3231 /* U-part */ 3232 pv = b->a + bdiag[i+1] + 1; 3233 pj = bj + bdiag[i+1] + 1; 3234 nzu = bdiag[i] - bdiag[i+1] - 1; 3235 for (j=0; j<nzu; j++) { 3236 pv[j] = rtmp[pj[j]]; 3237 /* printf(" (%d,%g),",pj[j],pv[j]); */ 3238 } 3239 /* printf("\n"); */ 3240 } 3241 3242 ierr = PetscFree(rtmp);CHKERRQ(ierr); 3243 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 3244 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 3245 3246 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 3247 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 3248 if (row_identity && col_identity) { 3249 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 3250 } else { 3251 C->ops->solve = MatSolve_SeqAIJ_newdatastruct; 3252 } 3253 C->ops->solveadd = 0; 3254 C->ops->solvetranspose = 0; 3255 C->ops->solvetransposeadd = 0; 3256 C->ops->matsolve = 0; 3257 C->assembled = PETSC_TRUE; 3258 C->preallocated = PETSC_TRUE; 3259 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 3260 PetscFunctionReturn(0); 3261 } 3262