/*
--------------------------------
purpose -- clear the data fields
created -- 95sep27, cca
--------------------------------
*/
void
Graph_clearData (
Graph *g
) {
#if MYTRACE > 0
fprintf(stdout, "\n just inside Graph_clearData(%p)", g) ;
fflush(stdout) ;
#endif
if ( g == NULL ) {
fprintf(stderr, "\n fatal error in Graph_clearData(%p)"
"\n graph is NULL\n", g) ;
spoolesFatal();
}
if ( g->adjIVL != NULL ) {
IVL_free(g->adjIVL) ;
}
if ( g->vwghts != NULL ) {
IVfree(g->vwghts) ;
}
if ( g->ewghtIVL != NULL ) {
IVL_free(g->ewghtIVL) ;
}
Graph_setDefaultFields(g) ;
#if MYTRACE > 0
fprintf(stdout, "\n leaving Graph_clearData(%p)", g) ;
fflush(stdout) ;
#endif
return ; }
PetscErrorCode MatDestroy_MPISBAIJSpooles(Mat A)
{
Mat_Spooles *lu = (Mat_Spooles*)A->spptr;
PetscErrorCode ierr;
PetscFunctionBegin;
if (lu->CleanUpSpooles) {
FrontMtx_free(lu->frontmtx);
IV_free(lu->newToOldIV);
IV_free(lu->oldToNewIV);
IV_free(lu->vtxmapIV);
InpMtx_free(lu->mtxA);
ETree_free(lu->frontETree);
IVL_free(lu->symbfacIVL);
SubMtxManager_free(lu->mtxmanager);
DenseMtx_free(lu->mtxX);
DenseMtx_free(lu->mtxY);
ierr = MPI_Comm_free(&(lu->comm_spooles));CHKERRQ(ierr);
if ( lu->scat ){
ierr = VecDestroy(lu->vec_spooles);CHKERRQ(ierr);
ierr = ISDestroy(lu->iden);CHKERRQ(ierr);
ierr = ISDestroy(lu->is_petsc);CHKERRQ(ierr);
ierr = VecScatterDestroy(lu->scat);CHKERRQ(ierr);
}
}
ierr = MatDestroy_MPISBAIJ(A);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
SpoolesSolver :: ~SpoolesSolver()
{
if ( msgFileCloseFlag ) {
fclose(msgFile);
}
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);
}
}
/*
-----------------------
clear the data fields
return value ---
1 -- normal return
-1 -- bridge is NULL
created -- 98sep18, cca
-----------------------
*/
int
Bridge_clearData (
Bridge *bridge
) {
if ( bridge == NULL ) {
fprintf(stderr, "\n fatal error in Bridge_clearData(%p)"
"\n bad input\n", bridge) ;
return(-1) ;
}
if ( bridge->frontmtx != NULL ) {
FrontMtx_free(bridge->frontmtx) ;
}
if ( bridge->frontETree != NULL ) {
ETree_free(bridge->frontETree) ;
}
if ( bridge->symbfacIVL != NULL ) {
IVL_free(bridge->symbfacIVL) ;
}
if ( bridge->mtxmanager != NULL ) {
SubMtxManager_free(bridge->mtxmanager) ;
}
if ( bridge->oldToNewIV != NULL ) {
IV_free(bridge->oldToNewIV) ;
}
if ( bridge->newToOldIV != NULL ) {
IV_free(bridge->newToOldIV) ;
}
Bridge_setDefaultFields(bridge) ;
return(1) ; }
PetscErrorCode MatDestroy_SeqAIJSpooles(Mat A)
{
Mat_Spooles *lu = (Mat_Spooles*)A->spptr;
PetscErrorCode ierr;
PetscFunctionBegin;
if (lu && lu->CleanUpSpooles) {
FrontMtx_free(lu->frontmtx);
IV_free(lu->newToOldIV);
IV_free(lu->oldToNewIV);
InpMtx_free(lu->mtxA);
ETree_free(lu->frontETree);
IVL_free(lu->symbfacIVL);
SubMtxManager_free(lu->mtxmanager);
Graph_free(lu->graph);
}
ierr = PetscFree(A->spptr);CHKERRQ(ierr);
ierr = MatDestroy_SeqAIJ(A);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*--------------------------------------------------------------------*/
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) ; }
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);
}
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) ; }
//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);
}
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 );
}
/*
-------------------------------------------------------------------
purpose --
given an InpMtx object that contains the structure of A, initialize
the bridge data structure for the serial factor's and solve's.
note: all parameters are pointers to be compatible with
fortran's call by reference.
return value --
1 -- normal return
-1 -- bridge is NULL
-2 -- mtxA is NULL
created -- 98sep17, cca
-------------------------------------------------------------------
*/
int
BridgeMT_setup (
BridgeMT *bridge,
InpMtx *mtxA
) {
double t0, t1, t2 ;
ETree *frontETree ;
FILE *msgFile ;
Graph *graph ;
int compressed, msglvl, nedges, neqns, Neqns ;
IV *eqmapIV ;
IVL *adjIVL, *symbfacIVL ;
MARKTIME(t0) ;
/*
--------------------
check the input data
--------------------
*/
if ( bridge == NULL ) {
fprintf(stderr, "\n fatal error in BridgeMT_setup()"
"\n data is NULL\n") ;
return(-1) ;
}
if ( mtxA == NULL ) {
fprintf(stderr, "\n fatal error in BridgeMT_setup()"
"\n A is NULL\n") ;
return(-2) ;
}
msglvl = bridge->msglvl ;
msgFile = bridge->msgFile ;
neqns = bridge->neqns ;
if ( ! (INPMTX_IS_BY_ROWS(mtxA) || INPMTX_IS_BY_COLUMNS(mtxA)) ) {
/*
------------------------------
change coordinate type to rows
------------------------------
*/
InpMtx_changeCoordType(mtxA, INPMTX_BY_ROWS) ;
}
if ( ! INPMTX_IS_BY_VECTORS(mtxA) ) {
/*
------------------------------
change storage mode to vectors
------------------------------
*/
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
}
/*
---------------------------
create a Graph object for A
---------------------------
*/
MARKTIME(t1) ;
graph = Graph_new() ;
adjIVL = InpMtx_fullAdjacency(mtxA);
nedges = bridge->nedges = IVL_tsize(adjIVL),
Graph_init2(graph, 0, neqns, 0, nedges,
neqns, nedges, adjIVL, NULL, NULL) ;
MARKTIME(t2) ;
bridge->cpus[0] += t2 - t1 ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n CPU %8.3f : time to create Graph", t2 - t1) ;
fflush(msgFile) ;
}
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n graph of the input matrix") ;
Graph_writeForHumanEye(graph, msgFile) ;
fflush(msgFile) ;
}
/*
------------------
compress the graph
------------------
*/
MARKTIME(t1) ;
eqmapIV = Graph_equivMap(graph) ;
Neqns = bridge->Neqns = 1 + IV_max(eqmapIV) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n graph's equivalence map") ;
IV_writeForHumanEye(eqmapIV, msgFile) ;
fflush(msgFile) ;
}
if ( Neqns < bridge->compressCutoff * neqns ) {
Graph *cgraph ;
/*
------------------
compress the graph
------------------
*/
cgraph = Graph_compress2(graph, eqmapIV, 1) ;
Graph_free(graph) ;
graph = cgraph ;
compressed = 1 ;
bridge->Nedges = graph->nedges ;
} else {
compressed = 0 ;
}
MARKTIME(t2) ;
bridge->cpus[1] += t2 - t1 ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n CPU %8.3f : time to create compressed graph",
t2 - t1) ;
fflush(msgFile) ;
}
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n graph to order") ;
Graph_writeForHumanEye(graph, msgFile) ;
fflush(msgFile) ;
}
/*
---------------
order the graph
---------------
*/
MARKTIME(t1) ;
if ( bridge->maxdomainsize <= 0 ) {
bridge->maxdomainsize = neqns/32 ;
}
if ( bridge->maxdomainsize <= 0 ) {
bridge->maxdomainsize = 1 ;
}
if ( bridge->maxnzeros < 0 ) {
bridge->maxnzeros = 0.01*neqns ;
}
if ( bridge->maxsize < 0 ) {
bridge->maxsize = neqns ;
}
frontETree = orderViaBestOfNDandMS(graph, bridge->maxdomainsize,
bridge->maxnzeros, bridge->maxsize,
bridge->seed, msglvl, msgFile) ;
MARKTIME(t2) ;
bridge->cpus[2] += t2 - t1 ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n CPU %8.3f : time to order graph", t2 - t1) ;
fflush(msgFile) ;
}
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n front tree from ordering") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fflush(msgFile) ;
}
MARKTIME(t1) ;
if ( compressed == 1 ) {
ETree *etree ;
IVL *tempIVL ;
/*
----------------------------------------------------------
compute the symbolic factorization of the compressed graph
----------------------------------------------------------
*/
tempIVL = SymbFac_initFromGraph(frontETree, graph) ;
/*
-------------------------------------------------------
expand the symbolic factorization to the original graph
-------------------------------------------------------
*/
symbfacIVL = IVL_expand(tempIVL, eqmapIV) ;
IVL_free(tempIVL) ;
/*
---------------------
expand the front tree
---------------------
*/
etree = ETree_expand(frontETree, eqmapIV) ;
ETree_free(frontETree) ;
frontETree = etree ;
} else {
/*
--------------------------------------------------------
compute the symbolic factorization of the original graph
--------------------------------------------------------
*/
symbfacIVL = SymbFac_initFromGraph(frontETree, graph) ;
}
MARKTIME(t2) ;
bridge->frontETree = frontETree ;
bridge->symbfacIVL = symbfacIVL ;
/*
----------------------------------------------
get the old-to-new and new-to-old permutations
----------------------------------------------
*/
bridge->oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ;
bridge->newToOldIV = ETree_newToOldVtxPerm(frontETree) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n old-to-new permutation") ;
IV_writeForHumanEye(bridge->oldToNewIV, msgFile) ;
fprintf(msgFile, "\n\n new-to-old permutation") ;
IV_writeForHumanEye(bridge->newToOldIV, msgFile) ;
fflush(msgFile) ;
}
/*
------------------------------------------------------
overwrite the symbolic factorization with the permuted
indices and sort the lists into ascending order
------------------------------------------------------
*/
IVL_overwrite(symbfacIVL, bridge->oldToNewIV) ;
IVL_sortUp(symbfacIVL) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n symbolic factorization") ;
IVL_writeForHumanEye(symbfacIVL, msgFile) ;
fflush(msgFile) ;
}
/*
--------------------------------------
permute the vertices in the front tree
--------------------------------------
*/
ETree_permuteVertices(frontETree, bridge->oldToNewIV) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n permuted front etree") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fflush(msgFile) ;
}
MARKTIME(t2) ;
bridge->cpus[3] += t2 - t1 ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n CPU %8.3f : time for symbolic factorization",
t2 - t1) ;
fflush(msgFile) ;
}
/*
------------------------
free the working storage
------------------------
*/
Graph_free(graph) ;
IV_free(eqmapIV) ;
MARKTIME(t2) ;
bridge->cpus[4] += t2 - t0 ;
return(1) ;
}
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
-------------------------------------------------
this program tests the IVL_MPI_allgather() method
(1) each process generates the same owners[n] map
(2) each process creates an IVL object
and fills its owned lists with random numbers
(3) the processes gather-all's the lists of ivl
created -- 98apr03, cca
-------------------------------------------------
*/
{
char *buffer ;
double chksum, globalsum, t1, t2 ;
Drand drand ;
int ilist, length, myid, msglvl, nlist,
nproc, rc, seed, size, tag ;
int *list, *owners, *vec ;
int stats[4], tstats[4] ;
IV *ownersIV ;
IVL *ivl ;
FILE *msgFile ;
/*
---------------------------------------------------------------
find out the identity of this process and the number of process
---------------------------------------------------------------
*/
MPI_Init(&argc, &argv) ;
MPI_Comm_rank(MPI_COMM_WORLD, &myid) ;
MPI_Comm_size(MPI_COMM_WORLD, &nproc) ;
fprintf(stdout, "\n process %d of %d, argc = %d", myid, nproc, argc) ;
fflush(stdout) ;
if ( argc != 5 ) {
fprintf(stdout,
"\n\n usage : %s msglvl msgFile n seed "
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n nlist -- number of lists in the IVL object"
"\n seed -- random number seed"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else {
length = strlen(argv[2]) + 1 + 4 ;
buffer = CVinit(length, '\0') ;
sprintf(buffer, "%s.%d", argv[2], myid) ;
if ( (msgFile = fopen(buffer, "w")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
CVfree(buffer) ;
}
nlist = atoi(argv[3]) ;
seed = atoi(argv[4]) ;
fprintf(msgFile,
"\n %s "
"\n msglvl -- %d"
"\n msgFile -- %s"
"\n nlist -- %d"
"\n seed -- %d"
"\n",
argv[0], msglvl, argv[2], nlist, seed) ;
fflush(msgFile) ;
/*
----------------------------
generate the ownersIV object
----------------------------
*/
MARKTIME(t1) ;
ownersIV = IV_new() ;
IV_init(ownersIV, nlist, NULL) ;
owners = IV_entries(ownersIV) ;
Drand_setDefaultFields(&drand) ;
Drand_setSeed(&drand, seed) ;
Drand_setUniform(&drand, 0, nproc) ;
Drand_fillIvector(&drand, nlist, owners) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : initialize the ownersIV object",
t2 - t1) ;
fflush(msgFile) ;
fprintf(msgFile, "\n\n ownersIV generated") ;
if ( msglvl > 2 ) {
IV_writeForHumanEye(ownersIV, msgFile) ;
} else {
IV_writeStats(ownersIV, msgFile) ;
}
fflush(msgFile) ;
/*
--------------------------------------------
set up the IVL object and fill owned entries
--------------------------------------------
*/
MARKTIME(t1) ;
ivl = IVL_new() ;
IVL_init1(ivl, IVL_CHUNKED, nlist) ;
vec = IVinit(nlist, -1) ;
Drand_setSeed(&drand, seed + myid) ;
Drand_setUniform(&drand, 0, nlist) ;
for ( ilist = 0 ; ilist < nlist ; ilist++ ) {
if ( owners[ilist] == myid ) {
size = (int) Drand_value(&drand) ;
Drand_fillIvector(&drand, size, vec) ;
IVL_setList(ivl, ilist, size, vec) ;
}
}
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : initialize the IVL object",
t2 - t1) ;
fflush(msgFile) ;
if ( msglvl > 2 ) {
IVL_writeForHumanEye(ivl, msgFile) ;
} else {
IVL_writeStats(ivl, msgFile) ;
}
fflush(msgFile) ;
/*
--------------------------------------------
compute the local checksum of the ivl object
--------------------------------------------
*/
for ( ilist = 0, chksum = 0.0 ; ilist < nlist ; ilist++ ) {
if ( owners[ilist] == myid ) {
IVL_listAndSize(ivl, ilist, &size, &list) ;
chksum += 1 + ilist + size + IVsum(size, list) ;
}
}
fprintf(msgFile, "\n\n local partial chksum = %12.4e", chksum) ;
fflush(msgFile) ;
/*
-----------------------
get the global checksum
-----------------------
*/
rc = MPI_Allreduce((void *) &chksum, (void *) &globalsum,
1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD) ;
/*
--------------------------------
execute the all-gather operation
--------------------------------
*/
tag = 47 ;
IVzero(4, stats) ;
IVL_MPI_allgather(ivl, ownersIV,
stats, msglvl, msgFile, tag, MPI_COMM_WORLD) ;
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n return from IVL_MPI_allgather()") ;
fprintf(msgFile,
"\n local send stats : %10d messages with %10d bytes"
"\n local recv stats : %10d messages with %10d bytes",
stats[0], stats[2], stats[1], stats[3]) ;
fflush(msgFile) ;
}
MPI_Reduce((void *) stats, (void *) tstats, 4, MPI_INT,
MPI_SUM, 0, MPI_COMM_WORLD) ;
if ( myid == 0 ) {
fprintf(msgFile,
"\n total send stats : %10d messages with %10d bytes"
"\n total recv stats : %10d messages with %10d bytes",
tstats[0], tstats[2], tstats[1], tstats[3]) ;
fflush(msgFile) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n ivl") ;
IVL_writeForHumanEye(ivl, msgFile) ;
fflush(msgFile) ;
}
/*
-----------------------------------------
compute the checksum of the entire object
-----------------------------------------
*/
for ( ilist = 0, chksum = 0.0 ; ilist < nlist ; ilist++ ) {
IVL_listAndSize(ivl, ilist, &size, &list) ;
chksum += 1 + ilist + size + IVsum(size, list) ;
}
fprintf(msgFile,
"\n globalsum = %12.4e, chksum = %12.4e, error = %12.4e",
globalsum, chksum, fabs(globalsum - chksum)) ;
fflush(msgFile) ;
/*
----------------
free the objects
----------------
*/
IV_free(ownersIV) ;
IVL_free(ivl) ;
/*
------------------------
exit the MPI environment
------------------------
*/
MPI_Finalize() ;
fprintf(msgFile, "\n") ;
fclose(msgFile) ;
return(0) ; }
/*
--------------------------------------------------
clear the data fields, releasing allocated storage
created -- 98may04, cca
--------------------------------------------------
*/
void
FrontMtx_clearData (
FrontMtx *frontmtx
) {
SubMtx *mtx ;
int ii, J, K, nadj, nfront ;
int *adj ;
/*
---------------
check the input
---------------
*/
if ( frontmtx == NULL ) {
fprintf(stderr, "\n fatal error in FrontMtx_clearData(%p)"
"\n bad input\n", frontmtx) ;
exit(-1) ;
}
nfront = frontmtx->nfront ;
/*
----------------------
free the owned storage
----------------------
*/
if ( frontmtx->frontsizesIV != NULL ) {
IV_free(frontmtx->frontsizesIV) ;
}
if ( frontmtx->rowadjIVL != NULL ) {
IVL_free(frontmtx->rowadjIVL) ;
}
if ( frontmtx->coladjIVL != NULL ) {
IVL_free(frontmtx->coladjIVL) ;
}
if ( frontmtx->p_mtxDJJ != NULL ) {
for ( J = 0 ; J < nfront ; J++ ) {
if ( (mtx = frontmtx->p_mtxDJJ[J]) != NULL ) {
SubMtx_free(mtx) ;
}
}
FREE(frontmtx->p_mtxDJJ) ;
}
if ( frontmtx->tree != NULL ) {
if ( frontmtx->frontETree == NULL
|| frontmtx->frontETree->tree != frontmtx->tree ) {
Tree_free(frontmtx->tree) ;
}
frontmtx->tree = NULL ;
}
if ( frontmtx->dataMode == FRONTMTX_1D_MODE ) {
if ( frontmtx->p_mtxUJJ != NULL ) {
for ( J = 0 ; J < nfront ; J++ ) {
if ( (mtx = frontmtx->p_mtxUJJ[J]) != NULL ) {
SubMtx_free(mtx) ;
}
}
FREE(frontmtx->p_mtxUJJ) ;
}
if ( frontmtx->p_mtxUJN != NULL ) {
for ( J = 0 ; J < nfront ; J++ ) {
if ( (mtx = frontmtx->p_mtxUJN[J]) != NULL ) {
SubMtx_free(mtx) ;
}
}
FREE(frontmtx->p_mtxUJN) ;
}
if ( frontmtx->p_mtxLJJ != NULL ) {
for ( J = 0 ; J < nfront ; J++ ) {
if ( (mtx = frontmtx->p_mtxLJJ[J]) != NULL ) {
SubMtx_free(mtx) ;
}
}
FREE(frontmtx->p_mtxLJJ) ;
}
if ( frontmtx->p_mtxLNJ != NULL ) {
for ( J = 0 ; J < nfront ; J++ ) {
if ( (mtx = frontmtx->p_mtxLNJ[J]) != NULL ) {
SubMtx_free(mtx) ;
}
}
FREE(frontmtx->p_mtxLNJ) ;
}
} else if ( frontmtx->dataMode == FRONTMTX_2D_MODE ) {
for ( J = 0 ; J < nfront ; J++ ) {
FrontMtx_upperAdjFronts(frontmtx, J, &nadj, &adj) ;
for ( ii = 0 ; ii < nadj ; ii++ ) {
K = adj[ii] ;
if ( (mtx = FrontMtx_upperMtx(frontmtx, J, K)) != NULL ) {
SubMtx_free(mtx) ;
}
}
}
if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) {
for ( J = 0 ; J < nfront ; J++ ) {
FrontMtx_lowerAdjFronts(frontmtx, J, &nadj, &adj) ;
for ( ii = 0 ; ii < nadj ; ii++ ) {
K = adj[ii] ;
if ( (mtx = FrontMtx_lowerMtx(frontmtx, K, J)) != NULL ) {
SubMtx_free(mtx) ;
}
}
}
}
if ( frontmtx->lowerblockIVL != NULL ) {
IVL_free(frontmtx->lowerblockIVL) ;
}
if ( frontmtx->upperblockIVL != NULL ) {
IVL_free(frontmtx->upperblockIVL) ;
}
if ( frontmtx->lowerhash != NULL ) {
I2Ohash_free(frontmtx->lowerhash) ;
}
if ( frontmtx->upperhash != NULL ) {
I2Ohash_free(frontmtx->upperhash) ;
}
}
if ( frontmtx->lock != NULL ) {
/*
-------------------------
destroy and free the lock
-------------------------
*/
Lock_free(frontmtx->lock) ;
}
/*
----------------------
set the default fields
----------------------
*/
FrontMtx_setDefaultFields(frontmtx) ;
return ; }
/*
--------------------------------------------------------------------
fill *pndom with ndom, the number of domains.
fill *pnseg with nseg, the number of segments.
domains are numbered in [0, ndom), segments in [ndom,ndom+nseg).
return -- an IV object that contains the map
from vertices to segments
created -- 99feb29, cca
--------------------------------------------------------------------
*/
IV *
GPart_domSegMap (
GPart *gpart,
int *pndom,
int *pnseg
) {
FILE *msgFile ;
Graph *g ;
int adjdom, count, d, first, ierr, ii, jj1, jj2, last, ndom,
msglvl, nextphi, nPhi, nPsi, nV, phi, phi0, phi1, phi2, phi3,
psi, sigma, size, size0, size1, size2, v, vsize, w ;
int *adj, *adj0, *adj1, *adj2, *compids, *dmark, *dsmap, *head,
*link, *list, *offsets, *PhiToPsi, *PhiToV, *PsiToSigma,
*vadj, *VtoPhi ;
IV *dsmapIV ;
IVL *PhiByPhi, *PhiByPowD, *PsiByPowD ;
/*
--------------------
set the initial time
--------------------
*/
icputimes = 0 ;
MARKTIME(cputimes[icputimes]) ;
/*
---------------
check the input
---------------
*/
if ( gpart == NULL
|| (g = gpart->g) == NULL
|| pndom == NULL
|| pnseg == NULL ) {
fprintf(stderr, "\n fatal error in GPart_domSegMap(%p,%p,%p)"
"\n bad input\n", gpart, pndom, pnseg) ;
exit(-1) ;
}
compids = IV_entries(&gpart->compidsIV) ;
msglvl = gpart->msglvl ;
msgFile = gpart->msgFile ;
/*
------------------------
create the map IV object
------------------------
*/
nV = g->nvtx ;
dsmapIV = IV_new() ;
IV_init(dsmapIV, nV, NULL) ;
dsmap = IV_entries(dsmapIV) ;
/*
----------------------------------
check compids[] and get the number
of domains and interface vertices
----------------------------------
*/
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
ndom = nPhi = 0 ;
for ( v = 0 ; v < nV ; v++ ) {
if ( (d = compids[v]) < 0 ) {
fprintf(stderr,
"\n fatal error in GPart_domSegMap(%p,%p,%p)"
"\n compids[%d] = %d\n", gpart, pndom, pnseg,
v, compids[v]) ;
exit(-1) ;
} else if ( d == 0 ) {
nPhi++ ;
} else if ( ndom < d ) {
ndom = d ;
}
}
*pndom = ndom ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n Inside GPart_domSegMap") ;
fprintf(msgFile, "\n %d domains, %d Phi vertices", ndom, nPhi) ;
}
if ( ndom == 1 ) {
IVfill(nV, dsmap, 0) ;
*pndom = 1 ;
*pnseg = 0 ;
return(dsmapIV) ;
}
/*
--------------------------------
get the maps
PhiToV : [0,nPhi) |---> [0,nV)
VtoPhi : [0,nV) |---> [0,nPhi)
--------------------------------
*/
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
PhiToV = IVinit(nPhi, -1) ;
VtoPhi = IVinit(nV, -1) ;
for ( v = 0, phi = 0 ; v < nV ; v++ ) {
if ( (d = compids[v]) == 0 ) {
PhiToV[phi] = v ;
VtoPhi[v] = phi++ ;
}
}
if ( phi != nPhi ) {
fprintf(stderr,
"\n fatal error in GPart_domSegMap(%p,%p,%p)"
"\n phi = %d != %d = nPhi\n",
gpart, pndom, pnseg, phi, nPhi) ;
exit(-1) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n PhiToV(%d) :", nPhi) ;
IVfp80(msgFile, nPhi, PhiToV, 15, &ierr) ;
fflush(msgFile) ;
}
if ( msglvl > 3 ) {
fprintf(msgFile, "\n VtoPhi(%d) :", nV) ;
IVfp80(msgFile, nV, VtoPhi, 15, &ierr) ;
fflush(msgFile) ;
}
/*
---------------------------------------------------
create an IVL object, PhiByPowD, to hold lists from
the interface vertices to their adjacent domains
---------------------------------------------------
*/
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
dmark = IVinit(ndom+1, -1) ;
if ( nPhi >= ndom ) {
list = IVinit(nPhi, -1) ;
} else {
list = IVinit(ndom, -1) ;
}
PhiByPowD = IVL_new() ;
IVL_init1(PhiByPowD, IVL_CHUNKED, nPhi) ;
for ( phi = 0 ; phi < nPhi ; phi++ ) {
v = PhiToV[phi] ;
Graph_adjAndSize(g, v, &vsize, &vadj) ;
/*
if ( phi == 0 ) {
int ierr ;
fprintf(msgFile, "\n adj(%d,%d) = ", v, phi) ;
IVfp80(msgFile, vsize, vadj, 15, &ierr) ;
fflush(msgFile) ;
}
*/
count = 0 ;
for ( ii = 0 ; ii < vsize ; ii++ ) {
if ( (w = vadj[ii]) < nV
&& (d = compids[w]) > 0
&& dmark[d] != phi ) {
dmark[d] = phi ;
list[count++] = d ;
}
}
if ( count > 0 ) {
IVqsortUp(count, list) ;
IVL_setList(PhiByPowD, phi, count, list) ;
}
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n PhiByPowD : interface x adjacent domains") ;
IVL_writeForHumanEye(PhiByPowD, msgFile) ;
fflush(msgFile) ;
}
/*
-------------------------------------------------------
create an IVL object, PhiByPhi to hold lists
from the interface vertices to interface vertices.
(s,t) are in the list if (s,t) is an edge in the graph
and s and t do not share an adjacent domain
-------------------------------------------------------
*/
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
PhiByPhi = IVL_new() ;
IVL_init1(PhiByPhi, IVL_CHUNKED, nPhi) ;
offsets = IVinit(nPhi, 0) ;
head = IVinit(nPhi, -1) ;
link = IVinit(nPhi, -1) ;
for ( phi1 = 0 ; phi1 < nPhi ; phi1++ ) {
count = 0 ;
if ( msglvl > 2 ) {
v = PhiToV[phi1] ;
Graph_adjAndSize(g, v, &vsize, &vadj) ;
fprintf(msgFile, "\n checking out phi = %d, v = %d", phi1, v) ;
fprintf(msgFile, "\n adj(%d) : ", v) ;
IVfp80(msgFile, vsize, vadj, 10, &ierr) ;
}
/*
-------------------------------------------------------------
get (phi1, phi0) edges that were previously put into the list
-------------------------------------------------------------
*/
if ( msglvl > 3 ) {
if ( head[phi1] == -1 ) {
fprintf(msgFile, "\n no previous edges") ;
} else {
fprintf(msgFile, "\n previous edges :") ;
}
}
for ( phi0 = head[phi1] ; phi0 != -1 ; phi0 = nextphi ) {
if ( msglvl > 3 ) {
fprintf(msgFile, " %d", phi0) ;
}
nextphi = link[phi0] ;
list[count++] = phi0 ;
IVL_listAndSize(PhiByPhi, phi0, &size0, &adj0) ;
if ( (ii = ++offsets[phi0]) < size0 ) {
/*
----------------------------
link phi0 into the next list
----------------------------
*/
phi2 = adj0[ii] ;
link[phi0] = head[phi2] ;
head[phi2] = phi0 ;
}
}
/*
--------------------------
get new edges (phi1, phi2)
--------------------------
*/
IVL_listAndSize(PhiByPowD, phi1, &size1, &adj1) ;
v = PhiToV[phi1] ;
Graph_adjAndSize(g, v, &vsize, &vadj) ;
for ( ii = 0 ; ii < vsize ; ii++ ) {
if ( (w = vadj[ii]) < nV
&& compids[w] == 0
&& (phi2 = VtoPhi[w]) > phi1 ) {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n checking out phi2 = %d", phi2) ;
}
/*
--------------------------------------------------
see if phi1 and phi2 have a common adjacent domain
--------------------------------------------------
*/
IVL_listAndSize(PhiByPowD, phi2, &size2, &adj2) ;
adjdom = 0 ;
jj1 = jj2 = 0 ;
while ( jj1 < size1 && jj2 < size2 ) {
if ( adj1[jj1] < adj2[jj2] ) {
jj1++ ;
} else if ( adj1[jj1] > adj2[jj2] ) {
jj2++ ;
} else {
if ( msglvl > 3 ) {
fprintf(msgFile, ", common adj domain %d", adj1[jj1]) ;
}
adjdom = 1 ;
break ;
}
}
if ( adjdom == 0 ) {
if ( msglvl > 3 ) {
fprintf(msgFile, ", no adjacent domain") ;
}
list[count++] = phi2 ;
}
}
}
if ( count > 0 ) {
/*
---------------------
set the list for phi1
---------------------
*/
IVqsortUp(count, list) ;
IVL_setList(PhiByPhi, phi1, count, list) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n edge list for %d :", phi1) ;
IVfp80(msgFile, count, list, 15, &ierr) ;
}
for ( ii = 0, phi2 = -1 ; ii < count ; ii++ ) {
if ( list[ii] > phi1 ) {
offsets[phi1] = ii ;
phi2 = list[ii] ;
break ;
}
}
if ( phi2 != -1 ) {
if ( msglvl > 2 ) {
fprintf(msgFile,
"\n linking %d into list for %d", phi1, phi2) ;
}
link[phi1] = head[phi2] ;
head[phi2] = phi1 ;
}
/*
phi2 = list[0] ;
link[phi1] = head[phi2] ;
head[phi2] = phi1 ;
*/
}
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n PhiByPhi : interface x interface") ;
IVL_writeForHumanEye(PhiByPhi, msgFile) ;
fflush(msgFile) ;
}
/*
--------------------
get the PhiToPsi map
--------------------
*/
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
PhiToPsi = IVinit(nPhi, -1) ;
nPsi = 0 ;
for ( phi = 0 ; phi < nPhi ; phi++ ) {
if ( PhiToPsi[phi] == -1 ) {
/*
---------------------------
phi not yet mapped to a psi
---------------------------
*/
first = last = 0 ;
list[0] = phi ;
PhiToPsi[phi] = nPsi ;
while ( first <= last ) {
phi2 = list[first++] ;
IVL_listAndSize(PhiByPhi, phi2, &size, &adj) ;
for ( ii = 0 ; ii < size ; ii++ ) {
phi3 = adj[ii] ;
if ( PhiToPsi[phi3] == -1 ) {
PhiToPsi[phi3] = nPsi ;
list[++last] = phi3 ;
}
}
}
nPsi++ ;
}
}
if ( msglvl > 1 ) {
fprintf(msgFile, "\n nPsi = %d", nPsi) ;
fflush(msgFile) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n PhiToPsi(%d) :", nPhi) ;
IVfp80(msgFile, nPhi, PhiToPsi, 15, &ierr) ;
fflush(msgFile) ;
}
/*
---------------------------------
create an IVL object, Psi --> 2^D
---------------------------------
*/
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
IVfill(nPsi, head, -1) ;
IVfill(nPhi, link, -1) ;
for ( phi = 0 ; phi < nPhi ; phi++ ) {
psi = PhiToPsi[phi] ;
link[phi] = head[psi] ;
head[psi] = phi ;
}
PsiByPowD = IVL_new() ;
IVL_init1(PsiByPowD, IVL_CHUNKED, nPsi) ;
IVfill(ndom+1, dmark, -1) ;
for ( psi = 0 ; psi < nPsi ; psi++ ) {
count = 0 ;
for ( phi = head[psi] ; phi != -1 ; phi = link[phi] ) {
v = PhiToV[phi] ;
Graph_adjAndSize(g, v, &size, &adj) ;
for ( ii = 0 ; ii < size ; ii++ ) {
if ( (w = adj[ii]) < nV
&& (d = compids[w]) > 0
&& dmark[d] != psi ) {
dmark[d] = psi ;
list[count++] = d ;
}
}
}
if ( count > 0 ) {
IVqsortUp(count, list) ;
IVL_setList(PsiByPowD, psi, count, list) ;
}
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n PsiByPowD(%d) :", nPhi) ;
IVL_writeForHumanEye(PsiByPowD, msgFile) ;
fflush(msgFile) ;
}
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
/*
-------------------------------------
now get the map Psi |---> Sigma that
is the equivalence map over PhiByPowD
-------------------------------------
*/
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
PsiToSigma = IVL_equivMap1(PsiByPowD) ;
*pnseg = 1 + IVmax(nPsi, PsiToSigma, &ii) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n nSigma = %d", *pnseg) ;
fprintf(msgFile, "\n PsiToSigma(%d) :", nPhi) ;
IVfp80(msgFile, nPsi, PsiToSigma, 15, &ierr) ;
fflush(msgFile) ;
}
/*
--------------------------------------------------------------
now fill the map from the vertices to the domains and segments
--------------------------------------------------------------
*/
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
for ( v = 0 ; v < nV ; v++ ) {
if ( (d = compids[v]) > 0 ) {
dsmap[v] = d - 1 ;
} else {
phi = VtoPhi[v] ;
psi = PhiToPsi[phi] ;
sigma = PsiToSigma[psi] ;
dsmap[v] = ndom + sigma ;
}
}
/*
------------------------
free the working storage
------------------------
*/
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
IVL_free(PhiByPhi) ;
IVL_free(PhiByPowD) ;
IVL_free(PsiByPowD) ;
IVfree(PhiToV) ;
IVfree(VtoPhi) ;
IVfree(dmark) ;
IVfree(list) ;
IVfree(PhiToPsi) ;
IVfree(head) ;
IVfree(link) ;
IVfree(offsets) ;
IVfree(PsiToSigma) ;
icputimes++ ;
MARKTIME(cputimes[icputimes]) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n domain/segment map timings split") ;
fprintf(msgFile,
"\n %9.5f : create the DSmap object"
"\n %9.5f : get numbers of domain and interface vertices"
"\n %9.5f : generate PhiToV and VtoPhi"
"\n %9.5f : generate PhiByPowD"
"\n %9.5f : generate PhiByPhi"
"\n %9.5f : generate PhiToPsi"
"\n %9.5f : generate PsiByPowD"
"\n %9.5f : generate PsiToSigma"
"\n %9.5f : generate dsmap"
"\n %9.5f : free working storage"
"\n %9.5f : total time",
cputimes[1] - cputimes[0],
cputimes[2] - cputimes[1],
cputimes[3] - cputimes[2],
cputimes[4] - cputimes[3],
cputimes[5] - cputimes[4],
cputimes[6] - cputimes[5],
cputimes[7] - cputimes[6],
cputimes[8] - cputimes[7],
cputimes[9] - cputimes[8],
cputimes[10] - cputimes[9],
cputimes[11] - cputimes[0]) ;
}
return(dsmapIV) ; }
/*
------------------------------------------------------------
purpose -- compute a QR factorization using multiple threads
created -- 98may29, cca
------------------------------------------------------------
*/
void
FrontMtx_MT_QR_factor (
FrontMtx *frontmtx,
InpMtx *mtxA,
ChvManager *chvmanager,
IV *ownersIV,
double cpus[],
double *pfacops,
int msglvl,
FILE *msgFile
) {
ChvList *updlist ;
double t0, t1 ;
IVL *rowsIVL ;
int ithread, myid, nthread, rc ;
int *firstnz ;
QR_factorData *data, *dataObjects ;
/*
---------------
check the input
---------------
*/
if ( frontmtx == NULL || mtxA == NULL || chvmanager == NULL
|| ownersIV == NULL || cpus == NULL || pfacops == NULL
|| (msglvl > 0 && msgFile == NULL) ) {
fprintf(stderr, "\n fatal error in FrontMtx_MT_QR_factor()"
"\n bad input\n") ;
exit(-1) ;
}
nthread = 1 + IV_max(ownersIV) ;
/*
----------------------------------------------------------------
create the update Chv list object
create the rowsIVL object, where
list(J) = list of rows that are assembled in front J
firstnz[irowA] = first column with nonzero element in A(irowA,*)
----------------------------------------------------------------
*/
MARKTIME(t0) ;
updlist = FrontMtx_postList(frontmtx, ownersIV, LOCK_IN_PROCESS) ;
FrontMtx_QR_setup(frontmtx, mtxA, &rowsIVL, &firstnz, msglvl, msgFile) ;
MARKTIME(t1) ;
cpus[0] = t1 - t0 ;
/*
------------------------------------
create and load nthread data objects
------------------------------------
*/
ALLOCATE(dataObjects, struct _QR_factorData, nthread) ;
for ( myid = 0, data = dataObjects ; myid < nthread ; myid++, data++ ) {
data->mtxA = mtxA ;
data->rowsIVL = rowsIVL ;
data->firstnz = firstnz ;
data->ownersIV = ownersIV ;
data->frontmtx = frontmtx ;
data->chvmanager = chvmanager ;
data->updlist = updlist ;
data->myid = myid ;
DVzero(7, data->cpus) ;
data->facops = 0.0 ;
data->msglvl = msglvl ;
if ( msglvl > 0 ) {
char buffer[20] ;
sprintf(buffer, "res.%d", myid) ;
if ( (data->msgFile = fopen(buffer, "w")) == NULL ) {
fprintf(stderr, "\n fatal error in FrontMtx_MT_QR_factor()"
"\n unable to open file %s", buffer) ;
exit(-1) ;
}
} else {
data->msgFile = NULL ;
}
}
#if THREAD_TYPE == TT_SOLARIS
/*
----------------------------------
Solaris threads.
(1) set the concurrency
(2) create nthread - 1 new threads
(3) execute own thread
(4) join the threads
----------------------------------
*/
thr_setconcurrency(nthread) ;
for ( myid = 0, data = dataObjects ;
myid < nthread - 1 ;
myid++, data++ ) {
rc = thr_create(NULL, 0, FrontMtx_QR_workerFactor, data, 0, NULL) ;
if ( rc != 0 ) {
fprintf(stderr,
"\n fatal error, myid = %d, rc = %d from thr_create()",
myid, rc) ;
exit(-1) ;
}
}
FrontMtx_QR_workerFactor(data) ;
for ( myid = 0 ; myid < nthread - 1 ; myid++ ) {
thr_join(0, 0, 0) ;
}
#endif
#if THREAD_TYPE == TT_POSIX
/*
----------------------------------
POSIX threads.
(1) if SGI, set the concurrency
(2) create nthread new threads
(3) join the threads
----------------------------------
*/
{
pthread_t *tids ;
pthread_attr_t attr ;
void *status ;
/*
---------------------------------------------------------
#### NOTE: for SGI machines, this command must be present
#### for the thread scheduling to be efficient.
#### this is NOT a POSIX call, but necessary
---------------------------------------------------------
pthread_setconcurrency(nthread) ;
*/
pthread_attr_init(&attr) ;
/*
pthread_attr_setscope(&attr, PTHREAD_SCOPE_SYSTEM) ;
*/
pthread_attr_setscope(&attr, PTHREAD_SCOPE_PROCESS) ;
ALLOCATE(tids, pthread_t, nthread) ;
for ( myid = 0 ; myid < nthread ; myid++ ) {
#ifdef _MSC_VER
tids[myid].p = 0;
tids[myid].x = 0;
#else
tids[myid] = 0 ;
#endif
}
for ( myid = 0, data = dataObjects ;
myid < nthread ;
myid++, data++ ) {
rc = pthread_create(&tids[myid], &attr,
FrontMtx_QR_workerFactor, data) ;
if ( rc != 0 ) {
fprintf(stderr,
"\n fatal error in FrontMtx_MT_QR_factor()"
"\n myid = %d, rc = %d from pthread_create()",
myid, rc) ;
exit(-1) ;
} else if ( msglvl > 2 ) {
fprintf(stderr, "\n thread %d created", myid) ;
}
}
for ( myid = 0 ; myid < nthread ; myid++ ) {
pthread_join(tids[myid], &status) ;
}
FREE(tids) ;
pthread_attr_destroy(&attr) ;
}
#endif
/*
----------------------------------------------
fill the cpu vector and factor operation count
----------------------------------------------
*/
*pfacops = 0 ;
for ( myid = 0, data = dataObjects ; myid < nthread ; myid++, data++ ) {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n thread %d cpus", myid) ;
DVfprintf(msgFile, 7, data->cpus) ;
}
for ( ithread = 0 ; ithread < 7 ; ithread++ ) {
cpus[ithread] += data->cpus[ithread] ;
}
*pfacops += data->facops ;
}
/*
-------------
free the data
-------------
*/
ChvList_free(updlist) ;
IVL_free(rowsIVL) ;
IVfree(firstnz) ;
FREE(dataObjects) ;
return ;
}
/*--------------------------------------------------------------------*/
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) ; }
/*--------------------------------------------------------------------*/
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) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
------------------------------------------------------
(1) read in an ETree object.
(2) read in an Graph object.
(3) find the optimal domain/schur complement partition
for a semi-implicit factorization
created -- 96oct03, cca
------------------------------------------------------
*/
{
char *inETreeFileName, *inGraphFileName, *outIVfileName ;
double alpha, nA21, nfent1, nfops1, nL11, nL22, nPhi, nV, t1, t2 ;
Graph *graph ;
int ii, inside, J, K, msglvl, nfind1, nfront, nJ, nleaves1,
nnode1, nvtx, rc, sizeJ, totalgain, vsize, v, w ;
int *adjJ, *compids, *nodwghts, *vadj, *vtxToFront, *vwghts ;
IV *compidsIV ;
IVL *symbfacIVL ;
ETree *etree ;
FILE *msgFile ;
Tree *tree ;
if ( argc != 7 ) {
fprintf(stdout,
"\n\n usage : %s msglvl msgFile inETreeFile inGraphFile alpha"
"\n outIVfile "
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n inETreeFile -- input file, must be *.etreef or *.etreeb"
"\n inGraphFile -- input file, must be *.graphf or *.graphb"
"\n alpha -- weight parameter"
"\n alpha = 0 --> minimize storage"
"\n alpha = 1 --> minimize solve ops"
"\n outIVfile -- output file for oldToNew vector,"
"\n must be *.ivf or *.ivb"
"\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) ;
}
inETreeFileName = argv[3] ;
inGraphFileName = argv[4] ;
alpha = atof(argv[5]) ;
outIVfileName = argv[6] ;
fprintf(msgFile,
"\n %s "
"\n msglvl -- %d"
"\n msgFile -- %s"
"\n inETreeFile -- %s"
"\n inGraphFile -- %s"
"\n alpha -- %f"
"\n outIVfile -- %s"
"\n",
argv[0], msglvl, argv[2],
inETreeFileName, inGraphFileName, alpha, outIVfileName) ;
fflush(msgFile) ;
/*
------------------------
read in the ETree object
------------------------
*/
if ( strcmp(inETreeFileName, "none") == 0 ) {
fprintf(msgFile, "\n no file to read from") ;
spoolesFatal();
}
etree = ETree_new() ;
MARKTIME(t1) ;
rc = ETree_readFromFile(etree, inETreeFileName) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %9.5f : read in etree from file %s",
t2 - t1, inETreeFileName) ;
if ( rc != 1 ) {
fprintf(msgFile, "\n return value %d from ETree_readFromFile(%p,%s)",
rc, etree, inETreeFileName) ;
spoolesFatal();
}
ETree_leftJustify(etree) ;
fprintf(msgFile, "\n\n after reading ETree object from file %s",
inETreeFileName) ;
if ( msglvl > 2 ) {
ETree_writeForHumanEye(etree, msgFile) ;
} else {
ETree_writeStats(etree, msgFile) ;
}
fflush(msgFile) ;
nfront = ETree_nfront(etree) ;
tree = ETree_tree(etree) ;
nodwghts = ETree_nodwghts(etree) ;
vtxToFront = ETree_vtxToFront(etree) ;
/*
------------------------
read in the Graph object
------------------------
*/
if ( strcmp(inGraphFileName, "none") == 0 ) {
fprintf(msgFile, "\n no file to read from") ;
spoolesFatal();
}
graph = Graph_new() ;
MARKTIME(t1) ;
rc = Graph_readFromFile(graph, inGraphFileName) ;
nvtx = graph->nvtx ;
vwghts = graph->vwghts ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %9.5f : read in graph from file %s",
t2 - t1, inGraphFileName) ;
if ( rc != 1 ) {
fprintf(msgFile, "\n return value %d from Graph_readFromFile(%p,%s)",
rc, graph, inGraphFileName) ;
spoolesFatal();
}
fprintf(msgFile, "\n\n after reading Graph object from file %s",
inGraphFileName) ;
if ( msglvl > 2 ) {
Graph_writeForHumanEye(graph, msgFile) ;
} else {
Graph_writeStats(graph, msgFile) ;
}
fflush(msgFile) ;
/*
----------------------
compute the statistics
----------------------
*/
nnode1 = etree->tree->n ;
nfind1 = ETree_nFactorIndices(etree) ;
nfent1 = ETree_nFactorEntries(etree, SPOOLES_SYMMETRIC) ;
nfops1 = ETree_nFactorOps(etree, SPOOLES_REAL, SPOOLES_SYMMETRIC) ;
nleaves1 = Tree_nleaves(etree->tree) ;
fprintf(stdout, "\n root front %d has %d vertices",
etree->tree->root,
etree->nodwghtsIV->vec[etree->tree->root]) ;
/*
---------------------------------
create the symbolic factorization
---------------------------------
*/
symbfacIVL = SymbFac_initFromGraph(etree, graph) ;
if ( msglvl > 2 ) {
IVL_writeForHumanEye(symbfacIVL, msgFile) ;
} else {
IVL_writeStats(symbfacIVL, msgFile) ;
}
fflush(msgFile) ;
/*
--------------------------
find the optimal partition
--------------------------
*/
compidsIV = ETree_optPart(etree, graph, symbfacIVL, alpha,
&totalgain, msglvl, msgFile) ;
if ( msglvl > 2 ) {
IV_writeForHumanEye(compidsIV, msgFile) ;
} else {
IV_writeStats(compidsIV, msgFile) ;
}
fflush(msgFile) ;
compids = IV_entries(compidsIV) ;
/*
------------------------------------------------------
compute the number of vertices in the schur complement
------------------------------------------------------
*/
for ( J = 0, nPhi = nV = 0. ; J < nfront ; J++ ) {
if ( compids[J] == 0 ) {
nPhi += nodwghts[J] ;
}
nV += nodwghts[J] ;
}
/*
--------------------------------------------
compute the number of entries in L11 and L22
--------------------------------------------
*/
nL11 = nL22 = 0 ;
for ( J = Tree_postOTfirst(tree) ;
J != -1 ;
J = Tree_postOTnext(tree, J) ) {
nJ = nodwghts[J] ;
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n front %d, nJ = %d", J, nJ) ;
}
IVL_listAndSize(symbfacIVL, J, &sizeJ, &adjJ) ;
for ( ii = 0, inside = 0 ; ii < sizeJ ; ii++ ) {
w = adjJ[ii] ;
K = vtxToFront[w] ;
if ( msglvl > 3 ) {
fprintf(msgFile, "\n w = %d, K = %d", w, K) ;
}
if ( K > J && compids[K] == compids[J] ) {
inside += (vwghts == NULL) ? 1 : vwghts[w] ;
if ( msglvl > 3 ) {
fprintf(msgFile, ", inside") ;
}
}
}
if ( compids[J] != 0 ) {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n inside = %d, adding %d to L11",
inside, nJ*nJ + 2*nJ*inside) ;
}
nL11 += (nJ*(nJ+1))/2 + nJ*inside ;
} else {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n inside = %d, adding %d to L22",
inside, (nJ*(nJ+1))/2 + nJ*inside) ;
}
nL22 += (nJ*(nJ+1))/2 + nJ*inside ;
}
}
if ( msglvl > 0 ) {
fprintf(msgFile, "\n |L| = %.0f, |L11| = %.0f, |L22| = %.0f",
nfent1, nL11, nL22) ;
}
/*
------------------------------------
compute the number of entries in A21
------------------------------------
*/
nA21 = 0 ;
if ( vwghts != NULL ) {
for ( v = 0 ; v < nvtx ; v++ ) {
J = vtxToFront[v] ;
if ( compids[J] != 0 ) {
Graph_adjAndSize(graph, v, &vsize, &vadj) ;
for ( ii = 0 ; ii < vsize ; ii++ ) {
w = vadj[ii] ;
K = vtxToFront[w] ;
if ( compids[K] == 0 ) {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n A21 : v = %d, w = %d", v, w) ;
}
nA21 += vwghts[v] * vwghts[w] ;
}
}
}
}
} else {
for ( v = 0 ; v < nvtx ; v++ ) {
J = vtxToFront[v] ;
if ( compids[J] != 0 ) {
Graph_adjAndSize(graph, v, &vsize, &vadj) ;
for ( ii = 0 ; ii < vsize ; ii++ ) {
w = vadj[ii] ;
K = vtxToFront[w] ;
if ( compids[K] == 0 ) {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n A21 : v = %d, w = %d", v, w) ;
}
nA21++ ;
}
}
}
}
}
if ( msglvl > 0 ) {
fprintf(msgFile,
"\n |L| = %.0f, |L11| = %.0f, |L22| = %.0f, |A21| = %.0f",
nfent1, nL11, nL22, nA21) ;
fprintf(msgFile,
"\n storage: explicit = %.0f, semi-implicit = %.0f, ratio = %.3f"
"\n opcount: explicit = %.0f, semi-implicit = %.0f, ratio = %.3f",
nfent1, nL11 + nA21 + nL22,
nfent1/(nL11 + nA21 + nL22),
2*nfent1, 4*nL11 + 2*nA21 + 2*nL22,
2*nfent1/(4*nL11 + 2*nA21 + 2*nL22)) ;
fprintf(msgFile, "\n ratios %8.3f %8.3f %8.3f",
nPhi/nV,
nfent1/(nL11 + nA21 + nL22),
2*nfent1/(4*nL11 + 2*nA21 + 2*nL22)) ;
}
/*
----------------
free the objects
----------------
*/
ETree_free(etree) ;
Graph_free(graph) ;
IVL_free(symbfacIVL) ;
fprintf(msgFile, "\n") ;
fclose(msgFile) ;
return(1) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
----------------------------------------
get statistics for a semi-implicit solve
created -- 97dec11, cca
----------------------------------------
*/
{
char *inGraphFileName, *inETreeFileName, *inMapFileName ;
double nA21, nL, nL11, nL22, nPhi, nV, t1, t2 ;
ETree *etree ;
int ii, inside, J, K, msglvl, nfront, nJ,
nvtx, rc, sizeJ, v, vsize, w ;
int *adjJ, *frontmap, *map, *nodwghts,
*vadj, *vtxToFront, *vwghts ;
IV *mapIV ;
IVL *symbfacIVL ;
Graph *graph ;
FILE *msgFile ;
Tree *tree ;
if ( argc != 6 ) {
fprintf(stdout,
"\n\n usage : %s msglvl msgFile GraphFile ETreeFile mapFile "
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n GraphFile -- input graph file, must be *.graphf or *.graphb"
"\n ETreeFile -- input ETree file, must be *.etreef or *.etreeb"
"\n mapFile -- input map IV file, must be *.ivf or *.ivb"
"\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) ;
}
inGraphFileName = argv[3] ;
inETreeFileName = argv[4] ;
inMapFileName = argv[5] ;
fprintf(msgFile,
"\n %s "
"\n msglvl -- %d"
"\n msgFile -- %s"
"\n GraphFile -- %s"
"\n ETreeFile -- %s"
"\n mapFile -- %s"
"\n",
argv[0], msglvl, argv[2],
inGraphFileName, inETreeFileName, inMapFileName) ;
fflush(msgFile) ;
/*
------------------------
read in the Graph object
------------------------
*/
graph = Graph_new() ;
if ( strcmp(inGraphFileName, "none") == 0 ) {
fprintf(msgFile, "\n no file to read from") ;
exit(0) ;
}
MARKTIME(t1) ;
rc = Graph_readFromFile(graph, inGraphFileName) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %9.5f : read in graph from file %s",
t2 - t1, inGraphFileName) ;
nvtx = graph->nvtx ;
vwghts = graph->vwghts ;
if ( rc != 1 ) {
fprintf(msgFile, "\n return value %d from Graph_readFromFile(%p,%s)",
rc, graph, inGraphFileName) ;
exit(-1) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n after reading Graph object from file %s",
inGraphFileName) ;
Graph_writeForHumanEye(graph, msgFile) ;
fflush(msgFile) ;
}
/*
------------------------
read in the ETree object
------------------------
*/
etree = ETree_new() ;
if ( strcmp(inETreeFileName, "none") == 0 ) {
fprintf(msgFile, "\n no file to read from") ;
exit(0) ;
}
MARKTIME(t1) ;
rc = ETree_readFromFile(etree, inETreeFileName) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %9.5f : read in etree from file %s",
t2 - t1, inETreeFileName) ;
nfront = ETree_nfront(etree) ;
tree = ETree_tree(etree) ;
vtxToFront = ETree_vtxToFront(etree) ;
nodwghts = ETree_nodwghts(etree) ;
nL = ETree_nFactorEntries(etree, 2) ;
if ( rc != 1 ) {
fprintf(msgFile, "\n return value %d from ETree_readFromFile(%p,%s)",
rc, etree, inETreeFileName) ;
exit(-1) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n after reading ETree object from file %s",
inETreeFileName) ;
ETree_writeForHumanEye(etree, msgFile) ;
fflush(msgFile) ;
}
/*
-------------------------
read in the map IV object
-------------------------
*/
mapIV = IV_new() ;
if ( strcmp(inMapFileName, "none") == 0 ) {
fprintf(msgFile, "\n no file to read from") ;
exit(0) ;
}
MARKTIME(t1) ;
rc = IV_readFromFile(mapIV, inMapFileName) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %9.5f : read in mapIV from file %s",
t2 - t1, inMapFileName) ;
map = IV_entries(mapIV) ;
if ( rc != 1 ) {
fprintf(msgFile, "\n return value %d from IV_readFromFile(%p,%s)",
rc, mapIV, inMapFileName) ;
exit(-1) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n after reading IV object from file %s",
inMapFileName) ;
IV_writeForHumanEye(mapIV, msgFile) ;
fflush(msgFile) ;
}
nV = nPhi = 0 ;
if ( vwghts == NULL ) {
for ( v = 0 ; v < nvtx ; v++ ) {
nV++ ;
if ( map[v] == 0 ) {
nPhi++ ;
}
}
} else {
for ( v = 0 ; v < nvtx ; v++ ) {
nV += vwghts[v] ;
if ( map[v] == 0 ) {
nPhi += vwghts[v] ;
}
}
}
fprintf(msgFile, "\n nPhi = %.0f, nV = %.0f", nPhi, nV) ;
/*
-------------------------
get the frontmap[] vector
-------------------------
*/
frontmap = IVinit(nfront, -1) ;
for ( v = 0 ; v < nvtx ; v++ ) {
J = vtxToFront[v] ;
if ( frontmap[J] == -1 ) {
frontmap[J] = map[v] ;
} else if ( frontmap[J] != map[v] ) {
fprintf(msgFile, "\n\n error, frontmap[%d] = %d, map[%d] = %d",
J, frontmap[J], v, map[v]) ;
}
}
/*
----------------------------------
compute the symbolic factorization
----------------------------------
*/
symbfacIVL = SymbFac_initFromGraph(etree, graph) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n symbolic factorization") ;
IVL_writeForHumanEye(symbfacIVL, msgFile) ;
fflush(msgFile) ;
}
/*
--------------------------------------------
compute the number of entries in L11 and L22
--------------------------------------------
*/
nL11 = nL22 = 0 ;
for ( J = Tree_postOTfirst(tree) ;
J != -1 ;
J = Tree_postOTnext(tree, J) ) {
nJ = nodwghts[J] ;
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n front %d, nJ = %d", J, nJ) ;
}
IVL_listAndSize(symbfacIVL, J, &sizeJ, &adjJ) ;
for ( ii = 0, inside = 0 ; ii < sizeJ ; ii++ ) {
w = adjJ[ii] ;
K = vtxToFront[w] ;
if ( msglvl > 3 ) {
fprintf(msgFile, "\n w = %d, K = %d", w, K) ;
}
if ( K > J && frontmap[K] == frontmap[J] ) {
inside += (vwghts == NULL) ? 1 : vwghts[w] ;
if ( msglvl > 3 ) {
fprintf(msgFile, ", inside") ;
}
}
}
if ( frontmap[J] != 0 ) {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n inside = %d, adding %d to L11",
inside, nJ*nJ + 2*nJ*inside) ;
}
nL11 += nJ*nJ + 2*nJ*inside ;
} else {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n inside = %d, adding %d to L22",
inside, nJ*nJ + 2*nJ*inside) ;
}
nL22 += nJ*nJ + 2*nJ*inside ;
}
}
if ( msglvl > 0 ) {
fprintf(msgFile, "\n |L| = %.0f, |L11| = %.0f, |L22| = %.0f",
nL, nL11, nL22) ;
}
/*
------------------------------------
compute the number of entries in A21
------------------------------------
*/
nA21 = 0 ;
if ( vwghts != NULL ) {
for ( v = 0 ; v < nvtx ; v++ ) {
if ( map[v] == 0 ) {
Graph_adjAndSize(graph, v, &vsize, &vadj) ;
for ( ii = 0 ; ii < vsize ; ii++ ) {
w = vadj[ii] ;
if ( map[v] != map[w] ) {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n A21 : v = %d, w = %d", v, w) ;
}
nA21 += vwghts[v] * vwghts[w] ;
}
}
}
}
} else {
for ( v = 0 ; v < nvtx ; v++ ) {
if ( map[v] == 0 ) {
Graph_adjAndSize(graph, v, &vsize, &vadj) ;
for ( ii = 0 ; ii < vsize ; ii++ ) {
w = vadj[ii] ;
if ( map[v] != map[w] ) {
if ( msglvl > 3 ) {
fprintf(msgFile, "\n A21 : v = %d, w = %d", v, w) ;
}
nA21++ ;
}
}
}
}
}
if ( msglvl > 0 ) {
fprintf(msgFile,
"\n |L| = %.0f, |L11| = %.0f, |L22| = %.0f, |A21| = %.0f",
nL, nL11, nL22, nA21) ;
fprintf(msgFile,
"\n storage: explicit = %.0f, semi-implicit = %.0f, ratio = %.3f"
"\n opcount: explicit = %.0f, semi-implicit = %.0f, ratio = %.3f",
nL, nL11 + nA21 + nL22,
nL/(nL11 + nA21 + nL22),
2*nL, 4*nL11 + 2*nA21 + 2*nL22,
2*nL/(4*nL11 + 2*nA21 + 2*nL22)) ;
fprintf(msgFile, "\n ratios %8.3f %8.3f %8.3f",
nPhi/nV,
nL/(nL11 + nA21 + nL22),
2*nL/(4*nL11 + 2*nA21 + 2*nL22)) ;
}
/*
------------------------
free the working storage
------------------------
*/
Graph_free(graph) ;
ETree_free(etree) ;
IV_free(mapIV) ;
IVL_free(symbfacIVL) ;
IVfree(frontmap) ;
fprintf(msgFile, "\n") ;
fclose(msgFile) ;
return(1) ; }