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);
}
SPOOLESSolverMT::~SPOOLESSolverMT()
{
BridgeMT_free((BridgeMT*) bridgeMTPointer);
InpMtx_free((InpMtx *) APointer);
DenseMtx_free((DenseMtx *) mtx_rhsPointer);
DenseMtx_free((DenseMtx *) mtx_xPointer);
fclose(msgFile);
}
void spooles_solve(void *ptr, double *b, long neq)
{
/* rhs vector B
* Note that there is only one rhs vector, thus
* a bit simpler that the AllInOne example
*/
long size = neq;
DenseMtx *mtxB,*mtxX;
struct factorinfo *pfi_ = ptr;
//printf(" Solving the system of equations using the symmetric spooles solver\n");
{
int i;
mtxB = DenseMtx_new();
DenseMtx_init(mtxB, SPOOLES_REAL, 0, 0, size, 1, 1, size);
DenseMtx_zero(mtxB);
for (i = 0; i < size; i++) {
DenseMtx_setRealEntry(mtxB, i, 0, b[i]);
}
if (DEBUG_LVL > 1) {
fprintf(msgFile, "\n\n rhs matrix in original ordering");
DenseMtx_writeForHumanEye(mtxB, msgFile);
fflush(msgFile);
}
}
#ifdef USE_MT
//printf(" Using up to %d cpu(s) for spooles.\n\n", num_cpus);
if (num_cpus > 1) {
/* do not use the multithreaded solver unless
* we have multiple threads - avoid the
* locking overhead
*/
mtxX=fsolve_MT(pfi_, mtxB);
} else {
mtxX=fsolve(pfi_, mtxB);
}
#else
//printf(" Using 1 cpu for spooles.\n\n");
mtxX=fsolve(pfi_, mtxB);
#endif
/* convert the result back to Calculix representation */
{
int i;
for (i = 0; i < size; i++) {
b[i] = DenseMtx_entries(mtxX)[i];
}
}
/* cleanup */
DenseMtx_free(mtxX);
}
DenseMtx *fsolve(struct factorinfo *pfi, DenseMtx *mtxB)
{
DenseMtx *mtxX;
/*
* STEP 6: permute the right hand side into the new ordering
*/
{
DenseMtx_permuteRows(mtxB, pfi->oldToNewIV);
if (DEBUG_LVL > 1) {
fprintf(pfi->msgFile,
"\n\n right hand side matrix in new ordering");
DenseMtx_writeForHumanEye(mtxB, pfi->msgFile);
fflush(pfi->msgFile);
}
}
/*
* STEP 7: solve the linear system
*/
{
mtxX = DenseMtx_new();
DenseMtx_init(mtxX, SPOOLES_REAL, 0, 0, pfi->size, 1, 1, pfi->size);
DenseMtx_zero(mtxX);
FrontMtx_solve(pfi->frontmtx, mtxX, mtxB, pfi->mtxmanager, pfi->cpus,
DEBUG_LVL, pfi->msgFile);
if (DEBUG_LVL > 1) {
fprintf(pfi->msgFile, "\n\n solution matrix in new ordering");
DenseMtx_writeForHumanEye(mtxX, pfi->msgFile);
fflush(pfi->msgFile);
}
}
/*
* STEP 8: permute the solution into the original ordering
*/
ssolve_permuteout(mtxX, pfi->newToOldIV, pfi->msgFile);
/* cleanup: */
DenseMtx_free(mtxB);
return mtxX;
}
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) ; }
/*
-------------------------------------------------------------
purpose --- to compute a matrix-vector multiply y[] = C * x[]
where C is the identity, A or B (depending on *pprbtype).
*pnrows -- # of rows in x[]
*pncols -- # of columns in x[]
*pprbtype -- problem type
*pprbtype = 1 --> vibration problem, matrix is A
*pprbtype = 2 --> buckling problem, matrix is B
*pprbtype = 3 --> matrix is identity, y[] = x[]
x[] -- vector to be multiplied
NOTE: the x[] vector is global, not a portion
y[] -- product vector
NOTE: the y[] vector is global, not a portion
created -- 98aug28, cca & jcp
-------------------------------------------------------------
*/
void
JimMatMulMPI (
int *pnrows,
int *pncols,
double x[],
double y[],
int *pprbtype,
void *data
) {
BridgeMPI *bridge = (BridgeMPI *) data ;
int ncols, nent, nrows ;
#if MYDEBUG > 0
double t1, t2 ;
count_JimMatMul++ ;
MARKTIME(t1) ;
if ( bridge->myid == 0 ) {
fprintf(stdout, "\n (%d) JimMatMulMPI() start", count_JimMatMul) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile,
"\n (%d) JimMatMulMPI() start", count_JimMatMul) ;
fflush(bridge->msgFile) ;
#endif
nrows = *pnrows ;
ncols = *pncols ;
nent = nrows*ncols ;
if ( *pprbtype == 3 ) {
/*
--------------------------
... matrix is the identity
--------------------------
*/
DVcopy(nent, y, x) ;
} else {
BridgeMPI *bridge = (BridgeMPI *) data ;
DenseMtx *mtx, *newmtx ;
int irow, jcol, jj, kk, myid, neqns, nowned, tag = 0 ;
int *vtxmap ;
int stats[4] ;
IV *mapIV ;
/*
---------------------------------------------
slide the owned rows of x[] down in the array
---------------------------------------------
*/
vtxmap = IV_entries(bridge->vtxmapIV) ;
neqns = bridge->neqns ;
myid = bridge->myid ;
nowned = IV_size(bridge->myownedIV) ;
for ( jcol = jj = kk = 0 ; jcol < ncols ; jcol++ ) {
for ( irow = 0 ; irow < neqns ; irow++, jj++ ) {
if ( vtxmap[irow] == myid ) {
y[kk++] = x[jj] ;
}
}
}
if ( kk != nowned * ncols ) {
fprintf(stderr, "\n proc %d : kk %d, nowned %d, ncols %d",
myid, kk, nowned, ncols) ;
exit(-1) ;
}
/*
----------------------------------------
call the method that assumes local input
----------------------------------------
*/
if ( bridge->msglvl > 2 ) {
fprintf(bridge->msgFile,
"\n inside JimMatMulMPI, calling MatMulMpi"
"\n prbtype %d, nrows %d, ncols %d, nowned %d",
*pprbtype, *pnrows, *pncols, nowned) ;
fflush(bridge->msgFile) ;
}
MatMulMPI(&nowned, pncols, y, y, pprbtype, data) ;
/*
-------------------------------------------------
gather all the entries of y[] onto processor zero
-------------------------------------------------
*/
mtx = DenseMtx_new() ;
DenseMtx_init(mtx, SPOOLES_REAL, 0, 0, nowned, ncols, 1, nowned) ;
DVcopy (nowned*ncols, DenseMtx_entries(mtx), y) ;
IVcopy(nowned, mtx->rowind, IV_entries(bridge->myownedIV)) ;
mapIV = IV_new() ;
IV_init(mapIV, neqns, NULL) ;
IV_fill(mapIV, 0) ;
IVfill(4, stats, 0) ;
if ( bridge->msglvl > 2 ) {
fprintf(bridge->msgFile, "\n mtx: %d rows x %d columns",
mtx->nrow, mtx->ncol) ;
fflush(bridge->msgFile) ;
}
newmtx = DenseMtx_MPI_splitByRows(mtx, mapIV, stats, bridge->msglvl,
bridge->msgFile, tag, bridge->comm) ;
if ( bridge->msglvl > 2 ) {
fprintf(bridge->msgFile, "\n newmtx: %d rows x %d columns",
newmtx->nrow, newmtx->ncol) ;
fflush(bridge->msgFile) ;
}
DenseMtx_free(mtx) ;
mtx = newmtx ;
IV_free(mapIV) ;
if ( myid == 0 ) {
if ( mtx->nrow != neqns || mtx->ncol != ncols ) {
fprintf(bridge->msgFile,
"\n\n WHOA: mtx->nrows %d, mtx->ncols %d"
", neqns %d, ncols %d", mtx->nrow, mtx->ncol,
neqns, ncols) ;
exit(-1) ;
}
DVcopy(neqns*ncols, y, DenseMtx_entries(mtx)) ;
}
DenseMtx_free(mtx) ;
/*
---------------------------------------------
broadcast the entries to the other processors
---------------------------------------------
*/
MPI_Bcast((void *) y, neqns*ncols, MPI_DOUBLE, 0, bridge->comm) ;
if ( bridge->msglvl > 2 ) {
fprintf(bridge->msgFile, "\n after the broadcast") ;
fflush(bridge->msgFile) ;
}
}
MPI_Barrier(bridge->comm) ;
#if MYDEBUG > 0
MARKTIME(t2) ;
time_JimMatMul += t2 - t1 ;
if ( bridge->myid == 0 ) {
fprintf(stdout, "\n (%d) JimMatMulMPI() end", count_JimMatMul) ;
fprintf(stdout, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_JimMatMul) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile,
"\n (%d) JimMatMulMPI() end", count_JimMatMul) ;
fprintf(bridge->msgFile, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_JimMatMul) ;
fflush(bridge->msgFile) ;
#endif
return ; }
NM_Status
SpoolesSolver :: solve(SparseMtrx *A, FloatArray *b, FloatArray *x)
{
int errorValue, mtxType, symmetryflag;
int seed = 30145, pivotingflag = 0;
int *oldToNew, *newToOld;
double droptol = 0.0, tau = 1.e300;
double cpus [ 10 ];
int stats [ 20 ];
ChvManager *chvmanager;
Chv *rootchv;
InpMtx *mtxA;
DenseMtx *mtxY, *mtxX;
// first check whether Lhs is defined
if ( !A ) {
_error("solveYourselfAt: unknown Lhs");
}
// and whether Rhs
if ( !b ) {
_error("solveYourselfAt: unknown Rhs");
}
// and whether previous Solution exist
if ( !x ) {
_error("solveYourselfAt: unknown solution array");
}
if ( x->giveSize() != b->giveSize() ) {
_error("solveYourselfAt: size mismatch");
}
Timer timer;
timer.startTimer();
if ( A->giveType() != SMT_SpoolesMtrx ) {
_error("solveYourselfAt: SpoolesSparseMtrx Expected");
}
mtxA = ( ( SpoolesSparseMtrx * ) A )->giveInpMtrx();
mtxType = ( ( SpoolesSparseMtrx * ) A )->giveValueType();
symmetryflag = ( ( SpoolesSparseMtrx * ) A )->giveSymmetryFlag();
int i;
int neqns = A->giveNumberOfRows();
int nrhs = 1;
/* convert right-hand side to DenseMtx */
mtxY = DenseMtx_new();
DenseMtx_init(mtxY, mtxType, 0, 0, neqns, nrhs, 1, neqns);
DenseMtx_zero(mtxY);
for ( i = 0; i < neqns; i++ ) {
DenseMtx_setRealEntry( mtxY, i, 0, b->at(i + 1) );
}
if ( ( Lhs != A ) || ( this->lhsVersion != A->giveVersion() ) ) {
//
// lhs has been changed -> new factorization
//
Lhs = A;
this->lhsVersion = A->giveVersion();
if ( frontmtx ) {
FrontMtx_free(frontmtx);
}
if ( newToOldIV ) {
IV_free(newToOldIV);
}
if ( oldToNewIV ) {
IV_free(oldToNewIV);
}
if ( frontETree ) {
ETree_free(frontETree);
}
if ( symbfacIVL ) {
IVL_free(symbfacIVL);
}
if ( mtxmanager ) {
SubMtxManager_free(mtxmanager);
}
if ( graph ) {
Graph_free(graph);
}
/*
* -------------------------------------------------
* STEP 3 : find a low-fill ordering
* (1) create the Graph object
* (2) order the graph using multiple minimum degree
* -------------------------------------------------
*/
int nedges;
graph = Graph_new();
adjIVL = InpMtx_fullAdjacency(mtxA);
nedges = IVL_tsize(adjIVL);
Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL,
NULL, NULL);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n graph of the input matrix");
Graph_writeForHumanEye(graph, msgFile);
fflush(msgFile);
}
frontETree = orderViaMMD(graph, seed, msglvl, msgFile);
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n front tree from ordering");
ETree_writeForHumanEye(frontETree, msgFile);
fflush(msgFile);
}
/*
* ----------------------------------------------------
* STEP 4: get the permutation, permute the front tree,
* permute the matrix and right hand side, and
* get the symbolic factorization
* ----------------------------------------------------
*/
oldToNewIV = ETree_oldToNewVtxPerm(frontETree);
oldToNew = IV_entries(oldToNewIV);
newToOldIV = ETree_newToOldVtxPerm(frontETree);
newToOld = IV_entries(newToOldIV);
ETree_permuteVertices(frontETree, oldToNewIV);
InpMtx_permute(mtxA, oldToNew, oldToNew);
if ( symmetryflag == SPOOLES_SYMMETRIC ||
symmetryflag == SPOOLES_HERMITIAN ) {
InpMtx_mapToUpperTriangle(mtxA);
}
InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS);
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS);
symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n old-to-new permutation vector");
IV_writeForHumanEye(oldToNewIV, msgFile);
fprintf(msgFile, "\n\n new-to-old permutation vector");
IV_writeForHumanEye(newToOldIV, msgFile);
fprintf(msgFile, "\n\n front tree after permutation");
ETree_writeForHumanEye(frontETree, msgFile);
fprintf(msgFile, "\n\n input matrix after permutation");
InpMtx_writeForHumanEye(mtxA, msgFile);
fprintf(msgFile, "\n\n symbolic factorization");
IVL_writeForHumanEye(symbfacIVL, msgFile);
fflush(msgFile);
}
Tree_writeToFile(frontETree->tree, (char*)"haggar.treef");
/*--------------------------------------------------------------------*/
/*
* ------------------------------------------
* STEP 5: initialize the front matrix object
* ------------------------------------------
*/
frontmtx = FrontMtx_new();
mtxmanager = SubMtxManager_new();
SubMtxManager_init(mtxmanager, NO_LOCK, 0);
FrontMtx_init(frontmtx, frontETree, symbfacIVL, mtxType, symmetryflag,
FRONTMTX_DENSE_FRONTS, pivotingflag, NO_LOCK, 0, NULL,
mtxmanager, msglvl, msgFile);
/*--------------------------------------------------------------------*/
/*
* -----------------------------------------
* STEP 6: compute the numeric factorization
* -----------------------------------------
*/
chvmanager = ChvManager_new();
ChvManager_init(chvmanager, NO_LOCK, 1);
DVfill(10, cpus, 0.0);
IVfill(20, stats, 0);
rootchv = FrontMtx_factorInpMtx(frontmtx, mtxA, tau, droptol,
chvmanager, & errorValue, cpus, stats, msglvl, msgFile);
ChvManager_free(chvmanager);
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n factor matrix");
FrontMtx_writeForHumanEye(frontmtx, msgFile);
fflush(msgFile);
}
if ( rootchv != NULL ) {
fprintf(msgFile, "\n\n matrix found to be singular\n");
exit(-1);
}
if ( errorValue >= 0 ) {
fprintf(msgFile, "\n\n error encountered at front %d", errorValue);
exit(-1);
}
/*--------------------------------------------------------------------*/
/*
* --------------------------------------
* STEP 7: post-process the factorization
* --------------------------------------
*/
FrontMtx_postProcess(frontmtx, msglvl, msgFile);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n factor matrix after post-processing");
FrontMtx_writeForHumanEye(frontmtx, msgFile);
fflush(msgFile);
}
/*--------------------------------------------------------------------*/
}
/*
* ----------------------------------------------------
* STEP 4: permute the right hand side
* ----------------------------------------------------
*/
DenseMtx_permuteRows(mtxY, oldToNewIV);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n right hand side matrix after permutation");
DenseMtx_writeForHumanEye(mtxY, msgFile);
}
/*
* -------------------------------
* STEP 8: solve the linear system
* -------------------------------
*/
mtxX = DenseMtx_new();
DenseMtx_init(mtxX, mtxType, 0, 0, neqns, nrhs, 1, neqns);
DenseMtx_zero(mtxX);
FrontMtx_solve(frontmtx, mtxX, mtxY, mtxmanager,
cpus, msglvl, msgFile);
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n solution matrix in new ordering");
DenseMtx_writeForHumanEye(mtxX, msgFile);
fflush(msgFile);
}
/*--------------------------------------------------------------------*/
/*
* -------------------------------------------------------
* STEP 9: permute the solution into the original ordering
* -------------------------------------------------------
*/
DenseMtx_permuteRows(mtxX, newToOldIV);
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n solution matrix in original ordering");
DenseMtx_writeForHumanEye(mtxX, msgFile);
fflush(msgFile);
}
// DenseMtx_writeForMatlab(mtxX, "x", msgFile) ;
/*--------------------------------------------------------------------*/
/* fetch data to oofem vectors */
double *xptr = x->givePointer();
for ( i = 0; i < neqns; i++ ) {
DenseMtx_realEntry(mtxX, i, 0, xptr + i);
// printf ("x(%d) = %e\n", i+1, *(xptr+i));
}
// DenseMtx_copyRowIntoVector(mtxX, 0, x->givePointer());
timer.stopTimer();
OOFEM_LOG_DEBUG( "SpoolesSolver info: user time consumed by solution: %.2fs\n", timer.getUtime() );
/*
* -----------
* free memory
* -----------
*/
DenseMtx_free(mtxX);
DenseMtx_free(mtxY);
/*--------------------------------------------------------------------*/
return ( 1 );
}
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
------------------------------------------------------
generate a random matrix and test the InpMtx_*_gmmm*()
matrix-matrix multiply methods.
the output is a matlab file to test correctness.
created -- 98nov14, cca
------------------------------------------------------
*/
{
DenseMtx *X, *Y ;
double alpha[2], beta[2] ;
double alphaImag, alphaReal, betaImag, betaReal ;
Drand *drand ;
int dataType, msglvl, ncolA, nitem, nrhs, nrowA, nrowX,
nrowY, seed, coordType, rc, symflag, transposeflag ;
InpMtx *A ;
FILE *msgFile ;
if ( argc != 16 ) {
fprintf(stdout,
"\n\n %% usage : %s msglvl msgFile symflag coordType transpose"
"\n %% nrow ncol nent nrhs seed "
"\n %% alphaReal alphaImag betaReal betaImag"
"\n %% msglvl -- message level"
"\n %% msgFile -- message file"
"\n %% dataType -- type of matrix entries"
"\n %% 1 -- real"
"\n %% 2 -- complex"
"\n %% symflag -- symmetry flag"
"\n %% 0 -- symmetric"
"\n %% 1 -- hermitian"
"\n %% 2 -- nonsymmetric"
"\n %% coordType -- storage mode"
"\n %% 1 -- by rows"
"\n %% 2 -- by columns"
"\n %% 3 -- by chevrons, (requires nrow = ncol)"
"\n %% transpose -- transpose flag"
"\n %% 0 -- Y := beta * Y + alpha * A * X"
"\n %% 1 -- Y := beta * Y + alpha * A^H * X, nonsymmetric only"
"\n %% 2 -- Y := beta * Y + alpha * A^T * X, nonsymmetric only"
"\n %% nrowA -- number of rows in A"
"\n %% ncolA -- number of columns in A"
"\n %% nitem -- number of items"
"\n %% nrhs -- number of right hand sides"
"\n %% seed -- random number seed"
"\n %% alphaReal -- y := beta*y + alpha*A*x"
"\n %% alphaImag -- y := beta*y + alpha*A*x"
"\n %% betaReal -- y := beta*y + alpha*A*x"
"\n %% betaImag -- y := beta*y + alpha*A*x"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
dataType = atoi(argv[3]) ;
symflag = atoi(argv[4]) ;
coordType = atoi(argv[5]) ;
transposeflag = atoi(argv[6]) ;
nrowA = atoi(argv[7]) ;
ncolA = atoi(argv[8]) ;
nitem = atoi(argv[9]) ;
nrhs = atoi(argv[10]) ;
seed = atoi(argv[11]) ;
alphaReal = atof(argv[12]) ;
alphaImag = atof(argv[13]) ;
betaReal = atof(argv[14]) ;
betaImag = atof(argv[15]) ;
fprintf(msgFile,
"\n %% %s "
"\n %% msglvl -- %d"
"\n %% msgFile -- %s"
"\n %% dataType -- %d"
"\n %% symflag -- %d"
"\n %% coordType -- %d"
"\n %% transposeflag -- %d"
"\n %% nrowA -- %d"
"\n %% ncolA -- %d"
"\n %% nitem -- %d"
"\n %% nrhs -- %d"
"\n %% seed -- %d"
"\n %% alphaReal -- %e"
"\n %% alphaImag -- %e"
"\n %% betaReal -- %e"
"\n %% betaImag -- %e"
"\n",
argv[0], msglvl, argv[2], dataType, symflag, coordType,
transposeflag, nrowA, ncolA, nitem, nrhs, seed,
alphaReal, alphaImag, betaReal, betaImag) ;
fflush(msgFile) ;
if ( dataType != 1 && dataType != 2 ) {
fprintf(stderr, "\n invalid value %d for dataType\n", dataType) ;
exit(-1) ;
}
if ( symflag != 0 && symflag != 1 && symflag != 2 ) {
fprintf(stderr, "\n invalid value %d for symflag\n", symflag) ;
exit(-1) ;
}
if ( coordType != 1 && coordType != 2 && coordType != 3 ) {
fprintf(stderr,
"\n invalid value %d for coordType\n", coordType) ;
exit(-1) ;
}
if ( transposeflag < 0
|| transposeflag > 2 ) {
fprintf(stderr, "\n error, transposeflag = %d, must be 0, 1 or 2",
transposeflag) ;
exit(-1) ;
}
if ( (transposeflag == 1 && symflag != 2)
|| (transposeflag == 2 && symflag != 2) ) {
fprintf(stderr, "\n error, transposeflag = %d, symflag = %d",
transposeflag, symflag) ;
exit(-1) ;
}
if ( transposeflag == 1 && dataType != 2 ) {
fprintf(stderr, "\n error, transposeflag = %d, dataType = %d",
transposeflag, dataType) ;
exit(-1) ;
}
if ( symflag == 1 && dataType != 2 ) {
fprintf(stderr,
"\n symflag = 1 (hermitian), dataType != 2 (complex)") ;
exit(-1) ;
}
if ( nrowA <= 0 || ncolA <= 0 || nitem <= 0 ) {
fprintf(stderr,
"\n invalid value: nrow = %d, ncol = %d, nitem = %d",
nrowA, ncolA, nitem) ;
exit(-1) ;
}
if ( symflag < 2 && nrowA != ncolA ) {
fprintf(stderr,
"\n invalid data: symflag = %d, nrow = %d, ncol = %d",
symflag, nrowA, ncolA) ;
exit(-1) ;
}
alpha[0] = alphaReal ;
alpha[1] = alphaImag ;
beta[0] = betaReal ;
beta[1] = betaImag ;
drand = Drand_new() ;
Drand_setSeed(drand, seed) ;
Drand_setUniform(drand, -1.0, 1.0) ;
/*
----------------------------
initialize the matrix object
and fill with random entries
----------------------------
*/
A = InpMtx_new() ;
InpMtx_init(A, coordType, dataType, 0, 0) ;
rc = InpMtx_randomMatrix(A, dataType, coordType, INPMTX_BY_VECTORS,
nrowA, ncolA, symflag, 1, nitem, seed) ;
if ( rc != 1 ) {
fprintf(stderr, "\n error return %d from InpMtx_randomMatrix()", rc);
exit(-1) ;
}
/*
-------------------------------------------
write the assembled matrix to a matlab file
-------------------------------------------
*/
InpMtx_writeForMatlab(A, "A", msgFile) ;
if ( symflag == 0 ) {
fprintf(msgFile,
"\n for k = 1:%d"
"\n for j = k+1:%d"
"\n A(j,k) = A(k,j) ;"
"\n end"
"\n end", nrowA, ncolA) ;
} else if ( symflag == 1 ) {
fprintf(msgFile,
"\n for k = 1:%d"
"\n for j = k+1:%d"
"\n A(j,k) = ctranspose(A(k,j)) ;"
"\n end"
"\n end", nrowA, ncolA) ;
}
/*
-------------------------------
generate dense matrices X and Y
-------------------------------
*/
if ( transposeflag == 0 ) {
nrowX = ncolA ;
nrowY = nrowA ;
} else {
nrowX = nrowA ;
nrowY = ncolA ;
}
X = DenseMtx_new() ;
Y = DenseMtx_new() ;
DenseMtx_init(X, dataType, 0, 0, nrowX, nrhs, 1, nrowX) ;
DenseMtx_fillRandomEntries(X, drand) ;
DenseMtx_init(Y, dataType, 0, 0, nrowY, nrhs, 1, nrowY) ;
DenseMtx_fillRandomEntries(Y, drand) ;
fprintf(msgFile, "\n X = zeros(%d,%d) ;", nrowX, nrhs) ;
DenseMtx_writeForMatlab(X, "X", msgFile) ;
fprintf(msgFile, "\n Y = zeros(%d,%d) ;", nrowY, nrhs) ;
DenseMtx_writeForMatlab(Y, "Y", msgFile) ;
/*
----------------------------------
perform the matrix-matrix multiply
----------------------------------
*/
fprintf(msgFile, "\n beta = %20.12e + %20.2e*i;", beta[0], beta[1]);
fprintf(msgFile, "\n alpha = %20.12e + %20.2e*i;", alpha[0], alpha[1]);
fprintf(msgFile, "\n Z = zeros(%d,1) ;", nrowY) ;
if ( transposeflag == 0 ) {
if ( symflag == 0 ) {
InpMtx_sym_gmmm(A, beta, Y, alpha, X) ;
} else if ( symflag == 1 ) {
InpMtx_herm_gmmm(A, beta, Y, alpha, X) ;
} else if ( symflag == 2 ) {
InpMtx_nonsym_gmmm(A, beta, Y, alpha, X) ;
}
DenseMtx_writeForMatlab(Y, "Z", msgFile) ;
fprintf(msgFile, "\n maxerr = max(Z - beta*Y - alpha*A*X) ") ;
fprintf(msgFile, "\n") ;
} else if ( transposeflag == 1 ) {
InpMtx_nonsym_gmmm_H(A, beta, Y, alpha, X) ;
DenseMtx_writeForMatlab(Y, "Z", msgFile) ;
fprintf(msgFile,
"\n maxerr = max(Z - beta*Y - alpha*ctranspose(A)*X) ") ;
fprintf(msgFile, "\n") ;
} else if ( transposeflag == 2 ) {
InpMtx_nonsym_gmmm_T(A, beta, Y, alpha, X) ;
DenseMtx_writeForMatlab(Y, "Z", msgFile) ;
fprintf(msgFile,
"\n maxerr = max(Z - beta*Y - alpha*transpose(A)*X) ") ;
fprintf(msgFile, "\n") ;
}
/*
------------------------
free the working storage
------------------------
*/
InpMtx_free(A) ;
DenseMtx_free(X) ;
DenseMtx_free(Y) ;
Drand_free(drand) ;
fclose(msgFile) ;
return(1) ; }
PetscErrorCode MatSolve_SeqSpooles(Mat A,Vec b,Vec x)
{
Mat_Spooles *lu = (Mat_Spooles*)A->spptr;
PetscScalar *array;
DenseMtx *mtxY, *mtxX ;
PetscErrorCode ierr;
PetscInt irow,neqns=A->cmap->n,nrow=A->rmap->n,*iv;
#if defined(PETSC_USE_COMPLEX)
double x_real,x_imag;
#else
double *entX;
#endif
PetscFunctionBegin;
mtxY = DenseMtx_new();
DenseMtx_init(mtxY, lu->options.typeflag, 0, 0, nrow, 1, 1, nrow); /* column major */
ierr = VecGetArray(b,&array);CHKERRQ(ierr);
if (lu->options.useQR) { /* copy b to mtxY */
for ( irow = 0 ; irow < nrow; irow++ )
#if !defined(PETSC_USE_COMPLEX)
DenseMtx_setRealEntry(mtxY, irow, 0, *array++);
#else
DenseMtx_setComplexEntry(mtxY, irow, 0, PetscRealPart(array[irow]), PetscImaginaryPart(array[irow]));
#endif
} else { /* copy permuted b to mtxY */
iv = IV_entries(lu->oldToNewIV);
for ( irow = 0 ; irow < nrow; irow++ )
#if !defined(PETSC_USE_COMPLEX)
DenseMtx_setRealEntry(mtxY, *iv++, 0, *array++);
#else
DenseMtx_setComplexEntry(mtxY,*iv++,0,PetscRealPart(array[irow]),PetscImaginaryPart(array[irow]));
#endif
}
ierr = VecRestoreArray(b,&array);CHKERRQ(ierr);
mtxX = DenseMtx_new();
DenseMtx_init(mtxX, lu->options.typeflag, 0, 0, neqns, 1, 1, neqns);
if (lu->options.useQR) {
FrontMtx_QR_solve(lu->frontmtx, lu->mtxA, mtxX, mtxY, lu->mtxmanager,
lu->cpus, lu->options.msglvl, lu->options.msgFile);
} else {
FrontMtx_solve(lu->frontmtx, mtxX, mtxY, lu->mtxmanager,
lu->cpus, lu->options.msglvl, lu->options.msgFile);
}
if ( lu->options.msglvl > 2 ) {
int err;
ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n right hand side matrix after permutation");CHKERRQ(ierr);
DenseMtx_writeForHumanEye(mtxY, lu->options.msgFile);
ierr = PetscFPrintf(PETSC_COMM_SELF,lu->options.msgFile, "\n\n solution matrix in new ordering");CHKERRQ(ierr);
DenseMtx_writeForHumanEye(mtxX, lu->options.msgFile);
err = fflush(lu->options.msgFile);
if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fflush() failed on file");
}
/* permute solution into original ordering, then copy to x */
DenseMtx_permuteRows(mtxX, lu->newToOldIV);
ierr = VecGetArray(x,&array);CHKERRQ(ierr);
#if !defined(PETSC_USE_COMPLEX)
entX = DenseMtx_entries(mtxX);
DVcopy(neqns, array, entX);
#else
for (irow=0; irow<nrow; irow++){
DenseMtx_complexEntry(mtxX,irow,0,&x_real,&x_imag);
array[irow] = x_real+x_imag*PETSC_i;
}
#endif
ierr = VecRestoreArray(x,&array);CHKERRQ(ierr);
/* free memory */
DenseMtx_free(mtxX);
DenseMtx_free(mtxY);
PetscFunctionReturn(0);
}
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
-----------------------------------------------
test the DenseMtx_frobNorm routine.
when msglvl > 1, the output of this program
can be fed into Matlab to check for errors
created -- 98dec03, ycp
-----------------------------------------------
*/
{
DenseMtx *A ;
double t1, t2, value ;
Drand *drand ;
FILE *msgFile ;
int inc1, inc2, msglvl, nrow, ncol, seed, type ;
if ( argc != 9 ) {
fprintf(stdout,
"\n\n usage : %s msglvl msgFile type nrow ncol inc1 inc2 "
"\n , seed "
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n type -- entries type"
"\n 1 -- real"
"\n 2 -- complex"
"\n nrow -- # of rows "
"\n ncol -- # of columns "
"\n inc1 -- row increment "
"\n inc2 -- column increment "
"\n seed -- random number seed"
"\n", argv[0]) ;
return(0) ;
}
if ( (msglvl = atoi(argv[1])) < 0 ) {
fprintf(stderr, "\n message level must be positive\n") ;
spoolesFatal();
}
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n unable to open file %s\n", argv[2]) ;
return(-1) ;
}
type = atoi(argv[3]) ;
nrow = atoi(argv[4]) ;
ncol = atoi(argv[5]) ;
inc1 = atoi(argv[6]) ;
inc2 = atoi(argv[7]) ;
if ( type < 1 || type > 2 || nrow < 0 || ncol < 0
|| inc1 < 1 || inc2 < 1 ) {
fprintf(stderr,
"\n fatal error, type %d, nrow %d, ncol %d, inc1 %d, inc2 %d",
type, nrow, ncol, inc1, inc2) ;
spoolesFatal();
}
seed = atoi(argv[8]) ;
fprintf(msgFile, "\n\n %% %s :"
"\n %% msglvl = %d"
"\n %% msgFile = %s"
"\n %% type = %d"
"\n %% nrow = %d"
"\n %% ncol = %d"
"\n %% inc1 = %d"
"\n %% inc2 = %d"
"\n %% seed = %d"
"\n",
argv[0], msglvl, argv[2], type, nrow, ncol, inc1, inc2, seed) ;
/*
----------------------------
initialize the matrix object
----------------------------
*/
MARKTIME(t1) ;
A = DenseMtx_new() ;
DenseMtx_init(A, type, 0, 0, nrow, ncol, inc1, inc2) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n %% CPU : %.3f to initialize matrix object",
t2 - t1) ;
MARKTIME(t1) ;
drand = Drand_new() ;
Drand_setSeed(drand, seed) ;
seed++ ;
Drand_setUniform(drand, -1.0, 1.0) ;
DenseMtx_fillRandomEntries(A, drand) ;
MARKTIME(t2) ;
fprintf(msgFile,
"\n %% CPU : %.3f to fill matrix with random numbers", t2 - t1) ;
if ( msglvl > 3 ) {
fprintf(msgFile, "\n matrix A") ;
DenseMtx_writeForHumanEye(A, msgFile) ;
}
if ( msglvl > 1 ) {
fprintf(msgFile, "\n %% matrix A") ;
fprintf(msgFile, "\n nrow = %d ;", nrow) ;
fprintf(msgFile, "\n ncol = %d ;", ncol) ;
fprintf(msgFile, "\n");
DenseMtx_writeForMatlab(A, "A", msgFile) ;
}
/*
--------------------------
compute the frobenius norm
--------------------------
*/
value = DenseMtx_frobNorm(A);
if ( msglvl > 1 ) {
fprintf(msgFile, "\n %% Frobenius Norm = %e", value) ;
fprintf(msgFile, "\n");
fflush(msgFile) ;
}
/*
------------------------
free the working storage
------------------------
*/
DenseMtx_free(A) ;
Drand_free(drand) ;
return(1) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
-----------------------------------------------------
test the factor method for a grid matrix
(0) read in matrix from source file
(1) conver data matrix to InpMtx object if necessary
(2) create Graph and ETree object if necessary
(3) read in/create an ETree object
(4) create a solution matrix object
(5) multiply the solution with the matrix
to get a right hand side matrix object
(6) factor the matrix
(7) solve the system
created -- 98dec30, jwu
-----------------------------------------------------
*/
{
char etreeFileName[80], mtxFileName[80], *cpt, rhsFileName[80],
srcFileName[80], ctemp[81], msgFileName[80], slnFileName[80] ;
Chv *chv, *rootchv ;
ChvManager *chvmanager ;
DenseMtx *mtxB, *mtxQ, *mtxX, *mtxZ ;
double one[2] = { 1.0, 0.0 } ;
FrontMtx *frontmtx ;
InpMtx *mtxA ;
SubMtxManager *mtxmanager ;
double cputotal, droptol, conv_tol, factorops ;
double cpus[9] ;
Drand drand ;
double nops, tau, t1, t2 ;
ETree *frontETree ;
Graph *graph ;
FILE *msgFile, *inFile ;
int error, loc, msglvl, neqns, nzf, iformat,
pivotingflag, rc, seed, sparsityflag, symmetryflag,
method[METHODS], type, nrhs, etreeflag ;
int stats[6] ;
int nnzA, Ik, itermax, zversion, iterout ;
IV *newToOldIV, *oldToNewIV ;
IVL *symbfacIVL ;
int i, j, k, m, n, imethod, maxdomainsize, maxzeros, maxsize;
int nouter,ninner ;
if ( argc != 2 ) {
fprintf(stdout,
"\n\n usage : %s inFile"
"\n inFile -- input filename"
"\n", argv[0]) ;
return(-1) ;
}
/* read input file */
inFile = fopen(argv[1], "r");
if (inFile == (FILE *)NULL) {
fprintf(stderr, "\n fatal error in %s: unable to open file %s\n",
argv[0], argv[1]) ;
return(-1) ;
}
for (i=0; i<METHODS; i++) method[i]=-1;
imethod=0;
k=0;
while (1) {
fgets(ctemp, 80, inFile);
if (ctemp[0] != '*') {
/*printf("l=%2d:%s\n", strlen(ctemp),ctemp);*/
if (strlen(ctemp)==80) {
fprintf(stderr, "\n fatal error in %s: input line contains more than "
"80 characters.\n",argv[0]);
exit(-1);
}
if (k==0) {
sscanf(ctemp, "%d", &iformat);
if (iformat < 0 || iformat > 2) {
fprintf(stderr, "\n fatal error in %s: "
"invalid source matrix format\n",argv[0]) ;
return(-1) ;
}
}
else if (k==1)
sscanf(ctemp, "%s", srcFileName);
else if (k==2)
sscanf(ctemp, "%s", mtxFileName);
else if (k==3) {
sscanf(ctemp, "%d", &etreeflag);
if (etreeflag < 0 || etreeflag > 4) {
fprintf(stderr, "\n fatal error in %s: "
"invalid etree file status\n",argv[0]) ;
return(-1) ;
}
}
else if (k==4)
sscanf(ctemp, "%s", etreeFileName);
else if (k==5)
sscanf(ctemp, "%s", rhsFileName);
else if (k==6)
sscanf(ctemp, "%s", slnFileName);
else if (k==7){
sscanf(ctemp, "%s", msgFileName);
if ( strcmp(msgFileName, "stdout") == 0 ) {
msgFile = stdout ;
}
else if ( (msgFile = fopen(msgFileName, "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n", argv[0], ctemp) ;
return(-1) ;
}
}
else if (k==8)
sscanf(ctemp, "%d %d %d %d %d %d",
&msglvl, &seed, &nrhs, &Ik, &itermax, &iterout);
else if (k==9)
sscanf(ctemp, "%d %d %d", &symmetryflag, &sparsityflag, &pivotingflag);
else if (k==10)
sscanf(ctemp, "%lf %lf %lf", &tau, &droptol, &conv_tol);
else if (k==11) {
/*
for (j=0; j<strlen(ctemp); j++) {
printf("j=%2d:%s",j,ctemp+j);
if (ctemp[j] == ' ' && ctemp[j+1] != ' ') {
sscanf(ctemp+j, "%d", method+imethod);
printf("method[%d]=%d\n",imethod,method[imethod]);
if (method[imethod] < 0) break;
imethod++;
}
}
*/
imethod = sscanf(ctemp,"%d %d %d %d %d %d %d %d %d %d",
method, method+1, method+2, method+3, method+4,
method+5, method+6, method+7, method+8, method+9);
/*printf("imethod=%d\n",imethod);*/
for (j=0; j<imethod; j++) {
/*printf("method[%d]=%d\n",j,method[j]);*/
if (method[j]<0) {
imethod=j;
break;
}
}
if (imethod == 0) {
fprintf(msgFile,"No method assigned in input file\n");
return(-1);
}
}
k++;
}
if (k==12) break;
}
fclose(inFile);
/* reset nrhs to 1 */
if (nrhs > 1) {
fprintf(msgFile,"*** Multiple right-hand-side vectors is not allowed yet.\n");
fprintf(msgFile,"*** nrhs is reset to 1.\n");
nrhs =1;
}
fprintf(msgFile,
"\n %s "
"\n srcFileName -- %s"
"\n mtxFileName -- %s"
"\n etreeFileName -- %s"
"\n rhsFileName -- %s"
"\n msglvl -- %d"
"\n seed -- %d"
"\n symmetryflag -- %d"
"\n sparsityflag -- %d"
"\n pivotingflag -- %d"
"\n tau -- %e"
"\n droptol -- %e"
"\n conv_tol -- %e"
"\n method -- ",
argv[0], srcFileName, mtxFileName, etreeFileName, rhsFileName,
msglvl, seed, symmetryflag, sparsityflag, pivotingflag,
tau, droptol, conv_tol) ;
for (k=0; k<imethod; k++)
fprintf(msgFile, "%d ", method[k]);
fprintf(msgFile, "\n ", method[k]);
fflush(msgFile) ;
/*
--------------------------------------
initialize the random number generator
--------------------------------------
*/
Drand_setDefaultFields(&drand) ;
Drand_init(&drand) ;
Drand_setSeed(&drand, seed) ;
/*Drand_setUniform(&drand, 0.0, 1.0) ;*/
Drand_setNormal(&drand, 0.0, 1.0) ;
/*
----------------------------------------------
read in or convert source to the InpMtx object
----------------------------------------------
*/
rc = 1;
if ( strcmp(srcFileName, "none") == 0 ) {
fprintf(msgFile, "\n no file to read from") ;
exit(-1) ;
}
mtxA = InpMtx_new() ;
MARKTIME(t1) ;
if (iformat == 0) { /* InpMtx source format */
rc = InpMtx_readFromFile(mtxA, srcFileName) ;
strcpy(mtxFileName, srcFileName);
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from InpMtx_readFromFile(%p,%s)",
rc, mtxA, srcFileName) ;
}
else if (iformat == 1) { /* HBF source format */
rc = InpMtx_readFromHBfile(mtxA, srcFileName) ;
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from InpMtx_readFromHBfile(%p,%s)",
rc, mtxA, srcFileName) ;
}
else { /* AIJ2 source format */
rc = InpMtx_readFromAIJ2file(mtxA, srcFileName) ;
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from InpMtx_readFromAIJ2file(%p,%s)",
rc, mtxA, srcFileName) ;
}
MARKTIME(t2) ;
if (iformat>0 && strcmp(mtxFileName, "none") != 0 ) {
rc = InpMtx_writeToFile(mtxA, mtxFileName) ;
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from InpMtx_writeToFile(%p,%s)",
rc, mtxA, mtxFileName) ;
}
fprintf(msgFile, "\n CPU %8.3f : read in (+ convert to) mtxA from file %s",
t2 - t1, mtxFileName) ;
if (rc != 1) {
goto end_read;
}
type = mtxA->inputMode ;
neqns = 1 + IVmax(mtxA->nent, InpMtx_ivec1(mtxA), &loc) ;
if ( INPMTX_IS_BY_ROWS(mtxA) ) {
fprintf(msgFile, "\n matrix coordinate type is rows") ;
} else if ( INPMTX_IS_BY_COLUMNS(mtxA) ) {
fprintf(msgFile, "\n matrix coordinate type is columns") ;
} else if ( INPMTX_IS_BY_CHEVRONS(mtxA) ) {
fprintf(msgFile, "\n matrix coordinate type is chevrons") ;
} else {
fprintf(msgFile, "\n\n, error, bad coordinate type") ;
rc=-1;
goto end_read;
}
if ( INPMTX_IS_RAW_DATA(mtxA) ) {
fprintf(msgFile, "\n matrix storage mode is raw data\n") ;
} else if ( INPMTX_IS_SORTED(mtxA) ) {
fprintf(msgFile, "\n matrix storage mode is sorted\n") ;
} else if ( INPMTX_IS_BY_VECTORS(mtxA) ) {
fprintf(msgFile, "\n matrix storage mode is by vectors\n") ;
} else {
fprintf(msgFile, "\n\n, error, bad storage mode") ;
rc=-1;
goto end_read;
}
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n after reading InpMtx object from file %s",
mtxFileName) ;
if ( msglvl == 2 ) {
InpMtx_writeStats(mtxA, msgFile) ;
} else {
InpMtx_writeForHumanEye(mtxA, msgFile) ;
}
fflush(msgFile) ;
}
/*
Get the nonzeros in matrix A and print it
*/
nnzA = InpMtx_nent( mtxA );
fprintf(msgFile, "\n\n Input matrix size %d NNZ %d",
neqns, nnzA) ;
/*
--------------------------------------------------------
generate the linear system
1. generate solution matrix and fill with random numbers
2. generate rhs matrix and fill with zeros
3. compute matrix-matrix multiply
--------------------------------------------------------
*/
MARKTIME(t1) ;
mtxX = DenseMtx_new() ;
DenseMtx_init(mtxX, type, 0, -1, neqns, nrhs, 1, neqns) ;
mtxB = DenseMtx_new() ;
if (strcmp(rhsFileName, "none")) {
rc = DenseMtx_readFromFile(mtxB, rhsFileName) ;
if ( rc != 1 )
fprintf(msgFile, "\n return value %d from DenseMtx_readFromFile(%p,%s)",
rc, mtxB, rhsFileName) ;
DenseMtx_zero(mtxX) ;
}
else {
DenseMtx_init(mtxB, type, 1, -1, neqns, nrhs, 1, neqns) ;
DenseMtx_fillRandomEntries(mtxX, &drand) ;
DenseMtx_zero(mtxB) ;
switch ( symmetryflag ) {
case SPOOLES_SYMMETRIC :
InpMtx_sym_mmm(mtxA, mtxB, one, mtxX) ;
break ;
case SPOOLES_HERMITIAN :
InpMtx_herm_mmm(mtxA, mtxB, one, mtxX) ;
break ;
case SPOOLES_NONSYMMETRIC :
InpMtx_nonsym_mmm(mtxA, mtxB, one, mtxX) ;
break ;
default :
break ;
}
}
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : set up the solution and rhs ",
t2 - t1) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n original mtxX") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fprintf(msgFile, "\n\n original mtxB") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
if (rc != 1) {
InpMtx_free(mtxA);
DenseMtx_free(mtxX);
DenseMtx_free(mtxB);
goto end_init;
}
/*
------------------------
read in/create the ETree object
------------------------
*/
MARKTIME(t1) ;
if (etreeflag == 0) { /* read in ETree from file */
if ( strcmp(etreeFileName, "none") == 0 )
fprintf(msgFile, "\n no file to read from") ;
frontETree = ETree_new() ;
rc = ETree_readFromFile(frontETree, etreeFileName) ;
if (rc!=1)
fprintf(msgFile, "\n return value %d from ETree_readFromFile(%p,%s)",
rc, frontETree, etreeFileName) ;
}
else {
graph = Graph_new() ;
rc = InpMtx_createGraph(mtxA, graph);
if (rc!=1) {
fprintf(msgFile, "\n return value %d from InpMtx_createGraph(%p,%p)",
rc, mtxA, graph) ;
Graph_free(graph);
goto end_tree;
}
if (etreeflag == 1) { /* Via BestOfNDandMS */
maxdomainsize = 500; maxzeros = 1000; maxsize = 64 ;
frontETree = orderViaBestOfNDandMS(graph, maxdomainsize, maxzeros,
maxsize, seed, msglvl, msgFile) ;
}
else if (etreeflag == 2) { /* Via MMD */
frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ;
}
else if (etreeflag == 3) { /* Via MS */
maxdomainsize = 500;
frontETree = orderViaMS(graph, maxdomainsize, seed, msglvl, msgFile) ;
}
else if (etreeflag == 4) { /* Via ND */
maxdomainsize = 500;
frontETree = orderViaND(graph, maxdomainsize, seed, msglvl, msgFile) ;
}
Graph_free(graph);
/* optionally write out the ETree object */
if ( strcmp(etreeFileName, "none") != 0 ) {
fprintf(msgFile, "\n\n writing out ETree to file %s",
etreeFileName) ;
ETree_writeToFile(frontETree, etreeFileName) ;
}
}
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : read in/create frontETree from file %s",
t2 - t1, etreeFileName) ;
if ( rc != 1 ) {
ETree_free(frontETree);
goto end_tree;
}
ETree_leftJustify(frontETree) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n after reading ETree object from file %s",
etreeFileName) ;
if ( msglvl == 2 ) {
ETree_writeStats(frontETree, msgFile) ;
} else {
ETree_writeForHumanEye(frontETree, msgFile) ;
}
}
fflush(msgFile) ;
/*
--------------------------------------------------
get the permutations, permute the matrix and the
front tree, and compute the symbolic factorization
--------------------------------------------------
*/
MARKTIME(t1) ;
oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ;
newToOldIV = ETree_newToOldVtxPerm(frontETree) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : get permutations", t2 - t1) ;
MARKTIME(t1) ;
ETree_permuteVertices(frontETree, oldToNewIV) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : permute front tree", t2 - t1) ;
MARKTIME(t1) ;
InpMtx_permute(mtxA, IV_entries(oldToNewIV), IV_entries(oldToNewIV)) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : permute mtxA", t2 - t1) ;
if ( symmetryflag == SPOOLES_SYMMETRIC
|| symmetryflag == SPOOLES_HERMITIAN ) {
MARKTIME(t1) ;
InpMtx_mapToUpperTriangle(mtxA) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : map to upper triangle", t2 - t1) ;
}
if ( ! INPMTX_IS_BY_CHEVRONS(mtxA) ) {
MARKTIME(t1) ;
InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : change coordinate type", t2 - t1) ;
}
if ( INPMTX_IS_RAW_DATA(mtxA) ) {
MARKTIME(t1) ;
InpMtx_changeStorageMode(mtxA, INPMTX_SORTED) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : sort entries ", t2 - t1) ;
}
if ( INPMTX_IS_SORTED(mtxA) ) {
MARKTIME(t1) ;
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : convert to vectors ", t2 - t1) ;
}
MARKTIME(t1) ;
symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : symbolic factorization", t2 - t1) ;
MARKTIME(t1) ;
DenseMtx_permuteRows(mtxB, oldToNewIV) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : permute rhs", t2 - t1) ;
/*
------------------------------
initialize the FrontMtx object
------------------------------
*/
MARKTIME(t1) ;
frontmtx = FrontMtx_new() ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, NO_LOCK, 0) ;
FrontMtx_init(frontmtx, frontETree, symbfacIVL,
type, symmetryflag, sparsityflag, pivotingflag,
NO_LOCK, 0, NULL, mtxmanager, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : initialize the front matrix",
t2 - t1) ;
if ( msglvl > 1 ) {
fprintf(msgFile,
"\n nendD = %d, nentL = %d, nentU = %d",
frontmtx->nentD, frontmtx->nentL, frontmtx->nentU) ;
SubMtxManager_writeForHumanEye(mtxmanager, msgFile) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n front matrix initialized") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*
-----------------
factor the matrix
-----------------
*/
nzf = ETree_nFactorEntries(frontETree, symmetryflag) ;
factorops = ETree_nFactorOps(frontETree, type, symmetryflag) ;
fprintf(msgFile,
"\n %d factor entries, %.0f factor ops, %8.3f ratio",
nzf, factorops, factorops/nzf) ;
IVzero(6, stats) ;
DVzero(9, cpus) ;
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, NO_LOCK, 1) ;
MARKTIME(t1) ;
rootchv = FrontMtx_factorInpMtx(frontmtx, mtxA, tau, droptol,
chvmanager, &error, cpus,
stats, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : factor matrix, %8.3f mflops",
t2 - t1, 1.e-6*factorops/(t2-t1)) ;
if ( rootchv != NULL ) {
fprintf(msgFile, "\n\n factorization did not complete") ;
for ( chv = rootchv ; chv != NULL ; chv = chv->next ) {
fprintf(stdout, "\n chv %d, nD = %d, nL = %d, nU = %d",
chv->id, chv->nD, chv->nL, chv->nU) ;
}
}
if ( error >= 0 ) {
fprintf(msgFile, "\n\n error encountered at front %d\n", error) ;
rc=error ;
goto end_front;
}
fprintf(msgFile,
"\n %8d pivots, %8d pivot tests, %8d delayed rows and columns",
stats[0], stats[1], stats[2]) ;
if ( frontmtx->rowadjIVL != NULL ) {
fprintf(msgFile,
"\n %d entries in rowadjIVL", frontmtx->rowadjIVL->tsize) ;
}
if ( frontmtx->coladjIVL != NULL ) {
fprintf(msgFile,
", %d entries in coladjIVL", frontmtx->coladjIVL->tsize) ;
}
if ( frontmtx->upperblockIVL != NULL ) {
fprintf(msgFile,
"\n %d fronts, %d entries in upperblockIVL",
frontmtx->nfront, frontmtx->upperblockIVL->tsize) ;
}
if ( frontmtx->lowerblockIVL != NULL ) {
fprintf(msgFile,
", %d entries in lowerblockIVL",
frontmtx->lowerblockIVL->tsize) ;
}
fprintf(msgFile,
"\n %d entries in D, %d entries in L, %d entries in U",
stats[3], stats[4], stats[5]) ;
fprintf(msgFile, "\n %d locks", frontmtx->nlocks) ;
if ( FRONTMTX_IS_SYMMETRIC(frontmtx)
|| FRONTMTX_IS_HERMITIAN(frontmtx) ) {
int nneg, npos, nzero ;
FrontMtx_inertia(frontmtx, &nneg, &nzero, &npos) ;
fprintf(msgFile,
"\n %d negative, %d zero and %d positive eigenvalues",
nneg, nzero, npos) ;
fflush(msgFile) ;
}
cputotal = cpus[8] ;
if ( cputotal > 0.0 ) {
fprintf(msgFile,
"\n initialize fronts %8.3f %6.2f"
"\n load original entries %8.3f %6.2f"
"\n update fronts %8.3f %6.2f"
"\n assemble postponed data %8.3f %6.2f"
"\n factor fronts %8.3f %6.2f"
"\n extract postponed data %8.3f %6.2f"
"\n store factor entries %8.3f %6.2f"
"\n miscellaneous %8.3f %6.2f"
"\n total time %8.3f",
cpus[0], 100.*cpus[0]/cputotal,
cpus[1], 100.*cpus[1]/cputotal,
cpus[2], 100.*cpus[2]/cputotal,
cpus[3], 100.*cpus[3]/cputotal,
cpus[4], 100.*cpus[4]/cputotal,
cpus[5], 100.*cpus[5]/cputotal,
cpus[6], 100.*cpus[6]/cputotal,
cpus[7], 100.*cpus[7]/cputotal, cputotal) ;
}
if ( msglvl > 1 ) {
SubMtxManager_writeForHumanEye(mtxmanager, msgFile) ;
ChvManager_writeForHumanEye(chvmanager, msgFile) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front factor matrix") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
}
/*
------------------------------
post-process the factor matrix
------------------------------
*/
MARKTIME(t1) ;
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : post-process the matrix", t2 - t1) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
}
fprintf(msgFile, "\n\n after post-processing") ;
if ( msglvl > 1 ) SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
/*
----------------
solve the system
----------------
*/
neqns = mtxB->nrow ;
mtxZ = DenseMtx_new() ;
DenseMtx_init(mtxZ, type, 0, 0, neqns, nrhs, 1, neqns) ;
zversion=INPMTX_IS_COMPLEX_ENTRIES(mtxA);
for (k=0; k<imethod; k++) {
DenseMtx_zero(mtxZ) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n rhs") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(stdout) ;
}
fprintf(msgFile, "\n\n itemax %d", itermax) ;
DVzero(6, cpus) ;
MARKTIME(t1) ;
switch ( method[k] ) {
case BiCGStabR :
if (zversion)
rc=zbicgstabr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=bicgstabr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case BiCGStabL :
if (zversion)
rc=zbicgstabl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=bicgstabl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case TFQMRR :
if (zversion)
rc=ztfqmrr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=tfqmrr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case TFQMRL :
if (zversion)
rc=ztfqmrl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=tfqmrl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case PCGR :
if (zversion)
rc=zpcgr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=pcgr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case PCGL :
if (zversion)
rc=zpcgl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
else
rc=pcgl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ, mtxB,
itermax, conv_tol, msglvl, msgFile);
break;
case MLBiCGStabR :
mtxQ = DenseMtx_new() ;
DenseMtx_init(mtxQ, type, 0, -1, neqns, Ik, 1, neqns) ;
Drand_setUniform(&drand, 0.0, 1.0) ;
DenseMtx_fillRandomEntries(mtxQ, &drand) ;
if (zversion)
rc=zmlbicgstabr(neqns, type, symmetryflag, mtxA, frontmtx, mtxQ, mtxZ,
mtxB, itermax, conv_tol, msglvl, msgFile);
else
rc=mlbicgstabr(neqns, type, symmetryflag, mtxA, frontmtx, mtxQ, mtxZ,
mtxB, itermax, conv_tol, msglvl, msgFile);
DenseMtx_free(mtxQ) ;
break;
case MLBiCGStabL :
mtxQ = DenseMtx_new() ;
DenseMtx_init(mtxQ, type, 0, -1, neqns, Ik, 1, neqns) ;
Drand_setUniform(&drand, 0.0, 1.0) ;
DenseMtx_fillRandomEntries(mtxQ, &drand) ;
if (zversion)
rc=zmlbicgstabl(neqns, type, symmetryflag, mtxA, frontmtx, mtxQ, mtxZ,
mtxB, itermax, conv_tol, msglvl, msgFile);
else
rc=mlbicgstabl(neqns, type, symmetryflag, mtxA, frontmtx, mtxQ, mtxZ,
mtxB, itermax, conv_tol, msglvl, msgFile);
DenseMtx_free(mtxQ) ;
break;
case BGMRESR:
if (zversion)
fprintf(msgFile, "\n\n *** BGMRESR complex version is not available "
"at this moment. ") ;
else
rc=bgmresr(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ,
mtxB, iterout, itermax, &nouter, &ninner, conv_tol,
msglvl, msgFile);
break;
case BGMRESL:
if (zversion)
fprintf(msgFile, "\n\n *** BGMRESR complex version is not available "
"at this moment. ") ;
else
rc=bgmresl(neqns, type, symmetryflag, mtxA, frontmtx, mtxZ,
mtxB, iterout, itermax, &nouter, &ninner, conv_tol,
msglvl, msgFile);
break;
default:
fprintf(msgFile, "\n\n *** Invalid method number ") ;
}
MARKTIME(t2) ;
fprintf(msgFile, "\n\n CPU %8.3f : solve the system", t2 - t1) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n computed solution") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
/*
-------------------------------------------------------------
permute the computed solution back into the original ordering
-------------------------------------------------------------
*/
MARKTIME(t1) ;
DenseMtx_permuteRows(mtxZ, newToOldIV) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : permute solution", t2 - t1) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n permuted solution") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
/*
-------------
save solution
-------------
*/
if ( strcmp(slnFileName, "none") != 0 ) {
DenseMtx_writeToFile(mtxZ, slnFileName) ;
}
/*
-----------------
compute the error
-----------------
*/
if (!strcmp(rhsFileName, "none")) {
DenseMtx_sub(mtxZ, mtxX) ;
if (method[k] <8) {
mtxQ = DenseMtx_new() ;
DenseMtx_init(mtxQ, type, 0, -1, neqns, 1, 1, neqns) ;
rc=DenseMtx_initAsSubmatrix (mtxQ, mtxZ, 0, neqns-1, 0, 0);
fprintf(msgFile, "\n\n maxabs error = %12.4e", DenseMtx_maxabs(mtxQ)) ;
DenseMtx_free(mtxQ) ;
}
else
fprintf(msgFile, "\n\n maxabs error = %12.4e", DenseMtx_maxabs(mtxZ)) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n error") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
if ( msglvl > 1 )
SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
}
fprintf(msgFile, "\n--------- End of Method %d -------\n",method[k]) ;
}
/*
------------------------
free the working storage
------------------------
*/
DenseMtx_free(mtxZ) ;
end_front:
ChvManager_free(chvmanager) ;
SubMtxManager_free(mtxmanager) ;
FrontMtx_free(frontmtx) ;
IVL_free(symbfacIVL) ;
IV_free(oldToNewIV) ;
IV_free(newToOldIV) ;
end_tree:
ETree_free(frontETree) ;
end_init:
DenseMtx_free(mtxB) ;
DenseMtx_free(mtxX) ;
end_read:
InpMtx_free(mtxA) ;
fprintf(msgFile, "\n") ;
fclose(msgFile) ;
return(rc) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] ) {
/*
--------------------------------------------------
all-in-one program to solve A X = B
using a multithreaded factorization and solve
We use a patch-and-go strategy
for the factorization without pivoting
(1) read in matrix entries and form DInpMtx object
(2) form Graph object
(3) order matrix and form front tree
(4) get the permutation, permute the matrix and
front tree and get the symbolic factorization
(5) compute the numeric factorization
(6) read in right hand side entries
(7) compute the solution
created -- 98jun04, cca
--------------------------------------------------
*/
/*--------------------------------------------------------------------*/
char *matrixFileName, *rhsFileName ;
DenseMtx *mtxB, *mtxX ;
Chv *rootchv ;
ChvManager *chvmanager ;
double fudge, imag, real, tau = 100., toosmall, value ;
double cpus[10] ;
DV *cumopsDV ;
ETree *frontETree ;
FrontMtx *frontmtx ;
FILE *inputFile, *msgFile ;
Graph *graph ;
InpMtx *mtxA ;
int error, ient, irow, jcol, jrhs, jrow, lookahead, msglvl,
ncol, nedges, nent, neqns, nfront, nrhs, nrow, nthread,
patchAndGoFlag, seed,
storeids, storevalues, symmetryflag, type ;
int *newToOld, *oldToNew ;
int stats[20] ;
IV *newToOldIV, *oldToNewIV, *ownersIV ;
IVL *adjIVL, *symbfacIVL ;
SolveMap *solvemap ;
SubMtxManager *mtxmanager ;
/*--------------------------------------------------------------------*/
/*
--------------------
get input parameters
--------------------
*/
if ( argc != 14 ) {
fprintf(stdout, "\n"
"\n usage: %s msglvl msgFile type symmetryflag patchAndGoFlag"
"\n fudge toosmall storeids storevalues"
"\n matrixFileName rhsFileName seed"
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n type -- type of entries"
"\n 1 (SPOOLES_REAL) -- real entries"
"\n 2 (SPOOLES_COMPLEX) -- complex entries"
"\n symmetryflag -- type of matrix"
"\n 0 (SPOOLES_SYMMETRIC) -- symmetric entries"
"\n 1 (SPOOLES_HERMITIAN) -- Hermitian entries"
"\n 2 (SPOOLES_NONSYMMETRIC) -- nonsymmetric entries"
"\n patchAndGoFlag -- flag for the patch-and-go strategy"
"\n 0 -- none, stop factorization"
"\n 1 -- optimization strategy"
"\n 2 -- structural analysis strategy"
"\n fudge -- perturbation parameter"
"\n toosmall -- upper bound on a small pivot"
"\n storeids -- flag to store ids of small pivots"
"\n storevalues -- flag to store perturbations"
"\n matrixFileName -- matrix file name, format"
"\n nrow ncol nent"
"\n irow jcol entry"
"\n ..."
"\n note: indices are zero based"
"\n rhsFileName -- right hand side file name, format"
"\n nrow nrhs "
"\n ..."
"\n jrow entry(jrow,0) ... entry(jrow,nrhs-1)"
"\n ..."
"\n seed -- random number seed, used for ordering"
"\n nthread -- number of threads"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
type = atoi(argv[3]) ;
symmetryflag = atoi(argv[4]) ;
patchAndGoFlag = atoi(argv[5]) ;
fudge = atof(argv[6]) ;
toosmall = atof(argv[7]) ;
storeids = atoi(argv[8]) ;
storevalues = atoi(argv[9]) ;
matrixFileName = argv[10] ;
rhsFileName = argv[11] ;
seed = atoi(argv[12]) ;
nthread = atoi(argv[13]) ;
/*--------------------------------------------------------------------*/
/*
--------------------------------------------
STEP 1: read the entries from the input file
and create the InpMtx object
--------------------------------------------
*/
if ( (inputFile = fopen(matrixFileName, "r")) == NULL ) {
fprintf(stderr, "\n unable to open file %s", matrixFileName) ;
spoolesFatal();
}
fscanf(inputFile, "%d %d %d", &nrow, &ncol, &nent) ;
neqns = nrow ;
mtxA = InpMtx_new() ;
InpMtx_init(mtxA, INPMTX_BY_ROWS, type, nent, 0) ;
if ( type == SPOOLES_REAL ) {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le", &irow, &jcol, &value) ;
InpMtx_inputRealEntry(mtxA, irow, jcol, value) ;
}
} else {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le %le", &irow, &jcol, &real, &imag) ;
InpMtx_inputComplexEntry(mtxA, irow, jcol, real, imag) ;
}
}
fclose(inputFile) ;
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n input matrix") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------------------------
STEP 2 : find a low-fill ordering
(1) create the Graph object
(2) order the graph using multiple minimum degree
-------------------------------------------------
*/
graph = Graph_new() ;
adjIVL = InpMtx_fullAdjacency(mtxA) ;
nedges = IVL_tsize(adjIVL) ;
Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL,
NULL, NULL) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n graph of the input matrix") ;
Graph_writeForHumanEye(graph, msgFile) ;
fflush(msgFile) ;
}
frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n front tree from ordering") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------------------
STEP 3: get the permutation, permute the matrix and
front tree and get the symbolic factorization
-----------------------------------------------------
*/
oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ;
oldToNew = IV_entries(oldToNewIV) ;
newToOldIV = ETree_newToOldVtxPerm(frontETree) ;
newToOld = IV_entries(newToOldIV) ;
ETree_permuteVertices(frontETree, oldToNewIV) ;
InpMtx_permute(mtxA, oldToNew, oldToNew) ;
InpMtx_mapToUpperTriangle(mtxA) ;
InpMtx_changeCoordType(mtxA, INPMTX_BY_CHEVRONS) ;
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
symbfacIVL = SymbFac_initFromInpMtx(frontETree, mtxA) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n old-to-new permutation vector") ;
IV_writeForHumanEye(oldToNewIV, msgFile) ;
fprintf(msgFile, "\n\n new-to-old permutation vector") ;
IV_writeForHumanEye(newToOldIV, msgFile) ;
fprintf(msgFile, "\n\n front tree after permutation") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fprintf(msgFile, "\n\n input matrix after permutation") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fprintf(msgFile, "\n\n symbolic factorization") ;
IVL_writeForHumanEye(symbfacIVL, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------------------------
STEP 4: initialize the front matrix object
and the PatchAndGoInfo object to handle
small pivots
------------------------------------------
*/
frontmtx = FrontMtx_new() ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, LOCK_IN_PROCESS, 0) ;
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type, symmetryflag,
FRONTMTX_DENSE_FRONTS, SPOOLES_NO_PIVOTING, LOCK_IN_PROCESS,
0, NULL, mtxmanager, msglvl, msgFile) ;
if ( patchAndGoFlag == 1 ) {
frontmtx->patchinfo = PatchAndGoInfo_new() ;
PatchAndGoInfo_init(frontmtx->patchinfo, 1, toosmall, fudge,
storeids, storevalues) ;
} else if ( patchAndGoFlag == 2 ) {
frontmtx->patchinfo = PatchAndGoInfo_new() ;
PatchAndGoInfo_init(frontmtx->patchinfo, 2, toosmall, fudge,
storeids, storevalues) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------------------------
STEP 5: setup the domain decomposition map
------------------------------------------
*/
if ( nthread > (nfront = FrontMtx_nfront(frontmtx)) ) {
nthread = nfront ;
}
cumopsDV = DV_new() ;
DV_init(cumopsDV, nthread, NULL) ;
ownersIV = ETree_ddMap(frontETree, type, symmetryflag,
cumopsDV, 1./(2.*nthread)) ;
DV_free(cumopsDV) ;
/*--------------------------------------------------------------------*/
/*
-----------------------------------------------------
STEP 6: compute the numeric factorization in parallel
-----------------------------------------------------
*/
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, LOCK_IN_PROCESS, 1) ;
DVfill(10, cpus, 0.0) ;
IVfill(20, stats, 0) ;
lookahead = 0 ;
rootchv = FrontMtx_MT_factorInpMtx(frontmtx, mtxA, tau, 0.0,
chvmanager, ownersIV, lookahead,
&error, cpus, stats, msglvl, msgFile) ;
if ( patchAndGoFlag == 1 ) {
if ( frontmtx->patchinfo->fudgeIV != NULL ) {
fprintf(msgFile, "\n small pivots found at these locations") ;
IV_writeForHumanEye(frontmtx->patchinfo->fudgeIV, msgFile) ;
}
PatchAndGoInfo_free(frontmtx->patchinfo) ;
} else if ( patchAndGoFlag == 2 ) {
if ( frontmtx->patchinfo->fudgeIV != NULL ) {
fprintf(msgFile, "\n small pivots found at these locations") ;
IV_writeForHumanEye(frontmtx->patchinfo->fudgeIV, msgFile) ;
}
if ( frontmtx->patchinfo->fudgeDV != NULL ) {
fprintf(msgFile, "\n perturbations") ;
DV_writeForHumanEye(frontmtx->patchinfo->fudgeDV, msgFile) ;
}
PatchAndGoInfo_free(frontmtx->patchinfo) ;
}
ChvManager_free(chvmanager) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
if ( rootchv != NULL ) {
fprintf(msgFile, "\n\n matrix found to be singular\n") ;
spoolesFatal();
}
if ( error >= 0 ) {
fprintf(msgFile, "\n\n fatal error at front %d\n", error) ;
spoolesFatal();
}
/*
--------------------------------------
STEP 7: post-process the factorization
--------------------------------------
*/
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------
STEP 8: read the right hand side matrix B
-----------------------------------------
*/
if ( (inputFile = fopen(rhsFileName, "r")) == NULL ) {
fprintf(stderr, "\n unable to open file %s", rhsFileName) ;
spoolesFatal();
}
fscanf(inputFile, "%d %d", &nrow, &nrhs) ;
mtxB = DenseMtx_new() ;
DenseMtx_init(mtxB, type, 0, 0, neqns, nrhs, 1, neqns) ;
DenseMtx_zero(mtxB) ;
if ( type == SPOOLES_REAL ) {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le", &value) ;
DenseMtx_setRealEntry(mtxB, jrow, jrhs, value) ;
}
}
} else {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le %le", &real, &imag) ;
DenseMtx_setComplexEntry(mtxB, jrow, jrhs, real, imag) ;
}
}
}
fclose(inputFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n rhs matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
--------------------------------------------------------------
STEP 9: permute the right hand side into the original ordering
--------------------------------------------------------------
*/
DenseMtx_permuteRows(mtxB, oldToNewIV) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n right hand side matrix in new ordering") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
--------------------------------------------------------
STEP 10: get the solve map object for the parallel solve
--------------------------------------------------------
*/
solvemap = SolveMap_new() ;
SolveMap_ddMap(solvemap, type, FrontMtx_upperBlockIVL(frontmtx),
FrontMtx_lowerBlockIVL(frontmtx), nthread, ownersIV,
FrontMtx_frontTree(frontmtx), seed, msglvl, msgFile) ;
/*--------------------------------------------------------------------*/
/*
--------------------------------------------
STEP 11: solve the linear system in parallel
--------------------------------------------
*/
mtxX = DenseMtx_new() ;
DenseMtx_init(mtxX, type, 0, 0, neqns, nrhs, 1, neqns) ;
DenseMtx_zero(mtxX) ;
FrontMtx_MT_solve(frontmtx, mtxX, mtxB, mtxmanager, solvemap,
cpus, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n solution matrix in new ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
--------------------------------------------------------
STEP 12: permute the solution into the original ordering
--------------------------------------------------------
*/
DenseMtx_permuteRows(mtxX, newToOldIV) ;
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n solution matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------
free memory
-----------
*/
FrontMtx_free(frontmtx) ;
DenseMtx_free(mtxX) ;
DenseMtx_free(mtxB) ;
IV_free(newToOldIV) ;
IV_free(oldToNewIV) ;
InpMtx_free(mtxA) ;
ETree_free(frontETree) ;
IVL_free(symbfacIVL) ;
SubMtxManager_free(mtxmanager) ;
Graph_free(graph) ;
SolveMap_free(solvemap) ;
IV_free(ownersIV) ;
/*--------------------------------------------------------------------*/
return(1) ; }
/*
--------------------------------------------------
purpose -- to solve a linear system
(A - sigma*B) sol[] = rhs[]
data -- pointer to bridge data object
*pnrows -- # of rows in x[] and y[]
*pncols -- # of columns in x[] and y[]
rhs[] -- vector that holds right hand sides
NOTE: the rhs[] vector is global, not a portion
sol[] -- vector to hold solutions
NOTE: the sol[] vector is global, not a portion
note: rhs[] and sol[] can be the same array.
on return, *perror holds an error code.
created -- 98aug28, cca & jcp
--------------------------------------------------
*/
void
JimSolveMPI (
int *pnrows,
int *pncols,
double rhs[],
double sol[],
void *data,
int *perror
) {
BridgeMPI *bridge = (BridgeMPI *) data ;
DenseMtx *mtx, *newmtx ;
int irow, jj, jcol, kk, myid, ncols = *pncols,
neqns, nowned, tag = 0 ;
int *vtxmap ;
int stats[4] ;
IV *mapIV ;
#if MYDEBUG > 0
double t1, t2 ;
count_JimSolve++ ;
MARKTIME(t1) ;
if ( bridge->myid == 0 ) {
fprintf(stdout, "\n (%d) JimSolve() start", count_JimSolve) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile, "\n (%d) JimSolve() start", count_JimSolve) ;
fflush(bridge->msgFile) ;
#endif
MPI_Barrier(bridge->comm) ;
/*
---------------------------------------------
slide the owned rows of rhs down in the array
---------------------------------------------
*/
vtxmap = IV_entries(bridge->vtxmapIV) ;
neqns = bridge->neqns ;
myid = bridge->myid ;
nowned = IV_size(bridge->myownedIV) ;
for ( jcol = jj = kk = 0 ; jcol < ncols ; jcol++ ) {
for ( irow = 0 ; irow < neqns ; irow++, jj++ ) {
if ( vtxmap[irow] == myid ) {
sol[kk++] = rhs[jj] ;
}
}
}
if ( kk != nowned * ncols ) {
fprintf(stderr, "\n proc %d : kk %d, nowned %d, ncols %d",
myid, kk, nowned, ncols) ;
exit(-1) ;
}
/*
----------------------------------------
call the method that assumes local input
----------------------------------------
*/
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n calling SolveMPI()") ;
fflush(bridge->msgFile) ;
}
SolveMPI(&nowned, pncols, sol, sol, data, perror) ;
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n return from SolveMPI()") ;
fflush(bridge->msgFile) ;
}
/*
------------------------------------------
gather all the entries onto processor zero
------------------------------------------
*/
mtx = DenseMtx_new() ;
DenseMtx_init(mtx, SPOOLES_REAL, 0, 0, nowned, ncols, 1, nowned) ;
DVcopy (nowned*ncols, DenseMtx_entries(mtx), sol) ;
IVcopy(nowned, mtx->rowind, IV_entries(bridge->myownedIV)) ;
mapIV = IV_new() ;
IV_init(mapIV, neqns, NULL) ;
IV_fill(mapIV, 0) ;
IVfill(4, stats, 0) ;
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n calling DenseMtx_split()()") ;
fflush(bridge->msgFile) ;
}
newmtx = DenseMtx_MPI_splitByRows(mtx, mapIV, stats, bridge->msglvl,
bridge->msgFile, tag, bridge->comm) ;
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n return from DenseMtx_split()()") ;
fflush(bridge->msgFile) ;
}
DenseMtx_free(mtx) ;
mtx = newmtx ;
IV_free(mapIV) ;
if ( myid == 0 ) {
DVcopy(neqns*ncols, sol, DenseMtx_entries(mtx)) ;
}
DenseMtx_free(mtx) ;
/*
---------------------------------------------
broadcast the entries to the other processors
---------------------------------------------
*/
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n calling MPI_Bcast()()") ;
fflush(bridge->msgFile) ;
}
MPI_Bcast((void *) sol, neqns*ncols, MPI_DOUBLE, 0, bridge->comm) ;
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n return from MPI_Bcast()()") ;
fflush(bridge->msgFile) ;
}
MPI_Barrier(bridge->comm) ;
/*
------------------------------------------------------------------
set the error. (this is simple since when the spooles codes detect
a fatal error, they print out a message to stderr and exit.)
------------------------------------------------------------------
*/
*perror = 0 ;
#if MYDEBUG > 0
MARKTIME(t2) ;
time_JimSolve += t2 - t1 ;
if ( bridge->myid == 0 ) {
fprintf(stdout, "\n (%d) JimSolve() end", count_JimSolve) ;
fprintf(stdout, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_JimSolve) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile, "\n (%d) JimSolve() end", count_JimSolve) ;
fprintf(bridge->msgFile, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_JimSolve) ;
fflush(bridge->msgFile) ;
#endif
return ; }
/*--------------------------------------------------------------------*/
static void
InpMtx_MT_mmm (
int flag,
InpMtx *A,
DenseMtx *Y,
double alpha[],
DenseMtx *X,
int nthread,
int msglvl,
FILE *msgFile
) {
double t1, t2 ;
int myid, nent, rc ;
MTmvmObj *MTmvmObjs, *obj ;
/*
-------------------------------
set up the nthread data objects
-------------------------------
*/
MARKTIME(t1) ;
MTmvmObjs = setup(A, Y, alpha, X, nthread) ;
MARKTIME(t2) ;
if ( msglvl > 0 ) {
fprintf(msgFile, "\n %% CPU %8.3f : setup time", t2 - t1) ;
}
#if THREAD_TYPE == TT_POSIX
{
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 SGI needs it anyway
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) ;
MARKTIME(t1) ;
for ( myid = 0, obj = MTmvmObjs ; myid < nthread ; myid++, obj++ ) {
switch ( flag ) {
case NONSYM :
rc = pthread_create(&tids[myid], &attr, worker_nonsym_mmm, obj) ;
break ;
case SYM :
rc = pthread_create(&tids[myid], &attr, worker_sym_mmm, obj) ;
break ;
case HERM :
rc = pthread_create(&tids[myid], &attr, worker_herm_mmm, obj) ;
break ;
case NONSYM_T :
rc = pthread_create(&tids[myid], &attr, worker_nonsym_mmm_T, obj);
break ;
case NONSYM_H :
rc = pthread_create(&tids[myid], &attr, worker_nonsym_mmm_H, obj);
break ;
}
if ( rc != 0 ) {
fprintf(stderr,
"\n fatal error, myid = %d, rc = %d from pthread_create",
myid, rc) ;
exit(-1) ;
} else if ( msglvl > 2 ) {
fprintf(stderr, "\n %% thread %d created", myid) ;
}
}
MARKTIME(t2) ;
if ( msglvl > 0 ) {
fprintf(msgFile, "\n %% CPU %8.3f : thread creation time", t2 - t1) ;
}
MARKTIME(t1) ;
for ( myid = 0 ; myid < nthread ; myid++ ) {
pthread_join(tids[myid], &status) ;
}
MARKTIME(t2) ;
if ( msglvl > 0 ) {
fprintf(msgFile, "\n %% CPU %8.3f : thread join time", t2 - t1) ;
}
FREE(tids) ;
pthread_attr_destroy(&attr) ;
}
#endif
/*
-------------------------------------
accumulate the rhs hand side matrices
-------------------------------------
*/
MARKTIME(t1) ;
nent = Y->nrow * Y->ncol ;
for ( myid = 1, obj = MTmvmObjs + 1 ;
myid < nthread ;
myid++, obj++ ) {
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
DVadd(nent, DenseMtx_entries(Y), DenseMtx_entries(obj->Y)) ;
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
DVadd(2*nent, DenseMtx_entries(Y), DenseMtx_entries(obj->Y)) ;
}
}
MARKTIME(t2) ;
if ( msglvl > 0 ) {
fprintf(msgFile,
"\n %% CPU %8.3f : time to accumulate rhs", t2 - t1) ;
}
/*
---------------------------
release the data structures
---------------------------
*/
MARKTIME(t1) ;
for ( myid = 0, obj = MTmvmObjs ; myid < nthread ; myid++, obj++ ) {
InpMtx_free(obj->A) ;
if ( myid > 0 ) {
DenseMtx_free(obj->Y) ;
}
}
FREE(MTmvmObjs) ;
MARKTIME(t2) ;
if ( msglvl > 0 ) {
fprintf(msgFile,
"\n %% CPU %8.3f : time to release and free data", t2 - t1) ;
}
return ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
-------------------------------------------------------
test the DenseMtx_mmm routine.
C = alpha*A*B + beta*C, where A, B and C are DenseMtx.
alpha and beta are scalars.
when msglvl > 1, the output of this program
can be fed into Matlab to check for errors
created -- 98dec14, ycp
-------------------------------------------------------
*/
{
DenseMtx *mtxA, *mtxB, *mtxC;
double t1, t2, value[2] = {1.0, 1.0} ;
Drand *drand ;
FILE *msgFile ;
int i, j, k, msglvl, nrow, nk, ncol, cnrow, cncol, seed, type ;
int ainc1, ainc2, binc1, binc2, cinc1, cinc2;
double alpha[2], beta[2], one[2] = {1.0, 0.0}, rvalue;
char A_opt[1]=" ", B_opt[1]=" ";
if ( argc != 20 ) {
fprintf(stdout,
"\n\n usage : %s msglvl msgFile type nrow nk ncol ainc1 ainc2 binc1 "
"\n binc2 cinc1 cinc2 A_opt B_opt ralpha ialpha rbeta ibeta seed "
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n type -- entries type"
"\n 1 -- real"
"\n 2 -- complex"
"\n nrow -- # of rows of mtxA "
"\n nk -- # of columns of mtxA "
"\n ncol -- # of columns of mtxB "
"\n ainc1 -- A row increment "
"\n ainc2 -- A column increment "
"\n binc1 -- B row increment "
"\n binc2 -- B column increment "
"\n binc1 -- C row increment "
"\n binc2 -- C column increment "
"\n A_opt -- A option "
"\n B_opt -- B option "
"\n ralpha -- real(alpha)"
"\n ialpha -- imag(alpha)"
"\n rbeta -- real(beta)"
"\n ibeta -- imag(beta)"
"\n seed -- random number seed"
"\n", argv[0]) ;
return(0) ;
}
if ( (msglvl = atoi(argv[1])) < 0 ) {
fprintf(stderr, "\n message level must be positive\n") ;
spoolesFatal();
}
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n unable to open file %s\n", argv[2]) ;
return(-1) ;
}
type = atoi(argv[3]) ;
nrow = atoi(argv[4]) ;
nk = atoi(argv[5]) ;
ncol = atoi(argv[6]) ;
ainc1= atoi(argv[7]) ;
ainc2= atoi(argv[8]) ;
binc1= atoi(argv[9]) ;
binc2= atoi(argv[10]) ;
cinc1= atoi(argv[11]) ;
cinc2= atoi(argv[12]) ;
if ( type < 1 || type > 2 || nrow < 0 || ncol < 0 ||
ainc1 < 1 || ainc2 < 1 || binc1 < 1 || binc2 < 1 ) {
fprintf(stderr,
"\n fatal error, type %d, nrow %d, ncol %d, ainc1 %d, ainc2 %d"
", binc1 %d, binc2 %d", type, nrow, ncol, ainc1, ainc2, binc1, binc2) ;
spoolesFatal();
}
A_opt[0] = *argv[13] ;
B_opt[0] = *argv[14] ;
alpha[0]= atof (argv[15]);
alpha[1]= atof (argv[16]);
beta[0] = atof (argv[17]);
beta[1] = atof (argv[18]);
seed = atoi (argv[19]) ;
fprintf(msgFile, "\n\n %% %s :"
"\n %% msglvl = %d"
"\n %% msgFile = %s"
"\n %% type = %d"
"\n %% nrow = %d"
"\n %% nk = %d"
"\n %% ncol = %d"
"\n %% ainc1 = %d"
"\n %% ainc2 = %d"
"\n %% binc1 = %d"
"\n %% binc2 = %d"
"\n %% cinc1 = %d"
"\n %% cinc2 = %d"
"\n %% a_opt = %c"
"\n %% b_opt = %c"
"\n %% ralpha = %e"
"\n %% ialpha = %e"
"\n %% rbeta = %e"
"\n %% ibeta = %e"
"\n %% seed = %d"
"\n",
argv[0], msglvl, argv[2], type, nrow, nk, ncol, ainc1, ainc2,
binc1, binc2, cinc1, cinc2, A_opt[0], B_opt[0], alpha[0],
alpha[1], beta[0], beta[1], seed) ;
/*
----------------------------
initialize the matrix object
----------------------------
*/
MARKTIME(t1) ;
mtxA = DenseMtx_new() ;
DenseMtx_init(mtxA, type, 0, 0, nrow, nk, ainc1, ainc2) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n %% CPU : %.3f to initialize matrix object",
t2 - t1) ;
MARKTIME(t1) ;
drand = Drand_new() ;
Drand_setSeed(drand, seed) ;
seed++ ;
Drand_setUniform(drand, -1.0, 1.0) ;
DenseMtx_fillRandomEntries(mtxA, drand) ;
MARKTIME(t2) ;
fprintf(msgFile,
"\n %% CPU : %.3f to fill matrix A with random numbers", t2 - t1) ;
MARKTIME(t1) ;
mtxB = DenseMtx_new() ;
DenseMtx_init(mtxB, type, 0, 0, nk, ncol, binc1, binc2) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n %% CPU : %.3f to initialize matrix object",
t2 - t1) ;
MARKTIME(t1) ;
drand = Drand_new() ;
Drand_setSeed(drand, seed) ;
seed++ ;
Drand_setUniform(drand, -1.0, 1.0) ;
DenseMtx_fillRandomEntries(mtxB, drand) ;
MARKTIME(t2) ;
fprintf(msgFile,
"\n %% CPU : %.3f to fill matrix B with random numbers", t2 - t1) ;
cnrow = nrow;
cncol = ncol;
MARKTIME(t1) ;
mtxC = DenseMtx_new() ;
if ( A_opt[0] == 't' || A_opt[0] == 'T' ||
A_opt[0] == 'c' || A_opt[0] == 'C') {
cnrow = nk;
}
if ( B_opt[0] == 't' || B_opt[0] == 'T' ||
B_opt[0] == 'c' || B_opt[0] == 'C') {
cncol = nk;
}
if ( cinc1 == 1 && cinc2 == nrow ){ /* stored by column */
cinc1 = 1;
cinc2 = cnrow;
} else { /* stored by row */
cinc1 = cncol;
cinc2 = 1;
}
DenseMtx_init(mtxC, type, 0, 0, cnrow, cncol, cinc1, cinc2) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n %% CPU : %.3f to initialize matrix object",
t2 - t1) ;
MARKTIME(t1) ;
drand = Drand_new() ;
Drand_setSeed(drand, seed) ;
seed++ ;
Drand_setUniform(drand, -1.0, 1.0) ;
DenseMtx_fillRandomEntries(mtxC, drand) ;
MARKTIME(t2) ;
fprintf(msgFile,
"\n %% CPU : %.3f to fill matrix C with random numbers", t2 - t1) ;
if ( msglvl > 3 ) {
fprintf(msgFile, "\n matrix A") ;
DenseMtx_writeForHumanEye(mtxA, msgFile) ;
fprintf(msgFile, "\n matrix B") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fprintf(msgFile, "\n matrix C") ;
DenseMtx_writeForHumanEye(mtxC, msgFile) ;
}
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n %% beta = (%f, %f)", beta[0], beta[1]) ;
fprintf(msgFile, "\n %% alpha = (%f, %f)\n", alpha[0], alpha[1]) ;
fprintf(msgFile, "\n %% matrix A") ;
fprintf(msgFile, "\n nrow = %d ;", nrow) ;
fprintf(msgFile, "\n ncol = %d ;", nk) ;
DenseMtx_writeForMatlab(mtxA, "A", msgFile) ;
fprintf(msgFile, "\n");
fprintf(msgFile, "\n %% matrix B") ;
fprintf(msgFile, "\n nrow = %d ;", nk) ;
fprintf(msgFile, "\n ncol = %d ;", ncol) ;
DenseMtx_writeForMatlab(mtxB, "B", msgFile) ;
fprintf(msgFile, "\n");
fprintf(msgFile, "\n %% matrix C") ;
fprintf(msgFile, "\n nrow = %d ;", cnrow) ;
fprintf(msgFile, "\n ncol = %d ;", cncol) ;
DenseMtx_writeForMatlab(mtxC, "C", msgFile) ;
}
/*
--------------------------
performs the matrix-matrix operations
C = alpha*(A)*(B) + beta*C
--------------------------
*/
DenseMtx_mmm(A_opt, B_opt, &beta, mtxC, &alpha, mtxA, mtxB);
if ( msglvl > 1 ) {
fprintf(msgFile, "\n");
fprintf(msgFile, "\n %% *** Output matrix C ***") ;
fprintf(msgFile, "\n nrow = %d ;", cnrow) ;
fprintf(msgFile, "\n ncol = %d ;", cncol) ;
DenseMtx_writeForMatlab(mtxC, "C", msgFile) ;
fprintf(msgFile, "\n");
fflush(msgFile) ;
}
/*
------------------------
free the working storage
------------------------
*/
DenseMtx_free(mtxA) ;
DenseMtx_free(mtxB) ;
DenseMtx_free(mtxC) ;
Drand_free(drand) ;
return(1) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
---------------------------------------------------
test the QR factor method for a FrontMtx object
on an n1 x n2 x n3 grid
(1) generate an overdetermined system AX = B
(2) factor the matrix
(3) solve the systems
created -- 97apr11, dkw
modified -- 98may28, cca
---------------------------------------------------
*/
{
ChvManager *chvmanager ;
DenseMtx *mtxB, *mtxX, *mtxZ ;
double cputotal, factorops ;
double cpus[9] ;
double nops, t1, t2 ;
ETree *frontETree ;
FILE *msgFile ;
FrontMtx *frontmtx ;
InpMtx *mtxA ;
int msglvl, neqns, nrhs, n1, n2, n3, seed, type ;
IVL *symbfacIVL ;
SubMtxManager *mtxmanager ;
if ( argc != 9 ) {
fprintf(stdout,
"\n\n usage : %s msglvl msgFile n1 n2 n3 seed nrhs "
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n n1 -- # of points in the first direction"
"\n n2 -- # of points in the second direction"
"\n n3 -- # of points in the third direction"
"\n seed -- random number seed"
"\n nrhs -- # of right hand sides"
"\n type -- type of linear system"
"\n 1 -- real"
"\n 2 -- complex"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
n1 = atoi(argv[3]) ;
n2 = atoi(argv[4]) ;
n3 = atoi(argv[5]) ;
seed = atoi(argv[6]) ;
nrhs = atoi(argv[7]) ;
type = atoi(argv[8]) ;
fprintf(msgFile,
"\n %s "
"\n msglvl -- %d"
"\n msgFile -- %s"
"\n n1 -- %d"
"\n n2 -- %d"
"\n n3 -- %d"
"\n seed -- %d"
"\n nrhs -- %d"
"\n type -- %d"
"\n",
argv[0], msglvl, argv[2], n1, n2, n3, seed, nrhs, type) ;
fflush(msgFile) ;
neqns = n1*n2*n3 ;
if ( type != SPOOLES_REAL && type != SPOOLES_COMPLEX ) {
fprintf(stderr, "\n fatal error, type must be real or complex") ;
exit(-1) ;
}
/*
------------------------------------------
generate the A X = B overdetermined system
------------------------------------------
*/
mkNDlinsysQR(n1, n2, n3, type, nrhs, seed, msglvl, msgFile,
&frontETree, &symbfacIVL, &mtxA, &mtxX, &mtxB) ;
/*
------------------------------
initialize the FrontMtx object
------------------------------
*/
MARKTIME(t1) ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, NO_LOCK, 0) ;
frontmtx = FrontMtx_new() ;
if ( type == SPOOLES_REAL ) {
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type,
SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS,
SPOOLES_NO_PIVOTING, NO_LOCK,
0, NULL, mtxmanager, msglvl, msgFile) ;
} else if ( type == SPOOLES_COMPLEX ) {
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type,
SPOOLES_HERMITIAN, FRONTMTX_DENSE_FRONTS,
SPOOLES_NO_PIVOTING, NO_LOCK,
0, NULL, mtxmanager, msglvl, msgFile) ;
}
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : FrontMtx initialized", t2 - t1) ;
fflush(msgFile) ;
/*
-----------------
factor the matrix
-----------------
*/
DVzero(6, cpus) ;
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, NO_LOCK, 0) ;
factorops = 0.0 ;
MARKTIME(t1) ;
FrontMtx_QR_factor(frontmtx, mtxA, chvmanager,
cpus, &factorops, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n after QR_factor() call, facops = %8.2f",factorops) ;
fprintf(msgFile, "\n CPU %8.3f : FrontMtx_QR_factor, %8.3f mflops",
t2 - t1, 1.e-6*factorops/(t2-t1)) ;
cputotal = t2 - t1 ;
if ( cputotal > 0.0 ) {
fprintf(msgFile, "\n"
"\n setup factorization %8.3f %6.2f"
"\n setup fronts %8.3f %6.2f"
"\n factor fronts %8.3f %6.2f"
"\n store factor %8.3f %6.2f"
"\n store update %8.3f %6.2f"
"\n miscellaneous %8.3f %6.2f"
"\n total time %8.3f",
cpus[0], 100.*cpus[0]/cputotal,
cpus[1], 100.*cpus[1]/cputotal,
cpus[2], 100.*cpus[2]/cputotal,
cpus[3], 100.*cpus[3]/cputotal,
cpus[4], 100.*cpus[4]/cputotal,
cpus[5], 100.*cpus[5]/cputotal, cputotal) ;
}
/*
------------------------------
post-process the factor matrix
------------------------------
*/
MARKTIME(t1) ;
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : post-process the matrix", t2 - t1)
;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
}
fprintf(msgFile, "\n\n after post-processing") ;
SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
/*
----------------
solve the system
----------------
*/
mtxZ = DenseMtx_new() ;
DenseMtx_init(mtxZ, type, 0, 0, neqns, nrhs, 1, neqns) ;
DenseMtx_zero(mtxZ) ;
if ( type == SPOOLES_REAL ) {
nops = frontmtx->nentD + 2*frontmtx->nentU ;
if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) {
nops += 2*frontmtx->nentL ;
} else {
nops += 2*frontmtx->nentU ;
}
} else if ( type == SPOOLES_COMPLEX ) {
nops = 8*frontmtx->nentD + 8*frontmtx->nentU ;
if ( FRONTMTX_IS_NONSYMMETRIC(frontmtx) ) {
nops += 8*frontmtx->nentL ;
} else {
nops += 8*frontmtx->nentU ;
}
}
nops *= nrhs ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n rhs") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(stdout) ;
}
DVzero(6, cpus) ;
MARKTIME(t1) ;
FrontMtx_QR_solve(frontmtx, mtxA, mtxZ, mtxB, mtxmanager,
cpus, msglvl, msgFile) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %8.3f : solve the system, %.3f mflops",
t2 - t1, 1.e-6*nops/(t2 - t1)) ;
cputotal = t2 - t1 ;
if ( cputotal > 0.0 ) {
fprintf(msgFile,
"\n CPU %%"
"\n A^TB matrix-matrix multiply %8.3f %6.2f"
"\n set up solves %8.3f %6.2f"
"\n load rhs and store solution %8.3f %6.2f"
"\n forward solve %8.3f %6.2f"
"\n diagonal solve %8.3f %6.2f"
"\n backward solve %8.3f %6.2f"
"\n total solve time %8.3f %6.2f"
"\n total QR solve time %8.3f",
cpus[6], 100.*cpus[6]/cputotal,
cpus[0], 100.*cpus[0]/cputotal,
cpus[1], 100.*cpus[1]/cputotal,
cpus[2], 100.*cpus[2]/cputotal,
cpus[3], 100.*cpus[3]/cputotal,
cpus[4], 100.*cpus[4]/cputotal,
cpus[5], 100.*cpus[5]/cputotal, cputotal) ;
}
if ( msglvl > 3 ) {
fprintf(msgFile, "\n\n computed solution") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
/*
-----------------
compute the error
-----------------
*/
DenseMtx_sub(mtxZ, mtxX) ;
fprintf(msgFile, "\n\n maxabs error = %12.4e",
DenseMtx_maxabs(mtxZ)) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n error") ;
DenseMtx_writeForHumanEye(mtxZ, msgFile) ;
fflush(stdout) ;
}
fprintf(msgFile, "\n\n after solve") ;
SubMtxManager_writeForHumanEye(frontmtx->manager, msgFile) ;
/*
------------------------
free the working storage
------------------------
*/
InpMtx_free(mtxA) ;
DenseMtx_free(mtxX) ;
DenseMtx_free(mtxZ) ;
DenseMtx_free(mtxB) ;
FrontMtx_free(frontmtx) ;
IVL_free(symbfacIVL) ;
ETree_free(frontETree) ;
SubMtxManager_free(mtxmanager) ;
ChvManager_free(chvmanager) ;
fprintf(msgFile, "\n") ;
fclose(msgFile) ;
return(1) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
------------------------------------------------------------------
generate a random matrix and test a matrix-matrix multiply method.
the output is a matlab file to test correctness.
created -- 98jan29, cca
--------------------------------------------------------------------
*/
{
DenseMtx *X, *Y, *Y2 ;
double alpha[2] ;
double alphaImag, alphaReal, t1, t2 ;
double *zvec ;
Drand *drand ;
int col, dataType, ii, msglvl, ncolA, nitem, nops, nrhs,
nrowA, nrowX, nrowY, nthread, row, seed,
storageMode, symflag, transposeflag ;
int *colids, *rowids ;
InpMtx *A ;
FILE *msgFile ;
if ( argc != 15 ) {
fprintf(stdout,
"\n\n %% usage : %s msglvl msgFile symflag storageMode "
"\n %% nrow ncol nent nrhs seed alphaReal alphaImag nthread"
"\n %% msglvl -- message level"
"\n %% msgFile -- message file"
"\n %% dataType -- type of matrix entries"
"\n %% 1 -- real"
"\n %% 2 -- complex"
"\n %% symflag -- symmetry flag"
"\n %% 0 -- symmetric"
"\n %% 1 -- hermitian"
"\n %% 2 -- nonsymmetric"
"\n %% storageMode -- storage mode"
"\n %% 1 -- by rows"
"\n %% 2 -- by columns"
"\n %% 3 -- by chevrons, (requires nrow = ncol)"
"\n %% transpose -- transpose flag"
"\n %% 0 -- Y := Y + alpha * A * X"
"\n %% 1 -- Y := Y + alpha * A^H * X, nonsymmetric only"
"\n %% 2 -- Y := Y + alpha * A^T * X, nonsymmetric only"
"\n %% nrowA -- number of rows in A"
"\n %% ncolA -- number of columns in A"
"\n %% nitem -- number of items"
"\n %% nrhs -- number of right hand sides"
"\n %% seed -- random number seed"
"\n %% alphaReal -- y := y + alpha*A*x"
"\n %% alphaImag -- y := y + alpha*A*x"
"\n %% nthread -- # of threads"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
dataType = atoi(argv[3]) ;
symflag = atoi(argv[4]) ;
storageMode = atoi(argv[5]) ;
transposeflag = atoi(argv[6]) ;
nrowA = atoi(argv[7]) ;
ncolA = atoi(argv[8]) ;
nitem = atoi(argv[9]) ;
nrhs = atoi(argv[10]) ;
seed = atoi(argv[11]) ;
alphaReal = atof(argv[12]) ;
alphaImag = atof(argv[13]) ;
nthread = atoi(argv[14]) ;
fprintf(msgFile,
"\n %% %s "
"\n %% msglvl -- %d"
"\n %% msgFile -- %s"
"\n %% dataType -- %d"
"\n %% symflag -- %d"
"\n %% storageMode -- %d"
"\n %% transposeflag -- %d"
"\n %% nrowA -- %d"
"\n %% ncolA -- %d"
"\n %% nitem -- %d"
"\n %% nrhs -- %d"
"\n %% seed -- %d"
"\n %% alphaReal -- %e"
"\n %% alphaImag -- %e"
"\n %% nthread -- %d"
"\n",
argv[0], msglvl, argv[2], dataType, symflag, storageMode,
transposeflag, nrowA, ncolA, nitem, nrhs, seed,
alphaReal, alphaImag, nthread) ;
fflush(msgFile) ;
if ( dataType != 1 && dataType != 2 ) {
fprintf(stderr, "\n invalid value %d for dataType\n", dataType) ;
spoolesFatal();
}
if ( symflag != 0 && symflag != 1 && symflag != 2 ) {
fprintf(stderr, "\n invalid value %d for symflag\n", symflag) ;
spoolesFatal();
}
if ( storageMode != 1 && storageMode != 2 && storageMode != 3 ) {
fprintf(stderr,
"\n invalid value %d for storageMode\n", storageMode) ;
spoolesFatal();
}
if ( transposeflag < 0
|| transposeflag > 2 ) {
fprintf(stderr, "\n error, transposeflag = %d, must be 0, 1 or 2",
transposeflag) ;
spoolesFatal();
}
if ( (transposeflag == 1 && symflag != 2)
|| (transposeflag == 2 && symflag != 2) ) {
fprintf(stderr, "\n error, transposeflag = %d, symflag = %d",
transposeflag, symflag) ;
spoolesFatal();
}
if ( transposeflag == 1 && dataType != 2 ) {
fprintf(stderr, "\n error, transposeflag = %d, dataType = %d",
transposeflag, dataType) ;
spoolesFatal();
}
if ( symflag == 1 && dataType != 2 ) {
fprintf(stderr,
"\n symflag = 1 (hermitian), dataType != 2 (complex)") ;
spoolesFatal();
}
if ( nrowA <= 0 || ncolA <= 0 || nitem <= 0 ) {
fprintf(stderr,
"\n invalid value: nrow = %d, ncol = %d, nitem = %d",
nrowA, ncolA, nitem) ;
spoolesFatal();
}
if ( symflag < 2 && nrowA != ncolA ) {
fprintf(stderr,
"\n invalid data: symflag = %d, nrow = %d, ncol = %d",
symflag, nrowA, ncolA) ;
spoolesFatal();
}
alpha[0] = alphaReal ;
alpha[1] = alphaImag ;
/*
----------------------------
initialize the matrix object
----------------------------
*/
A = InpMtx_new() ;
InpMtx_init(A, storageMode, dataType, 0, 0) ;
drand = Drand_new() ;
/*
----------------------------------
generate a vector of nitem triples
----------------------------------
*/
rowids = IVinit(nitem, -1) ;
Drand_setUniform(drand, 0, nrowA) ;
Drand_fillIvector(drand, nitem, rowids) ;
colids = IVinit(nitem, -1) ;
Drand_setUniform(drand, 0, ncolA) ;
Drand_fillIvector(drand, nitem, colids) ;
Drand_setUniform(drand, 0.0, 1.0) ;
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
zvec = DVinit(nitem, 0.0) ;
Drand_fillDvector(drand, nitem, zvec) ;
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
zvec = ZVinit(nitem, 0.0, 0.0) ;
Drand_fillDvector(drand, 2*nitem, zvec) ;
}
/*
-----------------------------------
assemble the entries entry by entry
-----------------------------------
*/
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n A = zeros(%d,%d) ;", nrowA, ncolA) ;
}
if ( symflag == 1 ) {
/*
----------------
hermitian matrix
----------------
*/
for ( ii = 0 ; ii < nitem ; ii++ ) {
if ( rowids[ii] == colids[ii] ) {
zvec[2*ii+1] = 0.0 ;
}
if ( rowids[ii] <= colids[ii] ) {
row = rowids[ii] ; col = colids[ii] ;
} else {
row = colids[ii] ; col = rowids[ii] ;
}
InpMtx_inputComplexEntry(A, row, col, zvec[2*ii], zvec[2*ii+1]) ;
}
} else if ( symflag == 0 ) {
/*
----------------
symmetric matrix
----------------
*/
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
for ( ii = 0 ; ii < nitem ; ii++ ) {
if ( rowids[ii] <= colids[ii] ) {
row = rowids[ii] ; col = colids[ii] ;
} else {
row = colids[ii] ; col = rowids[ii] ;
}
InpMtx_inputRealEntry(A, row, col, zvec[ii]) ;
}
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
for ( ii = 0 ; ii < nitem ; ii++ ) {
if ( rowids[ii] <= colids[ii] ) {
row = rowids[ii] ; col = colids[ii] ;
} else {
row = colids[ii] ; col = rowids[ii] ;
}
InpMtx_inputComplexEntry(A, row, col,
zvec[2*ii], zvec[2*ii+1]) ;
}
}
} else {
/*
-------------------
nonsymmetric matrix
-------------------
*/
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
for ( ii = 0 ; ii < nitem ; ii++ ) {
InpMtx_inputRealEntry(A, rowids[ii], colids[ii], zvec[ii]) ;
}
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
for ( ii = 0 ; ii < nitem ; ii++ ) {
InpMtx_inputComplexEntry(A, rowids[ii], colids[ii],
zvec[2*ii], zvec[2*ii+1]) ;
}
}
}
InpMtx_changeStorageMode(A, INPMTX_BY_VECTORS) ;
DVfree(zvec) ;
if ( symflag == 0 || symflag == 1 ) {
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
nops = 4*A->nent*nrhs ;
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
nops = 16*A->nent*nrhs ;
}
} else {
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
nops = 2*A->nent*nrhs ;
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
nops = 8*A->nent*nrhs ;
}
}
if ( msglvl > 1 ) {
/*
-------------------------------------------
write the assembled matrix to a matlab file
-------------------------------------------
*/
InpMtx_writeForMatlab(A, "A", msgFile) ;
if ( symflag == 0 ) {
fprintf(msgFile,
"\n for k = 1:%d"
"\n for j = k+1:%d"
"\n A(j,k) = A(k,j) ;"
"\n end"
"\n end", nrowA, ncolA) ;
} else if ( symflag == 1 ) {
fprintf(msgFile,
"\n for k = 1:%d"
"\n for j = k+1:%d"
"\n A(j,k) = ctranspose(A(k,j)) ;"
"\n end"
"\n end", nrowA, ncolA) ;
}
}
/*
-------------------------------
generate dense matrices X and Y
-------------------------------
*/
if ( transposeflag == 0 ) {
nrowX = ncolA ;
nrowY = nrowA ;
} else {
nrowX = nrowA ;
nrowY = ncolA ;
}
X = DenseMtx_new() ;
Y = DenseMtx_new() ;
Y2 = DenseMtx_new() ;
if ( INPMTX_IS_REAL_ENTRIES(A) ) {
DenseMtx_init(X, SPOOLES_REAL, 0, 0, nrowX, nrhs, 1, nrowX) ;
Drand_fillDvector(drand, nrowX*nrhs, DenseMtx_entries(X)) ;
DenseMtx_init(Y, SPOOLES_REAL, 0, 0, nrowY, nrhs, 1, nrowY) ;
Drand_fillDvector(drand, nrowY*nrhs, DenseMtx_entries(Y)) ;
DenseMtx_init(Y2, SPOOLES_REAL, 0, 0, nrowY, nrhs, 1, nrowY) ;
DVcopy(nrowY*nrhs, DenseMtx_entries(Y2), DenseMtx_entries(Y)) ;
} else if ( INPMTX_IS_COMPLEX_ENTRIES(A) ) {
DenseMtx_init(X, SPOOLES_COMPLEX, 0, 0, nrowX, nrhs, 1, nrowX) ;
Drand_fillDvector(drand, 2*nrowX*nrhs, DenseMtx_entries(X)) ;
DenseMtx_init(Y, SPOOLES_COMPLEX, 0, 0, nrowY, nrhs, 1, nrowY) ;
Drand_fillDvector(drand, 2*nrowY*nrhs, DenseMtx_entries(Y)) ;
DenseMtx_init(Y2, SPOOLES_COMPLEX, 0, 0, nrowY, nrhs, 1, nrowY) ;
DVcopy(2*nrowY*nrhs, DenseMtx_entries(Y2), DenseMtx_entries(Y)) ;
}
if ( msglvl > 1 ) {
fprintf(msgFile, "\n X = zeros(%d,%d) ;", nrowX, nrhs) ;
DenseMtx_writeForMatlab(X, "X", msgFile) ;
fprintf(msgFile, "\n Y = zeros(%d,%d) ;", nrowY, nrhs) ;
DenseMtx_writeForMatlab(Y, "Y", msgFile) ;
}
/*
--------------------------------------------
perform the matrix-matrix multiply in serial
--------------------------------------------
*/
if ( msglvl > 1 ) {
fprintf(msgFile, "\n alpha = %20.12e + %20.2e*i;",
alpha[0], alpha[1]);
fprintf(msgFile, "\n Z = zeros(%d,1) ;", nrowY) ;
}
if ( transposeflag == 0 ) {
MARKTIME(t1) ;
if ( symflag == 0 ) {
InpMtx_sym_mmm(A, Y, alpha, X) ;
} else if ( symflag == 1 ) {
InpMtx_herm_mmm(A, Y, alpha, X) ;
} else if ( symflag == 2 ) {
InpMtx_nonsym_mmm(A, Y, alpha, X) ;
}
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y, "Z", msgFile) ;
fprintf(msgFile, "\n maxerr = max(Z - Y - alpha*A*X) ") ;
fprintf(msgFile, "\n") ;
}
} else if ( transposeflag == 1 ) {
MARKTIME(t1) ;
InpMtx_nonsym_mmm_H(A, Y, alpha, X) ;
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y, "Z", msgFile) ;
fprintf(msgFile,
"\n maxerr = max(Z - Y - alpha*ctranspose(A)*X) ") ;
fprintf(msgFile, "\n") ;
}
} else if ( transposeflag == 2 ) {
MARKTIME(t1) ;
InpMtx_nonsym_mmm_T(A, Y, alpha, X) ;
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y, "Z", msgFile) ;
fprintf(msgFile,
"\n maxerr = max(Z - Y - alpha*transpose(A)*X) ") ;
fprintf(msgFile, "\n") ;
}
}
fprintf(msgFile, "\n %% %d ops, %.3f time, %.3f serial mflops",
nops, t2 - t1, 1.e-6*nops/(t2 - t1)) ;
/*
--------------------------------------------------------
perform the matrix-matrix multiply in multithreaded mode
--------------------------------------------------------
*/
if ( msglvl > 1 ) {
fprintf(msgFile,
"\n alpha = %20.12e + %20.2e*i;", alpha[0], alpha[1]);
fprintf(msgFile, "\n Z = zeros(%d,1) ;", nrowY) ;
}
if ( transposeflag == 0 ) {
MARKTIME(t1) ;
if ( symflag == 0 ) {
InpMtx_MT_sym_mmm(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
} else if ( symflag == 1 ) {
InpMtx_MT_herm_mmm(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
} else if ( symflag == 2 ) {
InpMtx_MT_nonsym_mmm(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
}
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y2, "Z2", msgFile) ;
fprintf(msgFile, "\n maxerr2 = max(Z2 - Y - alpha*A*X) ") ;
fprintf(msgFile, "\n") ;
}
} else if ( transposeflag == 1 ) {
MARKTIME(t1) ;
InpMtx_MT_nonsym_mmm_H(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y2, "Z2", msgFile) ;
fprintf(msgFile,
"\n maxerr2 = max(Z2 - Y - alpha*ctranspose(A)*X) ") ;
fprintf(msgFile, "\n") ;
}
} else if ( transposeflag == 2 ) {
MARKTIME(t1) ;
InpMtx_MT_nonsym_mmm_T(A, Y2, alpha, X, nthread, msglvl, msgFile) ;
MARKTIME(t2) ;
if ( msglvl > 1 ) {
DenseMtx_writeForMatlab(Y2, "Z2", msgFile) ;
fprintf(msgFile,
"\n maxerr2 = max(Z2 - Y - alpha*transpose(A)*X) ") ;
fprintf(msgFile, "\n") ;
}
}
fprintf(msgFile, "\n %% %d ops, %.3f time, %.3f MT mflops",
nops, t2 - t1, 1.e-6*nops/(t2 - t1)) ;
/*
------------------------
free the working storage
------------------------
*/
InpMtx_free(A) ;
DenseMtx_free(X) ;
DenseMtx_free(Y) ;
DenseMtx_free(Y2) ;
IVfree(rowids) ;
IVfree(colids) ;
Drand_free(drand) ;
fclose(msgFile) ;
return(1) ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] ) {
/*
--------------------------------------------------
QR all-in-one program
(1) read in matrix entries and form InpMtx object
of A and A^TA
(2) form Graph object of A^TA
(3) order matrix and form front tree
(4) get the permutation, permute the matrix and
front tree and get the symbolic factorization
(5) compute the numeric factorization
(6) read in right hand side entries
(7) compute the solution
created -- 98jun11, cca
--------------------------------------------------
*/
/*--------------------------------------------------------------------*/
char *matrixFileName, *rhsFileName ;
ChvManager *chvmanager ;
DenseMtx *mtxB, *mtxX ;
double facops, imag, real, value ;
double cpus[10] ;
ETree *frontETree ;
FILE *inputFile, *msgFile ;
FrontMtx *frontmtx ;
Graph *graph ;
int ient, irow, jcol, jrhs, jrow, msglvl, neqns,
nedges, nent, nrhs, nrow, seed, type ;
InpMtx *mtxA ;
IV *newToOldIV, *oldToNewIV ;
IVL *adjIVL, *symbfacIVL ;
SubMtxManager *mtxmanager ;
/*--------------------------------------------------------------------*/
/*
--------------------
get input parameters
--------------------
*/
if ( argc != 7 ) {
fprintf(stdout,
"\n usage: %s msglvl msgFile type matrixFileName rhsFileName seed"
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n type -- type of entries"
"\n 1 (SPOOLES_REAL) -- real entries"
"\n 2 (SPOOLES_COMPLEX) -- complex entries"
"\n matrixFileName -- matrix file name, format"
"\n nrow ncol nent"
"\n irow jcol entry"
"\n ..."
"\n note: indices are zero based"
"\n rhsFileName -- right hand side file name, format"
"\n nrow "
"\n entry[0]"
"\n ..."
"\n entry[nrow-1]"
"\n seed -- random number seed, used for ordering"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
type = atoi(argv[3]) ;
matrixFileName = argv[4] ;
rhsFileName = argv[5] ;
seed = atoi(argv[6]) ;
/*--------------------------------------------------------------------*/
/*
--------------------------------------------
STEP 1: read the entries from the input file
and create the InpMtx object of A
--------------------------------------------
*/
inputFile = fopen(matrixFileName, "r") ;
fscanf(inputFile, "%d %d %d", &nrow, &neqns, &nent) ;
mtxA = InpMtx_new() ;
InpMtx_init(mtxA, INPMTX_BY_ROWS, type, nent, 0) ;
if ( type == SPOOLES_REAL ) {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le", &irow, &jcol, &value) ;
InpMtx_inputRealEntry(mtxA, irow, jcol, value) ;
}
} else {
for ( ient = 0 ; ient < nent ; ient++ ) {
fscanf(inputFile, "%d %d %le %le", &irow, &jcol, &real, &imag) ;
InpMtx_inputComplexEntry(mtxA, irow, jcol, real, imag) ;
}
}
fclose(inputFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n input matrix") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
----------------------------------------
STEP 2: read the right hand side entries
----------------------------------------
*/
inputFile = fopen(rhsFileName, "r") ;
fscanf(inputFile, "%d %d", &nrow, &nrhs) ;
mtxB = DenseMtx_new() ;
DenseMtx_init(mtxB, type, 0, 0, nrow, nrhs, 1, nrow) ;
DenseMtx_zero(mtxB) ;
if ( type == SPOOLES_REAL ) {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le", &value) ;
DenseMtx_setRealEntry(mtxB, jrow, jrhs, value) ;
}
}
} else {
for ( irow = 0 ; irow < nrow ; irow++ ) {
fscanf(inputFile, "%d", &jrow) ;
for ( jrhs = 0 ; jrhs < nrhs ; jrhs++ ) {
fscanf(inputFile, "%le %le", &real, &imag) ;
DenseMtx_setComplexEntry(mtxB, jrow, jrhs, real, imag) ;
}
}
}
fclose(inputFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n rhs matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxB, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------------------------
STEP 3 : find a low-fill ordering
(1) create the Graph object for A^TA or A^HA
(2) order the graph using multiple minimum degree
-------------------------------------------------
*/
graph = Graph_new() ;
adjIVL = InpMtx_adjForATA(mtxA) ;
nedges = IVL_tsize(adjIVL) ;
Graph_init2(graph, 0, neqns, 0, nedges, neqns, nedges, adjIVL,
NULL, NULL) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n graph of A^T A") ;
Graph_writeForHumanEye(graph, msgFile) ;
fflush(msgFile) ;
}
frontETree = orderViaMMD(graph, seed, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n front tree from ordering") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------------------
STEP 4: get the permutation, permute the matrix and
front tree and get the symbolic factorization
-----------------------------------------------------
*/
oldToNewIV = ETree_oldToNewVtxPerm(frontETree) ;
newToOldIV = ETree_newToOldVtxPerm(frontETree) ;
InpMtx_permute(mtxA, NULL, IV_entries(oldToNewIV)) ;
InpMtx_changeStorageMode(mtxA, INPMTX_BY_VECTORS) ;
symbfacIVL = SymbFac_initFromGraph(frontETree, graph) ;
IVL_overwrite(symbfacIVL, oldToNewIV) ;
IVL_sortUp(symbfacIVL) ;
ETree_permuteVertices(frontETree, oldToNewIV) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n old-to-new permutation vector") ;
IV_writeForHumanEye(oldToNewIV, msgFile) ;
fprintf(msgFile, "\n\n new-to-old permutation vector") ;
IV_writeForHumanEye(newToOldIV, msgFile) ;
fprintf(msgFile, "\n\n front tree after permutation") ;
ETree_writeForHumanEye(frontETree, msgFile) ;
fprintf(msgFile, "\n\n input matrix after permutation") ;
InpMtx_writeForHumanEye(mtxA, msgFile) ;
fprintf(msgFile, "\n\n symbolic factorization") ;
IVL_writeForHumanEye(symbfacIVL, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------------------------
STEP 5: initialize the front matrix object
------------------------------------------
*/
frontmtx = FrontMtx_new() ;
mtxmanager = SubMtxManager_new() ;
SubMtxManager_init(mtxmanager, NO_LOCK, 0) ;
if ( type == SPOOLES_REAL ) {
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type,
SPOOLES_SYMMETRIC, FRONTMTX_DENSE_FRONTS,
SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL,
mtxmanager, msglvl, msgFile) ;
} else {
FrontMtx_init(frontmtx, frontETree, symbfacIVL, type,
SPOOLES_HERMITIAN, FRONTMTX_DENSE_FRONTS,
SPOOLES_NO_PIVOTING, NO_LOCK, 0, NULL,
mtxmanager, msglvl, msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-----------------------------------------
STEP 6: compute the numeric factorization
-----------------------------------------
*/
chvmanager = ChvManager_new() ;
ChvManager_init(chvmanager, NO_LOCK, 1) ;
DVzero(10, cpus) ;
facops = 0.0 ;
FrontMtx_QR_factor(frontmtx, mtxA, chvmanager,
cpus, &facops, msglvl, msgFile) ;
ChvManager_free(chvmanager) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix") ;
fprintf(msgFile, "\n facops = %9.2f", facops) ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
--------------------------------------
STEP 7: post-process the factorization
--------------------------------------
*/
FrontMtx_postProcess(frontmtx, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n factor matrix after post-processing") ;
FrontMtx_writeForHumanEye(frontmtx, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------
STEP 8: solve the linear system
-------------------------------
*/
mtxX = DenseMtx_new() ;
DenseMtx_init(mtxX, type, 0, 0, neqns, nrhs, 1, neqns) ;
FrontMtx_QR_solve(frontmtx, mtxA, mtxX, mtxB, mtxmanager,
cpus, msglvl, msgFile) ;
if ( msglvl > 1 ) {
fprintf(msgFile, "\n\n solution matrix in new ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
-------------------------------------------------------
STEP 9: permute the solution into the original ordering
-------------------------------------------------------
*/
DenseMtx_permuteRows(mtxX, newToOldIV) ;
if ( msglvl > 0 ) {
fprintf(msgFile, "\n\n solution matrix in original ordering") ;
DenseMtx_writeForHumanEye(mtxX, msgFile) ;
fflush(msgFile) ;
}
/*--------------------------------------------------------------------*/
/*
------------------------
free the working storage
------------------------
*/
InpMtx_free(mtxA) ;
FrontMtx_free(frontmtx) ;
Graph_free(graph) ;
DenseMtx_free(mtxX) ;
DenseMtx_free(mtxB) ;
ETree_free(frontETree) ;
IV_free(newToOldIV) ;
IV_free(oldToNewIV) ;
IVL_free(symbfacIVL) ;
SubMtxManager_free(mtxmanager) ;
/*--------------------------------------------------------------------*/
return(1) ; }
int
tfqmrl (
int n_matrixSize,
int type,
int symmetryflag,
InpMtx *mtxA,
FrontMtx *Precond,
DenseMtx *mtxX,
DenseMtx *mtxB,
int itermax,
double convergetol,
int msglvl,
FILE *msgFile
)
{
Chv *chv, *rootchv ;
ChvManager *chvmanager ;
DenseMtx *vecD, *vecR, *vecT, *vecU1, *vecU2, *vecV, *vecW;
DenseMtx *vecX, *vecY1, *vecY2 ;
double Alpha, Beta, Cee, Eta, Rho, Rho_new ;
double Sigma, Tau, Theta;
double Init_norm, ratio, Res_norm;
double error_trol, m, Rtmp;
double t1, t2, cpus[9] ;
double one[2] = {1.0, 0.0}, zero[2] ={0.0, 0.0} ;
double Tiny = 0.1e-28;
int Iter, Imv, neqns;
int stats[6] ;
neqns = n_matrixSize;
/*
--------------------
init the vectors in TFQMRL
--------------------
*/
vecD = DenseMtx_new() ;
DenseMtx_init(vecD, type, 0, 0, neqns, 1, 1, neqns) ;
vecR = DenseMtx_new() ;
DenseMtx_init(vecR, type, 0, 0, neqns, 1, 1, neqns) ;
vecT = DenseMtx_new() ;
DenseMtx_init(vecT, type, 0, 0, neqns, 1, 1, neqns) ;
vecU1 = DenseMtx_new() ;
DenseMtx_init(vecU1, type, 0, 0, neqns, 1, 1, neqns) ;
vecU2 = DenseMtx_new() ;
DenseMtx_init(vecU2, type, 0, 0, neqns, 1, 1, neqns) ;
vecV = DenseMtx_new() ;
DenseMtx_init(vecV, type, 0, 0, neqns, 1, 1, neqns) ;
vecW = DenseMtx_new() ;
DenseMtx_init(vecW, type, 0, 0, neqns, 1, 1, neqns) ;
vecX = DenseMtx_new() ;
DenseMtx_init(vecX, type, 0, 0, neqns, 1, 1, neqns) ;
vecY1 = DenseMtx_new() ;
DenseMtx_init(vecY1, type, 0, 0, neqns, 1, 1, neqns) ;
vecY2 = DenseMtx_new() ;
DenseMtx_init(vecY2, type, 0, 0, neqns, 1, 1, neqns) ;
/*
--------------------------
Initialize the iterations
--------------------------
*/
/* ---- Set initial guess as zero ---- */
DenseMtx_zero(vecX) ;
DenseMtx_colCopy(vecT, 0, mtxB, 0);
/* */
FrontMtx_solve(Precond, vecR, vecT, Precond->manager,
cpus, msglvl, msgFile) ;
/* */
Init_norm = DenseMtx_twoNormOfColumn(vecR,0);
if ( Init_norm == 0.0 ){
Init_norm = 1.0;
};
error_trol = Init_norm * convergetol ;
fprintf(msgFile, "\n TFQMRL Initial norml: %6.2e ", Init_norm ) ;
fprintf(msgFile, "\n TFQMRL Conveg. Control: %7.3e ", convergetol ) ;
fprintf(msgFile, "\n TFQMRL Convergen Control: %7.3e ",error_trol ) ;
DenseMtx_zero(vecD) ;
DenseMtx_zero(vecU1) ;
DenseMtx_zero(vecU2) ;
DenseMtx_zero(vecY2) ;
/* DenseMtx_copy(vecR, mtxB); */
DenseMtx_colCopy(vecW, 0, vecR, 0);
DenseMtx_colCopy(vecY1, 0, vecR, 0);
Iter = 0;
Imv = 0;
switch ( symmetryflag ) {
case SPOOLES_SYMMETRIC :
InpMtx_sym_gmmm(mtxA, zero, vecT, one, vecY1) ;
break ;
case SPOOLES_HERMITIAN :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
case SPOOLES_NONSYMMETRIC :
InpMtx_nonsym_gmmm(mtxA, zero, vecT, one, vecY1) ;
break ;
default :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
}
/* */
FrontMtx_solve(Precond, vecV, vecT, Precond->manager,
cpus, msglvl, msgFile) ;
/* */
Imv++;
DenseMtx_colCopy(vecU1, 0, vecV, 0);
/*
*/
Theta = 0.0;
Eta = 0.0;
Tau = Init_norm ;
Rho = Tau * Tau ;
/*
------------------------------
TFQMRL Iteration start
------------------------------
*/
MARKTIME(t1) ;
while ( Iter <= itermax )
{
Iter++;
DenseMtx_colDotProduct(vecV, 0, vecR, 0, &Sigma);
if (Sigma == 0){
fprintf(msgFile, "\n\n Fatal Error, \n"
" TFQMRL Breakdown, Sigma = 0 !!") ;
Imv = -1;
goto end;
};
Alpha = Rho/Sigma;
/*
----------------
Odd step
---------------
*/
m = 2 * Iter - 1;
Rtmp=-Alpha;
DenseMtx_colGenAxpy(one, vecW, 0, &Rtmp, vecU1, 0);
Rtmp = Theta * Theta * Eta / Alpha ;
DenseMtx_colGenAxpy(&Rtmp, vecD, 0, one, vecY1, 0);
Theta = DenseMtx_twoNormOfColumn(vecW,0)/Tau;
Cee = 1.0/sqrt(1.0 + Theta*Theta);
Tau = Tau * Theta * Cee ;
Eta = Cee * Cee * Alpha ;
DenseMtx_colGenAxpy(one, vecX, 0, &Eta, vecD, 0);
fprintf(msgFile, "\n\n Odd step at %d", Imv);
fprintf(msgFile, " \n Tau is : %7.3e", Tau) ;
/*
Debug purpose: Check the convergence history
for the true residual norm
*/
/*
DenseMtx_zero(vecT) ;
InpMtx_nonsym_mmm(mtxA, vecT, one, vecX) ;
DenseMtx_sub(vecT, mtxB) ;
Rtmp = DenseMtx_twoNormOfColumn(vecT,0);
fprintf(msgFile, "\n TFQMRL Residual norm: %6.2e ", Rtmp) ;
*/
/*
----------------
Convergence Test
---------------
*/
if (Tau * sqrt(m + 1) <= error_trol ) {
/* */
DenseMtx_colCopy(mtxX, 0, vecX, 0);
/*
DenseMtx_zero(vecT) ;
InpMtx_nonsym_mmm(mtxA, vecT, one, mtxX) ;
*/
switch ( symmetryflag ) {
case SPOOLES_SYMMETRIC :
InpMtx_sym_gmmm(mtxA, zero, vecT, one, mtxX) ;
break ;
case SPOOLES_HERMITIAN :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
case SPOOLES_NONSYMMETRIC :
InpMtx_nonsym_gmmm(mtxA, zero, vecT, one, mtxX) ;
break ;
default :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
}
DenseMtx_sub(vecT, mtxB) ;
Rtmp = DenseMtx_twoNormOfColumn(vecT,0);
fprintf(msgFile, "\n TFQMRL Residual norm: %6.2e ", Rtmp) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU : Converges in time: %8.3f ", t2 - t1) ;
fprintf(msgFile, "\n # iterations = %d", Imv) ;
fprintf(msgFile, "\n\n after TFQMRL") ;
goto end;
};
/*
----------------
Even step
---------------
*/
DenseMtx_colCopy(vecY2, 0, vecY1, 0);
Rtmp=-Alpha;
DenseMtx_colGenAxpy(one, vecY2, 0, &Rtmp, vecV, 0);
/*
DenseMtx_zero(vecT) ;
InpMtx_nonsym_mmm(mtxA, vecT, one, vecY2) ;
*/
switch ( symmetryflag ) {
case SPOOLES_SYMMETRIC :
InpMtx_sym_gmmm(mtxA, zero, vecT, one, vecY2) ;
break ;
case SPOOLES_HERMITIAN :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
case SPOOLES_NONSYMMETRIC :
InpMtx_nonsym_gmmm(mtxA, zero, vecT, one, vecY2) ;
break ;
default :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
}
FrontMtx_solve(Precond, vecU2, vecT, Precond->manager,
cpus, msglvl, msgFile) ;
Imv++;
m = 2 * Iter ;
Rtmp = -Alpha;
DenseMtx_colGenAxpy(one, vecW, 0, &Rtmp, vecU2, 0);
Rtmp = Theta * Theta * Eta / Alpha ;
DenseMtx_colGenAxpy(&Rtmp, vecD, 0, one, vecY2, 0);
Theta = DenseMtx_twoNormOfColumn(vecW,0)/Tau;
Cee = 1.0/sqrt(1.0 + Theta*Theta);
Tau = Tau * Theta * Cee ;
Eta = Cee * Cee * Alpha ;
DenseMtx_colGenAxpy(one, vecX, 0, &Eta, vecD, 0);
fprintf(msgFile, "\n\n Even step at %d", Imv) ;
/*
----------------
Convergence Test for even step
---------------
*/
if (Tau * sqrt(m + 1) <= error_trol ) {
DenseMtx_colCopy(mtxX, 0, vecX, 0);
/*
DenseMtx_zero(vecT) ;
InpMtx_nonsym_mmm(mtxA, vecT, one, mtxX) ;
*/
switch ( symmetryflag ) {
case SPOOLES_SYMMETRIC :
InpMtx_sym_gmmm(mtxA, zero, vecT, one, mtxX) ;
break ;
case SPOOLES_HERMITIAN :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
case SPOOLES_NONSYMMETRIC :
InpMtx_nonsym_gmmm(mtxA, zero, vecT, one, mtxX) ;
break ;
default :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
}
DenseMtx_sub(vecT, mtxB) ;
Rtmp = DenseMtx_twoNormOfColumn(vecT,0);
fprintf(msgFile, "\n TFQMRL Residual norm: %6.2e ", Rtmp) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU : Converges in time: %8.3f ", t2 - t1) ;
fprintf(msgFile, "\n # iterations = %d", Imv) ;
fprintf(msgFile, "\n\n after TFQMRL") ;
goto end;
};
if (Rho == 0){
fprintf(msgFile, "\n\n Fatal Error, \n"
" TFQMRL Breakdown, Rho = 0 !!") ;
Imv = -1;
goto end;
};
DenseMtx_colDotProduct(vecW, 0, vecR, 0, &Rho_new);
Beta = Rho_new / Rho;
Rho = Rho_new ;
DenseMtx_colCopy(vecY1, 0, vecY2, 0);
DenseMtx_colGenAxpy(&Beta, vecY1, 0, one, vecW, 0);
/*
DenseMtx_zero(vecT) ;
InpMtx_nonsym_mmm(mtxA, vecT, one, vecY1) ;
*/
switch ( symmetryflag ) {
case SPOOLES_SYMMETRIC :
InpMtx_sym_gmmm(mtxA, zero, vecT, one, vecY1) ;
break ;
case SPOOLES_HERMITIAN :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
case SPOOLES_NONSYMMETRIC :
InpMtx_nonsym_gmmm(mtxA, zero, vecT, one, vecY1) ;
break ;
default :
fprintf(msgFile, "\n TFQMRL Matrix type wrong");
fprintf(msgFile, "\n Fatal error");
goto end;
}
FrontMtx_solve(Precond, vecU1, vecT, Precond->manager,
cpus, msglvl, msgFile) ;
Imv++;
/* */
DenseMtx_colCopy(vecT, 0, vecU2, 0);
DenseMtx_colGenAxpy(one, vecT, 0, &Beta, vecV, 0);
DenseMtx_colCopy(vecV, 0, vecT, 0);
DenseMtx_colGenAxpy(&Beta, vecV, 0, one, vecU1, 0);
Rtmp = Tau*sqrt(m + 1)/Init_norm ;
fprintf(msgFile, "\n\n At iteration %d"
" the convergence ratio is %12.4e",
Imv, Rtmp) ;
}
/* End of while loop */
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU : Total iteration time is : %8.3f ", t2 - t1) ;
fprintf(msgFile, "\n # iterations = %d", Imv) ;
fprintf(msgFile, "\n\n TFQMRL did not Converge !") ;
fprintf(msgFile, "\n\n after TFQMRL") ;
DenseMtx_colCopy(mtxX, 0, vecX, 0);
/*
------------------------
free the working storage
------------------------
*/
end:
DenseMtx_free(vecD) ;
DenseMtx_free(vecR) ;
DenseMtx_free(vecT) ;
DenseMtx_free(vecU1) ;
DenseMtx_free(vecU2) ;
DenseMtx_free(vecV) ;
DenseMtx_free(vecW) ;
DenseMtx_free(vecX) ;
DenseMtx_free(vecY1) ;
DenseMtx_free(vecY2) ;
fprintf(msgFile, "\n") ;
return(Imv) ; }
int
zpcgr (
int n_matrixSize,
int type,
int symmetryflag,
InpMtx *mtxA,
FrontMtx *Precond,
DenseMtx *mtxX,
DenseMtx *mtxB,
int itermax,
double convergetol,
int msglvl,
FILE *msgFile
)
{
Chv *chv, *rootchv ;
ChvManager *chvmanager ;
DenseMtx *mtxZ ;
DenseMtx *vecP, *vecR, *vecQ ;
DenseMtx *vecX, *vecZ ;
double Alpha[2], Beta[2], Rho[2], Rho0[2], Rtmp[2];
double Init_norm, ratio, Res_norm ;
double t1, t2, cpus[9] ;
double one[2] = {1.0, 0.0}, zero[2] = {0.0, 0.0} ;
double Tiny[2] = {0.1e-28, 0.0};
int Iter, neqns;
int stats[6] ;
if (symmetryflag != SPOOLES_HERMITIAN){
fprintf(msgFile, "\n\n Fatal Error, \n"
" Matrix is not Hermitian in ZPCGR !!") ;
spoolesFatal();
};
neqns = n_matrixSize;
/*
--------------------
init the vectors in ZPCGR
--------------------
*/
vecP = DenseMtx_new() ;
DenseMtx_init(vecP, type, 0, 0, neqns, 1, 1, neqns) ;
vecR = DenseMtx_new() ;
DenseMtx_init(vecR, type, 0, 0, neqns, 1, 1, neqns) ;
vecX = DenseMtx_new() ;
DenseMtx_init(vecX, type, 0, 0, neqns, 1, 1, neqns) ;
vecQ = DenseMtx_new() ;
DenseMtx_init(vecQ, type, 0, 0, neqns, 1, 1, neqns) ;
vecZ = DenseMtx_new() ;
DenseMtx_init(vecZ, type, 0, 0, neqns, 1, 1, neqns) ;
/*
--------------------------
Initialize the iterations
--------------------------
*/
Init_norm = DenseMtx_twoNormOfColumn (mtxB, 0);
if ( Init_norm == 0.0 ){
Init_norm = 1.0; };
ratio = 1.0;
DenseMtx_zero(vecX) ;
DenseMtx_colCopy (vecR, 0, mtxB, 0);
MARKTIME(t1) ;
Iter = 0;
/*
------------------------------
Main Loop of the iterations
------------------------------
*/
while ( ratio > convergetol && Iter <= itermax )
{
Iter++;
/* */
FrontMtx_solve(Precond, vecZ, vecR, Precond->manager,
cpus, msglvl, msgFile) ;
DenseMtx_colDotProduct(vecR, 0, vecZ, 0, Rho);
if ( Rho[0] == 0.0 & Rho[1] == 0.0){
fprintf(stderr, "\n breakdown in ZPCGR !! "
"\n Fatal error \n");
spoolesFatal();
}
/* */
if ( Iter == 1 ) {
DenseMtx_colCopy (vecP, 0, vecZ, 0);
} else {
zdiv(Rho, Rho0, Beta);
DenseMtx_colGenAxpy (Beta, vecP, 0, one, vecZ, 0);
};
InpMtx_herm_gmmm(mtxA, zero, vecQ, one, vecP) ;
DenseMtx_colDotProduct (vecP, 0, vecQ,0, Rtmp);
zdiv(Rho, Rtmp, Alpha);
DenseMtx_colGenAxpy (one, vecX, 0, Alpha, vecP, 0);
Rtmp[0] = -Alpha[0];
Rtmp[1] = -Alpha[1];
DenseMtx_colGenAxpy (one, vecR, 0, Rtmp, vecQ, 0);
Rho0[0] = Rho[0];
Rho0[1] = Rho[1];
/* */
Res_norm = DenseMtx_twoNormOfColumn (vecR, 0);
ratio = Res_norm/Init_norm;
fprintf(msgFile, "\n\n At iteration %d"
" the convergence ratio is %12.4e",
Iter, ratio) ;
}
/* End of while loop */
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU : Converges in time: %8.3f ", t2 - t1) ;
fprintf(msgFile, "\n # iterations = %d", Iter) ;
fprintf(msgFile, "\n\n after ZPCGR") ;
DenseMtx_colCopy (mtxX, 0, vecX, 0);
/*
------------------------
free the working storage
------------------------
*/
DenseMtx_free(vecP) ;
DenseMtx_free(vecR) ;
DenseMtx_free(vecX) ;
DenseMtx_free(vecQ) ;
DenseMtx_free(vecZ) ;
fprintf(msgFile, "\n") ;
return(1) ; }