/*! \brief
*
* <pre>
* Purpose:
* ========
* Performs numeric block updates (sup-col) in topological order.
* It features: col-col, 2cols-col, 3cols-col, and sup-col updates.
* Special processing on the supernodal portion of L\U[*,j]
* Return value: 0 - successful return
* > 0 - number of bytes allocated when run out of space
* </pre>
*/
int
dcolumn_bmod (
const int jcol, /* in */
const int nseg, /* in */
double *dense, /* in */
double *tempv, /* working array */
int *segrep, /* in */
int *repfnz, /* in */
int fpanelc, /* in -- first column in the current panel */
GlobalLU_t *Glu, /* modified */
SuperLUStat_t *stat /* output */
)
{
#ifdef _CRAY
_fcd ftcs1 = _cptofcd("L", strlen("L")),
ftcs2 = _cptofcd("N", strlen("N")),
ftcs3 = _cptofcd("U", strlen("U"));
#endif
int incx = 1, incy = 1;
double alpha, beta;
/* krep = representative of current k-th supernode
* fsupc = first supernodal column
* nsupc = no of columns in supernode
* nsupr = no of rows in supernode (used as leading dimension)
* luptr = location of supernodal LU-block in storage
* kfnz = first nonz in the k-th supernodal segment
* no_zeros = no of leading zeros in a supernodal U-segment
*/
double ukj, ukj1, ukj2;
int luptr, luptr1, luptr2;
int fsupc, nsupc, nsupr, segsze;
int nrow; /* No of rows in the matrix of matrix-vector */
int jcolp1, jsupno, k, ksub, krep, krep_ind, ksupno;
register int lptr, kfnz, isub, irow, i;
register int no_zeros, new_next;
int ufirst, nextlu;
int fst_col; /* First column within small LU update */
int d_fsupc; /* Distance between the first column of the current
panel and the first column of the current snode. */
int *xsup, *supno;
int *lsub, *xlsub;
double *lusup;
int *xlusup;
int nzlumax;
double *tempv1;
double zero = 0.0;
double one = 1.0;
double none = -1.0;
int mem_error;
flops_t *ops = stat->ops;
xsup = Glu->xsup;
supno = Glu->supno;
lsub = Glu->lsub;
xlsub = Glu->xlsub;
lusup = (double *) Glu->lusup;
xlusup = Glu->xlusup;
nzlumax = Glu->nzlumax;
jcolp1 = jcol + 1;
jsupno = supno[jcol];
/*
* For each nonz supernode segment of U[*,j] in topological order
*/
k = nseg - 1;
for (ksub = 0; ksub < nseg; ksub++) {
krep = segrep[k];
k--;
ksupno = supno[krep];
if ( jsupno != ksupno ) { /* Outside the rectangular supernode */
fsupc = xsup[ksupno];
fst_col = SUPERLU_MAX ( fsupc, fpanelc );
/* Distance from the current supernode to the current panel;
d_fsupc=0 if fsupc > fpanelc. */
d_fsupc = fst_col - fsupc;
luptr = xlusup[fst_col] + d_fsupc;
lptr = xlsub[fsupc] + d_fsupc;
kfnz = repfnz[krep];
kfnz = SUPERLU_MAX ( kfnz, fpanelc );
segsze = krep - kfnz + 1;
nsupc = krep - fst_col + 1;
nsupr = xlsub[fsupc+1] - xlsub[fsupc]; /* Leading dimension */
nrow = nsupr - d_fsupc - nsupc;
krep_ind = lptr + nsupc - 1;
ops[TRSV] += segsze * (segsze - 1);
ops[GEMV] += 2 * nrow * segsze;
/*
* Case 1: Update U-segment of size 1 -- col-col update
*/
if ( segsze == 1 ) {
ukj = dense[lsub[krep_ind]];
luptr += nsupr*(nsupc-1) + nsupc;
for (i = lptr + nsupc; i < xlsub[fsupc+1]; ++i) {
irow = lsub[i];
dense[irow] -= ukj*lusup[luptr];
luptr++;
}
} else if ( segsze <= 3 ) {
ukj = dense[lsub[krep_ind]];
luptr += nsupr*(nsupc-1) + nsupc-1;
ukj1 = dense[lsub[krep_ind - 1]];
luptr1 = luptr - nsupr;
if ( segsze == 2 ) { /* Case 2: 2cols-col update */
ukj -= ukj1 * lusup[luptr1];
dense[lsub[krep_ind]] = ukj;
for (i = lptr + nsupc; i < xlsub[fsupc+1]; ++i) {
irow = lsub[i];
luptr++;
luptr1++;
dense[irow] -= ( ukj*lusup[luptr]
+ ukj1*lusup[luptr1] );
}
} else { /* Case 3: 3cols-col update */
ukj2 = dense[lsub[krep_ind - 2]];
luptr2 = luptr1 - nsupr;
ukj1 -= ukj2 * lusup[luptr2-1];
ukj = ukj - ukj1*lusup[luptr1] - ukj2*lusup[luptr2];
dense[lsub[krep_ind]] = ukj;
dense[lsub[krep_ind-1]] = ukj1;
for (i = lptr + nsupc; i < xlsub[fsupc+1]; ++i) {
irow = lsub[i];
luptr++;
luptr1++;
luptr2++;
dense[irow] -= ( ukj*lusup[luptr]
+ ukj1*lusup[luptr1] + ukj2*lusup[luptr2] );
}
}
} else {
/*
* Case: sup-col update
* Perform a triangular solve and block update,
* then scatter the result of sup-col update to dense
*/
no_zeros = kfnz - fst_col;
/* Copy U[*,j] segment from dense[*] to tempv[*] */
isub = lptr + no_zeros;
for (i = 0; i < segsze; i++) {
irow = lsub[isub];
tempv[i] = dense[irow];
++isub;
}
/* Dense triangular solve -- start effective triangle */
luptr += nsupr * no_zeros + no_zeros;
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
STRSV( ftcs1, ftcs2, ftcs3, &segsze, &lusup[luptr],
&nsupr, tempv, &incx );
#else
dtrsv_( "L", "N", "U", &segsze, &lusup[luptr],
&nsupr, tempv, &incx );
#endif
luptr += segsze; /* Dense matrix-vector */
tempv1 = &tempv[segsze];
alpha = one;
beta = zero;
#ifdef _CRAY
SGEMV( ftcs2, &nrow, &segsze, &alpha, &lusup[luptr],
&nsupr, tempv, &incx, &beta, tempv1, &incy );
#else
dgemv_( "N", &nrow, &segsze, &alpha, &lusup[luptr],
&nsupr, tempv, &incx, &beta, tempv1, &incy );
#endif
#else
dlsolve ( nsupr, segsze, &lusup[luptr], tempv );
luptr += segsze; /* Dense matrix-vector */
tempv1 = &tempv[segsze];
dmatvec (nsupr, nrow , segsze, &lusup[luptr], tempv, tempv1);
#endif
/* Scatter tempv[] into SPA dense[] as a temporary storage */
isub = lptr + no_zeros;
for (i = 0; i < segsze; i++) {
irow = lsub[isub];
dense[irow] = tempv[i];
tempv[i] = zero;
++isub;
}
/* Scatter tempv1[] into SPA dense[] */
for (i = 0; i < nrow; i++) {
irow = lsub[isub];
dense[irow] -= tempv1[i];
tempv1[i] = zero;
++isub;
}
}
} /* if jsupno ... */
} /* for each segment... */
/*
* Process the supernodal portion of L\U[*,j]
*/
nextlu = xlusup[jcol];
fsupc = xsup[jsupno];
/* Copy the SPA dense into L\U[*,j] */
new_next = nextlu + xlsub[fsupc+1] - xlsub[fsupc];
while ( new_next > nzlumax ) {
if (mem_error = dLUMemXpand(jcol, nextlu, LUSUP, &nzlumax, Glu))
return (mem_error);
lusup = (double *) Glu->lusup;
lsub = Glu->lsub;
}
for (isub = xlsub[fsupc]; isub < xlsub[fsupc+1]; isub++) {
irow = lsub[isub];
lusup[nextlu] = dense[irow];
dense[irow] = zero;
++nextlu;
}
xlusup[jcolp1] = nextlu; /* Close L\U[*,jcol] */
/* For more updates within the panel (also within the current supernode),
* should start from the first column of the panel, or the first column
* of the supernode, whichever is bigger. There are 2 cases:
* 1) fsupc < fpanelc, then fst_col := fpanelc
* 2) fsupc >= fpanelc, then fst_col := fsupc
*/
fst_col = SUPERLU_MAX ( fsupc, fpanelc );
if ( fst_col < jcol ) {
/* Distance between the current supernode and the current panel.
d_fsupc=0 if fsupc >= fpanelc. */
d_fsupc = fst_col - fsupc;
lptr = xlsub[fsupc] + d_fsupc;
luptr = xlusup[fst_col] + d_fsupc;
nsupr = xlsub[fsupc+1] - xlsub[fsupc]; /* Leading dimension */
nsupc = jcol - fst_col; /* Excluding jcol */
nrow = nsupr - d_fsupc - nsupc;
/* Points to the beginning of jcol in snode L\U(jsupno) */
ufirst = xlusup[jcol] + d_fsupc;
ops[TRSV] += nsupc * (nsupc - 1);
ops[GEMV] += 2 * nrow * nsupc;
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
STRSV( ftcs1, ftcs2, ftcs3, &nsupc, &lusup[luptr],
&nsupr, &lusup[ufirst], &incx );
#else
dtrsv_( "L", "N", "U", &nsupc, &lusup[luptr],
&nsupr, &lusup[ufirst], &incx );
#endif
alpha = none; beta = one; /* y := beta*y + alpha*A*x */
#ifdef _CRAY
SGEMV( ftcs2, &nrow, &nsupc, &alpha, &lusup[luptr+nsupc], &nsupr,
&lusup[ufirst], &incx, &beta, &lusup[ufirst+nsupc], &incy );
#else
dgemv_( "N", &nrow, &nsupc, &alpha, &lusup[luptr+nsupc], &nsupr,
&lusup[ufirst], &incx, &beta, &lusup[ufirst+nsupc], &incy );
#endif
#else
dlsolve ( nsupr, nsupc, &lusup[luptr], &lusup[ufirst] );
dmatvec ( nsupr, nrow, nsupc, &lusup[luptr+nsupc],
&lusup[ufirst], tempv );
/* Copy updates from tempv[*] into lusup[*] */
isub = ufirst + nsupc;
for (i = 0; i < nrow; i++) {
lusup[isub] -= tempv[i];
tempv[i] = 0.0;
++isub;
}
#endif
} /* if fst_col < jcol ... */
return 0;
}
void
dpanel_bmod (
const int m, /* in - number of rows in the matrix */
const int w, /* in */
const int jcol, /* in */
const int nseg, /* in */
double *dense, /* out, of size n by w */
double *tempv, /* working array */
int *segrep, /* in */
int *repfnz, /* in, of size n by w */
GlobalLU_t *Glu, /* modified */
SuperLUStat_t *stat /* output */
)
{
/*
* Purpose
* =======
*
* Performs numeric block updates (sup-panel) in topological order.
* It features: col-col, 2cols-col, 3cols-col, and sup-col updates.
* Special processing on the supernodal portion of L\U[*,j]
*
* Before entering this routine, the original nonzeros in the panel
* were already copied into the spa[m,w].
*
* Updated/Output parameters-
* dense[0:m-1,w]: L[*,j:j+w-1] and U[*,j:j+w-1] are returned
* collectively in the m-by-w vector dense[*].
*
*/
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
_fcd ftcs1 = _cptofcd("L", strlen("L")),
ftcs2 = _cptofcd("N", strlen("N")),
ftcs3 = _cptofcd("U", strlen("U"));
#endif
int incx = 1, incy = 1;
double alpha, beta;
#endif
register int k, ksub;
int fsupc, nsupc, nsupr, nrow;
int krep, krep_ind;
double ukj, ukj1, ukj2;
int luptr, luptr1, luptr2;
int segsze;
int block_nrow; /* no of rows in a block row */
register int lptr; /* Points to the row subscripts of a supernode */
int kfnz, irow, no_zeros;
register int isub, isub1, i;
register int jj; /* Index through each column in the panel */
int *xsup, *supno;
int *lsub, *xlsub;
double *lusup;
int *xlusup;
int *repfnz_col; /* repfnz[] for a column in the panel */
double *dense_col; /* dense[] for a column in the panel */
double *tempv1; /* Used in 1-D update */
double *TriTmp, *MatvecTmp; /* used in 2-D update */
double zero = 0.0;
double one = 1.0;
register int ldaTmp;
register int r_ind, r_hi;
static int first = 1, maxsuper, rowblk, colblk;
flops_t *ops = stat->ops;
xsup = Glu->xsup;
supno = Glu->supno;
lsub = Glu->lsub;
xlsub = Glu->xlsub;
lusup = Glu->lusup;
xlusup = Glu->xlusup;
if ( first ) {
maxsuper = sp_ienv(3);
rowblk = sp_ienv(4);
colblk = sp_ienv(5);
first = 0;
}
ldaTmp = maxsuper + rowblk;
/*
* For each nonz supernode segment of U[*,j] in topological order
*/
k = nseg - 1;
for (ksub = 0; ksub < nseg; ksub++) { /* for each updating supernode */
/* krep = representative of current k-th supernode
* fsupc = first supernodal column
* nsupc = no of columns in a supernode
* nsupr = no of rows in a supernode
*/
krep = segrep[k--];
fsupc = xsup[supno[krep]];
nsupc = krep - fsupc + 1;
nsupr = xlsub[fsupc+1] - xlsub[fsupc];
nrow = nsupr - nsupc;
lptr = xlsub[fsupc];
krep_ind = lptr + nsupc - 1;
repfnz_col = repfnz;
dense_col = dense;
if ( nsupc >= colblk && nrow > rowblk ) { /* 2-D block update */
TriTmp = tempv;
/* Sequence through each column in panel -- triangular solves */
for (jj = jcol; jj < jcol + w; jj++,
repfnz_col += m, dense_col += m, TriTmp += ldaTmp ) {
kfnz = repfnz_col[krep];
if ( kfnz == EMPTY ) continue; /* Skip any zero segment */
segsze = krep - kfnz + 1;
luptr = xlusup[fsupc];
ops[TRSV] += segsze * (segsze - 1);
ops[GEMV] += 2 * nrow * segsze;
/* Case 1: Update U-segment of size 1 -- col-col update */
if ( segsze == 1 ) {
ukj = dense_col[lsub[krep_ind]];
luptr += nsupr*(nsupc-1) + nsupc;
for (i = lptr + nsupc; i < xlsub[fsupc+1]; i++) {
irow = lsub[i];
dense_col[irow] -= ukj * lusup[luptr];
++luptr;
}
} else if ( segsze <= 3 ) {
ukj = dense_col[lsub[krep_ind]];
ukj1 = dense_col[lsub[krep_ind - 1]];
luptr += nsupr*(nsupc-1) + nsupc-1;
luptr1 = luptr - nsupr;
if ( segsze == 2 ) {
ukj -= ukj1 * lusup[luptr1];
dense_col[lsub[krep_ind]] = ukj;
for (i = lptr + nsupc; i < xlsub[fsupc+1]; ++i) {
irow = lsub[i];
luptr++; luptr1++;
dense_col[irow] -= (ukj*lusup[luptr]
+ ukj1*lusup[luptr1]);
}
} else {
ukj2 = dense_col[lsub[krep_ind - 2]];
luptr2 = luptr1 - nsupr;
ukj1 -= ukj2 * lusup[luptr2-1];
ukj = ukj - ukj1*lusup[luptr1] - ukj2*lusup[luptr2];
dense_col[lsub[krep_ind]] = ukj;
dense_col[lsub[krep_ind-1]] = ukj1;
for (i = lptr + nsupc; i < xlsub[fsupc+1]; ++i) {
irow = lsub[i];
luptr++; luptr1++; luptr2++;
dense_col[irow] -= ( ukj*lusup[luptr]
+ ukj1*lusup[luptr1] + ukj2*lusup[luptr2] );
}
}
} else { /* segsze >= 4 */
/* Copy U[*,j] segment from dense[*] to TriTmp[*], which
holds the result of triangular solves. */
no_zeros = kfnz - fsupc;
isub = lptr + no_zeros;
for (i = 0; i < segsze; ++i) {
irow = lsub[isub];
TriTmp[i] = dense_col[irow]; /* Gather */
++isub;
}
/* start effective triangle */
luptr += nsupr * no_zeros + no_zeros;
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
STRSV( ftcs1, ftcs2, ftcs3, &segsze, &lusup[luptr],
&nsupr, TriTmp, &incx );
#else
dtrsv_( "L", "N", "U", &segsze, &lusup[luptr],
&nsupr, TriTmp, &incx );
#endif
#else
dlsolve ( nsupr, segsze, &lusup[luptr], TriTmp );
#endif
} /* else ... */
} /* for jj ... end tri-solves */
/* Block row updates; push all the way into dense[*] block */
for ( r_ind = 0; r_ind < nrow; r_ind += rowblk ) {
r_hi = SUPERLU_MIN(nrow, r_ind + rowblk);
block_nrow = SUPERLU_MIN(rowblk, r_hi - r_ind);
luptr = xlusup[fsupc] + nsupc + r_ind;
isub1 = lptr + nsupc + r_ind;
repfnz_col = repfnz;
TriTmp = tempv;
dense_col = dense;
/* Sequence through each column in panel -- matrix-vector */
for (jj = jcol; jj < jcol + w; jj++,
repfnz_col += m, dense_col += m, TriTmp += ldaTmp) {
kfnz = repfnz_col[krep];
if ( kfnz == EMPTY ) continue; /* Skip any zero segment */
segsze = krep - kfnz + 1;
if ( segsze <= 3 ) continue; /* skip unrolled cases */
/* Perform a block update, and scatter the result of
matrix-vector to dense[]. */
no_zeros = kfnz - fsupc;
luptr1 = luptr + nsupr * no_zeros;
MatvecTmp = &TriTmp[maxsuper];
#ifdef USE_VENDOR_BLAS
alpha = one;
beta = zero;
#ifdef _CRAY
SGEMV(ftcs2, &block_nrow, &segsze, &alpha, &lusup[luptr1],
&nsupr, TriTmp, &incx, &beta, MatvecTmp, &incy);
#else
dgemv_("N", &block_nrow, &segsze, &alpha, &lusup[luptr1],
&nsupr, TriTmp, &incx, &beta, MatvecTmp, &incy);
#endif
#else
dmatvec(nsupr, block_nrow, segsze, &lusup[luptr1],
TriTmp, MatvecTmp);
#endif
/* Scatter MatvecTmp[*] into SPA dense[*] temporarily
* such that MatvecTmp[*] can be re-used for the
* the next blok row update. dense[] will be copied into
* global store after the whole panel has been finished.
*/
isub = isub1;
for (i = 0; i < block_nrow; i++) {
irow = lsub[isub];
dense_col[irow] -= MatvecTmp[i];
MatvecTmp[i] = zero;
++isub;
}
} /* for jj ... */
} /* for each block row ... */
/* Scatter the triangular solves into SPA dense[*] */
repfnz_col = repfnz;
TriTmp = tempv;
dense_col = dense;
for (jj = jcol; jj < jcol + w; jj++,
repfnz_col += m, dense_col += m, TriTmp += ldaTmp) {
kfnz = repfnz_col[krep];
if ( kfnz == EMPTY ) continue; /* Skip any zero segment */
segsze = krep - kfnz + 1;
if ( segsze <= 3 ) continue; /* skip unrolled cases */
no_zeros = kfnz - fsupc;
isub = lptr + no_zeros;
for (i = 0; i < segsze; i++) {
irow = lsub[isub];
dense_col[irow] = TriTmp[i];
TriTmp[i] = zero;
++isub;
}
} /* for jj ... */
} else { /* 1-D block modification */
/* Sequence through each column in the panel */
for (jj = jcol; jj < jcol + w; jj++,
repfnz_col += m, dense_col += m) {
kfnz = repfnz_col[krep];
if ( kfnz == EMPTY ) continue; /* Skip any zero segment */
segsze = krep - kfnz + 1;
luptr = xlusup[fsupc];
ops[TRSV] += segsze * (segsze - 1);
ops[GEMV] += 2 * nrow * segsze;
/* Case 1: Update U-segment of size 1 -- col-col update */
if ( segsze == 1 ) {
ukj = dense_col[lsub[krep_ind]];
luptr += nsupr*(nsupc-1) + nsupc;
for (i = lptr + nsupc; i < xlsub[fsupc+1]; i++) {
irow = lsub[i];
dense_col[irow] -= ukj * lusup[luptr];
++luptr;
}
} else if ( segsze <= 3 ) {
ukj = dense_col[lsub[krep_ind]];
luptr += nsupr*(nsupc-1) + nsupc-1;
ukj1 = dense_col[lsub[krep_ind - 1]];
luptr1 = luptr - nsupr;
if ( segsze == 2 ) {
ukj -= ukj1 * lusup[luptr1];
dense_col[lsub[krep_ind]] = ukj;
for (i = lptr + nsupc; i < xlsub[fsupc+1]; ++i) {
irow = lsub[i];
++luptr; ++luptr1;
dense_col[irow] -= (ukj*lusup[luptr]
+ ukj1*lusup[luptr1]);
}
} else {
ukj2 = dense_col[lsub[krep_ind - 2]];
luptr2 = luptr1 - nsupr;
ukj1 -= ukj2 * lusup[luptr2-1];
ukj = ukj - ukj1*lusup[luptr1] - ukj2*lusup[luptr2];
dense_col[lsub[krep_ind]] = ukj;
dense_col[lsub[krep_ind-1]] = ukj1;
for (i = lptr + nsupc; i < xlsub[fsupc+1]; ++i) {
irow = lsub[i];
++luptr; ++luptr1; ++luptr2;
dense_col[irow] -= ( ukj*lusup[luptr]
+ ukj1*lusup[luptr1] + ukj2*lusup[luptr2] );
}
}
} else { /* segsze >= 4 */
/*
* Perform a triangular solve and block update,
* then scatter the result of sup-col update to dense[].
*/
no_zeros = kfnz - fsupc;
/* Copy U[*,j] segment from dense[*] to tempv[*]:
* The result of triangular solve is in tempv[*];
* The result of matrix vector update is in dense_col[*]
*/
isub = lptr + no_zeros;
for (i = 0; i < segsze; ++i) {
irow = lsub[isub];
tempv[i] = dense_col[irow]; /* Gather */
++isub;
}
/* start effective triangle */
luptr += nsupr * no_zeros + no_zeros;
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
STRSV( ftcs1, ftcs2, ftcs3, &segsze, &lusup[luptr],
&nsupr, tempv, &incx );
#else
dtrsv_( "L", "N", "U", &segsze, &lusup[luptr],
&nsupr, tempv, &incx );
#endif
luptr += segsze; /* Dense matrix-vector */
tempv1 = &tempv[segsze];
alpha = one;
beta = zero;
#ifdef _CRAY
SGEMV( ftcs2, &nrow, &segsze, &alpha, &lusup[luptr],
&nsupr, tempv, &incx, &beta, tempv1, &incy );
#else
dgemv_( "N", &nrow, &segsze, &alpha, &lusup[luptr],
&nsupr, tempv, &incx, &beta, tempv1, &incy );
#endif
#else
dlsolve ( nsupr, segsze, &lusup[luptr], tempv );
luptr += segsze; /* Dense matrix-vector */
tempv1 = &tempv[segsze];
dmatvec (nsupr, nrow, segsze, &lusup[luptr], tempv, tempv1);
#endif
/* Scatter tempv[*] into SPA dense[*] temporarily, such
* that tempv[*] can be used for the triangular solve of
* the next column of the panel. They will be copied into
* ucol[*] after the whole panel has been finished.
*/
isub = lptr + no_zeros;
for (i = 0; i < segsze; i++) {
irow = lsub[isub];
dense_col[irow] = tempv[i];
tempv[i] = zero;
isub++;
}
/* Scatter the update from tempv1[*] into SPA dense[*] */
/* Start dense rectangular L */
for (i = 0; i < nrow; i++) {
irow = lsub[isub];
dense_col[irow] -= tempv1[i];
tempv1[i] = zero;
++isub;
}
} /* else segsze>=4 ... */
} /* for each column in the panel... */
} /* else 1-D update ... */
} /* for each updating supernode ... */
}
int
sp_dtrsv_dist(char *uplo, char *trans, char *diag, SuperMatrix *L,
SuperMatrix *U, double *x, int *info)
{
/*
* Purpose
* =======
*
* sp_dtrsv_dist() solves one of the systems of equations
* A*x = b, or A'*x = b,
* where b and x are n element vectors and A is a sparse unit , or
* non-unit, upper or lower triangular matrix.
* No test for singularity or near-singularity is included in this
* routine. Such tests must be performed before calling this routine.
*
* Parameters
* ==========
*
* uplo - (input) char*
* On entry, uplo specifies whether the matrix is an upper or
* lower triangular matrix as follows:
* uplo = 'U' or 'u' A is an upper triangular matrix.
* uplo = 'L' or 'l' A is a lower triangular matrix.
*
* trans - (input) char*
* On entry, trans specifies the equations to be solved as
* follows:
* trans = 'N' or 'n' A*x = b.
* trans = 'T' or 't' A'*x = b.
* trans = 'C' or 'c' A'*x = b.
*
* diag - (input) char*
* On entry, diag specifies whether or not A is unit
* triangular as follows:
* diag = 'U' or 'u' A is assumed to be unit triangular.
* diag = 'N' or 'n' A is not assumed to be unit
* triangular.
*
* L - (input) SuperMatrix*
* The factor L from the factorization Pr*A*Pc=L*U. Use
* compressed row subscripts storage for supernodes,
* i.e., L has types: Stype = SC, Dtype = D, Mtype = TRLU.
*
* U - (input) SuperMatrix*
* The factor U from the factorization Pr*A*Pc=L*U.
* U has types: Stype = NC, Dtype = D, Mtype = TRU.
*
* x - (input/output) double*
* Before entry, the incremented array X must contain the n
* element right-hand side vector b. On exit, X is overwritten
* with the solution vector x.
*
* info - (output) int*
* If *info = -i, the i-th argument had an illegal value.
*
*/
#ifdef _CRAY
_fcd ftcs1, ftcs2, ftcs3;
#endif
SCformat *Lstore;
NCformat *Ustore;
double *Lval, *Uval;
int incx = 1, incy = 1;
double alpha = 1.0, beta = 1.0;
int nrow;
int fsupc, nsupr, nsupc, luptr, istart, irow;
int i, k, iptr, jcol;
double *work;
flops_t solve_ops;
extern SuperLUStat_t SuperLUStat;
/* Test the input parameters */
*info = 0;
if ( !lsame_(uplo,"L") && !lsame_(uplo, "U") ) *info = -1;
else if ( !lsame_(trans, "N") && !lsame_(trans, "T") ) *info = -2;
else if ( !lsame_(diag, "U") && !lsame_(diag, "N") ) *info = -3;
else if ( L->nrow != L->ncol || L->nrow < 0 ) *info = -4;
else if ( U->nrow != U->ncol || U->nrow < 0 ) *info = -5;
if ( *info ) {
i = -(*info);
xerbla_("sp_dtrsv_dist", &i);
return 0;
}
Lstore = L->Store;
Lval = Lstore->nzval;
Ustore = U->Store;
Uval = Ustore->nzval;
solve_ops = 0;
if ( !(work = doubleCalloc_dist(L->nrow)) )
ABORT("Malloc fails for work in sp_dtrsv_dist().");
if ( lsame_(trans, "N") ) { /* Form x := inv(A)*x. */
if ( lsame_(uplo, "L") ) {
/* Form x := inv(L)*x */
if ( L->nrow == 0 ) return 0; /* Quick return */
for (k = 0; k <= Lstore->nsuper; k++) {
fsupc = L_FST_SUPC(k);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_START(fsupc+1) - istart;
nsupc = L_FST_SUPC(k+1) - fsupc;
luptr = L_NZ_START(fsupc);
nrow = nsupr - nsupc;
solve_ops += nsupc * (nsupc - 1);
solve_ops += 2 * nrow * nsupc;
if ( nsupc == 1 ) {
for (iptr=istart+1; iptr < L_SUB_START(fsupc+1); ++iptr) {
irow = L_SUB(iptr);
++luptr;
x[irow] -= x[fsupc] * Lval[luptr];
}
} else {
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("N", strlen("N"));
ftcs3 = _cptofcd("U", strlen("U"));
STRSV(ftcs1, ftcs2, ftcs3, &nsupc, &Lval[luptr], &nsupr,
&x[fsupc], &incx);
SGEMV(ftcs2, &nrow, &nsupc, &alpha, &Lval[luptr+nsupc],
&nsupr, &x[fsupc], &incx, &beta, &work[0], &incy);
#else
dtrsv_("L", "N", "U", &nsupc, &Lval[luptr], &nsupr,
&x[fsupc], &incx);
dgemv_("N", &nrow, &nsupc, &alpha, &Lval[luptr+nsupc],
&nsupr, &x[fsupc], &incx, &beta, &work[0], &incy);
#endif /* _CRAY */
#else
dlsolve ( nsupr, nsupc, &Lval[luptr], &x[fsupc]);
dmatvec ( nsupr, nsupr-nsupc, nsupc, &Lval[luptr+nsupc],
&x[fsupc], &work[0] );
#endif
iptr = istart + nsupc;
for (i = 0; i < nrow; ++i, ++iptr) {
irow = L_SUB(iptr);
x[irow] -= work[i]; /* Scatter */
work[i] = 0.0;
}
}
} /* for k ... */
} else {
/* Form x := inv(U)*x */
if ( U->nrow == 0 ) return 0; /* Quick return */
for (k = Lstore->nsuper; k >= 0; k--) {
fsupc = L_FST_SUPC(k);
nsupr = L_SUB_START(fsupc+1) - L_SUB_START(fsupc);
nsupc = L_FST_SUPC(k+1) - fsupc;
luptr = L_NZ_START(fsupc);
solve_ops += nsupc * (nsupc + 1);
if ( nsupc == 1 ) {
x[fsupc] /= Lval[luptr];
for (i = U_NZ_START(fsupc); i < U_NZ_START(fsupc+1); ++i) {
irow = U_SUB(i);
x[irow] -= x[fsupc] * Uval[i];
}
} else {
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
ftcs1 = _cptofcd("U", strlen("U"));
ftcs2 = _cptofcd("N", strlen("N"));
STRSV(ftcs1, ftcs2, ftcs2, &nsupc, &Lval[luptr], &nsupr,
&x[fsupc], &incx);
#else
dtrsv_("U", "N", "N", &nsupc, &Lval[luptr], &nsupr,
&x[fsupc], &incx);
#endif
#else
dusolve ( nsupr, nsupc, &Lval[luptr], &x[fsupc] );
#endif
for (jcol = fsupc; jcol < L_FST_SUPC(k+1); jcol++) {
solve_ops += 2*(U_NZ_START(jcol+1) - U_NZ_START(jcol));
for (i = U_NZ_START(jcol); i < U_NZ_START(jcol+1);
i++) {
irow = U_SUB(i);
x[irow] -= x[jcol] * Uval[i];
}
}
}
} /* for k ... */
}
} else { /* Form x := inv(A')*x */
if ( lsame_(uplo, "L") ) {
/* Form x := inv(L')*x */
if ( L->nrow == 0 ) return 0; /* Quick return */
for (k = Lstore->nsuper; k >= 0; --k) {
fsupc = L_FST_SUPC(k);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_START(fsupc+1) - istart;
nsupc = L_FST_SUPC(k+1) - fsupc;
luptr = L_NZ_START(fsupc);
solve_ops += 2 * (nsupr - nsupc) * nsupc;
for (jcol = fsupc; jcol < L_FST_SUPC(k+1); jcol++) {
iptr = istart + nsupc;
for (i = L_NZ_START(jcol) + nsupc;
i < L_NZ_START(jcol+1); i++) {
irow = L_SUB(iptr);
x[jcol] -= x[irow] * Lval[i];
iptr++;
}
}
if ( nsupc > 1 ) {
solve_ops += nsupc * (nsupc - 1);
#ifdef _CRAY
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("T", strlen("T"));
ftcs3 = _cptofcd("U", strlen("U"));
STRSV(ftcs1, ftcs2, ftcs3, &nsupc, &Lval[luptr], &nsupr,
&x[fsupc], &incx);
#else
dtrsv_("L", "T", "U", &nsupc, &Lval[luptr], &nsupr,
&x[fsupc], &incx);
#endif
}
}
} else {
/* Form x := inv(U')*x */
if ( U->nrow == 0 ) return 0; /* Quick return */
for (k = 0; k <= Lstore->nsuper; k++) {
fsupc = L_FST_SUPC(k);
nsupr = L_SUB_START(fsupc+1) - L_SUB_START(fsupc);
nsupc = L_FST_SUPC(k+1) - fsupc;
luptr = L_NZ_START(fsupc);
for (jcol = fsupc; jcol < L_FST_SUPC(k+1); jcol++) {
solve_ops += 2*(U_NZ_START(jcol+1) - U_NZ_START(jcol));
for (i = U_NZ_START(jcol); i < U_NZ_START(jcol+1); i++) {
irow = U_SUB(i);
x[jcol] -= x[irow] * Uval[i];
}
}
solve_ops += nsupc * (nsupc + 1);
if ( nsupc == 1 ) {
x[fsupc] /= Lval[luptr];
} else {
#ifdef _CRAY
ftcs1 = _cptofcd("U", strlen("U"));
ftcs2 = _cptofcd("T", strlen("T"));
ftcs3 = _cptofcd("N", strlen("N"));
STRSV( ftcs1, ftcs2, ftcs3, &nsupc, &Lval[luptr], &nsupr,
&x[fsupc], &incx);
#else
dtrsv_("U", "T", "N", &nsupc, &Lval[luptr], &nsupr,
&x[fsupc], &incx);
#endif
}
} /* for k ... */
}
}
SuperLUStat.ops[SOLVE] += solve_ops;
SUPERLU_FREE(work);
return 0;
}
/*! \brief Performs numeric block updates within the relaxed snode.
*/
int
dsnode_bmod (
const int jcol, /* in */
const int jsupno, /* in */
const int fsupc, /* in */
double *dense, /* in */
double *tempv, /* working array */
GlobalLU_t *Glu, /* modified */
SuperLUStat_t *stat /* output */
)
{
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
_fcd ftcs1 = _cptofcd("L", strlen("L")),
ftcs2 = _cptofcd("N", strlen("N")),
ftcs3 = _cptofcd("U", strlen("U"));
#endif
int incx = 1, incy = 1;
double alpha = -1.0, beta = 1.0;
#endif
int luptr, nsupc, nsupr, nrow;
int isub, irow, i, iptr;
register int ufirst, nextlu;
int *lsub, *xlsub;
double *lusup;
int *xlusup;
flops_t *ops = stat->ops;
lsub = Glu->lsub;
xlsub = Glu->xlsub;
lusup = Glu->lusup;
xlusup = Glu->xlusup;
nextlu = xlusup[jcol];
/*
* Process the supernodal portion of L\U[*,j]
*/
for (isub = xlsub[fsupc]; isub < xlsub[fsupc+1]; isub++) {
irow = lsub[isub];
lusup[nextlu] = dense[irow];
dense[irow] = 0;
++nextlu;
}
xlusup[jcol + 1] = nextlu; /* Initialize xlusup for next column */
if ( fsupc < jcol ) {
luptr = xlusup[fsupc];
nsupr = xlsub[fsupc+1] - xlsub[fsupc];
nsupc = jcol - fsupc; /* Excluding jcol */
ufirst = xlusup[jcol]; /* Points to the beginning of column
jcol in supernode L\U(jsupno). */
nrow = nsupr - nsupc;
ops[TRSV] += nsupc * (nsupc - 1);
ops[GEMV] += 2 * nrow * nsupc;
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
STRSV( ftcs1, ftcs2, ftcs3, &nsupc, &lusup[luptr], &nsupr,
&lusup[ufirst], &incx );
SGEMV( ftcs2, &nrow, &nsupc, &alpha, &lusup[luptr+nsupc], &nsupr,
&lusup[ufirst], &incx, &beta, &lusup[ufirst+nsupc], &incy );
#else
dtrsv_( "L", "N", "U", &nsupc, &lusup[luptr], &nsupr,
&lusup[ufirst], &incx );
dgemv_( "N", &nrow, &nsupc, &alpha, &lusup[luptr+nsupc], &nsupr,
&lusup[ufirst], &incx, &beta, &lusup[ufirst+nsupc], &incy );
#endif
#else
dlsolve ( nsupr, nsupc, &lusup[luptr], &lusup[ufirst] );
dmatvec ( nsupr, nrow, nsupc, &lusup[luptr+nsupc],
&lusup[ufirst], &tempv[0] );
/* Scatter tempv[*] into lusup[*] */
iptr = ufirst + nsupc;
for (i = 0; i < nrow; i++) {
lusup[iptr++] -= tempv[i];
tempv[i] = 0.0;
}
#endif
}
return 0;
}
void
dgstrs(trans_t trans, SuperMatrix *L, SuperMatrix *U,
int *perm_r, int *perm_c, SuperMatrix *B, Gstat_t *Gstat, int *info)
{
/*
* -- SuperLU MT routine (version 1.0) --
* Univ. of California Berkeley, Xerox Palo Alto Research Center,
* and Lawrence Berkeley National Lab.
* August 15, 1997
*
* Purpose
* =======
*
* dgstrs() solves a system of linear equations A*X=B or A'*X=B
* with A sparse and B dense, using the LU factorization computed by
* pdgstrf().
*
* Arguments
* =========
*
* trans (input) Specifies the form of the system of equations:
* = NOTRANS: A * X = B (No transpose)
* = TRANS: A'* X = B (Transpose)
*
* L (input) SuperMatrix*
* The factor L from the factorization Pr*A*Pc=L*U as computed by
* pdgstrf(). Use compressed row subscripts storage for supernodes,
* i.e., L has types: Stype = SCP, Dtype = _D, Mtype = TRLU.
*
* U (input) SuperMatrix*
* The factor U from the factorization Pr*A*Pc=L*U as computed by
* pdgstrf(). Use column-wise storage scheme, i.e., U has types:
* Stype = NCP, Dtype = _D, Mtype = TRU.
*
* perm_r (input) int*
* Row permutation vector of size L->nrow, which defines the
* permutation matrix Pr; perm_r[i] = j means row i of A is in
* position j in Pr*A.
*
* perm_c (int*) dimension A->ncol
* Column permutation vector, which defines the
* permutation matrix Pc; perm_c[i] = j means column i of A is
* in position j in A*Pc.
*
* B (input/output) SuperMatrix*
* B has types: Stype = DN, Dtype = _D, Mtype = GE.
* On entry, the right hand side matrix.
* On exit, the solution matrix if info = 0;
*
* Gstat (output) Gstat_t*
* Record all the statistics about the triangular solves;
* See Gstat_t structure defined in util.h.
*
* info (output) Diagnostics
* = 0: successful exit
* < 0: if info = -i, the i-th argument had an illegal value
*
*/
#if ( MACH==CRAY_PVP )
_fcd ftcs1, ftcs2, ftcs3, ftcs4;
#endif
#ifdef USE_VENDOR_BLAS
int incx = 1, incy = 1;
double alpha = 1.0, beta = 1.0;
#endif
register int j, k, jcol, iptr, luptr, ksupno, istart, irow, bptr;
register int fsupc, nsuper;
int i, n, nsupc, nsupr, nrow, nrhs, ldb;
int *supno;
DNformat *Bstore;
SCPformat *Lstore;
NCPformat *Ustore;
double *Lval, *Uval, *Bmat;
double *work, *work_col, *rhs_work, *soln;
flops_t solve_ops;
void dprint_soln();
/* Test input parameters ... */
*info = 0;
Bstore = B->Store;
ldb = Bstore->lda;
nrhs = B->ncol;
if ( trans != NOTRANS && trans != TRANS ) *info = -1;
else if ( L->nrow != L->ncol || L->nrow < 0 ) *info = -3;
else if ( U->nrow != U->ncol || U->nrow < 0 ) *info = -4;
else if ( ldb < MAX(0, L->nrow) ) *info = -6;
if ( *info ) {
i = -(*info);
xerbla_("dgstrs", &i);
return;
}
n = L->nrow;
work = doubleCalloc(n * nrhs);
if ( !work ) ABORT("Malloc fails for local work[].");
soln = doubleMalloc(n);
if ( !soln ) ABORT("Malloc fails for local soln[].");
Bmat = Bstore->nzval;
Lstore = L->Store;
Lval = Lstore->nzval;
Ustore = U->Store;
Uval = Ustore->nzval;
supno = Lstore->col_to_sup;
nsuper = Lstore->nsuper;
solve_ops = 0;
if ( trans == NOTRANS ) {
/* Permute right hand sides to form Pr*B */
for (i = 0, bptr = 0; i < nrhs; i++, bptr += ldb) {
rhs_work = &Bmat[bptr];
for (k = 0; k < n; k++) soln[perm_r[k]] = rhs_work[k];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
/* Forward solve PLy=Pb. */
/*>> for (k = 0; k < n; k += nsupc) {
ksupno = supno[k];
*/
for (ksupno = 0; ksupno <= nsuper; ++ksupno) {
fsupc = L_FST_SUPC(ksupno);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_END(fsupc) - istart;
nsupc = L_LAST_SUPC(ksupno) - fsupc;
nrow = nsupr - nsupc;
solve_ops += nsupc * (nsupc - 1) * nrhs;
solve_ops += 2 * nrow * nsupc * nrhs;
if ( nsupc == 1 ) {
for (j = 0, bptr = 0; j < nrhs; j++, bptr += ldb) {
rhs_work = &Bmat[bptr];
luptr = L_NZ_START(fsupc);
for (iptr=istart+1; iptr < L_SUB_END(fsupc); iptr++){
irow = L_SUB(iptr);
++luptr;
rhs_work[irow] -= rhs_work[fsupc] * Lval[luptr];
}
}
} else {
luptr = L_NZ_START(fsupc);
#ifdef USE_VENDOR_BLAS
#if ( MACH==CRAY_PVP )
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("N", strlen("N"));
ftcs3 = _cptofcd("U", strlen("U"));
STRSM(ftcs1, ftcs1, ftcs2, ftcs3, &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
SGEMM(ftcs2, ftcs2, &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#else
dtrsm_("L", "L", "N", "U", &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
dgemm_( "N", "N", &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#endif
for (j = 0, bptr = 0; j < nrhs; j++, bptr += ldb) {
rhs_work = &Bmat[bptr];
work_col = &work[j*n];
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
rhs_work[irow] -= work_col[i]; /* Scatter */
work_col[i] = 0.0;
iptr++;
}
}
#else
for (j = 0, bptr = 0; j < nrhs; j++, bptr += ldb) {
rhs_work = &Bmat[bptr];
dlsolve (nsupr, nsupc, &Lval[luptr], &rhs_work[fsupc]);
dmatvec (nsupr, nrow, nsupc, &Lval[luptr+nsupc],
&rhs_work[fsupc], &work[0] );
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
rhs_work[irow] -= work[i];
work[i] = 0.0;
iptr++;
}
}
#endif
} /* if-else: nsupc == 1 ... */
} /* for L-solve */
#if ( DEBUGlevel>=2 )
printf("After L-solve: y=\n");
dprint_soln(n, nrhs, Bmat);
#endif
/*
* Back solve Ux=y.
*/
/*>> for (k = n-1; k >= 0; k -= nsupc) {
ksupno = supno[k];
*/
for (ksupno = nsuper; ksupno >= 0; --ksupno) {
fsupc = L_FST_SUPC(ksupno);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_END(fsupc) - istart;
nsupc = L_LAST_SUPC(ksupno) - fsupc;
luptr = L_NZ_START(fsupc);
solve_ops += nsupc * (nsupc + 1) * nrhs;
/* dense triangular matrix */
if ( nsupc == 1 ) {
rhs_work = &Bmat[0];
for (j = 0; j < nrhs; j++) {
rhs_work[fsupc] /= Lval[luptr];
rhs_work += ldb;
}
} else {
#ifdef USE_VENDOR_BLAS
#if ( MACH==CRAY_PVP )
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("U", strlen("U"));
ftcs3 = _cptofcd("N", strlen("N"));
STRSM(ftcs1, ftcs2, ftcs3, ftcs3, &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
#else
dtrsm_("L", "U", "N", "N", &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
#endif
#else
for (j = 0, bptr = fsupc; j < nrhs; j++, bptr += ldb) {
dusolve (nsupr, nsupc, &Lval[luptr], &Bmat[bptr]);
}
#endif
}
/* matrix-vector update */
for (j = 0, bptr = 0; j < nrhs; ++j, bptr += ldb) {
rhs_work = &Bmat[bptr];
for (jcol = fsupc; jcol < fsupc + nsupc; jcol++) {
solve_ops += 2*(U_NZ_END(jcol) - U_NZ_START(jcol));
for (i = U_NZ_START(jcol); i < U_NZ_END(jcol); i++ ){
irow = U_SUB(i);
rhs_work[irow] -= rhs_work[jcol] * Uval[i];
}
}
}
} /* for U-solve */
#if ( DEBUGlevel>=2 )
printf("After U-solve: x=\n");
dprint_soln(n, nrhs, Bmat);
#endif
/* Compute the final solution X <= Pc*X. */
for (i = 0, bptr = 0; i < nrhs; i++, bptr += ldb) {
rhs_work = &Bmat[bptr];
for (k = 0; k < n; k++) soln[k] = rhs_work[perm_c[k]];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
} else { /* Solve A'*X=B */
/* Permute right hand sides to form Pc'*B. */
for (i = 0, bptr = 0; i < nrhs; i++, bptr += ldb) {
rhs_work = &Bmat[bptr];
for (k = 0; k < n; k++) soln[perm_c[k]] = rhs_work[k];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
for (k = 0; k < nrhs; ++k) {
/* Multiply by inv(U'). */
sp_dtrsv("U", "T", "N", L, U, &Bmat[k*ldb], info);
/* Multiply by inv(L'). */
sp_dtrsv("L", "T", "U", L, U, &Bmat[k*ldb], info);
}
/* Compute the final solution X <= Pr'*X (=inv(Pr)*X) */
for (i = 0, bptr = 0; i < nrhs; i++, bptr += ldb) {
rhs_work = &Bmat[bptr];
for (k = 0; k < n; k++) soln[k] = rhs_work[perm_r[k]];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
} /* if-else trans */
Gstat->ops[TRISOLVE] = solve_ops;
SUPERLU_FREE(work);
SUPERLU_FREE(soln);
}
int dfgmr(int n,
void (*dmatvec) (double, double[], double, double[]),
void (*dpsolve) (int, double[], double[]),
double *rhs, double *sol, double tol, int im, int *itmax, FILE * fits)
{
/*----------------------------------------------------------------------
| *** Preconditioned FGMRES ***
+-----------------------------------------------------------------------
| This is a simple version of the ARMS preconditioned FGMRES algorithm.
+-----------------------------------------------------------------------
| Y. S. Dec. 2000. -- Apr. 2008
+-----------------------------------------------------------------------
| on entry:
|----------
|
| rhs = real vector of length n containing the right hand side.
| sol = real vector of length n containing an initial guess to the
| solution on input.
| tol = tolerance for stopping iteration
| im = Krylov subspace dimension
| (itmax) = max number of iterations allowed.
| fits = NULL: no output
| != NULL: file handle to output " resid vs time and its"
|
| on return:
|----------
| fgmr int = 0 --> successful return.
| int = 1 --> convergence not achieved in itmax iterations.
| sol = contains an approximate solution (upon successful return).
| itmax = has changed. It now contains the number of steps required
| to converge --
+-----------------------------------------------------------------------
| internal work arrays:
|----------
| vv = work array of length [im+1][n] (used to store the Arnoldi
| basis)
| hh = work array of length [im][im+1] (Householder matrix)
| z = work array of length [im][n] to store preconditioned vectors
+-----------------------------------------------------------------------
| subroutines called :
| matvec - matrix-vector multiplication operation
| psolve - (right) preconditionning operation
| psolve can be a NULL pointer (GMRES without preconditioner)
+---------------------------------------------------------------------*/
int maxits = *itmax;
int i, i1, ii, j, k, k1, its, retval, i_1 = 1, i_2 = 2;
double beta, eps1 = 0.0, t, t0, gam;
double **hh, *c, *s, *rs;
double **vv, **z, tt;
double zero = 0.0;
double one = 1.0;
its = 0;
vv = (double **)SUPERLU_MALLOC((im + 1) * sizeof(double *));
for (i = 0; i <= im; i++) vv[i] = doubleMalloc(n);
z = (double **)SUPERLU_MALLOC(im * sizeof(double *));
hh = (double **)SUPERLU_MALLOC(im * sizeof(double *));
for (i = 0; i < im; i++)
{
hh[i] = doubleMalloc(i + 2);
z[i] = doubleMalloc(n);
}
c = doubleMalloc(im);
s = doubleMalloc(im);
rs = doubleMalloc(im + 1);
/*---- outer loop starts here ----*/
do
{
/*---- compute initial residual vector ----*/
dmatvec(one, sol, zero, vv[0]);
for (j = 0; j < n; j++)
vv[0][j] = rhs[j] - vv[0][j]; /* vv[0]= initial residual */
beta = dnrm2_(&n, vv[0], &i_1);
/*---- print info if fits != null ----*/
if (fits != NULL && its == 0)
fprintf(fits, "%8d %10.2e\n", its, beta);
/*if ( beta <= tol * dnrm2_(&n, rhs, &i_1) )*/
if ( !(beta > tol * dnrm2_(&n, rhs, &i_1)) )
break;
t = 1.0 / beta;
/*---- normalize: vv[0] = vv[0] / beta ----*/
for (j = 0; j < n; j++)
vv[0][j] = vv[0][j] * t;
if (its == 0)
eps1 = tol * beta;
/*---- initialize 1-st term of rhs of hessenberg system ----*/
rs[0] = beta;
for (i = 0; i < im; i++)
{
its++;
i1 = i + 1;
/*------------------------------------------------------------
| (Right) Preconditioning Operation z_{j} = M^{-1} v_{j}
+-----------------------------------------------------------*/
if (dpsolve)
dpsolve(n, z[i], vv[i]);
else
dcopy_(&n, vv[i], &i_1, z[i], &i_1);
/*---- matvec operation w = A z_{j} = A M^{-1} v_{j} ----*/
dmatvec(one, z[i], zero, vv[i1]);
/*------------------------------------------------------------
| modified gram - schmidt...
| h_{i,j} = (w,v_{i})
| w = w - h_{i,j} v_{i}
+------------------------------------------------------------*/
t0 = dnrm2_(&n, vv[i1], &i_1);
for (j = 0; j <= i; j++)
{
double negt;
tt = ddot_(&n, vv[j], &i_1, vv[i1], &i_1);
hh[i][j] = tt;
negt = -tt;
daxpy_(&n, &negt, vv[j], &i_1, vv[i1], &i_1);
}
/*---- h_{j+1,j} = ||w||_{2} ----*/
t = dnrm2_(&n, vv[i1], &i_1);
while (t < 0.5 * t0)
{
t0 = t;
for (j = 0; j <= i; j++)
{
double negt;
tt = ddot_(&n, vv[j], &i_1, vv[i1], &i_1);
hh[i][j] += tt;
negt = -tt;
daxpy_(&n, &negt, vv[j], &i_1, vv[i1], &i_1);
}
t = dnrm2_(&n, vv[i1], &i_1);
}
hh[i][i1] = t;
if (t != 0.0)
{
/*---- v_{j+1} = w / h_{j+1,j} ----*/
t = 1.0 / t;
for (k = 0; k < n; k++)
vv[i1][k] = vv[i1][k] * t;
}
/*---------------------------------------------------
| done with modified gram schimdt and arnoldi step
| now update factorization of hh
+--------------------------------------------------*/
/*--------------------------------------------------------
| perform previous transformations on i-th column of h
+-------------------------------------------------------*/
for (k = 1; k <= i; k++)
{
k1 = k - 1;
tt = hh[i][k1];
hh[i][k1] = c[k1] * tt + s[k1] * hh[i][k];
hh[i][k] = -s[k1] * tt + c[k1] * hh[i][k];
}
gam = sqrt(pow(hh[i][i], 2) + pow(hh[i][i1], 2));
/*---------------------------------------------------
| if gamma is zero then any small value will do
| affect only residual estimate
+--------------------------------------------------*/
/* if (gam == 0.0) gam = epsmac; */
/*---- get next plane rotation ---*/
if (gam == 0.0)
{
c[i] = one;
s[i] = zero;
}
else
{
c[i] = hh[i][i] / gam;
s[i] = hh[i][i1] / gam;
}
rs[i1] = -s[i] * rs[i];
rs[i] = c[i] * rs[i];
/*----------------------------------------------------
| determine residual norm and test for convergence
+---------------------------------------------------*/
hh[i][i] = c[i] * hh[i][i] + s[i] * hh[i][i1];
beta = fabs(rs[i1]);
if (fits != NULL)
fprintf(fits, "%8d %10.2e\n", its, beta);
if (beta <= eps1 || its >= maxits)
break;
}
if (i == im) i--;
/*---- now compute solution. 1st, solve upper triangular system ----*/
rs[i] = rs[i] / hh[i][i];
for (ii = 1; ii <= i; ii++)
{
k = i - ii;
k1 = k + 1;
tt = rs[k];
for (j = k1; j <= i; j++)
tt = tt - hh[j][k] * rs[j];
rs[k] = tt / hh[k][k];
}
/*---- linear combination of v[i]'s to get sol. ----*/
for (j = 0; j <= i; j++)
{
tt = rs[j];
for (k = 0; k < n; k++)
sol[k] += tt * z[j][k];
}
/* calculate the residual and output */
dmatvec(one, sol, zero, vv[0]);
for (j = 0; j < n; j++)
vv[0][j] = rhs[j] - vv[0][j]; /* vv[0]= initial residual */
/*---- print info if fits != null ----*/
beta = dnrm2_(&n, vv[0], &i_1);
/*---- restart outer loop if needed ----*/
/*if (beta >= eps1 / tol)*/
if ( !(beta < eps1 / tol) )
{
its = maxits + 10;
break;
}
if (beta <= eps1)
break;
} while(its < maxits);
retval = (its >= maxits);
for (i = 0; i <= im; i++)
SUPERLU_FREE(vv[i]);
SUPERLU_FREE(vv);
for (i = 0; i < im; i++)
{
SUPERLU_FREE(hh[i]);
SUPERLU_FREE(z[i]);
}
SUPERLU_FREE(hh);
SUPERLU_FREE(z);
SUPERLU_FREE(c);
SUPERLU_FREE(s);
SUPERLU_FREE(rs);
*itmax = its;
return retval;
} /*----end of fgmr ----*/
int
pdgstrf_column_bmod(
const int pnum, /* process number */
const int jcol, /* current column in the panel */
const int fpanelc,/* first column in the panel */
const int nseg, /* number of s-nodes to update jcol */
int *segrep,/* in */
int *repfnz,/* in */
double *dense, /* modified */
double *tempv, /* working array */
pxgstrf_shared_t *pxgstrf_shared, /* modified */
Gstat_t *Gstat /* modified */
)
{
/*
* -- SuperLU MT routine (version 1.0) --
* Univ. of California Berkeley, Xerox Palo Alto Research Center,
* and Lawrence Berkeley National Lab.
* August 15, 1997
*
* Purpose:
* ========
* Performs numeric block updates (sup-col) in topological order.
* It features: col-col, 2cols-col, 3cols-col, and sup-col updates.
* Special processing on the supernodal portion of L\U[*,j].
*
* Return value:
* =============
* 0 - successful return
* > 0 - number of bytes allocated when run out of space
*
*/
#if ( MACH==CRAY_PVP )
_fcd ftcs1 = _cptofcd("L", strlen("L")),
ftcs2 = _cptofcd("N", strlen("N")),
ftcs3 = _cptofcd("U", strlen("U"));
#endif
#ifdef USE_VENDOR_BLAS
int incx = 1, incy = 1;
double alpha, beta;
#endif
GlobalLU_t *Glu = pxgstrf_shared->Glu; /* modified */
/* krep = representative of current k-th supernode
* fsupc = first supernodal column
* nsupc = no of columns in supernode
* nsupr = no of rows in supernode (used as leading dimension)
* luptr = location of supernodal LU-block in storage
* kfnz = first nonz in the k-th supernodal segment
* no_zeros = no of leading zeros in a supernodal U-segment
*/
register double ukj, ukj1, ukj2;
register int lptr, kfnz, isub, irow, i, no_zeros;
register int luptr, luptr1, luptr2;
int fsupc, nsupc, nsupr, segsze;
int nrow; /* No of rows in the matrix of matrix-vector */
int jsupno, k, ksub, krep, krep_ind, ksupno;
int ufirst, nextlu;
int fst_col; /* First column within small LU update */
int d_fsupc; /* Distance between the first column of the current
panel and the first column of the current snode.*/
int *xsup, *supno;
int *lsub, *xlsub, *xlsub_end;
double *lusup;
int *xlusup, *xlusup_end;
double *tempv1;
int mem_error;
register float flopcnt;
xsup = Glu->xsup;
supno = Glu->supno;
lsub = Glu->lsub;
xlsub = Glu->xlsub;
xlsub_end = Glu->xlsub_end;
lusup = Glu->lusup;
xlusup = Glu->xlusup;
xlusup_end = Glu->xlusup_end;
jsupno = supno[jcol];
/*
* For each nonz supernode segment of U[*,j] in topological order
*/
k = nseg - 1;
for (ksub = 0; ksub < nseg; ksub++) {
krep = segrep[k];
k--;
ksupno = supno[krep];
#if ( DEBUGlvel>=2 )
if (jcol==BADCOL)
printf("(%d) pdgstrf_column_bmod[1]: %d, nseg %d, krep %d, jsupno %d, ksupno %d\n",
pnum, jcol, nseg, krep, jsupno, ksupno);
#endif
if ( jsupno != ksupno ) { /* Outside the rectangular supernode */
fsupc = xsup[ksupno];
fst_col = MAX ( fsupc, fpanelc );
/* Distance from the current supernode to the current panel;
d_fsupc=0 if fsupc >= fpanelc. */
d_fsupc = fst_col - fsupc;
luptr = xlusup[fst_col] + d_fsupc;
lptr = xlsub[fsupc] + d_fsupc;
kfnz = repfnz[krep];
kfnz = MAX ( kfnz, fpanelc );
segsze = krep - kfnz + 1;
nsupc = krep - fst_col + 1;
nsupr = xlsub_end[fsupc] - xlsub[fsupc]; /* Leading dimension */
nrow = nsupr - d_fsupc - nsupc;
krep_ind = lptr + nsupc - 1;
flopcnt = segsze * (segsze - 1) + 2 * nrow * segsze;
Gstat->procstat[pnum].fcops += flopcnt;
#if ( DEBUGlevel>=2 )
if (jcol==BADCOL)
printf("(%d) pdgstrf_column_bmod[2]: %d, krep %d, kfnz %d, segsze %d, d_fsupc %d,\
fsupc %d, nsupr %d, nsupc %d\n",
pnum, jcol, krep, kfnz, segsze, d_fsupc, fsupc, nsupr, nsupc);
#endif
/*
* Case 1: Update U-segment of size 1 -- col-col update
*/
if ( segsze == 1 ) {
ukj = dense[lsub[krep_ind]];
luptr += nsupr*(nsupc-1) + nsupc;
for (i = lptr + nsupc; i < xlsub_end[fsupc]; ++i) {
irow = lsub[i];
dense[irow] -= ukj*lusup[luptr];
luptr++;
}
} else if ( segsze <= 3 ) {
ukj = dense[lsub[krep_ind]];
luptr += nsupr*(nsupc-1) + nsupc-1;
ukj1 = dense[lsub[krep_ind - 1]];
luptr1 = luptr - nsupr;
if ( segsze == 2 ) { /* Case 2: 2cols-col update */
ukj -= ukj1 * lusup[luptr1];
dense[lsub[krep_ind]] = ukj;
for (i = lptr + nsupc; i < xlsub_end[fsupc]; ++i) {
irow = lsub[i];
luptr++;
luptr1++;
dense[irow] -= ( ukj*lusup[luptr]
+ ukj1*lusup[luptr1] );
}
} else { /* Case 3: 3cols-col update */
ukj2 = dense[lsub[krep_ind - 2]];
luptr2 = luptr1 - nsupr;
ukj1 -= ukj2 * lusup[luptr2-1];
ukj = ukj - ukj1*lusup[luptr1] - ukj2*lusup[luptr2];
dense[lsub[krep_ind]] = ukj;
dense[lsub[krep_ind-1]] = ukj1;
for (i = lptr + nsupc; i < xlsub_end[fsupc]; ++i) {
irow = lsub[i];
luptr++;
luptr1++;
luptr2++;
dense[irow] -= ( ukj*lusup[luptr]
+ ukj1*lusup[luptr1] + ukj2*lusup[luptr2] );
}
}
} else {
/*
* Case: sup-col update
* Perform a triangular solve and block update,
* then scatter the result of sup-col update to dense
*/
no_zeros = kfnz - fst_col;
/* Copy U[*,j] segment from dense[*] to tempv[*] */
isub = lptr + no_zeros;
for (i = 0; i < segsze; i++) {
irow = lsub[isub];
tempv[i] = dense[irow];
++isub;
}
/* Dense triangular solve -- start effective triangle */
luptr += nsupr * no_zeros + no_zeros;
#ifdef USE_VENDOR_BLAS
#if ( MACH==CRAY_PVP )
STRSV( ftcs1, ftcs2, ftcs3, &segsze, &lusup[luptr],
&nsupr, tempv, &incx );
#else
dtrsv_( "L", "N", "U", &segsze, &lusup[luptr],
&nsupr, tempv, &incx );
#endif
luptr += segsze; /* Dense matrix-vector */
tempv1 = &tempv[segsze];
alpha = 1.0; beta = 0.0;
#if ( MACH==CRAY_PVP )
SGEMV( ftcs2, &nrow, &segsze, &alpha, &lusup[luptr],
&nsupr, tempv, &incx, &beta, tempv1, &incy );
#else
dgemv_( "N", &nrow, &segsze, &alpha, &lusup[luptr],
&nsupr, tempv, &incx, &beta, tempv1, &incy );
#endif
#else
dlsolve ( nsupr, segsze, &lusup[luptr], tempv );
luptr += segsze; /* Dense matrix-vector */
tempv1 = &tempv[segsze];
dmatvec (nsupr, nrow , segsze, &lusup[luptr], tempv, tempv1);
#endif
/* Scatter tempv[] into SPA dense[*] */
isub = lptr + no_zeros;
for (i = 0; i < segsze; i++) {
irow = lsub[isub];
dense[irow] = tempv[i]; /* Scatter */
tempv[i] = 0.0;
isub++;
}
/* Scatter tempv1[] into SPA dense[*] */
for (i = 0; i < nrow; i++) {
irow = lsub[isub];
dense[irow] -= tempv1[i]; /* Scatter-add */
tempv1[i] = 0.0;
++isub;
}
} /* else segsze >= 4 */
} /* if jsupno ... */
} /* for each segment... */
/* ------------------------------------------
Process the supernodal portion of L\U[*,j]
------------------------------------------ */
fsupc = SUPER_FSUPC (jsupno);
nsupr = xlsub_end[fsupc] - xlsub[fsupc];
if ( (mem_error = Glu_alloc(pnum, jcol, nsupr, LUSUP, &nextlu,
pxgstrf_shared)) )
return mem_error;
xlusup[jcol] = nextlu;
lusup = Glu->lusup;
/* Gather the nonzeros from SPA dense[*,j] into L\U[*,j] */
for (isub = xlsub[fsupc]; isub < xlsub_end[fsupc]; ++isub) {
irow = lsub[isub];
lusup[nextlu] = dense[irow];
dense[irow] = 0.0;
#ifdef DEBUG
if (jcol == -1)
printf("(%d) pdgstrf_column_bmod[lusup] jcol %d, irow %d, lusup %.10e\n",
pnum, jcol, irow, lusup[nextlu]);
#endif
++nextlu;
}
xlusup_end[jcol] = nextlu; /* close L\U[*,jcol] */
#if ( DEBUGlevel>=2 )
if (jcol == -1) {
nrow = xlusup_end[jcol] - xlusup[jcol];
print_double_vec("before sup-col update", nrow, &lsub[xlsub[fsupc]],
&lusup[xlusup[jcol]]);
}
#endif
/*
* For more updates within the panel (also within the current supernode),
* should start from the first column of the panel, or the first column
* of the supernode, whichever is bigger. There are 2 cases:
* (1) fsupc < fpanelc, then fst_col := fpanelc
* (2) fsupc >= fpanelc, then fst_col := fsupc
*/
fst_col = MAX ( fsupc, fpanelc );
if ( fst_col < jcol ) {
/* distance between the current supernode and the current panel;
d_fsupc=0 if fsupc >= fpanelc. */
d_fsupc = fst_col - fsupc;
lptr = xlsub[fsupc] + d_fsupc;
luptr = xlusup[fst_col] + d_fsupc;
nsupr = xlsub_end[fsupc] - xlsub[fsupc]; /* Leading dimension */
nsupc = jcol - fst_col; /* Excluding jcol */
nrow = nsupr - d_fsupc - nsupc;
/* points to the beginning of jcol in supernode L\U[*,jsupno] */
ufirst = xlusup[jcol] + d_fsupc;
#if ( DEBUGlevel>=2 )
if (jcol==BADCOL)
printf("(%d) pdgstrf_column_bmod[3] jcol %d, fsupc %d, nsupr %d, nsupc %d, nrow %d\n",
pnum, jcol, fsupc, nsupr, nsupc, nrow);
#endif
flopcnt = nsupc * (nsupc - 1) + 2 * nrow * nsupc;
Gstat->procstat[pnum].fcops += flopcnt;
/* ops[TRSV] += nsupc * (nsupc - 1);
ops[GEMV] += 2 * nrow * nsupc; */
#ifdef USE_VENDOR_BLAS
alpha = -1.0; beta = 1.0; /* y := beta*y + alpha*A*x */
#if ( MACH==CRAY_PVP )
STRSV( ftcs1, ftcs2, ftcs3, &nsupc, &lusup[luptr],
&nsupr, &lusup[ufirst], &incx );
SGEMV( ftcs2, &nrow, &nsupc, &alpha, &lusup[luptr+nsupc], &nsupr,
&lusup[ufirst], &incx, &beta, &lusup[ufirst+nsupc], &incy );
#else
dtrsv_( "L", "N", "U", &nsupc, &lusup[luptr],
&nsupr, &lusup[ufirst], &incx );
dgemv_( "N", &nrow, &nsupc, &alpha, &lusup[luptr+nsupc], &nsupr,
&lusup[ufirst], &incx, &beta, &lusup[ufirst+nsupc], &incy );
#endif
#else
dlsolve ( nsupr, nsupc, &lusup[luptr], &lusup[ufirst] );
dmatvec ( nsupr, nrow, nsupc, &lusup[luptr+nsupc],
&lusup[ufirst], tempv );
/* Copy updates from tempv[*] into lusup[*] */
isub = ufirst + nsupc;
for (i = 0; i < nrow; i++) {
lusup[isub] -= tempv[i]; /* Scatter-add */
tempv[i] = 0.0;
++isub;
}
#endif
} /* if fst_col < jcol ... */
return 0;
}
void
dgstrsL(char *trans, SuperMatrix *L, int *perm_r, SuperMatrix *B, int *info)
{
/*
* Purpose
* =======
*
* DGSTRSL only performs the L-solve using the LU factorization computed
* by DGSTRF.
*
* See supermatrix.h for the definition of 'SuperMatrix' structure.
*
* Arguments
* =========
*
* trans (input) char*
* Specifies the form of the system of equations:
* = 'N': A * X = B (No transpose)
* = 'T': A'* X = B (Transpose)
* = 'C': A**H * X = B (Conjugate transpose)
*
* L (input) SuperMatrix*
* The factor L from the factorization Pr*A*Pc=L*U as computed by
* dgstrf(). Use compressed row subscripts storage for supernodes,
* i.e., L has types: Stype = SLU_SC, Dtype = SLU_D, Mtype = SLU_TRLU.
*
* U (input) SuperMatrix*
* The factor U from the factorization Pr*A*Pc=L*U as computed by
* dgstrf(). Use column-wise storage scheme, i.e., U has types:
* Stype = SLU_NC, Dtype = SLU_D, Mtype = SLU_TRU.
*
* perm_r (input) int*, dimension (L->nrow)
* Row permutation vector, which defines the permutation matrix Pr;
* perm_r[i] = j means row i of A is in position j in Pr*A.
*
* B (input/output) SuperMatrix*
* B has types: Stype = SLU_DN, Dtype = SLU_D, Mtype = SLU_GE.
* On entry, the right hand side matrix.
* On exit, the solution matrix if info = 0;
*
* info (output) int*
* = 0: successful exit
* < 0: if info = -i, the i-th argument had an illegal value
*
*/
#ifdef _CRAY
_fcd ftcs1, ftcs2, ftcs3, ftcs4;
#endif
int incx = 1, incy = 1;
double alpha = 1.0, beta = 1.0;
DNformat *Bstore;
double *Bmat;
SCformat *Lstore;
double *Lval, *Uval;
int nrow, notran;
int fsupc, nsupr, nsupc, luptr, istart, irow;
int i, j, k, iptr, jcol, n, ldb, nrhs;
double *work, *work_col, *rhs_work, *soln;
flops_t solve_ops;
extern SuperLUStat_t SuperLUStat;
void dprint_soln();
/* Test input parameters ... */
*info = 0;
Bstore = B->Store;
ldb = Bstore->lda;
nrhs = B->ncol;
notran = lsame_(trans, "N");
if ( !notran && !lsame_(trans, "T") && !lsame_(trans, "C") ) *info = -1;
else if ( L->nrow != L->ncol || L->nrow < 0 ||
L->Stype != SLU_SC || L->Dtype != SLU_D || L->Mtype != SLU_TRLU )
*info = -2;
else if ( ldb < SUPERLU_MAX(0, L->nrow) ||
B->Stype != SLU_DN || B->Dtype != SLU_D || B->Mtype != SLU_GE )
*info = -4;
if ( *info ) {
i = -(*info);
xerbla_("dgstrsL", &i);
return;
}
n = L->nrow;
work = doubleCalloc(n * nrhs);
if ( !work ) ABORT("Malloc fails for local work[].");
soln = doubleMalloc(n);
if ( !soln ) ABORT("Malloc fails for local soln[].");
Bmat = Bstore->nzval;
Lstore = L->Store;
Lval = Lstore->nzval;
solve_ops = 0;
if ( notran ) {
/* Permute right hand sides to form Pr*B */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[perm_r[k]] = rhs_work[k];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
/* Forward solve PLy=Pb. */
for (k = 0; k <= Lstore->nsuper; k++) {
fsupc = L_FST_SUPC(k);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_START(fsupc+1) - istart;
nsupc = L_FST_SUPC(k+1) - fsupc;
nrow = nsupr - nsupc;
solve_ops += nsupc * (nsupc - 1) * nrhs;
solve_ops += 2 * nrow * nsupc * nrhs;
if ( nsupc == 1 ) {
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
luptr = L_NZ_START(fsupc);
for (iptr=istart+1; iptr < L_SUB_START(fsupc+1); iptr++){
irow = L_SUB(iptr);
++luptr;
rhs_work[irow] -= rhs_work[fsupc] * Lval[luptr];
}
}
} else {
luptr = L_NZ_START(fsupc);
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("N", strlen("N"));
ftcs3 = _cptofcd("U", strlen("U"));
STRSM( ftcs1, ftcs1, ftcs2, ftcs3, &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
SGEMM( ftcs2, ftcs2, &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#else
dtrsm_("L", "L", "N", "U", &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
dgemm_( "N", "N", &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#endif
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
work_col = &work[j*n];
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
rhs_work[irow] -= work_col[i]; /* Scatter */
work_col[i] = 0.0;
iptr++;
}
}
#else
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
dlsolve (nsupr, nsupc, &Lval[luptr], &rhs_work[fsupc]);
dmatvec (nsupr, nrow, nsupc, &Lval[luptr+nsupc],
&rhs_work[fsupc], &work[0] );
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
rhs_work[irow] -= work[i];
work[i] = 0.0;
iptr++;
}
}
#endif
} /* else ... */
} /* for L-solve */
#ifdef DEBUG
printf("After L-solve: y=\n");
dprint_soln(n, nrhs, Bmat);
#endif
SuperLUStat.ops[SOLVE] = solve_ops;
} else {
printf("Transposed solve not implemented.\n");
exit(0);
}
SUPERLU_FREE(work);
SUPERLU_FREE(soln);
}
/*! \brief
*
* <pre>
* Purpose
* =======
*
* dgstrsL only performs the L-solve using the LU factorization computed
* by DGSTRF.
*
* See supermatrix.h for the definition of 'SuperMatrix' structure.
*
* Arguments
* =========
*
* trans (input) char*
* Specifies the form of the system of equations:
* = 'N': A * X = B (No transpose)
* = 'T': A'* X = B (Transpose)
* = 'C': A**H * X = B (Conjugate transpose)
*
* L (input) SuperMatrix*
* The factor L from the factorization Pr*A*Pc=L*U as computed by
* dgstrf(). Use compressed row subscripts storage for supernodes,
* i.e., L has types: Stype = SLU_SC, Dtype = SLU_D, Mtype = SLU_TRLU.
*
* U (input) SuperMatrix*
* The factor U from the factorization Pr*A*Pc=L*U as computed by
* dgstrf(). Use column-wise storage scheme, i.e., U has types:
* Stype = SLU_NC, Dtype = SLU_D, Mtype = SLU_TRU.
*
* perm_r (input) int*, dimension (L->nrow)
* Row permutation vector, which defines the permutation matrix Pr;
* perm_r[i] = j means row i of A is in position j in Pr*A.
*
* B (input/output) SuperMatrix*
* B has types: Stype = SLU_DN, Dtype = SLU_D, Mtype = SLU_GE.
* On entry, the right hand side matrix.
* On exit, the solution matrix if info = 0;
*
* info (output) int*
* = 0: successful exit
* < 0: if info = -i, the i-th argument had an illegal value
* </pre>
*/
void
dgstrsL(char *trans, SuperMatrix *L, int *perm_r, SuperMatrix *B, int *info)
{
#ifdef _CRAY
_fcd ftcs1, ftcs2, ftcs3, ftcs4;
#endif
int incx = 1, incy = 1;
double alpha = 1.0, beta = 1.0;
DNformat *Bstore;
double *Bmat;
SCformat *Lstore;
double *Lval, *Uval;
int nrow, notran;
int fsupc, nsupr, nsupc, luptr, istart, irow;
int i, j, k, iptr, jcol, n, ldb, nrhs;
double *work, *work_col, *rhs_work, *soln;
flops_t solve_ops;
extern SuperLUStat_t SuperLUStat;
void dprint_soln();
/* Test input parameters ... */
*info = 0;
Bstore = B->Store;
ldb = Bstore->lda;
nrhs = B->ncol;
notran = lsame_(trans, "N");
if ( !notran && !lsame_(trans, "T") && !lsame_(trans, "C") ) *info = -1;
else if ( L->nrow != L->ncol || L->nrow < 0 ||
L->Stype != SLU_SC || L->Dtype != SLU_D || L->Mtype != SLU_TRLU )
*info = -2;
else if ( ldb < SUPERLU_MAX(0, L->nrow) ||
B->Stype != SLU_DN || B->Dtype != SLU_D || B->Mtype != SLU_GE )
*info = -4;
if ( *info ) {
i = -(*info);
xerbla_("dgstrsL", &i);
return;
}
n = L->nrow;
work = doubleCalloc(n * nrhs);
if ( !work ) ABORT("Malloc fails for local work[].");
soln = doubleMalloc(n);
if ( !soln ) ABORT("Malloc fails for local soln[].");
Bmat = Bstore->nzval;
Lstore = L->Store;
Lval = Lstore->nzval;
solve_ops = 0;
if ( notran ) {
/* Permute right hand sides to form Pr*B */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[perm_r[k]] = rhs_work[k];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
/* Forward solve PLy=Pb. */
for (k = 0; k <= Lstore->nsuper; k++) {
fsupc = L_FST_SUPC(k);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_START(fsupc+1) - istart;
nsupc = L_FST_SUPC(k+1) - fsupc;
nrow = nsupr - nsupc;
solve_ops += nsupc * (nsupc - 1) * nrhs;
solve_ops += 2 * nrow * nsupc * nrhs;
if ( nsupc == 1 ) {
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
luptr = L_NZ_START(fsupc);
for (iptr=istart+1; iptr < L_SUB_START(fsupc+1); iptr++){
irow = L_SUB(iptr);
++luptr;
rhs_work[irow] -= rhs_work[fsupc] * Lval[luptr];
}
}
} else {
luptr = L_NZ_START(fsupc);
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("N", strlen("N"));
ftcs3 = _cptofcd("U", strlen("U"));
STRSM( ftcs1, ftcs1, ftcs2, ftcs3, &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
SGEMM( ftcs2, ftcs2, &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#else
dtrsm_("L", "L", "N", "U", &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
dgemm_( "N", "N", &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#endif
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
work_col = &work[j*n];
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
rhs_work[irow] -= work_col[i]; /* Scatter */
work_col[i] = 0.0;
iptr++;
}
}
#else
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
dlsolve (nsupr, nsupc, &Lval[luptr], &rhs_work[fsupc]);
dmatvec (nsupr, nrow, nsupc, &Lval[luptr+nsupc],
&rhs_work[fsupc], &work[0] );
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
rhs_work[irow] -= work[i];
work[i] = 0.0;
iptr++;
}
}
#endif
} /* else ... */
} /* for L-solve */
#ifdef DEBUG
printf("After L-solve: y=\n");
dprint_soln(n, nrhs, Bmat);
#endif
SuperLUStat.ops[SOLVE] = solve_ops;
} else {
printf("Transposed solve not implemented.\n");
exit(0);
}
SUPERLU_FREE(work);
SUPERLU_FREE(soln);
}