static void ssolve_creategraph(Graph ** graph, ETree ** frontETree, InpMtx * mtxA, int size, FILE * msgFile) { IVL *adjIVL; int nedges; *graph = Graph_new(); adjIVL = InpMtx_fullAdjacency(mtxA); nedges = IVL_tsize(adjIVL); Graph_init2(*graph, 0, size, 0, nedges, size, nedges, adjIVL, NULL, NULL); if (DEBUG_LVL > 1) { fprintf(msgFile, "\n\n graph of the input matrix"); Graph_writeForHumanEye(*graph, msgFile); fflush(msgFile); } /* (2) order the graph using multiple minimum degree */ /*maxdomainsize=neqns/100; */ /*if (maxdomainsize==0) maxdomainsize=1; */ /* *frontETree = orderViaMMD(*graph, RNDSEED, DEBUG_LVL, msgFile) ; */ /* *frontETree = orderViaND(*graph,maxdomainsize,RNDSEED,DEBUG_LVL,msgFile); */ /* *frontETree = orderViaMS(*graph,maxdomainsize,RNDSEED,DEBUG_LVL,msgFile); */ *frontETree = orderViaBestOfNDandMS(*graph, TUNE_MAXDOMAINSIZE, TUNE_MAXZEROS, TUNE_MAXSIZE, RNDSEED, DEBUG_LVL, msgFile); if (DEBUG_LVL > 1) { fprintf(msgFile, "\n\n front tree from ordering"); ETree_writeForHumanEye(*frontETree, msgFile); fflush(msgFile); } }
PetscErrorCode MatFactorNumeric_SeqSpooles(Mat F,Mat A,const MatFactorInfo *info) { Mat_Spooles *lu = (Mat_Spooles*)(F)->spptr; ChvManager *chvmanager ; Chv *rootchv ; IVL *adjIVL; PetscErrorCode ierr; PetscInt nz,nrow=A->rmap->n,irow,nedges,neqns=A->cmap->n,*ai,*aj,i,*diag=0,fierr; PetscScalar *av; double cputotal,facops; #if defined(PETSC_USE_COMPLEX) PetscInt nz_row,*aj_tmp; PetscScalar *av_tmp; #else PetscInt *ivec1,*ivec2,j; double *dvec; #endif PetscBool isSeqAIJ,isMPIAIJ; PetscFunctionBegin; if (lu->flg == DIFFERENT_NONZERO_PATTERN) { /* first numeric factorization */ (F)->ops->solve = MatSolve_SeqSpooles; (F)->assembled = PETSC_TRUE; /* set Spooles options */ ierr = SetSpoolesOptions(A, &lu->options);CHKERRQ(ierr); lu->mtxA = InpMtx_new(); } /* copy A to Spooles' InpMtx object */ ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isSeqAIJ);CHKERRQ(ierr); ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isMPIAIJ);CHKERRQ(ierr); if (isSeqAIJ){ Mat_SeqAIJ *mat = (Mat_SeqAIJ*)A->data; ai=mat->i; aj=mat->j; av=mat->a; if (lu->options.symflag == SPOOLES_NONSYMMETRIC) { nz=mat->nz; } else { /* SPOOLES_SYMMETRIC || SPOOLES_HERMITIAN */ nz=(mat->nz + A->rmap->n)/2; diag=mat->diag; } } else { /* A is SBAIJ */ Mat_SeqSBAIJ *mat = (Mat_SeqSBAIJ*)A->data; ai=mat->i; aj=mat->j; av=mat->a; nz=mat->nz; } InpMtx_init(lu->mtxA, INPMTX_BY_ROWS, lu->options.typeflag, nz, 0); #if defined(PETSC_USE_COMPLEX) for (irow=0; irow<nrow; irow++) { if ( lu->options.symflag == SPOOLES_NONSYMMETRIC || !(isSeqAIJ || isMPIAIJ)){ nz_row = ai[irow+1] - ai[irow]; aj_tmp = aj + ai[irow]; av_tmp = av + ai[irow]; } else { nz_row = ai[irow+1] - diag[irow]; aj_tmp = aj + diag[irow]; av_tmp = av + diag[irow]; } for (i=0; i<nz_row; i++){ InpMtx_inputComplexEntry(lu->mtxA, irow, *aj_tmp++,PetscRealPart(*av_tmp),PetscImaginaryPart(*av_tmp)); av_tmp++; } } #else ivec1 = InpMtx_ivec1(lu->mtxA); ivec2 = InpMtx_ivec2(lu->mtxA); dvec = InpMtx_dvec(lu->mtxA); if ( lu->options.symflag == SPOOLES_NONSYMMETRIC || !isSeqAIJ){ for (irow = 0; irow < nrow; irow++){ for (i = ai[irow]; i<ai[irow+1]; i++) ivec1[i] = irow; } IVcopy(nz, ivec2, aj); DVcopy(nz, dvec, av); } else { nz = 0; for (irow = 0; irow < nrow; irow++){ for (j = diag[irow]; j<ai[irow+1]; j++) { ivec1[nz] = irow; ivec2[nz] = aj[j]; dvec[nz] = av[j]; nz++; } } } InpMtx_inputRealTriples(lu->mtxA, nz, ivec1, ivec2, dvec); #endif InpMtx_changeStorageMode(lu->mtxA, INPMTX_BY_VECTORS); if ( lu->options.msglvl > 0 ) { int err; printf("\n\n input matrix"); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n input matrix");CHKERRQ(ierr); InpMtx_writeForHumanEye(lu->mtxA, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } if ( lu->flg == DIFFERENT_NONZERO_PATTERN){ /* first numeric factorization */ /*--------------------------------------------------- find a low-fill ordering (1) create the Graph object (2) order the graph -------------------------------------------------------*/ if (lu->options.useQR){ adjIVL = InpMtx_adjForATA(lu->mtxA); } else { adjIVL = InpMtx_fullAdjacency(lu->mtxA); } nedges = IVL_tsize(adjIVL); lu->graph = Graph_new(); Graph_init2(lu->graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL); if ( lu->options.msglvl > 2 ) { int err; if (lu->options.useQR){ ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n graph of A^T A");CHKERRQ(ierr); } else { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n graph of the input matrix");CHKERRQ(ierr); } Graph_writeForHumanEye(lu->graph, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } switch (lu->options.ordering) { case 0: lu->frontETree = orderViaBestOfNDandMS(lu->graph, lu->options.maxdomainsize, lu->options.maxzeros, lu->options.maxsize, lu->options.seed, lu->options.msglvl, lu->options.msgFile); break; case 1: lu->frontETree = orderViaMMD(lu->graph,lu->options.seed,lu->options.msglvl,lu->options.msgFile); break; case 2: lu->frontETree = orderViaMS(lu->graph, lu->options.maxdomainsize, lu->options.seed,lu->options.msglvl,lu->options.msgFile); break; case 3: lu->frontETree = orderViaND(lu->graph, lu->options.maxdomainsize, lu->options.seed,lu->options.msglvl,lu->options.msgFile); break; default: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown Spooles's ordering"); } if ( lu->options.msglvl > 0 ) { int err; ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n front tree from ordering");CHKERRQ(ierr); ETree_writeForHumanEye(lu->frontETree, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } /* get the permutation, permute the front tree */ lu->oldToNewIV = ETree_oldToNewVtxPerm(lu->frontETree); lu->oldToNew = IV_entries(lu->oldToNewIV); lu->newToOldIV = ETree_newToOldVtxPerm(lu->frontETree); if (!lu->options.useQR) ETree_permuteVertices(lu->frontETree, lu->oldToNewIV); /* permute the matrix */ if (lu->options.useQR){ InpMtx_permute(lu->mtxA, NULL, lu->oldToNew); } else { InpMtx_permute(lu->mtxA, lu->oldToNew, lu->oldToNew); if ( lu->options.symflag == SPOOLES_SYMMETRIC) { InpMtx_mapToUpperTriangle(lu->mtxA); } #if defined(PETSC_USE_COMPLEX) if ( lu->options.symflag == SPOOLES_HERMITIAN ) { InpMtx_mapToUpperTriangleH(lu->mtxA); } #endif InpMtx_changeCoordType(lu->mtxA, INPMTX_BY_CHEVRONS); } InpMtx_changeStorageMode(lu->mtxA, INPMTX_BY_VECTORS); /* get symbolic factorization */ if (lu->options.useQR){ lu->symbfacIVL = SymbFac_initFromGraph(lu->frontETree, lu->graph); IVL_overwrite(lu->symbfacIVL, lu->oldToNewIV); IVL_sortUp(lu->symbfacIVL); ETree_permuteVertices(lu->frontETree, lu->oldToNewIV); } else { lu->symbfacIVL = SymbFac_initFromInpMtx(lu->frontETree, lu->mtxA); } if ( lu->options.msglvl > 2 ) { int err; ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n old-to-new permutation vector");CHKERRQ(ierr); IV_writeForHumanEye(lu->oldToNewIV, lu->options.msgFile); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n new-to-old permutation vector");CHKERRQ(ierr); IV_writeForHumanEye(lu->newToOldIV, lu->options.msgFile); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n front tree after permutation");CHKERRQ(ierr); ETree_writeForHumanEye(lu->frontETree, lu->options.msgFile); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n input matrix after permutation");CHKERRQ(ierr); InpMtx_writeForHumanEye(lu->mtxA, lu->options.msgFile); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n symbolic factorization");CHKERRQ(ierr); IVL_writeForHumanEye(lu->symbfacIVL, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } lu->frontmtx = FrontMtx_new(); lu->mtxmanager = SubMtxManager_new(); SubMtxManager_init(lu->mtxmanager, NO_LOCK, 0); } else { /* new num factorization using previously computed symbolic factor */ if (lu->options.pivotingflag) { /* different FrontMtx is required */ FrontMtx_free(lu->frontmtx); lu->frontmtx = FrontMtx_new(); } else { FrontMtx_clearData (lu->frontmtx); } SubMtxManager_free(lu->mtxmanager); lu->mtxmanager = SubMtxManager_new(); SubMtxManager_init(lu->mtxmanager, NO_LOCK, 0); /* permute mtxA */ if (lu->options.useQR){ InpMtx_permute(lu->mtxA, NULL, lu->oldToNew); } else { InpMtx_permute(lu->mtxA, lu->oldToNew, lu->oldToNew); if ( lu->options.symflag == SPOOLES_SYMMETRIC ) { InpMtx_mapToUpperTriangle(lu->mtxA); } InpMtx_changeCoordType(lu->mtxA, INPMTX_BY_CHEVRONS); } InpMtx_changeStorageMode(lu->mtxA, INPMTX_BY_VECTORS); if ( lu->options.msglvl > 2 ) { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n input matrix after permutation");CHKERRQ(ierr); InpMtx_writeForHumanEye(lu->mtxA, lu->options.msgFile); } } /* end of if( lu->flg == DIFFERENT_NONZERO_PATTERN) */ if (lu->options.useQR){ FrontMtx_init(lu->frontmtx, lu->frontETree, lu->symbfacIVL, lu->options.typeflag, SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL, lu->mtxmanager, lu->options.msglvl, lu->options.msgFile); } else { FrontMtx_init(lu->frontmtx, lu->frontETree, lu->symbfacIVL, lu->options.typeflag, lu->options.symflag, FRONTMTX_DENSE_FRONTS, lu->options.pivotingflag, NO_LOCK, 0, NULL, lu->mtxmanager, lu->options.msglvl, lu->options.msgFile); } if ( lu->options.symflag == SPOOLES_SYMMETRIC ) { /* || SPOOLES_HERMITIAN ? */ if ( lu->options.patchAndGoFlag == 1 ) { lu->frontmtx->patchinfo = PatchAndGoInfo_new(); PatchAndGoInfo_init(lu->frontmtx->patchinfo, 1, lu->options.toosmall, lu->options.fudge, lu->options.storeids, lu->options.storevalues); } else if ( lu->options.patchAndGoFlag == 2 ) { lu->frontmtx->patchinfo = PatchAndGoInfo_new(); PatchAndGoInfo_init(lu->frontmtx->patchinfo, 2, lu->options.toosmall, lu->options.fudge, lu->options.storeids, lu->options.storevalues); } } /* numerical factorization */ chvmanager = ChvManager_new(); ChvManager_init(chvmanager, NO_LOCK, 1); DVfill(10, lu->cpus, 0.0); if (lu->options.useQR){ facops = 0.0 ; FrontMtx_QR_factor(lu->frontmtx, lu->mtxA, chvmanager, lu->cpus, &facops, lu->options.msglvl, lu->options.msgFile); if ( lu->options.msglvl > 1 ) { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n factor matrix");CHKERRQ(ierr); ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n facops = %9.2f", facops);CHKERRQ(ierr); } } else { IVfill(20, lu->stats, 0); rootchv = FrontMtx_factorInpMtx(lu->frontmtx, lu->mtxA, lu->options.tau, 0.0, chvmanager, &fierr, lu->cpus,lu->stats,lu->options.msglvl,lu->options.msgFile); if (rootchv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"\n matrix found to be singular"); if (fierr >= 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"\n error encountered at front %D", fierr); if(lu->options.FrontMtxInfo){ ierr = PetscPrintf(PETSC_COMM_SELF,"\n %8d pivots, %8d pivot tests, %8d delayed rows and columns\n",lu->stats[0], lu->stats[1], lu->stats[2]);CHKERRQ(ierr); cputotal = lu->cpus[8] ; if ( cputotal > 0.0 ) { ierr = PetscPrintf(PETSC_COMM_SELF, "\n cpus cpus/totaltime" "\n initialize fronts %8.3f %6.2f" "\n load original entries %8.3f %6.2f" "\n update fronts %8.3f %6.2f" "\n assemble postponed data %8.3f %6.2f" "\n factor fronts %8.3f %6.2f" "\n extract postponed data %8.3f %6.2f" "\n store factor entries %8.3f %6.2f" "\n miscellaneous %8.3f %6.2f" "\n total time %8.3f \n", lu->cpus[0], 100.*lu->cpus[0]/cputotal, lu->cpus[1], 100.*lu->cpus[1]/cputotal, lu->cpus[2], 100.*lu->cpus[2]/cputotal, lu->cpus[3], 100.*lu->cpus[3]/cputotal, lu->cpus[4], 100.*lu->cpus[4]/cputotal, lu->cpus[5], 100.*lu->cpus[5]/cputotal, lu->cpus[6], 100.*lu->cpus[6]/cputotal, lu->cpus[7], 100.*lu->cpus[7]/cputotal, cputotal);CHKERRQ(ierr); } } } ChvManager_free(chvmanager); if ( lu->options.msglvl > 0 ) { int err; ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n factor matrix");CHKERRQ(ierr); FrontMtx_writeForHumanEye(lu->frontmtx, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } if ( lu->options.symflag == SPOOLES_SYMMETRIC ) { /* || SPOOLES_HERMITIAN ? */ if ( lu->options.patchAndGoFlag == 1 ) { if ( lu->frontmtx->patchinfo->fudgeIV != NULL ) { if (lu->options.msglvl > 0 ){ ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n small pivots found at these locations");CHKERRQ(ierr); IV_writeForHumanEye(lu->frontmtx->patchinfo->fudgeIV, lu->options.msgFile); } } PatchAndGoInfo_free(lu->frontmtx->patchinfo); } else if ( lu->options.patchAndGoFlag == 2 ) { if (lu->options.msglvl > 0 ){ if ( lu->frontmtx->patchinfo->fudgeIV != NULL ) { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n small pivots found at these locations");CHKERRQ(ierr); IV_writeForHumanEye(lu->frontmtx->patchinfo->fudgeIV, lu->options.msgFile); } if ( lu->frontmtx->patchinfo->fudgeDV != NULL ) { ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n perturbations");CHKERRQ(ierr); DV_writeForHumanEye(lu->frontmtx->patchinfo->fudgeDV, lu->options.msgFile); } } PatchAndGoInfo_free(lu->frontmtx->patchinfo); } } /* post-process the factorization */ FrontMtx_postProcess(lu->frontmtx, lu->options.msglvl, lu->options.msgFile); if ( lu->options.msglvl > 2 ) { int err; ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n factor matrix after post-processing");CHKERRQ(ierr); FrontMtx_writeForHumanEye(lu->frontmtx, lu->options.msgFile); err = fflush(lu->options.msgFile); if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file"); } lu->flg = SAME_NONZERO_PATTERN; lu->CleanUpSpooles = PETSC_TRUE; PetscFunctionReturn(0); }
NM_Status SpoolesSolver :: solve(SparseMtrx *A, FloatArray *b, FloatArray *x) { int errorValue, mtxType, symmetryflag; int seed = 30145, pivotingflag = 0; int *oldToNew, *newToOld; double droptol = 0.0, tau = 1.e300; double cpus [ 10 ]; int stats [ 20 ]; ChvManager *chvmanager; Chv *rootchv; InpMtx *mtxA; DenseMtx *mtxY, *mtxX; // first check whether Lhs is defined if ( !A ) { _error("solveYourselfAt: unknown Lhs"); } // and whether Rhs if ( !b ) { _error("solveYourselfAt: unknown Rhs"); } // and whether previous Solution exist if ( !x ) { _error("solveYourselfAt: unknown solution array"); } if ( x->giveSize() != b->giveSize() ) { _error("solveYourselfAt: size mismatch"); } Timer timer; timer.startTimer(); if ( A->giveType() != SMT_SpoolesMtrx ) { _error("solveYourselfAt: SpoolesSparseMtrx Expected"); } mtxA = ( ( SpoolesSparseMtrx * ) A )->giveInpMtrx(); mtxType = ( ( SpoolesSparseMtrx * ) A )->giveValueType(); symmetryflag = ( ( SpoolesSparseMtrx * ) A )->giveSymmetryFlag(); int i; int neqns = A->giveNumberOfRows(); int nrhs = 1; /* convert right-hand side to DenseMtx */ mtxY = DenseMtx_new(); DenseMtx_init(mtxY, mtxType, 0, 0, neqns, nrhs, 1, neqns); DenseMtx_zero(mtxY); for ( i = 0; i < neqns; i++ ) { DenseMtx_setRealEntry( mtxY, i, 0, b->at(i + 1) ); } if ( ( Lhs != A ) || ( this->lhsVersion != A->giveVersion() ) ) { // // lhs has been changed -> new factorization // Lhs = A; this->lhsVersion = A->giveVersion(); if ( frontmtx ) { FrontMtx_free(frontmtx); } if ( newToOldIV ) { IV_free(newToOldIV); } if ( oldToNewIV ) { IV_free(oldToNewIV); } if ( frontETree ) { ETree_free(frontETree); } if ( symbfacIVL ) { IVL_free(symbfacIVL); } if ( mtxmanager ) { SubMtxManager_free(mtxmanager); } if ( graph ) { Graph_free(graph); } /* * ------------------------------------------------- * STEP 3 : find a low-fill ordering * (1) create the Graph object * (2) order the graph using multiple minimum degree * ------------------------------------------------- */ int nedges; graph = Graph_new(); adjIVL = InpMtx_fullAdjacency(mtxA); nedges = IVL_tsize(adjIVL); Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n graph of the input matrix"); Graph_writeForHumanEye(graph, msgFile); fflush(msgFile); } frontETree = orderViaMMD(graph, seed, msglvl, msgFile); if ( msglvl > 0 ) { fprintf(msgFile, "\n\n front tree from ordering"); ETree_writeForHumanEye(frontETree, msgFile); fflush(msgFile); } /* * ---------------------------------------------------- * STEP 4: get the permutation, permute the front tree, * permute the matrix and right hand side, and * get the symbolic factorization * ---------------------------------------------------- */ oldToNewIV = ETree_oldToNewVtxPerm(frontETree); oldToNew = IV_entries(oldToNewIV); newToOldIV = ETree_newToOldVtxPerm(frontETree); newToOld = IV_entries(newToOldIV); ETree_permuteVertices(frontETree, oldToNewIV); InpMtx_permute(mtxA, oldToNew, oldToNew); if ( symmetryflag == SPOOLES_SYMMETRIC || symmetryflag == SPOOLES_HERMITIAN ) { InpMtx_mapToUpperTriangle(mtxA); } InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS); InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS); symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n old-to-new permutation vector"); IV_writeForHumanEye(oldToNewIV, msgFile); fprintf(msgFile, "\n\n new-to-old permutation vector"); IV_writeForHumanEye(newToOldIV, msgFile); fprintf(msgFile, "\n\n front tree after permutation"); ETree_writeForHumanEye(frontETree, msgFile); fprintf(msgFile, "\n\n input matrix after permutation"); InpMtx_writeForHumanEye(mtxA, msgFile); fprintf(msgFile, "\n\n symbolic factorization"); IVL_writeForHumanEye(symbfacIVL, msgFile); fflush(msgFile); } Tree_writeToFile(frontETree->tree, (char*)"haggar.treef"); /*--------------------------------------------------------------------*/ /* * ------------------------------------------ * STEP 5: initialize the front matrix object * ------------------------------------------ */ frontmtx = FrontMtx_new(); mtxmanager = SubMtxManager_new(); SubMtxManager_init(mtxmanager, NO_LOCK, 0); FrontMtx_init(frontmtx, frontETree, symbfacIVL, mtxType, symmetryflag, FRONTMTX_DENSE_FRONTS, pivotingflag, NO_LOCK, 0, NULL, mtxmanager, msglvl, msgFile); /*--------------------------------------------------------------------*/ /* * ----------------------------------------- * STEP 6: compute the numeric factorization * ----------------------------------------- */ chvmanager = ChvManager_new(); ChvManager_init(chvmanager, NO_LOCK, 1); DVfill(10, cpus, 0.0); IVfill(20, stats, 0); rootchv = FrontMtx_factorInpMtx(frontmtx, mtxA, tau, droptol, chvmanager, & errorValue, cpus, stats, msglvl, msgFile); ChvManager_free(chvmanager); if ( msglvl > 0 ) { fprintf(msgFile, "\n\n factor matrix"); FrontMtx_writeForHumanEye(frontmtx, msgFile); fflush(msgFile); } if ( rootchv != NULL ) { fprintf(msgFile, "\n\n matrix found to be singular\n"); exit(-1); } if ( errorValue >= 0 ) { fprintf(msgFile, "\n\n error encountered at front %d", errorValue); exit(-1); } /*--------------------------------------------------------------------*/ /* * -------------------------------------- * STEP 7: post-process the factorization * -------------------------------------- */ FrontMtx_postProcess(frontmtx, msglvl, msgFile); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n factor matrix after post-processing"); FrontMtx_writeForHumanEye(frontmtx, msgFile); fflush(msgFile); } /*--------------------------------------------------------------------*/ } /* * ---------------------------------------------------- * STEP 4: permute the right hand side * ---------------------------------------------------- */ DenseMtx_permuteRows(mtxY, oldToNewIV); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n right hand side matrix after permutation"); DenseMtx_writeForHumanEye(mtxY, msgFile); } /* * ------------------------------- * STEP 8: solve the linear system * ------------------------------- */ mtxX = DenseMtx_new(); DenseMtx_init(mtxX, mtxType, 0, 0, neqns, nrhs, 1, neqns); DenseMtx_zero(mtxX); FrontMtx_solve(frontmtx, mtxX, mtxY, mtxmanager, cpus, msglvl, msgFile); if ( msglvl > 2 ) { fprintf(msgFile, "\n\n solution matrix in new ordering"); DenseMtx_writeForHumanEye(mtxX, msgFile); fflush(msgFile); } /*--------------------------------------------------------------------*/ /* * ------------------------------------------------------- * STEP 9: permute the solution into the original ordering * ------------------------------------------------------- */ DenseMtx_permuteRows(mtxX, newToOldIV); if ( msglvl > 0 ) { fprintf(msgFile, "\n\n solution matrix in original ordering"); DenseMtx_writeForHumanEye(mtxX, msgFile); fflush(msgFile); } // DenseMtx_writeForMatlab(mtxX, "x", msgFile) ; /*--------------------------------------------------------------------*/ /* fetch data to oofem vectors */ double *xptr = x->givePointer(); for ( i = 0; i < neqns; i++ ) { DenseMtx_realEntry(mtxX, i, 0, xptr + i); // printf ("x(%d) = %e\n", i+1, *(xptr+i)); } // DenseMtx_copyRowIntoVector(mtxX, 0, x->givePointer()); timer.stopTimer(); OOFEM_LOG_DEBUG( "SpoolesSolver info: user time consumed by solution: %.2fs\n", timer.getUtime() ); /* * ----------- * free memory * ----------- */ DenseMtx_free(mtxX); DenseMtx_free(mtxY); /*--------------------------------------------------------------------*/ return ( 1 ); }
/*--------------------------------------------------------------------*/ int main ( int argc, char *argv[] ) { /* -------------------------------------------------- QR all-in-one program (1) read in matrix entries and form InpMtx object of A and A^TA (2) form Graph object of A^TA (3) order matrix and form front tree (4) get the permutation, permute the matrix and front tree and get the symbolic factorization (5) compute the numeric factorization (6) read in right hand side entries (7) compute the solution created -- 98jun11, cca -------------------------------------------------- */ /*--------------------------------------------------------------------*/ char *matrixFileName, *rhsFileName ; ChvManager *chvmanager ; DenseMtx *mtxB, *mtxX ; double facops, imag, real, value ; double cpus[10] ; ETree *frontETree ; FILE *inputFile, *msgFile ; FrontMtx *frontmtx ; Graph *graph ; int ient, irow, jcol, jrhs, jrow, msglvl, neqns, nedges, nent, nrhs, nrow, seed, type ; InpMtx *mtxA ; IV *newToOldIV, *oldToNewIV ; IVL *adjIVL, *symbfacIVL ; SubMtxManager *mtxmanager ; /*--------------------------------------------------------------------*/ /* -------------------- get input parameters -------------------- */ if ( argc != 7 ) { fprintf(stdout, "\n usage: %s msglvl msgFile type matrixFileName rhsFileName seed" "\n msglvl -- message level" "\n msgFile -- message file" "\n type -- type of entries" "\n 1 (SPOOLES_REAL) -- real entries" "\n 2 (SPOOLES_COMPLEX) -- complex entries" "\n matrixFileName -- matrix file name, format" "\n nrow ncol nent" "\n irow jcol entry" "\n ..." "\n note: indices are zero based" "\n rhsFileName -- right hand side file name, format" "\n nrow " "\n entry[0]" "\n ..." "\n entry[nrow-1]" "\n seed -- random number seed, used for ordering" "\n", argv[0]) ; return(0) ; } msglvl = atoi(argv[1]) ; if ( strcmp(argv[2], "stdout") == 0 ) { msgFile = stdout ; } else if ( (msgFile = fopen(argv[2], "a")) == NULL ) { fprintf(stderr, "\n fatal error in %s" "\n unable to open file %s\n", argv[0], argv[2]) ; return(-1) ; } type = atoi(argv[3]) ; matrixFileName = argv[4] ; rhsFileName = argv[5] ; seed = atoi(argv[6]) ; /*--------------------------------------------------------------------*/ /* -------------------------------------------- STEP 1: read the entries from the input file and create the InpMtx object of A -------------------------------------------- */ inputFile = fopen(matrixFileName, "r") ; fscanf(inputFile, "%d %d %d", &nrow, &neqns, &nent) ; mtxA = InpMtx_new() ; InpMtx_init(mtxA, INPMTX_BY_ROWS, type, nent, 0) ; if ( type == SPOOLES_REAL ) { for ( ient = 0 ; ient < nent ; ient++ ) { fscanf(inputFile, "%d %d %le", &irow, &jcol, &value) ; InpMtx_inputRealEntry(mtxA, irow, jcol, value) ; } } else { for ( ient = 0 ; ient < nent ; ient++ ) { fscanf(inputFile, "%d %d %le %le", &irow, &jcol, &real, &imag) ; InpMtx_inputComplexEntry(mtxA, irow, jcol, real, imag) ; } } fclose(inputFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n input matrix") ; InpMtx_writeForHumanEye(mtxA, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ---------------------------------------- STEP 2: read the right hand side entries ---------------------------------------- */ inputFile = fopen(rhsFileName, "r") ; fscanf(inputFile, "%d %d", &nrow, &nrhs) ; mtxB = DenseMtx_new() ; DenseMtx_init(mtxB, type, 0, 0, nrow, nrhs, 1, nrow) ; DenseMtx_zero(mtxB) ; if ( type == SPOOLES_REAL ) { for ( irow = 0 ; irow < nrow ; irow++ ) { fscanf(inputFile, "%d", &jrow) ; for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) { fscanf(inputFile, "%le", &value) ; DenseMtx_setRealEntry(mtxB, jrow, jrhs, value) ; } } } else { for ( irow = 0 ; irow < nrow ; irow++ ) { fscanf(inputFile, "%d", &jrow) ; for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) { fscanf(inputFile, "%le %le", &real, &imag) ; DenseMtx_setComplexEntry(mtxB, jrow, jrhs, real, imag) ; } } } fclose(inputFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n rhs matrix in original ordering") ; DenseMtx_writeForHumanEye(mtxB, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------------- STEP 3 : find a low-fill ordering (1) create the Graph object for A^TA or A^HA (2) order the graph using multiple minimum degree ------------------------------------------------- */ graph = Graph_new() ; adjIVL = InpMtx_adjForATA(mtxA) ; nedges = IVL_tsize(adjIVL) ; Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n graph of A^T A") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n front tree from ordering") ; ETree_writeForHumanEye(frontETree, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------------------------------------------------- STEP 4: get the permutation, permute the matrix and front tree and get the symbolic factorization ----------------------------------------------------- */ oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ; newToOldIV = ETree_newToOldVtxPerm(frontETree) ; InpMtx_permute(mtxA, NULL, IV_entries(oldToNewIV)) ; InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ; symbfacIVL = SymbFac_initFromGraph(frontETree, graph) ; IVL_overwrite(symbfacIVL, oldToNewIV) ; IVL_sortUp(symbfacIVL) ; ETree_permuteVertices(frontETree, oldToNewIV) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n old-to-new permutation vector") ; IV_writeForHumanEye(oldToNewIV, msgFile) ; fprintf(msgFile, "\n\n new-to-old permutation vector") ; IV_writeForHumanEye(newToOldIV, msgFile) ; fprintf(msgFile, "\n\n front tree after permutation") ; ETree_writeForHumanEye(frontETree, msgFile) ; fprintf(msgFile, "\n\n input matrix after permutation") ; InpMtx_writeForHumanEye(mtxA, msgFile) ; fprintf(msgFile, "\n\n symbolic factorization") ; IVL_writeForHumanEye(symbfacIVL, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------ STEP 5: initialize the front matrix object ------------------------------------------ */ frontmtx = FrontMtx_new() ; mtxmanager = SubMtxManager_new() ; SubMtxManager_init(mtxmanager, NO_LOCK, 0) ; if ( type == SPOOLES_REAL ) { FrontMtx_init(frontmtx, frontETree, symbfacIVL, type, SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL, mtxmanager, msglvl, msgFile) ; } else { FrontMtx_init(frontmtx, frontETree, symbfacIVL, type, SPOOLES_HERMITIAN, FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL, mtxmanager, msglvl, msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------------------------------------- STEP 6: compute the numeric factorization ----------------------------------------- */ chvmanager = ChvManager_new() ; ChvManager_init(chvmanager, NO_LOCK, 1) ; DVzero(10, cpus) ; facops = 0.0 ; FrontMtx_QR_factor(frontmtx, mtxA, chvmanager, cpus, &facops, msglvl, msgFile) ; ChvManager_free(chvmanager) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n factor matrix") ; fprintf(msgFile, "\n facops = %9.2f", facops) ; FrontMtx_writeForHumanEye(frontmtx, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* -------------------------------------- STEP 7: post-process the factorization -------------------------------------- */ FrontMtx_postProcess(frontmtx, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n factor matrix after post-processing") ; FrontMtx_writeForHumanEye(frontmtx, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------- STEP 8: solve the linear system ------------------------------- */ mtxX = DenseMtx_new() ; DenseMtx_init(mtxX, type, 0, 0, neqns, nrhs, 1, neqns) ; FrontMtx_QR_solve(frontmtx, mtxA, mtxX, mtxB, mtxmanager, cpus, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n solution matrix in new ordering") ; DenseMtx_writeForHumanEye(mtxX, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------------------- STEP 9: permute the solution into the original ordering ------------------------------------------------------- */ DenseMtx_permuteRows(mtxX, newToOldIV) ; if ( msglvl > 0 ) { fprintf(msgFile, "\n\n solution matrix in original ordering") ; DenseMtx_writeForHumanEye(mtxX, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------ free the working storage ------------------------ */ InpMtx_free(mtxA) ; FrontMtx_free(frontmtx) ; Graph_free(graph) ; DenseMtx_free(mtxX) ; DenseMtx_free(mtxB) ; ETree_free(frontETree) ; IV_free(newToOldIV) ; IV_free(oldToNewIV) ; IVL_free(symbfacIVL) ; SubMtxManager_free(mtxmanager) ; /*--------------------------------------------------------------------*/ return(1) ; }
/*--------------------------------------------------------------------*/ int main ( int argc, char *argv[] ) { /* -------------------------------------------------- all-in-one program to solve A X = B using a multithreaded factorization and solve We use a patch-and-go strategy for the factorization without pivoting (1) read in matrix entries and form DInpMtx object (2) form Graph object (3) order matrix and form front tree (4) get the permutation, permute the matrix and front tree and get the symbolic factorization (5) compute the numeric factorization (6) read in right hand side entries (7) compute the solution created -- 98jun04, cca -------------------------------------------------- */ /*--------------------------------------------------------------------*/ char *matrixFileName, *rhsFileName ; DenseMtx *mtxB, *mtxX ; Chv *rootchv ; ChvManager *chvmanager ; double fudge, imag, real, tau = 100., toosmall, value ; double cpus[10] ; DV *cumopsDV ; ETree *frontETree ; FrontMtx *frontmtx ; FILE *inputFile, *msgFile ; Graph *graph ; InpMtx *mtxA ; int error, ient, irow, jcol, jrhs, jrow, lookahead, msglvl, ncol, nedges, nent, neqns, nfront, nrhs, nrow, nthread, patchAndGoFlag, seed, storeids, storevalues, symmetryflag, type ; int *newToOld, *oldToNew ; int stats[20] ; IV *newToOldIV, *oldToNewIV, *ownersIV ; IVL *adjIVL, *symbfacIVL ; SolveMap *solvemap ; SubMtxManager *mtxmanager ; /*--------------------------------------------------------------------*/ /* -------------------- get input parameters -------------------- */ if ( argc != 14 ) { fprintf(stdout, "\n" "\n usage: %s msglvl msgFile type symmetryflag patchAndGoFlag" "\n fudge toosmall storeids storevalues" "\n matrixFileName rhsFileName seed" "\n msglvl -- message level" "\n msgFile -- message file" "\n type -- type of entries" "\n 1 (SPOOLES_REAL) -- real entries" "\n 2 (SPOOLES_COMPLEX) -- complex entries" "\n symmetryflag -- type of matrix" "\n 0 (SPOOLES_SYMMETRIC) -- symmetric entries" "\n 1 (SPOOLES_HERMITIAN) -- Hermitian entries" "\n 2 (SPOOLES_NONSYMMETRIC) -- nonsymmetric entries" "\n patchAndGoFlag -- flag for the patch-and-go strategy" "\n 0 -- none, stop factorization" "\n 1 -- optimization strategy" "\n 2 -- structural analysis strategy" "\n fudge -- perturbation parameter" "\n toosmall -- upper bound on a small pivot" "\n storeids -- flag to store ids of small pivots" "\n storevalues -- flag to store perturbations" "\n matrixFileName -- matrix file name, format" "\n nrow ncol nent" "\n irow jcol entry" "\n ..." "\n note: indices are zero based" "\n rhsFileName -- right hand side file name, format" "\n nrow nrhs " "\n ..." "\n jrow entry(jrow,0) ... entry(jrow,nrhs-1)" "\n ..." "\n seed -- random number seed, used for ordering" "\n nthread -- number of threads" "\n", argv[0]) ; return(0) ; } msglvl = atoi(argv[1]) ; if ( strcmp(argv[2], "stdout") == 0 ) { msgFile = stdout ; } else if ( (msgFile = fopen(argv[2], "a")) == NULL ) { fprintf(stderr, "\n fatal error in %s" "\n unable to open file %s\n", argv[0], argv[2]) ; return(-1) ; } type = atoi(argv[3]) ; symmetryflag = atoi(argv[4]) ; patchAndGoFlag = atoi(argv[5]) ; fudge = atof(argv[6]) ; toosmall = atof(argv[7]) ; storeids = atoi(argv[8]) ; storevalues = atoi(argv[9]) ; matrixFileName = argv[10] ; rhsFileName = argv[11] ; seed = atoi(argv[12]) ; nthread = atoi(argv[13]) ; /*--------------------------------------------------------------------*/ /* -------------------------------------------- STEP 1: read the entries from the input file and create the InpMtx object -------------------------------------------- */ if ( (inputFile = fopen(matrixFileName, "r")) == NULL ) { fprintf(stderr, "\n unable to open file %s", matrixFileName) ; spoolesFatal(); } fscanf(inputFile, "%d %d %d", &nrow, &ncol, &nent) ; neqns = nrow ; mtxA = InpMtx_new() ; InpMtx_init(mtxA, INPMTX_BY_ROWS, type, nent, 0) ; if ( type == SPOOLES_REAL ) { for ( ient = 0 ; ient < nent ; ient++ ) { fscanf(inputFile, "%d %d %le", &irow, &jcol, &value) ; InpMtx_inputRealEntry(mtxA, irow, jcol, value) ; } } else { for ( ient = 0 ; ient < nent ; ient++ ) { fscanf(inputFile, "%d %d %le %le", &irow, &jcol, &real, &imag) ; InpMtx_inputComplexEntry(mtxA, irow, jcol, real, imag) ; } } fclose(inputFile) ; InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n input matrix") ; InpMtx_writeForHumanEye(mtxA, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------------- STEP 2 : find a low-fill ordering (1) create the Graph object (2) order the graph using multiple minimum degree ------------------------------------------------- */ graph = Graph_new() ; adjIVL = InpMtx_fullAdjacency(mtxA) ; nedges = IVL_tsize(adjIVL) ; Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n graph of the input matrix") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n front tree from ordering") ; ETree_writeForHumanEye(frontETree, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------------------------------------------------- STEP 3: get the permutation, permute the matrix and front tree and get the symbolic factorization ----------------------------------------------------- */ oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ; oldToNew = IV_entries(oldToNewIV) ; newToOldIV = ETree_newToOldVtxPerm(frontETree) ; newToOld = IV_entries(newToOldIV) ; ETree_permuteVertices(frontETree, oldToNewIV) ; InpMtx_permute(mtxA, oldToNew, oldToNew) ; InpMtx_mapToUpperTriangle(mtxA) ; InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS) ; InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ; symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n old-to-new permutation vector") ; IV_writeForHumanEye(oldToNewIV, msgFile) ; fprintf(msgFile, "\n\n new-to-old permutation vector") ; IV_writeForHumanEye(newToOldIV, msgFile) ; fprintf(msgFile, "\n\n front tree after permutation") ; ETree_writeForHumanEye(frontETree, msgFile) ; fprintf(msgFile, "\n\n input matrix after permutation") ; InpMtx_writeForHumanEye(mtxA, msgFile) ; fprintf(msgFile, "\n\n symbolic factorization") ; IVL_writeForHumanEye(symbfacIVL, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------ STEP 4: initialize the front matrix object and the PatchAndGoInfo object to handle small pivots ------------------------------------------ */ frontmtx = FrontMtx_new() ; mtxmanager = SubMtxManager_new() ; SubMtxManager_init(mtxmanager, LOCK_IN_PROCESS, 0) ; FrontMtx_init(frontmtx, frontETree, symbfacIVL, type, symmetryflag, FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, LOCK_IN_PROCESS, 0, NULL, mtxmanager, msglvl, msgFile) ; if ( patchAndGoFlag == 1 ) { frontmtx->patchinfo = PatchAndGoInfo_new() ; PatchAndGoInfo_init(frontmtx->patchinfo, 1, toosmall, fudge, storeids, storevalues) ; } else if ( patchAndGoFlag == 2 ) { frontmtx->patchinfo = PatchAndGoInfo_new() ; PatchAndGoInfo_init(frontmtx->patchinfo, 2, toosmall, fudge, storeids, storevalues) ; } /*--------------------------------------------------------------------*/ /* ------------------------------------------ STEP 5: setup the domain decomposition map ------------------------------------------ */ if ( nthread > (nfront = FrontMtx_nfront(frontmtx)) ) { nthread = nfront ; } cumopsDV = DV_new() ; DV_init(cumopsDV, nthread, NULL) ; ownersIV = ETree_ddMap(frontETree, type, symmetryflag, cumopsDV, 1./(2.*nthread)) ; DV_free(cumopsDV) ; /*--------------------------------------------------------------------*/ /* ----------------------------------------------------- STEP 6: compute the numeric factorization in parallel ----------------------------------------------------- */ chvmanager = ChvManager_new() ; ChvManager_init(chvmanager, LOCK_IN_PROCESS, 1) ; DVfill(10, cpus, 0.0) ; IVfill(20, stats, 0) ; lookahead = 0 ; rootchv = FrontMtx_MT_factorInpMtx(frontmtx, mtxA, tau, 0.0, chvmanager, ownersIV, lookahead, &error, cpus, stats, msglvl, msgFile) ; if ( patchAndGoFlag == 1 ) { if ( frontmtx->patchinfo->fudgeIV != NULL ) { fprintf(msgFile, "\n small pivots found at these locations") ; IV_writeForHumanEye(frontmtx->patchinfo->fudgeIV, msgFile) ; } PatchAndGoInfo_free(frontmtx->patchinfo) ; } else if ( patchAndGoFlag == 2 ) { if ( frontmtx->patchinfo->fudgeIV != NULL ) { fprintf(msgFile, "\n small pivots found at these locations") ; IV_writeForHumanEye(frontmtx->patchinfo->fudgeIV, msgFile) ; } if ( frontmtx->patchinfo->fudgeDV != NULL ) { fprintf(msgFile, "\n perturbations") ; DV_writeForHumanEye(frontmtx->patchinfo->fudgeDV, msgFile) ; } PatchAndGoInfo_free(frontmtx->patchinfo) ; } ChvManager_free(chvmanager) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n factor matrix") ; FrontMtx_writeForHumanEye(frontmtx, msgFile) ; fflush(msgFile) ; } if ( rootchv != NULL ) { fprintf(msgFile, "\n\n matrix found to be singular\n") ; spoolesFatal(); } if ( error >= 0 ) { fprintf(msgFile, "\n\n fatal error at front %d\n", error) ; spoolesFatal(); } /* -------------------------------------- STEP 7: post-process the factorization -------------------------------------- */ FrontMtx_postProcess(frontmtx, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n factor matrix after post-processing") ; FrontMtx_writeForHumanEye(frontmtx, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------------------------------------- STEP 8: read the right hand side matrix B ----------------------------------------- */ if ( (inputFile = fopen(rhsFileName, "r")) == NULL ) { fprintf(stderr, "\n unable to open file %s", rhsFileName) ; spoolesFatal(); } fscanf(inputFile, "%d %d", &nrow, &nrhs) ; mtxB = DenseMtx_new() ; DenseMtx_init(mtxB, type, 0, 0, neqns, nrhs, 1, neqns) ; DenseMtx_zero(mtxB) ; if ( type == SPOOLES_REAL ) { for ( irow = 0 ; irow < nrow ; irow++ ) { fscanf(inputFile, "%d", &jrow) ; for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) { fscanf(inputFile, "%le", &value) ; DenseMtx_setRealEntry(mtxB, jrow, jrhs, value) ; } } } else { for ( irow = 0 ; irow < nrow ; irow++ ) { fscanf(inputFile, "%d", &jrow) ; for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) { fscanf(inputFile, "%le %le", &real, &imag) ; DenseMtx_setComplexEntry(mtxB, jrow, jrhs, real, imag) ; } } } fclose(inputFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n rhs matrix in original ordering") ; DenseMtx_writeForHumanEye(mtxB, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* -------------------------------------------------------------- STEP 9: permute the right hand side into the original ordering -------------------------------------------------------------- */ DenseMtx_permuteRows(mtxB, oldToNewIV) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n right hand side matrix in new ordering") ; DenseMtx_writeForHumanEye(mtxB, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* -------------------------------------------------------- STEP 10: get the solve map object for the parallel solve -------------------------------------------------------- */ solvemap = SolveMap_new() ; SolveMap_ddMap(solvemap, type, FrontMtx_upperBlockIVL(frontmtx), FrontMtx_lowerBlockIVL(frontmtx), nthread, ownersIV, FrontMtx_frontTree(frontmtx), seed, msglvl, msgFile) ; /*--------------------------------------------------------------------*/ /* -------------------------------------------- STEP 11: solve the linear system in parallel -------------------------------------------- */ mtxX = DenseMtx_new() ; DenseMtx_init(mtxX, type, 0, 0, neqns, nrhs, 1, neqns) ; DenseMtx_zero(mtxX) ; FrontMtx_MT_solve(frontmtx, mtxX, mtxB, mtxmanager, solvemap, cpus, msglvl, msgFile) ; if ( msglvl > 1 ) { fprintf(msgFile, "\n\n solution matrix in new ordering") ; DenseMtx_writeForHumanEye(mtxX, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* -------------------------------------------------------- STEP 12: permute the solution into the original ordering -------------------------------------------------------- */ DenseMtx_permuteRows(mtxX, newToOldIV) ; if ( msglvl > 0 ) { fprintf(msgFile, "\n\n solution matrix in original ordering") ; DenseMtx_writeForHumanEye(mtxX, msgFile) ; fflush(msgFile) ; } /*--------------------------------------------------------------------*/ /* ----------- free memory ----------- */ FrontMtx_free(frontmtx) ; DenseMtx_free(mtxX) ; DenseMtx_free(mtxB) ; IV_free(newToOldIV) ; IV_free(oldToNewIV) ; InpMtx_free(mtxA) ; ETree_free(frontETree) ; IVL_free(symbfacIVL) ; SubMtxManager_free(mtxmanager) ; Graph_free(graph) ; SolveMap_free(solvemap) ; IV_free(ownersIV) ; /*--------------------------------------------------------------------*/ return(1) ; }
/* ---------------------------------------------------------------- purpose -- given InpMtx objects that contain A and B, initialize the bridge data structure for the serial factor's, solve's and mvm's. data -- pointer to a Bridge object pprbtype -- pointer to value containing problem type *prbtype = 1 --> A X = B X Lambda, vibration problem *prbtype = 2 --> A X = B X Lambda, buckling problem *prbtype = 3 --> A X = X Lambda, simple eigenvalue problem pneqns -- pointer to value containing number of equations pmxbsz -- pointer to value containing blocksize A -- pointer to InpMtx object containing A B -- pointer to InpMtx object containing B pseed -- pointer to value containing a random number seed pmsglvl -- pointer to value containing a message level msgFile -- message file pointer return value -- 1 -- normal return -1 -- data is NULL -2 -- pprbtype is NULL -3 -- *pprbtype is invalid -4 -- pneqns is NULL -5 -- *pneqns is invalid -6 -- pmxbsz is NULL -7 -- *pmxbsz is invalid -8 -- A and B are NULL -9 -- pseed is NULL -10 -- pmsglvl is NULL -11 -- *pmsglvl > 0 and msgFile is NULL created -- 98aug10, cca ---------------------------------------------------------------- */ int Setup ( void *data, int *pprbtype, int *pneqns, int *pmxbsz, InpMtx *A, InpMtx *B, int *pseed, int *pmsglvl, FILE *msgFile ) { Bridge *bridge = (Bridge *) data ; double sigma[2] ; Graph *graph ; int maxdomainsize, maxsize, maxzeros, msglvl, mxbsz, nedges, neqns, prbtype, seed ; IVL *adjIVL ; #if MYDEBUG > 0 double t1, t2 ; MARKTIME(t1) ; count_Setup++ ; fprintf(stdout, "\n (%d) Setup()", count_Setup) ; fflush(stdout) ; #endif /* -------------------- check the input data -------------------- */ if ( data == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n data is NULL\n") ; return(-1) ; } if ( pprbtype == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n prbtype is NULL\n") ; return(-2) ; } prbtype = *pprbtype ; if ( prbtype < 1 || prbtype > 3 ) { fprintf(stderr, "\n fatal error in Setup()" "\n prbtype = %d, is invalid\n", prbtype) ; return(-3) ; } if ( pneqns == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n pneqns is NULL\n") ; return(-4) ; } neqns = *pneqns ; if ( neqns <= 0 ) { fprintf(stderr, "\n fatal error in Setup()" "\n neqns = %d, is invalid\n", neqns) ; return(-5) ; } if ( pmxbsz == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n pmxbsz is NULL\n") ; return(-6) ; } mxbsz = *pmxbsz ; if ( mxbsz <= 0 ) { fprintf(stderr, "\n fatal error in Setup()" "\n *pmxbsz = %d, is invalid\n", mxbsz) ; return(-7) ; } if ( A == NULL && B == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n A and B are NULL\n") ; return(-8) ; } if ( pseed == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n pseed is NULL\n") ; return(-9) ; } seed = *pseed ; if ( pmsglvl == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n pmsglvl is NULL\n") ; return(-10) ; } msglvl = *pmsglvl ; if ( msglvl > 0 && msgFile == NULL ) { fprintf(stderr, "\n fatal error in Setup()" "\n msglvl = %d, msgFile = NULL\n", msglvl) ; return(-11) ; } bridge->msglvl = msglvl ; bridge->msgFile = msgFile ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n inside Setup()" "\n neqns = %d, prbtype = %d, mxbsz = %d, seed = %d", neqns, prbtype, mxbsz, seed) ; if ( A != NULL ) { fprintf(msgFile, "\n\n matrix A") ; InpMtx_writeForHumanEye(A, msgFile) ; } if ( B != NULL ) { fprintf(msgFile, "\n\n matrix B") ; InpMtx_writeForHumanEye(B, msgFile) ; } fflush(msgFile) ; } bridge->prbtype = prbtype ; bridge->neqns = neqns ; bridge->mxbsz = mxbsz ; bridge->A = A ; bridge->B = B ; bridge->seed = seed ; /* ---------------------------- create and initialize pencil ---------------------------- */ sigma[0] = 1.0; sigma[1] = 0.0; bridge->pencil = Pencil_new() ; Pencil_setDefaultFields(bridge->pencil) ; Pencil_init(bridge->pencil, SPOOLES_REAL, SPOOLES_SYMMETRIC, A, sigma, B) ; /* -------------------------------- convert to row or column vectors -------------------------------- */ if ( A != NULL ) { if ( ! INPMTX_IS_BY_ROWS(A) && ! INPMTX_IS_BY_COLUMNS(A) ) { InpMtx_changeCoordType(A, INPMTX_BY_ROWS) ; } if ( ! INPMTX_IS_BY_VECTORS(A) ) { InpMtx_changeStorageMode(A, INPMTX_BY_VECTORS) ; } } if ( B != NULL ) { if ( ! INPMTX_IS_BY_ROWS(B) && ! INPMTX_IS_BY_COLUMNS(B) ) { InpMtx_changeCoordType(B, INPMTX_BY_ROWS) ; } if ( ! INPMTX_IS_BY_VECTORS(B) ) { InpMtx_changeStorageMode(B, INPMTX_BY_VECTORS) ; } } /* ------------------------------- create a Graph object for A + B ------------------------------- */ graph = Graph_new() ; adjIVL = Pencil_fullAdjacency(bridge->pencil) ; nedges = IVL_tsize(adjIVL), Graph_init2(graph, 0, bridge->neqns, 0, nedges, bridge->neqns, nedges, adjIVL, NULL, NULL) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n graph of the input matrix") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } /* --------------- order the graph --------------- */ maxdomainsize = neqns / 64 ; if ( maxdomainsize == 0 ) { maxdomainsize = 1 ; } maxzeros = (int) (0.01*neqns) ; maxsize = 64 ; bridge->frontETree = orderViaBestOfNDandMS(graph, maxdomainsize, maxzeros, maxsize, bridge->seed, msglvl, msgFile) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n front tree from ordering") ; ETree_writeForHumanEye(bridge->frontETree, msgFile) ; fflush(msgFile) ; } /* ---------------------------------------------- get the old-to-new and new-to-old permutations ---------------------------------------------- */ bridge->oldToNewIV = ETree_oldToNewVtxPerm(bridge->frontETree) ; bridge->newToOldIV = ETree_newToOldVtxPerm(bridge->frontETree) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n old-to-new permutation") ; IV_writeForHumanEye(bridge->oldToNewIV, msgFile) ; fprintf(msgFile, "\n\n new-to-old permutation") ; IV_writeForHumanEye(bridge->newToOldIV, msgFile) ; fflush(msgFile) ; } /* -------------------------------------- permute the vertices in the front tree -------------------------------------- */ ETree_permuteVertices(bridge->frontETree, bridge->oldToNewIV) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n permuted front etree") ; ETree_writeForHumanEye(bridge->frontETree, msgFile) ; fflush(msgFile) ; } /* ------------------------------------------- permute the entries in the pencil. note, after the permutation the entries are mapped into the upper triangle. ------------------------------------------- */ Pencil_permute(bridge->pencil, bridge->oldToNewIV, bridge->oldToNewIV) ; Pencil_mapToUpperTriangle(bridge->pencil) ; Pencil_changeCoordType(bridge->pencil, INPMTX_BY_CHEVRONS) ; Pencil_changeStorageMode(bridge->pencil, INPMTX_BY_VECTORS) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n permuted pencil") ; Pencil_writeForHumanEye(bridge->pencil, msgFile) ; fflush(msgFile) ; } /* ---------------------------------- compute the symbolic factorization ---------------------------------- */ bridge->symbfacIVL = SymbFac_initFromPencil(bridge->frontETree, bridge->pencil) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n symbolic factorization") ; IVL_writeForHumanEye(bridge->symbfacIVL, msgFile) ; fflush(msgFile) ; } /* -------------------------------------------------- create a FrontMtx object to hold the factorization -------------------------------------------------- */ bridge->frontmtx = FrontMtx_new() ; /* ------------------------------------------------------------ create a SubMtxManager object to hold the factor submatrices ------------------------------------------------------------ */ bridge->mtxmanager = SubMtxManager_new() ; SubMtxManager_init(bridge->mtxmanager, NO_LOCK, 0) ; /* ------------------------------------------------------------ allocate the working objects X and Y for the matrix multiply ------------------------------------------------------------ */ bridge->X = DenseMtx_new() ; DenseMtx_init(bridge->X, SPOOLES_REAL, 0, 0, neqns, mxbsz, 1, neqns) ; bridge->Y = DenseMtx_new() ; DenseMtx_init(bridge->Y, SPOOLES_REAL, 0, 0, neqns, mxbsz, 1, neqns) ; /* ------------------------ free the working storage ------------------------ */ Graph_free(graph) ; #if MYDEBUG > 0 MARKTIME(t2) ; time_Setup += t2 - t1 ; fprintf(stdout, ", %8.3f seconds, %8.3f total time", t2 - t1, time_Setup) ; fflush(stdout) ; #endif return(1) ; }
/* ------------------------------------------------------------------- purpose -- given an InpMtx object that contains the structure of A, initialize the bridge data structure for the serial factor's and solve's. note: all parameters are pointers to be compatible with fortran's call by reference. return value -- 1 -- normal return -1 -- bridge is NULL -2 -- mtxA is NULL created -- 98sep17, cca ------------------------------------------------------------------- */ int BridgeMT_setup ( BridgeMT *bridge, InpMtx *mtxA ) { double t0, t1, t2 ; ETree *frontETree ; FILE *msgFile ; Graph *graph ; int compressed, msglvl, nedges, neqns, Neqns ; IV *eqmapIV ; IVL *adjIVL, *symbfacIVL ; MARKTIME(t0) ; /* -------------------- check the input data -------------------- */ if ( bridge == NULL ) { fprintf(stderr, "\n fatal error in BridgeMT_setup()" "\n data is NULL\n") ; return(-1) ; } if ( mtxA == NULL ) { fprintf(stderr, "\n fatal error in BridgeMT_setup()" "\n A is NULL\n") ; return(-2) ; } msglvl = bridge->msglvl ; msgFile = bridge->msgFile ; neqns = bridge->neqns ; if ( ! (INPMTX_IS_BY_ROWS(mtxA) || INPMTX_IS_BY_COLUMNS(mtxA)) ) { /* ------------------------------ change coordinate type to rows ------------------------------ */ InpMtx_changeCoordType(mtxA, INPMTX_BY_ROWS) ; } if ( ! INPMTX_IS_BY_VECTORS(mtxA) ) { /* ------------------------------ change storage mode to vectors ------------------------------ */ InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ; } /* --------------------------- create a Graph object for A --------------------------- */ MARKTIME(t1) ; graph = Graph_new() ; adjIVL = InpMtx_fullAdjacency(mtxA); nedges = bridge->nedges = IVL_tsize(adjIVL), Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL, NULL, NULL) ; MARKTIME(t2) ; bridge->cpus[0] += t2 - t1 ; if ( msglvl > 1 ) { fprintf(msgFile, "\n CPU %8.3f : time to create Graph", t2 - t1) ; fflush(msgFile) ; } if ( msglvl > 3 ) { fprintf(msgFile, "\n\n graph of the input matrix") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } /* ------------------ compress the graph ------------------ */ MARKTIME(t1) ; eqmapIV = Graph_equivMap(graph) ; Neqns = bridge->Neqns = 1 + IV_max(eqmapIV) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n graph's equivalence map") ; IV_writeForHumanEye(eqmapIV, msgFile) ; fflush(msgFile) ; } if ( Neqns < bridge->compressCutoff * neqns ) { Graph *cgraph ; /* ------------------ compress the graph ------------------ */ cgraph = Graph_compress2(graph, eqmapIV, 1) ; Graph_free(graph) ; graph = cgraph ; compressed = 1 ; bridge->Nedges = graph->nedges ; } else { compressed = 0 ; } MARKTIME(t2) ; bridge->cpus[1] += t2 - t1 ; if ( msglvl > 1 ) { fprintf(msgFile, "\n CPU %8.3f : time to create compressed graph", t2 - t1) ; fflush(msgFile) ; } if ( msglvl > 3 ) { fprintf(msgFile, "\n\n graph to order") ; Graph_writeForHumanEye(graph, msgFile) ; fflush(msgFile) ; } /* --------------- order the graph --------------- */ MARKTIME(t1) ; if ( bridge->maxdomainsize <= 0 ) { bridge->maxdomainsize = neqns/32 ; } if ( bridge->maxdomainsize <= 0 ) { bridge->maxdomainsize = 1 ; } if ( bridge->maxnzeros < 0 ) { bridge->maxnzeros = 0.01*neqns ; } if ( bridge->maxsize < 0 ) { bridge->maxsize = neqns ; } frontETree = orderViaBestOfNDandMS(graph, bridge->maxdomainsize, bridge->maxnzeros, bridge->maxsize, bridge->seed, msglvl, msgFile) ; MARKTIME(t2) ; bridge->cpus[2] += t2 - t1 ; if ( msglvl > 1 ) { fprintf(msgFile, "\n CPU %8.3f : time to order graph", t2 - t1) ; fflush(msgFile) ; } if ( msglvl > 3 ) { fprintf(msgFile, "\n\n front tree from ordering") ; ETree_writeForHumanEye(frontETree, msgFile) ; fflush(msgFile) ; } MARKTIME(t1) ; if ( compressed == 1 ) { ETree *etree ; IVL *tempIVL ; /* ---------------------------------------------------------- compute the symbolic factorization of the compressed graph ---------------------------------------------------------- */ tempIVL = SymbFac_initFromGraph(frontETree, graph) ; /* ------------------------------------------------------- expand the symbolic factorization to the original graph ------------------------------------------------------- */ symbfacIVL = IVL_expand(tempIVL, eqmapIV) ; IVL_free(tempIVL) ; /* --------------------- expand the front tree --------------------- */ etree = ETree_expand(frontETree, eqmapIV) ; ETree_free(frontETree) ; frontETree = etree ; } else { /* -------------------------------------------------------- compute the symbolic factorization of the original graph -------------------------------------------------------- */ symbfacIVL = SymbFac_initFromGraph(frontETree, graph) ; } MARKTIME(t2) ; bridge->frontETree = frontETree ; bridge->symbfacIVL = symbfacIVL ; /* ---------------------------------------------- get the old-to-new and new-to-old permutations ---------------------------------------------- */ bridge->oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ; bridge->newToOldIV = ETree_newToOldVtxPerm(frontETree) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n old-to-new permutation") ; IV_writeForHumanEye(bridge->oldToNewIV, msgFile) ; fprintf(msgFile, "\n\n new-to-old permutation") ; IV_writeForHumanEye(bridge->newToOldIV, msgFile) ; fflush(msgFile) ; } /* ------------------------------------------------------ overwrite the symbolic factorization with the permuted indices and sort the lists into ascending order ------------------------------------------------------ */ IVL_overwrite(symbfacIVL, bridge->oldToNewIV) ; IVL_sortUp(symbfacIVL) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n symbolic factorization") ; IVL_writeForHumanEye(symbfacIVL, msgFile) ; fflush(msgFile) ; } /* -------------------------------------- permute the vertices in the front tree -------------------------------------- */ ETree_permuteVertices(frontETree, bridge->oldToNewIV) ; if ( msglvl > 2 ) { fprintf(msgFile, "\n\n permuted front etree") ; ETree_writeForHumanEye(frontETree, msgFile) ; fflush(msgFile) ; } MARKTIME(t2) ; bridge->cpus[3] += t2 - t1 ; if ( msglvl > 1 ) { fprintf(msgFile, "\n CPU %8.3f : time for symbolic factorization", t2 - t1) ; fflush(msgFile) ; } /* ------------------------ free the working storage ------------------------ */ Graph_free(graph) ; IV_free(eqmapIV) ; MARKTIME(t2) ; bridge->cpus[4] += t2 - t0 ; return(1) ; }