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[] ) {
/*
--------------------------------------------------
QR all-in-one program
(1) read in matrix entries and form InpMtx object
of A and A^TA
(2) form Graph object of A^TA
(3) order matrix and form front tree
(4) get the permutation, permute the matrix and
front tree and get the symbolic factorization
(5) compute the numeric factorization
(6) read in right hand side entries
(7) compute the solution
created -- 98jun11, cca
--------------------------------------------------
*/
/*--------------------------------------------------------------------*/
char *matrixFileName, *rhsFileName ;
ChvManager *chvmanager ;
DenseMtx *mtxB, *mtxX ;
double facops, imag, real, value ;
double cpus[10] ;
ETree *frontETree ;
FILE *inputFile, *msgFile ;
FrontMtx *frontmtx ;
Graph *graph ;
int ient, irow, jcol, jrhs, jrow, msglvl, neqns,
nedges, nent, nrhs, nrow, seed, type ;
InpMtx *mtxA ;
IV *newToOldIV, *oldToNewIV ;
IVL *adjIVL, *symbfacIVL ;
SubMtxManager *mtxmanager ;
/*--------------------------------------------------------------------*/
/*
--------------------
get input parameters
--------------------
*/
if ( argc != 7 ) {
fprintf(stdout,
"\n usage: %s msglvl msgFile type matrixFileName rhsFileName seed"
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n type -- type of entries"
"\n 1 (SPOOLES_REAL) -- real entries"
"\n 2 (SPOOLES_COMPLEX) -- complex entries"
"\n matrixFileName -- matrix file name, format"
"\n nrow ncol nent"
"\n irow jcol entry"
"\n ..."
"\n note: indices are zero based"
"\n rhsFileName -- right hand side file name, format"
"\n nrow "
"\n entry[0]"
"\n ..."
"\n entry[nrow-1]"
"\n seed -- random number seed, used for ordering"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
type = atoi(argv[3]) ;
matrixFileName = argv[4] ;
rhsFileName = argv[5] ;
seed = atoi(argv[6]) ;
/*--------------------------------------------------------------------*/
/*
--------------------------------------------
STEP 1: read the entries from the input file
and create the InpMtx object of A
--------------------------------------------
*/
inputFile = fopen(matrixFileName, "r") ;
fscanf(inputFile, "%d %d %d", &nrow, &neqns, &nent) ;
mtxA = InpMtx_new() ;
InpMtx_init(mtxA, INPMTX_BY_ROWS, type, nent, 0) ;
if ( type == SPOOLES_REAL ) {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le", &irow, &jcol, &value) ;
InpMtx_inputRealEntry(mtxA, irow, jcol, value) ;
}
} else {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le %le", &irow, &jcol, &real, &imag) ;
InpMtx_inputComplexEntry(mtxA, irow, jcol, real, imag) ;
}
}
fclose(inputFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n input matrix") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
----------------------------------------
STEP 2: read the right hand side entries
----------------------------------------
*/
inputFile = fopen(rhsFileName, "r") ;
fscanf(inputFile, "%d %d", &nrow, &nrhs) ;
mtxB = DenseMtx_new() ;
DenseMtx_init(mtxB, type, 0, 0, nrow, nrhs, 1, nrow) ;
DenseMtx_zero(mtxB) ;
if ( type == SPOOLES_REAL ) {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le", &value) ;
DenseMtx_setRealEntry(mtxB, jrow, jrhs, value) ;
}
}
} else {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le %le", &real, &imag) ;
DenseMtx_setComplexEntry(mtxB, jrow, jrhs, real, imag) ;
}
}
}
fclose(inputFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n rhs matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------------------------
STEP 3 : find a low-fill ordering
(1) create the Graph object for A^TA or A^HA
(2) order the graph using multiple minimum degree
-------------------------------------------------
*/
graph = Graph_new() ;
adjIVL = InpMtx_adjForATA(mtxA) ;
nedges = IVL_tsize(adjIVL) ;
Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL,
NULL, NULL) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n graph of A^T A") ;
Graph_writeForHumanEye(graph, msgFile) ;
fflush(msgFile) ;
}
frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n front tree from ordering") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------------------
STEP 4: get the permutation, permute the matrix and
front tree and get the symbolic factorization
-----------------------------------------------------
*/
oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ;
newToOldIV = ETree_newToOldVtxPerm(frontETree) ;
InpMtx_permute(mtxA, NULL, IV_entries(oldToNewIV)) ;
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
symbfacIVL = SymbFac_initFromGraph(frontETree, graph) ;
IVL_overwrite(symbfacIVL, oldToNewIV) ;
IVL_sortUp(symbfacIVL) ;
ETree_permuteVertices(frontETree, oldToNewIV) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n old-to-new permutation vector") ;
IV_writeForHumanEye(oldToNewIV, msgFile) ;
fprintf(msgFile, "\n\n new-to-old permutation vector") ;
IV_writeForHumanEye(newToOldIV, msgFile) ;
fprintf(msgFile, "\n\n front tree after permutation") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fprintf(msgFile, "\n\n input matrix after permutation") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fprintf(msgFile, "\n\n symbolic factorization") ;
IVL_writeForHumanEye(symbfacIVL, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------------------------
STEP 5: initialize the front matrix object
------------------------------------------
*/
frontmtx = FrontMtx_new() ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, NO_LOCK, 0) ;
if ( type == SPOOLES_REAL ) {
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type,
SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS,
SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL,
mtxmanager, msglvl, msgFile) ;
} else {
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type,
SPOOLES_HERMITIAN, FRONTMTX_DENSE_FRONTS,
SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL,
mtxmanager, msglvl, msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------
STEP 6: compute the numeric factorization
-----------------------------------------
*/
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, NO_LOCK, 1) ;
DVzero(10, cpus) ;
facops = 0.0 ;
FrontMtx_QR_factor(frontmtx, mtxA, chvmanager,
cpus, &facops, msglvl, msgFile) ;
ChvManager_free(chvmanager) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix") ;
fprintf(msgFile, "\n facops = %9.2f", facops) ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
--------------------------------------
STEP 7: post-process the factorization
--------------------------------------
*/
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------
STEP 8: solve the linear system
-------------------------------
*/
mtxX = DenseMtx_new() ;
DenseMtx_init(mtxX, type, 0, 0, neqns, nrhs, 1, neqns) ;
FrontMtx_QR_solve(frontmtx, mtxA, mtxX, mtxB, mtxmanager,
cpus, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n solution matrix in new ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------------------------------
STEP 9: permute the solution into the original ordering
-------------------------------------------------------
*/
DenseMtx_permuteRows(mtxX, newToOldIV) ;
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n solution matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------
free the working storage
------------------------
*/
InpMtx_free(mtxA) ;
FrontMtx_free(frontmtx) ;
Graph_free(graph) ;
DenseMtx_free(mtxX) ;
DenseMtx_free(mtxB) ;
ETree_free(frontETree) ;
IV_free(newToOldIV) ;
IV_free(oldToNewIV) ;
IVL_free(symbfacIVL) ;
SubMtxManager_free(mtxmanager) ;
/*--------------------------------------------------------------------*/
return(1) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
------------------------------------------------------------------
generate a random matrix and test a matrix-matrix multiply method.
the output is a matlab file to test correctness.
created -- 98jan29, cca
--------------------------------------------------------------------
*/
{
DenseMtx *X, *Y, *Y2 ;
double alpha[2] ;
double alphaImag, alphaReal, t1, t2 ;
double *zvec ;
Drand *drand ;
int col, dataType, ii, msglvl, ncolA, nitem, nops, nrhs,
nrowA, nrowX, nrowY, nthread, row, seed,
storageMode, symflag, transposeflag ;
int *colids, *rowids ;
InpMtx *A ;
FILE *msgFile ;
if ( argc != 15 ) {
fprintf(stdout,
"\n\n %% usage : %s msglvl msgFile symflag storageMode "
"\n %% nrow ncol nent nrhs seed alphaReal alphaImag nthread"
"\n %% msglvl -- message level"
"\n %% msgFile -- message file"
"\n %% dataType -- type of matrix entries"
"\n %% 1 -- real"
"\n %% 2 -- complex"
"\n %% symflag -- symmetry flag"
"\n %% 0 -- symmetric"
"\n %% 1 -- hermitian"
"\n %% 2 -- nonsymmetric"
"\n %% storageMode -- storage mode"
"\n %% 1 -- by rows"
"\n %% 2 -- by columns"
"\n %% 3 -- by chevrons, (requires nrow = ncol)"
"\n %% transpose -- transpose flag"
"\n %% 0 -- Y := Y + alpha * A * X"
"\n %% 1 -- Y := Y + alpha * A^H * X, nonsymmetric only"
"\n %% 2 -- Y := Y + alpha * A^T * X, nonsymmetric only"
"\n %% nrowA -- number of rows in A"
"\n %% ncolA -- number of columns in A"
"\n %% nitem -- number of items"
"\n %% nrhs -- number of right hand sides"
"\n %% seed -- random number seed"
"\n %% alphaReal -- y := y + alpha*A*x"
"\n %% alphaImag -- y := y + alpha*A*x"
"\n %% nthread -- # 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) ;
}
dataType = atoi(argv[3]) ;
symflag = atoi(argv[4]) ;
storageMode = atoi(argv[5]) ;
transposeflag = atoi(argv[6]) ;
nrowA = atoi(argv[7]) ;
ncolA = atoi(argv[8]) ;
nitem = atoi(argv[9]) ;
nrhs = atoi(argv[10]) ;
seed = atoi(argv[11]) ;
alphaReal = atof(argv[12]) ;
alphaImag = atof(argv[13]) ;
nthread = atoi(argv[14]) ;
fprintf(msgFile,
"\n %% %s "
"\n %% msglvl -- %d"
"\n %% msgFile -- %s"
"\n %% dataType -- %d"
"\n %% symflag -- %d"
"\n %% storageMode -- %d"
"\n %% transposeflag -- %d"
"\n %% nrowA -- %d"
"\n %% ncolA -- %d"
"\n %% nitem -- %d"
"\n %% nrhs -- %d"
"\n %% seed -- %d"
"\n %% alphaReal -- %e"
"\n %% alphaImag -- %e"
"\n %% nthread -- %d"
"\n",
argv[0], msglvl, argv[2], dataType, symflag, storageMode,
transposeflag, nrowA, ncolA, nitem, nrhs, seed,
alphaReal, alphaImag, nthread) ;
fflush(msgFile) ;
if ( dataType != 1 && dataType != 2 ) {
fprintf(stderr, "\n invalid value %d for dataType\n", dataType) ;
spoolesFatal();
}
if ( symflag != 0 && symflag != 1 && symflag != 2 ) {
fprintf(stderr, "\n invalid value %d for symflag\n", symflag) ;
spoolesFatal();
}
if ( storageMode != 1 && storageMode != 2 && storageMode != 3 ) {
fprintf(stderr,
"\n invalid value %d for storageMode\n", storageMode) ;
spoolesFatal();
}
if ( transposeflag < 0
|| transposeflag > 2 ) {
fprintf(stderr, "\n error, transposeflag = %d, must be 0, 1 or 2",
transposeflag) ;
spoolesFatal();
}
if ( (transposeflag == 1 && symflag != 2)
|| (transposeflag == 2 && symflag != 2) ) {
fprintf(stderr, "\n error, transposeflag = %d, symflag = %d",
transposeflag, symflag) ;
spoolesFatal();
}
if ( transposeflag == 1 && dataType != 2 ) {
fprintf(stderr, "\n error, transposeflag = %d, dataType = %d",
transposeflag, dataType) ;
spoolesFatal();
}
if ( symflag == 1 && dataType != 2 ) {
fprintf(stderr,
"\n symflag = 1 (hermitian), dataType != 2 (complex)") ;
spoolesFatal();
}
if ( nrowA <= 0 || ncolA <= 0 || nitem <= 0 ) {
fprintf(stderr,
"\n invalid value: nrow = %d, ncol = %d, nitem = %d",
nrowA, ncolA, nitem) ;
spoolesFatal();
}
if ( symflag < 2 && nrowA != ncolA ) {
fprintf(stderr,
"\n invalid data: symflag = %d, nrow = %d, ncol = %d",
symflag, nrowA, ncolA) ;
spoolesFatal();
}
alpha[0] = alphaReal ;
alpha[1] = alphaImag ;
/*
----------------------------
initialize the matrix object
----------------------------
*/
A = InpMtx_new() ;
InpMtx_init(A, storageMode, dataType, 0, 0) ;
drand = Drand_new() ;
/*
----------------------------------
generate a vector of nitem triples
----------------------------------
*/
rowids = IVinit(nitem, -1) ;
Drand_setUniform(drand, 0, nrowA) ;
Drand_fillIvector(drand, nitem, rowids) ;
colids = IVinit(nitem, -1) ;
Drand_setUniform(drand, 0, ncolA) ;
Drand_fillIvector(drand, nitem, colids) ;
Drand_setUniform(drand, 0.0, 1.0) ;
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
zvec = DVinit(nitem, 0.0) ;
Drand_fillDvector(drand, nitem, zvec) ;
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
zvec = ZVinit(nitem, 0.0, 0.0) ;
Drand_fillDvector(drand, 2*nitem, zvec) ;
}
/*
-----------------------------------
assemble the entries entry by entry
-----------------------------------
*/
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n A = zeros(%d,%d) ;", nrowA, ncolA) ;
}
if ( symflag == 1 ) {
/*
----------------
hermitian matrix
----------------
*/
for ( ii = 0 ; ii < nitem ; ii++ ) {
if ( rowids[ii] == colids[ii] ) {
zvec[2*ii+1] = 0.0 ;
}
if ( rowids[ii] <= colids[ii] ) {
row = rowids[ii] ; col = colids[ii] ;
} else {
row = colids[ii] ; col = rowids[ii] ;
}
InpMtx_inputComplexEntry(A, row, col, zvec[2*ii], zvec[2*ii+1]) ;
}
} else if ( symflag == 0 ) {
/*
----------------
symmetric matrix
----------------
*/
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
for ( ii = 0 ; ii < nitem ; ii++ ) {
if ( rowids[ii] <= colids[ii] ) {
row = rowids[ii] ; col = colids[ii] ;
} else {
row = colids[ii] ; col = rowids[ii] ;
}
InpMtx_inputRealEntry(A, row, col, zvec[ii]) ;
}
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
for ( ii = 0 ; ii < nitem ; ii++ ) {
if ( rowids[ii] <= colids[ii] ) {
row = rowids[ii] ; col = colids[ii] ;
} else {
row = colids[ii] ; col = rowids[ii] ;
}
InpMtx_inputComplexEntry(A, row, col,
zvec[2*ii], zvec[2*ii+1]) ;
}
}
} else {
/*
-------------------
nonsymmetric matrix
-------------------
*/
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
for ( ii = 0 ; ii < nitem ; ii++ ) {
InpMtx_inputRealEntry(A, rowids[ii], colids[ii], zvec[ii]) ;
}
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
for ( ii = 0 ; ii < nitem ; ii++ ) {
InpMtx_inputComplexEntry(A, rowids[ii], colids[ii],
zvec[2*ii], zvec[2*ii+1]) ;
}
}
}
InpMtx_changeStorageMode(A, INPMTX_BY_VECTORS) ;
DVfree(zvec) ;
if ( symflag == 0 || symflag == 1 ) {
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
nops = 4*A->nent*nrhs ;
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
nops = 16*A->nent*nrhs ;
}
} else {
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
nops = 2*A->nent*nrhs ;
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
nops = 8*A->nent*nrhs ;
}
}
if ( msglvl > 1 ) {
/*
-------------------------------------------
write the assembled matrix to a matlab file
-------------------------------------------
*/
InpMtx_writeForMatlab(A, "A", msgFile) ;
if ( symflag == 0 ) {
fprintf(msgFile,
"\n for k = 1:%d"
"\n for j = k+1:%d"
"\n A(j,k) = A(k,j) ;"
"\n end"
"\n end", nrowA, ncolA) ;
} else if ( symflag == 1 ) {
fprintf(msgFile,
"\n for k = 1:%d"
"\n for j = k+1:%d"
"\n A(j,k) = ctranspose(A(k,j)) ;"
"\n end"
"\n end", nrowA, ncolA) ;
}
}
/*
-------------------------------
generate dense matrices X and Y
-------------------------------
*/
if ( transposeflag == 0 ) {
nrowX = ncolA ;
nrowY = nrowA ;
} else {
nrowX = nrowA ;
nrowY = ncolA ;
}
X = DenseMtx_new() ;
Y = DenseMtx_new() ;
Y2 = DenseMtx_new() ;
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
DenseMtx_init(X, SPOOLES_REAL, 0, 0, nrowX, nrhs, 1, nrowX) ;
Drand_fillDvector(drand, nrowX*nrhs, DenseMtx_entries(X)) ;
DenseMtx_init(Y, SPOOLES_REAL, 0, 0, nrowY, nrhs, 1, nrowY) ;
Drand_fillDvector(drand, nrowY*nrhs, DenseMtx_entries(Y)) ;
DenseMtx_init(Y2, SPOOLES_REAL, 0, 0, nrowY, nrhs, 1, nrowY) ;
DVcopy(nrowY*nrhs, DenseMtx_entries(Y2), DenseMtx_entries(Y)) ;
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
DenseMtx_init(X, SPOOLES_COMPLEX, 0, 0, nrowX, nrhs, 1, nrowX) ;
Drand_fillDvector(drand, 2*nrowX*nrhs, DenseMtx_entries(X)) ;
DenseMtx_init(Y, SPOOLES_COMPLEX, 0, 0, nrowY, nrhs, 1, nrowY) ;
Drand_fillDvector(drand, 2*nrowY*nrhs, DenseMtx_entries(Y)) ;
DenseMtx_init(Y2, SPOOLES_COMPLEX, 0, 0, nrowY, nrhs, 1, nrowY) ;
DVcopy(2*nrowY*nrhs, DenseMtx_entries(Y2), DenseMtx_entries(Y)) ;
}
if ( msglvl > 1 ) {
fprintf(msgFile, "\n X = zeros(%d,%d) ;", nrowX, nrhs) ;
DenseMtx_writeForMatlab(X, "X", msgFile) ;
fprintf(msgFile, "\n Y = zeros(%d,%d) ;", nrowY, nrhs) ;
DenseMtx_writeForMatlab(Y, "Y", msgFile) ;
}
/*
--------------------------------------------
perform the matrix-matrix multiply in serial
--------------------------------------------
*/
if ( msglvl > 1 ) {
fprintf(msgFile, "\n alpha = %20.12e + %20.2e*i;",
alpha[0], alpha[1]);
fprintf(msgFile, "\n Z = zeros(%d,1) ;", nrowY) ;
}
if ( transposeflag == 0 ) {
MARKTIME(t1) ;
if ( symflag == 0 ) {
InpMtx_sym_mmm(A, Y, alpha, X) ;
} else if ( symflag == 1 ) {
InpMtx_herm_mmm(A, Y, alpha, X) ;
} else if ( symflag == 2 ) {
InpMtx_nonsym_mmm(A, Y, alpha, X) ;
}
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y, "Z", msgFile) ;
fprintf(msgFile, "\n maxerr = max(Z - Y - alpha*A*X) ") ;
fprintf(msgFile, "\n") ;
}
} else if ( transposeflag == 1 ) {
MARKTIME(t1) ;
InpMtx_nonsym_mmm_H(A, Y, alpha, X) ;
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y, "Z", msgFile) ;
fprintf(msgFile,
"\n maxerr = max(Z - Y - alpha*ctranspose(A)*X) ") ;
fprintf(msgFile, "\n") ;
}
} else if ( transposeflag == 2 ) {
MARKTIME(t1) ;
InpMtx_nonsym_mmm_T(A, Y, alpha, X) ;
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y, "Z", msgFile) ;
fprintf(msgFile,
"\n maxerr = max(Z - Y - alpha*transpose(A)*X) ") ;
fprintf(msgFile, "\n") ;
}
}
fprintf(msgFile, "\n %% %d ops, %.3f time, %.3f serial mflops",
nops, t2 - t1, 1.e-6*nops/(t2 - t1)) ;
/*
--------------------------------------------------------
perform the matrix-matrix multiply in multithreaded mode
--------------------------------------------------------
*/
if ( msglvl > 1 ) {
fprintf(msgFile,
"\n alpha = %20.12e + %20.2e*i;", alpha[0], alpha[1]);
fprintf(msgFile, "\n Z = zeros(%d,1) ;", nrowY) ;
}
if ( transposeflag == 0 ) {
MARKTIME(t1) ;
if ( symflag == 0 ) {
InpMtx_MT_sym_mmm(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
} else if ( symflag == 1 ) {
InpMtx_MT_herm_mmm(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
} else if ( symflag == 2 ) {
InpMtx_MT_nonsym_mmm(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
}
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y2, "Z2", msgFile) ;
fprintf(msgFile, "\n maxerr2 = max(Z2 - Y - alpha*A*X) ") ;
fprintf(msgFile, "\n") ;
}
} else if ( transposeflag == 1 ) {
MARKTIME(t1) ;
InpMtx_MT_nonsym_mmm_H(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y2, "Z2", msgFile) ;
fprintf(msgFile,
"\n maxerr2 = max(Z2 - Y - alpha*ctranspose(A)*X) ") ;
fprintf(msgFile, "\n") ;
}
} else if ( transposeflag == 2 ) {
MARKTIME(t1) ;
InpMtx_MT_nonsym_mmm_T(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y2, "Z2", msgFile) ;
fprintf(msgFile,
"\n maxerr2 = max(Z2 - Y - alpha*transpose(A)*X) ") ;
fprintf(msgFile, "\n") ;
}
}
fprintf(msgFile, "\n %% %d ops, %.3f time, %.3f MT mflops",
nops, t2 - t1, 1.e-6*nops/(t2 - t1)) ;
/*
------------------------
free the working storage
------------------------
*/
InpMtx_free(A) ;
DenseMtx_free(X) ;
DenseMtx_free(Y) ;
DenseMtx_free(Y2) ;
IVfree(rowids) ;
IVfree(colids) ;
Drand_free(drand) ;
fclose(msgFile) ;
return(1) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] ) {
/*
--------------------------------------------------
all-in-one program to solve A X = B
using a multithreaded factorization and solve
We use a patch-and-go strategy
for the factorization without pivoting
(1) read in matrix entries and form DInpMtx object
(2) form Graph object
(3) order matrix and form front tree
(4) get the permutation, permute the matrix and
front tree and get the symbolic factorization
(5) compute the numeric factorization
(6) read in right hand side entries
(7) compute the solution
created -- 98jun04, cca
--------------------------------------------------
*/
/*--------------------------------------------------------------------*/
char *matrixFileName, *rhsFileName ;
DenseMtx *mtxB, *mtxX ;
Chv *rootchv ;
ChvManager *chvmanager ;
double fudge, imag, real, tau = 100., toosmall, value ;
double cpus[10] ;
DV *cumopsDV ;
ETree *frontETree ;
FrontMtx *frontmtx ;
FILE *inputFile, *msgFile ;
Graph *graph ;
InpMtx *mtxA ;
int error, ient, irow, jcol, jrhs, jrow, lookahead, msglvl,
ncol, nedges, nent, neqns, nfront, nrhs, nrow, nthread,
patchAndGoFlag, seed,
storeids, storevalues, symmetryflag, type ;
int *newToOld, *oldToNew ;
int stats[20] ;
IV *newToOldIV, *oldToNewIV, *ownersIV ;
IVL *adjIVL, *symbfacIVL ;
SolveMap *solvemap ;
SubMtxManager *mtxmanager ;
/*--------------------------------------------------------------------*/
/*
--------------------
get input parameters
--------------------
*/
if ( argc != 14 ) {
fprintf(stdout, "\n"
"\n usage: %s msglvl msgFile type symmetryflag patchAndGoFlag"
"\n fudge toosmall storeids storevalues"
"\n matrixFileName rhsFileName seed"
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n type -- type of entries"
"\n 1 (SPOOLES_REAL) -- real entries"
"\n 2 (SPOOLES_COMPLEX) -- complex entries"
"\n symmetryflag -- type of matrix"
"\n 0 (SPOOLES_SYMMETRIC) -- symmetric entries"
"\n 1 (SPOOLES_HERMITIAN) -- Hermitian entries"
"\n 2 (SPOOLES_NONSYMMETRIC) -- nonsymmetric entries"
"\n patchAndGoFlag -- flag for the patch-and-go strategy"
"\n 0 -- none, stop factorization"
"\n 1 -- optimization strategy"
"\n 2 -- structural analysis strategy"
"\n fudge -- perturbation parameter"
"\n toosmall -- upper bound on a small pivot"
"\n storeids -- flag to store ids of small pivots"
"\n storevalues -- flag to store perturbations"
"\n matrixFileName -- matrix file name, format"
"\n nrow ncol nent"
"\n irow jcol entry"
"\n ..."
"\n note: indices are zero based"
"\n rhsFileName -- right hand side file name, format"
"\n nrow nrhs "
"\n ..."
"\n jrow entry(jrow,0) ... entry(jrow,nrhs-1)"
"\n ..."
"\n seed -- random number seed, used for ordering"
"\n nthread -- number of threads"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
type = atoi(argv[3]) ;
symmetryflag = atoi(argv[4]) ;
patchAndGoFlag = atoi(argv[5]) ;
fudge = atof(argv[6]) ;
toosmall = atof(argv[7]) ;
storeids = atoi(argv[8]) ;
storevalues = atoi(argv[9]) ;
matrixFileName = argv[10] ;
rhsFileName = argv[11] ;
seed = atoi(argv[12]) ;
nthread = atoi(argv[13]) ;
/*--------------------------------------------------------------------*/
/*
--------------------------------------------
STEP 1: read the entries from the input file
and create the InpMtx object
--------------------------------------------
*/
if ( (inputFile = fopen(matrixFileName, "r")) == NULL ) {
fprintf(stderr, "\n unable to open file %s", matrixFileName) ;
spoolesFatal();
}
fscanf(inputFile, "%d %d %d", &nrow, &ncol, &nent) ;
neqns = nrow ;
mtxA = InpMtx_new() ;
InpMtx_init(mtxA, INPMTX_BY_ROWS, type, nent, 0) ;
if ( type == SPOOLES_REAL ) {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le", &irow, &jcol, &value) ;
InpMtx_inputRealEntry(mtxA, irow, jcol, value) ;
}
} else {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le %le", &irow, &jcol, &real, &imag) ;
InpMtx_inputComplexEntry(mtxA, irow, jcol, real, imag) ;
}
}
fclose(inputFile) ;
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n input matrix") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------------------------
STEP 2 : find a low-fill ordering
(1) create the Graph object
(2) order the graph using multiple minimum degree
-------------------------------------------------
*/
graph = Graph_new() ;
adjIVL = InpMtx_fullAdjacency(mtxA) ;
nedges = IVL_tsize(adjIVL) ;
Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL,
NULL, NULL) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n graph of the input matrix") ;
Graph_writeForHumanEye(graph, msgFile) ;
fflush(msgFile) ;
}
frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n front tree from ordering") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------------------
STEP 3: get the permutation, permute the matrix and
front tree and get the symbolic factorization
-----------------------------------------------------
*/
oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ;
oldToNew = IV_entries(oldToNewIV) ;
newToOldIV = ETree_newToOldVtxPerm(frontETree) ;
newToOld = IV_entries(newToOldIV) ;
ETree_permuteVertices(frontETree, oldToNewIV) ;
InpMtx_permute(mtxA, oldToNew, oldToNew) ;
InpMtx_mapToUpperTriangle(mtxA) ;
InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS) ;
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n old-to-new permutation vector") ;
IV_writeForHumanEye(oldToNewIV, msgFile) ;
fprintf(msgFile, "\n\n new-to-old permutation vector") ;
IV_writeForHumanEye(newToOldIV, msgFile) ;
fprintf(msgFile, "\n\n front tree after permutation") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fprintf(msgFile, "\n\n input matrix after permutation") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fprintf(msgFile, "\n\n symbolic factorization") ;
IVL_writeForHumanEye(symbfacIVL, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------------------------
STEP 4: initialize the front matrix object
and the PatchAndGoInfo object to handle
small pivots
------------------------------------------
*/
frontmtx = FrontMtx_new() ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, LOCK_IN_PROCESS, 0) ;
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type, symmetryflag,
FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, LOCK_IN_PROCESS,
0, NULL, mtxmanager, msglvl, msgFile) ;
if ( patchAndGoFlag == 1 ) {
frontmtx->patchinfo = PatchAndGoInfo_new() ;
PatchAndGoInfo_init(frontmtx->patchinfo, 1, toosmall, fudge,
storeids, storevalues) ;
} else if ( patchAndGoFlag == 2 ) {
frontmtx->patchinfo = PatchAndGoInfo_new() ;
PatchAndGoInfo_init(frontmtx->patchinfo, 2, toosmall, fudge,
storeids, storevalues) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------------------------
STEP 5: setup the domain decomposition map
------------------------------------------
*/
if ( nthread > (nfront = FrontMtx_nfront(frontmtx)) ) {
nthread = nfront ;
}
cumopsDV = DV_new() ;
DV_init(cumopsDV, nthread, NULL) ;
ownersIV = ETree_ddMap(frontETree, type, symmetryflag,
cumopsDV, 1./(2.*nthread)) ;
DV_free(cumopsDV) ;
/*--------------------------------------------------------------------*/
/*
-----------------------------------------------------
STEP 6: compute the numeric factorization in parallel
-----------------------------------------------------
*/
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, LOCK_IN_PROCESS, 1) ;
DVfill(10, cpus, 0.0) ;
IVfill(20, stats, 0) ;
lookahead = 0 ;
rootchv = FrontMtx_MT_factorInpMtx(frontmtx, mtxA, tau, 0.0,
chvmanager, ownersIV, lookahead,
&error, cpus, stats, msglvl, msgFile) ;
if ( patchAndGoFlag == 1 ) {
if ( frontmtx->patchinfo->fudgeIV != NULL ) {
fprintf(msgFile, "\n small pivots found at these locations") ;
IV_writeForHumanEye(frontmtx->patchinfo->fudgeIV, msgFile) ;
}
PatchAndGoInfo_free(frontmtx->patchinfo) ;
} else if ( patchAndGoFlag == 2 ) {
if ( frontmtx->patchinfo->fudgeIV != NULL ) {
fprintf(msgFile, "\n small pivots found at these locations") ;
IV_writeForHumanEye(frontmtx->patchinfo->fudgeIV, msgFile) ;
}
if ( frontmtx->patchinfo->fudgeDV != NULL ) {
fprintf(msgFile, "\n perturbations") ;
DV_writeForHumanEye(frontmtx->patchinfo->fudgeDV, msgFile) ;
}
PatchAndGoInfo_free(frontmtx->patchinfo) ;
}
ChvManager_free(chvmanager) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
if ( rootchv != NULL ) {
fprintf(msgFile, "\n\n matrix found to be singular\n") ;
spoolesFatal();
}
if ( error >= 0 ) {
fprintf(msgFile, "\n\n fatal error at front %d\n", error) ;
spoolesFatal();
}
/*
--------------------------------------
STEP 7: post-process the factorization
--------------------------------------
*/
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------
STEP 8: read the right hand side matrix B
-----------------------------------------
*/
if ( (inputFile = fopen(rhsFileName, "r")) == NULL ) {
fprintf(stderr, "\n unable to open file %s", rhsFileName) ;
spoolesFatal();
}
fscanf(inputFile, "%d %d", &nrow, &nrhs) ;
mtxB = DenseMtx_new() ;
DenseMtx_init(mtxB, type, 0, 0, neqns, nrhs, 1, neqns) ;
DenseMtx_zero(mtxB) ;
if ( type == SPOOLES_REAL ) {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le", &value) ;
DenseMtx_setRealEntry(mtxB, jrow, jrhs, value) ;
}
}
} else {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le %le", &real, &imag) ;
DenseMtx_setComplexEntry(mtxB, jrow, jrhs, real, imag) ;
}
}
}
fclose(inputFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n rhs matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
--------------------------------------------------------------
STEP 9: permute the right hand side into the original ordering
--------------------------------------------------------------
*/
DenseMtx_permuteRows(mtxB, oldToNewIV) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n right hand side matrix in new ordering") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
--------------------------------------------------------
STEP 10: get the solve map object for the parallel solve
--------------------------------------------------------
*/
solvemap = SolveMap_new() ;
SolveMap_ddMap(solvemap, type, FrontMtx_upperBlockIVL(frontmtx),
FrontMtx_lowerBlockIVL(frontmtx), nthread, ownersIV,
FrontMtx_frontTree(frontmtx), seed, msglvl, msgFile) ;
/*--------------------------------------------------------------------*/
/*
--------------------------------------------
STEP 11: solve the linear system in parallel
--------------------------------------------
*/
mtxX = DenseMtx_new() ;
DenseMtx_init(mtxX, type, 0, 0, neqns, nrhs, 1, neqns) ;
DenseMtx_zero(mtxX) ;
FrontMtx_MT_solve(frontmtx, mtxX, mtxB, mtxmanager, solvemap,
cpus, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n solution matrix in new ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
--------------------------------------------------------
STEP 12: permute the solution into the original ordering
--------------------------------------------------------
*/
DenseMtx_permuteRows(mtxX, newToOldIV) ;
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n solution matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------
free memory
-----------
*/
FrontMtx_free(frontmtx) ;
DenseMtx_free(mtxX) ;
DenseMtx_free(mtxB) ;
IV_free(newToOldIV) ;
IV_free(oldToNewIV) ;
InpMtx_free(mtxA) ;
ETree_free(frontETree) ;
IVL_free(symbfacIVL) ;
SubMtxManager_free(mtxmanager) ;
Graph_free(graph) ;
SolveMap_free(solvemap) ;
IV_free(ownersIV) ;
/*--------------------------------------------------------------------*/
return(1) ; }
