void spooles_solve(void *ptr, double *b, long neq)
{
/* rhs vector B
* Note that there is only one rhs vector, thus
* a bit simpler that the AllInOne example
*/
long size = neq;
DenseMtx *mtxB,*mtxX;
struct factorinfo *pfi_ = ptr;
//printf(" Solving the system of equations using the symmetric spooles solver\n");
{
int i;
mtxB = DenseMtx_new();
DenseMtx_init(mtxB, SPOOLES_REAL, 0, 0, size, 1, 1, size);
DenseMtx_zero(mtxB);
for (i = 0; i < size; i++) {
DenseMtx_setRealEntry(mtxB, i, 0, b[i]);
}
if (DEBUG_LVL > 1) {
fprintf(msgFile, "\n\n rhs matrix in original ordering");
DenseMtx_writeForHumanEye(mtxB, msgFile);
fflush(msgFile);
}
}
#ifdef USE_MT
//printf(" Using up to %d cpu(s) for spooles.\n\n", num_cpus);
if (num_cpus > 1) {
/* do not use the multithreaded solver unless
* we have multiple threads - avoid the
* locking overhead
*/
mtxX=fsolve_MT(pfi_, mtxB);
} else {
mtxX=fsolve(pfi_, mtxB);
}
#else
//printf(" Using 1 cpu for spooles.\n\n");
mtxX=fsolve(pfi_, mtxB);
#endif
/* convert the result back to Calculix representation */
{
int i;
for (i = 0; i < size; i++) {
b[i] = DenseMtx_entries(mtxX)[i];
}
}
/* cleanup */
DenseMtx_free(mtxX);
}
PetscErrorCode MatSolve_SeqSpooles(Mat A,Vec b,Vec x)
{
Mat_Spooles *lu = (Mat_Spooles*)A->spptr;
PetscScalar *array;
DenseMtx *mtxY, *mtxX ;
PetscErrorCode ierr;
PetscInt irow,neqns=A->cmap->n,nrow=A->rmap->n,*iv;
#if defined(PETSC_USE_COMPLEX)
double x_real,x_imag;
#else
double *entX;
#endif
PetscFunctionBegin;
mtxY = DenseMtx_new();
DenseMtx_init(mtxY, lu->options.typeflag, 0, 0, nrow, 1, 1, nrow); /* column major */
ierr = VecGetArray(b,&array);CHKERRQ(ierr);
if (lu->options.useQR) { /* copy b to mtxY */
for ( irow = 0 ; irow < nrow; irow++ )
#if !defined(PETSC_USE_COMPLEX)
DenseMtx_setRealEntry(mtxY, irow, 0, *array++);
#else
DenseMtx_setComplexEntry(mtxY, irow, 0, PetscRealPart(array[irow]), PetscImaginaryPart(array[irow]));
#endif
} else { /* copy permuted b to mtxY */
iv = IV_entries(lu->oldToNewIV);
for ( irow = 0 ; irow < nrow; irow++ )
#if !defined(PETSC_USE_COMPLEX)
DenseMtx_setRealEntry(mtxY, *iv++, 0, *array++);
#else
DenseMtx_setComplexEntry(mtxY,*iv++,0,PetscRealPart(array[irow]),PetscImaginaryPart(array[irow]));
#endif
}
ierr = VecRestoreArray(b,&array);CHKERRQ(ierr);
mtxX = DenseMtx_new();
DenseMtx_init(mtxX, lu->options.typeflag, 0, 0, neqns, 1, 1, neqns);
if (lu->options.useQR) {
FrontMtx_QR_solve(lu->frontmtx, lu->mtxA, mtxX, mtxY, lu->mtxmanager,
lu->cpus, lu->options.msglvl, lu->options.msgFile);
} else {
FrontMtx_solve(lu->frontmtx, mtxX, mtxY, lu->mtxmanager,
lu->cpus, lu->options.msglvl, lu->options.msgFile);
}
if ( lu->options.msglvl > 2 ) {
int err;
ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n right hand side matrix after permutation");CHKERRQ(ierr);
DenseMtx_writeForHumanEye(mtxY, lu->options.msgFile);
ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n solution matrix in new ordering");CHKERRQ(ierr);
DenseMtx_writeForHumanEye(mtxX, lu->options.msgFile);
err = fflush(lu->options.msgFile);
if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file");
}
/* permute solution into original ordering, then copy to x */
DenseMtx_permuteRows(mtxX, lu->newToOldIV);
ierr = VecGetArray(x,&array);CHKERRQ(ierr);
#if !defined(PETSC_USE_COMPLEX)
entX = DenseMtx_entries(mtxX);
DVcopy(neqns, array, entX);
#else
for (irow=0; irow<nrow; irow++){
DenseMtx_complexEntry(mtxX,irow,0,&x_real,&x_imag);
array[irow] = x_real+x_imag*PETSC_i;
}
#endif
ierr = VecRestoreArray(x,&array);CHKERRQ(ierr);
/* free memory */
DenseMtx_free(mtxX);
DenseMtx_free(mtxY);
PetscFunctionReturn(0);
}
NM_Status
SpoolesSolver :: solve(SparseMtrx *A, FloatArray *b, FloatArray *x)
{
int errorValue, mtxType, symmetryflag;
int seed = 30145, pivotingflag = 0;
int *oldToNew, *newToOld;
double droptol = 0.0, tau = 1.e300;
double cpus [ 10 ];
int stats [ 20 ];
ChvManager *chvmanager;
Chv *rootchv;
InpMtx *mtxA;
DenseMtx *mtxY, *mtxX;
// first check whether Lhs is defined
if ( !A ) {
_error("solveYourselfAt: unknown Lhs");
}
// and whether Rhs
if ( !b ) {
_error("solveYourselfAt: unknown Rhs");
}
// and whether previous Solution exist
if ( !x ) {
_error("solveYourselfAt: unknown solution array");
}
if ( x->giveSize() != b->giveSize() ) {
_error("solveYourselfAt: size mismatch");
}
Timer timer;
timer.startTimer();
if ( A->giveType() != SMT_SpoolesMtrx ) {
_error("solveYourselfAt: SpoolesSparseMtrx Expected");
}
mtxA = ( ( SpoolesSparseMtrx * ) A )->giveInpMtrx();
mtxType = ( ( SpoolesSparseMtrx * ) A )->giveValueType();
symmetryflag = ( ( SpoolesSparseMtrx * ) A )->giveSymmetryFlag();
int i;
int neqns = A->giveNumberOfRows();
int nrhs = 1;
/* convert right-hand side to DenseMtx */
mtxY = DenseMtx_new();
DenseMtx_init(mtxY, mtxType, 0, 0, neqns, nrhs, 1, neqns);
DenseMtx_zero(mtxY);
for ( i = 0; i < neqns; i++ ) {
DenseMtx_setRealEntry( mtxY, i, 0, b->at(i + 1) );
}
if ( ( Lhs != A ) || ( this->lhsVersion != A->giveVersion() ) ) {
//
// lhs has been changed -> new factorization
//
Lhs = A;
this->lhsVersion = A->giveVersion();
if ( frontmtx ) {
FrontMtx_free(frontmtx);
}
if ( newToOldIV ) {
IV_free(newToOldIV);
}
if ( oldToNewIV ) {
IV_free(oldToNewIV);
}
if ( frontETree ) {
ETree_free(frontETree);
}
if ( symbfacIVL ) {
IVL_free(symbfacIVL);
}
if ( mtxmanager ) {
SubMtxManager_free(mtxmanager);
}
if ( graph ) {
Graph_free(graph);
}
/*
* -------------------------------------------------
* STEP 3 : find a low-fill ordering
* (1) create the Graph object
* (2) order the graph using multiple minimum degree
* -------------------------------------------------
*/
int nedges;
graph = Graph_new();
adjIVL = InpMtx_fullAdjacency(mtxA);
nedges = IVL_tsize(adjIVL);
Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL,
NULL, NULL);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n graph of the input matrix");
Graph_writeForHumanEye(graph, msgFile);
fflush(msgFile);
}
frontETree = orderViaMMD(graph, seed, msglvl, msgFile);
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n front tree from ordering");
ETree_writeForHumanEye(frontETree, msgFile);
fflush(msgFile);
}
/*
* ----------------------------------------------------
* STEP 4: get the permutation, permute the front tree,
* permute the matrix and right hand side, and
* get the symbolic factorization
* ----------------------------------------------------
*/
oldToNewIV = ETree_oldToNewVtxPerm(frontETree);
oldToNew = IV_entries(oldToNewIV);
newToOldIV = ETree_newToOldVtxPerm(frontETree);
newToOld = IV_entries(newToOldIV);
ETree_permuteVertices(frontETree, oldToNewIV);
InpMtx_permute(mtxA, oldToNew, oldToNew);
if ( symmetryflag == SPOOLES_SYMMETRIC ||
symmetryflag == SPOOLES_HERMITIAN ) {
InpMtx_mapToUpperTriangle(mtxA);
}
InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS);
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS);
symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n old-to-new permutation vector");
IV_writeForHumanEye(oldToNewIV, msgFile);
fprintf(msgFile, "\n\n new-to-old permutation vector");
IV_writeForHumanEye(newToOldIV, msgFile);
fprintf(msgFile, "\n\n front tree after permutation");
ETree_writeForHumanEye(frontETree, msgFile);
fprintf(msgFile, "\n\n input matrix after permutation");
InpMtx_writeForHumanEye(mtxA, msgFile);
fprintf(msgFile, "\n\n symbolic factorization");
IVL_writeForHumanEye(symbfacIVL, msgFile);
fflush(msgFile);
}
Tree_writeToFile(frontETree->tree, (char*)"haggar.treef");
/*--------------------------------------------------------------------*/
/*
* ------------------------------------------
* STEP 5: initialize the front matrix object
* ------------------------------------------
*/
frontmtx = FrontMtx_new();
mtxmanager = SubMtxManager_new();
SubMtxManager_init(mtxmanager, NO_LOCK, 0);
FrontMtx_init(frontmtx, frontETree, symbfacIVL, mtxType, symmetryflag,
FRONTMTX_DENSE_FRONTS, pivotingflag, NO_LOCK, 0, NULL,
mtxmanager, msglvl, msgFile);
/*--------------------------------------------------------------------*/
/*
* -----------------------------------------
* STEP 6: compute the numeric factorization
* -----------------------------------------
*/
chvmanager = ChvManager_new();
ChvManager_init(chvmanager, NO_LOCK, 1);
DVfill(10, cpus, 0.0);
IVfill(20, stats, 0);
rootchv = FrontMtx_factorInpMtx(frontmtx, mtxA, tau, droptol,
chvmanager, & errorValue, cpus, stats, msglvl, msgFile);
ChvManager_free(chvmanager);
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n factor matrix");
FrontMtx_writeForHumanEye(frontmtx, msgFile);
fflush(msgFile);
}
if ( rootchv != NULL ) {
fprintf(msgFile, "\n\n matrix found to be singular\n");
exit(-1);
}
if ( errorValue >= 0 ) {
fprintf(msgFile, "\n\n error encountered at front %d", errorValue);
exit(-1);
}
/*--------------------------------------------------------------------*/
/*
* --------------------------------------
* STEP 7: post-process the factorization
* --------------------------------------
*/
FrontMtx_postProcess(frontmtx, msglvl, msgFile);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n factor matrix after post-processing");
FrontMtx_writeForHumanEye(frontmtx, msgFile);
fflush(msgFile);
}
/*--------------------------------------------------------------------*/
}
/*
* ----------------------------------------------------
* STEP 4: permute the right hand side
* ----------------------------------------------------
*/
DenseMtx_permuteRows(mtxY, oldToNewIV);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n right hand side matrix after permutation");
DenseMtx_writeForHumanEye(mtxY, msgFile);
}
/*
* -------------------------------
* STEP 8: solve the linear system
* -------------------------------
*/
mtxX = DenseMtx_new();
DenseMtx_init(mtxX, mtxType, 0, 0, neqns, nrhs, 1, neqns);
DenseMtx_zero(mtxX);
FrontMtx_solve(frontmtx, mtxX, mtxY, mtxmanager,
cpus, msglvl, msgFile);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n solution matrix in new ordering");
DenseMtx_writeForHumanEye(mtxX, msgFile);
fflush(msgFile);
}
/*--------------------------------------------------------------------*/
/*
* -------------------------------------------------------
* STEP 9: permute the solution into the original ordering
* -------------------------------------------------------
*/
DenseMtx_permuteRows(mtxX, newToOldIV);
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n solution matrix in original ordering");
DenseMtx_writeForHumanEye(mtxX, msgFile);
fflush(msgFile);
}
// DenseMtx_writeForMatlab(mtxX, "x", msgFile) ;
/*--------------------------------------------------------------------*/
/* fetch data to oofem vectors */
double *xptr = x->givePointer();
for ( i = 0; i < neqns; i++ ) {
DenseMtx_realEntry(mtxX, i, 0, xptr + i);
// printf ("x(%d) = %e\n", i+1, *(xptr+i));
}
// DenseMtx_copyRowIntoVector(mtxX, 0, x->givePointer());
timer.stopTimer();
OOFEM_LOG_DEBUG( "SpoolesSolver info: user time consumed by solution: %.2fs\n", timer.getUtime() );
/*
* -----------
* free memory
* -----------
*/
DenseMtx_free(mtxX);
DenseMtx_free(mtxY);
/*--------------------------------------------------------------------*/
return ( 1 );
}
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] ) {
/*
--------------------------------------------------
QR all-in-one program
(1) read in matrix entries and form InpMtx object
of A and A^TA
(2) form Graph object of A^TA
(3) order matrix and form front tree
(4) get the permutation, permute the matrix and
front tree and get the symbolic factorization
(5) compute the numeric factorization
(6) read in right hand side entries
(7) compute the solution
created -- 98jun11, cca
--------------------------------------------------
*/
/*--------------------------------------------------------------------*/
char *matrixFileName, *rhsFileName ;
ChvManager *chvmanager ;
DenseMtx *mtxB, *mtxX ;
double facops, imag, real, value ;
double cpus[10] ;
ETree *frontETree ;
FILE *inputFile, *msgFile ;
FrontMtx *frontmtx ;
Graph *graph ;
int ient, irow, jcol, jrhs, jrow, msglvl, neqns,
nedges, nent, nrhs, nrow, seed, type ;
InpMtx *mtxA ;
IV *newToOldIV, *oldToNewIV ;
IVL *adjIVL, *symbfacIVL ;
SubMtxManager *mtxmanager ;
/*--------------------------------------------------------------------*/
/*
--------------------
get input parameters
--------------------
*/
if ( argc != 7 ) {
fprintf(stdout,
"\n usage: %s msglvl msgFile type matrixFileName rhsFileName seed"
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n type -- type of entries"
"\n 1 (SPOOLES_REAL) -- real entries"
"\n 2 (SPOOLES_COMPLEX) -- complex entries"
"\n matrixFileName -- matrix file name, format"
"\n nrow ncol nent"
"\n irow jcol entry"
"\n ..."
"\n note: indices are zero based"
"\n rhsFileName -- right hand side file name, format"
"\n nrow "
"\n entry[0]"
"\n ..."
"\n entry[nrow-1]"
"\n seed -- random number seed, used for ordering"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
type = atoi(argv[3]) ;
matrixFileName = argv[4] ;
rhsFileName = argv[5] ;
seed = atoi(argv[6]) ;
/*--------------------------------------------------------------------*/
/*
--------------------------------------------
STEP 1: read the entries from the input file
and create the InpMtx object of A
--------------------------------------------
*/
inputFile = fopen(matrixFileName, "r") ;
fscanf(inputFile, "%d %d %d", &nrow, &neqns, &nent) ;
mtxA = InpMtx_new() ;
InpMtx_init(mtxA, INPMTX_BY_ROWS, type, nent, 0) ;
if ( type == SPOOLES_REAL ) {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le", &irow, &jcol, &value) ;
InpMtx_inputRealEntry(mtxA, irow, jcol, value) ;
}
} else {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le %le", &irow, &jcol, &real, &imag) ;
InpMtx_inputComplexEntry(mtxA, irow, jcol, real, imag) ;
}
}
fclose(inputFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n input matrix") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
----------------------------------------
STEP 2: read the right hand side entries
----------------------------------------
*/
inputFile = fopen(rhsFileName, "r") ;
fscanf(inputFile, "%d %d", &nrow, &nrhs) ;
mtxB = DenseMtx_new() ;
DenseMtx_init(mtxB, type, 0, 0, nrow, nrhs, 1, nrow) ;
DenseMtx_zero(mtxB) ;
if ( type == SPOOLES_REAL ) {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le", &value) ;
DenseMtx_setRealEntry(mtxB, jrow, jrhs, value) ;
}
}
} else {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le %le", &real, &imag) ;
DenseMtx_setComplexEntry(mtxB, jrow, jrhs, real, imag) ;
}
}
}
fclose(inputFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n rhs matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------------------------
STEP 3 : find a low-fill ordering
(1) create the Graph object for A^TA or A^HA
(2) order the graph using multiple minimum degree
-------------------------------------------------
*/
graph = Graph_new() ;
adjIVL = InpMtx_adjForATA(mtxA) ;
nedges = IVL_tsize(adjIVL) ;
Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL,
NULL, NULL) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n graph of A^T A") ;
Graph_writeForHumanEye(graph, msgFile) ;
fflush(msgFile) ;
}
frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n front tree from ordering") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------------------
STEP 4: get the permutation, permute the matrix and
front tree and get the symbolic factorization
-----------------------------------------------------
*/
oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ;
newToOldIV = ETree_newToOldVtxPerm(frontETree) ;
InpMtx_permute(mtxA, NULL, IV_entries(oldToNewIV)) ;
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
symbfacIVL = SymbFac_initFromGraph(frontETree, graph) ;
IVL_overwrite(symbfacIVL, oldToNewIV) ;
IVL_sortUp(symbfacIVL) ;
ETree_permuteVertices(frontETree, oldToNewIV) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n old-to-new permutation vector") ;
IV_writeForHumanEye(oldToNewIV, msgFile) ;
fprintf(msgFile, "\n\n new-to-old permutation vector") ;
IV_writeForHumanEye(newToOldIV, msgFile) ;
fprintf(msgFile, "\n\n front tree after permutation") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fprintf(msgFile, "\n\n input matrix after permutation") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fprintf(msgFile, "\n\n symbolic factorization") ;
IVL_writeForHumanEye(symbfacIVL, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------------------------
STEP 5: initialize the front matrix object
------------------------------------------
*/
frontmtx = FrontMtx_new() ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, NO_LOCK, 0) ;
if ( type == SPOOLES_REAL ) {
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type,
SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS,
SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL,
mtxmanager, msglvl, msgFile) ;
} else {
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type,
SPOOLES_HERMITIAN, FRONTMTX_DENSE_FRONTS,
SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL,
mtxmanager, msglvl, msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------
STEP 6: compute the numeric factorization
-----------------------------------------
*/
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, NO_LOCK, 1) ;
DVzero(10, cpus) ;
facops = 0.0 ;
FrontMtx_QR_factor(frontmtx, mtxA, chvmanager,
cpus, &facops, msglvl, msgFile) ;
ChvManager_free(chvmanager) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix") ;
fprintf(msgFile, "\n facops = %9.2f", facops) ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
--------------------------------------
STEP 7: post-process the factorization
--------------------------------------
*/
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------
STEP 8: solve the linear system
-------------------------------
*/
mtxX = DenseMtx_new() ;
DenseMtx_init(mtxX, type, 0, 0, neqns, nrhs, 1, neqns) ;
FrontMtx_QR_solve(frontmtx, mtxA, mtxX, mtxB, mtxmanager,
cpus, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n solution matrix in new ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------------------------------
STEP 9: permute the solution into the original ordering
-------------------------------------------------------
*/
DenseMtx_permuteRows(mtxX, newToOldIV) ;
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n solution matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------
free the working storage
------------------------
*/
InpMtx_free(mtxA) ;
FrontMtx_free(frontmtx) ;
Graph_free(graph) ;
DenseMtx_free(mtxX) ;
DenseMtx_free(mtxB) ;
ETree_free(frontETree) ;
IV_free(newToOldIV) ;
IV_free(oldToNewIV) ;
IVL_free(symbfacIVL) ;
SubMtxManager_free(mtxmanager) ;
/*--------------------------------------------------------------------*/
return(1) ; }
int SPOOLESSolverMT::SolveLinearSystem(double * x, const double * rhs)
{
BridgeMT * bridgeMT = (BridgeMT*) bridgeMTPointer;
DenseMtx * mtx_rhs = (DenseMtx*) mtx_rhsPointer;
DenseMtx * mtx_x = (DenseMtx*) mtx_xPointer;
// set mtx_rhs
DenseMtx_zero(mtx_rhs);
for(int i=0; i < n; i++)
DenseMtx_setRealEntry(mtx_rhs, i, 0, rhs[i]);
/*
FILE * fout = fopen("bla.txt","w");
DenseMtx_writeForHumanEye(mtx_rhs, fout);
fclose(fout);
*/
// zero the solution
DenseMtx_zero(mtx_x);
// solve
if (verbose >= 2)
printf("Solving the linear system...\n");
// setup the solve
int rc = BridgeMT_solveSetup(bridgeMT) ;
fprintf(msgFile, "\n\n ----- PARALLEL SOLVE SETUP -----\n") ;
fprintf(msgFile,
"\n CPU %8.3f : time to setup parallel solve",
bridgeMT->cpus[11]) ;
// solve the system
int permuteflag = 1;
rc = BridgeMT_solve(bridgeMT, permuteflag, mtx_x, mtx_rhs) ;
if (rc != 1)
{
printf("Error: linear system solve failed. BridgeMT_solve exit code: %d.\n", rc);
return (rc == 0) ? 1 : rc;
}
fprintf(msgFile, "\n\n ----- SOLVE -----\n") ;
fprintf(msgFile,
"\n CPU %8.3f : time to permute rhs into new ordering"
"\n CPU %8.3f : time to solve linear system"
"\n CPU %8.3f : time to permute solution into old ordering"
"\n CPU %8.3f : total solve time\n",
bridgeMT->cpus[12], bridgeMT->cpus[13],
bridgeMT->cpus[14], bridgeMT->cpus[15]) ;
fprintf(msgFile,
"\n\n solve: raw mflops %8.3f, overall mflops %8.3f",
1.e-6*nsolveops/bridgeMT->cpus[13],
1.e-6*nsolveops/bridgeMT->cpus[15]) ;
fflush(msgFile) ;
if (verbose >= 2)
printf("Solve completed.\n");
// store result
for(int i=0; i < n; i++)
DenseMtx_realEntry(mtx_x, i, 0, &x[i]);
/*
fout = fopen("ble.txt","w");
DenseMtx_writeForHumanEye(mtx_x, fout);
fclose(fout);
*/
return 0;
}
/*--------------------------------------------------------------------*/
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) ; }