xref: /petsc/src/mat/impls/aij/seq/klu/klu.c (revision bcee047adeeb73090d7e36cc71e39fc287cdbb97)
1 
2 /*
3    Provides an interface to the KLUv1.2 sparse solver
4 
5    When build with PETSC_USE_64BIT_INDICES this will use SuiteSparse_long as the
6    integer type in KLU, otherwise it will use int. This means
7    all integers in this file are simply declared as PetscInt. Also it means
8    that KLU SuiteSparse_long version MUST be built with 64-bit integers when used.
9 
10 */
11 #include <../src/mat/impls/aij/seq/aij.h>
12 
13 #if defined(PETSC_USE_64BIT_INDICES)
14   #define klu_K_defaults                        klu_l_defaults
15   #define klu_K_analyze(a, b, c, d)             klu_l_analyze((SuiteSparse_long)a, (SuiteSparse_long *)b, (SuiteSparse_long *)c, d)
16   #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)
17   #define klu_K_free_symbolic                   klu_l_free_symbolic
18   #define klu_K_free_numeric                    klu_l_free_numeric
19   #define klu_K_common                          klu_l_common
20   #define klu_K_symbolic                        klu_l_symbolic
21   #define klu_K_numeric                         klu_l_numeric
22   #if defined(PETSC_USE_COMPLEX)
23     #define klu_K_factor(a, b, c, d, e) klu_zl_factor((SuiteSparse_long *)a, (SuiteSparse_long *)b, c, d, e);
24     #define klu_K_solve                 klu_zl_solve
25     #define klu_K_tsolve                klu_zl_tsolve
26     #define klu_K_refactor              klu_zl_refactor
27     #define klu_K_sort                  klu_zl_sort
28     #define klu_K_flops                 klu_zl_flops
29     #define klu_K_rgrowth               klu_zl_rgrowth
30     #define klu_K_condest               klu_zl_condest
31     #define klu_K_rcond                 klu_zl_rcond
32     #define klu_K_scale                 klu_zl_scale
33   #else
34     #define klu_K_factor(a, b, c, d, e) klu_l_factor((SuiteSparse_long *)a, (SuiteSparse_long *)b, c, d, e);
35     #define klu_K_solve                 klu_l_solve
36     #define klu_K_tsolve                klu_l_tsolve
37     #define klu_K_refactor              klu_l_refactor
38     #define klu_K_sort                  klu_l_sort
39     #define klu_K_flops                 klu_l_flops
40     #define klu_K_rgrowth               klu_l_rgrowth
41     #define klu_K_condest               klu_l_condest
42     #define klu_K_rcond                 klu_l_rcond
43     #define klu_K_scale                 klu_l_scale
44   #endif
45 #else
46   #define klu_K_defaults      klu_defaults
47   #define klu_K_analyze       klu_analyze
48   #define klu_K_analyze_given klu_analyze_given
49   #define klu_K_free_symbolic klu_free_symbolic
50   #define klu_K_free_numeric  klu_free_numeric
51   #define klu_K_common        klu_common
52   #define klu_K_symbolic      klu_symbolic
53   #define klu_K_numeric       klu_numeric
54   #if defined(PETSC_USE_COMPLEX)
55     #define klu_K_factor   klu_z_factor
56     #define klu_K_solve    klu_z_solve
57     #define klu_K_tsolve   klu_z_tsolve
58     #define klu_K_refactor klu_z_refactor
59     #define klu_K_sort     klu_z_sort
60     #define klu_K_flops    klu_z_flops
61     #define klu_K_rgrowth  klu_z_rgrowth
62     #define klu_K_condest  klu_z_condest
63     #define klu_K_rcond    klu_z_rcond
64     #define klu_K_scale    klu_z_scale
65   #else
66     #define klu_K_factor   klu_factor
67     #define klu_K_solve    klu_solve
68     #define klu_K_tsolve   klu_tsolve
69     #define klu_K_refactor klu_refactor
70     #define klu_K_sort     klu_sort
71     #define klu_K_flops    klu_flops
72     #define klu_K_rgrowth  klu_rgrowth
73     #define klu_K_condest  klu_condest
74     #define klu_K_rcond    klu_rcond
75     #define klu_K_scale    klu_scale
76   #endif
77 #endif
78 
79 EXTERN_C_BEGIN
80 #include <klu.h>
81 EXTERN_C_END
82 
83 static const char *KluOrderingTypes[] = {"AMD", "COLAMD"};
84 static const char *scale[]            = {"NONE", "SUM", "MAX"};
85 
86 typedef struct {
87   klu_K_common    Common;
88   klu_K_symbolic *Symbolic;
89   klu_K_numeric  *Numeric;
90   PetscInt       *perm_c, *perm_r;
91   MatStructure    flg;
92   PetscBool       PetscMatOrdering;
93   PetscBool       CleanUpKLU;
94 } Mat_KLU;
95 
96 static PetscErrorCode MatDestroy_KLU(Mat A)
97 {
98   Mat_KLU *lu = (Mat_KLU *)A->data;
99 
100   PetscFunctionBegin;
101   if (lu->CleanUpKLU) {
102     klu_K_free_symbolic(&lu->Symbolic, &lu->Common);
103     klu_K_free_numeric(&lu->Numeric, &lu->Common);
104     PetscCall(PetscFree2(lu->perm_r, lu->perm_c));
105   }
106   PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
107   PetscCall(PetscFree(A->data));
108   PetscFunctionReturn(PETSC_SUCCESS);
109 }
110 
111 static PetscErrorCode MatSolveTranspose_KLU(Mat A, Vec b, Vec x)
112 {
113   Mat_KLU     *lu = (Mat_KLU *)A->data;
114   PetscScalar *xa;
115   PetscInt     status;
116 
117   PetscFunctionBegin;
118   /* KLU uses a column major format, solve Ax = b by klu_*_solve */
119   PetscCall(VecCopy(b, x)); /* klu_solve stores the solution in rhs */
120   PetscCall(VecGetArray(x, &xa));
121   status = klu_K_solve(lu->Symbolic, lu->Numeric, A->rmap->n, 1, (PetscReal *)xa, &lu->Common);
122   PetscCheck(status == 1, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Solve failed");
123   PetscCall(VecRestoreArray(x, &xa));
124   PetscFunctionReturn(PETSC_SUCCESS);
125 }
126 
127 static PetscErrorCode MatSolve_KLU(Mat A, Vec b, Vec x)
128 {
129   Mat_KLU     *lu = (Mat_KLU *)A->data;
130   PetscScalar *xa;
131   PetscInt     status;
132 
133   PetscFunctionBegin;
134   /* KLU uses a column major format, solve A^Tx = b by klu_*_tsolve */
135   PetscCall(VecCopy(b, x)); /* klu_solve stores the solution in rhs */
136   PetscCall(VecGetArray(x, &xa));
137 #if defined(PETSC_USE_COMPLEX)
138   PetscInt conj_solve = 1;
139   status              = klu_K_tsolve(lu->Symbolic, lu->Numeric, A->rmap->n, 1, (PetscReal *)xa, conj_solve, &lu->Common); /* conjugate solve */
140 #else
141   status = klu_K_tsolve(lu->Symbolic, lu->Numeric, A->rmap->n, 1, xa, &lu->Common);
142 #endif
143   PetscCheck(status == 1, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Solve failed");
144   PetscCall(VecRestoreArray(x, &xa));
145   PetscFunctionReturn(PETSC_SUCCESS);
146 }
147 
148 static PetscErrorCode MatLUFactorNumeric_KLU(Mat F, Mat A, const MatFactorInfo *info)
149 {
150   Mat_KLU     *lu = (Mat_KLU *)(F)->data;
151   Mat_SeqAIJ  *a  = (Mat_SeqAIJ *)A->data;
152   PetscInt    *ai = a->i, *aj = a->j;
153   PetscScalar *av = a->a;
154 
155   PetscFunctionBegin;
156   /* numeric factorization of A' */
157 
158   if (lu->flg == SAME_NONZERO_PATTERN && lu->Numeric) klu_K_free_numeric(&lu->Numeric, &lu->Common);
159   lu->Numeric = klu_K_factor(ai, aj, (PetscReal *)av, lu->Symbolic, &lu->Common);
160   PetscCheck(lu->Numeric, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Numeric factorization failed");
161 
162   lu->flg                = SAME_NONZERO_PATTERN;
163   lu->CleanUpKLU         = PETSC_TRUE;
164   F->ops->solve          = MatSolve_KLU;
165   F->ops->solvetranspose = MatSolveTranspose_KLU;
166   PetscFunctionReturn(PETSC_SUCCESS);
167 }
168 
169 static PetscErrorCode MatLUFactorSymbolic_KLU(Mat F, Mat A, IS r, IS c, const MatFactorInfo *info)
170 {
171   Mat_SeqAIJ     *a  = (Mat_SeqAIJ *)A->data;
172   Mat_KLU        *lu = (Mat_KLU *)(F->data);
173   PetscInt        i, *ai = a->i, *aj = a->j, m = A->rmap->n, n = A->cmap->n;
174   const PetscInt *ra, *ca;
175 
176   PetscFunctionBegin;
177   if (lu->PetscMatOrdering) {
178     PetscCall(ISGetIndices(r, &ra));
179     PetscCall(ISGetIndices(c, &ca));
180     PetscCall(PetscMalloc2(m, &lu->perm_r, n, &lu->perm_c));
181     /* we cannot simply memcpy on 64-bit archs */
182     for (i = 0; i < m; i++) lu->perm_r[i] = ra[i];
183     for (i = 0; i < n; i++) lu->perm_c[i] = ca[i];
184     PetscCall(ISRestoreIndices(r, &ra));
185     PetscCall(ISRestoreIndices(c, &ca));
186   }
187 
188   /* symbolic factorization of A' */
189   if (r) {
190     lu->PetscMatOrdering = PETSC_TRUE;
191     lu->Symbolic         = klu_K_analyze_given(n, ai, aj, lu->perm_c, lu->perm_r, &lu->Common);
192   } else { /* use klu internal ordering */
193     lu->Symbolic = klu_K_analyze(n, ai, aj, &lu->Common);
194   }
195   PetscCheck(lu->Symbolic, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Symbolic Factorization failed");
196 
197   lu->flg                   = DIFFERENT_NONZERO_PATTERN;
198   lu->CleanUpKLU            = PETSC_TRUE;
199   (F)->ops->lufactornumeric = MatLUFactorNumeric_KLU;
200   PetscFunctionReturn(PETSC_SUCCESS);
201 }
202 
203 static PetscErrorCode MatView_Info_KLU(Mat A, PetscViewer viewer)
204 {
205   Mat_KLU       *lu      = (Mat_KLU *)A->data;
206   klu_K_numeric *Numeric = (klu_K_numeric *)lu->Numeric;
207 
208   PetscFunctionBegin;
209   PetscCall(PetscViewerASCIIPrintf(viewer, "KLU stats:\n"));
210   PetscCall(PetscViewerASCIIPrintf(viewer, "  Number of diagonal blocks: %" PetscInt_FMT "\n", (PetscInt)(Numeric->nblocks)));
211   PetscCall(PetscViewerASCIIPrintf(viewer, "  Total nonzeros=%" PetscInt_FMT "\n", (PetscInt)(Numeric->lnz + Numeric->unz)));
212   PetscCall(PetscViewerASCIIPrintf(viewer, "KLU runtime parameters:\n"));
213   /* Control parameters used by numeric factorization */
214   PetscCall(PetscViewerASCIIPrintf(viewer, "  Partial pivoting tolerance: %g\n", lu->Common.tol));
215   /* BTF preordering */
216   PetscCall(PetscViewerASCIIPrintf(viewer, "  BTF preordering enabled: %" PetscInt_FMT "\n", (PetscInt)(lu->Common.btf)));
217   /* mat ordering */
218   if (!lu->PetscMatOrdering) PetscCall(PetscViewerASCIIPrintf(viewer, "  Ordering: %s (not using the PETSc ordering)\n", KluOrderingTypes[(int)lu->Common.ordering]));
219   /* matrix row scaling */
220   PetscCall(PetscViewerASCIIPrintf(viewer, "  Matrix row scaling: %s\n", scale[(int)lu->Common.scale]));
221   PetscFunctionReturn(PETSC_SUCCESS);
222 }
223 
224 static PetscErrorCode MatView_KLU(Mat A, PetscViewer viewer)
225 {
226   PetscBool         iascii;
227   PetscViewerFormat format;
228 
229   PetscFunctionBegin;
230   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
231   if (iascii) {
232     PetscCall(PetscViewerGetFormat(viewer, &format));
233     if (format == PETSC_VIEWER_ASCII_INFO) PetscCall(MatView_Info_KLU(A, viewer));
234   }
235   PetscFunctionReturn(PETSC_SUCCESS);
236 }
237 
238 PetscErrorCode MatFactorGetSolverType_seqaij_klu(Mat A, MatSolverType *type)
239 {
240   PetscFunctionBegin;
241   *type = MATSOLVERKLU;
242   PetscFunctionReturn(PETSC_SUCCESS);
243 }
244 
245 /*MC
246   MATSOLVERKLU = "klu" - A matrix type providing direct solvers, LU, for sequential matrices
247   via the external package KLU.
248 
249   `./configure --download-suitesparse` to install PETSc to use KLU
250 
251   Use `-pc_type lu` `-pc_factor_mat_solver_type klu` to use this direct solver
252 
253   Consult KLU documentation for more information on the options database keys below.
254 
255   Options Database Keys:
256 + -mat_klu_pivot_tol <0.001>                  - Partial pivoting tolerance
257 . -mat_klu_use_btf <1>                        - Use BTF preordering
258 . -mat_klu_ordering <AMD>                     - KLU reordering scheme to reduce fill-in (choose one of) `AMD`, `COLAMD`, `PETSC`
259 - -mat_klu_row_scale <NONE>                   - Matrix row scaling (choose one of) `NONE`, `SUM`, `MAX`
260 
261    Level: beginner
262 
263    Note:
264    KLU is part of SuiteSparse http://faculty.cse.tamu.edu/davis/suitesparse.html
265 
266 .seealso: [](ch_matrices), `Mat`, `PCLU`, `MATSOLVERUMFPACK`, `MATSOLVERCHOLMOD`, `PCFactorSetMatSolverType()`, `MatSolverType`
267 M*/
268 
269 PETSC_INTERN PetscErrorCode MatGetFactor_seqaij_klu(Mat A, MatFactorType ftype, Mat *F)
270 {
271   Mat       B;
272   Mat_KLU  *lu;
273   PetscInt  m = A->rmap->n, n = A->cmap->n, idx = 0, status;
274   PetscBool flg;
275 
276   PetscFunctionBegin;
277   /* Create the factorization matrix F */
278   PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
279   PetscCall(MatSetSizes(B, PETSC_DECIDE, PETSC_DECIDE, m, n));
280   PetscCall(PetscStrallocpy("klu", &((PetscObject)B)->type_name));
281   PetscCall(MatSetUp(B));
282 
283   PetscCall(PetscNew(&lu));
284 
285   B->data                  = lu;
286   B->ops->getinfo          = MatGetInfo_External;
287   B->ops->lufactorsymbolic = MatLUFactorSymbolic_KLU;
288   B->ops->destroy          = MatDestroy_KLU;
289   B->ops->view             = MatView_KLU;
290 
291   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorGetSolverType_C", MatFactorGetSolverType_seqaij_klu));
292 
293   B->factortype   = MAT_FACTOR_LU;
294   B->assembled    = PETSC_TRUE; /* required by -ksp_view */
295   B->preallocated = PETSC_TRUE;
296 
297   PetscCall(PetscFree(B->solvertype));
298   PetscCall(PetscStrallocpy(MATSOLVERKLU, &B->solvertype));
299   B->canuseordering = PETSC_TRUE;
300   PetscCall(PetscStrallocpy(MATORDERINGEXTERNAL, (char **)&B->preferredordering[MAT_FACTOR_LU]));
301 
302   /* initializations */
303   /* get the default control parameters */
304   status = klu_K_defaults(&lu->Common);
305   PetscCheck(status > 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "KLU Initialization failed");
306 
307   lu->Common.scale = 0; /* No row scaling */
308 
309   PetscOptionsBegin(PetscObjectComm((PetscObject)B), ((PetscObject)B)->prefix, "KLU Options", "Mat");
310   /* Partial pivoting tolerance */
311   PetscCall(PetscOptionsReal("-mat_klu_pivot_tol", "Partial pivoting tolerance", "None", lu->Common.tol, &lu->Common.tol, NULL));
312   /* BTF pre-ordering */
313   PetscCall(PetscOptionsInt("-mat_klu_use_btf", "Enable BTF preordering", "None", (PetscInt)lu->Common.btf, (PetscInt *)&lu->Common.btf, NULL));
314   /* Matrix reordering */
315   PetscCall(PetscOptionsEList("-mat_klu_ordering", "Internal ordering method", "None", KluOrderingTypes, PETSC_STATIC_ARRAY_LENGTH(KluOrderingTypes), KluOrderingTypes[0], &idx, &flg));
316   lu->Common.ordering = (int)idx;
317   /* Matrix row scaling */
318   PetscCall(PetscOptionsEList("-mat_klu_row_scale", "Matrix row scaling", "None", scale, 3, scale[0], &idx, &flg));
319   PetscOptionsEnd();
320   *F = B;
321   PetscFunctionReturn(PETSC_SUCCESS);
322 }
323