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);
}
int
main ( int argc, char *argv[] )
/*
-----------------------------------------------------
test the factor method for a grid matrix
(1) construct a linear system for a nested dissection
ordering on a regular grid
(2) create a solution matrix object
(3) multiply the solution with the matrix
to get a right hand side matrix object
(4) factor the matrix
(5) solve the system
created -- 98may16, cca
-----------------------------------------------------
*/
{
Chv *chv, *rootchv ;
ChvManager *chvmanager ;
DenseMtx *mtxB, *mtxX, *mtxZ ;
FrontMtx *frontmtx ;
InpMtx *mtxA ;
SubMtxManager *mtxmanager ;
double cputotal, droptol, factorops ;
double cpus[9] ;
Drand drand ;
double nops, tau, t1, t2 ;
ETree *frontETree ;
FILE *msgFile ;
int error, lockflag, maxsize, maxzeros, msglvl, neqns,
n1, n2, n3, nrhs, nzf, pivotingflag,
seed, sparsityflag, symmetryflag, type ;
int stats[6] ;
IVL *symbfacIVL ;
if ( argc != 17 ) {
fprintf(stdout,
"\n\n usage : %s msglvl msgFile n1 n2 n3 maxzeros maxsize"
"\n seed type symmetryflag sparsityflag "
"\n pivotingflag tau droptol lockflag nrhs"
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n n1 -- number of grid points in the first direction"
"\n n2 -- number of grid points in the second direction"
"\n n3 -- number of grid points in the third direction"
"\n maxzeros -- max number of zeroes in a front"
"\n maxsize -- max number of internal nodes in a front"
"\n seed -- random number seed"
"\n type -- type of entries"
"\n 1 --> real"
"\n 2 --> complex"
"\n symmetryflag -- symmetry flag"
"\n 0 --> symmetric "
"\n 1 --> hermitian"
"\n 2 --> nonsymmetric"
"\n sparsityflag -- sparsity flag"
"\n 0 --> store dense fronts"
"\n 1 --> store sparse fronts, use droptol to drop entries"
"\n pivotingflag -- pivoting flag"
"\n 0 --> do not pivot"
"\n 1 --> enable pivoting"
"\n tau -- upper bound on factor entries"
"\n used only with pivoting"
"\n droptol -- lower bound on factor entries"
"\n used only with sparse fronts"
"\n lockflag -- flag to specify lock status"
"\n 0 --> mutex lock is not allocated or initialized"
"\n 1 --> mutex lock is allocated and it can synchronize"
"\n only threads in this process."
"\n 2 --> mutex lock is allocated and it can synchronize"
"\n only threads in this and other processes."
"\n nrhs -- # of right hand sides"
"\n", argv[0]) ;
return(-1) ;
}
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) ;
}
n1 = atoi(argv[3]) ;
n2 = atoi(argv[4]) ;
n3 = atoi(argv[5]) ;
maxzeros = atoi(argv[6]) ;
maxsize = atoi(argv[7]) ;
seed = atoi(argv[8]) ;
type = atoi(argv[9]) ;
symmetryflag = atoi(argv[10]) ;
sparsityflag = atoi(argv[11]) ;
pivotingflag = atoi(argv[12]) ;
tau = atof(argv[13]) ;
droptol = atof(argv[14]) ;
lockflag = atoi(argv[15]) ;
nrhs = atoi(argv[16]) ;
fprintf(msgFile,
"\n %s "
"\n msglvl -- %d"
"\n msgFile -- %s"
"\n n1 -- %d"
"\n n2 -- %d"
"\n n3 -- %d"
"\n maxzeros -- %d"
"\n maxsize -- %d"
"\n seed -- %d"
"\n type -- %d"
"\n symmetryflag -- %d"
"\n sparsityflag -- %d"
"\n pivotingflag -- %d"
"\n tau -- %e"
"\n droptol -- %e"
"\n lockflag -- %d"
"\n nrhs -- %d"
"\n",
argv[0], msglvl, argv[2], n1, n2, n3, maxzeros, maxsize,
seed, type, symmetryflag, sparsityflag, pivotingflag,
tau, droptol, lockflag, nrhs) ;
fflush(msgFile) ;
neqns = n1 * n2 * n3 ;
/*
--------------------------------------
initialize the random number generator
--------------------------------------
*/
Drand_setDefaultFields(&drand) ;
Drand_init(&drand) ;
Drand_setSeed(&drand, seed) ;
/*
Drand_setUniform(&drand, 0.0, 1.0) ;
*/
Drand_setNormal(&drand, 0.0, 1.0) ;
/*
--------------------------
generate the linear system
--------------------------
*/
mkNDlinsys(n1, n2, n3, maxzeros, maxsize, type,
symmetryflag, nrhs, seed, msglvl, msgFile,
&frontETree, &symbfacIVL, &mtxA, &mtxX, &mtxB) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n mtxA") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fprintf(msgFile, "\n mtxX") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fprintf(msgFile, "\n mtxB") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
/*
fprintf(msgFile, "\n neqns = %d ;", n1*n2*n3) ;
fprintf(msgFile, "\n nrhs = %d ;", nrhs) ;
fprintf(msgFile, "\n A = zeros(neqns, neqns) ;") ;
fprintf(msgFile, "\n X = zeros(neqns, nrhs) ;") ;
fprintf(msgFile, "\n B = zeros(neqns, nrhs) ;") ;
InpMtx_writeForMatlab(mtxA, "A", msgFile) ;
DenseMtx_writeForMatlab(mtxX, "X", msgFile) ;
DenseMtx_writeForMatlab(mtxB, "B", msgFile) ;
{
int *ivec1 = InpMtx_ivec1(mtxA) ;
int *ivec2 = InpMtx_ivec2(mtxA) ;
double *dvec = InpMtx_dvec(mtxA) ;
int ichv, ii, col, offset, row, nent = InpMtx_nent(mtxA) ;
fprintf(msgFile, "\n coordType = %d", mtxA->coordType) ;
fprintf(msgFile, "\n start of matrix output file") ;
fprintf(msgFile, "\n %d %d %d", n1*n2*n3, n1*n2*n3, nent) ;
for ( ii = 0 ; ii < nent ; ii++ ) {
ichv = ivec1[ii] ;
if ( (offset = ivec2[ii]) >= 0 ) {
row = ichv, col = row + offset ;
} else {
col = ichv, row = col - offset ;
}
fprintf(msgFile, "\n %d %d %24.16e %24.16e",
row, col, dvec[2*ii], dvec[2*ii+1]) ;
}
}
{
int ii, jj ;
double imag, real ;
fprintf(msgFile, "\n start of rhs output file") ;
fprintf(msgFile, "\n %d %d", n1*n2*n3, nrhs) ;
for ( ii = 0 ; ii < n1*n2*n3 ; ii++ ) {
fprintf(msgFile, "\n %d ", ii) ;
for ( jj = 0 ; jj < nrhs ; jj++ ) {
DenseMtx_complexEntry(mtxB, ii, jj, &real, &imag) ;
fprintf(msgFile, " %24.16e %24.16e", real, imag) ;
}
}
}
*/
/*
------------------------------
initialize the FrontMtx object
------------------------------
*/
MARKTIME(t1) ;
frontmtx = FrontMtx_new() ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, lockflag, 0) ;
FrontMtx_init(frontmtx, frontETree, symbfacIVL,
type, symmetryflag, sparsityflag, pivotingflag,
lockflag, 0, NULL, mtxmanager, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : initialize the front matrix",
t2 - t1) ;
if ( msglvl > 1 ) {
fprintf(msgFile,
"\n nendD = %d, nentL = %d, nentU = %d",
frontmtx->nentD, frontmtx->nentL, frontmtx->nentU) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n front matrix initialized") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
SubMtxManager_writeForHumanEye(mtxmanager, msgFile) ;
/*
-----------------
factor the matrix
-----------------
*/
nzf = ETree_nFactorEntries(frontETree, symmetryflag) ;
factorops = ETree_nFactorOps(frontETree, type, symmetryflag) ;
fprintf(msgFile,
"\n %d factor entries, %.0f factor ops, %8.3f ratio",
nzf, factorops, factorops/nzf) ;
IVzero(6, stats) ;
DVzero(9, cpus) ;
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, lockflag, 1) ;
MARKTIME(t1) ;
rootchv = FrontMtx_factorInpMtx(frontmtx, mtxA, tau, droptol,
chvmanager, &error, cpus,
stats, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : factor matrix, %8.3f mflops",
t2 - t1, 1.e-6*factorops/(t2-t1)) ;
if ( rootchv != NULL ) {
fprintf(msgFile, "\n\n factorization did not complete") ;
for ( chv = rootchv ; chv != NULL ; chv = chv->next ) {
fprintf(stdout, "\n chv %d, nD = %d, nL = %d, nU = %d",
chv->id, chv->nD, chv->nL, chv->nU) ;
}
}
if ( error >= 0 ) {
fprintf(msgFile, "\n\n error encountered at front %d\n", error) ;
exit(-1) ;
}
fprintf(msgFile,
"\n %8d pivots, %8d pivot tests, %8d delayed rows and columns",
stats[0], stats[1], stats[2]) ;
if ( frontmtx->rowadjIVL != NULL ) {
fprintf(msgFile,
"\n %d entries in rowadjIVL", frontmtx->rowadjIVL->tsize) ;
}
if ( frontmtx->coladjIVL != NULL ) {
fprintf(msgFile,
", %d entries in coladjIVL", frontmtx->coladjIVL->tsize) ;
}
if ( frontmtx->upperblockIVL != NULL ) {
fprintf(msgFile,
"\n %d fronts, %d entries in upperblockIVL",
frontmtx->nfront, frontmtx->upperblockIVL->tsize) ;
}
if ( frontmtx->lowerblockIVL != NULL ) {
fprintf(msgFile,
", %d entries in lowerblockIVL",
frontmtx->lowerblockIVL->tsize) ;
}
fprintf(msgFile,
"\n %d entries in D, %d entries in L, %d entries in U",
stats[3], stats[4], stats[5]) ;
fprintf(msgFile, "\n %d locks", frontmtx->nlocks) ;
cputotal = cpus[8] ;
if ( cputotal > 0.0 ) {
fprintf(msgFile,
"\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",
cpus[0], 100.*cpus[0]/cputotal,
cpus[1], 100.*cpus[1]/cputotal,
cpus[2], 100.*cpus[2]/cputotal,
cpus[3], 100.*cpus[3]/cputotal,
cpus[4], 100.*cpus[4]/cputotal,
cpus[5], 100.*cpus[5]/cputotal,
cpus[6], 100.*cpus[6]/cputotal,
cpus[7], 100.*cpus[7]/cputotal, cputotal) ;
}
SubMtxManager_writeForHumanEye(mtxmanager, msgFile) ;
ChvManager_writeForHumanEye(chvmanager, msgFile) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front factor matrix") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n %% MATLAB file: front factor matrix") ;
FrontMtx_writeForMatlab(frontmtx, "L", "D", "U", msgFile) ;
}
/*
------------------------------
post-process the factor matrix
------------------------------
*/
MARKTIME(t1) ;
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : post-process the matrix", t2 - t1) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
}
fprintf(msgFile, "\n\n after post-processing") ;
SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
/*
code to test out the IO methods.
write the matrix to a file, free it,
then read it back in.
note: formatted files do not have much accuracy.
*/
/*
FrontMtx_writeToFile(frontmtx, "temp.frontmtxb") ;
FrontMtx_free(frontmtx) ;
frontmtx = FrontMtx_new() ;
FrontMtx_readFromFile(frontmtx, "temp.frontmtxb") ;
frontmtx->manager = mtxmanager ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
*/
/*
----------------
solve the system
----------------
*/
neqns = mtxB->nrow ;
nrhs = mtxB->ncol ;
mtxZ = DenseMtx_new() ;
DenseMtx_init(mtxZ, type, 0, 0, neqns, nrhs, 1, neqns) ;
DenseMtx_zero(mtxZ) ;
if ( type == SPOOLES_REAL ) {
nops = frontmtx->nentD + 2*frontmtx->nentU ;
if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) {
nops += 2*frontmtx->nentL ;
} else {
nops += 2*frontmtx->nentU ;
}
} else if ( type == SPOOLES_COMPLEX ) {
nops = 8*frontmtx->nentD + 8*frontmtx->nentU ;
if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) {
nops += 8*frontmtx->nentL ;
} else {
nops += 8*frontmtx->nentU ;
}
}
nops *= nrhs ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n rhs") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(stdout) ;
}
DVzero(6, cpus) ;
MARKTIME(t1) ;
FrontMtx_solve(frontmtx, mtxZ, mtxB, mtxmanager,
cpus, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : solve the system, %.3f mflops",
t2 - t1, 1.e-6*nops/(t2 - t1)) ;
cputotal = t2 - t1 ;
if ( cputotal > 0.0 ) {
fprintf(msgFile,
"\n set up solves %8.3f %6.2f"
"\n load rhs and store solution %8.3f %6.2f"
"\n forward solve %8.3f %6.2f"
"\n diagonal solve %8.3f %6.2f"
"\n backward solve %8.3f %6.2f"
"\n total time %8.3f",
cpus[0], 100.*cpus[0]/cputotal,
cpus[1], 100.*cpus[1]/cputotal,
cpus[2], 100.*cpus[2]/cputotal,
cpus[3], 100.*cpus[3]/cputotal,
cpus[4], 100.*cpus[4]/cputotal, cputotal) ;
}
/*
fprintf(msgFile, "\n Z = zeros(neqns, nrhs) ;") ;
DenseMtx_writeForMatlab(mtxZ, "Z", msgFile) ;
*/
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n computed solution") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
DenseMtx_sub(mtxZ, mtxX) ;
fprintf(msgFile, "\n\n maxabs error = %12.4e", DenseMtx_maxabs(mtxZ)) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n error") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
fprintf(msgFile, "\n\n after solve") ;
SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
/*
------------------------
free the working storage
------------------------
*/
InpMtx_free(mtxA) ;
DenseMtx_free(mtxX) ;
DenseMtx_free(mtxB) ;
DenseMtx_free(mtxZ) ;
FrontMtx_free(frontmtx) ;
ETree_free(frontETree) ;
IVL_free(symbfacIVL) ;
ChvManager_free(chvmanager) ;
SubMtxManager_free(mtxmanager) ;
fprintf(msgFile, "\n") ;
fclose(msgFile) ;
return(1) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
---------------------------------------------------
test the QR factor method for a FrontMtx object
on an n1 x n2 x n3 grid
(1) generate an overdetermined system AX = B
(2) factor the matrix
(3) solve the systems
created -- 97apr11, dkw
modified -- 98may28, cca
---------------------------------------------------
*/
{
ChvManager *chvmanager ;
DenseMtx *mtxB, *mtxX, *mtxZ ;
double cputotal, factorops ;
double cpus[9] ;
double nops, t1, t2 ;
ETree *frontETree ;
FILE *msgFile ;
FrontMtx *frontmtx ;
InpMtx *mtxA ;
int msglvl, neqns, nrhs, n1, n2, n3, seed, type ;
IVL *symbfacIVL ;
SubMtxManager *mtxmanager ;
if ( argc != 9 ) {
fprintf(stdout,
"\n\n usage : %s msglvl msgFile n1 n2 n3 seed nrhs "
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n n1 -- # of points in the first direction"
"\n n2 -- # of points in the second direction"
"\n n3 -- # of points in the third direction"
"\n seed -- random number seed"
"\n nrhs -- # of right hand sides"
"\n type -- type of linear system"
"\n 1 -- real"
"\n 2 -- complex"
"\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) ;
}
n1 = atoi(argv[3]) ;
n2 = atoi(argv[4]) ;
n3 = atoi(argv[5]) ;
seed = atoi(argv[6]) ;
nrhs = atoi(argv[7]) ;
type = atoi(argv[8]) ;
fprintf(msgFile,
"\n %s "
"\n msglvl -- %d"
"\n msgFile -- %s"
"\n n1 -- %d"
"\n n2 -- %d"
"\n n3 -- %d"
"\n seed -- %d"
"\n nrhs -- %d"
"\n type -- %d"
"\n",
argv[0], msglvl, argv[2], n1, n2, n3, seed, nrhs, type) ;
fflush(msgFile) ;
neqns = n1*n2*n3 ;
if ( type != SPOOLES_REAL && type != SPOOLES_COMPLEX ) {
fprintf(stderr, "\n fatal error, type must be real or complex") ;
exit(-1) ;
}
/*
------------------------------------------
generate the A X = B overdetermined system
------------------------------------------
*/
mkNDlinsysQR(n1, n2, n3, type, nrhs, seed, msglvl, msgFile,
&frontETree, &symbfacIVL, &mtxA, &mtxX, &mtxB) ;
/*
------------------------------
initialize the FrontMtx object
------------------------------
*/
MARKTIME(t1) ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, NO_LOCK, 0) ;
frontmtx = FrontMtx_new() ;
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 if ( type == SPOOLES_COMPLEX ) {
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type,
SPOOLES_HERMITIAN, FRONTMTX_DENSE_FRONTS,
SPOOLES_NO_PIVOTING, NO_LOCK,
0, NULL, mtxmanager, msglvl, msgFile) ;
}
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : FrontMtx initialized", t2 - t1) ;
fflush(msgFile) ;
/*
-----------------
factor the matrix
-----------------
*/
DVzero(6, cpus) ;
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, NO_LOCK, 0) ;
factorops = 0.0 ;
MARKTIME(t1) ;
FrontMtx_QR_factor(frontmtx, mtxA, chvmanager,
cpus, &factorops, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n after QR_factor() call, facops = %8.2f",factorops) ;
fprintf(msgFile, "\n CPU %8.3f : FrontMtx_QR_factor, %8.3f mflops",
t2 - t1, 1.e-6*factorops/(t2-t1)) ;
cputotal = t2 - t1 ;
if ( cputotal > 0.0 ) {
fprintf(msgFile, "\n"
"\n setup factorization %8.3f %6.2f"
"\n setup fronts %8.3f %6.2f"
"\n factor fronts %8.3f %6.2f"
"\n store factor %8.3f %6.2f"
"\n store update %8.3f %6.2f"
"\n miscellaneous %8.3f %6.2f"
"\n total time %8.3f",
cpus[0], 100.*cpus[0]/cputotal,
cpus[1], 100.*cpus[1]/cputotal,
cpus[2], 100.*cpus[2]/cputotal,
cpus[3], 100.*cpus[3]/cputotal,
cpus[4], 100.*cpus[4]/cputotal,
cpus[5], 100.*cpus[5]/cputotal, cputotal) ;
}
/*
------------------------------
post-process the factor matrix
------------------------------
*/
MARKTIME(t1) ;
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : post-process the matrix", t2 - t1)
;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
}
fprintf(msgFile, "\n\n after post-processing") ;
SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
/*
----------------
solve the system
----------------
*/
mtxZ = DenseMtx_new() ;
DenseMtx_init(mtxZ, type, 0, 0, neqns, nrhs, 1, neqns) ;
DenseMtx_zero(mtxZ) ;
if ( type == SPOOLES_REAL ) {
nops = frontmtx->nentD + 2*frontmtx->nentU ;
if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) {
nops += 2*frontmtx->nentL ;
} else {
nops += 2*frontmtx->nentU ;
}
} else if ( type == SPOOLES_COMPLEX ) {
nops = 8*frontmtx->nentD + 8*frontmtx->nentU ;
if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) {
nops += 8*frontmtx->nentL ;
} else {
nops += 8*frontmtx->nentU ;
}
}
nops *= nrhs ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n rhs") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(stdout) ;
}
DVzero(6, cpus) ;
MARKTIME(t1) ;
FrontMtx_QR_solve(frontmtx, mtxA, mtxZ, mtxB, mtxmanager,
cpus, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : solve the system, %.3f mflops",
t2 - t1, 1.e-6*nops/(t2 - t1)) ;
cputotal = t2 - t1 ;
if ( cputotal > 0.0 ) {
fprintf(msgFile,
"\n CPU %%"
"\n A^TB matrix-matrix multiply %8.3f %6.2f"
"\n set up solves %8.3f %6.2f"
"\n load rhs and store solution %8.3f %6.2f"
"\n forward solve %8.3f %6.2f"
"\n diagonal solve %8.3f %6.2f"
"\n backward solve %8.3f %6.2f"
"\n total solve time %8.3f %6.2f"
"\n total QR solve time %8.3f",
cpus[6], 100.*cpus[6]/cputotal,
cpus[0], 100.*cpus[0]/cputotal,
cpus[1], 100.*cpus[1]/cputotal,
cpus[2], 100.*cpus[2]/cputotal,
cpus[3], 100.*cpus[3]/cputotal,
cpus[4], 100.*cpus[4]/cputotal,
cpus[5], 100.*cpus[5]/cputotal, cputotal) ;
}
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n computed solution") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
/*
-----------------
compute the error
-----------------
*/
DenseMtx_sub(mtxZ, mtxX) ;
fprintf(msgFile, "\n\n maxabs error = %12.4e",
DenseMtx_maxabs(mtxZ)) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n error") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
fprintf(msgFile, "\n\n after solve") ;
SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
/*
------------------------
free the working storage
------------------------
*/
InpMtx_free(mtxA) ;
DenseMtx_free(mtxX) ;
DenseMtx_free(mtxZ) ;
DenseMtx_free(mtxB) ;
FrontMtx_free(frontmtx) ;
IVL_free(symbfacIVL) ;
ETree_free(frontETree) ;
SubMtxManager_free(mtxmanager) ;
ChvManager_free(chvmanager) ;
fprintf(msgFile, "\n") ;
fclose(msgFile) ;
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) ; }
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[] )
/*
-----------------------------------------------------
test the factor method for a grid matrix
(0) read in matrix from source file
(1) conver data matrix to InpMtx object if necessary
(2) create Graph and ETree object if necessary
(3) read in/create an ETree object
(4) create a solution matrix object
(5) multiply the solution with the matrix
to get a right hand side matrix object
(6) factor the matrix
(7) solve the system
created -- 98dec30, jwu
-----------------------------------------------------
*/
{
char etreeFileName[80], mtxFileName[80], *cpt, rhsFileName[80],
srcFileName[80], ctemp[81], msgFileName[80], slnFileName[80] ;
Chv *chv, *rootchv ;
ChvManager *chvmanager ;
DenseMtx *mtxB, *mtxQ, *mtxX, *mtxZ ;
double one[2] = { 1.0, 0.0 } ;
FrontMtx *frontmtx ;
InpMtx *mtxA ;
SubMtxManager *mtxmanager ;
double cputotal, droptol, conv_tol, factorops ;
double cpus[9] ;
Drand drand ;
double nops, tau, t1, t2 ;
ETree *frontETree ;
Graph *graph ;
FILE *msgFile, *inFile ;
int error, loc, msglvl, neqns, nzf, iformat,
pivotingflag, rc, seed, sparsityflag, symmetryflag,
method[METHODS], type, nrhs, etreeflag ;
int stats[6] ;
int nnzA, Ik, itermax, zversion, iterout ;
IV *newToOldIV, *oldToNewIV ;
IVL *symbfacIVL ;
int i, j, k, m, n, imethod, maxdomainsize, maxzeros, maxsize;
int nouter,ninner ;
if ( argc != 2 ) {
fprintf(stdout,
"\n\n usage : %s inFile"
"\n inFile -- input filename"
"\n", argv[0]) ;
return(-1) ;
}
/* read input file */
inFile = fopen(argv[1], "r");
if (inFile == (FILE *)NULL) {
fprintf(stderr, "\n fatal error in %s: unable to open file %s\n",
argv[0], argv[1]) ;
return(-1) ;
}
for (i=0; i<METHODS; i++) method[i]=-1;
imethod=0;
k=0;
while (1) {
fgets(ctemp, 80, inFile);
if (ctemp[0] != '*') {
/*printf("l=%2d:%s\n", strlen(ctemp),ctemp);*/
if (strlen(ctemp)==80) {
fprintf(stderr, "\n fatal error in %s: input line contains more than "
"80 characters.\n",argv[0]);
exit(-1);
}
if (k==0) {
sscanf(ctemp, "%d", &iformat);
if (iformat < 0 || iformat > 2) {
fprintf(stderr, "\n fatal error in %s: "
"invalid source matrix format\n",argv[0]) ;
return(-1) ;
}
}
else if (k==1)
sscanf(ctemp, "%s", srcFileName);
else if (k==2)
sscanf(ctemp, "%s", mtxFileName);
else if (k==3) {
sscanf(ctemp, "%d", &etreeflag);
if (etreeflag < 0 || etreeflag > 4) {
fprintf(stderr, "\n fatal error in %s: "
"invalid etree file status\n",argv[0]) ;
return(-1) ;
}
}
else if (k==4)
sscanf(ctemp, "%s", etreeFileName);
else if (k==5)
sscanf(ctemp, "%s", rhsFileName);
else if (k==6)
sscanf(ctemp, "%s", slnFileName);
else if (k==7){
sscanf(ctemp, "%s", msgFileName);
if ( strcmp(msgFileName, "stdout") == 0 ) {
msgFile = stdout ;
}
else if ( (msgFile = fopen(msgFileName, "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n", argv[0], ctemp) ;
return(-1) ;
}
}
else if (k==8)
sscanf(ctemp, "%d %d %d %d %d %d",
&msglvl, &seed, &nrhs, &Ik, &itermax, &iterout);
else if (k==9)
sscanf(ctemp, "%d %d %d", &symmetryflag, &sparsityflag, &pivotingflag);
else if (k==10)
sscanf(ctemp, "%lf %lf %lf", &tau, &droptol, &conv_tol);
else if (k==11) {
/*
for (j=0; j<strlen(ctemp); j++) {
printf("j=%2d:%s",j,ctemp+j);
if (ctemp[j] == ' ' && ctemp[j+1] != ' ') {
sscanf(ctemp+j, "%d", method+imethod);
printf("method[%d]=%d\n",imethod,method[imethod]);
if (method[imethod] < 0) break;
imethod++;
}
}
*/
imethod = sscanf(ctemp,"%d %d %d %d %d %d %d %d %d %d",
method, method+1, method+2, method+3, method+4,
method+5, method+6, method+7, method+8, method+9);
/*printf("imethod=%d\n",imethod);*/
for (j=0; j<imethod; j++) {
/*printf("method[%d]=%d\n",j,method[j]);*/
if (method[j]<0) {
imethod=j;
break;
}
}
if (imethod == 0) {
fprintf(msgFile,"No method assigned in input file\n");
return(-1);
}
}
k++;
}
if (k==12) break;
}
fclose(inFile);
/* reset nrhs to 1 */
if (nrhs > 1) {
fprintf(msgFile,"*** Multiple right-hand-side vectors is not allowed yet.\n");
fprintf(msgFile,"*** nrhs is reset to 1.\n");
nrhs =1;
}
fprintf(msgFile,
"\n %s "
"\n srcFileName -- %s"
"\n mtxFileName -- %s"
"\n etreeFileName -- %s"
"\n rhsFileName -- %s"
"\n msglvl -- %d"
"\n seed -- %d"
"\n symmetryflag -- %d"
"\n sparsityflag -- %d"
"\n pivotingflag -- %d"
"\n tau -- %e"
"\n droptol -- %e"
"\n conv_tol -- %e"
"\n method -- ",
argv[0], srcFileName, mtxFileName, etreeFileName, rhsFileName,
msglvl, seed, symmetryflag, sparsityflag, pivotingflag,
tau, droptol, conv_tol) ;
for (k=0; k<imethod; k++)
fprintf(msgFile, "%d ", method[k]);
fprintf(msgFile, "\n ", method[k]);
fflush(msgFile) ;
/*
--------------------------------------
initialize the random number generator
--------------------------------------
*/
Drand_setDefaultFields(&drand) ;
Drand_init(&drand) ;
Drand_setSeed(&drand, seed) ;
/*Drand_setUniform(&drand, 0.0, 1.0) ;*/
Drand_setNormal(&drand, 0.0, 1.0) ;
/*
----------------------------------------------
read in or convert source to the InpMtx object
----------------------------------------------
*/
rc = 1;
if ( strcmp(srcFileName, "none") == 0 ) {
fprintf(msgFile, "\n no file to read from") ;
exit(-1) ;
}
mtxA = InpMtx_new() ;
MARKTIME(t1) ;
if (iformat == 0) { /* InpMtx source format */
rc = InpMtx_readFromFile(mtxA, srcFileName) ;
strcpy(mtxFileName, srcFileName);
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from InpMtx_readFromFile(%p,%s)",
rc, mtxA, srcFileName) ;
}
else if (iformat == 1) { /* HBF source format */
rc = InpMtx_readFromHBfile(mtxA, srcFileName) ;
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from InpMtx_readFromHBfile(%p,%s)",
rc, mtxA, srcFileName) ;
}
else { /* AIJ2 source format */
rc = InpMtx_readFromAIJ2file(mtxA, srcFileName) ;
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from InpMtx_readFromAIJ2file(%p,%s)",
rc, mtxA, srcFileName) ;
}
MARKTIME(t2) ;
if (iformat>0 && strcmp(mtxFileName, "none") != 0 ) {
rc = InpMtx_writeToFile(mtxA, mtxFileName) ;
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from InpMtx_writeToFile(%p,%s)",
rc, mtxA, mtxFileName) ;
}
fprintf(msgFile, "\n CPU %8.3f : read in (+ convert to) mtxA from file %s",
t2 - t1, mtxFileName) ;
if (rc != 1) {
goto end_read;
}
type = mtxA->inputMode ;
neqns = 1 + IVmax(mtxA->nent, InpMtx_ivec1(mtxA), &loc) ;
if ( INPMTX_IS_BY_ROWS(mtxA) ) {
fprintf(msgFile, "\n matrix coordinate type is rows") ;
} else if ( INPMTX_IS_BY_COLUMNS(mtxA) ) {
fprintf(msgFile, "\n matrix coordinate type is columns") ;
} else if ( INPMTX_IS_BY_CHEVRONS(mtxA) ) {
fprintf(msgFile, "\n matrix coordinate type is chevrons") ;
} else {
fprintf(msgFile, "\n\n, error, bad coordinate type") ;
rc=-1;
goto end_read;
}
if ( INPMTX_IS_RAW_DATA(mtxA) ) {
fprintf(msgFile, "\n matrix storage mode is raw data\n") ;
} else if ( INPMTX_IS_SORTED(mtxA) ) {
fprintf(msgFile, "\n matrix storage mode is sorted\n") ;
} else if ( INPMTX_IS_BY_VECTORS(mtxA) ) {
fprintf(msgFile, "\n matrix storage mode is by vectors\n") ;
} else {
fprintf(msgFile, "\n\n, error, bad storage mode") ;
rc=-1;
goto end_read;
}
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n after reading InpMtx object from file %s",
mtxFileName) ;
if ( msglvl == 2 ) {
InpMtx_writeStats(mtxA, msgFile) ;
} else {
InpMtx_writeForHumanEye(mtxA, msgFile) ;
}
fflush(msgFile) ;
}
/*
Get the nonzeros in matrix A and print it
*/
nnzA = InpMtx_nent( mtxA );
fprintf(msgFile, "\n\n Input matrix size %d NNZ %d",
neqns, nnzA) ;
/*
--------------------------------------------------------
generate the linear system
1. generate solution matrix and fill with random numbers
2. generate rhs matrix and fill with zeros
3. compute matrix-matrix multiply
--------------------------------------------------------
*/
MARKTIME(t1) ;
mtxX = DenseMtx_new() ;
DenseMtx_init(mtxX, type, 0, -1, neqns, nrhs, 1, neqns) ;
mtxB = DenseMtx_new() ;
if (strcmp(rhsFileName, "none")) {
rc = DenseMtx_readFromFile(mtxB, rhsFileName) ;
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from DenseMtx_readFromFile(%p,%s)",
rc, mtxB, rhsFileName) ;
DenseMtx_zero(mtxX) ;
}
else {
DenseMtx_init(mtxB, type, 1, -1, neqns, nrhs, 1, neqns) ;
DenseMtx_fillRandomEntries(mtxX, &drand) ;
DenseMtx_zero(mtxB) ;
switch ( symmetryflag ) {
case SPOOLES_SYMMETRIC :
InpMtx_sym_mmm(mtxA, mtxB, one, mtxX) ;
break ;
case SPOOLES_HERMITIAN :
InpMtx_herm_mmm(mtxA, mtxB, one, mtxX) ;
break ;
case SPOOLES_NONSYMMETRIC :
InpMtx_nonsym_mmm(mtxA, mtxB, one, mtxX) ;
break ;
default :
break ;
}
}
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : set up the solution and rhs ",
t2 - t1) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n original mtxX") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fprintf(msgFile, "\n\n original mtxB") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
if (rc != 1) {
InpMtx_free(mtxA);
DenseMtx_free(mtxX);
DenseMtx_free(mtxB);
goto end_init;
}
/*
------------------------
read in/create the ETree object
------------------------
*/
MARKTIME(t1) ;
if (etreeflag == 0) { /* read in ETree from file */
if ( strcmp(etreeFileName, "none") == 0 )
fprintf(msgFile, "\n no file to read from") ;
frontETree = ETree_new() ;
rc = ETree_readFromFile(frontETree, etreeFileName) ;
if (rc!=1)
fprintf(msgFile, "\n return value %d from ETree_readFromFile(%p,%s)",
rc, frontETree, etreeFileName) ;
}
else {
graph = Graph_new() ;
rc = InpMtx_createGraph(mtxA, graph);
if (rc!=1) {
fprintf(msgFile, "\n return value %d from InpMtx_createGraph(%p,%p)",
rc, mtxA, graph) ;
Graph_free(graph);
goto end_tree;
}
if (etreeflag == 1) { /* Via BestOfNDandMS */
maxdomainsize = 500; maxzeros = 1000; maxsize = 64 ;
frontETree = orderViaBestOfNDandMS(graph, maxdomainsize, maxzeros,
maxsize, seed, msglvl, msgFile) ;
}
else if (etreeflag == 2) { /* Via MMD */
frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ;
}
else if (etreeflag == 3) { /* Via MS */
maxdomainsize = 500;
frontETree = orderViaMS(graph, maxdomainsize, seed, msglvl, msgFile) ;
}
else if (etreeflag == 4) { /* Via ND */
maxdomainsize = 500;
frontETree = orderViaND(graph, maxdomainsize, seed, msglvl, msgFile) ;
}
Graph_free(graph);
/* optionally write out the ETree object */
if ( strcmp(etreeFileName, "none") != 0 ) {
fprintf(msgFile, "\n\n writing out ETree to file %s",
etreeFileName) ;
ETree_writeToFile(frontETree, etreeFileName) ;
}
}
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : read in/create frontETree from file %s",
t2 - t1, etreeFileName) ;
if ( rc != 1 ) {
ETree_free(frontETree);
goto end_tree;
}
ETree_leftJustify(frontETree) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n after reading ETree object from file %s",
etreeFileName) ;
if ( msglvl == 2 ) {
ETree_writeStats(frontETree, msgFile) ;
} else {
ETree_writeForHumanEye(frontETree, msgFile) ;
}
}
fflush(msgFile) ;
/*
--------------------------------------------------
get the permutations, permute the matrix and the
front tree, and compute the symbolic factorization
--------------------------------------------------
*/
MARKTIME(t1) ;
oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ;
newToOldIV = ETree_newToOldVtxPerm(frontETree) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : get permutations", t2 - t1) ;
MARKTIME(t1) ;
ETree_permuteVertices(frontETree, oldToNewIV) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : permute front tree", t2 - t1) ;
MARKTIME(t1) ;
InpMtx_permute(mtxA, IV_entries(oldToNewIV), IV_entries(oldToNewIV)) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : permute mtxA", t2 - t1) ;
if ( symmetryflag == SPOOLES_SYMMETRIC
|| symmetryflag == SPOOLES_HERMITIAN ) {
MARKTIME(t1) ;
InpMtx_mapToUpperTriangle(mtxA) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : map to upper triangle", t2 - t1) ;
}
if ( ! INPMTX_IS_BY_CHEVRONS(mtxA) ) {
MARKTIME(t1) ;
InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : change coordinate type", t2 - t1) ;
}
if ( INPMTX_IS_RAW_DATA(mtxA) ) {
MARKTIME(t1) ;
InpMtx_changeStorageMode(mtxA, INPMTX_SORTED) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : sort entries ", t2 - t1) ;
}
if ( INPMTX_IS_SORTED(mtxA) ) {
MARKTIME(t1) ;
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : convert to vectors ", t2 - t1) ;
}
MARKTIME(t1) ;
symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : symbolic factorization", t2 - t1) ;
MARKTIME(t1) ;
DenseMtx_permuteRows(mtxB, oldToNewIV) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : permute rhs", t2 - t1) ;
/*
------------------------------
initialize the FrontMtx object
------------------------------
*/
MARKTIME(t1) ;
frontmtx = FrontMtx_new() ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, NO_LOCK, 0) ;
FrontMtx_init(frontmtx, frontETree, symbfacIVL,
type, symmetryflag, sparsityflag, pivotingflag,
NO_LOCK, 0, NULL, mtxmanager, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : initialize the front matrix",
t2 - t1) ;
if ( msglvl > 1 ) {
fprintf(msgFile,
"\n nendD = %d, nentL = %d, nentU = %d",
frontmtx->nentD, frontmtx->nentL, frontmtx->nentU) ;
SubMtxManager_writeForHumanEye(mtxmanager, msgFile) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n front matrix initialized") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*
-----------------
factor the matrix
-----------------
*/
nzf = ETree_nFactorEntries(frontETree, symmetryflag) ;
factorops = ETree_nFactorOps(frontETree, type, symmetryflag) ;
fprintf(msgFile,
"\n %d factor entries, %.0f factor ops, %8.3f ratio",
nzf, factorops, factorops/nzf) ;
IVzero(6, stats) ;
DVzero(9, cpus) ;
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, NO_LOCK, 1) ;
MARKTIME(t1) ;
rootchv = FrontMtx_factorInpMtx(frontmtx, mtxA, tau, droptol,
chvmanager, &error, cpus,
stats, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : factor matrix, %8.3f mflops",
t2 - t1, 1.e-6*factorops/(t2-t1)) ;
if ( rootchv != NULL ) {
fprintf(msgFile, "\n\n factorization did not complete") ;
for ( chv = rootchv ; chv != NULL ; chv = chv->next ) {
fprintf(stdout, "\n chv %d, nD = %d, nL = %d, nU = %d",
chv->id, chv->nD, chv->nL, chv->nU) ;
}
}
if ( error >= 0 ) {
fprintf(msgFile, "\n\n error encountered at front %d\n", error) ;
rc=error ;
goto end_front;
}
fprintf(msgFile,
"\n %8d pivots, %8d pivot tests, %8d delayed rows and columns",
stats[0], stats[1], stats[2]) ;
if ( frontmtx->rowadjIVL != NULL ) {
fprintf(msgFile,
"\n %d entries in rowadjIVL", frontmtx->rowadjIVL->tsize) ;
}
if ( frontmtx->coladjIVL != NULL ) {
fprintf(msgFile,
", %d entries in coladjIVL", frontmtx->coladjIVL->tsize) ;
}
if ( frontmtx->upperblockIVL != NULL ) {
fprintf(msgFile,
"\n %d fronts, %d entries in upperblockIVL",
frontmtx->nfront, frontmtx->upperblockIVL->tsize) ;
}
if ( frontmtx->lowerblockIVL != NULL ) {
fprintf(msgFile,
", %d entries in lowerblockIVL",
frontmtx->lowerblockIVL->tsize) ;
}
fprintf(msgFile,
"\n %d entries in D, %d entries in L, %d entries in U",
stats[3], stats[4], stats[5]) ;
fprintf(msgFile, "\n %d locks", frontmtx->nlocks) ;
if ( FRONTMTX_IS_SYMMETRIC(frontmtx)
|| FRONTMTX_IS_HERMITIAN(frontmtx) ) {
int nneg, npos, nzero ;
FrontMtx_inertia(frontmtx, &nneg, &nzero, &npos) ;
fprintf(msgFile,
"\n %d negative, %d zero and %d positive eigenvalues",
nneg, nzero, npos) ;
fflush(msgFile) ;
}
cputotal = cpus[8] ;
if ( cputotal > 0.0 ) {
fprintf(msgFile,
"\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",
cpus[0], 100.*cpus[0]/cputotal,
cpus[1], 100.*cpus[1]/cputotal,
cpus[2], 100.*cpus[2]/cputotal,
cpus[3], 100.*cpus[3]/cputotal,
cpus[4], 100.*cpus[4]/cputotal,
cpus[5], 100.*cpus[5]/cputotal,
cpus[6], 100.*cpus[6]/cputotal,
cpus[7], 100.*cpus[7]/cputotal, cputotal) ;
}
if ( msglvl > 1 ) {
SubMtxManager_writeForHumanEye(mtxmanager, msgFile) ;
ChvManager_writeForHumanEye(chvmanager, msgFile) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front factor matrix") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
}
/*
------------------------------
post-process the factor matrix
------------------------------
*/
MARKTIME(t1) ;
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : post-process the matrix", t2 - t1) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
}
fprintf(msgFile, "\n\n after post-processing") ;
if ( msglvl > 1 ) SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
/*
----------------
solve the system
----------------
*/
neqns = mtxB->nrow ;
mtxZ = DenseMtx_new() ;
DenseMtx_init(mtxZ, type, 0, 0, neqns, nrhs, 1, neqns) ;
zversion=INPMTX_IS_COMPLEX_ENTRIES(mtxA);
for (k=0; k<imethod; k++) {
DenseMtx_zero(mtxZ) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n rhs") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(stdout) ;
}
fprintf(msgFile, "\n\n itemax %d", itermax) ;
DVzero(6, cpus) ;
MARKTIME(t1) ;
switch ( method[k] ) {
case BiCGStabR :
if (zversion)
rc=zbicgstabr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=bicgstabr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case BiCGStabL :
if (zversion)
rc=zbicgstabl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=bicgstabl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case TFQMRR :
if (zversion)
rc=ztfqmrr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=tfqmrr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case TFQMRL :
if (zversion)
rc=ztfqmrl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=tfqmrl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case PCGR :
if (zversion)
rc=zpcgr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=pcgr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case PCGL :
if (zversion)
rc=zpcgl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=pcgl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case MLBiCGStabR :
mtxQ = DenseMtx_new() ;
DenseMtx_init(mtxQ, type, 0, -1, neqns, Ik, 1, neqns) ;
Drand_setUniform(&drand, 0.0, 1.0) ;
DenseMtx_fillRandomEntries(mtxQ, &drand) ;
if (zversion)
rc=zmlbicgstabr(neqns, type, symmetryflag, mtxA, frontmtx, mtxQ, mtxZ,
mtxB, itermax, conv_tol, msglvl, msgFile);
else
rc=mlbicgstabr(neqns, type, symmetryflag, mtxA, frontmtx, mtxQ, mtxZ,
mtxB, itermax, conv_tol, msglvl, msgFile);
DenseMtx_free(mtxQ) ;
break;
case MLBiCGStabL :
mtxQ = DenseMtx_new() ;
DenseMtx_init(mtxQ, type, 0, -1, neqns, Ik, 1, neqns) ;
Drand_setUniform(&drand, 0.0, 1.0) ;
DenseMtx_fillRandomEntries(mtxQ, &drand) ;
if (zversion)
rc=zmlbicgstabl(neqns, type, symmetryflag, mtxA, frontmtx, mtxQ, mtxZ,
mtxB, itermax, conv_tol, msglvl, msgFile);
else
rc=mlbicgstabl(neqns, type, symmetryflag, mtxA, frontmtx, mtxQ, mtxZ,
mtxB, itermax, conv_tol, msglvl, msgFile);
DenseMtx_free(mtxQ) ;
break;
case BGMRESR:
if (zversion)
fprintf(msgFile, "\n\n *** BGMRESR complex version is not available "
"at this moment. ") ;
else
rc=bgmresr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ,
mtxB, iterout, itermax, &nouter, &ninner, conv_tol,
msglvl, msgFile);
break;
case BGMRESL:
if (zversion)
fprintf(msgFile, "\n\n *** BGMRESR complex version is not available "
"at this moment. ") ;
else
rc=bgmresl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ,
mtxB, iterout, itermax, &nouter, &ninner, conv_tol,
msglvl, msgFile);
break;
default:
fprintf(msgFile, "\n\n *** Invalid method number ") ;
}
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : solve the system", t2 - t1) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n computed solution") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
/*
-------------------------------------------------------------
permute the computed solution back into the original ordering
-------------------------------------------------------------
*/
MARKTIME(t1) ;
DenseMtx_permuteRows(mtxZ, newToOldIV) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : permute solution", t2 - t1) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n permuted solution") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
/*
-------------
save solution
-------------
*/
if ( strcmp(slnFileName, "none") != 0 ) {
DenseMtx_writeToFile(mtxZ, slnFileName) ;
}
/*
-----------------
compute the error
-----------------
*/
if (!strcmp(rhsFileName, "none")) {
DenseMtx_sub(mtxZ, mtxX) ;
if (method[k] <8) {
mtxQ = DenseMtx_new() ;
DenseMtx_init(mtxQ, type, 0, -1, neqns, 1, 1, neqns) ;
rc=DenseMtx_initAsSubmatrix (mtxQ, mtxZ, 0, neqns-1, 0, 0);
fprintf(msgFile, "\n\n maxabs error = %12.4e", DenseMtx_maxabs(mtxQ)) ;
DenseMtx_free(mtxQ) ;
}
else
fprintf(msgFile, "\n\n maxabs error = %12.4e", DenseMtx_maxabs(mtxZ)) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n error") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
if ( msglvl > 1 )
SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
}
fprintf(msgFile, "\n--------- End of Method %d -------\n",method[k]) ;
}
/*
------------------------
free the working storage
------------------------
*/
DenseMtx_free(mtxZ) ;
end_front:
ChvManager_free(chvmanager) ;
SubMtxManager_free(mtxmanager) ;
FrontMtx_free(frontmtx) ;
IVL_free(symbfacIVL) ;
IV_free(oldToNewIV) ;
IV_free(newToOldIV) ;
end_tree:
ETree_free(frontETree) ;
end_init:
DenseMtx_free(mtxB) ;
DenseMtx_free(mtxX) ;
end_read:
InpMtx_free(mtxA) ;
fprintf(msgFile, "\n") ;
fclose(msgFile) ;
return(rc) ; }
/*
---------------------------------------------------------------------
purpose -- to compute the factorization of A - sigma * B
note: all variables in the calling sequence are references
to allow call from fortran.
input parameters
data -- pointer to bridge data object
psigma -- shift for the matrix pencil
ppvttol -- pivot tolerance
*ppvttol = 0.0 --> no pivoting used
*ppvttol != 0.0 --> pivoting used, entries in factor are
bounded above by 1/pvttol in magnitude
output parameters
*pinertia -- on return contains the number of negative eigenvalues
*perror -- on return contains an error code
1 -- error found during factorization
0 -- normal return
-1 -- psigma is NULL
-2 -- ppvttol is NULL
-3 -- data is NULL
-4 -- pinertia is NULL
created -- 98aug10, cca & jcp
---------------------------------------------------------------------
*/
void
FactorMPI (
double *psigma,
double *ppvttol,
void *data,
int *pinertia,
int *perror
) {
BridgeMPI *bridge = (BridgeMPI *) data ;
Chv *rootchv ;
ChvManager *chvmanager ;
double droptol=0.0, tau ;
double cpus[20] ;
FILE *msgFile ;
int recvtemp[3], sendtemp[3], stats[20] ;
int msglvl, nnegative, nzero, npositive, pivotingflag, tag ;
MPI_Comm comm ;
int nproc ;
#if MYDEBUG > 0
double t1, t2 ;
count_Factor++ ;
MARKTIME(t1) ;
if ( bridge->myid == 0 ) {
fprintf(stdout, "\n (%d) FactorMPI()", count_Factor) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile, "\n (%d) FactorMPI()", count_Factor) ;
fflush(bridge->msgFile) ;
#endif
nproc = bridge->nproc ;
/*
---------------
check the input
---------------
*/
if ( psigma == NULL ) {
fprintf(stderr, "\n error in FactorMPI()"
"\n psigma is NULL\n") ;
*perror = -1 ; return ;
}
if ( ppvttol == NULL ) {
fprintf(stderr, "\n error in FactorMPI()"
"\n ppvttol is NULL\n") ;
*perror = -2 ; return ;
}
if ( data == NULL ) {
fprintf(stderr, "\n error in FactorMPI()"
"\n data is NULL\n") ;
*perror = -3 ; return ;
}
if ( pinertia == NULL ) {
fprintf(stderr, "\n error in FactorMPI()"
"\n pinertia is NULL\n") ;
*perror = -4 ; return ;
}
if ( perror == NULL ) {
fprintf(stderr, "\n error in FactorMPI()"
"\n perror is NULL\n") ;
return ;
}
comm = bridge->comm ;
msglvl = bridge->msglvl ;
msgFile = bridge->msgFile ;
/*
----------------------------------
set the shift in the pencil object
----------------------------------
*/
bridge->pencil->sigma[0] = -(*psigma) ;
bridge->pencil->sigma[1] = 0.0 ;
/*
-----------------------------------------
if the matrices are in local coordinates
(i.e., this is the first factorization
following a matrix-vector multiply) then
map the matrix into global coordinates
-----------------------------------------
*/
if ( bridge->coordFlag == LOCAL ) {
if ( bridge->prbtype == 1 ) {
MatMul_setGlobalIndices(bridge->info, bridge->B) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n matrix B in local coordinates") ;
InpMtx_writeForHumanEye(bridge->B, msgFile) ;
fflush(msgFile) ;
}
}
if ( bridge->prbtype == 2 ) {
MatMul_setGlobalIndices(bridge->info, bridge->A) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n matrix A in local coordinates") ;
InpMtx_writeForHumanEye(bridge->A, msgFile) ;
fflush(msgFile) ;
}
}
bridge->coordFlag = GLOBAL ;
}
/*
-----------------------------------------------------
clear the front matrix and submatrix mananger objects
-----------------------------------------------------
*/
FrontMtx_clearData(bridge->frontmtx);
SubMtxManager_clearData(bridge->mtxmanager);
SolveMap_clearData(bridge->solvemap) ;
if ( bridge->rowmapIV != NULL ) {
IV_free(bridge->rowmapIV) ;
bridge->rowmapIV = NULL ;
}
/*
-----------------------------------------------------------
set the pivot tolerance.
NOTE: spooles's "tau" parameter is a bound on the magnitude
of the factor entries, and is the recipricol of that of the
pivot tolerance of the lanczos code
-----------------------------------------------------------
*/
if ( *ppvttol == 0.0 ) {
tau = 10.0 ;
pivotingflag = SPOOLES_NO_PIVOTING ;
} else {
tau = (1.0)/(*ppvttol) ;
pivotingflag = SPOOLES_PIVOTING ;
}
/*
----------------------------------
initialize the front matrix object
----------------------------------
*/
FrontMtx_init(bridge->frontmtx, bridge->frontETree, bridge->symbfacIVL,
SPOOLES_REAL, SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS,
pivotingflag, NO_LOCK, bridge->myid, bridge->ownersIV,
bridge->mtxmanager, bridge->msglvl, bridge->msgFile) ;
/*
-------------------------
compute the factorization
-------------------------
*/
tag = 0 ;
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, NO_LOCK, 0);
IVfill(20, stats, 0) ;
DVfill(20, cpus, 0.0) ;
rootchv = FrontMtx_MPI_factorPencil(bridge->frontmtx, bridge->pencil,
tau, droptol, chvmanager, bridge->ownersIV,
0, perror, cpus, stats, bridge->msglvl,
bridge->msgFile, tag, comm) ;
ChvManager_free(chvmanager);
tag += 3*FrontMtx_nfront(bridge->frontmtx) + 2 ;
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n numeric factorization") ;
FrontMtx_writeForHumanEye(bridge->frontmtx, bridge->msgFile) ;
fflush(bridge->msgFile) ;
}
/*
----------------------------
if matrix is singular then
set error flag and return
----------------------------
*/
if ( rootchv != NULL ) {
fprintf(msgFile, "\n WHOA NELLY!, matrix is singular") ;
fflush(msgFile) ;
*perror = 1 ;
return ;
}
/*
------------------------------------------------------------------
post-process the factor matrix, convert from fronts to submatrices
------------------------------------------------------------------
*/
FrontMtx_MPI_postProcess(bridge->frontmtx, bridge->ownersIV, stats,
bridge->msglvl, bridge->msgFile, tag, comm);
tag += 5*bridge->nproc ;
/*
-------------------
compute the inertia
-------------------
*/
FrontMtx_inertia(bridge->frontmtx, &nnegative, &nzero, &npositive) ;
sendtemp[0] = nnegative ;
sendtemp[1] = nzero ;
sendtemp[2] = npositive ;
if ( bridge->msglvl > 2 && bridge->msgFile != NULL ) {
fprintf(bridge->msgFile, "\n local inertia = < %d, %d, %d >",
nnegative, nzero, npositive) ;
fflush(bridge->msgFile) ;
}
MPI_Allreduce((void *) sendtemp, (void *) recvtemp, 3, MPI_INT,
MPI_SUM, comm) ;
nnegative = recvtemp[0] ;
nzero = recvtemp[1] ;
npositive = recvtemp[2] ;
if ( bridge->msglvl > 2 && bridge->msgFile != NULL ) {
fprintf(bridge->msgFile, "\n global inertia = < %d, %d, %d >",
nnegative, nzero, npositive) ;
fflush(bridge->msgFile) ;
}
*pinertia = nnegative;
/*
---------------------------
create the solve map object
---------------------------
*/
SolveMap_ddMap(bridge->solvemap, SPOOLES_REAL,
FrontMtx_upperBlockIVL(bridge->frontmtx),
FrontMtx_lowerBlockIVL(bridge->frontmtx), nproc,
bridge->ownersIV, FrontMtx_frontTree(bridge->frontmtx),
bridge->seed, bridge->msglvl, bridge->msgFile) ;
/*
-------------------------------
redistribute the front matrices
-------------------------------
*/
FrontMtx_MPI_split(bridge->frontmtx, bridge->solvemap, stats,
bridge->msglvl, bridge->msgFile, tag, comm) ;
if ( *ppvttol != 0.0 ) {
/*
-------------------------------------------------------------
pivoting for stability may have taken place. create rowmapIV,
the map from rows in the factorization to processes.
-------------------------------------------------------------
*/
bridge->rowmapIV = FrontMtx_MPI_rowmapIV(bridge->frontmtx,
bridge->ownersIV, bridge->msglvl,
bridge->msgFile, bridge->comm) ;
if ( bridge->msglvl > 2 && bridge->msgFile != NULL ) {
fprintf(bridge->msgFile, "\n\n bridge->rowmapIV") ;
IV_writeForHumanEye(bridge->rowmapIV, bridge->msgFile) ;
fflush(bridge->msgFile) ;
}
} else {
bridge->rowmapIV = NULL ;
}
/*
------------------------------------------------------------------
set the error. (this is simple since when the spooles codes detect
a fatal error, they print out a message to stderr and exit.)
------------------------------------------------------------------
*/
*perror = 0 ;
#if MYDEBUG > 0
MARKTIME(t2) ;
time_Factor += t2 - t1 ;
if ( bridge->myid == 0 ) {
fprintf(stdout, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_Factor) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_Factor) ;
fflush(bridge->msgFile) ;
#endif
return; }
/*--------------------------------------------------------------------*/
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 -- to permute (if necessary) the original matrix,
and to initialize, factor and postprocess the factor matrix
return value ---
1 -- normal return, factorization complete
0 -- factorization did not complete, see error flag
-1 -- bridge is NULL
-2 -- mtxA is NULL
-3 -- perror is NULL
created -- 98sep18, cca
--------------------------------------------------------------
*/
int
BridgeMT_factor (
BridgeMT *bridge,
InpMtx *mtxA,
int permuteflag,
int *perror
) {
Chv *rootchv ;
ChvManager *chvmanager ;
double cputotal, nfops, t0, t1, t2 ;
double cpus[11] ;
int msglvl, nzf ;
int stats[16] ;
FILE *msgFile ;
FrontMtx *frontmtx ;
SubMtxManager *mtxmanager ;
/*--------------------------------------------------------------------*/
MARKTIME(t0) ;
/*
---------------
check the input
---------------
*/
if ( bridge == NULL ) {
fprintf(stderr, "\n error in BridgeMT_factor()"
"\n bridge is NULL\n") ;
return(-1) ;
}
if ( mtxA == NULL ) {
fprintf(stderr, "\n error in BridgeMT_factor()"
"\n mtxA is NULL\n") ;
return(-2) ;
}
if ( perror == NULL ) {
fprintf(stderr, "\n error in BridgeMT_factor()"
"\n perror is NULL\n") ;
return(-3) ;
}
msglvl = bridge->msglvl ;
msgFile = bridge->msgFile ;
/*--------------------------------------------------------------------*/
MARKTIME(t1) ;
if ( permuteflag == 1 ) {
int *oldToNew = IV_entries(bridge->oldToNewIV) ;
/*
------------------------------------------------
permute the input matrix and convert to chevrons
------------------------------------------------
*/
InpMtx_permute(mtxA, oldToNew, oldToNew) ;
if ( bridge->symmetryflag == SPOOLES_SYMMETRIC
|| bridge->symmetryflag == SPOOLES_HERMITIAN ) {
InpMtx_mapToUpperTriangle(mtxA) ;
}
}
if ( ! INPMTX_IS_BY_CHEVRONS(mtxA) ) {
InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS) ;
}
if ( ! INPMTX_IS_BY_VECTORS(mtxA) ) {
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
}
MARKTIME(t2) ;
bridge->cpus[6] += t2 - t1 ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n CPU %8.3f : permute and format A", t2 - t1) ;
fflush(msgFile) ;
}
/*
---------------------------
initialize the front matrix
---------------------------
*/
MARKTIME(t1) ;
if ( (mtxmanager = bridge->mtxmanager) == NULL ) {
mtxmanager = bridge->mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, LOCK_IN_PROCESS, 0) ;
}
if ( (frontmtx = bridge->frontmtx) == NULL ) {
frontmtx = bridge->frontmtx = FrontMtx_new() ;
} else {
FrontMtx_clearData(frontmtx) ;
}
FrontMtx_init(frontmtx, bridge->frontETree, bridge->symbfacIVL,
bridge->type, bridge->symmetryflag, bridge->sparsityflag,
bridge->pivotingflag, LOCK_IN_PROCESS, 0, NULL,
mtxmanager, msglvl, msgFile) ;
frontmtx->patchinfo = bridge->patchinfo ;
MARKTIME(t2) ;
bridge->cpus[7] += t2 - t1 ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n CPU %8.3f : initialize front matrix", t2 - t1) ;
fflush(msgFile) ;
}
/*
-----------------
factor the matrix
-----------------
*/
nzf = ETree_nFactorEntries(bridge->frontETree, bridge->symmetryflag) ;
nfops = ETree_nFactorOps(bridge->frontETree,
bridge->type, bridge->symmetryflag) ;
if ( msglvl > 1 ) {
fprintf(msgFile,
"\n %d factor entries, %.0f factor ops, %8.3f ratio",
nzf, nfops, nfops/nzf) ;
fflush(msgFile) ;
}
IVzero(16, stats) ;
DVzero(11, cpus) ;
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, LOCK_IN_PROCESS, 1) ;
MARKTIME(t1) ;
rootchv = FrontMtx_MT_factorInpMtx(frontmtx, mtxA, bridge->tau,
bridge->droptol, chvmanager, bridge->ownersIV,
bridge->lookahead, perror, cpus, stats, msglvl, msgFile) ;
MARKTIME(t2) ;
IVcopy(6, bridge->stats, stats) ;
bridge->cpus[8] += t2 - t1 ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n CPU %8.3f : factor matrix, %8.3f mflops",
t2 - t1, 1.e-6*nfops/(t2-t1)) ;
fprintf(msgFile,
"\n %8d pivots, %8d pivot tests, %8d delayed vertices"
"\n %d entries in D, %d entries in L, %d entries in U",
stats[0], stats[1], stats[2], stats[3], stats[4], stats[5]) ;
cputotal = cpus[8] ;
if ( cputotal > 0.0 ) {
fprintf(msgFile,
"\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",
cpus[0], 100.*cpus[0]/cputotal,
cpus[1], 100.*cpus[1]/cputotal,
cpus[2], 100.*cpus[2]/cputotal,
cpus[3], 100.*cpus[3]/cputotal,
cpus[4], 100.*cpus[4]/cputotal,
cpus[5], 100.*cpus[5]/cputotal,
cpus[6], 100.*cpus[6]/cputotal,
cpus[7], 100.*cpus[7]/cputotal, cputotal) ;
}
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n submatrix mananger after factorization") ;
SubMtxManager_writeForHumanEye(mtxmanager, msgFile) ;
fprintf(msgFile, "\n\n chevron mananger after factorization") ;
ChvManager_writeForHumanEye(chvmanager, msgFile) ;
fflush(msgFile) ;
}
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n front factor matrix") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
ChvManager_free(chvmanager) ;
if ( *perror >= 0 ) {
return(0) ;
}
/*
-----------------------------
post-process the front matrix
-----------------------------
*/
MARKTIME(t1) ;
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
MARKTIME(t2) ;
bridge->cpus[9] += t2 - t1 ;
if ( msglvl > 1 ) {
fprintf(msgFile,
"\n\n CPU %8.3f : post-process the matrix", t2 - t1) ;
fflush(msgFile) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n submatrix mananger after post-processing") ;
SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
fflush(msgFile) ;
}
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n front factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
MARKTIME(t2) ;
bridge->cpus[10] += t2 - t0 ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n CPU %8.3f : total factor time", t2 - t0) ;
fflush(msgFile) ;
}
return(1) ; }
//static void factor_MT(struct factorinfo *pfi, InpMtx *mtxA, int size, FILE *msgFile, int symmetryflag)
void factor_MT(struct factorinfo *pfi, InpMtx *mtxA, int size, FILE *msgFile, int symmetryflag)
{
Graph *graph;
IV *ownersIV;
IVL *symbfacIVL;
Chv *rootchv;
/* Initialize pfi: */
pfi->size = size;
pfi->msgFile = msgFile;
DVfill(10, pfi->cpus, 0.0);
/*
* STEP 1 : find a low-fill ordering
* (1) create the Graph object
*/
ssolve_creategraph(&graph, &pfi->frontETree, mtxA, size, msgFile);
/*
* STEP 2: get the permutation, permute the matrix and
* front tree and get the symbolic factorization
*/
ssolve_permuteA(&pfi->oldToNewIV, &pfi->newToOldIV, &symbfacIVL, pfi->frontETree,
mtxA, msgFile, symmetryflag);
/*
* STEP 3: Prepare distribution to multiple threads/cpus
*/
{
DV *cumopsDV;
int nfront;
nfront = ETree_nfront(pfi->frontETree);
pfi->nthread = num_cpus;
if (pfi->nthread > nfront)
pfi->nthread = nfront;
cumopsDV = DV_new();
DV_init(cumopsDV, pfi->nthread, NULL);
ownersIV = ETree_ddMap(pfi->frontETree, SPOOLES_REAL, symmetryflag,
cumopsDV, 1. / (2. * pfi->nthread));
if (DEBUG_LVL > 1) {
fprintf(msgFile,
"\n\n map from fronts to threads");
IV_writeForHumanEye(ownersIV, msgFile);
fprintf(msgFile,
"\n\n factor operations for each front");
DV_writeForHumanEye(cumopsDV, msgFile);
fflush(msgFile);
} else {
fprintf(msgFile, "\n\n Using %d threads\n",
pfi->nthread);
}
DV_free(cumopsDV);
}
/*
* STEP 4: initialize the front matrix object
*/
{
pfi->frontmtx = FrontMtx_new();
pfi->mtxmanager = SubMtxManager_new();
SubMtxManager_init(pfi->mtxmanager, LOCK_IN_PROCESS, 0);
FrontMtx_init(pfi->frontmtx, pfi->frontETree, symbfacIVL, SPOOLES_REAL,
symmetryflag, FRONTMTX_DENSE_FRONTS,
SPOOLES_PIVOTING, LOCK_IN_PROCESS, 0, NULL,
pfi->mtxmanager, DEBUG_LVL, pfi->msgFile);
}
/*
* STEP 5: compute the numeric factorization in parallel
*/
{
ChvManager *chvmanager;
int stats[20];
int error;
chvmanager = ChvManager_new();
ChvManager_init(chvmanager, LOCK_IN_PROCESS, 1);
IVfill(20, stats, 0);
rootchv = FrontMtx_MT_factorInpMtx(pfi->frontmtx, mtxA, MAGIC_TAU, MAGIC_DTOL,
chvmanager, ownersIV, 0,
&error, pfi->cpus, stats, DEBUG_LVL,
pfi->msgFile);
ChvManager_free(chvmanager);
if (DEBUG_LVL > 1) {
fprintf(msgFile, "\n\n factor matrix");
FrontMtx_writeForHumanEye(pfi->frontmtx, pfi->msgFile);
fflush(pfi->msgFile);
}
if (rootchv != NULL) {
fprintf(pfi->msgFile, "\n\n matrix found to be singular\n");
exit(-1);
}
if (error >= 0) {
fprintf(pfi->msgFile, "\n\n fatal error at front %d", error);
exit(-1);
}
}
/*
* STEP 6: post-process the factorization
*/
ssolve_postfactor(pfi->frontmtx, pfi->msgFile);
/*
* STEP 7: get the solve map object for the parallel solve
*/
{
pfi->solvemap = SolveMap_new();
SolveMap_ddMap(pfi->solvemap, symmetryflag,
FrontMtx_upperBlockIVL(pfi->frontmtx),
FrontMtx_lowerBlockIVL(pfi->frontmtx), pfi->nthread, ownersIV,
FrontMtx_frontTree(pfi->frontmtx), RNDSEED, DEBUG_LVL,
pfi->msgFile);
}
/* cleanup: */
InpMtx_free(mtxA);
IVL_free(symbfacIVL);
Graph_free(graph);
IV_free(ownersIV);
}
void factor(struct factorinfo *pfi, InpMtx *mtxA, int size, FILE *msgFile,
int symmetryflag)
{
Graph *graph;
IVL *symbfacIVL;
Chv *rootchv;
/* Initialize pfi: */
pfi->size = size;
pfi->msgFile = msgFile;
pfi->solvemap = NULL;
DVfill(10, pfi->cpus, 0.0);
/*
* STEP 1 : find a low-fill ordering
* (1) create the Graph object
*/
ssolve_creategraph(&graph, &pfi->frontETree, mtxA, size, pfi->msgFile);
/*
* STEP 2: get the permutation, permute the matrix and
* front tree and get the symbolic factorization
*/
ssolve_permuteA(&pfi->oldToNewIV, &pfi->newToOldIV, &symbfacIVL, pfi->frontETree,
mtxA, pfi->msgFile, symmetryflag);
/*
* STEP 3: initialize the front matrix object
*/
{
pfi->frontmtx = FrontMtx_new();
pfi->mtxmanager = SubMtxManager_new();
SubMtxManager_init(pfi->mtxmanager, NO_LOCK, 0);
FrontMtx_init(pfi->frontmtx, pfi->frontETree, symbfacIVL, SPOOLES_REAL,
symmetryflag, FRONTMTX_DENSE_FRONTS,
SPOOLES_PIVOTING, NO_LOCK, 0, NULL,
pfi->mtxmanager, DEBUG_LVL, pfi->msgFile);
}
/*
* STEP 4: compute the numeric factorization
*/
{
ChvManager *chvmanager;
int stats[20];
int error;
chvmanager = ChvManager_new();
ChvManager_init(chvmanager, NO_LOCK, 1);
IVfill(20, stats, 0);
rootchv = FrontMtx_factorInpMtx(pfi->frontmtx, mtxA, MAGIC_TAU, MAGIC_DTOL,
chvmanager, &error, pfi->cpus,
stats, DEBUG_LVL, pfi->msgFile);
ChvManager_free(chvmanager);
if (DEBUG_LVL > 1) {
fprintf(msgFile, "\n\n factor matrix");
FrontMtx_writeForHumanEye(pfi->frontmtx, pfi->msgFile);
fflush(msgFile);
}
if (rootchv != NULL) {
fprintf(pfi->msgFile, "\n\n matrix found to be singular\n");
exit(-1);
}
if (error >= 0) {
fprintf(pfi->msgFile, "\n\nerror encountered at front %d",
error);
exit(-1);
}
}
/*
* STEP 5: post-process the factorization
*/
ssolve_postfactor(pfi->frontmtx, pfi->msgFile);
/* cleanup: */
IVL_free(symbfacIVL);
InpMtx_free(mtxA);
Graph_free(graph);
}
/*
---------------------------------------------------------------------
purpose -- to compute the factorization of A - sigma * B
note: all variables in the calling sequence are references
to allow call from fortran.
input parameters
data -- pointer to bridge data object
psigma -- shift for the matrix pencil
ppvttol -- pivot tolerance
*ppvttol = 0.0 --> no pivoting used
*ppvttol != 0.0 --> pivoting used, entries in factor are
bounded above by 1/pvttol in magnitude
output parameters
*pinertia -- on return contains the number of negative eigenvalues
*perror -- on return contains an error code
1 -- error found during factorization
0 -- normal return
-1 -- psigma is NULL
-2 -- ppvttol is NULL
-3 -- data is NULL
-4 -- pinertia is NULL
created -- 98aug10, cca & jcp
---------------------------------------------------------------------
*/
void
Factor (
double *psigma,
double *ppvttol,
void *data,
int *pinertia,
int *perror
) {
Bridge *bridge = (Bridge *) data ;
Chv *rootchv ;
ChvManager *chvmanager ;
double droptol=0.0, tau ;
double cpus[10] ;
int stats[20] ;
int nnegative, nzero, npositive, pivotingflag ;
#if MYDEBUG > 0
double t1, t2 ;
MARKTIME(t1) ;
count_Factor++ ;
fprintf(stdout, "\n (%d) Factor()", count_Factor) ;
fflush(stdout) ;
#endif
/*
---------------
check the input
---------------
*/
if ( psigma == NULL ) {
fprintf(stderr, "\n error in Factor()"
"\n psigma is NULL\n") ;
*perror = -1 ; return ;
}
if ( ppvttol == NULL ) {
fprintf(stderr, "\n error in Factor()"
"\n ppvttol is NULL\n") ;
*perror = -2 ; return ;
}
if ( data == NULL ) {
fprintf(stderr, "\n error in Factor()"
"\n data is NULL\n") ;
*perror = -3 ; return ;
}
if ( pinertia == NULL ) {
fprintf(stderr, "\n error in Factor()"
"\n pinertia is NULL\n") ;
*perror = -4 ; return ;
}
if ( perror == NULL ) {
fprintf(stderr, "\n error in Factor()"
"\n perror is NULL\n") ;
return ;
}
/*
----------------------------------
set the shift in the pencil object
----------------------------------
*/
bridge->pencil->sigma[0] = -(*psigma) ;
bridge->pencil->sigma[1] = 0.0 ;
/*
-----------------------------------------------------
clear the front matrix and submatrix mananger objects
-----------------------------------------------------
*/
FrontMtx_clearData(bridge->frontmtx);
SubMtxManager_clearData(bridge->mtxmanager);
/*
-----------------------------------------------------------
set the pivot tolerance.
NOTE: spooles's "tau" parameter is a bound on the magnitude
of the factor entries, and is the recipricol of that of the
pivot tolerance of the lanczos code
-----------------------------------------------------------
*/
if ( *ppvttol == 0.0 ) {
tau = 10.0 ;
pivotingflag = SPOOLES_NO_PIVOTING ;
} else {
tau = (1.0)/(*ppvttol) ;
pivotingflag = SPOOLES_PIVOTING ;
}
/*
----------------------------------
initialize the front matrix object
----------------------------------
*/
FrontMtx_init(bridge->frontmtx, bridge->frontETree, bridge->symbfacIVL,
SPOOLES_REAL, SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS,
pivotingflag, NO_LOCK, 0, NULL, bridge->mtxmanager,
bridge->msglvl, bridge->msgFile) ;
/*
-------------------------
compute the factorization
-------------------------
*/
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, NO_LOCK, 1);
IVfill(20, stats, 0) ;
DVfill(10, cpus, 0.0) ;
rootchv = FrontMtx_factorPencil(bridge->frontmtx, bridge->pencil, tau,
droptol, chvmanager, perror, cpus,
stats, bridge->msglvl, bridge->msgFile);
ChvManager_free(chvmanager);
/*
----------------------------
if matrix is singular then
set error flag and return
----------------------------
*/
if ( rootchv != NULL ) {
*perror = 1 ;
return ;
}
/*
------------------------------------------------------------------
post-process the factor matrix, convert from fronts to submatrices
------------------------------------------------------------------
*/
FrontMtx_postProcess(bridge->frontmtx, bridge->msglvl, bridge->msgFile);
/*
-------------------
compute the inertia
-------------------
*/
FrontMtx_inertia(bridge->frontmtx, &nnegative, &nzero, &npositive) ;
*pinertia = nnegative;
/*
------------------------------------------------------------------
set the error. (this is simple since when the spooles codes detect
a fatal error, they print out a message to stderr and exit.)
------------------------------------------------------------------
*/
*perror = 0 ;
#if MYDEBUG > 0
MARKTIME(t2) ;
time_Factor += t2 - t1 ;
fprintf(stdout, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_Factor) ;
fflush(stdout) ;
#endif
return ; }
