xref: /petsc/src/mat/impls/aij/seq/bas/basfactor.c (revision 66af8762ec03dbef0e079729eb2a1734a35ed7ff)
1 #include <../src/mat/impls/aij/seq/aij.h>
2 #include <../src/mat/impls/sbaij/seq/sbaij.h>
3 #include <../src/mat/impls/aij/seq/bas/spbas.h>
4 
5 static PetscErrorCode MatICCFactorSymbolic_SeqAIJ_Bas(Mat fact, Mat A, IS perm, const MatFactorInfo *info)
6 {
7   Mat_SeqAIJ     *a = (Mat_SeqAIJ *)A->data;
8   Mat_SeqSBAIJ   *b;
9   PetscBool       perm_identity, missing;
10   PetscInt        reallocs = 0, i, *ai = a->i, *aj = a->j, am = A->rmap->n, *ui;
11   const PetscInt *rip, *riip;
12   PetscInt        j;
13   PetscInt        d;
14   PetscInt        ncols, *cols, *uj;
15   PetscReal       fill = info->fill, levels = info->levels;
16   IS              iperm;
17   spbas_matrix    Pattern_0, Pattern_P;
18 
19   PetscFunctionBegin;
20   PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Must be square matrix, rows %" PetscInt_FMT " columns %" PetscInt_FMT, A->rmap->n, A->cmap->n);
21   PetscCall(MatMissingDiagonal(A, &missing, &d));
22   PetscCheck(!missing, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entry %" PetscInt_FMT, d);
23   PetscCall(ISIdentity(perm, &perm_identity));
24   PetscCall(ISInvertPermutation(perm, PETSC_DECIDE, &iperm));
25 
26   /* ICC(0) without matrix ordering: simply copies fill pattern */
27   if (!levels && perm_identity) {
28     PetscCall(PetscMalloc1(am + 1, &ui));
29     ui[0] = 0;
30 
31     for (i = 0; i < am; i++) ui[i + 1] = ui[i] + ai[i + 1] - a->diag[i];
32     PetscCall(PetscMalloc1(ui[am] + 1, &uj));
33     cols = uj;
34     for (i = 0; i < am; i++) {
35       aj    = a->j + a->diag[i];
36       ncols = ui[i + 1] - ui[i];
37       for (j = 0; j < ncols; j++) *cols++ = *aj++;
38     }
39   } else { /* case: levels>0 || (levels=0 && !perm_identity) */
40     PetscCall(ISGetIndices(iperm, &riip));
41     PetscCall(ISGetIndices(perm, &rip));
42 
43     /* Create spbas_matrix for pattern */
44     PetscCall(spbas_pattern_only(am, am, ai, aj, &Pattern_0));
45 
46     /* Apply the permutation */
47     PetscCall(spbas_apply_reordering(&Pattern_0, rip, riip));
48 
49     /* Raise the power */
50     PetscCall(spbas_power(Pattern_0, (int)levels + 1, &Pattern_P));
51     PetscCall(spbas_delete(Pattern_0));
52 
53     /* Keep only upper triangle of pattern */
54     PetscCall(spbas_keep_upper(&Pattern_P));
55 
56     /* Convert to Sparse Row Storage  */
57     PetscCall(spbas_matrix_to_crs(Pattern_P, NULL, &ui, &uj));
58     PetscCall(spbas_delete(Pattern_P));
59   } /* end of case: levels>0 || (levels=0 && !perm_identity) */
60 
61   /* put together the new matrix in MATSEQSBAIJ format */
62 
63   b               = (Mat_SeqSBAIJ *)(fact)->data;
64   b->singlemalloc = PETSC_FALSE;
65 
66   PetscCall(PetscMalloc1(ui[am] + 1, &b->a));
67 
68   b->j    = uj;
69   b->i    = ui;
70   b->diag = NULL;
71   b->ilen = NULL;
72   b->imax = NULL;
73   b->row  = perm;
74   b->col  = perm;
75 
76   PetscCall(PetscObjectReference((PetscObject)perm));
77   PetscCall(PetscObjectReference((PetscObject)perm));
78 
79   b->icol          = iperm;
80   b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */
81   PetscCall(PetscMalloc1(am + 1, &b->solve_work));
82   b->maxnz = b->nz = ui[am];
83   b->free_a        = PETSC_TRUE;
84   b->free_ij       = PETSC_TRUE;
85 
86   (fact)->info.factor_mallocs   = reallocs;
87   (fact)->info.fill_ratio_given = fill;
88   if (ai[am] != 0) {
89     (fact)->info.fill_ratio_needed = ((PetscReal)ui[am]) / ((PetscReal)ai[am]);
90   } else {
91     (fact)->info.fill_ratio_needed = 0.0;
92   }
93   /*  (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_inplace; */
94   PetscFunctionReturn(PETSC_SUCCESS);
95 }
96 
97 static PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_Bas(Mat B, Mat A, const MatFactorInfo *info)
98 {
99   Mat             C  = B;
100   Mat_SeqSBAIJ   *b  = (Mat_SeqSBAIJ *)C->data;
101   IS              ip = b->row, iip = b->icol;
102   const PetscInt *rip, *riip;
103   PetscInt        mbs = A->rmap->n, *bi = b->i, *bj = b->j;
104   MatScalar      *ba      = b->a;
105   PetscReal       shiftnz = info->shiftamount;
106   PetscReal       droptol = -1;
107   PetscBool       perm_identity;
108   spbas_matrix    Pattern, matrix_L, matrix_LT;
109   PetscReal       mem_reduction;
110 
111   PetscFunctionBegin;
112   /* Reduce memory requirements:   erase values of B-matrix */
113   PetscCall(PetscFree(ba));
114   /*   Compress (maximum) sparseness pattern of B-matrix */
115   PetscCall(spbas_compress_pattern(bi, bj, mbs, mbs, SPBAS_DIAGONAL_OFFSETS, &Pattern, &mem_reduction));
116   PetscCall(PetscFree(bi));
117   PetscCall(PetscFree(bj));
118 
119   PetscCall(PetscInfo(NULL, "    compression rate for spbas_compress_pattern %g \n", (double)mem_reduction));
120 
121   /* Make Cholesky decompositions with larger Manteuffel shifts until no more    negative diagonals are found. */
122   PetscCall(ISGetIndices(ip, &rip));
123   PetscCall(ISGetIndices(iip, &riip));
124 
125   if (info->usedt) droptol = info->dt;
126 
127   for (int ierr = NEGATIVE_DIAGONAL; ierr == NEGATIVE_DIAGONAL;) {
128     PetscBool success;
129 
130     ierr = (int)spbas_incomplete_cholesky(A, rip, riip, Pattern, droptol, shiftnz, &matrix_LT, &success);
131     if (!success) {
132       shiftnz *= 1.5;
133       if (shiftnz < 1e-5) shiftnz = 1e-5;
134       PetscCall(PetscInfo(NULL, "spbas_incomplete_cholesky found a negative diagonal. Trying again with Manteuffel shift=%g\n", (double)shiftnz));
135     }
136   }
137   PetscCall(spbas_delete(Pattern));
138 
139   PetscCall(PetscInfo(NULL, "    memory_usage for  spbas_incomplete_cholesky  %g bytes per row\n", (double)(PetscReal)(spbas_memory_requirement(matrix_LT) / (PetscReal)mbs)));
140 
141   PetscCall(ISRestoreIndices(ip, &rip));
142   PetscCall(ISRestoreIndices(iip, &riip));
143 
144   /* Convert spbas_matrix to compressed row storage */
145   PetscCall(spbas_transpose(matrix_LT, &matrix_L));
146   PetscCall(spbas_delete(matrix_LT));
147   PetscCall(spbas_matrix_to_crs(matrix_L, &ba, &bi, &bj));
148   b->i = bi;
149   b->j = bj;
150   b->a = ba;
151   PetscCall(spbas_delete(matrix_L));
152 
153   /* Set the appropriate solution functions */
154   PetscCall(ISIdentity(ip, &perm_identity));
155   if (perm_identity) {
156     (B)->ops->solve          = MatSolve_SeqSBAIJ_1_NaturalOrdering_inplace;
157     (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1_NaturalOrdering_inplace;
158     (B)->ops->forwardsolve   = MatForwardSolve_SeqSBAIJ_1_NaturalOrdering_inplace;
159     (B)->ops->backwardsolve  = MatBackwardSolve_SeqSBAIJ_1_NaturalOrdering_inplace;
160   } else {
161     (B)->ops->solve          = MatSolve_SeqSBAIJ_1_inplace;
162     (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1_inplace;
163     (B)->ops->forwardsolve   = MatForwardSolve_SeqSBAIJ_1_inplace;
164     (B)->ops->backwardsolve  = MatBackwardSolve_SeqSBAIJ_1_inplace;
165   }
166 
167   C->assembled    = PETSC_TRUE;
168   C->preallocated = PETSC_TRUE;
169 
170   PetscCall(PetscLogFlops(C->rmap->n));
171   PetscFunctionReturn(PETSC_SUCCESS);
172 }
173 
174 static PetscErrorCode MatFactorGetSolverType_seqaij_bas(Mat A, MatSolverType *type)
175 {
176   PetscFunctionBegin;
177   *type = MATSOLVERBAS;
178   PetscFunctionReturn(PETSC_SUCCESS);
179 }
180 
181 PETSC_INTERN PetscErrorCode MatGetFactor_seqaij_bas(Mat A, MatFactorType ftype, Mat *B)
182 {
183   PetscInt n = A->rmap->n;
184 
185   PetscFunctionBegin;
186   PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
187   PetscCall(MatSetSizes(*B, n, n, n, n));
188   if (ftype == MAT_FACTOR_ICC) {
189     PetscCall(MatSetType(*B, MATSEQSBAIJ));
190     PetscCall(MatSeqSBAIJSetPreallocation(*B, 1, MAT_SKIP_ALLOCATION, NULL));
191 
192     (*B)->ops->iccfactorsymbolic     = MatICCFactorSymbolic_SeqAIJ_Bas;
193     (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_Bas;
194     PetscCall(PetscObjectComposeFunction((PetscObject)*B, "MatFactorGetSolverType_C", MatFactorGetSolverType_seqaij_bas));
195     PetscCall(PetscStrallocpy(MATORDERINGND, (char **)&(*B)->preferredordering[MAT_FACTOR_LU]));
196     PetscCall(PetscStrallocpy(MATORDERINGND, (char **)&(*B)->preferredordering[MAT_FACTOR_CHOLESKY]));
197   } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Factor type not supported");
198   (*B)->factortype = ftype;
199 
200   PetscCall(PetscFree((*B)->solvertype));
201   PetscCall(PetscStrallocpy(MATSOLVERBAS, &(*B)->solvertype));
202   (*B)->canuseordering = PETSC_TRUE;
203   PetscCall(PetscStrallocpy(MATORDERINGNATURAL, (char **)&(*B)->preferredordering[MAT_FACTOR_ICC]));
204   PetscFunctionReturn(PETSC_SUCCESS);
205 }
206