xref: /petsc/src/mat/impls/aij/seq/klu/klu.c (revision 2ff79c18c26c94ed8cb599682f680f231dca6444)
1 /*
2    Provides an interface to the KLUv1.2 sparse solver
3 
4    When build with PETSC_USE_64BIT_INDICES this will use SuiteSparse_long as the
5    integer type in KLU, otherwise it will use int. This means
6    all integers in this file are simply declared as PetscInt. Also it means
7    that KLU SuiteSparse_long version MUST be built with 64-bit integers when used.
8 
9 */
10 #include <../src/mat/impls/aij/seq/aij.h>
11 
12 #if defined(PETSC_USE_64BIT_INDICES)
13   #define klu_K_defaults                        klu_l_defaults
14   #define klu_K_analyze(a, b, c, d)             klu_l_analyze((SuiteSparse_long)a, (SuiteSparse_long *)b, (SuiteSparse_long *)c, d)
15   #define klu_K_analyze_given(a, b, c, d, e, f) klu_l_analyze_given((SuiteSparse_long)a, (SuiteSparse_long *)b, (SuiteSparse_long *)c, (SuiteSparse_long *)d, (SuiteSparse_long *)e, f)
16   #define klu_K_free_symbolic                   klu_l_free_symbolic
17   #define klu_K_free_numeric                    klu_l_free_numeric
18   #define klu_K_common                          klu_l_common
19   #define klu_K_symbolic                        klu_l_symbolic
20   #define klu_K_numeric                         klu_l_numeric
21   #if defined(PETSC_USE_COMPLEX)
22     #define klu_K_factor(a, b, c, d, e) klu_zl_factor((SuiteSparse_long *)a, (SuiteSparse_long *)b, c, d, e);
23     #define klu_K_solve                 klu_zl_solve
24     #define klu_K_tsolve                klu_zl_tsolve
25     #define klu_K_refactor              klu_zl_refactor
26     #define klu_K_sort                  klu_zl_sort
27     #define klu_K_flops                 klu_zl_flops
28     #define klu_K_rgrowth               klu_zl_rgrowth
29     #define klu_K_condest               klu_zl_condest
30     #define klu_K_rcond                 klu_zl_rcond
31     #define klu_K_scale                 klu_zl_scale
32   #else
33     #define klu_K_factor(a, b, c, d, e) klu_l_factor((SuiteSparse_long *)a, (SuiteSparse_long *)b, c, d, e);
34     #define klu_K_solve                 klu_l_solve
35     #define klu_K_tsolve                klu_l_tsolve
36     #define klu_K_refactor              klu_l_refactor
37     #define klu_K_sort                  klu_l_sort
38     #define klu_K_flops                 klu_l_flops
39     #define klu_K_rgrowth               klu_l_rgrowth
40     #define klu_K_condest               klu_l_condest
41     #define klu_K_rcond                 klu_l_rcond
42     #define klu_K_scale                 klu_l_scale
43   #endif
44 #else
45   #define klu_K_defaults      klu_defaults
46   #define klu_K_analyze       klu_analyze
47   #define klu_K_analyze_given klu_analyze_given
48   #define klu_K_free_symbolic klu_free_symbolic
49   #define klu_K_free_numeric  klu_free_numeric
50   #define klu_K_common        klu_common
51   #define klu_K_symbolic      klu_symbolic
52   #define klu_K_numeric       klu_numeric
53   #if defined(PETSC_USE_COMPLEX)
54     #define klu_K_factor   klu_z_factor
55     #define klu_K_solve    klu_z_solve
56     #define klu_K_tsolve   klu_z_tsolve
57     #define klu_K_refactor klu_z_refactor
58     #define klu_K_sort     klu_z_sort
59     #define klu_K_flops    klu_z_flops
60     #define klu_K_rgrowth  klu_z_rgrowth
61     #define klu_K_condest  klu_z_condest
62     #define klu_K_rcond    klu_z_rcond
63     #define klu_K_scale    klu_z_scale
64   #else
65     #define klu_K_factor   klu_factor
66     #define klu_K_solve    klu_solve
67     #define klu_K_tsolve   klu_tsolve
68     #define klu_K_refactor klu_refactor
69     #define klu_K_sort     klu_sort
70     #define klu_K_flops    klu_flops
71     #define klu_K_rgrowth  klu_rgrowth
72     #define klu_K_condest  klu_condest
73     #define klu_K_rcond    klu_rcond
74     #define klu_K_scale    klu_scale
75   #endif
76 #endif
77 
78 EXTERN_C_BEGIN
79 #include <klu.h>
80 EXTERN_C_END
81 
82 static const char *KluOrderingTypes[] = {"AMD", "COLAMD"};
83 static const char *scale[]            = {"NONE", "SUM", "MAX"};
84 
85 typedef struct {
86   klu_K_common    Common;
87   klu_K_symbolic *Symbolic;
88   klu_K_numeric  *Numeric;
89   PetscInt       *perm_c, *perm_r;
90   MatStructure    flg;
91   PetscBool       PetscMatOrdering;
92   PetscBool       CleanUpKLU;
93 } Mat_KLU;
94 
95 static PetscErrorCode MatDestroy_KLU(Mat A)
96 {
97   Mat_KLU *lu = (Mat_KLU *)A->data;
98 
99   PetscFunctionBegin;
100   if (lu->CleanUpKLU) {
101     klu_K_free_symbolic(&lu->Symbolic, &lu->Common);
102     klu_K_free_numeric(&lu->Numeric, &lu->Common);
103     PetscCall(PetscFree2(lu->perm_r, lu->perm_c));
104   }
105   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
106   PetscCall(PetscFree(A->data));
107   PetscFunctionReturn(PETSC_SUCCESS);
108 }
109 
110 static PetscErrorCode MatSolveTranspose_KLU(Mat A, Vec b, Vec x)
111 {
112   Mat_KLU     *lu = (Mat_KLU *)A->data;
113   PetscScalar *xa;
114   PetscInt     status;
115 
116   PetscFunctionBegin;
117   /* KLU uses a column major format, solve Ax = b by klu_*_solve */
118   PetscCall(VecCopy(b, x)); /* klu_solve stores the solution in rhs */
119   PetscCall(VecGetArray(x, &xa));
120   status = klu_K_solve(lu->Symbolic, lu->Numeric, A->rmap->n, 1, (PetscReal *)xa, &lu->Common);
121   PetscCheck(status == 1, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Solve failed");
122   PetscCall(VecRestoreArray(x, &xa));
123   PetscFunctionReturn(PETSC_SUCCESS);
124 }
125 
126 static PetscErrorCode MatSolve_KLU(Mat A, Vec b, Vec x)
127 {
128   Mat_KLU     *lu = (Mat_KLU *)A->data;
129   PetscScalar *xa;
130   PetscInt     status;
131 
132   PetscFunctionBegin;
133   /* KLU uses a column major format, solve A^Tx = b by klu_*_tsolve */
134   PetscCall(VecCopy(b, x)); /* klu_solve stores the solution in rhs */
135   PetscCall(VecGetArray(x, &xa));
136 #if defined(PETSC_USE_COMPLEX)
137   PetscInt conj_solve = 1;
138   status              = klu_K_tsolve(lu->Symbolic, lu->Numeric, A->rmap->n, 1, (PetscReal *)xa, conj_solve, &lu->Common); /* conjugate solve */
139 #else
140   status = klu_K_tsolve(lu->Symbolic, lu->Numeric, A->rmap->n, 1, xa, &lu->Common);
141 #endif
142   PetscCheck(status == 1, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Solve failed");
143   PetscCall(VecRestoreArray(x, &xa));
144   PetscFunctionReturn(PETSC_SUCCESS);
145 }
146 
147 static PetscErrorCode MatLUFactorNumeric_KLU(Mat F, Mat A, const MatFactorInfo *info)
148 {
149   Mat_KLU     *lu = (Mat_KLU *)F->data;
150   Mat_SeqAIJ  *a  = (Mat_SeqAIJ *)A->data;
151   PetscInt    *ai = a->i, *aj = a->j;
152   PetscScalar *av = a->a;
153 
154   PetscFunctionBegin;
155   /* numeric factorization of A' */
156 
157   if (lu->flg == SAME_NONZERO_PATTERN && lu->Numeric) klu_K_free_numeric(&lu->Numeric, &lu->Common);
158   lu->Numeric = klu_K_factor(ai, aj, (PetscReal *)av, lu->Symbolic, &lu->Common);
159   PetscCheck(lu->Numeric, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Numeric factorization failed");
160 
161   lu->flg                = SAME_NONZERO_PATTERN;
162   lu->CleanUpKLU         = PETSC_TRUE;
163   F->ops->solve          = MatSolve_KLU;
164   F->ops->solvetranspose = MatSolveTranspose_KLU;
165   PetscFunctionReturn(PETSC_SUCCESS);
166 }
167 
168 static PetscErrorCode MatLUFactorSymbolic_KLU(Mat F, Mat A, IS r, IS c, const MatFactorInfo *info)
169 {
170   Mat_SeqAIJ     *a  = (Mat_SeqAIJ *)A->data;
171   Mat_KLU        *lu = (Mat_KLU *)F->data;
172   PetscInt        i, *ai = a->i, *aj = a->j, m = A->rmap->n, n = A->cmap->n;
173   const PetscInt *ra, *ca;
174 
175   PetscFunctionBegin;
176   if (lu->PetscMatOrdering) {
177     PetscCall(ISGetIndices(r, &ra));
178     PetscCall(ISGetIndices(c, &ca));
179     PetscCall(PetscMalloc2(m, &lu->perm_r, n, &lu->perm_c));
180     /* we cannot simply memcpy on 64-bit archs */
181     for (i = 0; i < m; i++) lu->perm_r[i] = ra[i];
182     for (i = 0; i < n; i++) lu->perm_c[i] = ca[i];
183     PetscCall(ISRestoreIndices(r, &ra));
184     PetscCall(ISRestoreIndices(c, &ca));
185   }
186 
187   /* symbolic factorization of A' */
188   if (r) {
189     lu->PetscMatOrdering = PETSC_TRUE;
190     lu->Symbolic         = klu_K_analyze_given(n, ai, aj, lu->perm_c, lu->perm_r, &lu->Common);
191   } else { /* use klu internal ordering */
192     lu->Symbolic = klu_K_analyze(n, ai, aj, &lu->Common);
193   }
194   PetscCheck(lu->Symbolic, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Symbolic Factorization failed");
195 
196   lu->flg                 = DIFFERENT_NONZERO_PATTERN;
197   lu->CleanUpKLU          = PETSC_TRUE;
198   F->ops->lufactornumeric = MatLUFactorNumeric_KLU;
199   PetscFunctionReturn(PETSC_SUCCESS);
200 }
201 
202 static PetscErrorCode MatView_Info_KLU(Mat A, PetscViewer viewer)
203 {
204   Mat_KLU       *lu      = (Mat_KLU *)A->data;
205   klu_K_numeric *Numeric = (klu_K_numeric *)lu->Numeric;
206 
207   PetscFunctionBegin;
208   PetscCall(PetscViewerASCIIPrintf(viewer, "KLU stats:\n"));
209   PetscCall(PetscViewerASCIIPrintf(viewer, "  Number of diagonal blocks: %" PetscInt_FMT "\n", (PetscInt)Numeric->nblocks));
210   PetscCall(PetscViewerASCIIPrintf(viewer, "  Total nonzeros=%" PetscInt_FMT "\n", (PetscInt)(Numeric->lnz + Numeric->unz)));
211   PetscCall(PetscViewerASCIIPrintf(viewer, "KLU runtime parameters:\n"));
212   /* Control parameters used by numeric factorization */
213   PetscCall(PetscViewerASCIIPrintf(viewer, "  Partial pivoting tolerance: %g\n", lu->Common.tol));
214   /* BTF preordering */
215   PetscCall(PetscViewerASCIIPrintf(viewer, "  BTF preordering enabled: %" PetscInt_FMT "\n", (PetscInt)lu->Common.btf));
216   /* mat ordering */
217   if (!lu->PetscMatOrdering) PetscCall(PetscViewerASCIIPrintf(viewer, "  Ordering: %s (not using the PETSc ordering)\n", KluOrderingTypes[(int)lu->Common.ordering]));
218   /* matrix row scaling */
219   PetscCall(PetscViewerASCIIPrintf(viewer, "  Matrix row scaling: %s\n", scale[(int)lu->Common.scale]));
220   PetscFunctionReturn(PETSC_SUCCESS);
221 }
222 
223 static PetscErrorCode MatView_KLU(Mat A, PetscViewer viewer)
224 {
225   PetscBool         isascii;
226   PetscViewerFormat format;
227 
228   PetscFunctionBegin;
229   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
230   if (isascii) {
231     PetscCall(PetscViewerGetFormat(viewer, &format));
232     if (format == PETSC_VIEWER_ASCII_INFO) PetscCall(MatView_Info_KLU(A, viewer));
233   }
234   PetscFunctionReturn(PETSC_SUCCESS);
235 }
236 
237 static PetscErrorCode MatFactorGetSolverType_seqaij_klu(Mat A, MatSolverType *type)
238 {
239   PetscFunctionBegin;
240   *type = MATSOLVERKLU;
241   PetscFunctionReturn(PETSC_SUCCESS);
242 }
243 
244 /*MC
245   MATSOLVERKLU = "klu" - A matrix type providing direct solvers, LU, for sequential matrices
246   via the external package KLU.
247 
248   `./configure --download-suitesparse` to install PETSc to use KLU
249 
250   Use `-pc_type lu` `-pc_factor_mat_solver_type klu` to use this direct solver
251 
252   Consult KLU documentation for more information on the options database keys below.
253 
254   Options Database Keys:
255 + -mat_klu_pivot_tol <0.001>                  - Partial pivoting tolerance
256 . -mat_klu_use_btf <1>                        - Use BTF preordering
257 . -mat_klu_ordering <AMD>                     - KLU reordering scheme to reduce fill-in (choose one of) `AMD`, `COLAMD`, `PETSC`
258 - -mat_klu_row_scale <NONE>                   - Matrix row scaling (choose one of) `NONE`, `SUM`, `MAX`
259 
260    Level: beginner
261 
262    Note:
263    KLU is part of SuiteSparse <http://faculty.cse.tamu.edu/davis/suitesparse.html>
264 
265 .seealso: [](ch_matrices), `Mat`, `PCLU`, `MATSOLVERUMFPACK`, `MATSOLVERCHOLMOD`, `PCFactorSetMatSolverType()`, `MatSolverType`
266 M*/
267 
268 PETSC_INTERN PetscErrorCode MatGetFactor_seqaij_klu(Mat A, MatFactorType ftype, Mat *F)
269 {
270   Mat       B;
271   Mat_KLU  *lu;
272   PetscInt  m = A->rmap->n, n = A->cmap->n, idx = 0, status;
273   PetscBool flg;
274 
275   PetscFunctionBegin;
276   /* Create the factorization matrix F */
277   PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
278   PetscCall(MatSetSizes(B, PETSC_DECIDE, PETSC_DECIDE, m, n));
279   PetscCall(PetscStrallocpy("klu", &((PetscObject)B)->type_name));
280   PetscCall(MatSetUp(B));
281 
282   PetscCall(PetscNew(&lu));
283 
284   B->data                  = lu;
285   B->ops->getinfo          = MatGetInfo_External;
286   B->ops->lufactorsymbolic = MatLUFactorSymbolic_KLU;
287   B->ops->destroy          = MatDestroy_KLU;
288   B->ops->view             = MatView_KLU;
289 
290   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorGetSolverType_C", MatFactorGetSolverType_seqaij_klu));
291 
292   B->factortype   = MAT_FACTOR_LU;
293   B->assembled    = PETSC_TRUE; /* required by -ksp_view */
294   B->preallocated = PETSC_TRUE;
295 
296   PetscCall(PetscFree(B->solvertype));
297   PetscCall(PetscStrallocpy(MATSOLVERKLU, &B->solvertype));
298   B->canuseordering = PETSC_TRUE;
299   PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&B->preferredordering[MAT_FACTOR_LU]));
300 
301   /* initializations */
302   /* get the default control parameters */
303   status = klu_K_defaults(&lu->Common);
304   PetscCheck(status > 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Initialization failed");
305 
306   lu->Common.scale = 0; /* No row scaling */
307 
308   PetscOptionsBegin(PetscObjectComm((PetscObject)B), ((PetscObject)B)->prefix, "KLU Options", "Mat");
309   /* Partial pivoting tolerance */
310   PetscCall(PetscOptionsReal("-mat_klu_pivot_tol", "Partial pivoting tolerance", "None", lu->Common.tol, &lu->Common.tol, NULL));
311   /* BTF pre-ordering */
312   PetscCall(PetscOptionsInt("-mat_klu_use_btf", "Enable BTF preordering", "None", (PetscInt)lu->Common.btf, (PetscInt *)&lu->Common.btf, NULL));
313   /* Matrix reordering */
314   PetscCall(PetscOptionsEList("-mat_klu_ordering", "Internal ordering method", "None", KluOrderingTypes, PETSC_STATIC_ARRAY_LENGTH(KluOrderingTypes), KluOrderingTypes[0], &idx, &flg));
315   lu->Common.ordering = (int)idx;
316   /* Matrix row scaling */
317   PetscCall(PetscOptionsEList("-mat_klu_row_scale", "Matrix row scaling", "None", scale, 3, scale[0], &idx, &flg));
318   PetscOptionsEnd();
319   *F = B;
320   PetscFunctionReturn(PETSC_SUCCESS);
321 }
322