void tlin::factorize(SuperMatrix *A, SuperFactors *&F, superlu_options_t *opt) {
assert(A->nrow == A->ncol);
int n = A->nrow;
if (!F) F = (SuperFactors *)SUPERLU_MALLOC(sizeof(SuperFactors));
if (!opt) opt = &defaultOpt;
F->perm_c = intMalloc(n);
get_perm_c(3, A, F->perm_c);
SuperMatrix AC;
int *etree = intMalloc(n);
sp_preorder(opt, A, F->perm_c, etree, &AC);
F->L = (SuperMatrix *)SUPERLU_MALLOC(sizeof(SuperMatrix));
F->U = (SuperMatrix *)SUPERLU_MALLOC(sizeof(SuperMatrix));
F->perm_r = intMalloc(n);
SuperLUStat_t stat;
StatInit(&stat);
int result;
dgstrf(opt, &AC, sp_ienv(1), sp_ienv(2), etree, NULL, 0, F->perm_c, F->perm_r,
F->L, F->U, &stat, &result);
StatFree(&stat);
Destroy_CompCol_Permuted(&AC);
SUPERLU_FREE(etree);
if (result != 0) freeF(F), F = 0;
}
int HYPRE_ParCSR_SuperLUSetup(HYPRE_Solver solver, HYPRE_ParCSRMatrix A_csr,
HYPRE_ParVector b, HYPRE_ParVector x )
{
#ifdef HAVE_SUPERLU
int startRow, endRow, nrows, *partition, *AdiagI, *AdiagJ, nnz;
int irow, colNum, index, *cscI, *cscJ, jcol, *colLengs;
int *etree, permcSpec, lwork, panelSize, relax, info;
double *AdiagA, *cscA, diagPivotThresh, dropTol;
char refact[1];
hypre_CSRMatrix *Adiag;
HYPRE_SuperLU *sluPtr;
SuperMatrix sluAmat, auxAmat;
superlu_options_t slu_options;
SuperLUStat_t slu_stat;
/* ---------------------------------------------------------------- */
/* get matrix information */
/* ---------------------------------------------------------------- */
sluPtr = (HYPRE_SuperLU *) solver;
assert ( sluPtr != NULL );
HYPRE_ParCSRMatrixGetRowPartitioning( A_csr, &partition );
startRow = partition[0];
endRow = partition[1] - 1;
nrows = endRow - startRow + 1;
free( partition );
if ( startRow != 0 )
{
printf("HYPRE_ParCSR_SuperLUSetup ERROR - start row != 0.\n");
return -1;
}
/* ---------------------------------------------------------------- */
/* get hypre matrix */
/* ---------------------------------------------------------------- */
Adiag = hypre_ParCSRMatrixDiag((hypre_ParCSRMatrix *) A_csr);
AdiagI = hypre_CSRMatrixI(Adiag);
AdiagJ = hypre_CSRMatrixJ(Adiag);
AdiagA = hypre_CSRMatrixData(Adiag);
nnz = AdiagI[nrows];
/* ---------------------------------------------------------------- */
/* convert the csr matrix into csc matrix */
/* ---------------------------------------------------------------- */
colLengs = (int *) malloc(nrows * sizeof(int));
for ( irow = 0; irow < nrows; irow++ ) colLengs[irow] = 0;
for ( irow = 0; irow < nrows; irow++ )
for ( jcol = AdiagI[irow]; jcol < AdiagI[irow+1]; jcol++ )
colLengs[AdiagJ[jcol]]++;
cscJ = (int *) malloc( (nrows+1) * sizeof(int) );
cscI = (int *) malloc( nnz * sizeof(int) );
cscA = (double *) malloc( nnz * sizeof(double) );
cscJ[0] = 0;
nnz = 0;
for ( jcol = 1; jcol <= nrows; jcol++ )
{
nnz += colLengs[jcol-1];
cscJ[jcol] = nnz;
}
for ( irow = 0; irow < nrows; irow++ )
{
for ( jcol = AdiagI[irow]; jcol < AdiagI[irow+1]; jcol++ )
{
colNum = AdiagJ[jcol];
index = cscJ[colNum]++;
cscI[index] = irow;
cscA[index] = AdiagA[jcol];
}
}
cscJ[0] = 0;
nnz = 0;
for ( jcol = 1; jcol <= nrows; jcol++ )
{
nnz += colLengs[jcol-1];
cscJ[jcol] = nnz;
}
free(colLengs);
/* ---------------------------------------------------------------- */
/* create SuperMatrix */
/* ---------------------------------------------------------------- */
dCreate_CompCol_Matrix(&sluAmat,nrows,nrows,cscJ[nrows],cscA,cscI,
cscJ, SLU_NC, SLU_D, SLU_GE);
etree = (int *) malloc(nrows * sizeof(int));
sluPtr->permC_ = (int *) malloc(nrows * sizeof(int));
sluPtr->permR_ = (int *) malloc(nrows * sizeof(int));
permcSpec = 0;
get_perm_c(permcSpec, &sluAmat, sluPtr->permC_);
slu_options.Fact = DOFACT;
slu_options.SymmetricMode = NO;
sp_preorder(&slu_options, &sluAmat, sluPtr->permC_, etree, &auxAmat);
diagPivotThresh = 1.0;
dropTol = 0.0;
panelSize = sp_ienv(1);
relax = sp_ienv(2);
StatInit(&slu_stat);
lwork = 0;
slu_options.ColPerm = MY_PERMC;
slu_options.DiagPivotThresh = diagPivotThresh;
dgstrf(&slu_options, &auxAmat, dropTol, relax, panelSize,
etree, NULL, lwork, sluPtr->permC_, sluPtr->permR_,
&(sluPtr->SLU_Lmat), &(sluPtr->SLU_Umat), &slu_stat, &info);
Destroy_CompCol_Permuted(&auxAmat);
Destroy_CompCol_Matrix(&sluAmat);
free(etree);
sluPtr->factorized_ = 1;
StatFree(&slu_stat);
return 0;
#else
printf("HYPRE_ParCSR_SuperLUSetup ERROR - SuperLU not enabled.\n");
*solver = (HYPRE_Solver) NULL;
return -1;
#endif
}
static PyObject* superluWrappersSparseFactorPrepare(PyObject* self,
PyObject* args)
{
int i,n,relax=1,panel_size=10,lwork=0,info=0,permc_spec=3;
double drop_tol=-1.0;/* not used by superlu */
void *work=NULL;
PyObject *mat,*sparseFactor;
if(!PyArg_ParseTuple(args,"OO",
&mat,
&sparseFactor))
return NULL;
SFP(sparseFactor)->storeA.nnz = ((SparseMatrix*)mat)->A.nnz;
SFP(sparseFactor)->storeA.nzval = ((SparseMatrix*)mat)->A.nzval;
SFP(sparseFactor)->storeA.colptr = ((SparseMatrix*)mat)->A.rowptr;
SFP(sparseFactor)->storeA.rowind = ((SparseMatrix*)mat)->A.colind;
/* calc column permutation */
if ( SFP(sparseFactor)->use_same_perm_c == 0)
{
get_perm_c(permc_spec,
&SFP(sparseFactor)->A,
SFP(sparseFactor)->perm_c);
SFP(sparseFactor)->use_same_perm_c = 1;
}
if ( SFP(sparseFactor)->use_same_sparsity == 0)
{
if (SFP(sparseFactor)->AC.Store != NULL)
{
Destroy_CompCol_Permuted(&SFP(sparseFactor)->AC);
Destroy_SuperNode_Matrix(&SFP(sparseFactor)->L);
Destroy_CompCol_Matrix(&SFP(sparseFactor)->U);
}
/* apply column permutation and build AC and etree*/
sp_preorder(&SFP(sparseFactor)->options,
&SFP(sparseFactor)->A,
SFP(sparseFactor)->perm_c,
SFP(sparseFactor)->etree,
&SFP(sparseFactor)->AC);
SFP(sparseFactor)->use_same_sparsity = 1;
}
else
{
/* apply column permutation */
SFP(sparseFactor)->options.Fact = SamePattern_SameRowPerm;
n = SFP(sparseFactor)->A.ncol;
for (i = 0; i < n; i++)
{
((NCPformat*)SFP(sparseFactor)->AC.Store)->colbeg[SFP(sparseFactor)->perm_c[i]] = ((NCformat*)SFP(sparseFactor)->A.Store)->colptr[i];
((NCPformat*)SFP(sparseFactor)->AC.Store)->colend[SFP(sparseFactor)->perm_c[i]] = ((NCformat*)SFP(sparseFactor)->A.Store)->colptr[i+1];
}
}
dgstrf(&SFP(sparseFactor)->options,
&SFP(sparseFactor)->AC,
relax,
panel_size,
SFP(sparseFactor)->etree,
work,
lwork,
SFP(sparseFactor)->perm_c,
SFP(sparseFactor)->perm_r,
&SFP(sparseFactor)->L,
&SFP(sparseFactor)->U,
&SFP(sparseFactor)->Glu,
&SFP(sparseFactor)->stat,
&info);
Py_INCREF(Py_None);
return Py_None;
}
void
dgssv(superlu_options_t *options, SuperMatrix *A, int *perm_c, int *perm_r,
SuperMatrix *L, SuperMatrix *U, SuperMatrix *B,
SuperLUStat_t *stat, int *info )
{
DNformat *Bstore;
SuperMatrix *AA;/* A in SLU_NC format used by the factorization routine.*/
SuperMatrix AC; /* Matrix postmultiplied by Pc */
int lwork = 0, *etree, i;
/* Set default values for some parameters */
double drop_tol = 0.;
int panel_size; /* panel size */
int relax; /* no of columns in a relaxed snodes */
int permc_spec;
trans_t trans = NOTRANS;
double *utime;
double t; /* Temporary time */
/* Test the input parameters ... */
*info = 0;
Bstore = B->Store;
if ( options->Fact != DOFACT ) *info = -1;
else if ( A->nrow != A->ncol || A->nrow < 0 ||
(A->Stype != SLU_NC && A->Stype != SLU_NR) ||
A->Dtype != SLU_D || A->Mtype != SLU_GE )
*info = -2;
else if ( B->ncol < 0 || Bstore->lda < SUPERLU_MAX(0, A->nrow) ||
B->Stype != SLU_DN || B->Dtype != SLU_D || B->Mtype != SLU_GE )
*info = -7;
if ( *info != 0 ) {
i = -(*info);
xerbla_("dgssv", &i);
return;
}
utime = stat->utime;
/* Convert A to SLU_NC format when necessary. */
if ( A->Stype == SLU_NR ) {
NRformat *Astore = A->Store;
AA = (SuperMatrix *) SUPERLU_MALLOC( sizeof(SuperMatrix) );
dCreate_CompCol_Matrix(AA, A->ncol, A->nrow, Astore->nnz,
Astore->nzval, Astore->colind, Astore->rowptr,
SLU_NC, A->Dtype, A->Mtype);
trans = TRANS;
} else {
if ( A->Stype == SLU_NC ) AA = A;
}
t = SuperLU_timer_();
/*
* Get column permutation vector perm_c[], according to permc_spec:
* permc_spec = NATURAL: natural ordering
* permc_spec = MMD_AT_PLUS_A: minimum degree on structure of A'+A
* permc_spec = MMD_ATA: minimum degree on structure of A'*A
* permc_spec = COLAMD: approximate minimum degree column ordering
* permc_spec = MY_PERMC: the ordering already supplied in perm_c[]
*/
permc_spec = options->ColPerm;
if ( permc_spec != MY_PERMC && options->Fact == DOFACT )
get_perm_c(permc_spec, AA, perm_c);
utime[COLPERM] = SuperLU_timer_() - t;
etree = intMalloc(A->ncol);
t = SuperLU_timer_();
sp_preorder(options, AA, perm_c, etree, &AC);
utime[ETREE] = SuperLU_timer_() - t;
panel_size = sp_ienv(1);
relax = sp_ienv(2);
/*printf("Factor PA = LU ... relax %d\tw %d\tmaxsuper %d\trowblk %d\n",
relax, panel_size, sp_ienv(3), sp_ienv(4));*/
t = SuperLU_timer_();
/* Compute the LU factorization of A. */
dgstrf(options, &AC, drop_tol, relax, panel_size,
etree, NULL, lwork, perm_c, perm_r, L, U, stat, info);
utime[FACT] = SuperLU_timer_() - t;
t = SuperLU_timer_();
if ( *info == 0 ) {
/* Solve the system A*X=B, overwriting B with X. */
dgstrs (trans, L, U, perm_c, perm_r, B, stat, info);
}
utime[SOLVE] = SuperLU_timer_() - t;
SUPERLU_FREE (etree);
Destroy_CompCol_Permuted(&AC);
if ( A->Stype == SLU_NR ) {
Destroy_SuperMatrix_Store(AA);
SUPERLU_FREE(AA);
}
}
void
sgssv(superlu_options_t *options, SuperMatrix *A, int *perm_c, int *perm_r,
SuperMatrix *L, SuperMatrix *U, SuperMatrix *B,
SuperLUStat_t *stat, int *info )
{
/*
* Purpose
* =======
*
* SGSSV solves the system of linear equations A*X=B, using the
* LU factorization from SGSTRF. It performs the following steps:
*
* 1. If A is stored column-wise (A->Stype = SLU_NC):
*
* 1.1. Permute the columns of A, forming A*Pc, where Pc
* is a permutation matrix. For more details of this step,
* see sp_preorder.c.
*
* 1.2. Factor A as Pr*A*Pc=L*U with the permutation Pr determined
* by Gaussian elimination with partial pivoting.
* L is unit lower triangular with offdiagonal entries
* bounded by 1 in magnitude, and U is upper triangular.
*
* 1.3. Solve the system of equations A*X=B using the factored
* form of A.
*
* 2. If A is stored row-wise (A->Stype = SLU_NR), apply the
* above algorithm to the transpose of A:
*
* 2.1. Permute columns of transpose(A) (rows of A),
* forming transpose(A)*Pc, where Pc is a permutation matrix.
* For more details of this step, see sp_preorder.c.
*
* 2.2. Factor A as Pr*transpose(A)*Pc=L*U with the permutation Pr
* determined by Gaussian elimination with partial pivoting.
* L is unit lower triangular with offdiagonal entries
* bounded by 1 in magnitude, and U is upper triangular.
*
* 2.3. Solve the system of equations A*X=B using the factored
* form of A.
*
* See supermatrix.h for the definition of 'SuperMatrix' structure.
*
* Arguments
* =========
*
* options (input) superlu_options_t*
* The structure defines the input parameters to control
* how the LU decomposition will be performed and how the
* system will be solved.
*
* A (input) SuperMatrix*
* Matrix A in A*X=B, of dimension (A->nrow, A->ncol). The number
* of linear equations is A->nrow. Currently, the type of A can be:
* Stype = SLU_NC or SLU_NR; Dtype = SLU_S; Mtype = SLU_GE.
* In the future, more general A may be handled.
*
* perm_c (input/output) int*
* If A->Stype = SLU_NC, column permutation vector of size A->ncol
* which defines the permutation matrix Pc; perm_c[i] = j means
* column i of A is in position j in A*Pc.
* If A->Stype = SLU_NR, column permutation vector of size A->nrow
* which describes permutation of columns of transpose(A)
* (rows of A) as described above.
*
* If options->ColPerm = MY_PERMC or options->Fact = SamePattern or
* options->Fact = SamePattern_SameRowPerm, it is an input argument.
* On exit, perm_c may be overwritten by the product of the input
* perm_c and a permutation that postorders the elimination tree
* of Pc'*A'*A*Pc; perm_c is not changed if the elimination tree
* is already in postorder.
* Otherwise, it is an output argument.
*
* perm_r (input/output) int*
* If A->Stype = SLU_NC, row permutation vector of size A->nrow,
* which defines the permutation matrix Pr, and is determined
* by partial pivoting. perm_r[i] = j means row i of A is in
* position j in Pr*A.
* If A->Stype = SLU_NR, permutation vector of size A->ncol, which
* determines permutation of rows of transpose(A)
* (columns of A) as described above.
*
* If options->RowPerm = MY_PERMR or
* options->Fact = SamePattern_SameRowPerm, perm_r is an
* input argument.
* otherwise it is an output argument.
*
* L (output) SuperMatrix*
* The factor L from the factorization
* Pr*A*Pc=L*U (if A->Stype = SLU_NC) or
* Pr*transpose(A)*Pc=L*U (if A->Stype = SLU_NR).
* Uses compressed row subscripts storage for supernodes, i.e.,
* L has types: Stype = SLU_SC, Dtype = SLU_S, Mtype = SLU_TRLU.
*
* U (output) SuperMatrix*
* The factor U from the factorization
* Pr*A*Pc=L*U (if A->Stype = SLU_NC) or
* Pr*transpose(A)*Pc=L*U (if A->Stype = SLU_NR).
* Uses column-wise storage scheme, i.e., U has types:
* Stype = SLU_NC, Dtype = SLU_S, Mtype = SLU_TRU.
*
* B (input/output) SuperMatrix*
* B has types: Stype = SLU_DN, Dtype = SLU_S, Mtype = SLU_GE.
* On entry, the right hand side matrix.
* On exit, the solution matrix if info = 0;
*
* stat (output) SuperLUStat_t*
* Record the statistics on runtime and doubleing-point operation count.
* See util.h for the definition of 'SuperLUStat_t'.
*
* info (output) int*
* = 0: successful exit
* > 0: if info = i, and i is
* <= A->ncol: U(i,i) is exactly zero. The factorization has
* been completed, but the factor U is exactly singular,
* so the solution could not be computed.
* > A->ncol: number of bytes allocated when memory allocation
* failure occurred, plus A->ncol.
*
*/
DNformat *Bstore;
SuperMatrix *AA = NULL;/* A in SLU_NC format used by the factorization routine.*/
SuperMatrix AC; /* Matrix postmultiplied by Pc */
int lwork = 0, *etree, i;
/* Set default values for some parameters */
int panel_size; /* panel size */
int relax; /* no of columns in a relaxed snodes */
int permc_spec;
trans_t trans = NOTRANS;
double *utime;
double t; /* Temporary time */
/* Test the input parameters ... */
*info = 0;
Bstore = B->Store;
if ( options->Fact != DOFACT ) *info = -1;
else if ( A->nrow != A->ncol || A->nrow < 0 ||
(A->Stype != SLU_NC && A->Stype != SLU_NR) ||
A->Dtype != SLU_S || A->Mtype != SLU_GE )
*info = -2;
else if ( B->ncol < 0 || Bstore->lda < SUPERLU_MAX(0, A->nrow) ||
B->Stype != SLU_DN || B->Dtype != SLU_S || B->Mtype != SLU_GE )
*info = -7;
if ( *info != 0 ) {
i = -(*info);
xerbla_("sgssv", &i);
return;
}
utime = stat->utime;
/* Convert A to SLU_NC format when necessary. */
if ( A->Stype == SLU_NR ) {
NRformat *Astore = A->Store;
AA = (SuperMatrix *) SUPERLU_MALLOC( sizeof(SuperMatrix) );
sCreate_CompCol_Matrix(AA, A->ncol, A->nrow, Astore->nnz,
Astore->nzval, Astore->colind, Astore->rowptr,
SLU_NC, A->Dtype, A->Mtype);
trans = TRANS;
} else {
if ( A->Stype == SLU_NC ) AA = A;
}
t = SuperLU_timer_();
/*
* Get column permutation vector perm_c[], according to permc_spec:
* permc_spec = NATURAL: natural ordering
* permc_spec = MMD_AT_PLUS_A: minimum degree on structure of A'+A
* permc_spec = MMD_ATA: minimum degree on structure of A'*A
* permc_spec = COLAMD: approximate minimum degree column ordering
* permc_spec = MY_PERMC: the ordering already supplied in perm_c[]
*/
permc_spec = options->ColPerm;
if ( permc_spec != MY_PERMC && options->Fact == DOFACT )
get_perm_c(permc_spec, AA, perm_c);
utime[COLPERM] = SuperLU_timer_() - t;
etree = intMalloc(A->ncol);
t = SuperLU_timer_();
sp_preorder(options, AA, perm_c, etree, &AC);
utime[ETREE] = SuperLU_timer_() - t;
panel_size = sp_ienv(1);
relax = sp_ienv(2);
/*printf("Factor PA = LU ... relax %d\tw %d\tmaxsuper %d\trowblk %d\n",
relax, panel_size, sp_ienv(3), sp_ienv(4));*/
t = SuperLU_timer_();
/* Compute the LU factorization of A. */
sgstrf(options, &AC, relax, panel_size,
etree, NULL, lwork, perm_c, perm_r, L, U, stat, info);
utime[FACT] = SuperLU_timer_() - t;
t = SuperLU_timer_();
if ( *info == 0 ) {
/* Solve the system A*X=B, overwriting B with X. */
sgstrs (trans, L, U, perm_c, perm_r, B, stat, info);
}
utime[SOLVE] = SuperLU_timer_() - t;
SUPERLU_FREE (etree);
Destroy_CompCol_Permuted(&AC);
if ( A->Stype == SLU_NR ) {
Destroy_SuperMatrix_Store(AA);
SUPERLU_FREE(AA);
}
}
int main ( int argc, char *argv[] )
/******************************************************************************/
/*
Purpose:
MAIN is the main program for PSLINSOL.
Licensing:
This code is distributed under the GNU LGPL license.
Modified:
10 February 2014
Author:
Xiaoye Li
*/
{
SuperMatrix A;
NCformat *Astore;
float *a;
int *asub, *xa;
int *perm_r; /* row permutations from partial pivoting */
int *perm_c; /* column permutation vector */
SuperMatrix L; /* factor L */
SCPformat *Lstore;
SuperMatrix U; /* factor U */
NCPformat *Ustore;
SuperMatrix B;
int nrhs, ldx, info, m, n, nnz, b;
int nprocs; /* maximum number of processors to use. */
int panel_size, relax, maxsup;
int permc_spec;
trans_t trans;
float *xact, *rhs;
superlu_memusage_t superlu_memusage;
void parse_command_line();
timestamp ( );
printf ( "\n" );
printf ( "PSLINSOL:\n" );
printf ( " C/OpenMP version\n" );
printf ( " Call the OpenMP version of SuperLU to solve a linear system.\n" );
nrhs = 1;
trans = NOTRANS;
nprocs = 1;
n = 1000;
b = 1;
panel_size = sp_ienv(1);
relax = sp_ienv(2);
maxsup = sp_ienv(3);
/*
Check for any commandline input.
*/
parse_command_line ( argc, argv, &nprocs, &n, &b, &panel_size,
&relax, &maxsup );
#if ( PRNTlevel>=1 || DEBUGlevel>=1 )
cpp_defs();
#endif
#define HB
#if defined( DEN )
m = n;
nnz = n * n;
sband(n, n, nnz, &a, &asub, &xa);
#elif defined( BAND )
m = n;
nnz = (2*b+1) * n;
sband(n, b, nnz, &a, &asub, &xa);
#elif defined( BD )
nb = 5;
bs = 200;
m = n = bs * nb;
nnz = bs * bs * nb;
sblockdiag(nb, bs, nnz, &a, &asub, &xa);
#elif defined( HB )
sreadhb(&m, &n, &nnz, &a, &asub, &xa);
#else
sreadmt(&m, &n, &nnz, &a, &asub, &xa);
#endif
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if (!(rhs = floatMalloc(m * nrhs))) SUPERLU_ABORT("Malloc fails for rhs[].");
sCreate_Dense_Matrix(&B, m, nrhs, rhs, m, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
ldx = n;
sGenXtrue(n, nrhs, xact, ldx);
sFillRHS(trans, nrhs, xact, ldx, &A, &B);
if (!(perm_r = intMalloc(m))) SUPERLU_ABORT("Malloc fails for perm_r[].");
if (!(perm_c = intMalloc(n))) SUPERLU_ABORT("Malloc fails for perm_c[].");
/*
* Get column permutation vector perm_c[], according to permc_spec:
* permc_spec = 0: natural ordering
* permc_spec = 1: minimum degree ordering on structure of A'*A
* permc_spec = 2: minimum degree ordering on structure of A'+A
* permc_spec = 3: approximate minimum degree for unsymmetric matrices
*/
permc_spec = 1;
get_perm_c(permc_spec, &A, perm_c);
psgssv(nprocs, &A, perm_c, perm_r, &L, &U, &B, &info);
if ( info == 0 ) {
sinf_norm_error(nrhs, &B, xact); /* Inf. norm of the error */
Lstore = (SCPformat *) L.Store;
Ustore = (NCPformat *) U.Store;
printf("#NZ in factor L = %d\n", Lstore->nnz);
printf("#NZ in factor U = %d\n", Ustore->nnz);
printf("#NZ in L+U = %d\n", Lstore->nnz + Ustore->nnz - L.ncol);
superlu_sQuerySpace(nprocs, &L, &U, panel_size, &superlu_memusage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
superlu_memusage.for_lu/1024/1024,
superlu_memusage.total_needed/1024/1024,
superlu_memusage.expansions);
}
SUPERLU_FREE (rhs);
SUPERLU_FREE (xact);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperNode_SCP(&L);
Destroy_CompCol_NCP(&U);
/*
Terminate.
*/
printf ( "\n" );
printf ( "PSLINSOL:\n" );
printf ( " Normal end of execution.\n" );
printf ( "\n" );
timestamp ( );
return 0;
}
/* Here is a driver inspired by A. Sheffer's "cow flattener". */
static NLboolean __nlSolve_SUPERLU( NLboolean do_perm) {
/* OpenNL Context */
__NLSparseMatrix* M = &(__nlCurrentContext->M);
NLfloat* b = __nlCurrentContext->b;
NLfloat* x = __nlCurrentContext->x;
/* Compressed Row Storage matrix representation */
NLuint n = __nlCurrentContext->n;
NLuint nnz = __nlSparseMatrixNNZ(M); /* Number of Non-Zero coeffs */
NLint* xa = __NL_NEW_ARRAY(NLint, n+1);
NLfloat* rhs = __NL_NEW_ARRAY(NLfloat, n);
NLfloat* a = __NL_NEW_ARRAY(NLfloat, nnz);
NLint* asub = __NL_NEW_ARRAY(NLint, nnz);
/* Permutation vector */
NLint* perm_r = __NL_NEW_ARRAY(NLint, n);
NLint* perm = __NL_NEW_ARRAY(NLint, n);
/* SuperLU variables */
SuperMatrix A, B; /* System */
SuperMatrix L, U; /* Inverse of A */
NLint info; /* status code */
DNformat *vals = NULL; /* access to result */
float *rvals = NULL; /* access to result */
/* SuperLU options and stats */
superlu_options_t options;
SuperLUStat_t stat;
/* Temporary variables */
__NLRowColumn* Ri = NULL;
NLuint i,jj,count;
__nl_assert(!(M->storage & __NL_SYMMETRIC));
__nl_assert(M->storage & __NL_ROWS);
__nl_assert(M->m == M->n);
/*
* Step 1: convert matrix M into SuperLU compressed column
* representation.
* -------------------------------------------------------
*/
count = 0;
for(i=0; i<n; i++) {
Ri = &(M->row[i]);
xa[i] = count;
for(jj=0; jj<Ri->size; jj++) {
a[count] = Ri->coeff[jj].value;
asub[count] = Ri->coeff[jj].index;
count++;
}
}
xa[n] = nnz;
/* Save memory for SuperLU */
__nlSparseMatrixClear(M);
/*
* Rem: symmetric storage does not seem to work with
* SuperLU ... (->deactivated in main SLS::Solver driver)
*/
sCreate_CompCol_Matrix(
&A, n, n, nnz, a, asub, xa,
SLU_NR, /* Row_wise, no supernode */
SLU_S, /* floats */
SLU_GE /* general storage */
);
/* Step 2: create vector */
sCreate_Dense_Matrix(
&B, n, 1, b, n,
SLU_DN, /* Fortran-type column-wise storage */
SLU_S, /* floats */
SLU_GE /* general */
);
/* Step 3: get permutation matrix
* ------------------------------
* com_perm: 0 -> no re-ordering
* 1 -> re-ordering for A^t.A
* 2 -> re-ordering for A^t+A
* 3 -> approximate minimum degree ordering
*/
get_perm_c(do_perm ? 3 : 0, &A, perm);
/* Step 4: call SuperLU main routine
* ---------------------------------
*/
set_default_options(&options);
options.ColPerm = MY_PERMC;
StatInit(&stat);
sgssv(&options, &A, perm, perm_r, &L, &U, &B, &stat, &info);
/* Step 5: get the solution
* ------------------------
* Fortran-type column-wise storage
*/
vals = (DNformat*)B.Store;
rvals = (float*)(vals->nzval);
if(info == 0) {
for(i = 0; i < n; i++){
x[i] = rvals[i];
}
}
/* Step 6: cleanup
* ---------------
*/
/*
* For these two ones, only the "store" structure
* needs to be deallocated (the arrays have been allocated
* by us).
*/
Destroy_SuperMatrix_Store(&A);
Destroy_SuperMatrix_Store(&B);
/*
* These ones need to be fully deallocated (they have been
* allocated by SuperLU).
*/
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
StatFree(&stat);
__NL_DELETE_ARRAY(xa);
__NL_DELETE_ARRAY(rhs);
__NL_DELETE_ARRAY(a);
__NL_DELETE_ARRAY(asub);
__NL_DELETE_ARRAY(perm_r);
__NL_DELETE_ARRAY(perm);
return (info == 0);
}
void
zgsisx(superlu_options_t *options, SuperMatrix *A, int *perm_c, int *perm_r,
int *etree, char *equed, double *R, double *C,
SuperMatrix *L, SuperMatrix *U, void *work, int lwork,
SuperMatrix *B, SuperMatrix *X,
double *recip_pivot_growth, double *rcond,
mem_usage_t *mem_usage, SuperLUStat_t *stat, int *info)
{
DNformat *Bstore, *Xstore;
doublecomplex *Bmat, *Xmat;
int ldb, ldx, nrhs;
SuperMatrix *AA;/* A in SLU_NC format used by the factorization routine.*/
SuperMatrix AC; /* Matrix postmultiplied by Pc */
int colequ, equil, nofact, notran, rowequ, permc_spec, mc64;
trans_t trant;
char norm[1];
int i, j, info1;
double amax, anorm, bignum, smlnum, colcnd, rowcnd, rcmax, rcmin;
int relax, panel_size;
double diag_pivot_thresh;
double t0; /* temporary time */
double *utime;
int *perm = NULL;
/* External functions */
extern double zlangs(char *, SuperMatrix *);
Bstore = B->Store;
Xstore = X->Store;
Bmat = Bstore->nzval;
Xmat = Xstore->nzval;
ldb = Bstore->lda;
ldx = Xstore->lda;
nrhs = B->ncol;
*info = 0;
nofact = (options->Fact != FACTORED);
equil = (options->Equil == YES);
notran = (options->Trans == NOTRANS);
mc64 = (options->RowPerm == LargeDiag);
if ( nofact ) {
*(unsigned char *)equed = 'N';
rowequ = FALSE;
colequ = FALSE;
} else {
rowequ = lsame_(equed, "R") || lsame_(equed, "B");
colequ = lsame_(equed, "C") || lsame_(equed, "B");
smlnum = dlamch_("Safe minimum");
bignum = 1. / smlnum;
}
/* Test the input parameters */
if (!nofact && options->Fact != DOFACT && options->Fact != SamePattern &&
options->Fact != SamePattern_SameRowPerm &&
!notran && options->Trans != TRANS && options->Trans != CONJ &&
!equil && options->Equil != NO)
*info = -1;
else if ( A->nrow != A->ncol || A->nrow < 0 ||
(A->Stype != SLU_NC && A->Stype != SLU_NR) ||
A->Dtype != SLU_Z || A->Mtype != SLU_GE )
*info = -2;
else if (options->Fact == FACTORED &&
!(rowequ || colequ || lsame_(equed, "N")))
*info = -6;
else {
if (rowequ) {
rcmin = bignum;
rcmax = 0.;
for (j = 0; j < A->nrow; ++j) {
rcmin = SUPERLU_MIN(rcmin, R[j]);
rcmax = SUPERLU_MAX(rcmax, R[j]);
}
if (rcmin <= 0.) *info = -7;
else if ( A->nrow > 0)
rowcnd = SUPERLU_MAX(rcmin,smlnum) / SUPERLU_MIN(rcmax,bignum);
else rowcnd = 1.;
}
if (colequ && *info == 0) {
rcmin = bignum;
rcmax = 0.;
for (j = 0; j < A->nrow; ++j) {
rcmin = SUPERLU_MIN(rcmin, C[j]);
rcmax = SUPERLU_MAX(rcmax, C[j]);
}
if (rcmin <= 0.) *info = -8;
else if (A->nrow > 0)
colcnd = SUPERLU_MAX(rcmin,smlnum) / SUPERLU_MIN(rcmax,bignum);
else colcnd = 1.;
}
if (*info == 0) {
if ( lwork < -1 ) *info = -12;
else if ( B->ncol < 0 || Bstore->lda < SUPERLU_MAX(0, A->nrow) ||
B->Stype != SLU_DN || B->Dtype != SLU_Z ||
B->Mtype != SLU_GE )
*info = -13;
else if ( X->ncol < 0 || Xstore->lda < SUPERLU_MAX(0, A->nrow) ||
(B->ncol != 0 && B->ncol != X->ncol) ||
X->Stype != SLU_DN ||
X->Dtype != SLU_Z || X->Mtype != SLU_GE )
*info = -14;
}
}
if (*info != 0) {
i = -(*info);
xerbla_("zgsisx", &i);
return;
}
/* Initialization for factor parameters */
panel_size = sp_ienv(1);
relax = sp_ienv(2);
diag_pivot_thresh = options->DiagPivotThresh;
utime = stat->utime;
/* Convert A to SLU_NC format when necessary. */
if ( A->Stype == SLU_NR ) {
NRformat *Astore = A->Store;
AA = (SuperMatrix *) SUPERLU_MALLOC( sizeof(SuperMatrix) );
zCreate_CompCol_Matrix(AA, A->ncol, A->nrow, Astore->nnz,
Astore->nzval, Astore->colind, Astore->rowptr,
SLU_NC, A->Dtype, A->Mtype);
if ( notran ) { /* Reverse the transpose argument. */
trant = TRANS;
notran = 0;
} else {
trant = NOTRANS;
notran = 1;
}
} else { /* A->Stype == SLU_NC */
trant = options->Trans;
AA = A;
}
if ( nofact ) {
register int i, j;
NCformat *Astore = AA->Store;
int nnz = Astore->nnz;
int *colptr = Astore->colptr;
int *rowind = Astore->rowind;
doublecomplex *nzval = (doublecomplex *)Astore->nzval;
int n = AA->nrow;
if ( mc64 ) {
*equed = 'B';
rowequ = colequ = 1;
t0 = SuperLU_timer_();
if ((perm = intMalloc(n)) == NULL)
ABORT("SUPERLU_MALLOC fails for perm[]");
info1 = zldperm(5, n, nnz, colptr, rowind, nzval, perm, R, C);
if (info1 > 0) { /* MC64 fails, call zgsequ() later */
mc64 = 0;
SUPERLU_FREE(perm);
perm = NULL;
} else {
for (i = 0; i < n; i++) {
R[i] = exp(R[i]);
C[i] = exp(C[i]);
}
/* permute and scale the matrix */
for (j = 0; j < n; j++) {
for (i = colptr[j]; i < colptr[j + 1]; i++) {
zd_mult(&nzval[i], &nzval[i], R[rowind[i]] * C[j]);
rowind[i] = perm[rowind[i]];
}
}
}
utime[EQUIL] = SuperLU_timer_() - t0;
}
if ( !mc64 & equil ) {
t0 = SuperLU_timer_();
/* Compute row and column scalings to equilibrate the matrix A. */
zgsequ(AA, R, C, &rowcnd, &colcnd, &amax, &info1);
if ( info1 == 0 ) {
/* Equilibrate matrix A. */
zlaqgs(AA, R, C, rowcnd, colcnd, amax, equed);
rowequ = lsame_(equed, "R") || lsame_(equed, "B");
colequ = lsame_(equed, "C") || lsame_(equed, "B");
}
utime[EQUIL] = SuperLU_timer_() - t0;
}
}
if ( nrhs > 0 ) {
/* Scale the right hand side if equilibration was performed. */
if ( notran ) {
if ( rowequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i) {
zd_mult(&Bmat[i+j*ldb], &Bmat[i+j*ldb], R[i]);
}
}
} else if ( colequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i) {
zd_mult(&Bmat[i+j*ldb], &Bmat[i+j*ldb], C[i]);
}
}
}
if ( nofact ) {
t0 = SuperLU_timer_();
/*
* Gnet column permutation vector perm_c[], according to permc_spec:
* permc_spec = NATURAL: natural ordering
* permc_spec = MMD_AT_PLUS_A: minimum degree on structure of A'+A
* permc_spec = MMD_ATA: minimum degree on structure of A'*A
* permc_spec = COLAMD: approximate minimum degree column ordering
* permc_spec = MY_PERMC: the ordering already supplied in perm_c[]
*/
permc_spec = options->ColPerm;
if ( permc_spec != MY_PERMC && options->Fact == DOFACT )
get_perm_c(permc_spec, AA, perm_c);
utime[COLPERM] = SuperLU_timer_() - t0;
t0 = SuperLU_timer_();
sp_preorder(options, AA, perm_c, etree, &AC);
utime[ETREE] = SuperLU_timer_() - t0;
/* Compute the LU factorization of A*Pc. */
t0 = SuperLU_timer_();
zgsitrf(options, &AC, relax, panel_size, etree, work, lwork,
perm_c, perm_r, L, U, stat, info);
utime[FACT] = SuperLU_timer_() - t0;
if ( lwork == -1 ) {
mem_usage->total_needed = *info - A->ncol;
return;
}
}
if ( options->PivotGrowth ) {
if ( *info > 0 ) return;
/* Compute the reciprocal pivot growth factor *recip_pivot_growth. */
*recip_pivot_growth = zPivotGrowth(A->ncol, AA, perm_c, L, U);
}
if ( options->ConditionNumber ) {
/* Estimate the reciprocal of the condition number of A. */
t0 = SuperLU_timer_();
if ( notran ) {
*(unsigned char *)norm = '1';
} else {
*(unsigned char *)norm = 'I';
}
anorm = zlangs(norm, AA);
zgscon(norm, L, U, anorm, rcond, stat, &info1);
utime[RCOND] = SuperLU_timer_() - t0;
}
if ( nrhs > 0 ) {
/* Compute the solution matrix X. */
for (j = 0; j < nrhs; j++) /* Save a copy of the right hand sides */
for (i = 0; i < B->nrow; i++)
Xmat[i + j*ldx] = Bmat[i + j*ldb];
t0 = SuperLU_timer_();
zgstrs (trant, L, U, perm_c, perm_r, X, stat, &info1);
utime[SOLVE] = SuperLU_timer_() - t0;
/* Transform the solution matrix X to a solution of the original
system. */
if ( notran ) {
if ( colequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i) {
zd_mult(&Xmat[i+j*ldx], &Xmat[i+j*ldx], C[i]);
}
}
} else {
if ( rowequ ) {
if (perm) {
doublecomplex *tmp;
int n = A->nrow;
if ((tmp = doublecomplexMalloc(n)) == NULL)
ABORT("SUPERLU_MALLOC fails for tmp[]");
for (j = 0; j < nrhs; j++) {
for (i = 0; i < n; i++)
tmp[i] = Xmat[i + j * ldx]; /*dcopy*/
for (i = 0; i < n; i++)
zd_mult(&Xmat[i+j*ldx], &tmp[perm[i]], R[i]);
}
SUPERLU_FREE(tmp);
} else {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i) {
zd_mult(&Xmat[i+j*ldx], &Xmat[i+j*ldx], R[i]);
}
}
}
}
} /* end if nrhs > 0 */
if ( options->ConditionNumber ) {
/* Set INFO = A->ncol+1 if the matrix is singular to working precision. */
if ( *rcond < dlamch_("E") && *info == 0) *info = A->ncol + 1;
}
if (perm) SUPERLU_FREE(perm);
if ( nofact ) {
ilu_zQuerySpace(L, U, mem_usage);
Destroy_CompCol_Permuted(&AC);
}
if ( A->Stype == SLU_NR ) {
Destroy_SuperMatrix_Store(AA);
SUPERLU_FREE(AA);
}
}
PyObject *newSuperLUObject(SuperMatrix * A, PyObject * option_dict,
int intype, int ilu)
{
/* A must be in SLU_NC format used by the factorization routine. */
SuperLUObject *self;
SuperMatrix AC = { 0 }; /* Matrix postmultiplied by Pc */
int lwork = 0;
int *etree = NULL;
int info;
int n;
superlu_options_t options;
SuperLUStat_t stat = { 0 };
int panel_size, relax;
n = A->ncol;
if (!set_superlu_options_from_dict(&options, ilu, option_dict,
&panel_size, &relax)) {
return NULL;
}
/* Create SLUObject */
self = PyObject_New(SuperLUObject, &SuperLUType);
if (self == NULL)
return PyErr_NoMemory();
self->m = A->nrow;
self->n = n;
self->perm_r = NULL;
self->perm_c = NULL;
self->L.Store = NULL;
self->U.Store = NULL;
self->cached_U = NULL;
self->cached_L = NULL;
self->type = intype;
if (setjmp(_superlu_py_jmpbuf))
goto fail;
/* Calculate and apply minimum degree ordering */
etree = intMalloc(n);
self->perm_r = intMalloc(n);
self->perm_c = intMalloc(n);
StatInit(&stat);
get_perm_c(options.ColPerm, A, self->perm_c); /* calc column permutation */
sp_preorder(&options, A, self->perm_c, etree, &AC); /* apply column
* permutation */
/* Perform factorization */
if (!CHECK_SLU_TYPE(SLU_TYPECODE_TO_NPY(A->Dtype))) {
PyErr_SetString(PyExc_ValueError, "Invalid type in SuperMatrix.");
goto fail;
}
if (ilu) {
gsitrf(SLU_TYPECODE_TO_NPY(A->Dtype),
&options, &AC, relax, panel_size,
etree, NULL, lwork, self->perm_c, self->perm_r,
&self->L, &self->U, &stat, &info);
}
else {
gstrf(SLU_TYPECODE_TO_NPY(A->Dtype),
&options, &AC, relax, panel_size,
etree, NULL, lwork, self->perm_c, self->perm_r,
&self->L, &self->U, &stat, &info);
}
if (info) {
if (info < 0)
PyErr_SetString(PyExc_SystemError,
"gstrf was called with invalid arguments");
else {
if (info <= n)
PyErr_SetString(PyExc_RuntimeError,
"Factor is exactly singular");
else
PyErr_NoMemory();
}
goto fail;
}
/* free memory */
SUPERLU_FREE(etree);
Destroy_CompCol_Permuted(&AC);
StatFree(&stat);
return (PyObject *) self;
fail:
SUPERLU_FREE(etree);
XDestroy_CompCol_Permuted(&AC);
XStatFree(&stat);
Py_DECREF(self);
return NULL;
}
main(int argc, char *argv[])
{
SuperMatrix A;
NCformat *Astore;
doublecomplex *a;
int *asub, *xa;
int *perm_r; /* row permutations from partial pivoting */
int *perm_c; /* column permutation vector */
SuperMatrix L; /* factor L */
SCformat *Lstore;
SuperMatrix U; /* factor U */
NCformat *Ustore;
SuperMatrix B;
int nrhs, ldx, info, panel_size, m, n, nnz, permc_spec;
char trans[1];
doublecomplex *xact, *rhs;
mem_usage_t mem_usage;
nrhs = 1;
*trans = 'N';
zreadhb(&m, &n, &nnz, &a, &asub, &xa);
zCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_Z, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if ( !(rhs = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhs[].");
zCreate_Dense_Matrix(&B, m, nrhs, rhs, m, SLU_DN, SLU_Z, SLU_GE);
xact = doublecomplexMalloc(n * nrhs);
ldx = n;
zGenXtrue(n, nrhs, xact, ldx);
zFillRHS(trans, nrhs, xact, ldx, &A, &B);
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
/*
* Get column permutation vector perm_c[], according to permc_spec:
* permc_spec = 0: natural ordering
* permc_spec = 1: minimum degree on structure of A'*A
* permc_spec = 2: minimum degree on structure of A'+A
* permc_spec = 3: approximate minimum degree for unsymmetric matrices
*/
permc_spec = 1;
get_perm_c(permc_spec, &A, perm_c);
panel_size = sp_ienv(1);
zgssv(&A, perm_c, perm_r, &L, &U, &B, &info);
if ( info == 0 ) {
zinf_norm_error(nrhs, &B, xact); /* Inf. norm of the error */
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
zQuerySpace(&L, &U, panel_size, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions);
} else {
printf("zgssv() error returns INFO= %d\n", info);
if ( info <= n ) { /* factorization completes */
zQuerySpace(&L, &U, panel_size, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions);
}
}
SUPERLU_FREE (rhs);
SUPERLU_FREE (xact);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
}
main(int argc, char *argv[])
{
SuperMatrix A, AC, L, U, B;
NCformat *Astore;
SCPformat *Lstore;
NCPformat *Ustore;
superlumt_options_t superlumt_options;
pxgstrf_shared_t pxgstrf_shared;
pdgstrf_threadarg_t *pdgstrf_threadarg;
int nprocs;
fact_t fact;
trans_t trans;
yes_no_t refact, usepr;
double u, drop_tol;
double *a;
int *asub, *xa;
int *perm_c; /* column permutation vector */
int *perm_r; /* row permutations from partial pivoting */
void *work;
int info, lwork, nrhs, ldx;
int m, n, nnz, permc_spec, panel_size, relax;
int i, firstfact;
double *rhsb, *xact;
Gstat_t Gstat;
flops_t flopcnt;
void parse_command_line();
/* Default parameters to control factorization. */
nprocs = 1;
fact = EQUILIBRATE;
trans = NOTRANS;
panel_size = sp_ienv(1);
relax = sp_ienv(2);
u = 1.0;
usepr = NO;
drop_tol = 0.0;
work = NULL;
lwork = 0;
nrhs = 1;
/* Get the number of processes from command line. */
parse_command_line(argc, argv, &nprocs);
/* Read the input matrix stored in Harwell-Boeing format. */
dreadhb(&m, &n, &nnz, &a, &asub, &xa);
/* Set up the sparse matrix data structure for A. */
dCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_D, SLU_GE);
if (!(rhsb = doubleMalloc(m * nrhs))) SUPERLU_ABORT("Malloc fails for rhsb[].");
dCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_D, SLU_GE);
xact = doubleMalloc(n * nrhs);
ldx = n;
dGenXtrue(n, nrhs, xact, ldx);
dFillRHS(trans, nrhs, xact, ldx, &A, &B);
if (!(perm_r = intMalloc(m))) SUPERLU_ABORT("Malloc fails for perm_r[].");
if (!(perm_c = intMalloc(n))) SUPERLU_ABORT("Malloc fails for perm_c[].");
/********************************
* THE FIRST TIME FACTORIZATION *
********************************/
/* ------------------------------------------------------------
Allocate storage and initialize statistics variables.
------------------------------------------------------------*/
StatAlloc(n, nprocs, panel_size, relax, &Gstat);
StatInit(n, nprocs, &Gstat);
/* ------------------------------------------------------------
Get column permutation vector perm_c[], according to permc_spec:
permc_spec = 0: natural ordering
permc_spec = 1: minimum degree ordering on structure of A'*A
permc_spec = 2: minimum degree ordering on structure of A'+A
permc_spec = 3: approximate minimum degree for unsymmetric matrices
------------------------------------------------------------*/
permc_spec = 1;
get_perm_c(permc_spec, &A, perm_c);
/* ------------------------------------------------------------
Initialize the option structure superlumt_options using the
user-input parameters;
Apply perm_c to the columns of original A to form AC.
------------------------------------------------------------*/
refact= NO;
pdgstrf_init(nprocs, fact, trans, refact, panel_size, relax,
u, usepr, drop_tol, perm_c, perm_r,
work, lwork, &A, &AC, &superlumt_options, &Gstat);
/* ------------------------------------------------------------
Compute the LU factorization of A.
The following routine will create nprocs threads.
------------------------------------------------------------*/
pdgstrf(&superlumt_options, &AC, perm_r, &L, &U, &Gstat, &info);
flopcnt = 0;
for (i = 0; i < nprocs; ++i) flopcnt += Gstat.procstat[i].fcops;
Gstat.ops[FACT] = flopcnt;
/* ------------------------------------------------------------
Solve the system A*X=B, overwriting B with X.
------------------------------------------------------------*/
dgstrs(trans, &L, &U, perm_r, perm_c, &B, &Gstat, &info);
printf("\n** Result of sparse LU **\n");
dinf_norm_error(nrhs, &B, xact); /* Check inf. norm of the error */
Destroy_CompCol_Permuted(&AC); /* Free extra arrays in AC. */
/*********************************
* THE SUBSEQUENT FACTORIZATIONS *
*********************************/
/* ------------------------------------------------------------
Re-initialize statistics variables and options used by the
factorization routine pdgstrf().
------------------------------------------------------------*/
StatInit(n, nprocs, &Gstat);
refact= YES;
pdgstrf_init(nprocs, fact, trans, refact, panel_size, relax,
u, usepr, drop_tol, perm_c, perm_r,
work, lwork, &A, &AC, &superlumt_options, &Gstat);
/* ------------------------------------------------------------
Compute the LU factorization of A.
The following routine will create nprocs threads.
------------------------------------------------------------*/
pdgstrf(&superlumt_options, &AC, perm_r, &L, &U, &Gstat, &info);
flopcnt = 0;
for (i = 0; i < nprocs; ++i) flopcnt += Gstat.procstat[i].fcops;
Gstat.ops[FACT] = flopcnt;
/* ------------------------------------------------------------
Re-generate right-hand side B, then solve A*X= B.
------------------------------------------------------------*/
dFillRHS(trans, nrhs, xact, ldx, &A, &B);
dgstrs(trans, &L, &U, perm_r, perm_c, &B, &Gstat, &info);
/* ------------------------------------------------------------
Deallocate storage after factorization.
------------------------------------------------------------*/
pxgstrf_finalize(&superlumt_options, &AC);
printf("\n** Result of sparse LU **\n");
dinf_norm_error(nrhs, &B, xact); /* Check inf. norm of the error */
Lstore = (SCPformat *) L.Store;
Ustore = (NCPformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
fflush(stdout);
SUPERLU_FREE (rhsb);
SUPERLU_FREE (xact);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
if ( lwork >= 0 ) {
Destroy_SuperNode_SCP(&L);
Destroy_CompCol_NCP(&U);
}
StatFree(&Gstat);
}
bool SparseMatrix::solveSLU (Vector& B)
{
int ierr = ncol+1;
if (!factored) this->optimiseSLU();
#ifdef HAS_SUPERLU_MT
if (!slu) {
// Create a new SuperLU matrix
slu = new SuperLUdata;
slu->perm_c = new int[ncol];
slu->perm_r = new int[nrow];
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
else {
Destroy_SuperMatrix_Store(&slu->A);
Destroy_SuperNode_Matrix(&slu->L);
Destroy_CompCol_Matrix(&slu->U);
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
// Get column permutation vector perm_c[], according to permc_spec:
// permc_spec = 0: natural ordering
// permc_spec = 1: minimum degree ordering on structure of A'*A
// permc_spec = 2: minimum degree ordering on structure of A'+A
// permc_spec = 3: approximate minimum degree for unsymmetric matrices
int permc_spec = 1;
get_perm_c(permc_spec, &slu->A, slu->perm_c);
// Create right-hand-side/solution vector(s)
size_t nrhs = B.size() / nrow;
SuperMatrix Bmat;
dCreate_Dense_Matrix(&Bmat, nrow, nrhs, B.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
// Invoke the simple driver
pdgssv(numThreads, &slu->A, slu->perm_c, slu->perm_r,
&slu->L, &slu->U, &Bmat, &ierr);
if (ierr > 0)
std::cerr <<"SuperLU_MT Failure "<< ierr << std::endl;
Destroy_SuperMatrix_Store(&Bmat);
#elif defined(HAS_SUPERLU)
if (!slu) {
// Create a new SuperLU matrix
slu = new SuperLUdata(1);
slu->perm_c = new int[ncol];
slu->perm_r = new int[nrow];
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
else if (factored)
slu->opts->Fact = FACTORED; // Re-use previous factorization
else {
Destroy_SuperMatrix_Store(&slu->A);
Destroy_SuperNode_Matrix(&slu->L);
Destroy_CompCol_Matrix(&slu->U);
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
// Create right-hand-side/solution vector(s)
size_t nrhs = B.size() / nrow;
SuperMatrix Bmat;
dCreate_Dense_Matrix(&Bmat, nrow, nrhs, B.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
SuperLUStat_t stat;
StatInit(&stat);
// Invoke the simple driver
dgssv(slu->opts, &slu->A, slu->perm_c, slu->perm_r,
&slu->L, &slu->U, &Bmat, &stat, &ierr);
if (ierr > 0)
std::cerr <<"SuperLU Failure "<< ierr << std::endl;
else
factored = true;
if (printSLUstat)
StatPrint(&stat);
StatFree(&stat);
Destroy_SuperMatrix_Store(&Bmat);
#else
std::cerr <<"SparseMatrix::solve: SuperLU solver not available"<< std::endl;
#endif
return ierr == 0;
}
bool SparseMatrix::solveSLUx (Vector& B, Real* rcond)
{
int ierr = ncol+1;
if (!factored) this->optimiseSLU();
#ifdef HAS_SUPERLU_MT
if (!slu) {
// Create a new SuperLU matrix
slu = new SuperLUdata(numThreads);
slu->equed = NOEQUIL;
slu->perm_c = new int[ncol];
slu->perm_r = new int[nrow];
slu->C = new Real[ncol];
slu->R = new Real[nrow];
slu->opts->etree = new int[ncol];
slu->opts->colcnt_h = new int[ncol];
slu->opts->part_super_h = new int[ncol];
memset(slu->opts->colcnt_h, 0, ncol*sizeof(int));
memset(slu->opts->part_super_h, 0, ncol*sizeof(int));
memset(slu->opts->etree, 0, ncol*sizeof(int));
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
// Get column permutation vector perm_c[], according to permc_spec:
// permc_spec = 0: natural ordering
// permc_spec = 1: minimum degree ordering on structure of A'*A
// permc_spec = 2: minimum degree ordering on structure of A'+A
// permc_spec = 3: approximate minimum degree for unsymmetric matrices
int permc_spec = 1;
get_perm_c(permc_spec, &slu->A, slu->perm_c);
}
else if (factored)
slu->opts->fact = FACTORED; // Re-use previous factorization
else
slu->opts->refact = YES; // Re-use previous ordering
// Create right-hand-side and solution vector(s)
Vector X(B.size());
SuperMatrix Bmat, Xmat;
const size_t nrhs = B.size() / nrow;
dCreate_Dense_Matrix(&Bmat, nrow, nrhs, B.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
dCreate_Dense_Matrix(&Xmat, nrow, nrhs, X.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
Real ferr[nrhs], berr[nrhs];
superlu_memusage_t mem_usage;
// Invoke the expert driver
pdgssvx(numThreads, slu->opts, &slu->A, slu->perm_c, slu->perm_r,
&slu->equed, slu->R, slu->C, &slu->L, &slu->U, &Bmat, &Xmat,
&slu->rpg, &slu->rcond, ferr, berr, &mem_usage, &ierr);
B.swap(X);
if (ierr > 0)
std::cerr <<"SuperLU_MT Failure "<< ierr << std::endl;
else if (!factored)
{
factored = true;
if (rcond)
*rcond = slu->rcond;
}
Destroy_SuperMatrix_Store(&Bmat);
Destroy_SuperMatrix_Store(&Xmat);
#elif defined(HAS_SUPERLU)
if (!slu) {
// Create a new SuperLU matrix
slu = new SuperLUdata(1);
slu->perm_c = new int[ncol];
slu->perm_r = new int[nrow];
slu->etree = new int[ncol];
slu->C = new Real[ncol];
slu->R = new Real[nrow];
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
else if (factored)
slu->opts->Fact = FACTORED; // Re-use previous factorization
else {
Destroy_SuperMatrix_Store(&slu->A);
Destroy_SuperNode_Matrix(&slu->L);
Destroy_CompCol_Matrix(&slu->U);
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
// Create right-hand-side vector and solution vector
Vector X(B.size());
SuperMatrix Bmat, Xmat;
const size_t nrhs = B.size() / nrow;
dCreate_Dense_Matrix(&Bmat, nrow, nrhs, B.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
dCreate_Dense_Matrix(&Xmat, nrow, nrhs, X.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
slu->opts->ConditionNumber = printSLUstat || rcond ? YES : NO;
slu->opts->PivotGrowth = printSLUstat ? YES : NO;
void* work = 0;
int lwork = 0;
Real ferr[nrhs], berr[nrhs];
mem_usage_t mem_usage;
SuperLUStat_t stat;
StatInit(&stat);
// Invoke the expert driver
#if SUPERLU_VERSION == 5
GlobalLU_t Glu;
dgssvx(slu->opts, &slu->A, slu->perm_c, slu->perm_r, slu->etree, slu->equed,
slu->R, slu->C, &slu->L, &slu->U, work, lwork, &Bmat, &Xmat,
&slu->rpg, &slu->rcond, ferr, berr, &Glu, &mem_usage, &stat, &ierr);
#else
dgssvx(slu->opts, &slu->A, slu->perm_c, slu->perm_r, slu->etree, slu->equed,
slu->R, slu->C, &slu->L, &slu->U, work, lwork, &Bmat, &Xmat,
&slu->rpg, &slu->rcond, ferr, berr, &mem_usage, &stat, &ierr);
#endif
B.swap(X);
if (ierr > 0)
std::cerr <<"SuperLU Failure "<< ierr << std::endl;
else if (!factored)
{
factored = true;
if (rcond)
*rcond = slu->rcond;
}
if (printSLUstat)
{
StatPrint(&stat);
IFEM::cout <<"Reciprocal condition number = "<< slu->rcond
<<"\nReciprocal pivot growth = "<< slu->rpg << std::endl;
}
StatFree(&stat);
Destroy_SuperMatrix_Store(&Bmat);
Destroy_SuperMatrix_Store(&Xmat);
#else
std::cerr <<"SparseMatrix::solve: SuperLU solver not available"<< std::endl;
#endif
return ierr == 0;
}
void
sgssvx(superlu_options_t *options, SuperMatrix *A, int *perm_c, int *perm_r,
int *etree, char *equed, float *R, float *C,
SuperMatrix *L, SuperMatrix *U, void *work, int lwork,
SuperMatrix *B, SuperMatrix *X, float *recip_pivot_growth,
float *rcond, float *ferr, float *berr,
mem_usage_t *mem_usage, SuperLUStat_t *stat, int *info )
{
DNformat *Bstore, *Xstore;
float *Bmat, *Xmat;
int ldb, ldx, nrhs;
SuperMatrix *AA;/* A in SLU_NC format used by the factorization routine.*/
SuperMatrix AC; /* Matrix postmultiplied by Pc */
int colequ, equil, nofact, notran, rowequ, permc_spec;
trans_t trant;
char norm[1];
int i, j, info1;
float amax, anorm, bignum, smlnum, colcnd, rowcnd, rcmax, rcmin;
int relax, panel_size;
float diag_pivot_thresh;
double t0; /* temporary time */
double *utime;
/* External functions */
extern float slangs(char *, SuperMatrix *);
Bstore = B->Store;
Xstore = X->Store;
Bmat = Bstore->nzval;
Xmat = Xstore->nzval;
ldb = Bstore->lda;
ldx = Xstore->lda;
nrhs = B->ncol;
*info = 0;
nofact = (options->Fact != FACTORED);
equil = (options->Equil == YES);
notran = (options->Trans == NOTRANS);
if ( nofact ) {
*(unsigned char *)equed = 'N';
rowequ = FALSE;
colequ = FALSE;
} else {
rowequ = lsame_(equed, "R") || lsame_(equed, "B");
colequ = lsame_(equed, "C") || lsame_(equed, "B");
smlnum = slamch_("Safe minimum");
bignum = 1. / smlnum;
}
#if 0
printf("dgssvx: Fact=%4d, Trans=%4d, equed=%c\n",
options->Fact, options->Trans, *equed);
#endif
/* Test the input parameters */
if (options->Fact != DOFACT && options->Fact != SamePattern &&
options->Fact != SamePattern_SameRowPerm &&
options->Fact != FACTORED &&
options->Trans != NOTRANS && options->Trans != TRANS &&
options->Trans != CONJ &&
options->Equil != NO && options->Equil != YES)
*info = -1;
else if ( A->nrow != A->ncol || A->nrow < 0 ||
(A->Stype != SLU_NC && A->Stype != SLU_NR) ||
A->Dtype != SLU_S || A->Mtype != SLU_GE )
*info = -2;
else if (options->Fact == FACTORED &&
!(rowequ || colequ || lsame_(equed, "N")))
*info = -6;
else {
if (rowequ) {
rcmin = bignum;
rcmax = 0.;
for (j = 0; j < A->nrow; ++j) {
rcmin = SUPERLU_MIN(rcmin, R[j]);
rcmax = SUPERLU_MAX(rcmax, R[j]);
}
if (rcmin <= 0.) *info = -7;
else if ( A->nrow > 0)
rowcnd = SUPERLU_MAX(rcmin,smlnum) / SUPERLU_MIN(rcmax,bignum);
else rowcnd = 1.;
}
if (colequ && *info == 0) {
rcmin = bignum;
rcmax = 0.;
for (j = 0; j < A->nrow; ++j) {
rcmin = SUPERLU_MIN(rcmin, C[j]);
rcmax = SUPERLU_MAX(rcmax, C[j]);
}
if (rcmin <= 0.) *info = -8;
else if (A->nrow > 0)
colcnd = SUPERLU_MAX(rcmin,smlnum) / SUPERLU_MIN(rcmax,bignum);
else colcnd = 1.;
}
if (*info == 0) {
if ( lwork < -1 ) *info = -12;
else if ( B->ncol < 0 ) *info = -13;
else if ( B->ncol > 0 ) { /* no checking if B->ncol=0 */
if ( Bstore->lda < SUPERLU_MAX(0, A->nrow) ||
B->Stype != SLU_DN || B->Dtype != SLU_S ||
B->Mtype != SLU_GE )
*info = -13;
}
if ( X->ncol < 0 ) *info = -14;
else if ( X->ncol > 0 ) { /* no checking if X->ncol=0 */
if ( Xstore->lda < SUPERLU_MAX(0, A->nrow) ||
(B->ncol != 0 && B->ncol != X->ncol) ||
X->Stype != SLU_DN ||
X->Dtype != SLU_S || X->Mtype != SLU_GE )
*info = -14;
}
}
}
if (*info != 0) {
i = -(*info);
xerbla_("sgssvx", &i);
return;
}
/* Initialization for factor parameters */
panel_size = sp_ienv(1);
relax = sp_ienv(2);
diag_pivot_thresh = options->DiagPivotThresh;
utime = stat->utime;
/* Convert A to SLU_NC format when necessary. */
if ( A->Stype == SLU_NR ) {
NRformat *Astore = A->Store;
AA = (SuperMatrix *) SUPERLU_MALLOC( sizeof(SuperMatrix) );
sCreate_CompCol_Matrix(AA, A->ncol, A->nrow, Astore->nnz,
Astore->nzval, Astore->colind, Astore->rowptr,
SLU_NC, A->Dtype, A->Mtype);
if ( notran ) { /* Reverse the transpose argument. */
trant = TRANS;
notran = 0;
} else {
trant = NOTRANS;
notran = 1;
}
} else { /* A->Stype == SLU_NC */
trant = options->Trans;
AA = A;
}
if ( nofact && equil ) {
t0 = SuperLU_timer_();
/* Compute row and column scalings to equilibrate the matrix A. */
sgsequ(AA, R, C, &rowcnd, &colcnd, &amax, &info1);
if ( info1 == 0 ) {
/* Equilibrate matrix A. */
slaqgs(AA, R, C, rowcnd, colcnd, amax, equed);
rowequ = lsame_(equed, "R") || lsame_(equed, "B");
colequ = lsame_(equed, "C") || lsame_(equed, "B");
}
utime[EQUIL] = SuperLU_timer_() - t0;
}
if ( nofact ) {
t0 = SuperLU_timer_();
/*
* Gnet column permutation vector perm_c[], according to permc_spec:
* permc_spec = NATURAL: natural ordering
* permc_spec = MMD_AT_PLUS_A: minimum degree on structure of A'+A
* permc_spec = MMD_ATA: minimum degree on structure of A'*A
* permc_spec = COLAMD: approximate minimum degree column ordering
* permc_spec = MY_PERMC: the ordering already supplied in perm_c[]
*/
permc_spec = options->ColPerm;
if ( permc_spec != MY_PERMC && options->Fact == DOFACT )
get_perm_c(permc_spec, AA, perm_c);
utime[COLPERM] = SuperLU_timer_() - t0;
t0 = SuperLU_timer_();
sp_preorder(options, AA, perm_c, etree, &AC);
utime[ETREE] = SuperLU_timer_() - t0;
/* printf("Factor PA = LU ... relax %d\tw %d\tmaxsuper %d\trowblk %d\n",
relax, panel_size, sp_ienv(3), sp_ienv(4));
fflush(stdout); */
/* Compute the LU factorization of A*Pc. */
t0 = SuperLU_timer_();
sgstrf(options, &AC, relax, panel_size, etree,
work, lwork, perm_c, perm_r, L, U, stat, info);
utime[FACT] = SuperLU_timer_() - t0;
if ( lwork == -1 ) {
mem_usage->total_needed = *info - A->ncol;
return;
}
}
if ( options->PivotGrowth ) {
if ( *info > 0 ) {
if ( *info <= A->ncol ) {
/* Compute the reciprocal pivot growth factor of the leading
rank-deficient *info columns of A. */
*recip_pivot_growth = sPivotGrowth(*info, AA, perm_c, L, U);
}
return;
}
/* Compute the reciprocal pivot growth factor *recip_pivot_growth. */
*recip_pivot_growth = sPivotGrowth(A->ncol, AA, perm_c, L, U);
}
if ( options->ConditionNumber ) {
/* Estimate the reciprocal of the condition number of A. */
t0 = SuperLU_timer_();
if ( notran ) {
*(unsigned char *)norm = '1';
} else {
*(unsigned char *)norm = 'I';
}
anorm = slangs(norm, AA);
sgscon(norm, L, U, anorm, rcond, stat, info);
utime[RCOND] = SuperLU_timer_() - t0;
}
if ( nrhs > 0 ) {
/* Scale the right hand side if equilibration was performed. */
if ( notran ) {
if ( rowequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i)
Bmat[i + j*ldb] *= R[i];
}
} else if ( colequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i)
Bmat[i + j*ldb] *= C[i];
}
/* Compute the solution matrix X. */
for (j = 0; j < nrhs; j++) /* Save a copy of the right hand sides */
for (i = 0; i < B->nrow; i++)
Xmat[i + j*ldx] = Bmat[i + j*ldb];
t0 = SuperLU_timer_();
sgstrs (trant, L, U, perm_c, perm_r, X, stat, info);
utime[SOLVE] = SuperLU_timer_() - t0;
/* Use iterative refinement to improve the computed solution and compute
error bounds and backward error estimates for it. */
t0 = SuperLU_timer_();
if ( options->IterRefine != NOREFINE ) {
sgsrfs(trant, AA, L, U, perm_c, perm_r, equed, R, C, B,
X, ferr, berr, stat, info);
} else {
for (j = 0; j < nrhs; ++j) ferr[j] = berr[j] = 1.0;
}
utime[REFINE] = SuperLU_timer_() - t0;
/* Transform the solution matrix X to a solution of the original system. */
if ( notran ) {
if ( colequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i)
Xmat[i + j*ldx] *= C[i];
}
} else if ( rowequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i)
Xmat[i + j*ldx] *= R[i];
}
} /* end if nrhs > 0 */
if ( options->ConditionNumber ) {
/* Set INFO = A->ncol+1 if the matrix is singular to working precision. */
if ( *rcond < slamch_("E") ) *info = A->ncol + 1;
}
if ( nofact ) {
sQuerySpace(L, U, mem_usage);
Destroy_CompCol_Permuted(&AC);
}
if ( A->Stype == SLU_NR ) {
Destroy_SuperMatrix_Store(AA);
SUPERLU_FREE(AA);
}
}
main(int argc, char *argv[])
{
char fact[1], equed[1], trans[1], refact[1];
SuperMatrix A, L, U;
SuperMatrix B, X;
NCformat *Astore;
NCformat *Ustore;
SCformat *Lstore;
complex *a;
int *asub, *xa;
int *perm_r; /* row permutations from partial pivoting */
int *perm_c; /* column permutation vector */
int *etree;
void *work;
factor_param_t iparam;
int info, lwork, nrhs, ldx, panel_size, relax;
int m, n, nnz, permc_spec;
complex *rhsb, *rhsx, *xact;
float *R, *C;
float *ferr, *berr;
float u, rpg, rcond;
int i, firstfact;
mem_usage_t mem_usage;
void parse_command_line();
/* Defaults */
lwork = 0;
*fact = 'E';
*equed = 'N';
*trans = 'N';
*refact = 'N';
nrhs = 1;
panel_size = sp_ienv(1);
relax = sp_ienv(2);
u = 1.0;
parse_command_line(argc, argv, &lwork, &panel_size, &relax, &u,
fact, trans, refact);
firstfact = lsame_(fact, "F") || lsame_(refact, "Y");
iparam.panel_size = panel_size;
iparam.relax = relax;
iparam.diag_pivot_thresh = u;
iparam.drop_tol = -1;
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) {
ABORT("CLINSOLX: cannot allocate work[]");
}
}
creadhb(&m, &n, &nnz, &a, &asub, &xa);
cCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_C, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if ( !(rhsb = complexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[].");
if ( !(rhsx = complexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[].");
cCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_C, SLU_GE);
cCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_C, SLU_GE);
xact = complexMalloc(n * nrhs);
ldx = n;
cGenXtrue(n, nrhs, xact, ldx);
cFillRHS(trans, nrhs, xact, ldx, &A, &B);
if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
/*
* Get column permutation vector perm_c[], according to permc_spec:
* permc_spec = 0: natural ordering
* permc_spec = 1: minimum degree on structure of A'*A
* permc_spec = 2: minimum degree on structure of A'+A
* permc_spec = 3: approximate minimum degree for unsymmetric matrices
*/
permc_spec = 1;
get_perm_c(permc_spec, &A, perm_c);
if ( !(R = (float *) SUPERLU_MALLOC(A.nrow * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for R[].");
if ( !(C = (float *) SUPERLU_MALLOC(A.ncol * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for C[].");
if ( !(ferr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for ferr[].");
if ( !(berr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for berr[].");
/* Solve the system and compute the condition number
and error bounds using dgssvx. */
cgssvx(fact, trans, refact, &A, &iparam, perm_c, perm_r, etree,
equed, R, C, &L, &U, work, lwork, &B, &X, &rpg, &rcond,
ferr, berr, &mem_usage, &info);
printf("cgssvx(): info %d\n", info);
if ( info == 0 || info == n+1 ) {
printf("Recip. pivot growth = %e\n", rpg);
printf("Recip. condition number = %e\n", rcond);
printf("%8s%16s%16s\n", "rhs", "FERR", "BERR");
for (i = 0; i < nrhs; ++i) {
printf("%8d%16e%16e\n", i+1, ferr[i], berr[i]);
}
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions);
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
SUPERLU_FREE (rhsb);
SUPERLU_FREE (rhsx);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
SUPERLU_FREE (ferr);
SUPERLU_FREE (berr);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperMatrix_Store(&X);
if ( lwork >= 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
}
}
void
dgssvx(superlu_options_t *options, SuperMatrix *A, int *perm_c, int *perm_r,
int *etree, char *equed, double *R, double *C,
SuperMatrix *L, SuperMatrix *U, void *work, int lwork,
SuperMatrix *B, SuperMatrix *X, double *recip_pivot_growth,
double *rcond, double *ferr, double *berr,
mem_usage_t *mem_usage, SuperLUStat_t *stat, int *info )
{
/*
* Purpose
* =======
*
* DGSSVX solves the system of linear equations A*X=B or A'*X=B, using
* the LU factorization from dgstrf(). Error bounds on the solution and
* a condition estimate are also provided. It performs the following steps:
*
* 1. If A is stored column-wise (A->Stype = SLU_NC):
*
* 1.1. If options->Equil = YES, scaling factors are computed to
* equilibrate the system:
* options->Trans = NOTRANS:
* diag(R)*A*diag(C) *inv(diag(C))*X = diag(R)*B
* options->Trans = TRANS:
* (diag(R)*A*diag(C))**T *inv(diag(R))*X = diag(C)*B
* options->Trans = CONJ:
* (diag(R)*A*diag(C))**H *inv(diag(R))*X = diag(C)*B
* Whether or not the system will be equilibrated depends on the
* scaling of the matrix A, but if equilibration is used, A is
* overwritten by diag(R)*A*diag(C) and B by diag(R)*B
* (if options->Trans=NOTRANS) or diag(C)*B (if options->Trans
* = TRANS or CONJ).
*
* 1.2. Permute columns of A, forming A*Pc, where Pc is a permutation
* matrix that usually preserves sparsity.
* For more details of this step, see sp_preorder.c.
*
* 1.3. If options->Fact != FACTORED, the LU decomposition is used to
* factor the matrix A (after equilibration if options->Equil = YES)
* as Pr*A*Pc = L*U, with Pr determined by partial pivoting.
*
* 1.4. Compute the reciprocal pivot growth factor.
*
* 1.5. If some U(i,i) = 0, so that U is exactly singular, then the
* routine returns with info = i. Otherwise, the factored form of
* A is used to estimate the condition number of the matrix A. If
* the reciprocal of the condition number is less than machine
* precision, info = A->ncol+1 is returned as a warning, but the
* routine still goes on to solve for X and computes error bounds
* as described below.
*
* 1.6. The system of equations is solved for X using the factored form
* of A.
*
* 1.7. If options->IterRefine != NOREFINE, iterative refinement is
* applied to improve the computed solution matrix and calculate
* error bounds and backward error estimates for it.
*
* 1.8. If equilibration was used, the matrix X is premultiplied by
* diag(C) (if options->Trans = NOTRANS) or diag(R)
* (if options->Trans = TRANS or CONJ) so that it solves the
* original system before equilibration.
*
* 2. If A is stored row-wise (A->Stype = SLU_NR), apply the above algorithm
* to the transpose of A:
*
* 2.1. If options->Equil = YES, scaling factors are computed to
* equilibrate the system:
* options->Trans = NOTRANS:
* diag(R)*A*diag(C) *inv(diag(C))*X = diag(R)*B
* options->Trans = TRANS:
* (diag(R)*A*diag(C))**T *inv(diag(R))*X = diag(C)*B
* options->Trans = CONJ:
* (diag(R)*A*diag(C))**H *inv(diag(R))*X = diag(C)*B
* Whether or not the system will be equilibrated depends on the
* scaling of the matrix A, but if equilibration is used, A' is
* overwritten by diag(R)*A'*diag(C) and B by diag(R)*B
* (if trans='N') or diag(C)*B (if trans = 'T' or 'C').
*
* 2.2. Permute columns of transpose(A) (rows of A),
* forming transpose(A)*Pc, where Pc is a permutation matrix that
* usually preserves sparsity.
* For more details of this step, see sp_preorder.c.
*
* 2.3. If options->Fact != FACTORED, the LU decomposition is used to
* factor the transpose(A) (after equilibration if
* options->Fact = YES) as Pr*transpose(A)*Pc = L*U with the
* permutation Pr determined by partial pivoting.
*
* 2.4. Compute the reciprocal pivot growth factor.
*
* 2.5. If some U(i,i) = 0, so that U is exactly singular, then the
* routine returns with info = i. Otherwise, the factored form
* of transpose(A) is used to estimate the condition number of the
* matrix A. If the reciprocal of the condition number
* is less than machine precision, info = A->nrow+1 is returned as
* a warning, but the routine still goes on to solve for X and
* computes error bounds as described below.
*
* 2.6. The system of equations is solved for X using the factored form
* of transpose(A).
*
* 2.7. If options->IterRefine != NOREFINE, iterative refinement is
* applied to improve the computed solution matrix and calculate
* error bounds and backward error estimates for it.
*
* 2.8. If equilibration was used, the matrix X is premultiplied by
* diag(C) (if options->Trans = NOTRANS) or diag(R)
* (if options->Trans = TRANS or CONJ) so that it solves the
* original system before equilibration.
*
* See supermatrix.h for the definition of 'SuperMatrix' structure.
*
* Arguments
* =========
*
* options (input) superlu_options_t*
* The structure defines the input parameters to control
* how the LU decomposition will be performed and how the
* system will be solved.
*
* A (input/output) SuperMatrix*
* Matrix A in A*X=B, of dimension (A->nrow, A->ncol). The number
* of the linear equations is A->nrow. Currently, the type of A can be:
* Stype = SLU_NC or SLU_NR, Dtype = SLU_D, Mtype = SLU_GE.
* In the future, more general A may be handled.
*
* On entry, If options->Fact = FACTORED and equed is not 'N',
* then A must have been equilibrated by the scaling factors in
* R and/or C.
* On exit, A is not modified if options->Equil = NO, or if
* options->Equil = YES but equed = 'N' on exit.
* Otherwise, if options->Equil = YES and equed is not 'N',
* A is scaled as follows:
* If A->Stype = SLU_NC:
* equed = 'R': A := diag(R) * A
* equed = 'C': A := A * diag(C)
* equed = 'B': A := diag(R) * A * diag(C).
* If A->Stype = SLU_NR:
* equed = 'R': transpose(A) := diag(R) * transpose(A)
* equed = 'C': transpose(A) := transpose(A) * diag(C)
* equed = 'B': transpose(A) := diag(R) * transpose(A) * diag(C).
*
* perm_c (input/output) int*
* If A->Stype = SLU_NC, Column permutation vector of size A->ncol,
* which defines the permutation matrix Pc; perm_c[i] = j means
* column i of A is in position j in A*Pc.
* On exit, perm_c may be overwritten by the product of the input
* perm_c and a permutation that postorders the elimination tree
* of Pc'*A'*A*Pc; perm_c is not changed if the elimination tree
* is already in postorder.
*
* If A->Stype = SLU_NR, column permutation vector of size A->nrow,
* which describes permutation of columns of transpose(A)
* (rows of A) as described above.
*
* perm_r (input/output) int*
* If A->Stype = SLU_NC, row permutation vector of size A->nrow,
* which defines the permutation matrix Pr, and is determined
* by partial pivoting. perm_r[i] = j means row i of A is in
* position j in Pr*A.
*
* If A->Stype = SLU_NR, permutation vector of size A->ncol, which
* determines permutation of rows of transpose(A)
* (columns of A) as described above.
*
* If options->Fact = SamePattern_SameRowPerm, the pivoting routine
* will try to use the input perm_r, unless a certain threshold
* criterion is violated. In that case, perm_r is overwritten by a
* new permutation determined by partial pivoting or diagonal
* threshold pivoting.
* Otherwise, perm_r is output argument.
*
* etree (input/output) int*, dimension (A->ncol)
* Elimination tree of Pc'*A'*A*Pc.
* If options->Fact != FACTORED and options->Fact != DOFACT,
* etree is an input argument, otherwise it is an output argument.
* Note: etree is a vector of parent pointers for a forest whose
* vertices are the integers 0 to A->ncol-1; etree[root]==A->ncol.
*
* equed (input/output) char*
* Specifies the form of equilibration that was done.
* = 'N': No equilibration.
* = 'R': Row equilibration, i.e., A was premultiplied by diag(R).
* = 'C': Column equilibration, i.e., A was postmultiplied by diag(C).
* = 'B': Both row and column equilibration, i.e., A was replaced
* by diag(R)*A*diag(C).
* If options->Fact = FACTORED, equed is an input argument,
* otherwise it is an output argument.
*
* R (input/output) double*, dimension (A->nrow)
* The row scale factors for A or transpose(A).
* If equed = 'R' or 'B', A (if A->Stype = SLU_NC) or transpose(A)
* (if A->Stype = SLU_NR) is multiplied on the left by diag(R).
* If equed = 'N' or 'C', R is not accessed.
* If options->Fact = FACTORED, R is an input argument,
* otherwise, R is output.
* If options->zFact = FACTORED and equed = 'R' or 'B', each element
* of R must be positive.
*
* C (input/output) double*, dimension (A->ncol)
* The column scale factors for A or transpose(A).
* If equed = 'C' or 'B', A (if A->Stype = SLU_NC) or transpose(A)
* (if A->Stype = SLU_NR) is multiplied on the right by diag(C).
* If equed = 'N' or 'R', C is not accessed.
* If options->Fact = FACTORED, C is an input argument,
* otherwise, C is output.
* If options->Fact = FACTORED and equed = 'C' or 'B', each element
* of C must be positive.
*
* L (output) SuperMatrix*
* The factor L from the factorization
* Pr*A*Pc=L*U (if A->Stype SLU_= NC) or
* Pr*transpose(A)*Pc=L*U (if A->Stype = SLU_NR).
* Uses compressed row subscripts storage for supernodes, i.e.,
* L has types: Stype = SLU_SC, Dtype = SLU_D, Mtype = SLU_TRLU.
*
* U (output) SuperMatrix*
* The factor U from the factorization
* Pr*A*Pc=L*U (if A->Stype = SLU_NC) or
* Pr*transpose(A)*Pc=L*U (if A->Stype = SLU_NR).
* Uses column-wise storage scheme, i.e., U has types:
* Stype = SLU_NC, Dtype = SLU_D, Mtype = SLU_TRU.
*
* work (workspace/output) void*, size (lwork) (in bytes)
* User supplied workspace, should be large enough
* to hold data structures for factors L and U.
* On exit, if fact is not 'F', L and U point to this array.
*
* lwork (input) int
* Specifies the size of work array in bytes.
* = 0: allocate space internally by system malloc;
* > 0: use user-supplied work array of length lwork in bytes,
* returns error if space runs out.
* = -1: the routine guesses the amount of space needed without
* performing the factorization, and returns it in
* mem_usage->total_needed; no other side effects.
*
* See argument 'mem_usage' for memory usage statistics.
*
* B (input/output) SuperMatrix*
* B has types: Stype = SLU_DN, Dtype = SLU_D, Mtype = SLU_GE.
* On entry, the right hand side matrix.
* If B->ncol = 0, only LU decomposition is performed, the triangular
* solve is skipped.
* On exit,
* if equed = 'N', B is not modified; otherwise
* if A->Stype = SLU_NC:
* if options->Trans = NOTRANS and equed = 'R' or 'B',
* B is overwritten by diag(R)*B;
* if options->Trans = TRANS or CONJ and equed = 'C' of 'B',
* B is overwritten by diag(C)*B;
* if A->Stype = SLU_NR:
* if options->Trans = NOTRANS and equed = 'C' or 'B',
* B is overwritten by diag(C)*B;
* if options->Trans = TRANS or CONJ and equed = 'R' of 'B',
* B is overwritten by diag(R)*B.
*
* X (output) SuperMatrix*
* X has types: Stype = SLU_DN, Dtype = SLU_D, Mtype = SLU_GE.
* If info = 0 or info = A->ncol+1, X contains the solution matrix
* to the original system of equations. Note that A and B are modified
* on exit if equed is not 'N', and the solution to the equilibrated
* system is inv(diag(C))*X if options->Trans = NOTRANS and
* equed = 'C' or 'B', or inv(diag(R))*X if options->Trans = 'T' or 'C'
* and equed = 'R' or 'B'.
*
* recip_pivot_growth (output) double*
* The reciprocal pivot growth factor max_j( norm(A_j)/norm(U_j) ).
* The infinity norm is used. If recip_pivot_growth is much less
* than 1, the stability of the LU factorization could be poor.
*
* rcond (output) double*
* The estimate of the reciprocal condition number of the matrix A
* after equilibration (if done). If rcond is less than the machine
* precision (in particular, if rcond = 0), the matrix is singular
* to working precision. This condition is indicated by a return
* code of info > 0.
*
* FERR (output) double*, dimension (B->ncol)
* The estimated forward error bound for each solution vector
* X(j) (the j-th column of the solution matrix X).
* If XTRUE is the true solution corresponding to X(j), FERR(j)
* is an estimated upper bound for the magnitude of the largest
* element in (X(j) - XTRUE) divided by the magnitude of the
* largest element in X(j). The estimate is as reliable as
* the estimate for RCOND, and is almost always a slight
* overestimate of the true error.
* If options->IterRefine = NOREFINE, ferr = 1.0.
*
* BERR (output) double*, dimension (B->ncol)
* The componentwise relative backward error of each solution
* vector X(j) (i.e., the smallest relative change in
* any element of A or B that makes X(j) an exact solution).
* If options->IterRefine = NOREFINE, berr = 1.0.
*
* mem_usage (output) mem_usage_t*
* Record the memory usage statistics, consisting of following fields:
* - for_lu (float)
* The amount of space used in bytes for L\U data structures.
* - total_needed (float)
* The amount of space needed in bytes to perform factorization.
* - expansions (int)
* The number of memory expansions during the LU factorization.
*
* stat (output) SuperLUStat_t*
* Record the statistics on runtime and floating-point operation count.
* See util.h for the definition of 'SuperLUStat_t'.
*
* info (output) int*
* = 0: successful exit
* < 0: if info = -i, the i-th argument had an illegal value
* > 0: if info = i, and i is
* <= A->ncol: U(i,i) is exactly zero. The factorization has
* been completed, but the factor U is exactly
* singular, so the solution and error bounds
* could not be computed.
* = A->ncol+1: U is nonsingular, but RCOND is less than machine
* precision, meaning that the matrix is singular to
* working precision. Nevertheless, the solution and
* error bounds are computed because there are a number
* of situations where the computed solution can be more
* accurate than the value of RCOND would suggest.
* > A->ncol+1: number of bytes allocated when memory allocation
* failure occurred, plus A->ncol.
*
*/
DNformat *Bstore, *Xstore;
double *Bmat, *Xmat;
int ldb, ldx, nrhs;
SuperMatrix *AA;/* A in SLU_NC format used by the factorization routine.*/
SuperMatrix AC; /* Matrix postmultiplied by Pc */
int colequ, equil, nofact, notran, rowequ, permc_spec;
trans_t trant;
char norm[1];
int i, j, info1;
double amax, anorm, bignum, smlnum, colcnd, rowcnd, rcmax, rcmin;
int relax, panel_size;
double drop_tol;
double t0; /* temporary time */
double *utime;
/* External functions */
extern double dlangs(char *, SuperMatrix *);
extern double hypre_F90_NAME_LAPACK(dlamch,DLAMCH)(const char *);
Bstore = (DNformat*) B->Store;
Xstore = (DNformat*) X->Store;
Bmat = ( double*) Bstore->nzval;
Xmat = ( double*) Xstore->nzval;
ldb = Bstore->lda;
ldx = Xstore->lda;
nrhs = B->ncol;
*info = 0;
nofact = (options->Fact != FACTORED);
equil = (options->Equil == YES);
notran = (options->Trans == NOTRANS);
if ( nofact ) {
*(unsigned char *)equed = 'N';
rowequ = FALSE;
colequ = FALSE;
} else {
rowequ = superlu_lsame(equed, "R") || superlu_lsame(equed, "B");
colequ = superlu_lsame(equed, "C") || superlu_lsame(equed, "B");
smlnum = hypre_F90_NAME_LAPACK(dlamch,DLAMCH)("Safe minimum");
bignum = 1. / smlnum;
}
#if 0
printf("dgssvx: Fact=%4d, Trans=%4d, equed=%c\n",
options->Fact, options->Trans, *equed);
#endif
/* Test the input parameters */
if (!nofact && options->Fact != DOFACT && options->Fact != SamePattern &&
options->Fact != SamePattern_SameRowPerm &&
!notran && options->Trans != TRANS && options->Trans != CONJ &&
!equil && options->Equil != NO)
*info = -1;
else if ( A->nrow != A->ncol || A->nrow < 0 ||
(A->Stype != SLU_NC && A->Stype != SLU_NR) ||
A->Dtype != SLU_D || A->Mtype != SLU_GE )
*info = -2;
else if (options->Fact == FACTORED &&
!(rowequ || colequ || superlu_lsame(equed, "N")))
*info = -6;
else {
if (rowequ) {
rcmin = bignum;
rcmax = 0.;
for (j = 0; j < A->nrow; ++j) {
rcmin = SUPERLU_MIN(rcmin, R[j]);
rcmax = SUPERLU_MAX(rcmax, R[j]);
}
if (rcmin <= 0.) *info = -7;
else if ( A->nrow > 0)
rowcnd = SUPERLU_MAX(rcmin,smlnum) / SUPERLU_MIN(rcmax,bignum);
else rowcnd = 1.;
}
if (colequ && *info == 0) {
rcmin = bignum;
rcmax = 0.;
for (j = 0; j < A->nrow; ++j) {
rcmin = SUPERLU_MIN(rcmin, C[j]);
rcmax = SUPERLU_MAX(rcmax, C[j]);
}
if (rcmin <= 0.) *info = -8;
else if (A->nrow > 0)
colcnd = SUPERLU_MAX(rcmin,smlnum) / SUPERLU_MIN(rcmax,bignum);
else colcnd = 1.;
}
if (*info == 0) {
if ( lwork < -1 ) *info = -12;
else if ( B->ncol < 0 || Bstore->lda < SUPERLU_MAX(0, A->nrow) ||
B->Stype != SLU_DN || B->Dtype != SLU_D ||
B->Mtype != SLU_GE )
*info = -13;
else if ( X->ncol < 0 || Xstore->lda < SUPERLU_MAX(0, A->nrow) ||
(B->ncol != 0 && B->ncol != X->ncol) ||
X->Stype != SLU_DN ||
X->Dtype != SLU_D || X->Mtype != SLU_GE )
*info = -14;
}
}
if (*info != 0) {
i = -(*info);
superlu_xerbla("dgssvx", &i);
return;
}
/* Initialization for factor parameters */
panel_size = sp_ienv(1);
relax = sp_ienv(2);
drop_tol = 0.0;
utime = stat->utime;
/* Convert A to SLU_NC format when necessary. */
if ( A->Stype == SLU_NR ) {
NRformat *Astore = (NRformat*) A->Store;
AA = (SuperMatrix *) SUPERLU_MALLOC( sizeof(SuperMatrix) );
dCreate_CompCol_Matrix(AA, A->ncol, A->nrow, Astore->nnz,
(double*) Astore->nzval, Astore->colind, Astore->rowptr,
SLU_NC, A->Dtype, A->Mtype);
if ( notran ) { /* Reverse the transpose argument. */
trant = TRANS;
notran = 0;
} else {
trant = NOTRANS;
notran = 1;
}
} else { /* A->Stype == SLU_NC */
trant = options->Trans;
AA = A;
}
if ( nofact && equil ) {
t0 = SuperLU_timer_();
/* Compute row and column scalings to equilibrate the matrix A. */
dgsequ(AA, R, C, &rowcnd, &colcnd, &amax, &info1);
if ( info1 == 0 ) {
/* Equilibrate matrix A. */
dlaqgs(AA, R, C, rowcnd, colcnd, amax, equed);
rowequ = superlu_lsame(equed, "R") || superlu_lsame(equed, "B");
colequ = superlu_lsame(equed, "C") || superlu_lsame(equed, "B");
}
utime[EQUIL] = SuperLU_timer_() - t0;
}
if ( nrhs > 0 ) {
/* Scale the right hand side if equilibration was performed. */
if ( notran ) {
if ( rowequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i) {
Bmat[i + j*ldb] *= R[i];
}
}
} else if ( colequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i) {
Bmat[i + j*ldb] *= C[i];
}
}
}
if ( nofact ) {
t0 = SuperLU_timer_();
/*
* Gnet column permutation vector perm_c[], according to permc_spec:
* permc_spec = NATURAL: natural ordering
* permc_spec = MMD_AT_PLUS_A: minimum degree on structure of A'+A
* permc_spec = MMD_ATA: minimum degree on structure of A'*A
* permc_spec = COLAMD: approximate minimum degree column ordering
* permc_spec = MY_PERMC: the ordering already supplied in perm_c[]
*/
permc_spec = options->ColPerm;
if ( permc_spec != MY_PERMC && options->Fact == DOFACT )
get_perm_c(permc_spec, AA, perm_c);
utime[COLPERM] = SuperLU_timer_() - t0;
t0 = SuperLU_timer_();
sp_preorder(options, AA, perm_c, etree, &AC);
utime[ETREE] = SuperLU_timer_() - t0;
/* printf("Factor PA = LU ... relax %d\tw %d\tmaxsuper %d\trowblk %d\n",
relax, panel_size, sp_ienv(3), sp_ienv(4));
fflush(stdout); */
/* Compute the LU factorization of A*Pc. */
t0 = SuperLU_timer_();
dgstrf(options, &AC, drop_tol, relax, panel_size,
etree, work, lwork, perm_c, perm_r, L, U, stat, info);
utime[FACT] = SuperLU_timer_() - t0;
if ( lwork == -1 ) {
mem_usage->total_needed = *info - A->ncol;
return;
}
}
if ( options->PivotGrowth ) {
if ( *info > 0 ) {
if ( *info <= A->ncol ) {
/* Compute the reciprocal pivot growth factor of the leading
rank-deficient *info columns of A. */
*recip_pivot_growth = dPivotGrowth(*info, AA, perm_c, L, U);
}
return;
}
/* Compute the reciprocal pivot growth factor *recip_pivot_growth. */
*recip_pivot_growth = dPivotGrowth(A->ncol, AA, perm_c, L, U);
}
if ( options->ConditionNumber ) {
/* Estimate the reciprocal of the condition number of A. */
t0 = SuperLU_timer_();
if ( notran ) {
*(unsigned char *)norm = '1';
} else {
*(unsigned char *)norm = 'I';
}
anorm = dlangs(norm, AA);
dgscon(norm, L, U, anorm, rcond, stat, info);
utime[RCOND] = SuperLU_timer_() - t0;
}
if ( nrhs > 0 ) {
/* Compute the solution matrix X. */
for (j = 0; j < nrhs; j++) /* Save a copy of the right hand sides */
for (i = 0; i < B->nrow; i++)
Xmat[i + j*ldx] = Bmat[i + j*ldb];
t0 = SuperLU_timer_();
dgstrs (trant, L, U, perm_c, perm_r, X, stat, info);
utime[SOLVE] = SuperLU_timer_() - t0;
/* Use iterative refinement to improve the computed solution and compute
error bounds and backward error estimates for it. */
t0 = SuperLU_timer_();
if ( options->IterRefine != NOREFINE ) {
dgsrfs(trant, AA, L, U, perm_c, perm_r, equed, R, C, B,
X, ferr, berr, stat, info);
} else {
for (j = 0; j < nrhs; ++j) ferr[j] = berr[j] = 1.0;
}
utime[REFINE] = SuperLU_timer_() - t0;
/* Transform the solution matrix X to a solution of the original system. */
if ( notran ) {
if ( colequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i) {
Xmat[i + j*ldx] *= C[i];
}
}
} else if ( rowequ ) {
for (j = 0; j < nrhs; ++j)
for (i = 0; i < A->nrow; ++i) {
Xmat[i + j*ldx] *= R[i];
}
}
} /* end if nrhs > 0 */
if ( options->ConditionNumber ) {
/* Set INFO = A->ncol+1 if the matrix is singular to working precision. */
if (*rcond < hypre_F90_NAME_LAPACK(dlamch,DLAMCH)("E")) *info=A->ncol+1;
}
if ( nofact ) {
dQuerySpace(L, U, mem_usage);
Destroy_CompCol_Permuted(&AC);
}
if ( A->Stype == SLU_NR ) {
Destroy_SuperMatrix_Store(AA);
SUPERLU_FREE(AA);
}
}
/* Here is a driver inspired by A. Sheffer's "cow flattener". */
static NLboolean __nlFactorize_SUPERLU(__NLContext *context, NLint *permutation) {
/* OpenNL Context */
__NLSparseMatrix* M = (context->least_squares)? &context->MtM: &context->M;
NLuint n = context->n;
NLuint nnz = __nlSparseMatrixNNZ(M); /* number of non-zero coeffs */
/* Compressed Row Storage matrix representation */
NLint *xa = __NL_NEW_ARRAY(NLint, n+1);
NLfloat *rhs = __NL_NEW_ARRAY(NLfloat, n);
NLfloat *a = __NL_NEW_ARRAY(NLfloat, nnz);
NLint *asub = __NL_NEW_ARRAY(NLint, nnz);
NLint *etree = __NL_NEW_ARRAY(NLint, n);
/* SuperLU variables */
SuperMatrix At, AtP;
NLint info, panel_size, relax;
superlu_options_t options;
/* Temporary variables */
NLuint i, jj, count;
__nl_assert(!(M->storage & __NL_SYMMETRIC));
__nl_assert(M->storage & __NL_ROWS);
__nl_assert(M->m == M->n);
/* Convert M to compressed column format */
for(i=0, count=0; i<n; i++) {
__NLRowColumn *Ri = M->row + i;
xa[i] = count;
for(jj=0; jj<Ri->size; jj++, count++) {
a[count] = Ri->coeff[jj].value;
asub[count] = Ri->coeff[jj].index;
}
}
xa[n] = nnz;
/* Free M, don't need it anymore at this point */
__nlSparseMatrixClear(M);
/* Create superlu A matrix transposed */
sCreate_CompCol_Matrix(
&At, n, n, nnz, a, asub, xa,
SLU_NC, /* Colum wise, no supernode */
SLU_S, /* floats */
SLU_GE /* general storage */
);
/* Set superlu options */
set_default_options(&options);
options.ColPerm = MY_PERMC;
options.Fact = DOFACT;
StatInit(&(context->slu.stat));
panel_size = sp_ienv(1); /* sp_ienv give us the defaults */
relax = sp_ienv(2);
/* Compute permutation and permuted matrix */
context->slu.perm_r = __NL_NEW_ARRAY(NLint, n);
context->slu.perm_c = __NL_NEW_ARRAY(NLint, n);
if ((permutation == NULL) || (*permutation == -1)) {
get_perm_c(3, &At, context->slu.perm_c);
if (permutation)
memcpy(permutation, context->slu.perm_c, sizeof(NLint)*n);
}
else
memcpy(context->slu.perm_c, permutation, sizeof(NLint)*n);
sp_preorder(&options, &At, context->slu.perm_c, etree, &AtP);
/* Decompose into L and U */
sgstrf(&options, &AtP, relax, panel_size,
etree, NULL, 0, context->slu.perm_c, context->slu.perm_r,
&(context->slu.L), &(context->slu.U), &(context->slu.stat), &info);
/* Cleanup */
Destroy_SuperMatrix_Store(&At);
Destroy_CompCol_Permuted(&AtP);
__NL_DELETE_ARRAY(etree);
__NL_DELETE_ARRAY(xa);
__NL_DELETE_ARRAY(rhs);
__NL_DELETE_ARRAY(a);
__NL_DELETE_ARRAY(asub);
context->slu.alloc_slu = NL_TRUE;
return (info == 0);
}
void
c_fortran_zgssv_(int *iopt, int *n, int *nnz, int *nrhs,
doublecomplex *values, int *rowind, int *colptr,
doublecomplex *b, int *ldb,
fptr *f_factors, /* a handle containing the address
pointing to the factored matrices */
int *info)
{
/*
* This routine can be called from Fortran.
*
* iopt (input) int
* Specifies the operation:
* = 1, performs LU decomposition for the first time
* = 2, performs triangular solve
* = 3, free all the storage in the end
*
* f_factors (input/output) fptr*
* If iopt == 1, it is an output and contains the pointer pointing to
* the structure of the factored matrices.
* Otherwise, it it an input.
*
*/
SuperMatrix A, AC, B;
SuperMatrix *L, *U;
int *perm_r; /* row permutations from partial pivoting */
int *perm_c; /* column permutation vector */
int *etree; /* column elimination tree */
SCformat *Lstore;
NCformat *Ustore;
int i, panel_size, permc_spec, relax;
trans_t trans;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
factors_t *LUfactors;
trans = TRANS;
if ( *iopt == 1 ) { /* LU decomposition */
/* Set the default input options. */
set_default_options(&options);
/* Initialize the statistics variables. */
StatInit(&stat);
/* Adjust to 0-based indexing */
for (i = 0; i < *nnz; ++i) --rowind[i];
for (i = 0; i <= *n; ++i) --colptr[i];
zCreate_CompCol_Matrix(&A, *n, *n, *nnz, values, rowind, colptr,
SLU_NC, SLU_Z, SLU_GE);
L = (SuperMatrix *) SUPERLU_MALLOC( sizeof(SuperMatrix) );
U = (SuperMatrix *) SUPERLU_MALLOC( sizeof(SuperMatrix) );
if ( !(perm_r = intMalloc(*n)) ) ABORT("Malloc fails for perm_r[].");
if ( !(perm_c = intMalloc(*n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(etree = intMalloc(*n)) ) ABORT("Malloc fails for etree[].");
/*
* Get column permutation vector perm_c[], according to permc_spec:
* permc_spec = 0: natural ordering
* permc_spec = 1: minimum degree on structure of A'*A
* permc_spec = 2: minimum degree on structure of A'+A
* permc_spec = 3: approximate minimum degree for unsymmetric matrices
*/
permc_spec = options.ColPerm;
get_perm_c(permc_spec, &A, perm_c);
sp_preorder(&options, &A, perm_c, etree, &AC);
panel_size = sp_ienv(1);
relax = sp_ienv(2);
zgstrf(&options, &AC, relax, panel_size, etree,
NULL, 0, perm_c, perm_r, L, U, &stat, info);
if ( *info == 0 ) {
Lstore = (SCformat *) L->Store;
Ustore = (NCformat *) U->Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz);
zQuerySpace(L, U, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
} else {
printf("zgstrf() error returns INFO= %d\n", *info);
if ( *info <= *n ) { /* factorization completes */
zQuerySpace(L, U, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
}
}
/* Restore to 1-based indexing */
for (i = 0; i < *nnz; ++i) ++rowind[i];
for (i = 0; i <= *n; ++i) ++colptr[i];
/* Save the LU factors in the factors handle */
LUfactors = (factors_t*) SUPERLU_MALLOC(sizeof(factors_t));
LUfactors->L = L;
LUfactors->U = U;
LUfactors->perm_c = perm_c;
LUfactors->perm_r = perm_r;
*f_factors = (fptr) LUfactors;
/* Free un-wanted storage */
SUPERLU_FREE(etree);
Destroy_SuperMatrix_Store(&A);
Destroy_CompCol_Permuted(&AC);
StatFree(&stat);
} else if ( *iopt == 2 ) { /* Triangular solve */
/* Initialize the statistics variables. */
StatInit(&stat);
/* Extract the LU factors in the factors handle */
LUfactors = (factors_t*) *f_factors;
L = LUfactors->L;
U = LUfactors->U;
perm_c = LUfactors->perm_c;
perm_r = LUfactors->perm_r;
zCreate_Dense_Matrix(&B, *n, *nrhs, b, *ldb, SLU_DN, SLU_Z, SLU_GE);
/* Solve the system A*X=B, overwriting B with X. */
zgstrs (trans, L, U, perm_c, perm_r, &B, &stat, info);
Destroy_SuperMatrix_Store(&B);
StatFree(&stat);
} else if ( *iopt == 3 ) { /* Free storage */
/* Free the LU factors in the factors handle */
LUfactors = (factors_t*) *f_factors;
SUPERLU_FREE (LUfactors->perm_r);
SUPERLU_FREE (LUfactors->perm_c);
Destroy_SuperNode_Matrix(LUfactors->L);
Destroy_CompCol_Matrix(LUfactors->U);
SUPERLU_FREE (LUfactors->L);
SUPERLU_FREE (LUfactors->U);
SUPERLU_FREE (LUfactors);
} else {
fprintf(stderr,"Invalid iopt=%d passed to c_fortran_zgssv()\n",*iopt);
exit(-1);
}
}
NLboolean nlFactorize_SUPERLU() {
/* OpenNL Context */
NLSparseMatrix* M = &(nlCurrentContext->M) ;
NLuint n = nlCurrentContext->n ;
NLuint nnz = nlSparseMatrixNNZ(M) ; /* Number of Non-Zero coeffs */
superlu_context* context = (superlu_context*)(nlCurrentContext->direct_solver_context) ;
if(context == NULL) {
nlCurrentContext->direct_solver_context = malloc(sizeof(superlu_context)) ;
context = (superlu_context*)(nlCurrentContext->direct_solver_context) ;
}
/* SUPERLU variables */
NLint info ;
SuperMatrix A, AC ;
/* Temporary variables */
NLRowColumn* Ci = NULL ;
NLuint i,j,count ;
/* Sanity checks */
nl_assert(!(M->storage & NL_MATRIX_STORE_SYMMETRIC)) ;
nl_assert(M->storage & NL_MATRIX_STORE_ROWS) ;
nl_assert(M->m == M->n) ;
set_default_options(&(context->options)) ;
switch(nlCurrentContext->solver) {
case NL_SUPERLU_EXT: {
context->options.ColPerm = NATURAL ;
} break ;
case NL_PERM_SUPERLU_EXT: {
context->options.ColPerm = COLAMD ;
} break ;
case NL_SYMMETRIC_SUPERLU_EXT: {
context->options.ColPerm = MMD_AT_PLUS_A ;
context->options.SymmetricMode = YES ;
} break ;
default: {
nl_assert_not_reached ;
} break ;
}
StatInit(&(context->stat)) ;
/*
* Step 1: convert matrix M into SUPERLU compressed column representation
* ----------------------------------------------------------------------
*/
NLint* xa = NL_NEW_ARRAY(NLint, n+1) ;
NLdouble* a = NL_NEW_ARRAY(NLdouble, nnz) ;
NLint* asub = NL_NEW_ARRAY(NLint, nnz) ;
count = 0 ;
for(i = 0; i < n; i++) {
Ci = &(M->row[i]) ;
xa[i] = count ;
for(j = 0; j < Ci->size; j++) {
a[count] = Ci->coeff[j].value ;
asub[count] = Ci->coeff[j].index ;
count++ ;
}
}
xa[n] = nnz ;
dCreate_CompCol_Matrix(
&A, n, n, nnz, a, asub, xa,
SLU_NR, /* Row wise */
SLU_D, /* doubles */
SLU_GE /* general storage */
);
/*
* Step 2: factorize matrix
* ------------------------
*/
context->perm_c = NL_NEW_ARRAY(NLint, n) ;
context->perm_r = NL_NEW_ARRAY(NLint, n) ;
NLint* etree = NL_NEW_ARRAY(NLint, n) ;
get_perm_c(context->options.ColPerm, &A, context->perm_c) ;
sp_preorder(&(context->options), &A, context->perm_c, etree, &AC) ;
int panel_size = sp_ienv(1) ;
int relax = sp_ienv(2) ;
dgstrf(&(context->options),
&AC,
relax,
panel_size,
etree,
NULL,
0,
context->perm_c,
context->perm_r,
&(context->L),
&(context->U),
&(context->stat),
&info) ;
/*
* Step 3: cleanup
* ---------------
*/
NL_DELETE_ARRAY(xa) ;
NL_DELETE_ARRAY(a) ;
NL_DELETE_ARRAY(asub) ;
NL_DELETE_ARRAY(etree) ;
Destroy_SuperMatrix_Store(&A);
Destroy_CompCol_Permuted(&AC);
StatFree(&(context->stat));
return NL_TRUE ;
}
