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 klu_l_analyze 16 #define klu_K_analyze_given klu_l_analyze_given 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 klu_zl_factor 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 klu_l_factor 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 80 #define SuiteSparse_long long long 81 #define SuiteSparse_long_max LONG_LONG_MAX 82 #define SuiteSparse_long_id "%lld" 83 84 EXTERN_C_BEGIN 85 #include <klu.h> 86 EXTERN_C_END 87 88 static const char *KluOrderingTypes[] = {"AMD","COLAMD","PETSC"}; 89 static const char *scale[] ={"NONE","SUM","MAX"}; 90 91 typedef struct { 92 klu_K_common Common; 93 klu_K_symbolic *Symbolic; 94 klu_K_numeric *Numeric; 95 PetscInt *perm_c,*perm_r; 96 MatStructure flg; 97 PetscBool PetscMatOrdering; 98 99 /* Flag to clean up KLU objects during Destroy */ 100 PetscBool CleanUpKLU; 101 } Mat_KLU; 102 103 #undef __FUNCT__ 104 #define __FUNCT__ "MatDestroy_KLU" 105 static PetscErrorCode MatDestroy_KLU(Mat A) 106 { 107 PetscErrorCode ierr; 108 Mat_KLU *lu=(Mat_KLU*)A->spptr; 109 110 PetscFunctionBegin; 111 if (lu && lu->CleanUpKLU) { 112 klu_K_free_symbolic(&lu->Symbolic,&lu->Common); 113 klu_K_free_numeric(&lu->Numeric,&lu->Common); 114 ierr = PetscFree2(lu->perm_r,lu->perm_c);CHKERRQ(ierr); 115 } 116 ierr = PetscFree(A->spptr);CHKERRQ(ierr); 117 ierr = MatDestroy_SeqAIJ(A);CHKERRQ(ierr); 118 PetscFunctionReturn(0); 119 } 120 121 #undef __FUNCT__ 122 #define __FUNCT__ "MatSolveTranspose_KLU" 123 static PetscErrorCode MatSolveTranspose_KLU(Mat A,Vec b,Vec x) 124 { 125 Mat_KLU *lu = (Mat_KLU*)A->spptr; 126 PetscScalar *xa; 127 PetscErrorCode ierr; 128 PetscInt status; 129 130 PetscFunctionBegin; 131 /* KLU uses a column major format, solve Ax = b by klu_*_solve */ 132 /* ----------------------------------*/ 133 ierr = VecCopy(b,x); /* klu_solve stores the solution in rhs */ 134 ierr = VecGetArray(x,&xa); 135 status = klu_K_solve(lu->Symbolic,lu->Numeric,A->rmap->n,1,(PetscReal*)xa,&lu->Common); 136 if (status != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"KLU Solve failed"); 137 ierr = VecRestoreArray(x,&xa);CHKERRQ(ierr); 138 PetscFunctionReturn(0); 139 } 140 141 #undef __FUNCT__ 142 #define __FUNCT__ "MatSolve_KLU" 143 static PetscErrorCode MatSolve_KLU(Mat A,Vec b,Vec x) 144 { 145 Mat_KLU *lu = (Mat_KLU*)A->spptr; 146 PetscScalar *xa; 147 PetscErrorCode ierr; 148 PetscInt status; 149 150 PetscFunctionBegin; 151 /* KLU uses a column major format, solve A^Tx = b by klu_*_tsolve */ 152 /* ----------------------------------*/ 153 ierr = VecCopy(b,x); /* klu_solve stores the solution in rhs */ 154 ierr = VecGetArray(x,&xa); 155 #if defined(PETSC_USE_COMPLEX) 156 PetscInt conj_solve=1; 157 status = klu_K_tsolve(lu->Symbolic,lu->Numeric,A->rmap->n,1,(PetscReal*)xa,conj_solve,&lu->Common); /* conjugate solve */ 158 #else 159 status = klu_K_tsolve(lu->Symbolic,lu->Numeric,A->rmap->n,1,xa,&lu->Common); 160 #endif 161 if (status != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"KLU Solve failed"); 162 ierr = VecRestoreArray(x,&xa);CHKERRQ(ierr); 163 PetscFunctionReturn(0); 164 } 165 166 #undef __FUNCT__ 167 #define __FUNCT__ "MatLUFactorNumeric_KLU" 168 static PetscErrorCode MatLUFactorNumeric_KLU(Mat F,Mat A,const MatFactorInfo *info) 169 { 170 Mat_KLU *lu = (Mat_KLU*)(F)->spptr; 171 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 172 PetscInt *ai = a->i,*aj=a->j; 173 PetscScalar *av = a->a; 174 175 PetscFunctionBegin; 176 /* numeric factorization of A' */ 177 /* ----------------------------*/ 178 179 if (lu->flg == SAME_NONZERO_PATTERN && lu->Numeric) { 180 klu_K_free_numeric(&lu->Numeric,&lu->Common); 181 } 182 lu->Numeric = klu_K_factor(ai,aj,(PetscReal*)av,lu->Symbolic,&lu->Common); 183 if(!lu->Numeric) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"KLU Numeric factorization failed"); 184 185 lu->flg = SAME_NONZERO_PATTERN; 186 lu->CleanUpKLU = PETSC_TRUE; 187 F->ops->solve = MatSolve_KLU; 188 F->ops->solvetranspose = MatSolveTranspose_KLU; 189 PetscFunctionReturn(0); 190 } 191 192 #undef __FUNCT__ 193 #define __FUNCT__ "MatLUFactorSymbolic_KLU" 194 static PetscErrorCode MatLUFactorSymbolic_KLU(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info) 195 { 196 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 197 Mat_KLU *lu = (Mat_KLU*)(F->spptr); 198 PetscErrorCode ierr; 199 PetscInt i,*ai = a->i,*aj = a->j,m=A->rmap->n,n=A->cmap->n; 200 const PetscInt *ra,*ca; 201 202 PetscFunctionBegin; 203 if (lu->PetscMatOrdering) { 204 ierr = ISGetIndices(r,&ra);CHKERRQ(ierr); 205 ierr = ISGetIndices(c,&ca);CHKERRQ(ierr); 206 ierr = PetscMalloc2(m,&lu->perm_r,n,&lu->perm_c);CHKERRQ(ierr); 207 /* we cannot simply memcpy on 64 bit archs */ 208 for (i = 0; i < m; i++) lu->perm_r[i] = ra[i]; 209 for (i = 0; i < n; i++) lu->perm_c[i] = ca[i]; 210 ierr = ISRestoreIndices(r,&ra);CHKERRQ(ierr); 211 ierr = ISRestoreIndices(c,&ca);CHKERRQ(ierr); 212 } 213 214 /* symbolic factorization of A' */ 215 /* ---------------------------------------------------------------------- */ 216 if (lu->PetscMatOrdering) { /* use Petsc ordering */ 217 lu->Symbolic = klu_K_analyze_given(n,ai,aj,lu->perm_c,lu->perm_r,&lu->Common); 218 } else { /* use klu internal ordering */ 219 lu->Symbolic = klu_K_analyze(n,ai,aj,&lu->Common); 220 } 221 if (!lu->Symbolic) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"KLU Symbolic Factorization failed"); 222 223 lu->flg = DIFFERENT_NONZERO_PATTERN; 224 lu->CleanUpKLU = PETSC_TRUE; 225 (F)->ops->lufactornumeric = MatLUFactorNumeric_KLU; 226 PetscFunctionReturn(0); 227 } 228 229 #undef __FUNCT__ 230 #define __FUNCT__ "MatFactorInfo_KLU" 231 static PetscErrorCode MatFactorInfo_KLU(Mat A,PetscViewer viewer) 232 { 233 Mat_KLU *lu= (Mat_KLU*)A->spptr; 234 klu_K_numeric *Numeric=(klu_K_numeric*)lu->Numeric; 235 PetscErrorCode ierr; 236 237 PetscFunctionBegin; 238 /* check if matrix is KLU type */ 239 if (A->ops->solve != MatSolve_KLU) PetscFunctionReturn(0); 240 241 ierr = PetscViewerASCIIPrintf(viewer,"KLU stats:\n");CHKERRQ(ierr); 242 ierr = PetscViewerASCIIPrintf(viewer," Number of diagonal blocks: %d\n",Numeric->nblocks); 243 ierr = PetscViewerASCIIPrintf(viewer," Total nonzeros=%d\n",Numeric->lnz+Numeric->unz);CHKERRQ(ierr); 244 245 ierr = PetscViewerASCIIPrintf(viewer,"KLU runtime parameters:\n");CHKERRQ(ierr); 246 247 /* Control parameters used by numeric factorization */ 248 ierr = PetscViewerASCIIPrintf(viewer," Partial pivoting tolerance: %g\n",lu->Common.tol);CHKERRQ(ierr); 249 /* BTF preordering */ 250 ierr = PetscViewerASCIIPrintf(viewer," BTF preordering enabled: %d\n",lu->Common.btf);CHKERRQ(ierr); 251 /* mat ordering */ 252 if (!lu->PetscMatOrdering) { 253 ierr = PetscViewerASCIIPrintf(viewer," Ordering: %s (not using the PETSc ordering)\n",KluOrderingTypes[(int)lu->Common.ordering]);CHKERRQ(ierr); 254 } else { 255 ierr = PetscViewerASCIIPrintf(viewer," Using PETSc ordering\n");CHKERRQ(ierr); 256 } 257 /* matrix row scaling */ 258 ierr = PetscViewerASCIIPrintf(viewer, " Matrix row scaling: %s\n",scale[(int)lu->Common.scale]);CHKERRQ(ierr); 259 PetscFunctionReturn(0); 260 } 261 262 #undef __FUNCT__ 263 #define __FUNCT__ "MatView_KLU" 264 static PetscErrorCode MatView_KLU(Mat A,PetscViewer viewer) 265 { 266 PetscErrorCode ierr; 267 PetscBool iascii; 268 PetscViewerFormat format; 269 270 PetscFunctionBegin; 271 ierr = MatView_SeqAIJ(A,viewer);CHKERRQ(ierr); 272 273 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 274 if (iascii) { 275 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 276 if (format == PETSC_VIEWER_ASCII_INFO) { 277 ierr = MatFactorInfo_KLU(A,viewer);CHKERRQ(ierr); 278 } 279 } 280 PetscFunctionReturn(0); 281 } 282 283 #undef __FUNCT__ 284 #define __FUNCT__ "MatFactorGetSolverPackage_seqaij_klu" 285 PetscErrorCode MatFactorGetSolverPackage_seqaij_klu(Mat A,const MatSolverPackage *type) 286 { 287 PetscFunctionBegin; 288 *type = MATSOLVERKLU; 289 PetscFunctionReturn(0); 290 } 291 292 293 /*MC 294 MATSOLVERKLU = "klu" - A matrix type providing direct solvers (LU) for sequential matrices 295 via the external package KLU. 296 297 ./configure --download-suitesparse to install PETSc to use KLU 298 299 Use -pc_type lu -pc_factor_mat_solver_package klu to us this direct solver 300 301 Consult KLU documentation for more information on the options database keys below. 302 303 Options Database Keys: 304 + -mat_klu_pivot_tol <0.001> - Partial pivoting tolerance 305 . -mat_klu_use_btf <1> - Use BTF preordering 306 . -mat_klu_ordering <AMD> - KLU reordering scheme to reduce fill-in (choose one of) AMD COLAMD PETSC 307 - -mat_klu_row_scale <NONE> - Matrix row scaling (choose one of) NONE SUM MAX 308 309 Note: KLU is part of SuiteSparse http://faculty.cse.tamu.edu/davis/suitesparse.html 310 311 Level: beginner 312 313 .seealso: PCLU, MATSOLVERUMFPACK, MATSOLVERCHOLMOD, PCFactorSetMatSolverPackage(), MatSolverPackage 314 M*/ 315 316 #undef __FUNCT__ 317 #define __FUNCT__ "MatGetFactor_seqaij_klu" 318 PETSC_INTERN PetscErrorCode MatGetFactor_seqaij_klu(Mat A,MatFactorType ftype,Mat *F) 319 { 320 Mat B; 321 Mat_KLU *lu; 322 PetscErrorCode ierr; 323 PetscInt m=A->rmap->n,n=A->cmap->n,idx,status; 324 PetscBool flg; 325 326 PetscFunctionBegin; 327 /* Create the factorization matrix F */ 328 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 329 ierr = MatSetSizes(B,PETSC_DECIDE,PETSC_DECIDE,m,n);CHKERRQ(ierr); 330 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 331 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 332 ierr = PetscNewLog(B,&lu);CHKERRQ(ierr); 333 334 B->spptr = lu; 335 B->ops->lufactorsymbolic = MatLUFactorSymbolic_KLU; 336 B->ops->destroy = MatDestroy_KLU; 337 B->ops->view = MatView_KLU; 338 339 ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_seqaij_klu);CHKERRQ(ierr); 340 341 B->factortype = MAT_FACTOR_LU; 342 B->assembled = PETSC_TRUE; /* required by -ksp_view */ 343 B->preallocated = PETSC_TRUE; 344 345 /* initializations */ 346 /* ------------------------------------------------*/ 347 /* get the default control parameters */ 348 status = klu_K_defaults(&lu->Common); 349 if(status <= 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"KLU Initialization failed"); 350 351 lu->Common.scale = 0; /* No row scaling */ 352 353 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"KLU Options","Mat");CHKERRQ(ierr); 354 /* Partial pivoting tolerance */ 355 ierr = PetscOptionsReal("-mat_klu_pivot_tol","Partial pivoting tolerance","None",lu->Common.tol,&lu->Common.tol,NULL);CHKERRQ(ierr); 356 /* BTF pre-ordering */ 357 ierr = PetscOptionsInt("-mat_klu_use_btf","Enable BTF preordering","None",lu->Common.btf,&lu->Common.btf,NULL);CHKERRQ(ierr); 358 /* Matrix reordering */ 359 ierr = PetscOptionsEList("-mat_klu_ordering","Internal ordering method","None",KluOrderingTypes,sizeof(KluOrderingTypes)/sizeof(KluOrderingTypes[0]),KluOrderingTypes[0],&idx,&flg);CHKERRQ(ierr); 360 if (flg) { 361 if ((int)idx == 2) lu->PetscMatOrdering = PETSC_TRUE; /* use Petsc mat ordering (note: size is for the transpose, and PETSc r = Klu perm_c) */ 362 else lu->Common.ordering = (int)idx; 363 } 364 /* Matrix row scaling */ 365 ierr = PetscOptionsEList("-mat_klu_row_scale","Matrix row scaling","None",scale,3,scale[0],&idx,&flg);CHKERRQ(ierr); 366 PetscOptionsEnd(); 367 *F = B; 368 PetscFunctionReturn(0); 369 } 370