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 = PetscMalloc(n*sizeof(PetscInt),&order);CHKERRQ(ierr); 42 ierr = PetscMemzero(done,n*sizeof(PetscTruth));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 PetscTruth newdatastruct_v2 = PETSC_FALSE; 150 151 PetscFunctionBegin; 152 ierr = PetscOptionsGetTruth(PETSC_NULL,"-lu_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 153 if(newdatastruct){ 154 ierr = MatLUFactorSymbolic_SeqAIJ_newdatastruct(B,A,isrow,iscol,info);CHKERRQ(ierr); 155 PetscFunctionReturn(0); 156 } 157 ierr = PetscOptionsGetTruth(PETSC_NULL,"-lu_new_v2",&newdatastruct_v2,PETSC_NULL);CHKERRQ(ierr); 158 if(newdatastruct_v2){ 159 ierr = MatLUFactorSymbolic_SeqAIJ_newdatastruct_v2(B,A,isrow,iscol,info);CHKERRQ(ierr); 160 PetscFunctionReturn(0); 161 } 162 163 164 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 165 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 166 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 167 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 168 169 /* get new row pointers */ 170 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 171 bi[0] = 0; 172 173 /* bdiag is location of diagonal in factor */ 174 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 175 bdiag[0] = 0; 176 177 /* linked list for storing column indices of the active row */ 178 nlnk = n + 1; 179 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 180 181 ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr); 182 183 /* initial FreeSpace size is f*(ai[n]+1) */ 184 f = info->fill; 185 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 186 current_space = free_space; 187 188 for (i=0; i<n; i++) { 189 /* copy previous fill into linked list */ 190 nzi = 0; 191 nnz = ai[r[i]+1] - ai[r[i]]; 192 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 193 ajtmp = aj + ai[r[i]]; 194 ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 195 nzi += nlnk; 196 197 /* add pivot rows into linked list */ 198 row = lnk[n]; 199 while (row < i) { 200 nzbd = bdiag[row] - bi[row] + 1; /* num of entries in the row with column index <= row */ 201 ajtmp = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */ 202 ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr); 203 nzi += nlnk; 204 row = lnk[row]; 205 } 206 bi[i+1] = bi[i] + nzi; 207 im[i] = nzi; 208 209 /* mark bdiag */ 210 nzbd = 0; 211 nnz = nzi; 212 k = lnk[n]; 213 while (nnz-- && k < i){ 214 nzbd++; 215 k = lnk[k]; 216 } 217 bdiag[i] = bi[i] + nzbd; 218 219 /* if free space is not available, make more free space */ 220 if (current_space->local_remaining<nzi) { 221 nnz = (n - i)*nzi; /* estimated and max additional space needed */ 222 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 223 reallocs++; 224 } 225 226 /* copy data into free space, then initialize lnk */ 227 ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 228 bi_ptr[i] = current_space->array; 229 current_space->array += nzi; 230 current_space->local_used += nzi; 231 current_space->local_remaining -= nzi; 232 } 233 #if defined(PETSC_USE_INFO) 234 if (ai[n] != 0) { 235 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 236 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 237 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 238 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr); 239 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 240 } else { 241 ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr); 242 } 243 #endif 244 245 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 246 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 247 248 /* destroy list of free space and other temporary array(s) */ 249 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 250 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 251 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 252 ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr); 253 254 /* put together the new matrix */ 255 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 256 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 257 b = (Mat_SeqAIJ*)(B)->data; 258 b->free_a = PETSC_TRUE; 259 b->free_ij = PETSC_TRUE; 260 b->singlemalloc = PETSC_FALSE; 261 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 262 b->j = bj; 263 b->i = bi; 264 b->diag = bdiag; 265 b->ilen = 0; 266 b->imax = 0; 267 b->row = isrow; 268 b->col = iscol; 269 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 270 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 271 b->icol = isicol; 272 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 273 274 /* In b structure: Free imax, ilen, old a, old j. Allocate solve_work, new a, new j */ 275 ierr = PetscLogObjectMemory(B,(bi[n]-n)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 276 b->maxnz = b->nz = bi[n] ; 277 278 (B)->factor = MAT_FACTOR_LU; 279 (B)->info.factor_mallocs = reallocs; 280 (B)->info.fill_ratio_given = f; 281 282 if (ai[n]) { 283 (B)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]); 284 } else { 285 (B)->info.fill_ratio_needed = 0.0; 286 } 287 (B)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ; 288 (B)->ops->solve = MatSolve_SeqAIJ; 289 (B)->ops->solvetranspose = MatSolveTranspose_SeqAIJ; 290 /* switch to inodes if appropriate */ 291 ierr = MatLUFactorSymbolic_Inode(B,A,isrow,iscol,info);CHKERRQ(ierr); 292 PetscFunctionReturn(0); 293 } 294 295 #undef __FUNCT__ 296 #define __FUNCT__ "MatLUFactorSymbolic_SeqAIJ_newdatastruct" 297 PetscErrorCode MatLUFactorSymbolic_SeqAIJ_newdatastruct(Mat B,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 298 { 299 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 300 IS isicol; 301 PetscErrorCode ierr; 302 const PetscInt *r,*ic; 303 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j; 304 PetscInt *bi,*bj,*ajtmp; 305 PetscInt *bdiag,row,nnz,nzi,reallocs=0,nzbd,*im; 306 PetscReal f; 307 PetscInt nlnk,*lnk,k,**bi_ptr; 308 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 309 PetscBT lnkbt; 310 311 PetscFunctionBegin; 312 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 313 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 314 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 315 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 316 317 /* get new row pointers */ 318 ierr = PetscMalloc((2*n+2)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 319 bi[0] = 0; 320 321 /* bdiag is location of diagonal in factor */ 322 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 323 bdiag[0] = 0; 324 325 /* linked list for storing column indices of the active row */ 326 nlnk = n + 1; 327 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 328 329 ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr); 330 331 /* initial FreeSpace size is f*(ai[n]+1) */ 332 f = info->fill; 333 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 334 current_space = free_space; 335 336 for (i=0; i<n; i++) { 337 /* copy previous fill into linked list */ 338 nzi = 0; 339 nnz = ai[r[i]+1] - ai[r[i]]; 340 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 341 ajtmp = aj + ai[r[i]]; 342 ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 343 nzi += nlnk; 344 345 /* add pivot rows into linked list */ 346 row = lnk[n]; 347 while (row < i){ 348 nzbd = bdiag[row] + 1; /* num of entries in the row with column index <= row */ 349 ajtmp = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */ 350 ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr); 351 nzi += nlnk; 352 row = lnk[row]; 353 } 354 bi[i+1] = bi[i] + nzi; 355 im[i] = nzi; 356 357 /* mark bdiag */ 358 nzbd = 0; 359 nnz = nzi; 360 k = lnk[n]; 361 while (nnz-- && k < i){ 362 nzbd++; 363 k = lnk[k]; 364 } 365 bdiag[i] = nzbd; /* note: bdiag[i] = nnzL as input for PetscFreeSpaceContiguous_LU() */ 366 367 /* if free space is not available, make more free space */ 368 if (current_space->local_remaining<nzi) { 369 nnz = 2*(n - i)*nzi; /* estimated and max additional space needed */ 370 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 371 reallocs++; 372 } 373 374 /* copy data into free space, then initialize lnk */ 375 ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 376 bi_ptr[i] = current_space->array; 377 current_space->array += nzi; 378 current_space->local_used += nzi; 379 current_space->local_remaining -= nzi; 380 } 381 #if defined(PETSC_USE_INFO) 382 if (ai[n] != 0) { 383 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 384 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 385 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 386 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr); 387 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 388 } else { 389 ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr); 390 } 391 #endif 392 393 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 394 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 395 396 /* destroy list of free space and other temporary array(s) */ 397 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 398 ierr = PetscFreeSpaceContiguous_LU(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 399 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 400 ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr); 401 402 /* put together the new matrix */ 403 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 404 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 405 b = (Mat_SeqAIJ*)(B)->data; 406 b->free_a = PETSC_TRUE; 407 b->free_ij = PETSC_TRUE; 408 b->singlemalloc = PETSC_FALSE; 409 ierr = PetscMalloc((bi[2*n+1])*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 410 b->j = bj; 411 b->i = bi; 412 b->diag = bdiag; 413 b->ilen = 0; 414 b->imax = 0; 415 b->row = isrow; 416 b->col = iscol; 417 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 418 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 419 b->icol = isicol; 420 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 421 422 /* In b structure: Free imax, ilen, old a, old j. Allocate solve_work, new a, new j */ 423 ierr = PetscLogObjectMemory(B,bi[2*n+1]*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 424 b->maxnz = b->nz = bi[2*n+1] ; 425 426 (B)->factor = MAT_FACTOR_LU; 427 (B)->info.factor_mallocs = reallocs; 428 (B)->info.fill_ratio_given = f; 429 430 if (ai[n]) { 431 (B)->info.fill_ratio_needed = ((PetscReal)bi[2*n+1])/((PetscReal)ai[n]); 432 } else { 433 (B)->info.fill_ratio_needed = 0.0; 434 } 435 (B)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 436 PetscFunctionReturn(0); 437 } 438 439 #undef __FUNCT__ 440 #define __FUNCT__ "MatLUFactorSymbolic_SeqAIJ_newdatastruct_v2" 441 PetscErrorCode MatLUFactorSymbolic_SeqAIJ_newdatastruct_v2(Mat B,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 442 { 443 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 444 IS isicol; 445 PetscErrorCode ierr; 446 const PetscInt *r,*ic; 447 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j; 448 PetscInt *bi,*bj,*ajtmp; 449 PetscInt *bdiag,row,nnz,nzi,reallocs=0,nzbd,*im; 450 PetscReal f; 451 PetscInt nlnk,*lnk,k,**bi_ptr; 452 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 453 PetscBT lnkbt; 454 455 PetscFunctionBegin; 456 if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 457 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 458 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 459 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 460 461 /* get new row pointers */ 462 /* ierr = PetscMalloc((2*n+2)*sizeof(PetscInt),&bi);CHKERRQ(ierr); */ 463 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 464 bi[0] = 0; 465 466 /* bdiag is location of diagonal in factor */ 467 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 468 bdiag[0] = 0; 469 470 /* linked list for storing column indices of the active row */ 471 nlnk = n + 1; 472 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 473 474 ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr); 475 476 /* initial FreeSpace size is f*(ai[n]+1) */ 477 f = info->fill; 478 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 479 current_space = free_space; 480 481 for (i=0; i<n; i++) { 482 /* copy previous fill into linked list */ 483 nzi = 0; 484 nnz = ai[r[i]+1] - ai[r[i]]; 485 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 486 ajtmp = aj + ai[r[i]]; 487 ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 488 nzi += nlnk; 489 490 /* add pivot rows into linked list */ 491 row = lnk[n]; 492 while (row < i){ 493 nzbd = bdiag[row] + 1; /* num of entries in the row with column index <= row */ 494 ajtmp = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */ 495 ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr); 496 nzi += nlnk; 497 row = lnk[row]; 498 } 499 bi[i+1] = bi[i] + nzi; 500 im[i] = nzi; 501 502 /* mark bdiag */ 503 nzbd = 0; 504 nnz = nzi; 505 k = lnk[n]; 506 while (nnz-- && k < i){ 507 nzbd++; 508 k = lnk[k]; 509 } 510 bdiag[i] = nzbd; /* note: bdiag[i] = nnzL as input for PetscFreeSpaceContiguous_LU() */ 511 512 /* if free space is not available, make more free space */ 513 if (current_space->local_remaining<nzi) { 514 nnz = 2*(n - i)*nzi; /* estimated and max additional space needed */ 515 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 516 reallocs++; 517 } 518 519 /* copy data into free space, then initialize lnk */ 520 ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 521 bi_ptr[i] = current_space->array; 522 current_space->array += nzi; 523 current_space->local_used += nzi; 524 current_space->local_remaining -= nzi; 525 } 526 #if defined(PETSC_USE_INFO) 527 if (ai[n] != 0) { 528 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 529 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 530 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 531 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr); 532 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 533 } else { 534 ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr); 535 } 536 #endif 537 538 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 539 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 540 541 /* destroy list of free space and other temporary array(s) */ 542 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 543 /* ierr = PetscFreeSpaceContiguous_LU(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); */ 544 ierr = PetscFreeSpaceContiguous_LU_v2(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 545 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 546 ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr); 547 548 /* put together the new matrix */ 549 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 550 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 551 b = (Mat_SeqAIJ*)(B)->data; 552 b->free_a = PETSC_TRUE; 553 b->free_ij = PETSC_TRUE; 554 b->singlemalloc = PETSC_FALSE; 555 /*ierr = PetscMalloc((bi[2*n+1])*sizeof(PetscScalar),&b->a);CHKERRQ(ierr);*/ 556 ierr = PetscMalloc((bdiag[0]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 557 b->j = bj; 558 b->i = bi; 559 b->diag = bdiag; 560 b->ilen = 0; 561 b->imax = 0; 562 b->row = isrow; 563 b->col = iscol; 564 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 565 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 566 b->icol = isicol; 567 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 568 569 /* In b structure: Free imax, ilen, old a, old j. Allocate solve_work, new a, new j */ 570 /* ierr = PetscLogObjectMemory(B,bi[2*n+1]*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 571 b->maxnz = b->nz = bi[2*n+1] ; */ 572 ierr = PetscLogObjectMemory(B,(bdiag[0]+1)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 573 b->maxnz = b->nz = bdiag[0]+1; 574 (B)->factor = MAT_FACTOR_LU; 575 (B)->info.factor_mallocs = reallocs; 576 (B)->info.fill_ratio_given = f; 577 578 if (ai[n]) { 579 /* (B)->info.fill_ratio_needed = ((PetscReal)bi[2*n+1])/((PetscReal)ai[n]); */ 580 (B)->info.fill_ratio_needed = ((PetscReal)(bdiag[0]+1))/((PetscReal)ai[n]); 581 } else { 582 (B)->info.fill_ratio_needed = 0.0; 583 } 584 /* (B)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct;*/ 585 (B)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct_v2; 586 PetscFunctionReturn(0); 587 } 588 589 /* 590 Trouble in factorization, should we dump the original matrix? 591 */ 592 #undef __FUNCT__ 593 #define __FUNCT__ "MatFactorDumpMatrix" 594 PetscErrorCode MatFactorDumpMatrix(Mat A) 595 { 596 PetscErrorCode ierr; 597 PetscTruth flg = PETSC_FALSE; 598 599 PetscFunctionBegin; 600 ierr = PetscOptionsGetTruth(PETSC_NULL,"-mat_factor_dump_on_error",&flg,PETSC_NULL);CHKERRQ(ierr); 601 if (flg) { 602 PetscViewer viewer; 603 char filename[PETSC_MAX_PATH_LEN]; 604 605 ierr = PetscSNPrintf(filename,PETSC_MAX_PATH_LEN,"matrix_factor_error.%d",PetscGlobalRank);CHKERRQ(ierr); 606 ierr = PetscViewerBinaryOpen(((PetscObject)A)->comm,filename,FILE_MODE_WRITE,&viewer);CHKERRQ(ierr); 607 ierr = MatView(A,viewer);CHKERRQ(ierr); 608 ierr = PetscViewerDestroy(viewer);CHKERRQ(ierr); 609 } 610 PetscFunctionReturn(0); 611 } 612 613 extern PetscErrorCode MatSolve_Inode(Mat,Vec,Vec); 614 615 /* ----------------------------------------------------------- */ 616 extern PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_newdatastruct(Mat,Vec,Vec); 617 extern PetscErrorCode MatSolve_SeqAIJ_newdatastruct(Mat,Vec,Vec); 618 619 #undef __FUNCT__ 620 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_newdatastruct" 621 PetscErrorCode MatLUFactorNumeric_SeqAIJ_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 622 { 623 Mat C=B; 624 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 625 IS isrow = b->row,isicol = b->icol; 626 PetscErrorCode ierr; 627 const PetscInt *r,*ic,*ics; 628 PetscInt i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 629 PetscInt *ajtmp,*bjtmp,nz,nzL,row,*bdiag=b->diag,*pj; 630 MatScalar *rtmp,*pc,multiplier,*v,*pv,*aa=a->a; 631 PetscTruth row_identity,col_identity; 632 633 LUShift_Ctx sctx; 634 PetscInt *ddiag,newshift; 635 PetscReal rs; 636 MatScalar d; 637 638 PetscFunctionBegin; 639 /* MatPivotSetUp(): initialize shift context sctx */ 640 ierr = PetscMemzero(&sctx,sizeof(LUShift_Ctx));CHKERRQ(ierr); 641 642 /* if both shift schemes are chosen by user, only use info->shiftpd */ 643 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 644 ddiag = a->diag; 645 sctx.shift_top = info->zeropivot; 646 for (i=0; i<n; i++) { 647 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 648 d = (aa)[ddiag[i]]; 649 rs = -PetscAbsScalar(d) - PetscRealPart(d); 650 v = aa+ai[i]; 651 nz = ai[i+1] - ai[i]; 652 for (j=0; j<nz; j++) 653 rs += PetscAbsScalar(v[j]); 654 if (rs>sctx.shift_top) sctx.shift_top = rs; 655 } 656 sctx.shift_top *= 1.1; 657 sctx.nshift_max = 5; 658 sctx.shift_lo = 0.; 659 sctx.shift_hi = 1.; 660 } 661 662 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 663 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 664 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 665 ics = ic; 666 667 do { 668 sctx.lushift = PETSC_FALSE; 669 for (i=0; i<n; i++){ 670 /* zero rtmp */ 671 /* L part */ 672 nz = bi[i+1] - bi[i]; 673 bjtmp = bj + bi[i]; 674 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 675 676 /* U part */ 677 nz = bi[2*n-i+1] - bi[2*n-i]; 678 bjtmp = bj + bi[2*n-i]; 679 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 680 681 /* load in initial (unfactored row) */ 682 nz = ai[r[i]+1] - ai[r[i]]; 683 ajtmp = aj + ai[r[i]]; 684 v = aa + ai[r[i]]; 685 for (j=0; j<nz; j++) { 686 rtmp[ics[ajtmp[j]]] = v[j]; 687 } 688 /* ZeropivotApply() */ 689 rtmp[ics[r[i]]] += sctx.shift_amount; /* shift the diagonal of the matrix */ 690 691 /* elimination */ 692 bjtmp = bj + bi[i]; 693 row = *bjtmp++; 694 nzL = bi[i+1] - bi[i]; 695 k = 0; 696 while (k < nzL) { 697 pc = rtmp + row; 698 if (*pc != 0.0) { 699 pv = b->a + bdiag[row]; 700 multiplier = *pc * (*pv); 701 *pc = multiplier; 702 pj = b->j + bi[2*n-row]; /* begining of U(row,:) */ 703 pv = b->a + bi[2*n-row]; 704 nz = bi[2*n-row+1] - bi[2*n-row] - 1; /* num of entries in U(row,:), excluding diag */ 705 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 706 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 707 } 708 row = *bjtmp++; k++; 709 } 710 711 /* finished row so stick it into b->a */ 712 rs = 0.0; 713 /* L part */ 714 pv = b->a + bi[i] ; 715 pj = b->j + bi[i] ; 716 nz = bi[i+1] - bi[i]; 717 for (j=0; j<nz; j++) { 718 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 719 } 720 721 /* U part */ 722 pv = b->a + bi[2*n-i]; 723 pj = b->j + bi[2*n-i]; 724 nz = bi[2*n-i+1] - bi[2*n-i] - 1; 725 for (j=0; j<nz; j++) { 726 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 727 } 728 729 /* MatPivotCheck() */ 730 sctx.rs = rs; 731 sctx.pv = rtmp[i]; 732 /* ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); */ 733 if (info->shiftnz){ 734 ierr = MatPivotCheck_nz(info,sctx,i,newshift);CHKERRQ(ierr); 735 } else if (info->shiftpd){ 736 ierr = MatPivotCheck_pd(info,sctx,i,newshift);CHKERRQ(ierr); 737 } else if (info->shiftinblocks){ 738 ierr = MatPivotCheck_inblocks(info,sctx,i,newshift);CHKERRQ(ierr); 739 } else { 740 ierr = MatPivotCheck_none(info,sctx,i,newshift);CHKERRQ(ierr); 741 } 742 rtmp[i] = sctx.pv; 743 if (newshift == 1) break; 744 745 /* Mark diagonal and invert diagonal for simplier triangular solves */ 746 pv = b->a + bdiag[i]; 747 *pv = 1.0/rtmp[i]; 748 749 } /* endof for (i=0; i<n; i++){ */ 750 751 /* MatPivotRefine() */ 752 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max){ 753 /* 754 * if no shift in this attempt & shifting & started shifting & can refine, 755 * then try lower shift 756 */ 757 sctx.shift_hi = sctx.shift_fraction; 758 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 759 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 760 sctx.lushift = PETSC_TRUE; 761 sctx.nshift++; 762 } 763 } while (sctx.lushift); 764 765 ierr = PetscFree(rtmp);CHKERRQ(ierr); 766 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 767 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 768 769 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 770 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 771 if (row_identity && col_identity) { 772 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 773 } else { 774 C->ops->solve = MatSolve_SeqAIJ_newdatastruct; 775 } 776 777 C->ops->solveadd = 0; 778 C->ops->solvetranspose = 0; 779 C->ops->solvetransposeadd = 0; 780 C->ops->matsolve = 0; 781 C->assembled = PETSC_TRUE; 782 C->preallocated = PETSC_TRUE; 783 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 784 785 /* MatPivotView() */ 786 if (sctx.nshift){ 787 if (info->shiftpd) { 788 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); 789 } else if (info->shiftnz) { 790 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 791 } 792 } 793 PetscFunctionReturn(0); 794 } 795 796 /* ----------------------------------------------------------- */ 797 extern PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_newdatastruct_v2(Mat,Vec,Vec); 798 extern PetscErrorCode MatSolve_SeqAIJ_newdatastruct_v2(Mat,Vec,Vec); 799 800 #undef __FUNCT__ 801 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_newdatastruct_v2" 802 PetscErrorCode MatLUFactorNumeric_SeqAIJ_newdatastruct_v2(Mat B,Mat A,const MatFactorInfo *info) 803 { 804 Mat C=B; 805 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 806 IS isrow = b->row,isicol = b->icol; 807 PetscErrorCode ierr; 808 const PetscInt *r,*ic,*ics; 809 PetscInt i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 810 PetscInt *ajtmp,*bjtmp,nz,nzL,row,*bdiag=b->diag,*pj; 811 MatScalar *rtmp,*pc,multiplier,*v,*pv,*aa=a->a; 812 PetscTruth row_identity,col_identity; 813 814 LUShift_Ctx sctx; 815 PetscInt *ddiag,newshift; 816 PetscReal rs; 817 MatScalar d; 818 819 PetscFunctionBegin; 820 /* ZeropivotSetUp(): initialize shift context sctx */ 821 sctx.nshift = 0; 822 sctx.nshift_max = 0; 823 sctx.shift_top = 0.0; 824 sctx.shift_lo = 0.0; 825 sctx.shift_hi = 0.0; 826 sctx.shift_fraction = 0.0; 827 sctx.shift_amount = 0.0; 828 829 /* if both shift schemes are chosen by user, only use info->shiftpd */ 830 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 831 ddiag = a->diag; 832 sctx.shift_top = info->zeropivot; 833 for (i=0; i<n; i++) { 834 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 835 d = (aa)[ddiag[i]]; 836 rs = -PetscAbsScalar(d) - PetscRealPart(d); 837 v = aa+ai[i]; 838 nz = ai[i+1] - ai[i]; 839 for (j=0; j<nz; j++) 840 rs += PetscAbsScalar(v[j]); 841 if (rs>sctx.shift_top) sctx.shift_top = rs; 842 } 843 sctx.shift_top *= 1.1; 844 sctx.nshift_max = 5; 845 sctx.shift_lo = 0.; 846 sctx.shift_hi = 1.; 847 } 848 849 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 850 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 851 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 852 ics = ic; 853 854 do { 855 sctx.lushift = PETSC_FALSE; 856 for (i=0; i<n; i++){ 857 /* zero rtmp */ 858 /* L part */ 859 nz = bi[i+1] - bi[i]; 860 bjtmp = bj + bi[i]; 861 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 862 863 /* U part */ 864 /* nz = bi[2*n-i+1] - bi[2*n-i]; 865 bjtmp = bj + bi[2*n-i]; */ 866 nz = bdiag[i]-bdiag[i+1]; 867 bjtmp = bj + bdiag[i+1]+1; 868 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 869 870 /* load in initial (unfactored row) */ 871 nz = ai[r[i]+1] - ai[r[i]]; 872 ajtmp = aj + ai[r[i]]; 873 v = aa + ai[r[i]]; 874 for (j=0; j<nz; j++) { 875 rtmp[ics[ajtmp[j]]] = v[j]; 876 } 877 /* ZeropivotApply() */ 878 rtmp[ics[r[i]]] += sctx.shift_amount; /* shift the diagonal of the matrix */ 879 880 /* elimination */ 881 bjtmp = bj + bi[i]; 882 row = *bjtmp++; 883 nzL = bi[i+1] - bi[i]; 884 for(k=0; k < nzL;k++) { 885 pc = rtmp + row; 886 if (*pc != 0.0) { 887 pv = b->a + bdiag[row]; 888 multiplier = *pc * (*pv); 889 *pc = multiplier; 890 /* pj = b->j + bi[2*n-row]; */ /* begining of U(row,:) */ 891 /* pv = b->a + bi[2*n-row]; 892 nz = bi[2*n-row+1] - bi[2*n-row] - 1; */ /* num of entries in U(row,:), excluding diag */ 893 pj = b->j + bdiag[row+1]+1; /* beginning of U(row,:) */ 894 pv = b->a + bdiag[row+1]+1; 895 nz = bdiag[row]-bdiag[row+1]-1; /* num of entries in U(row,:) excluding diag */ 896 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 897 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 898 } 899 row = *bjtmp++; 900 } 901 902 /* finished row so stick it into b->a */ 903 rs = 0.0; 904 /* L part */ 905 pv = b->a + bi[i] ; 906 pj = b->j + bi[i] ; 907 nz = bi[i+1] - bi[i]; 908 for (j=0; j<nz; j++) { 909 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 910 } 911 912 /* U part */ 913 /* pv = b->a + bi[2*n-i]; 914 pj = b->j + bi[2*n-i]; 915 nz = bi[2*n-i+1] - bi[2*n-i] - 1; 916 */ 917 pv = b->a + bdiag[i+1]+1; 918 pj = b->j + bdiag[i+1]+1; 919 nz = bdiag[i] - bdiag[i+1]-1; 920 for (j=0; j<nz; j++) { 921 pv[j] = rtmp[pj[j]]; rs += PetscAbsScalar(pv[j]); 922 } 923 924 /* ZeropivotCheck() */ 925 sctx.rs = rs; 926 sctx.pv = rtmp[i]; 927 ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); 928 if (newshift == 1) break; 929 930 /* Mark diagonal and invert diagonal for simplier triangular solves */ 931 pv = b->a + bdiag[i]; 932 *pv = 1.0/rtmp[i]; 933 934 } /* endof for (i=0; i<n; i++){ */ 935 936 /* ZeropivotRefine() */ 937 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max){ 938 /* 939 * if no shift in this attempt & shifting & started shifting & can refine, 940 * then try lower shift 941 */ 942 sctx.shift_hi = sctx.shift_fraction; 943 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 944 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 945 sctx.lushift = PETSC_TRUE; 946 sctx.nshift++; 947 } 948 } while (sctx.lushift); 949 950 ierr = PetscFree(rtmp);CHKERRQ(ierr); 951 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 952 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 953 954 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 955 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 956 if (row_identity && col_identity) { 957 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct_v2; 958 } else { 959 C->ops->solve = MatSolve_SeqAIJ_newdatastruct_v2; 960 } 961 962 C->ops->solveadd = 0; 963 C->ops->solvetranspose = 0; 964 C->ops->solvetransposeadd = 0; 965 C->ops->matsolve = 0; 966 C->assembled = PETSC_TRUE; 967 C->preallocated = PETSC_TRUE; 968 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 969 970 /* ZeropivotView() */ 971 if (sctx.nshift){ 972 if (info->shiftpd) { 973 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); 974 } else if (info->shiftnz) { 975 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 976 } 977 } 978 PetscFunctionReturn(0); 979 } 980 981 #undef __FUNCT__ 982 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ" 983 PetscErrorCode MatLUFactorNumeric_SeqAIJ(Mat B,Mat A,const MatFactorInfo *info) 984 { 985 Mat C=B; 986 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 987 IS isrow = b->row,isicol = b->icol; 988 PetscErrorCode ierr; 989 const PetscInt *r,*ic,*ics; 990 PetscInt nz,row,i,j,n=A->rmap->n,diag; 991 const PetscInt *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 992 const PetscInt *ajtmp,*bjtmp,*diag_offset = b->diag,*pj; 993 MatScalar *pv,*rtmp,*pc,multiplier,d; 994 const MatScalar *v,*aa=a->a; 995 PetscReal rs=0.0; 996 LUShift_Ctx sctx; 997 PetscInt newshift,*ddiag; 998 999 PetscFunctionBegin; 1000 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1001 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 1002 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 1003 ics = ic; 1004 1005 /* initialize shift context sctx */ 1006 sctx.nshift = 0; 1007 sctx.nshift_max = 0; 1008 sctx.shift_top = 0.0; 1009 sctx.shift_lo = 0.0; 1010 sctx.shift_hi = 0.0; 1011 sctx.shift_fraction = 0.0; 1012 sctx.shift_amount = 0.0; 1013 1014 /* if both shift schemes are chosen by user, only use info->shiftpd */ 1015 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 1016 ddiag = a->diag; 1017 sctx.shift_top = info->zeropivot; 1018 for (i=0; i<n; i++) { 1019 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 1020 d = (aa)[ddiag[i]]; 1021 rs = -PetscAbsScalar(d) - PetscRealPart(d); 1022 v = aa+ai[i]; 1023 nz = ai[i+1] - ai[i]; 1024 for (j=0; j<nz; j++) 1025 rs += PetscAbsScalar(v[j]); 1026 if (rs>sctx.shift_top) sctx.shift_top = rs; 1027 } 1028 sctx.shift_top *= 1.1; 1029 sctx.nshift_max = 5; 1030 sctx.shift_lo = 0.; 1031 sctx.shift_hi = 1.; 1032 } 1033 1034 do { 1035 sctx.lushift = PETSC_FALSE; 1036 for (i=0; i<n; i++){ 1037 nz = bi[i+1] - bi[i]; 1038 bjtmp = bj + bi[i]; 1039 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 1040 1041 /* load in initial (unfactored row) */ 1042 nz = ai[r[i]+1] - ai[r[i]]; 1043 ajtmp = aj + ai[r[i]]; 1044 v = aa + ai[r[i]]; 1045 for (j=0; j<nz; j++) { 1046 rtmp[ics[ajtmp[j]]] = v[j]; 1047 } 1048 rtmp[ics[r[i]]] += sctx.shift_amount; /* shift the diagonal of the matrix */ 1049 /* if (sctx.shift_amount > 0.0) printf("row %d, shift %g\n",i,sctx.shift_amount); */ 1050 1051 row = *bjtmp++; 1052 while (row < i) { 1053 pc = rtmp + row; 1054 if (*pc != 0.0) { 1055 pv = b->a + diag_offset[row]; 1056 pj = b->j + diag_offset[row] + 1; 1057 multiplier = *pc / *pv++; 1058 *pc = multiplier; 1059 nz = bi[row+1] - diag_offset[row] - 1; 1060 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 1061 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 1062 } 1063 row = *bjtmp++; 1064 } 1065 /* finished row so stick it into b->a */ 1066 pv = b->a + bi[i] ; 1067 pj = b->j + bi[i] ; 1068 nz = bi[i+1] - bi[i]; 1069 diag = diag_offset[i] - bi[i]; 1070 rs = 0.0; 1071 for (j=0; j<nz; j++) { 1072 pv[j] = rtmp[pj[j]]; 1073 rs += PetscAbsScalar(pv[j]); 1074 } 1075 rs -= PetscAbsScalar(pv[diag]); 1076 1077 /* 9/13/02 Victor Eijkhout suggested scaling zeropivot by rs for matrices with funny scalings */ 1078 sctx.rs = rs; 1079 sctx.pv = pv[diag]; 1080 ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); 1081 if (newshift == 1) break; 1082 } 1083 1084 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max) { 1085 /* 1086 * if no shift in this attempt & shifting & started shifting & can refine, 1087 * then try lower shift 1088 */ 1089 sctx.shift_hi = sctx.shift_fraction; 1090 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 1091 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 1092 sctx.lushift = PETSC_TRUE; 1093 sctx.nshift++; 1094 } 1095 } while (sctx.lushift); 1096 1097 /* invert diagonal entries for simplier triangular solves */ 1098 for (i=0; i<n; i++) { 1099 b->a[diag_offset[i]] = 1.0/b->a[diag_offset[i]]; 1100 } 1101 ierr = PetscFree(rtmp);CHKERRQ(ierr); 1102 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 1103 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 1104 if (b->inode.use) { 1105 C->ops->solve = MatSolve_Inode; 1106 } else { 1107 PetscTruth row_identity, col_identity; 1108 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 1109 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 1110 if (row_identity && col_identity) { 1111 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering; 1112 } else { 1113 C->ops->solve = MatSolve_SeqAIJ; 1114 } 1115 } 1116 C->ops->solveadd = MatSolveAdd_SeqAIJ; 1117 C->ops->solvetranspose = MatSolveTranspose_SeqAIJ; 1118 C->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ; 1119 C->ops->matsolve = MatMatSolve_SeqAIJ; 1120 C->assembled = PETSC_TRUE; 1121 C->preallocated = PETSC_TRUE; 1122 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 1123 if (sctx.nshift){ 1124 if (info->shiftpd) { 1125 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); 1126 } else if (info->shiftnz) { 1127 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 1128 } 1129 } 1130 PetscFunctionReturn(0); 1131 } 1132 1133 /* 1134 This routine implements inplace ILU(0) with row or/and column permutations. 1135 Input: 1136 A - original matrix 1137 Output; 1138 A - a->i (rowptr) is same as original rowptr, but factored i-the row is stored in rowperm[i] 1139 a->j (col index) is permuted by the inverse of colperm, then sorted 1140 a->a reordered accordingly with a->j 1141 a->diag (ptr to diagonal elements) is updated. 1142 */ 1143 #undef __FUNCT__ 1144 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_InplaceWithPerm" 1145 PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat B,Mat A,const MatFactorInfo *info) 1146 { 1147 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 1148 IS isrow = a->row,isicol = a->icol; 1149 PetscErrorCode ierr; 1150 const PetscInt *r,*ic,*ics; 1151 PetscInt i,j,n=A->rmap->n,*ai=a->i,*aj=a->j; 1152 PetscInt *ajtmp,nz,row; 1153 PetscInt *diag = a->diag,nbdiag,*pj; 1154 PetscScalar *rtmp,*pc,multiplier,d; 1155 MatScalar *v,*pv; 1156 PetscReal rs; 1157 LUShift_Ctx sctx; 1158 PetscInt newshift; 1159 1160 PetscFunctionBegin; 1161 if (A != B) SETERRQ(PETSC_ERR_ARG_INCOMP,"input and output matrix must have same address"); 1162 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1163 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 1164 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&rtmp);CHKERRQ(ierr); 1165 ierr = PetscMemzero(rtmp,(n+1)*sizeof(PetscScalar));CHKERRQ(ierr); 1166 ics = ic; 1167 1168 sctx.shift_top = 0; 1169 sctx.nshift_max = 0; 1170 sctx.shift_lo = 0; 1171 sctx.shift_hi = 0; 1172 sctx.shift_fraction = 0; 1173 1174 /* if both shift schemes are chosen by user, only use info->shiftpd */ 1175 if (info->shiftpd) { /* set sctx.shift_top=max{rs} */ 1176 sctx.shift_top = 0; 1177 for (i=0; i<n; i++) { 1178 /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */ 1179 d = (a->a)[diag[i]]; 1180 rs = -PetscAbsScalar(d) - PetscRealPart(d); 1181 v = a->a+ai[i]; 1182 nz = ai[i+1] - ai[i]; 1183 for (j=0; j<nz; j++) 1184 rs += PetscAbsScalar(v[j]); 1185 if (rs>sctx.shift_top) sctx.shift_top = rs; 1186 } 1187 if (sctx.shift_top < info->zeropivot) sctx.shift_top = info->zeropivot; 1188 sctx.shift_top *= 1.1; 1189 sctx.nshift_max = 5; 1190 sctx.shift_lo = 0.; 1191 sctx.shift_hi = 1.; 1192 } 1193 1194 sctx.shift_amount = 0; 1195 sctx.nshift = 0; 1196 do { 1197 sctx.lushift = PETSC_FALSE; 1198 for (i=0; i<n; i++){ 1199 /* load in initial unfactored row */ 1200 nz = ai[r[i]+1] - ai[r[i]]; 1201 ajtmp = aj + ai[r[i]]; 1202 v = a->a + ai[r[i]]; 1203 /* sort permuted ajtmp and values v accordingly */ 1204 for (j=0; j<nz; j++) ajtmp[j] = ics[ajtmp[j]]; 1205 ierr = PetscSortIntWithScalarArray(nz,ajtmp,v);CHKERRQ(ierr); 1206 1207 diag[r[i]] = ai[r[i]]; 1208 for (j=0; j<nz; j++) { 1209 rtmp[ajtmp[j]] = v[j]; 1210 if (ajtmp[j] < i) diag[r[i]]++; /* update a->diag */ 1211 } 1212 rtmp[r[i]] += sctx.shift_amount; /* shift the diagonal of the matrix */ 1213 1214 row = *ajtmp++; 1215 while (row < i) { 1216 pc = rtmp + row; 1217 if (*pc != 0.0) { 1218 pv = a->a + diag[r[row]]; 1219 pj = aj + diag[r[row]] + 1; 1220 1221 multiplier = *pc / *pv++; 1222 *pc = multiplier; 1223 nz = ai[r[row]+1] - diag[r[row]] - 1; 1224 for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j]; 1225 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 1226 } 1227 row = *ajtmp++; 1228 } 1229 /* finished row so overwrite it onto a->a */ 1230 pv = a->a + ai[r[i]] ; 1231 pj = aj + ai[r[i]] ; 1232 nz = ai[r[i]+1] - ai[r[i]]; 1233 nbdiag = diag[r[i]] - ai[r[i]]; /* num of entries before the diagonal */ 1234 1235 rs = 0.0; 1236 for (j=0; j<nz; j++) { 1237 pv[j] = rtmp[pj[j]]; 1238 if (j != nbdiag) rs += PetscAbsScalar(pv[j]); 1239 } 1240 1241 /* 9/13/02 Victor Eijkhout suggested scaling zeropivot by rs for matrices with funny scalings */ 1242 sctx.rs = rs; 1243 sctx.pv = pv[nbdiag]; 1244 ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr); 1245 if (newshift == 1) break; 1246 } 1247 1248 if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max) { 1249 /* 1250 * if no shift in this attempt & shifting & started shifting & can refine, 1251 * then try lower shift 1252 */ 1253 sctx.shift_hi = sctx.shift_fraction; 1254 sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.; 1255 sctx.shift_amount = sctx.shift_fraction * sctx.shift_top; 1256 sctx.lushift = PETSC_TRUE; 1257 sctx.nshift++; 1258 } 1259 } while (sctx.lushift); 1260 1261 /* invert diagonal entries for simplier triangular solves */ 1262 for (i=0; i<n; i++) { 1263 a->a[diag[r[i]]] = 1.0/a->a[diag[r[i]]]; 1264 } 1265 1266 ierr = PetscFree(rtmp);CHKERRQ(ierr); 1267 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 1268 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 1269 A->ops->solve = MatSolve_SeqAIJ_InplaceWithPerm; 1270 A->ops->solveadd = MatSolveAdd_SeqAIJ; 1271 A->ops->solvetranspose = MatSolveTranspose_SeqAIJ; 1272 A->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ; 1273 A->assembled = PETSC_TRUE; 1274 A->preallocated = PETSC_TRUE; 1275 ierr = PetscLogFlops(A->cmap->n);CHKERRQ(ierr); 1276 if (sctx.nshift){ 1277 if (info->shiftpd) { 1278 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); 1279 } else if (info->shiftnz) { 1280 ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 1281 } 1282 } 1283 PetscFunctionReturn(0); 1284 } 1285 1286 /* ----------------------------------------------------------- */ 1287 #undef __FUNCT__ 1288 #define __FUNCT__ "MatLUFactor_SeqAIJ" 1289 PetscErrorCode MatLUFactor_SeqAIJ(Mat A,IS row,IS col,const MatFactorInfo *info) 1290 { 1291 PetscErrorCode ierr; 1292 Mat C; 1293 1294 PetscFunctionBegin; 1295 ierr = MatGetFactor(A,MAT_SOLVER_PETSC,MAT_FACTOR_LU,&C);CHKERRQ(ierr); 1296 ierr = MatLUFactorSymbolic(C,A,row,col,info);CHKERRQ(ierr); 1297 ierr = MatLUFactorNumeric(C,A,info);CHKERRQ(ierr); 1298 A->ops->solve = C->ops->solve; 1299 A->ops->solvetranspose = C->ops->solvetranspose; 1300 ierr = MatHeaderCopy(A,C);CHKERRQ(ierr); 1301 ierr = PetscLogObjectParent(A,((Mat_SeqAIJ*)(A->data))->icol);CHKERRQ(ierr); 1302 PetscFunctionReturn(0); 1303 } 1304 /* ----------------------------------------------------------- */ 1305 1306 1307 #undef __FUNCT__ 1308 #define __FUNCT__ "MatSolve_SeqAIJ" 1309 PetscErrorCode MatSolve_SeqAIJ(Mat A,Vec bb,Vec xx) 1310 { 1311 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1312 IS iscol = a->col,isrow = a->row; 1313 PetscErrorCode ierr; 1314 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 1315 PetscInt nz; 1316 const PetscInt *rout,*cout,*r,*c; 1317 PetscScalar *x,*tmp,*tmps,sum; 1318 const PetscScalar *b; 1319 const MatScalar *aa = a->a,*v; 1320 1321 PetscFunctionBegin; 1322 if (!n) PetscFunctionReturn(0); 1323 1324 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1325 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1326 tmp = a->solve_work; 1327 1328 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1329 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1330 1331 /* forward solve the lower triangular */ 1332 tmp[0] = b[*r++]; 1333 tmps = tmp; 1334 for (i=1; i<n; i++) { 1335 v = aa + ai[i] ; 1336 vi = aj + ai[i] ; 1337 nz = a->diag[i] - ai[i]; 1338 sum = b[*r++]; 1339 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1340 tmp[i] = sum; 1341 } 1342 1343 /* backward solve the upper triangular */ 1344 for (i=n-1; i>=0; i--){ 1345 v = aa + a->diag[i] + 1; 1346 vi = aj + a->diag[i] + 1; 1347 nz = ai[i+1] - a->diag[i] - 1; 1348 sum = tmp[i]; 1349 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1350 x[*c--] = tmp[i] = sum*aa[a->diag[i]]; 1351 } 1352 1353 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1354 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1355 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1356 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1357 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1358 PetscFunctionReturn(0); 1359 } 1360 1361 #undef __FUNCT__ 1362 #define __FUNCT__ "MatMatSolve_SeqAIJ" 1363 PetscErrorCode MatMatSolve_SeqAIJ(Mat A,Mat B,Mat X) 1364 { 1365 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1366 IS iscol = a->col,isrow = a->row; 1367 PetscErrorCode ierr; 1368 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 1369 PetscInt nz,neq; 1370 const PetscInt *rout,*cout,*r,*c; 1371 PetscScalar *x,*b,*tmp,*tmps,sum; 1372 const MatScalar *aa = a->a,*v; 1373 PetscTruth bisdense,xisdense; 1374 1375 PetscFunctionBegin; 1376 if (!n) PetscFunctionReturn(0); 1377 1378 ierr = PetscTypeCompare((PetscObject)B,MATSEQDENSE,&bisdense);CHKERRQ(ierr); 1379 if (!bisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"B matrix must be a SeqDense matrix"); 1380 ierr = PetscTypeCompare((PetscObject)X,MATSEQDENSE,&xisdense);CHKERRQ(ierr); 1381 if (!xisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"X matrix must be a SeqDense matrix"); 1382 1383 ierr = MatGetArray(B,&b);CHKERRQ(ierr); 1384 ierr = MatGetArray(X,&x);CHKERRQ(ierr); 1385 1386 tmp = a->solve_work; 1387 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1388 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1389 1390 for (neq=0; neq<B->cmap->n; neq++){ 1391 /* forward solve the lower triangular */ 1392 tmp[0] = b[r[0]]; 1393 tmps = tmp; 1394 for (i=1; i<n; i++) { 1395 v = aa + ai[i] ; 1396 vi = aj + ai[i] ; 1397 nz = a->diag[i] - ai[i]; 1398 sum = b[r[i]]; 1399 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1400 tmp[i] = sum; 1401 } 1402 /* backward solve the upper triangular */ 1403 for (i=n-1; i>=0; i--){ 1404 v = aa + a->diag[i] + 1; 1405 vi = aj + a->diag[i] + 1; 1406 nz = ai[i+1] - a->diag[i] - 1; 1407 sum = tmp[i]; 1408 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1409 x[c[i]] = tmp[i] = sum*aa[a->diag[i]]; 1410 } 1411 1412 b += n; 1413 x += n; 1414 } 1415 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1416 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1417 ierr = MatRestoreArray(B,&b);CHKERRQ(ierr); 1418 ierr = MatRestoreArray(X,&x);CHKERRQ(ierr); 1419 ierr = PetscLogFlops(B->cmap->n*(2.0*a->nz - n));CHKERRQ(ierr); 1420 PetscFunctionReturn(0); 1421 } 1422 1423 #undef __FUNCT__ 1424 #define __FUNCT__ "MatSolve_SeqAIJ_InplaceWithPerm" 1425 PetscErrorCode MatSolve_SeqAIJ_InplaceWithPerm(Mat A,Vec bb,Vec xx) 1426 { 1427 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1428 IS iscol = a->col,isrow = a->row; 1429 PetscErrorCode ierr; 1430 const PetscInt *r,*c,*rout,*cout; 1431 PetscInt i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j; 1432 PetscInt nz,row; 1433 PetscScalar *x,*b,*tmp,*tmps,sum; 1434 const MatScalar *aa = a->a,*v; 1435 1436 PetscFunctionBegin; 1437 if (!n) PetscFunctionReturn(0); 1438 1439 ierr = VecGetArray(bb,&b);CHKERRQ(ierr); 1440 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1441 tmp = a->solve_work; 1442 1443 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1444 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1445 1446 /* forward solve the lower triangular */ 1447 tmp[0] = b[*r++]; 1448 tmps = tmp; 1449 for (row=1; row<n; row++) { 1450 i = rout[row]; /* permuted row */ 1451 v = aa + ai[i] ; 1452 vi = aj + ai[i] ; 1453 nz = a->diag[i] - ai[i]; 1454 sum = b[*r++]; 1455 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1456 tmp[row] = sum; 1457 } 1458 1459 /* backward solve the upper triangular */ 1460 for (row=n-1; row>=0; row--){ 1461 i = rout[row]; /* permuted row */ 1462 v = aa + a->diag[i] + 1; 1463 vi = aj + a->diag[i] + 1; 1464 nz = ai[i+1] - a->diag[i] - 1; 1465 sum = tmp[row]; 1466 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 1467 x[*c--] = tmp[row] = sum*aa[a->diag[i]]; 1468 } 1469 1470 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1471 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1472 ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr); 1473 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1474 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1475 PetscFunctionReturn(0); 1476 } 1477 1478 /* ----------------------------------------------------------- */ 1479 #include "../src/mat/impls/aij/seq/ftn-kernels/fsolve.h" 1480 #undef __FUNCT__ 1481 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering" 1482 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat A,Vec bb,Vec xx) 1483 { 1484 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1485 PetscErrorCode ierr; 1486 PetscInt n = A->rmap->n; 1487 const PetscInt *ai = a->i,*aj = a->j,*adiag = a->diag; 1488 PetscScalar *x; 1489 const PetscScalar *b; 1490 const MatScalar *aa = a->a; 1491 #if !defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 1492 PetscInt adiag_i,i,nz,ai_i; 1493 const PetscInt *vi; 1494 const MatScalar *v; 1495 PetscScalar sum; 1496 #endif 1497 1498 PetscFunctionBegin; 1499 if (!n) PetscFunctionReturn(0); 1500 1501 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1502 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1503 1504 #if defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 1505 fortransolveaij_(&n,x,ai,aj,adiag,aa,b); 1506 #else 1507 /* forward solve the lower triangular */ 1508 x[0] = b[0]; 1509 for (i=1; i<n; i++) { 1510 ai_i = ai[i]; 1511 v = aa + ai_i; 1512 vi = aj + ai_i; 1513 nz = adiag[i] - ai_i; 1514 sum = b[i]; 1515 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 1516 x[i] = sum; 1517 } 1518 1519 /* backward solve the upper triangular */ 1520 for (i=n-1; i>=0; i--){ 1521 adiag_i = adiag[i]; 1522 v = aa + adiag_i + 1; 1523 vi = aj + adiag_i + 1; 1524 nz = ai[i+1] - adiag_i - 1; 1525 sum = x[i]; 1526 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 1527 x[i] = sum*aa[adiag_i]; 1528 } 1529 #endif 1530 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 1531 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1532 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1533 PetscFunctionReturn(0); 1534 } 1535 1536 #undef __FUNCT__ 1537 #define __FUNCT__ "MatSolveAdd_SeqAIJ" 1538 PetscErrorCode MatSolveAdd_SeqAIJ(Mat A,Vec bb,Vec yy,Vec xx) 1539 { 1540 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1541 IS iscol = a->col,isrow = a->row; 1542 PetscErrorCode ierr; 1543 PetscInt i, n = A->rmap->n,j; 1544 PetscInt nz; 1545 const PetscInt *rout,*cout,*r,*c,*vi,*ai = a->i,*aj = a->j; 1546 PetscScalar *x,*tmp,sum; 1547 const PetscScalar *b; 1548 const MatScalar *aa = a->a,*v; 1549 1550 PetscFunctionBegin; 1551 if (yy != xx) {ierr = VecCopy(yy,xx);CHKERRQ(ierr);} 1552 1553 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1554 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1555 tmp = a->solve_work; 1556 1557 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1558 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 1559 1560 /* forward solve the lower triangular */ 1561 tmp[0] = b[*r++]; 1562 for (i=1; i<n; i++) { 1563 v = aa + ai[i] ; 1564 vi = aj + ai[i] ; 1565 nz = a->diag[i] - ai[i]; 1566 sum = b[*r++]; 1567 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1568 tmp[i] = sum; 1569 } 1570 1571 /* backward solve the upper triangular */ 1572 for (i=n-1; i>=0; i--){ 1573 v = aa + a->diag[i] + 1; 1574 vi = aj + a->diag[i] + 1; 1575 nz = ai[i+1] - a->diag[i] - 1; 1576 sum = tmp[i]; 1577 for (j=0; j<nz; j++) sum -= v[j]*tmp[vi[j]]; 1578 tmp[i] = sum*aa[a->diag[i]]; 1579 x[*c--] += tmp[i]; 1580 } 1581 1582 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1583 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1584 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1585 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1586 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1587 1588 PetscFunctionReturn(0); 1589 } 1590 1591 #undef __FUNCT__ 1592 #define __FUNCT__ "MatSolveTranspose_SeqAIJ" 1593 PetscErrorCode MatSolveTranspose_SeqAIJ(Mat A,Vec bb,Vec xx) 1594 { 1595 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1596 IS iscol = a->col,isrow = a->row; 1597 PetscErrorCode ierr; 1598 const PetscInt *rout,*cout,*r,*c,*diag = a->diag,*ai = a->i,*aj = a->j,*vi; 1599 PetscInt i,n = A->rmap->n,j; 1600 PetscInt nz; 1601 PetscScalar *x,*tmp,s1; 1602 const MatScalar *aa = a->a,*v; 1603 const PetscScalar *b; 1604 1605 PetscFunctionBegin; 1606 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1607 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1608 tmp = a->solve_work; 1609 1610 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1611 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1612 1613 /* copy the b into temp work space according to permutation */ 1614 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1615 1616 /* forward solve the U^T */ 1617 for (i=0; i<n; i++) { 1618 v = aa + diag[i] ; 1619 vi = aj + diag[i] + 1; 1620 nz = ai[i+1] - diag[i] - 1; 1621 s1 = tmp[i]; 1622 s1 *= (*v++); /* multiply by inverse of diagonal entry */ 1623 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1624 tmp[i] = s1; 1625 } 1626 1627 /* backward solve the L^T */ 1628 for (i=n-1; i>=0; i--){ 1629 v = aa + diag[i] - 1 ; 1630 vi = aj + diag[i] - 1 ; 1631 nz = diag[i] - ai[i]; 1632 s1 = tmp[i]; 1633 for (j=0; j>-nz; j--) tmp[vi[j]] -= s1*v[j]; 1634 } 1635 1636 /* copy tmp into x according to permutation */ 1637 for (i=0; i<n; i++) x[r[i]] = tmp[i]; 1638 1639 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1640 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1641 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1642 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1643 1644 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1645 PetscFunctionReturn(0); 1646 } 1647 1648 #undef __FUNCT__ 1649 #define __FUNCT__ "MatSolveTransposeAdd_SeqAIJ" 1650 PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat A,Vec bb,Vec zz,Vec xx) 1651 { 1652 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1653 IS iscol = a->col,isrow = a->row; 1654 PetscErrorCode ierr; 1655 const PetscInt *rout,*cout,*r,*c,*diag = a->diag,*ai = a->i,*aj = a->j,*vi; 1656 PetscInt i,n = A->rmap->n,j; 1657 PetscInt nz; 1658 PetscScalar *x,*tmp,s1; 1659 const MatScalar *aa = a->a,*v; 1660 const PetscScalar *b; 1661 1662 PetscFunctionBegin; 1663 if (zz != xx) {ierr = VecCopy(zz,xx);CHKERRQ(ierr);} 1664 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1665 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1666 tmp = a->solve_work; 1667 1668 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 1669 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 1670 1671 /* copy the b into temp work space according to permutation */ 1672 for (i=0; i<n; i++) tmp[i] = b[c[i]]; 1673 1674 /* forward solve the U^T */ 1675 for (i=0; i<n; i++) { 1676 v = aa + diag[i] ; 1677 vi = aj + diag[i] + 1; 1678 nz = ai[i+1] - diag[i] - 1; 1679 s1 = tmp[i]; 1680 s1 *= (*v++); /* multiply by inverse of diagonal entry */ 1681 for (j=0; j<nz; j++) tmp[vi[j]] -= s1*v[j]; 1682 tmp[i] = s1; 1683 } 1684 1685 /* backward solve the L^T */ 1686 for (i=n-1; i>=0; i--){ 1687 v = aa + diag[i] - 1 ; 1688 vi = aj + diag[i] - 1 ; 1689 nz = diag[i] - ai[i]; 1690 s1 = tmp[i]; 1691 for (j=0; j>-nz; j--) tmp[vi[j]] -= s1*v[j]; 1692 } 1693 1694 /* copy tmp into x according to permutation */ 1695 for (i=0; i<n; i++) x[r[i]] += tmp[i]; 1696 1697 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 1698 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 1699 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 1700 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1701 1702 ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr); 1703 PetscFunctionReturn(0); 1704 } 1705 1706 /* ----------------------------------------------------------------*/ 1707 EXTERN PetscErrorCode Mat_CheckInode(Mat,PetscTruth); 1708 EXTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat,Mat,MatDuplicateOption,PetscTruth); 1709 1710 /* 1711 ilu(0) with natural ordering under new data structure. 1712 Factored arrays bj and ba are stored as 1713 L(0,:), L(1,:), ...,L(n-1,:), U(n-1,:),...,U(i,:),U(i-1,:),...,U(0,:) 1714 1715 bi=fact->i is an array of size 2n+2, in which 1716 bi+ 1717 bi[i] -> 1st entry of L(i,:),i=0,...,i-1 1718 bi[n] -> points to L(n-1,:)+1 1719 bi[n+1] -> 1st entry of U(n-1,:) 1720 bi[2n-i] -> 1st entry of U(i,:) 1721 bi[2n-i+1] -> end of U(i,:)+1, the 1st entry of U(i-1,:) 1722 bi[2n] -> 1st entry of U(0,:) 1723 bi[2n+1] -> points to U(0,:)+1 1724 1725 U(i,:) contains diag[i] as its last entry, i.e., 1726 U(i,:) = (u[i,i+1],...,u[i,n-1],diag[i]) 1727 */ 1728 #undef __FUNCT__ 1729 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct" 1730 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1731 { 1732 1733 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1734 PetscErrorCode ierr; 1735 PetscInt n=A->rmap->n,*ai=a->i,*aj,*adiag=a->diag; 1736 PetscInt i,j,nz,*bi,*bj,*bdiag; 1737 1738 PetscFunctionBegin; 1739 ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES,PETSC_FALSE);CHKERRQ(ierr); 1740 b = (Mat_SeqAIJ*)(fact)->data; 1741 1742 /* allocate matrix arrays for new data structure */ 1743 ierr = PetscMalloc3(ai[n]+1,PetscScalar,&b->a,ai[n]+1,PetscInt,&b->j,2*n+2,PetscInt,&b->i);CHKERRQ(ierr); 1744 ierr = PetscLogObjectMemory(fact,ai[n]*(sizeof(PetscScalar)+sizeof(PetscInt))+(2*n+2)*sizeof(PetscInt));CHKERRQ(ierr); 1745 b->singlemalloc = PETSC_TRUE; 1746 if (!b->diag){ 1747 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&b->diag);CHKERRQ(ierr); 1748 } 1749 bdiag = b->diag; 1750 1751 if (n > 0) { 1752 ierr = PetscMemzero(b->a,(ai[n])*sizeof(MatScalar));CHKERRQ(ierr); 1753 } 1754 1755 /* set bi and bj with new data structure */ 1756 bi = b->i; 1757 bj = b->j; 1758 1759 /* L part */ 1760 bi[0] = 0; 1761 for (i=0; i<n; i++){ 1762 nz = adiag[i] - ai[i]; 1763 bi[i+1] = bi[i] + nz; 1764 aj = a->j + ai[i]; 1765 for (j=0; j<nz; j++){ 1766 *bj = aj[j]; bj++; 1767 } 1768 } 1769 1770 /* U part */ 1771 bi[n+1] = bi[n]; 1772 for (i=n-1; i>=0; i--){ 1773 nz = ai[i+1] - adiag[i] - 1; 1774 bi[2*n-i+1] = bi[2*n-i] + nz + 1; 1775 aj = a->j + adiag[i] + 1; 1776 for (j=0; j<nz; j++){ 1777 *bj = aj[j]; bj++; 1778 } 1779 /* diag[i] */ 1780 *bj = i; bj++; 1781 bdiag[i] = bi[2*n-i+1]-1; 1782 } 1783 PetscFunctionReturn(0); 1784 } 1785 1786 #undef __FUNCT__ 1787 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_newdatastruct" 1788 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 1789 { 1790 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 1791 IS isicol; 1792 PetscErrorCode ierr; 1793 const PetscInt *r,*ic; 1794 PetscInt n=A->rmap->n,*ai=a->i,*aj=a->j,d; 1795 PetscInt *bi,*cols,nnz,*cols_lvl; 1796 PetscInt *bdiag,prow,fm,nzbd,reallocs=0,dcount=0; 1797 PetscInt i,levels,diagonal_fill; 1798 PetscTruth col_identity,row_identity; 1799 PetscReal f; 1800 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL; 1801 PetscBT lnkbt; 1802 PetscInt nzi,*bj,**bj_ptr,**bjlvl_ptr; 1803 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 1804 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 1805 PetscTruth missing; 1806 1807 PetscFunctionBegin; 1808 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); 1809 f = info->fill; 1810 levels = (PetscInt)info->levels; 1811 diagonal_fill = (PetscInt)info->diagonal_fill; 1812 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 1813 1814 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 1815 ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 1816 1817 if (!levels && row_identity && col_identity) { 1818 /* special case: ilu(0) with natural ordering */ 1819 ierr = MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct(fact,A,isrow,iscol,info);CHKERRQ(ierr); 1820 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 1821 1822 fact->factor = MAT_FACTOR_ILU; 1823 (fact)->info.factor_mallocs = 0; 1824 (fact)->info.fill_ratio_given = info->fill; 1825 (fact)->info.fill_ratio_needed = 1.0; 1826 b = (Mat_SeqAIJ*)(fact)->data; 1827 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 1828 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 1829 b->row = isrow; 1830 b->col = iscol; 1831 b->icol = isicol; 1832 ierr = PetscMalloc(((fact)->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1833 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1834 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1835 /* ierr = MatILUFactorSymbolic_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); */ 1836 PetscFunctionReturn(0); 1837 } 1838 1839 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 1840 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 1841 1842 /* get new row pointers */ 1843 ierr = PetscMalloc((2*n+2)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 1844 bi[0] = 0; 1845 /* bdiag is location of diagonal in factor */ 1846 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 1847 bdiag[0] = 0; 1848 1849 ierr = PetscMalloc((2*n+1)*sizeof(PetscInt**),&bj_ptr);CHKERRQ(ierr); 1850 bjlvl_ptr = (PetscInt**)(bj_ptr + n); 1851 1852 /* create a linked list for storing column indices of the active row */ 1853 nlnk = n + 1; 1854 ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1855 1856 /* initial FreeSpace size is f*(ai[n]+1) */ 1857 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 1858 current_space = free_space; 1859 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr); 1860 current_space_lvl = free_space_lvl; 1861 1862 for (i=0; i<n; i++) { 1863 nzi = 0; 1864 /* copy current row into linked list */ 1865 nnz = ai[r[i]+1] - ai[r[i]]; 1866 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 1867 cols = aj + ai[r[i]]; 1868 lnk[i] = -1; /* marker to indicate if diagonal exists */ 1869 ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 1870 nzi += nlnk; 1871 1872 /* make sure diagonal entry is included */ 1873 if (diagonal_fill && lnk[i] == -1) { 1874 fm = n; 1875 while (lnk[fm] < i) fm = lnk[fm]; 1876 lnk[i] = lnk[fm]; /* insert diagonal into linked list */ 1877 lnk[fm] = i; 1878 lnk_lvl[i] = 0; 1879 nzi++; dcount++; 1880 } 1881 1882 /* add pivot rows into the active row */ 1883 nzbd = 0; 1884 prow = lnk[n]; 1885 while (prow < i) { 1886 nnz = bdiag[prow]; 1887 cols = bj_ptr[prow] + nnz + 1; 1888 cols_lvl = bjlvl_ptr[prow] + nnz + 1; 1889 nnz = bi[prow+1] - bi[prow] - nnz - 1; 1890 ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr); 1891 nzi += nlnk; 1892 prow = lnk[prow]; 1893 nzbd++; 1894 } 1895 bdiag[i] = nzbd; 1896 bi[i+1] = bi[i] + nzi; 1897 1898 /* if free space is not available, make more free space */ 1899 if (current_space->local_remaining<nzi) { 1900 nnz = 2*nzi*(n - i); /* estimated and max additional space needed */ 1901 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 1902 ierr = PetscFreeSpaceGet(nnz,¤t_space_lvl);CHKERRQ(ierr); 1903 reallocs++; 1904 } 1905 1906 /* copy data into free_space and free_space_lvl, then initialize lnk */ 1907 ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 1908 bj_ptr[i] = current_space->array; 1909 bjlvl_ptr[i] = current_space_lvl->array; 1910 1911 /* make sure the active row i has diagonal entry */ 1912 if (*(bj_ptr[i]+bdiag[i]) != i) { 1913 SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\ 1914 try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i); 1915 } 1916 1917 current_space->array += nzi; 1918 current_space->local_used += nzi; 1919 current_space->local_remaining -= nzi; 1920 current_space_lvl->array += nzi; 1921 current_space_lvl->local_used += nzi; 1922 current_space_lvl->local_remaining -= nzi; 1923 } 1924 1925 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 1926 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 1927 1928 /* destroy list of free space and other temporary arrays */ 1929 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 1930 1931 /* copy free_space into bj and free free_space; set bi, bj, bdiag in new datastructure; */ 1932 ierr = PetscFreeSpaceContiguous_LU(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 1933 1934 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 1935 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 1936 ierr = PetscFree(bj_ptr);CHKERRQ(ierr); 1937 1938 #if defined(PETSC_USE_INFO) 1939 { 1940 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 1941 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 1942 ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 1943 ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr); 1944 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 1945 if (diagonal_fill) { 1946 ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr); 1947 } 1948 } 1949 #endif 1950 1951 /* put together the new matrix */ 1952 ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 1953 ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr); 1954 b = (Mat_SeqAIJ*)(fact)->data; 1955 b->free_a = PETSC_TRUE; 1956 b->free_ij = PETSC_TRUE; 1957 b->singlemalloc = PETSC_FALSE; 1958 ierr = PetscMalloc( (bi[2*n+1] )*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 1959 b->j = bj; 1960 b->i = bi; 1961 b->diag = bdiag; 1962 b->ilen = 0; 1963 b->imax = 0; 1964 b->row = isrow; 1965 b->col = iscol; 1966 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1967 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 1968 b->icol = isicol; 1969 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 1970 /* In b structure: Free imax, ilen, old a, old j. 1971 Allocate bdiag, solve_work, new a, new j */ 1972 ierr = PetscLogObjectMemory(fact,bi[2*n+1] * (sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 1973 b->maxnz = b->nz = bi[2*n+1] ; 1974 (fact)->info.factor_mallocs = reallocs; 1975 (fact)->info.fill_ratio_given = f; 1976 (fact)->info.fill_ratio_needed = ((PetscReal)bi[2*n+1])/((PetscReal)ai[n]); 1977 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; 1978 /* ierr = MatILUFactorSymbolic_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); */ 1979 PetscFunctionReturn(0); 1980 } 1981 1982 1983 /* 1984 ilu(0) with natural ordering under revised new data structure. 1985 Factored arrays bj and ba are stored as 1986 L(0,:), L(1,:), ...,L(n-1,:), U(n-1,:),...,U(i,:),U(i-1,:),...,U(0,:) 1987 1988 bi=fact->i is an array of size n+1, in which 1989 bi+ 1990 bi[i]: points to 1st entry of L(i,:),i=0,...,n-1 1991 bi[n]: points to L(n-1,n-1)+1 1992 1993 bdiag=fact->diag is an array of size n+1,in which 1994 bdiag[i]: points to diagonal of U(i,:), i=0,...,n-1 1995 bdiag[n]: points to diagonal of U(n-1,0)-1 1996 1997 U(i,:) contains bdiag[i] as its last entry, i.e., 1998 U(i,:) = (u[i,i+1],...,u[i,n-1],diag[i]) 1999 */ 2000 #undef __FUNCT__ 2001 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct_v2" 2002 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct_v2(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 2003 { 2004 2005 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 2006 PetscErrorCode ierr; 2007 PetscInt n=A->rmap->n,*ai=a->i,*aj,*adiag=a->diag; 2008 PetscInt i,j,nz,*bi,*bj,*bdiag,bi_temp; 2009 2010 PetscFunctionBegin; 2011 ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES,PETSC_FALSE);CHKERRQ(ierr); 2012 b = (Mat_SeqAIJ*)(fact)->data; 2013 2014 /* allocate matrix arrays for new data structure */ 2015 ierr = PetscMalloc3(ai[n]+1,PetscScalar,&b->a,ai[n]+1,PetscInt,&b->j,n+1,PetscInt,&b->i);CHKERRQ(ierr); 2016 ierr = PetscLogObjectMemory(fact,ai[n]*(sizeof(PetscScalar)+sizeof(PetscInt))+(n+1)*sizeof(PetscInt));CHKERRQ(ierr); 2017 b->singlemalloc = PETSC_TRUE; 2018 if (!b->diag){ 2019 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&b->diag);CHKERRQ(ierr); 2020 ierr = PetscLogObjectMemory(fact,(n+1)*sizeof(PetscInt));CHKERRQ(ierr); 2021 } 2022 bdiag = b->diag; 2023 2024 if (n > 0) { 2025 ierr = PetscMemzero(b->a,(ai[n])*sizeof(MatScalar));CHKERRQ(ierr); 2026 } 2027 2028 /* set bi and bj with new data structure */ 2029 bi = b->i; 2030 bj = b->j; 2031 2032 /* L part */ 2033 bi[0] = 0; 2034 for (i=0; i<n; i++){ 2035 nz = adiag[i] - ai[i]; 2036 bi[i+1] = bi[i] + nz; 2037 aj = a->j + ai[i]; 2038 for (j=0; j<nz; j++){ 2039 *bj = aj[j]; bj++; 2040 } 2041 } 2042 2043 /* U part */ 2044 /* bi[n+1] = bi[n]; */ 2045 bi_temp = bi[n]; 2046 bdiag[n] = bi[n]-1; 2047 for (i=n-1; i>=0; i--){ 2048 nz = ai[i+1] - adiag[i] - 1; 2049 /* bi[2*n-i+1] = bi[2*n-i] + nz + 1; */ 2050 bi_temp = bi_temp + nz + 1; 2051 aj = a->j + adiag[i] + 1; 2052 for (j=0; j<nz; j++){ 2053 *bj = aj[j]; bj++; 2054 } 2055 /* diag[i] */ 2056 *bj = i; bj++; 2057 /* bdiag[i] = bi[2*n-i+1]-1; */ 2058 bdiag[i] = bi_temp - 1; 2059 } 2060 PetscFunctionReturn(0); 2061 } 2062 2063 #undef __FUNCT__ 2064 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ_newdatastruct_v2" 2065 PetscErrorCode MatILUFactorSymbolic_SeqAIJ_newdatastruct_v2(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 2066 { 2067 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 2068 IS isicol; 2069 PetscErrorCode ierr; 2070 const PetscInt *r,*ic; 2071 PetscInt n=A->rmap->n,*ai=a->i,*aj=a->j,d; 2072 PetscInt *bi,*cols,nnz,*cols_lvl; 2073 PetscInt *bdiag,prow,fm,nzbd,reallocs=0,dcount=0; 2074 PetscInt i,levels,diagonal_fill; 2075 PetscTruth col_identity,row_identity; 2076 PetscReal f; 2077 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL; 2078 PetscBT lnkbt; 2079 PetscInt nzi,*bj,**bj_ptr,**bjlvl_ptr; 2080 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2081 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2082 PetscTruth missing; 2083 2084 PetscFunctionBegin; 2085 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); 2086 f = info->fill; 2087 levels = (PetscInt)info->levels; 2088 diagonal_fill = (PetscInt)info->diagonal_fill; 2089 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 2090 2091 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 2092 ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 2093 2094 if (!levels && row_identity && col_identity) { 2095 /* special case: ilu(0) with natural ordering */ 2096 ierr = MatILUFactorSymbolic_SeqAIJ_ilu0_newdatastruct_v2(fact,A,isrow,iscol,info);CHKERRQ(ierr); 2097 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct_v2; 2098 2099 fact->factor = MAT_FACTOR_ILU; 2100 (fact)->info.factor_mallocs = 0; 2101 (fact)->info.fill_ratio_given = info->fill; 2102 (fact)->info.fill_ratio_needed = 1.0; 2103 b = (Mat_SeqAIJ*)(fact)->data; 2104 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2105 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2106 b->row = isrow; 2107 b->col = iscol; 2108 b->icol = isicol; 2109 ierr = PetscMalloc(((fact)->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2110 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2111 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2112 /* ierr = MatILUFactorSymbolic_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); */ 2113 PetscFunctionReturn(0); 2114 } 2115 2116 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 2117 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 2118 2119 /* get new row pointers */ 2120 /* ierr = PetscMalloc((2*n+2)*sizeof(PetscInt),&bi);CHKERRQ(ierr); */ 2121 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 2122 bi[0] = 0; 2123 /* bdiag is location of diagonal in factor */ 2124 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 2125 bdiag[0] = 0; 2126 2127 ierr = PetscMalloc((2*n+1)*sizeof(PetscInt**),&bj_ptr);CHKERRQ(ierr); 2128 bjlvl_ptr = (PetscInt**)(bj_ptr + n); 2129 2130 /* create a linked list for storing column indices of the active row */ 2131 nlnk = n + 1; 2132 ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2133 2134 /* initial FreeSpace size is f*(ai[n]+1) */ 2135 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 2136 current_space = free_space; 2137 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr); 2138 current_space_lvl = free_space_lvl; 2139 2140 for (i=0; i<n; i++) { 2141 nzi = 0; 2142 /* copy current row into linked list */ 2143 nnz = ai[r[i]+1] - ai[r[i]]; 2144 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 2145 cols = aj + ai[r[i]]; 2146 lnk[i] = -1; /* marker to indicate if diagonal exists */ 2147 ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2148 nzi += nlnk; 2149 2150 /* make sure diagonal entry is included */ 2151 if (diagonal_fill && lnk[i] == -1) { 2152 fm = n; 2153 while (lnk[fm] < i) fm = lnk[fm]; 2154 lnk[i] = lnk[fm]; /* insert diagonal into linked list */ 2155 lnk[fm] = i; 2156 lnk_lvl[i] = 0; 2157 nzi++; dcount++; 2158 } 2159 2160 /* add pivot rows into the active row */ 2161 nzbd = 0; 2162 prow = lnk[n]; 2163 while (prow < i) { 2164 nnz = bdiag[prow]; 2165 cols = bj_ptr[prow] + nnz + 1; 2166 cols_lvl = bjlvl_ptr[prow] + nnz + 1; 2167 nnz = bi[prow+1] - bi[prow] - nnz - 1; 2168 ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr); 2169 nzi += nlnk; 2170 prow = lnk[prow]; 2171 nzbd++; 2172 } 2173 bdiag[i] = nzbd; 2174 bi[i+1] = bi[i] + nzi; 2175 2176 /* if free space is not available, make more free space */ 2177 if (current_space->local_remaining<nzi) { 2178 nnz = 2*nzi*(n - i); /* estimated and max additional space needed */ 2179 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 2180 ierr = PetscFreeSpaceGet(nnz,¤t_space_lvl);CHKERRQ(ierr); 2181 reallocs++; 2182 } 2183 2184 /* copy data into free_space and free_space_lvl, then initialize lnk */ 2185 ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 2186 bj_ptr[i] = current_space->array; 2187 bjlvl_ptr[i] = current_space_lvl->array; 2188 2189 /* make sure the active row i has diagonal entry */ 2190 if (*(bj_ptr[i]+bdiag[i]) != i) { 2191 SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\ 2192 try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i); 2193 } 2194 2195 current_space->array += nzi; 2196 current_space->local_used += nzi; 2197 current_space->local_remaining -= nzi; 2198 current_space_lvl->array += nzi; 2199 current_space_lvl->local_used += nzi; 2200 current_space_lvl->local_remaining -= nzi; 2201 } 2202 2203 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 2204 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 2205 2206 /* destroy list of free space and other temporary arrays */ 2207 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 2208 2209 /* copy free_space into bj and free free_space; set bi, bj, bdiag in new datastructure; */ 2210 /* ierr = PetscFreeSpaceContiguous_LU(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); */ 2211 ierr = PetscFreeSpaceContiguous_LU_v2(&free_space,bj,n,bi,bdiag);CHKERRQ(ierr); 2212 2213 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2214 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2215 ierr = PetscFree(bj_ptr);CHKERRQ(ierr); 2216 2217 #if defined(PETSC_USE_INFO) 2218 { 2219 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 2220 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 2221 ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2222 ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr); 2223 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 2224 if (diagonal_fill) { 2225 ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr); 2226 } 2227 } 2228 #endif 2229 2230 /* put together the new matrix */ 2231 ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 2232 ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr); 2233 b = (Mat_SeqAIJ*)(fact)->data; 2234 b->free_a = PETSC_TRUE; 2235 b->free_ij = PETSC_TRUE; 2236 b->singlemalloc = PETSC_FALSE; 2237 /* ierr = PetscMalloc( (bi[2*n+1] )*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); */ 2238 ierr = PetscMalloc((bdiag[0]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 2239 b->j = bj; 2240 b->i = bi; 2241 b->diag = bdiag; 2242 b->ilen = 0; 2243 b->imax = 0; 2244 b->row = isrow; 2245 b->col = iscol; 2246 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2247 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2248 b->icol = isicol; 2249 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2250 /* In b structure: Free imax, ilen, old a, old j. 2251 Allocate bdiag, solve_work, new a, new j */ 2252 2253 /* ierr = PetscLogObjectMemory(fact,bi[2*n+1] * (sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 2254 b->maxnz = b->nz = bi[2*n+1] ; */ 2255 ierr = PetscLogObjectMemory(fact,(bdiag[0]+1)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 2256 b->maxnz = b->nz = bdiag[0]+1; 2257 (fact)->info.factor_mallocs = reallocs; 2258 (fact)->info.fill_ratio_given = f; 2259 /* (fact)->info.fill_ratio_needed = ((PetscReal)bi[2*n+1])/((PetscReal)ai[n]); */ 2260 (fact)->info.fill_ratio_needed = ((PetscReal)(bdiag[0]+1))/((PetscReal)ai[n]); 2261 /* (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct; */ 2262 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ_newdatastruct_v2; 2263 /* ierr = MatILUFactorSymbolic_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); */ 2264 PetscFunctionReturn(0); 2265 } 2266 2267 #undef __FUNCT__ 2268 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ" 2269 PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info) 2270 { 2271 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 2272 IS isicol; 2273 PetscErrorCode ierr; 2274 const PetscInt *r,*ic; 2275 PetscInt n=A->rmap->n,*ai=a->i,*aj=a->j,d; 2276 PetscInt *bi,*cols,nnz,*cols_lvl; 2277 PetscInt *bdiag,prow,fm,nzbd,reallocs=0,dcount=0; 2278 PetscInt i,levels,diagonal_fill; 2279 PetscTruth col_identity,row_identity; 2280 PetscReal f; 2281 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL; 2282 PetscBT lnkbt; 2283 PetscInt nzi,*bj,**bj_ptr,**bjlvl_ptr; 2284 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2285 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2286 PetscTruth missing; 2287 PetscTruth newdatastruct=PETSC_FALSE,newdatastruct_v2=PETSC_FALSE; 2288 2289 PetscFunctionBegin; 2290 ierr = PetscOptionsGetTruth(PETSC_NULL,"-ilu_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 2291 if (newdatastruct){ 2292 ierr = MatILUFactorSymbolic_SeqAIJ_newdatastruct(fact,A,isrow,iscol,info);CHKERRQ(ierr); 2293 PetscFunctionReturn(0); 2294 } 2295 ierr = PetscOptionsGetTruth(PETSC_NULL,"-ilu_new_v2",&newdatastruct_v2,PETSC_NULL);CHKERRQ(ierr); 2296 if(newdatastruct_v2){ 2297 ierr = MatILUFactorSymbolic_SeqAIJ_newdatastruct_v2(fact,A,isrow,iscol,info);CHKERRQ(ierr); 2298 PetscFunctionReturn(0); 2299 } 2300 2301 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); 2302 f = info->fill; 2303 levels = (PetscInt)info->levels; 2304 diagonal_fill = (PetscInt)info->diagonal_fill; 2305 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 2306 2307 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 2308 ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 2309 if (!levels && row_identity && col_identity) { /* special case: ilu(0) with natural ordering */ 2310 ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES,PETSC_TRUE);CHKERRQ(ierr); 2311 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ; 2312 2313 fact->factor = MAT_FACTOR_ILU; 2314 (fact)->info.factor_mallocs = 0; 2315 (fact)->info.fill_ratio_given = info->fill; 2316 (fact)->info.fill_ratio_needed = 1.0; 2317 b = (Mat_SeqAIJ*)(fact)->data; 2318 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2319 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2320 b->row = isrow; 2321 b->col = iscol; 2322 b->icol = isicol; 2323 ierr = PetscMalloc(((fact)->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2324 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2325 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2326 ierr = MatILUFactorSymbolic_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); 2327 PetscFunctionReturn(0); 2328 } 2329 2330 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 2331 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 2332 2333 /* get new row pointers */ 2334 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 2335 bi[0] = 0; 2336 /* bdiag is location of diagonal in factor */ 2337 ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 2338 bdiag[0] = 0; 2339 2340 ierr = PetscMalloc((2*n+1)*sizeof(PetscInt**),&bj_ptr);CHKERRQ(ierr); 2341 bjlvl_ptr = (PetscInt**)(bj_ptr + n); 2342 2343 /* create a linked list for storing column indices of the active row */ 2344 nlnk = n + 1; 2345 ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2346 2347 /* initial FreeSpace size is f*(ai[n]+1) */ 2348 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr); 2349 current_space = free_space; 2350 ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr); 2351 current_space_lvl = free_space_lvl; 2352 2353 for (i=0; i<n; i++) { 2354 nzi = 0; 2355 /* copy current row into linked list */ 2356 nnz = ai[r[i]+1] - ai[r[i]]; 2357 if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 2358 cols = aj + ai[r[i]]; 2359 lnk[i] = -1; /* marker to indicate if diagonal exists */ 2360 ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2361 nzi += nlnk; 2362 2363 /* make sure diagonal entry is included */ 2364 if (diagonal_fill && lnk[i] == -1) { 2365 fm = n; 2366 while (lnk[fm] < i) fm = lnk[fm]; 2367 lnk[i] = lnk[fm]; /* insert diagonal into linked list */ 2368 lnk[fm] = i; 2369 lnk_lvl[i] = 0; 2370 nzi++; dcount++; 2371 } 2372 2373 /* add pivot rows into the active row */ 2374 nzbd = 0; 2375 prow = lnk[n]; 2376 while (prow < i) { 2377 nnz = bdiag[prow]; 2378 cols = bj_ptr[prow] + nnz + 1; 2379 cols_lvl = bjlvl_ptr[prow] + nnz + 1; 2380 nnz = bi[prow+1] - bi[prow] - nnz - 1; 2381 ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr); 2382 nzi += nlnk; 2383 prow = lnk[prow]; 2384 nzbd++; 2385 } 2386 bdiag[i] = nzbd; 2387 bi[i+1] = bi[i] + nzi; 2388 2389 /* if free space is not available, make more free space */ 2390 if (current_space->local_remaining<nzi) { 2391 nnz = nzi*(n - i); /* estimated and max additional space needed */ 2392 ierr = PetscFreeSpaceGet(nnz,¤t_space);CHKERRQ(ierr); 2393 ierr = PetscFreeSpaceGet(nnz,¤t_space_lvl);CHKERRQ(ierr); 2394 reallocs++; 2395 } 2396 2397 /* copy data into free_space and free_space_lvl, then initialize lnk */ 2398 ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 2399 bj_ptr[i] = current_space->array; 2400 bjlvl_ptr[i] = current_space_lvl->array; 2401 2402 /* make sure the active row i has diagonal entry */ 2403 if (*(bj_ptr[i]+bdiag[i]) != i) { 2404 SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\ 2405 try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i); 2406 } 2407 2408 current_space->array += nzi; 2409 current_space->local_used += nzi; 2410 current_space->local_remaining -= nzi; 2411 current_space_lvl->array += nzi; 2412 current_space_lvl->local_used += nzi; 2413 current_space_lvl->local_remaining -= nzi; 2414 } 2415 2416 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 2417 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 2418 2419 /* destroy list of free space and other temporary arrays */ 2420 ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 2421 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); /* copy free_space -> bj */ 2422 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2423 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2424 ierr = PetscFree(bj_ptr);CHKERRQ(ierr); 2425 2426 #if defined(PETSC_USE_INFO) 2427 { 2428 PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]); 2429 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr); 2430 ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2431 ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr); 2432 ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr); 2433 if (diagonal_fill) { 2434 ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr); 2435 } 2436 } 2437 #endif 2438 2439 /* put together the new matrix */ 2440 ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 2441 ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr); 2442 b = (Mat_SeqAIJ*)(fact)->data; 2443 b->free_a = PETSC_TRUE; 2444 b->free_ij = PETSC_TRUE; 2445 b->singlemalloc = PETSC_FALSE; 2446 ierr = PetscMalloc( (bi[n] )*sizeof(PetscScalar),&b->a);CHKERRQ(ierr); 2447 b->j = bj; 2448 b->i = bi; 2449 for (i=0; i<n; i++) bdiag[i] += bi[i]; 2450 b->diag = bdiag; 2451 b->ilen = 0; 2452 b->imax = 0; 2453 b->row = isrow; 2454 b->col = iscol; 2455 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2456 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2457 b->icol = isicol; 2458 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2459 /* In b structure: Free imax, ilen, old a, old j. 2460 Allocate bdiag, solve_work, new a, new j */ 2461 ierr = PetscLogObjectMemory(fact,(bi[n]-n) * (sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr); 2462 b->maxnz = b->nz = bi[n] ; 2463 (fact)->info.factor_mallocs = reallocs; 2464 (fact)->info.fill_ratio_given = f; 2465 (fact)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]); 2466 (fact)->ops->lufactornumeric = MatLUFactorNumeric_SeqAIJ; 2467 ierr = MatILUFactorSymbolic_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr); 2468 PetscFunctionReturn(0); 2469 } 2470 2471 #undef __FUNCT__ 2472 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ_newdatastruct" 2473 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 2474 { 2475 Mat C = B; 2476 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 2477 Mat_SeqSBAIJ *b=(Mat_SeqSBAIJ*)C->data; 2478 IS ip=b->row,iip = b->icol; 2479 PetscErrorCode ierr; 2480 const PetscInt *rip,*riip; 2481 PetscInt i,j,mbs=A->rmap->n,*bi=b->i,*bj=b->j,*bdiag=b->diag,*bjtmp; 2482 PetscInt *ai=a->i,*aj=a->j; 2483 PetscInt k,jmin,jmax,*c2r,*il,col,nexti,ili,nz; 2484 MatScalar *rtmp,*ba=b->a,*bval,*aa=a->a,dk,uikdi; 2485 PetscReal zeropivot,shiftnz; 2486 PetscReal shiftpd; 2487 PetscTruth perm_identity; 2488 2489 PetscFunctionBegin; 2490 2491 shiftnz = info->shiftnz; 2492 shiftpd = info->shiftpd; 2493 zeropivot = info->zeropivot; 2494 2495 ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 2496 ierr = ISGetIndices(iip,&riip);CHKERRQ(ierr); 2497 2498 /* allocate working arrays 2499 c2r: linked list, keep track of pivot rows for a given column. c2r[col]: head of the list for a given col 2500 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 2501 */ 2502 nz = (2*mbs+1)*sizeof(PetscInt)+mbs*sizeof(MatScalar); 2503 ierr = PetscMalloc(nz,&il);CHKERRQ(ierr); 2504 c2r = il + mbs; 2505 rtmp = (MatScalar*)(c2r + mbs); 2506 2507 for (i=0; i<mbs; i++) { 2508 c2r[i] = mbs; 2509 } 2510 il[0] = 0; 2511 2512 for (k = 0; k<mbs; k++){ 2513 /* zero rtmp */ 2514 nz = bi[k+1] - bi[k]; 2515 bjtmp = bj + bi[k]; 2516 for (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0; 2517 2518 /* load in initial unfactored row */ 2519 bval = ba + bi[k]; 2520 jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 2521 for (j = jmin; j < jmax; j++){ 2522 col = riip[aj[j]]; 2523 if (col >= k){ /* only take upper triangular entry */ 2524 rtmp[col] = aa[j]; 2525 *bval++ = 0.0; /* for in-place factorization */ 2526 } 2527 } 2528 /* shift the diagonal of the matrix */ 2529 2530 /* modify k-th row by adding in those rows i with U(i,k)!=0 */ 2531 dk = rtmp[k]; 2532 i = c2r[k]; /* first row to be added to k_th row */ 2533 2534 while (i < k){ 2535 nexti = c2r[i]; /* next row to be added to k_th row */ 2536 2537 /* compute multiplier, update diag(k) and U(i,k) */ 2538 ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2539 uikdi = - ba[ili]*ba[bdiag[i]]; /* diagonal(k) */ 2540 dk += uikdi*ba[ili]; /* update diag[k] */ 2541 ba[ili] = uikdi; /* -U(i,k) */ 2542 2543 /* add multiple of row i to k-th row */ 2544 jmin = ili + 1; jmax = bi[i+1]; 2545 if (jmin < jmax){ 2546 for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 2547 /* update il and c2r for row i */ 2548 il[i] = jmin; 2549 j = bj[jmin]; c2r[i] = c2r[j]; c2r[j] = i; 2550 } 2551 i = nexti; 2552 } 2553 2554 /* copy data into U(k,:) */ 2555 ba[bdiag[k]] = 1.0/dk; /* U(k,k) */ 2556 jmin = bi[k]; jmax = bi[k+1]-1; 2557 if (jmin < jmax) { 2558 for (j=jmin; j<jmax; j++){ 2559 col = bj[j]; ba[j] = rtmp[col]; 2560 } 2561 /* add the k-th row into il and c2r */ 2562 il[k] = jmin; 2563 i = bj[jmin]; c2r[k] = c2r[i]; c2r[i] = k; 2564 } 2565 } 2566 2567 ierr = PetscFree(il);CHKERRQ(ierr); 2568 ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 2569 ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr); 2570 2571 ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr); 2572 if (perm_identity){ 2573 (B)->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering_newdatastruct; 2574 (B)->ops->solvetranspose = 0; 2575 (B)->ops->forwardsolve = 0; 2576 (B)->ops->backwardsolve = 0; 2577 } else { 2578 (B)->ops->solve = MatSolve_SeqSBAIJ_1_newdatastruct; 2579 (B)->ops->solvetranspose = 0; 2580 (B)->ops->forwardsolve = 0; 2581 (B)->ops->backwardsolve = 0; 2582 } 2583 2584 C->assembled = PETSC_TRUE; 2585 C->preallocated = PETSC_TRUE; 2586 ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr); 2587 PetscFunctionReturn(0); 2588 } 2589 2590 #undef __FUNCT__ 2591 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ" 2592 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat B,Mat A,const MatFactorInfo *info) 2593 { 2594 Mat C = B; 2595 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 2596 Mat_SeqSBAIJ *b=(Mat_SeqSBAIJ*)C->data; 2597 IS ip=b->row,iip = b->icol; 2598 PetscErrorCode ierr; 2599 const PetscInt *rip,*riip; 2600 PetscInt i,j,mbs=A->rmap->n,*bi=b->i,*bj=b->j,*bcol; 2601 PetscInt *ai=a->i,*aj=a->j; 2602 PetscInt k,jmin,jmax,*jl,*il,col,nexti,ili,nz; 2603 MatScalar *rtmp,*ba=b->a,*bval,*aa=a->a,dk,uikdi; 2604 PetscReal zeropivot,rs,shiftnz; 2605 PetscReal shiftpd; 2606 ChShift_Ctx sctx; 2607 PetscInt newshift; 2608 PetscTruth perm_identity; 2609 2610 PetscFunctionBegin; 2611 shiftnz = info->shiftnz; 2612 shiftpd = info->shiftpd; 2613 zeropivot = info->zeropivot; 2614 2615 ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 2616 ierr = ISGetIndices(iip,&riip);CHKERRQ(ierr); 2617 2618 /* initialization */ 2619 nz = (2*mbs+1)*sizeof(PetscInt)+mbs*sizeof(MatScalar); 2620 ierr = PetscMalloc(nz,&il);CHKERRQ(ierr); 2621 jl = il + mbs; 2622 rtmp = (MatScalar*)(jl + mbs); 2623 2624 sctx.shift_amount = 0; 2625 sctx.nshift = 0; 2626 do { 2627 sctx.chshift = PETSC_FALSE; 2628 for (i=0; i<mbs; i++) { 2629 rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0; 2630 } 2631 2632 for (k = 0; k<mbs; k++){ 2633 bval = ba + bi[k]; 2634 /* initialize k-th row by the perm[k]-th row of A */ 2635 jmin = ai[rip[k]]; jmax = ai[rip[k]+1]; 2636 for (j = jmin; j < jmax; j++){ 2637 col = riip[aj[j]]; 2638 if (col >= k){ /* only take upper triangular entry */ 2639 rtmp[col] = aa[j]; 2640 *bval++ = 0.0; /* for in-place factorization */ 2641 } 2642 } 2643 /* shift the diagonal of the matrix */ 2644 if (sctx.nshift) rtmp[k] += sctx.shift_amount; 2645 2646 /* modify k-th row by adding in those rows i with U(i,k)!=0 */ 2647 dk = rtmp[k]; 2648 i = jl[k]; /* first row to be added to k_th row */ 2649 2650 while (i < k){ 2651 nexti = jl[i]; /* next row to be added to k_th row */ 2652 2653 /* compute multiplier, update diag(k) and U(i,k) */ 2654 ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 2655 uikdi = - ba[ili]*ba[bi[i]]; /* diagonal(k) */ 2656 dk += uikdi*ba[ili]; 2657 ba[ili] = uikdi; /* -U(i,k) */ 2658 2659 /* add multiple of row i to k-th row */ 2660 jmin = ili + 1; jmax = bi[i+1]; 2661 if (jmin < jmax){ 2662 for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j]; 2663 /* update il and jl for row i */ 2664 il[i] = jmin; 2665 j = bj[jmin]; jl[i] = jl[j]; jl[j] = i; 2666 } 2667 i = nexti; 2668 } 2669 2670 /* shift the diagonals when zero pivot is detected */ 2671 /* compute rs=sum of abs(off-diagonal) */ 2672 rs = 0.0; 2673 jmin = bi[k]+1; 2674 nz = bi[k+1] - jmin; 2675 bcol = bj + jmin; 2676 for (j=0; j<nz; j++) { 2677 rs += PetscAbsScalar(rtmp[bcol[j]]); 2678 } 2679 2680 sctx.rs = rs; 2681 sctx.pv = dk; 2682 ierr = MatCholeskyCheckShift_inline(info,sctx,k,newshift);CHKERRQ(ierr); 2683 2684 if (newshift == 1) { 2685 if (!sctx.shift_amount) { 2686 sctx.shift_amount = 1e-5; 2687 } 2688 break; 2689 } 2690 2691 /* copy data into U(k,:) */ 2692 ba[bi[k]] = 1.0/dk; /* U(k,k) */ 2693 jmin = bi[k]+1; jmax = bi[k+1]; 2694 if (jmin < jmax) { 2695 for (j=jmin; j<jmax; j++){ 2696 col = bj[j]; ba[j] = rtmp[col]; rtmp[col] = 0.0; 2697 } 2698 /* add the k-th row into il and jl */ 2699 il[k] = jmin; 2700 i = bj[jmin]; jl[k] = jl[i]; jl[i] = k; 2701 } 2702 } 2703 } while (sctx.chshift); 2704 ierr = PetscFree(il);CHKERRQ(ierr); 2705 2706 ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 2707 ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr); 2708 2709 ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr); 2710 if (perm_identity){ 2711 (B)->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering; 2712 (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1_NaturalOrdering; 2713 (B)->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1_NaturalOrdering; 2714 (B)->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1_NaturalOrdering; 2715 } else { 2716 (B)->ops->solve = MatSolve_SeqSBAIJ_1; 2717 (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1; 2718 (B)->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1; 2719 (B)->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1; 2720 } 2721 2722 C->assembled = PETSC_TRUE; 2723 C->preallocated = PETSC_TRUE; 2724 ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr); 2725 if (sctx.nshift){ 2726 if (shiftnz) { 2727 ierr = PetscInfo2(A,"number of shiftnz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2728 } else if (shiftpd) { 2729 ierr = PetscInfo2(A,"number of shiftpd tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr); 2730 } 2731 } 2732 PetscFunctionReturn(0); 2733 } 2734 2735 #undef __FUNCT__ 2736 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ_newdatastruct" 2737 PetscErrorCode MatICCFactorSymbolic_SeqAIJ_newdatastruct(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2738 { 2739 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2740 Mat_SeqSBAIJ *b; 2741 PetscErrorCode ierr; 2742 PetscTruth perm_identity,missing; 2743 PetscInt reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui,*udiag; 2744 const PetscInt *rip,*riip; 2745 PetscInt jmin,jmax,nzk,k,j,*jl,prow,*il,nextprow; 2746 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL,d; 2747 PetscInt ncols,ncols_upper,*cols,*ajtmp,*uj,**uj_ptr,**uj_lvl_ptr; 2748 PetscReal fill=info->fill,levels=info->levels; 2749 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2750 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2751 PetscBT lnkbt; 2752 IS iperm; 2753 2754 PetscFunctionBegin; 2755 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); 2756 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2757 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2758 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2759 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2760 2761 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2762 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2763 ui[0] = 0; 2764 2765 /* ICC(0) without matrix ordering: simply copies fill pattern */ 2766 if (!levels && perm_identity) { 2767 2768 for (i=0; i<am; i++) { 2769 ui[i+1] = ui[i] + ai[i+1] - a->diag[i]; 2770 udiag[i] = ui[i+1] - 1; /* points to the last entry of U(i,:) */ 2771 } 2772 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2773 cols = uj; 2774 for (i=0; i<am; i++) { 2775 aj = a->j + a->diag[i] + 1; /* 1st entry of U(i,:) without diagonal */ 2776 ncols = ui[i+1] - ui[i] - 1; 2777 for (j=0; j<ncols; j++) *cols++ = aj[j]; 2778 *cols++ = i; /* diagoanl is located as the last entry of U(i,:) */ 2779 } 2780 } else { /* case: levels>0 || (levels=0 && !perm_identity) */ 2781 2782 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2783 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2784 2785 /* initialization */ 2786 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ajtmp);CHKERRQ(ierr); 2787 2788 /* jl: linked list for storing indices of the pivot rows 2789 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 2790 ierr = PetscMalloc((2*am+1)*sizeof(PetscInt)+2*am*sizeof(PetscInt**),&jl);CHKERRQ(ierr); 2791 il = jl + am; 2792 uj_ptr = (PetscInt**)(il + am); 2793 uj_lvl_ptr = (PetscInt**)(uj_ptr + am); 2794 for (i=0; i<am; i++){ 2795 jl[i] = am; il[i] = 0; 2796 } 2797 2798 /* create and initialize a linked list for storing column indices of the active row k */ 2799 nlnk = am + 1; 2800 ierr = PetscIncompleteLLCreate(am,am,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2801 2802 /* initial FreeSpace size is fill*(ai[am]+1) */ 2803 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 2804 current_space = free_space; 2805 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space_lvl);CHKERRQ(ierr); 2806 current_space_lvl = free_space_lvl; 2807 2808 for (k=0; k<am; k++){ /* for each active row k */ 2809 /* initialize lnk by the column indices of row rip[k] of A */ 2810 nzk = 0; 2811 ncols = ai[rip[k]+1] - ai[rip[k]]; 2812 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 2813 ncols_upper = 0; 2814 for (j=0; j<ncols; j++){ 2815 i = *(aj + ai[rip[k]] + j); /* unpermuted column index */ 2816 if (riip[i] >= k){ /* only take upper triangular entry */ 2817 ajtmp[ncols_upper] = i; 2818 ncols_upper++; 2819 } 2820 } 2821 ierr = PetscIncompleteLLInit(ncols_upper,ajtmp,am,riip,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 2822 nzk += nlnk; 2823 2824 /* update lnk by computing fill-in for each pivot row to be merged in */ 2825 prow = jl[k]; /* 1st pivot row */ 2826 2827 while (prow < k){ 2828 nextprow = jl[prow]; 2829 2830 /* merge prow into k-th row */ 2831 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 2832 jmax = ui[prow+1]; 2833 ncols = jmax-jmin; 2834 i = jmin - ui[prow]; 2835 cols = uj_ptr[prow] + i; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 2836 uj = uj_lvl_ptr[prow] + i; /* levels of cols */ 2837 j = *(uj - 1); 2838 ierr = PetscICCLLAddSorted(ncols,cols,levels,uj,am,nlnk,lnk,lnk_lvl,lnkbt,j);CHKERRQ(ierr); 2839 nzk += nlnk; 2840 2841 /* update il and jl for prow */ 2842 if (jmin < jmax){ 2843 il[prow] = jmin; 2844 j = *cols; jl[prow] = jl[j]; jl[j] = prow; 2845 } 2846 prow = nextprow; 2847 } 2848 2849 /* if free space is not available, make more free space */ 2850 if (current_space->local_remaining<nzk) { 2851 i = am - k + 1; /* num of unfactored rows */ 2852 i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 2853 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 2854 ierr = PetscFreeSpaceGet(i,¤t_space_lvl);CHKERRQ(ierr); 2855 reallocs++; 2856 } 2857 2858 /* copy data into free_space and free_space_lvl, then initialize lnk */ 2859 if (nzk == 0) SETERRQ1(PETSC_ERR_ARG_WRONG,"Empty row %D in ICC matrix factor",k); 2860 ierr = PetscIncompleteLLClean(am,am,nzk,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 2861 2862 /* add the k-th row into il and jl */ 2863 if (nzk > 1){ 2864 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 2865 jl[k] = jl[i]; jl[i] = k; 2866 il[k] = ui[k] + 1; 2867 } 2868 uj_ptr[k] = current_space->array; 2869 uj_lvl_ptr[k] = current_space_lvl->array; 2870 2871 current_space->array += nzk; 2872 current_space->local_used += nzk; 2873 current_space->local_remaining -= nzk; 2874 2875 current_space_lvl->array += nzk; 2876 current_space_lvl->local_used += nzk; 2877 current_space_lvl->local_remaining -= nzk; 2878 2879 ui[k+1] = ui[k] + nzk; 2880 } 2881 2882 #if defined(PETSC_USE_INFO) 2883 if (ai[am] != 0) { 2884 PetscReal af = (PetscReal)ui[am]/((PetscReal)ai[am]); 2885 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 2886 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 2887 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 2888 } else { 2889 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 2890 } 2891 #endif 2892 2893 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 2894 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 2895 ierr = PetscFree(jl);CHKERRQ(ierr); 2896 ierr = PetscFree(ajtmp);CHKERRQ(ierr); 2897 2898 /* destroy list of free space and other temporary array(s) */ 2899 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2900 ierr = PetscFreeSpaceContiguous_Cholesky(&free_space,uj,am,ui,udiag);CHKERRQ(ierr); 2901 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 2902 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 2903 2904 } /* end of case: levels>0 || (levels=0 && !perm_identity) */ 2905 2906 /* put together the new matrix in MATSEQSBAIJ format */ 2907 2908 b = (Mat_SeqSBAIJ*)(fact)->data; 2909 b->singlemalloc = PETSC_FALSE; 2910 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 2911 b->j = uj; 2912 b->i = ui; 2913 b->diag = udiag; 2914 b->free_diag = PETSC_TRUE; 2915 b->ilen = 0; 2916 b->imax = 0; 2917 b->row = perm; 2918 b->col = perm; 2919 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2920 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 2921 b->icol = iperm; 2922 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 2923 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 2924 ierr = PetscLogObjectMemory((fact),(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 2925 b->maxnz = b->nz = ui[am]; 2926 b->free_a = PETSC_TRUE; 2927 b->free_ij = PETSC_TRUE; 2928 2929 (fact)->info.factor_mallocs = reallocs; 2930 (fact)->info.fill_ratio_given = fill; 2931 if (ai[am] != 0) { 2932 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 2933 } else { 2934 (fact)->info.fill_ratio_needed = 0.0; 2935 } 2936 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_newdatastruct; 2937 PetscFunctionReturn(0); 2938 } 2939 2940 #undef __FUNCT__ 2941 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ" 2942 PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 2943 { 2944 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2945 Mat_SeqSBAIJ *b; 2946 PetscErrorCode ierr; 2947 PetscTruth perm_identity,missing; 2948 PetscInt reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui,*udiag; 2949 const PetscInt *rip,*riip; 2950 PetscInt jmin,jmax,nzk,k,j,*jl,prow,*il,nextprow; 2951 PetscInt nlnk,*lnk,*lnk_lvl=PETSC_NULL,d; 2952 PetscInt ncols,ncols_upper,*cols,*ajtmp,*uj,**uj_ptr,**uj_lvl_ptr; 2953 PetscReal fill=info->fill,levels=info->levels; 2954 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 2955 PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL; 2956 PetscBT lnkbt; 2957 IS iperm; 2958 PetscTruth newdatastruct=PETSC_FALSE; 2959 2960 PetscFunctionBegin; 2961 ierr = PetscOptionsGetTruth(PETSC_NULL,"-icc_new",&newdatastruct,PETSC_NULL);CHKERRQ(ierr); 2962 if(newdatastruct){ 2963 ierr = MatICCFactorSymbolic_SeqAIJ_newdatastruct(fact,A,perm,info);CHKERRQ(ierr); 2964 PetscFunctionReturn(0); 2965 } 2966 2967 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); 2968 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 2969 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 2970 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 2971 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 2972 2973 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 2974 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&udiag);CHKERRQ(ierr); 2975 ui[0] = 0; 2976 2977 /* ICC(0) without matrix ordering: simply copies fill pattern */ 2978 if (!levels && perm_identity) { 2979 2980 for (i=0; i<am; i++) { 2981 ui[i+1] = ui[i] + ai[i+1] - a->diag[i]; 2982 udiag[i] = ui[i]; 2983 } 2984 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 2985 cols = uj; 2986 for (i=0; i<am; i++) { 2987 aj = a->j + a->diag[i]; 2988 ncols = ui[i+1] - ui[i]; 2989 for (j=0; j<ncols; j++) *cols++ = *aj++; 2990 } 2991 } else { /* case: levels>0 || (levels=0 && !perm_identity) */ 2992 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 2993 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 2994 2995 /* initialization */ 2996 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ajtmp);CHKERRQ(ierr); 2997 2998 /* jl: linked list for storing indices of the pivot rows 2999 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 3000 ierr = PetscMalloc((2*am+1)*sizeof(PetscInt)+2*am*sizeof(PetscInt**),&jl);CHKERRQ(ierr); 3001 il = jl + am; 3002 uj_ptr = (PetscInt**)(il + am); 3003 uj_lvl_ptr = (PetscInt**)(uj_ptr + am); 3004 for (i=0; i<am; i++){ 3005 jl[i] = am; il[i] = 0; 3006 } 3007 3008 /* create and initialize a linked list for storing column indices of the active row k */ 3009 nlnk = am + 1; 3010 ierr = PetscIncompleteLLCreate(am,am,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 3011 3012 /* initial FreeSpace size is fill*(ai[am]+1) */ 3013 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 3014 current_space = free_space; 3015 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space_lvl);CHKERRQ(ierr); 3016 current_space_lvl = free_space_lvl; 3017 3018 for (k=0; k<am; k++){ /* for each active row k */ 3019 /* initialize lnk by the column indices of row rip[k] of A */ 3020 nzk = 0; 3021 ncols = ai[rip[k]+1] - ai[rip[k]]; 3022 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 3023 ncols_upper = 0; 3024 for (j=0; j<ncols; j++){ 3025 i = *(aj + ai[rip[k]] + j); /* unpermuted column index */ 3026 if (riip[i] >= k){ /* only take upper triangular entry */ 3027 ajtmp[ncols_upper] = i; 3028 ncols_upper++; 3029 } 3030 } 3031 ierr = PetscIncompleteLLInit(ncols_upper,ajtmp,am,riip,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr); 3032 nzk += nlnk; 3033 3034 /* update lnk by computing fill-in for each pivot row to be merged in */ 3035 prow = jl[k]; /* 1st pivot row */ 3036 3037 while (prow < k){ 3038 nextprow = jl[prow]; 3039 3040 /* merge prow into k-th row */ 3041 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 3042 jmax = ui[prow+1]; 3043 ncols = jmax-jmin; 3044 i = jmin - ui[prow]; 3045 cols = uj_ptr[prow] + i; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 3046 uj = uj_lvl_ptr[prow] + i; /* levels of cols */ 3047 j = *(uj - 1); 3048 ierr = PetscICCLLAddSorted(ncols,cols,levels,uj,am,nlnk,lnk,lnk_lvl,lnkbt,j);CHKERRQ(ierr); 3049 nzk += nlnk; 3050 3051 /* update il and jl for prow */ 3052 if (jmin < jmax){ 3053 il[prow] = jmin; 3054 j = *cols; jl[prow] = jl[j]; jl[j] = prow; 3055 } 3056 prow = nextprow; 3057 } 3058 3059 /* if free space is not available, make more free space */ 3060 if (current_space->local_remaining<nzk) { 3061 i = am - k + 1; /* num of unfactored rows */ 3062 i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 3063 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 3064 ierr = PetscFreeSpaceGet(i,¤t_space_lvl);CHKERRQ(ierr); 3065 reallocs++; 3066 } 3067 3068 /* copy data into free_space and free_space_lvl, then initialize lnk */ 3069 if (nzk == 0) SETERRQ1(PETSC_ERR_ARG_WRONG,"Empty row %D in ICC matrix factor",k); 3070 ierr = PetscIncompleteLLClean(am,am,nzk,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr); 3071 3072 /* add the k-th row into il and jl */ 3073 if (nzk > 1){ 3074 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 3075 jl[k] = jl[i]; jl[i] = k; 3076 il[k] = ui[k] + 1; 3077 } 3078 uj_ptr[k] = current_space->array; 3079 uj_lvl_ptr[k] = current_space_lvl->array; 3080 3081 current_space->array += nzk; 3082 current_space->local_used += nzk; 3083 current_space->local_remaining -= nzk; 3084 3085 current_space_lvl->array += nzk; 3086 current_space_lvl->local_used += nzk; 3087 current_space_lvl->local_remaining -= nzk; 3088 3089 ui[k+1] = ui[k] + nzk; 3090 } 3091 3092 #if defined(PETSC_USE_INFO) 3093 if (ai[am] != 0) { 3094 PetscReal af = (PetscReal)ui[am]/((PetscReal)ai[am]); 3095 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 3096 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 3097 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 3098 } else { 3099 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 3100 } 3101 #endif 3102 3103 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 3104 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 3105 ierr = PetscFree(jl);CHKERRQ(ierr); 3106 ierr = PetscFree(ajtmp);CHKERRQ(ierr); 3107 3108 /* destroy list of free space and other temporary array(s) */ 3109 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 3110 ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr); 3111 ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 3112 ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr); 3113 3114 } /* end of case: levels>0 || (levels=0 && !perm_identity) */ 3115 3116 /* put together the new matrix in MATSEQSBAIJ format */ 3117 3118 b = (Mat_SeqSBAIJ*)(fact)->data; 3119 b->singlemalloc = PETSC_FALSE; 3120 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 3121 b->j = uj; 3122 b->i = ui; 3123 b->diag = udiag; 3124 b->free_diag = PETSC_TRUE; 3125 b->ilen = 0; 3126 b->imax = 0; 3127 b->row = perm; 3128 b->col = perm; 3129 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 3130 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 3131 b->icol = iperm; 3132 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 3133 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 3134 ierr = PetscLogObjectMemory((fact),(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 3135 b->maxnz = b->nz = ui[am]; 3136 b->free_a = PETSC_TRUE; 3137 b->free_ij = PETSC_TRUE; 3138 3139 (fact)->info.factor_mallocs = reallocs; 3140 (fact)->info.fill_ratio_given = fill; 3141 if (ai[am] != 0) { 3142 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 3143 } else { 3144 (fact)->info.fill_ratio_needed = 0.0; 3145 } 3146 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ; 3147 PetscFunctionReturn(0); 3148 } 3149 3150 #undef __FUNCT__ 3151 #define __FUNCT__ "MatCholeskyFactorSymbolic_SeqAIJ" 3152 PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 3153 { 3154 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3155 Mat_SeqSBAIJ *b; 3156 PetscErrorCode ierr; 3157 PetscTruth perm_identity; 3158 PetscReal fill = info->fill; 3159 const PetscInt *rip,*riip; 3160 PetscInt i,am=A->rmap->n,*ai=a->i,*aj=a->j,reallocs=0,prow; 3161 PetscInt *jl,jmin,jmax,nzk,*ui,k,j,*il,nextprow; 3162 PetscInt nlnk,*lnk,ncols,ncols_upper,*cols,*uj,**ui_ptr,*uj_ptr; 3163 PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 3164 PetscBT lnkbt; 3165 IS iperm; 3166 3167 PetscFunctionBegin; 3168 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); 3169 /* check whether perm is the identity mapping */ 3170 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 3171 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 3172 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 3173 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 3174 3175 /* initialization */ 3176 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 3177 ui[0] = 0; 3178 3179 /* jl: linked list for storing indices of the pivot rows 3180 il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */ 3181 ierr = PetscMalloc((3*am+1)*sizeof(PetscInt)+am*sizeof(PetscInt**),&jl);CHKERRQ(ierr); 3182 il = jl + am; 3183 cols = il + am; 3184 ui_ptr = (PetscInt**)(cols + am); 3185 for (i=0; i<am; i++){ 3186 jl[i] = am; il[i] = 0; 3187 } 3188 3189 /* create and initialize a linked list for storing column indices of the active row k */ 3190 nlnk = am + 1; 3191 ierr = PetscLLCreate(am,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3192 3193 /* initial FreeSpace size is fill*(ai[am]+1) */ 3194 ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr); 3195 current_space = free_space; 3196 3197 for (k=0; k<am; k++){ /* for each active row k */ 3198 /* initialize lnk by the column indices of row rip[k] of A */ 3199 nzk = 0; 3200 ncols = ai[rip[k]+1] - ai[rip[k]]; 3201 if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k); 3202 ncols_upper = 0; 3203 for (j=0; j<ncols; j++){ 3204 i = riip[*(aj + ai[rip[k]] + j)]; 3205 if (i >= k){ /* only take upper triangular entry */ 3206 cols[ncols_upper] = i; 3207 ncols_upper++; 3208 } 3209 } 3210 ierr = PetscLLAdd(ncols_upper,cols,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3211 nzk += nlnk; 3212 3213 /* update lnk by computing fill-in for each pivot row to be merged in */ 3214 prow = jl[k]; /* 1st pivot row */ 3215 3216 while (prow < k){ 3217 nextprow = jl[prow]; 3218 /* merge prow into k-th row */ 3219 jmin = il[prow] + 1; /* index of the 2nd nzero entry in U(prow,k:am-1) */ 3220 jmax = ui[prow+1]; 3221 ncols = jmax-jmin; 3222 uj_ptr = ui_ptr[prow] + jmin - ui[prow]; /* points to the 2nd nzero entry in U(prow,k:am-1) */ 3223 ierr = PetscLLAddSorted(ncols,uj_ptr,am,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3224 nzk += nlnk; 3225 3226 /* update il and jl for prow */ 3227 if (jmin < jmax){ 3228 il[prow] = jmin; 3229 j = *uj_ptr; jl[prow] = jl[j]; jl[j] = prow; 3230 } 3231 prow = nextprow; 3232 } 3233 3234 /* if free space is not available, make more free space */ 3235 if (current_space->local_remaining<nzk) { 3236 i = am - k + 1; /* num of unfactored rows */ 3237 i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */ 3238 ierr = PetscFreeSpaceGet(i,¤t_space);CHKERRQ(ierr); 3239 reallocs++; 3240 } 3241 3242 /* copy data into free space, then initialize lnk */ 3243 ierr = PetscLLClean(am,am,nzk,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 3244 3245 /* add the k-th row into il and jl */ 3246 if (nzk-1 > 0){ 3247 i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */ 3248 jl[k] = jl[i]; jl[i] = k; 3249 il[k] = ui[k] + 1; 3250 } 3251 ui_ptr[k] = current_space->array; 3252 current_space->array += nzk; 3253 current_space->local_used += nzk; 3254 current_space->local_remaining -= nzk; 3255 3256 ui[k+1] = ui[k] + nzk; 3257 } 3258 3259 #if defined(PETSC_USE_INFO) 3260 if (ai[am] != 0) { 3261 PetscReal af = (PetscReal)(ui[am])/((PetscReal)ai[am]); 3262 ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr); 3263 ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr); 3264 ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr); 3265 } else { 3266 ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr); 3267 } 3268 #endif 3269 3270 ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr); 3271 ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr); 3272 ierr = PetscFree(jl);CHKERRQ(ierr); 3273 3274 /* destroy list of free space and other temporary array(s) */ 3275 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 3276 ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr); 3277 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 3278 3279 /* put together the new matrix in MATSEQSBAIJ format */ 3280 3281 b = (Mat_SeqSBAIJ*)(fact)->data; 3282 b->singlemalloc = PETSC_FALSE; 3283 b->free_a = PETSC_TRUE; 3284 b->free_ij = PETSC_TRUE; 3285 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 3286 b->j = uj; 3287 b->i = ui; 3288 b->diag = 0; 3289 b->ilen = 0; 3290 b->imax = 0; 3291 b->row = perm; 3292 b->col = perm; 3293 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 3294 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 3295 b->icol = iperm; 3296 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 3297 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 3298 ierr = PetscLogObjectMemory(fact,(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 3299 b->maxnz = b->nz = ui[am]; 3300 3301 (fact)->info.factor_mallocs = reallocs; 3302 (fact)->info.fill_ratio_given = fill; 3303 if (ai[am] != 0) { 3304 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 3305 } else { 3306 (fact)->info.fill_ratio_needed = 0.0; 3307 } 3308 (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ; 3309 PetscFunctionReturn(0); 3310 } 3311 3312 #undef __FUNCT__ 3313 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering_newdatastruct" 3314 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_newdatastruct(Mat A,Vec bb,Vec xx) 3315 { 3316 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3317 PetscErrorCode ierr; 3318 PetscInt n = A->rmap->n; 3319 const PetscInt *ai = a->i,*aj = a->j,*vi; 3320 PetscScalar *x,sum; 3321 const PetscScalar *b; 3322 const MatScalar *aa = a->a,*v; 3323 PetscInt i,nz; 3324 3325 PetscFunctionBegin; 3326 if (!n) PetscFunctionReturn(0); 3327 3328 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3329 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 3330 3331 /* forward solve the lower triangular */ 3332 x[0] = b[0]; 3333 v = aa; 3334 vi = aj; 3335 for (i=1; i<n; i++) { 3336 nz = ai[i+1] - ai[i]; 3337 sum = b[i]; 3338 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 3339 v += nz; 3340 vi += nz; 3341 x[i] = sum; 3342 } 3343 3344 /* backward solve the upper triangular */ 3345 v = aa + ai[n+1]; 3346 vi = aj + ai[n+1]; 3347 for (i=n-1; i>=0; i--){ 3348 nz = ai[2*n-i +1] - ai[2*n-i]-1; 3349 sum = x[i]; 3350 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 3351 v += nz; 3352 vi += nz; vi++; 3353 x[i] = *v++ *sum; /* x[i]=aa[adiag[i]]*sum; v++; */ 3354 } 3355 3356 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 3357 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3358 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 3359 PetscFunctionReturn(0); 3360 } 3361 3362 #undef __FUNCT__ 3363 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering_newdatastruct_v2" 3364 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_newdatastruct_v2(Mat A,Vec bb,Vec xx) 3365 { 3366 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3367 PetscErrorCode ierr; 3368 PetscInt n = A->rmap->n; 3369 const PetscInt *ai = a->i,*aj = a->j,*adiag = a->diag,*vi; 3370 PetscScalar *x,sum; 3371 const PetscScalar *b; 3372 const MatScalar *aa = a->a,*v; 3373 PetscInt i,nz; 3374 3375 PetscFunctionBegin; 3376 if (!n) PetscFunctionReturn(0); 3377 3378 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3379 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 3380 3381 /* forward solve the lower triangular */ 3382 x[0] = b[0]; 3383 v = aa; 3384 vi = aj; 3385 for (i=1; i<n; i++) { 3386 nz = ai[i+1] - ai[i]; 3387 sum = b[i]; 3388 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 3389 v += nz; 3390 vi += nz; 3391 x[i] = sum; 3392 } 3393 3394 /* backward solve the upper triangular */ 3395 /* v = aa + ai[n+1]; 3396 vi = aj + ai[n+1]; */ 3397 v = aa + adiag[n-1]; 3398 vi = aj + adiag[n-1]; 3399 for (i=n-1; i>=0; i--){ 3400 nz = adiag[i] - adiag[i+1]-1; 3401 sum = x[i]; 3402 PetscSparseDenseMinusDot(sum,x,v,vi,nz); 3403 v += nz; 3404 vi += nz; vi++; 3405 x[i] = *v++ *sum; /* x[i]=aa[adiag[i]]*sum; v++; */ 3406 } 3407 3408 ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr); 3409 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3410 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 3411 PetscFunctionReturn(0); 3412 } 3413 3414 #undef __FUNCT__ 3415 #define __FUNCT__ "MatSolve_SeqAIJ_newdatastruct" 3416 PetscErrorCode MatSolve_SeqAIJ_newdatastruct(Mat A,Vec bb,Vec xx) 3417 { 3418 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3419 IS iscol = a->col,isrow = a->row; 3420 PetscErrorCode ierr; 3421 PetscInt i,n=A->rmap->n,*vi,*ai=a->i,*aj=a->j,nz,k; 3422 const PetscInt *rout,*cout,*r,*c; 3423 PetscScalar *x,*tmp,*tmps,sum; 3424 const PetscScalar *b; 3425 const MatScalar *aa = a->a,*v; 3426 3427 PetscFunctionBegin; 3428 if (!n) PetscFunctionReturn(0); 3429 3430 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3431 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 3432 tmp = a->solve_work; 3433 3434 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 3435 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 3436 3437 /* forward solve the lower triangular */ 3438 tmp[0] = b[*r++]; 3439 tmps = tmp; 3440 v = aa; 3441 vi = aj; 3442 for (i=1; i<n; i++) { 3443 nz = ai[i+1] - ai[i]; 3444 sum = b[*r++]; 3445 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 3446 tmp[i] = sum; 3447 v += nz; vi += nz; 3448 } 3449 3450 /* backward solve the upper triangular */ 3451 k = n+1; 3452 v = aa + ai[k]; /* 1st entry of U(n-1,:) */ 3453 vi = aj + ai[k]; 3454 for (i=n-1; i>=0; i--){ 3455 k = 2*n-i; 3456 nz = ai[k +1] - ai[k] - 1; 3457 sum = tmp[i]; 3458 PetscSparseDenseMinusDot(sum,tmps,v,vi,nz); 3459 x[c[i]] = tmp[i] = sum*v[nz]; /* v[nz] = aa[adiag[i]] */ 3460 v += nz+1; vi += nz+1; 3461 } 3462 3463 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 3464 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 3465 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3466 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 3467 ierr = PetscLogFlops(2*a->nz - A->cmap->n);CHKERRQ(ierr); 3468 PetscFunctionReturn(0); 3469 } 3470 3471 #undef __FUNCT__ 3472 #define __FUNCT__ "MatSolve_SeqAIJ_newdatastruct_v2" 3473 PetscErrorCode MatSolve_SeqAIJ_newdatastruct_v2(Mat A,Vec bb,Vec xx) 3474 { 3475 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3476 IS iscol = a->col,isrow = a->row; 3477 PetscErrorCode ierr; 3478 PetscInt i,n=A->rmap->n,*vi,*ai=a->i,*aj=a->j,*adiag = a->diag,nz; 3479 const PetscInt *rout,*cout,*r,*c; 3480 PetscScalar *x,*tmp,sum; 3481 const PetscScalar *b; 3482 const MatScalar *aa = a->a,*v; 3483 3484 PetscFunctionBegin; 3485 if (!n) PetscFunctionReturn(0); 3486 3487 ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3488 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 3489 tmp = a->solve_work; 3490 3491 ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 3492 ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 3493 3494 /* forward solve the lower triangular */ 3495 /* tmp[0] = b[*r++]; */ 3496 tmp[0] = b[r[0]]; 3497 v = aa; 3498 vi = aj; 3499 for (i=1; i<n; i++) { 3500 nz = ai[i+1] - ai[i]; 3501 sum = b[r[i]]; 3502 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 3503 tmp[i] = sum; 3504 v += nz; vi += nz; 3505 } 3506 3507 /* backward solve the upper triangular */ 3508 /* v = aa + ai[k]; *//* 1st entry of U(n-1,:) */ 3509 /* vi = aj + ai[k]; */ 3510 v = aa + adiag[n-1]; 3511 vi = aj + adiag[n-1]; 3512 for (i=n-1; i>=0; i--){ 3513 /* nz = ai[k +1] - ai[k] - 1;*/ 3514 nz = adiag[i]-adiag[i+1]-1; 3515 sum = tmp[i]; 3516 PetscSparseDenseMinusDot(sum,tmp,v,vi,nz); 3517 x[c[i]] = tmp[i] = sum*v[nz]; /* v[nz] = aa[adiag[i]] */ 3518 v += nz+1; vi += nz+1; 3519 } 3520 3521 ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 3522 ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 3523 ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr); 3524 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 3525 ierr = PetscLogFlops(2*a->nz - A->cmap->n);CHKERRQ(ierr); 3526 PetscFunctionReturn(0); 3527 } 3528 3529 #undef __FUNCT__ 3530 #define __FUNCT__ "MatILUDTFactor_SeqAIJ" 3531 PetscErrorCode MatILUDTFactor_SeqAIJ(Mat A,IS isrow,IS iscol,const MatFactorInfo *info,Mat *fact) 3532 { 3533 Mat B = *fact; 3534 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b; 3535 IS isicol; 3536 PetscErrorCode ierr; 3537 const PetscInt *r,*ic; 3538 PetscInt i,n=A->rmap->n,*ai=a->i,*aj=a->j,*ajtmp,*adiag; 3539 PetscInt *bi,*bj,*bdiag,*bdiag_rev; 3540 PetscInt row,nzi,nzi_bl,nzi_bu,*im,nzi_al,nzi_au; 3541 PetscInt nlnk,*lnk; 3542 PetscBT lnkbt; 3543 PetscTruth row_identity,icol_identity,both_identity; 3544 MatScalar *aatmp,*pv,*batmp,*ba,*rtmp,*pc,multiplier,*vtmp,diag_tmp; 3545 const PetscInt *ics; 3546 PetscInt j,nz,*pj,*bjtmp,k,ncut,*jtmp; 3547 PetscReal dt=info->dt,dtcol=info->dtcol,shift=info->shiftinblocks; 3548 PetscInt dtcount=(PetscInt)info->dtcount,nnz_max; 3549 PetscTruth missing; 3550 3551 PetscFunctionBegin; 3552 3553 if (dt == PETSC_DEFAULT) dt = 0.005; 3554 if (dtcol == PETSC_DEFAULT) dtcol = 0.01; /* XXX unused! */ 3555 if (dtcount == PETSC_DEFAULT) dtcount = (PetscInt)(1.5*a->rmax); 3556 3557 /* ------- symbolic factorization, can be reused ---------*/ 3558 ierr = MatMissingDiagonal(A,&missing,&i);CHKERRQ(ierr); 3559 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",i); 3560 adiag=a->diag; 3561 3562 ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 3563 3564 /* bdiag is location of diagonal in factor */ 3565 ierr = PetscMalloc((2*n+2)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr); 3566 bdiag_rev = bdiag + n+1; 3567 3568 /* allocate row pointers bi */ 3569 ierr = PetscMalloc((2*n+2)*sizeof(PetscInt),&bi);CHKERRQ(ierr); 3570 3571 /* allocate bj and ba; max num of nonzero entries is (ai[n]+2*n*dtcount+2) */ 3572 if (dtcount > n-1) dtcount = n-1; /* diagonal is excluded */ 3573 nnz_max = ai[n]+2*n*dtcount+2; 3574 3575 ierr = PetscMalloc((nnz_max+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr); 3576 ierr = PetscMalloc((nnz_max+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr); 3577 3578 /* put together the new matrix */ 3579 ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 3580 ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr); 3581 b = (Mat_SeqAIJ*)(B)->data; 3582 b->free_a = PETSC_TRUE; 3583 b->free_ij = PETSC_TRUE; 3584 b->singlemalloc = PETSC_FALSE; 3585 b->a = ba; 3586 b->j = bj; 3587 b->i = bi; 3588 b->diag = bdiag; 3589 b->ilen = 0; 3590 b->imax = 0; 3591 b->row = isrow; 3592 b->col = iscol; 3593 ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 3594 ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 3595 b->icol = isicol; 3596 ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 3597 3598 ierr = PetscLogObjectMemory(B,nnz_max*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 3599 b->maxnz = nnz_max; 3600 3601 (B)->factor = MAT_FACTOR_ILUDT; 3602 (B)->info.factor_mallocs = 0; 3603 (B)->info.fill_ratio_given = ((PetscReal)nnz_max)/((PetscReal)ai[n]); 3604 CHKMEMQ; 3605 /* ------- end of symbolic factorization ---------*/ 3606 3607 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 3608 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 3609 ics = ic; 3610 3611 /* linked list for storing column indices of the active row */ 3612 nlnk = n + 1; 3613 ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3614 3615 /* im: used by PetscLLAddSortedLU(); jtmp: working array for column indices of active row */ 3616 ierr = PetscMalloc((2*n+1)*sizeof(PetscInt),&im);CHKERRQ(ierr); 3617 jtmp = im + n; 3618 /* rtmp, vtmp: working arrays for sparse and contiguous row entries of active row */ 3619 ierr = PetscMalloc((2*n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 3620 ierr = PetscMemzero(rtmp,(n+1)*sizeof(PetscScalar));CHKERRQ(ierr); 3621 vtmp = rtmp + n; 3622 3623 bi[0] = 0; 3624 bdiag[0] = nnz_max-1; /* location of diag[0] in factor B */ 3625 bdiag_rev[n] = bdiag[0]; 3626 bi[2*n+1] = bdiag[0]+1; /* endof bj and ba array */ 3627 for (i=0; i<n; i++) { 3628 /* copy initial fill into linked list */ 3629 nzi = 0; /* nonzeros for active row i */ 3630 nzi = ai[r[i]+1] - ai[r[i]]; 3631 if (!nzi) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i); 3632 nzi_al = adiag[r[i]] - ai[r[i]]; 3633 nzi_au = ai[r[i]+1] - adiag[r[i]] -1; 3634 ajtmp = aj + ai[r[i]]; 3635 ierr = PetscLLAddPerm(nzi,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr); 3636 3637 /* load in initial (unfactored row) */ 3638 aatmp = a->a + ai[r[i]]; 3639 for (j=0; j<nzi; j++) { 3640 rtmp[ics[*ajtmp++]] = *aatmp++; 3641 } 3642 3643 /* add pivot rows into linked list */ 3644 row = lnk[n]; 3645 while (row < i ) { 3646 nzi_bl = bi[row+1] - bi[row] + 1; 3647 bjtmp = bj + bdiag[row+1]+1; /* points to 1st column next to the diagonal in U */ 3648 ierr = PetscLLAddSortedLU(bjtmp,row,nlnk,lnk,lnkbt,i,nzi_bl,im);CHKERRQ(ierr); 3649 nzi += nlnk; 3650 row = lnk[row]; 3651 } 3652 3653 /* copy data from lnk into jtmp, then initialize lnk */ 3654 ierr = PetscLLClean(n,n,nzi,lnk,jtmp,lnkbt);CHKERRQ(ierr); 3655 3656 /* numerical factorization */ 3657 bjtmp = jtmp; 3658 row = *bjtmp++; /* 1st pivot row */ 3659 while ( row < i ) { 3660 pc = rtmp + row; 3661 pv = ba + bdiag[row]; /* 1./(diag of the pivot row) */ 3662 multiplier = (*pc) * (*pv); 3663 *pc = multiplier; 3664 if (PetscAbsScalar(*pc) > dt){ /* apply tolerance dropping rule */ 3665 pj = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */ 3666 pv = ba + bdiag[row+1] + 1; 3667 /* if (multiplier < -1.0 or multiplier >1.0) printf("row/prow %d, %d, multiplier %g\n",i,row,multiplier); */ 3668 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */ 3669 for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++); 3670 ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); 3671 } 3672 row = *bjtmp++; 3673 } 3674 3675 /* copy sparse rtmp into contiguous vtmp; separate L and U part */ 3676 diag_tmp = rtmp[i]; /* save diagonal value - may not needed?? */ 3677 nzi_bl = 0; j = 0; 3678 while (jtmp[j] < i){ /* Note: jtmp is sorted */ 3679 vtmp[j] = rtmp[jtmp[j]]; rtmp[jtmp[j]]=0.0; 3680 nzi_bl++; j++; 3681 } 3682 nzi_bu = nzi - nzi_bl -1; 3683 while (j < nzi){ 3684 vtmp[j] = rtmp[jtmp[j]]; rtmp[jtmp[j]]=0.0; 3685 j++; 3686 } 3687 3688 bjtmp = bj + bi[i]; 3689 batmp = ba + bi[i]; 3690 /* apply level dropping rule to L part */ 3691 ncut = nzi_al + dtcount; 3692 if (ncut < nzi_bl){ 3693 ierr = PetscSortSplit(ncut,nzi_bl,vtmp,jtmp);CHKERRQ(ierr); 3694 ierr = PetscSortIntWithScalarArray(ncut,jtmp,vtmp);CHKERRQ(ierr); 3695 } else { 3696 ncut = nzi_bl; 3697 } 3698 for (j=0; j<ncut; j++){ 3699 bjtmp[j] = jtmp[j]; 3700 batmp[j] = vtmp[j]; 3701 /* printf(" (%d,%g),",bjtmp[j],batmp[j]); */ 3702 } 3703 bi[i+1] = bi[i] + ncut; 3704 nzi = ncut + 1; 3705 3706 /* apply level dropping rule to U part */ 3707 ncut = nzi_au + dtcount; 3708 if (ncut < nzi_bu){ 3709 ierr = PetscSortSplit(ncut,nzi_bu,vtmp+nzi_bl+1,jtmp+nzi_bl+1);CHKERRQ(ierr); 3710 ierr = PetscSortIntWithScalarArray(ncut,jtmp+nzi_bl+1,vtmp+nzi_bl+1);CHKERRQ(ierr); 3711 } else { 3712 ncut = nzi_bu; 3713 } 3714 nzi += ncut; 3715 3716 /* mark bdiagonal */ 3717 bdiag[i+1] = bdiag[i] - (ncut + 1); 3718 bdiag_rev[n-i-1] = bdiag[i+1]; 3719 bi[2*n - i] = bi[2*n - i +1] - (ncut + 1); 3720 bjtmp = bj + bdiag[i]; 3721 batmp = ba + bdiag[i]; 3722 *bjtmp = i; 3723 *batmp = diag_tmp; /* rtmp[i]; */ 3724 if (*batmp == 0.0) { 3725 *batmp = dt+shift; 3726 /* printf(" row %d add shift %g\n",i,shift); */ 3727 } 3728 *batmp = 1.0/(*batmp); /* invert diagonal entries for simplier triangular solves */ 3729 /* printf(" (%d,%g),",*bjtmp,*batmp); */ 3730 3731 bjtmp = bj + bdiag[i+1]+1; 3732 batmp = ba + bdiag[i+1]+1; 3733 for (k=0; k<ncut; k++){ 3734 bjtmp[k] = jtmp[nzi_bl+1+k]; 3735 batmp[k] = vtmp[nzi_bl+1+k]; 3736 /* printf(" (%d,%g),",bjtmp[k],batmp[k]); */ 3737 } 3738 /* printf("\n"); */ 3739 3740 im[i] = nzi; /* used by PetscLLAddSortedLU() */ 3741 /* 3742 printf("row %d: bi %d, bdiag %d\n",i,bi[i],bdiag[i]); 3743 printf(" ----------------------------\n"); 3744 */ 3745 } /* for (i=0; i<n; i++) */ 3746 /* printf("end of L %d, beginning of U %d\n",bi[n],bdiag[n]); */ 3747 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]); 3748 3749 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 3750 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 3751 3752 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 3753 ierr = PetscFree(im);CHKERRQ(ierr); 3754 ierr = PetscFree(rtmp);CHKERRQ(ierr); 3755 3756 ierr = PetscLogFlops(B->cmap->n);CHKERRQ(ierr); 3757 b->maxnz = b->nz = bi[n] + bdiag[0] - bdiag[n]; 3758 3759 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 3760 ierr = ISIdentity(isicol,&icol_identity);CHKERRQ(ierr); 3761 both_identity = (PetscTruth) (row_identity && icol_identity); 3762 if (row_identity && icol_identity) { 3763 B->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 3764 } else { 3765 B->ops->solve = MatSolve_SeqAIJ_newdatastruct; 3766 } 3767 3768 B->ops->lufactorsymbolic = MatILUDTFactorSymbolic_SeqAIJ; 3769 B->ops->lufactornumeric = MatILUDTFactorNumeric_SeqAIJ; 3770 B->ops->solveadd = 0; 3771 B->ops->solvetranspose = 0; 3772 B->ops->solvetransposeadd = 0; 3773 B->ops->matsolve = 0; 3774 B->assembled = PETSC_TRUE; 3775 B->preallocated = PETSC_TRUE; 3776 PetscFunctionReturn(0); 3777 } 3778 3779 /* a wraper of MatILUDTFactor_SeqAIJ() */ 3780 #undef __FUNCT__ 3781 #define __FUNCT__ "MatILUDTFactorSymbolic_SeqAIJ" 3782 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS row,IS col,const MatFactorInfo *info) 3783 { 3784 PetscErrorCode ierr; 3785 3786 PetscFunctionBegin; 3787 ierr = MatILUDTFactor_SeqAIJ(A,row,col,info,&fact);CHKERRQ(ierr); 3788 3789 fact->ops->lufactornumeric = MatILUDTFactorNumeric_SeqAIJ; 3790 PetscFunctionReturn(0); 3791 } 3792 3793 /* 3794 same as MatLUFactorNumeric_SeqAIJ(), except using contiguous array matrix factors 3795 - intend to replace existing MatLUFactorNumeric_SeqAIJ() 3796 */ 3797 #undef __FUNCT__ 3798 #define __FUNCT__ "MatILUDTFactorNumeric_SeqAIJ" 3799 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorNumeric_SeqAIJ(Mat fact,Mat A,const MatFactorInfo *info) 3800 { 3801 Mat C=fact; 3802 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data; 3803 IS isrow = b->row,isicol = b->icol; 3804 PetscErrorCode ierr; 3805 const PetscInt *r,*ic,*ics; 3806 PetscInt i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 3807 PetscInt *ajtmp,*bjtmp,nz,nzl,nzu,row,*bdiag = b->diag,*pj; 3808 MatScalar *rtmp,*pc,multiplier,*v,*pv,*aa=a->a; 3809 PetscReal dt=info->dt,shift=info->shiftinblocks; 3810 PetscTruth row_identity, col_identity; 3811 3812 PetscFunctionBegin; 3813 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 3814 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 3815 ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 3816 ics = ic; 3817 3818 for (i=0; i<n; i++){ 3819 /* initialize rtmp array */ 3820 nzl = bi[i+1] - bi[i]; /* num of nozeros in L(i,:) */ 3821 bjtmp = bj + bi[i]; 3822 for (j=0; j<nzl; j++) rtmp[*bjtmp++] = 0.0; 3823 rtmp[i] = 0.0; 3824 nzu = bdiag[i] - bdiag[i+1]; /* num of nozeros in U(i,:) */ 3825 bjtmp = bj + bdiag[i+1] + 1; 3826 for (j=0; j<nzu; j++) rtmp[*bjtmp++] = 0.0; 3827 3828 /* load in initial unfactored row of A */ 3829 /* printf("row %d\n",i); */ 3830 nz = ai[r[i]+1] - ai[r[i]]; 3831 ajtmp = aj + ai[r[i]]; 3832 v = aa + ai[r[i]]; 3833 for (j=0; j<nz; j++) { 3834 rtmp[ics[*ajtmp++]] = v[j]; 3835 /* printf(" (%d,%g),",ics[ajtmp[j]],rtmp[ics[ajtmp[j]]]); */ 3836 } 3837 /* printf("\n"); */ 3838 3839 /* numerical factorization */ 3840 bjtmp = bj + bi[i]; /* point to 1st entry of L(i,:) */ 3841 nzl = bi[i+1] - bi[i]; /* num of entries in L(i,:) */ 3842 k = 0; 3843 while (k < nzl){ 3844 row = *bjtmp++; 3845 /* printf(" prow %d\n",row); */ 3846 pc = rtmp + row; 3847 pv = b->a + bdiag[row]; /* 1./(diag of the pivot row) */ 3848 multiplier = (*pc) * (*pv); 3849 *pc = multiplier; 3850 if (PetscAbsScalar(multiplier) > dt){ 3851 pj = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */ 3852 pv = b->a + bdiag[row+1] + 1; 3853 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */ 3854 for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++); 3855 /* ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); */ 3856 } 3857 k++; 3858 } 3859 3860 /* finished row so stick it into b->a */ 3861 /* L-part */ 3862 pv = b->a + bi[i] ; 3863 pj = bj + bi[i] ; 3864 nzl = bi[i+1] - bi[i]; 3865 for (j=0; j<nzl; j++) { 3866 pv[j] = rtmp[pj[j]]; 3867 /* printf(" (%d,%g),",pj[j],pv[j]); */ 3868 } 3869 3870 /* diagonal: invert diagonal entries for simplier triangular solves */ 3871 if (rtmp[i] == 0.0) rtmp[i] = dt+shift; 3872 b->a[bdiag[i]] = 1.0/rtmp[i]; 3873 /* printf(" (%d,%g),",i,b->a[bdiag[i]]); */ 3874 3875 /* U-part */ 3876 pv = b->a + bdiag[i+1] + 1; 3877 pj = bj + bdiag[i+1] + 1; 3878 nzu = bdiag[i] - bdiag[i+1] - 1; 3879 for (j=0; j<nzu; j++) { 3880 pv[j] = rtmp[pj[j]]; 3881 /* printf(" (%d,%g),",pj[j],pv[j]); */ 3882 } 3883 /* printf("\n"); */ 3884 } 3885 3886 ierr = PetscFree(rtmp);CHKERRQ(ierr); 3887 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 3888 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 3889 3890 ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 3891 ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr); 3892 if (row_identity && col_identity) { 3893 C->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_newdatastruct; 3894 } else { 3895 C->ops->solve = MatSolve_SeqAIJ_newdatastruct; 3896 } 3897 C->ops->solveadd = 0; 3898 C->ops->solvetranspose = 0; 3899 C->ops->solvetransposeadd = 0; 3900 C->ops->matsolve = 0; 3901 C->assembled = PETSC_TRUE; 3902 C->preallocated = PETSC_TRUE; 3903 ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr); 3904 PetscFunctionReturn(0); 3905 } 3906