1 2 #include <petsc.h> 3 #include <petscksp.h> 4 #include "private/kspimpl.h" 5 #include "petscpc.h" 6 #include "../src/mat/impls/aij/seq/aij.h" 7 #include "../src/mat/impls/sbaij/seq/sbaij.h" 8 #include "../src/mat/impls/aij/seq/bas/spbas.h" 9 10 #undef __FUNCT__ 11 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ_Bas" 12 PetscErrorCode MatICCFactorSymbolic_SeqAIJ_Bas(Mat fact,Mat A,IS perm,const MatFactorInfo *info) 13 { 14 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 15 Mat_SeqSBAIJ *b; 16 PetscErrorCode ierr; 17 PetscTruth perm_identity,missing; 18 PetscInt reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui; 19 const PetscInt *rip,*riip; 20 PetscInt j; 21 PetscInt d; 22 PetscInt ncols,*cols,*uj; 23 PetscReal fill=info->fill,levels=info->levels; 24 IS iperm; 25 spbas_matrix Pattern_0, Pattern_P; 26 27 PetscFunctionBegin; 28 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); 29 ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr); 30 if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d); 31 ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr); 32 ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr); 33 34 35 /* ICC(0) without matrix ordering: simply copies fill pattern */ 36 if (!levels && perm_identity) { 37 ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr); 38 ui[0] = 0; 39 40 for (i=0; i<am; i++) { 41 ui[i+1] = ui[i] + ai[i+1] - a->diag[i]; 42 } 43 ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr); 44 cols = uj; 45 for (i=0; i<am; i++) { 46 aj = a->j + a->diag[i]; 47 ncols = ui[i+1] - ui[i]; 48 for (j=0; j<ncols; j++) *cols++ = *aj++; 49 } 50 } else { /* case: levels>0 || (levels=0 && !perm_identity) */ 51 ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr); 52 ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr); 53 54 // Create spbas_matrix for pattern 55 ierr = spbas_pattern_only(am, am, ai, aj, &Pattern_0); CHKERRQ(ierr); 56 57 // Apply the permutation 58 ierr = spbas_apply_reordering( &Pattern_0, rip, riip); CHKERRQ(ierr); 59 60 // Raise the power 61 ierr = spbas_power( Pattern_0, (int) levels+1, &Pattern_P); 62 CHKERRQ(ierr); 63 ierr = spbas_delete( Pattern_0 ); CHKERRQ(ierr); 64 65 // Keep only upper triangle of pattern 66 ierr = spbas_keep_upper( &Pattern_P ); 67 68 // Convert to Sparse Row Storage 69 ierr = spbas_matrix_to_crs(Pattern_P, NULL, &ui, &uj); CHKERRQ(ierr); 70 ierr = spbas_delete(Pattern_P);CHKERRQ(ierr); 71 } /* end of case: levels>0 || (levels=0 && !perm_identity) */ 72 73 /* put together the new matrix in MATSEQSBAIJ format */ 74 75 b = (Mat_SeqSBAIJ*)(fact)->data; 76 b->singlemalloc = PETSC_FALSE; 77 ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr); 78 b->j = uj; 79 b->i = ui; 80 b->diag = 0; 81 b->ilen = 0; 82 b->imax = 0; 83 b->row = perm; 84 b->col = perm; 85 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 86 ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr); 87 b->icol = iperm; 88 b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */ 89 ierr = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr); 90 ierr = PetscLogObjectMemory((fact),(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr); 91 b->maxnz = b->nz = ui[am]; 92 b->free_a = PETSC_TRUE; 93 b->free_ij = PETSC_TRUE; 94 95 (fact)->info.factor_mallocs = reallocs; 96 (fact)->info.fill_ratio_given = fill; 97 if (ai[am] != 0) { 98 (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]); 99 } else { 100 (fact)->info.fill_ratio_needed = 0.0; 101 } 102 /* (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_inplace; */ 103 PetscFunctionReturn(0); 104 } 105 106 107 #undef __FUNCT__ 108 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ_Bas" 109 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_Bas(Mat B,Mat A,const MatFactorInfo *info) 110 { 111 Mat C = B; 112 Mat_SeqSBAIJ *b=(Mat_SeqSBAIJ*)C->data; 113 IS ip=b->row,iip = b->icol; 114 PetscErrorCode ierr; 115 const PetscInt *rip,*riip; 116 PetscInt mbs=A->rmap->n,*bi=b->i,*bj=b->j; 117 118 MatScalar *ba=b->a; 119 PetscReal shiftnz = info->shiftnz; 120 PetscReal droptol = -1; 121 PetscTruth perm_identity; 122 spbas_matrix Pattern, matrix_L,matrix_LT; 123 PetscScalar mem_reduction; 124 125 PetscFunctionBegin; 126 // Reduce memory requirements: 127 // erase values of B-matrix 128 ierr = PetscFree(ba); CHKERRQ(ierr); 129 // Compress (maximum) sparseness pattern of B-matrix 130 ierr = spbas_compress_pattern(bi, bj, mbs, mbs, SPBAS_DIAGONAL_OFFSETS, 131 &Pattern, &mem_reduction);CHKERRQ(ierr); 132 ierr = PetscFree(bi); CHKERRQ(ierr); 133 ierr = PetscFree(bj); CHKERRQ(ierr); 134 135 printf("Results from spbas_compress_pattern:\n"); 136 printf(" compression rate %6.2f %%\n",mem_reduction); 137 ierr=7; 138 139 // Make Cholesky decompositions with larger Manteuffel shifts until no more 140 // negative diagonals are found. 141 ierr = ISGetIndices(ip,&rip);CHKERRQ(ierr); 142 ierr = ISGetIndices(iip,&riip);CHKERRQ(ierr); 143 144 if (info->usedt) { 145 droptol = info->dt; 146 } 147 for (ierr = NEGATIVE_DIAGONAL; ierr == NEGATIVE_DIAGONAL; ) 148 { 149 ierr = spbas_incomplete_cholesky( A, rip, riip, Pattern, droptol, shiftnz, 150 &matrix_LT); 151 if (ierr == NEGATIVE_DIAGONAL) 152 { 153 shiftnz *= 1.5; 154 printf("spbas_incomplete_cholesky found a negative diagonal.\n"); 155 printf(" Trying again with Manteuffel shift=%e\n",shiftnz); 156 } 157 } 158 CHKERRQ(ierr); 159 ierr = spbas_delete(Pattern); CHKERRQ(ierr); 160 161 printf("Results from spbas_incomplete_cholesky:\n"); 162 printf(" memory_usage: %6.2f bytes per row\n", 163 (PetscScalar) spbas_memory_requirement( matrix_LT)/ (PetscScalar) mbs); 164 165 ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr); 166 ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr); 167 168 // Convert spbas_matrix to compressed row storage 169 ierr = spbas_transpose(matrix_LT, &matrix_L); CHKERRQ(ierr); 170 ierr = spbas_delete(matrix_LT); CHKERRQ(ierr); 171 #if defined(foo) 172 { ierr = spbas_dump("factorL",matrix_L); CHKERRQ(ierr);} 173 #endif 174 ierr = spbas_matrix_to_crs(matrix_L, &ba, &bi, &bj); CHKERRQ(ierr); 175 b->i=bi; b->j=bj; b->a=ba; 176 ierr = spbas_delete(matrix_L); CHKERRQ(ierr); 177 178 // Set the appropriate solution functions 179 ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr); 180 if (perm_identity){ 181 (B)->ops->solve = MatSolve_SeqSBAIJ_1_NaturalOrdering_inplace; 182 (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1_NaturalOrdering_inplace; 183 (B)->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1_NaturalOrdering_inplace; 184 (B)->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1_NaturalOrdering_inplace; 185 } else { 186 (B)->ops->solve = MatSolve_SeqSBAIJ_1_inplace; 187 (B)->ops->solvetranspose = MatSolve_SeqSBAIJ_1_inplace; 188 (B)->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_1_inplace; 189 (B)->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_1_inplace; 190 } 191 192 C->assembled = PETSC_TRUE; 193 C->preallocated = PETSC_TRUE; 194 ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr); 195 196 // Optionally, print the factor matrix to file 197 #if defined(foo) 198 { 199 FILE * factfile = fopen("factorL","w"); 200 if (!factfile) CHKERRQ((PetscErrorCode) 10); 201 for (i=0; i<mbs; i++) 202 { 203 for (j=bi[i]; j<bi[i+1]; j++) 204 { 205 fprintf(factfile,"%d %d %e\n",i,bj[j],ba[j]); 206 } 207 } 208 fclose(factfile); 209 } 210 #endif 211 PetscFunctionReturn(0); 212 } 213 214 #undef __FUNCT__ 215 #define __FUNCT__ "MatGetFactor_seqaij_bas" 216 PetscErrorCode MatGetFactor_seqaij_bas(Mat A,MatFactorType ftype,Mat *B) 217 { 218 PetscInt n = A->rmap->n; 219 PetscErrorCode ierr; 220 221 PetscFunctionBegin; 222 ierr = MatCreate(((PetscObject)A)->comm,B);CHKERRQ(ierr); 223 ierr = MatSetSizes(*B,n,n,n,n);CHKERRQ(ierr); 224 if (ftype == MAT_FACTOR_ICC) { 225 ierr = MatSetType(*B,MATSEQSBAIJ);CHKERRQ(ierr); 226 ierr = MatSeqSBAIJSetPreallocation(*B,1,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr); 227 (*B)->ops->iccfactorsymbolic = MatICCFactorSymbolic_SeqAIJ_Bas; 228 (*B)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ_Bas; 229 } else SETERRQ(PETSC_ERR_SUP,"Factor type not supported"); 230 (*B)->factor = ftype; 231 PetscFunctionReturn(0); 232 } 233 234 235 236