int
pcgstrf_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 */
complex *dense, /* modified */
complex *tempv, /* working array */
pxgstrf_shared_t *pxgstrf_shared, /* modified */
Gstat_t *Gstat /* modified */
)
{
/*
* -- SuperLU MT routine (version 2.0) --
* Lawrence Berkeley National Lab, Univ. of California Berkeley,
* and Xerox Palo Alto Research Center.
* September 10, 2007
*
* 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;
complex 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
*/
complex 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;
complex *lusup;
int *xlusup, *xlusup_end;
complex *tempv1;
int mem_error;
register float flopcnt;
complex zero = {0.0, 0.0};
complex one = {1.0, 0.0};
complex none = {-1.0, 0.0};
complex comp_temp, comp_temp1;
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) pcgstrf_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 = 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_end[fsupc] - xlsub[fsupc]; /* Leading dimension */
nrow = nsupr - d_fsupc - nsupc;
krep_ind = lptr + nsupc - 1;
flopcnt = segsze * (segsze - 1) + 2 * nrow * segsze;//sj
Gstat->procstat[pnum].fcops += flopcnt;
#if ( DEBUGlevel>=2 )
if (jcol==BADCOL)
printf("(%d) pcgstrf_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];
cc_mult(&comp_temp, &ukj, &lusup[luptr]);
c_sub(&dense[irow], &dense[irow], &comp_temp);
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 */
cc_mult(&comp_temp, &ukj1, &lusup[luptr1]);
c_sub(&ukj, &ukj, &comp_temp);
dense[lsub[krep_ind]] = ukj;
for (i = lptr + nsupc; i < xlsub_end[fsupc]; ++i) {
irow = lsub[i];
luptr++;
luptr1++;
cc_mult(&comp_temp, &ukj, &lusup[luptr]);
cc_mult(&comp_temp1, &ukj1, &lusup[luptr1]);
c_add(&comp_temp, &comp_temp, &comp_temp1);
c_sub(&dense[irow], &dense[irow], &comp_temp);
}
} else { /* Case 3: 3cols-col update */
ukj2 = dense[lsub[krep_ind - 2]];
luptr2 = luptr1 - nsupr;
cc_mult(&comp_temp, &ukj2, &lusup[luptr2-1]);
c_sub(&ukj1, &ukj1, &comp_temp);
cc_mult(&comp_temp, &ukj1, &lusup[luptr1]);
cc_mult(&comp_temp1, &ukj2, &lusup[luptr2]);
c_add(&comp_temp, &comp_temp, &comp_temp1);
c_sub(&ukj, &ukj, &comp_temp);
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++;
cc_mult(&comp_temp, &ukj, &lusup[luptr]);
cc_mult(&comp_temp1, &ukj1, &lusup[luptr1]);
c_add(&comp_temp, &comp_temp, &comp_temp1);
cc_mult(&comp_temp1, &ukj2, &lusup[luptr2]);
c_add(&comp_temp, &comp_temp, &comp_temp1);
c_sub(&dense[irow], &dense[irow], &comp_temp);
}
}
} 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 )
CTRSV( ftcs1, ftcs2, ftcs3, &segsze, &lusup[luptr],
&nsupr, tempv, &incx );
#else
ctrsv_( "L", "N", "U", &segsze, &lusup[luptr],
&nsupr, tempv, &incx );
#endif
luptr += segsze; /* Dense matrix-vector */
tempv1 = &tempv[segsze];
alpha = one;
beta = zero;
#if ( MACH==CRAY_PVP )
CGEMV( ftcs2, &nrow, &segsze, &alpha, &lusup[luptr],
&nsupr, tempv, &incx, &beta, tempv1, &incy );
#else
cgemv_( "N", &nrow, &segsze, &alpha, &lusup[luptr],
&nsupr, tempv, &incx, &beta, tempv1, &incy );
#endif
#else
clsolve ( nsupr, segsze, &lusup[luptr], tempv );
luptr += segsze; /* Dense matrix-vector */
tempv1 = &tempv[segsze];
cmatvec (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] = zero;
isub++;
}
/* Scatter tempv1[] into SPA dense[*] */
for (i = 0; i < nrow; i++) {
irow = lsub[isub];
c_sub(&dense[irow], &dense[irow], &tempv1[i]);
tempv1[i] = zero;
++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] = zero;
#ifdef DEBUG
if (jcol == -1)
printf("(%d) pcgstrf_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 = 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_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) pcgstrf_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; //sj
Gstat->procstat[pnum].fcops += flopcnt;
/* ops[TRSV] += nsupc * (nsupc - 1);
ops[GEMV] += 2 * nrow * nsupc; */
#ifdef USE_VENDOR_BLAS
alpha = none; beta = one; /* y := beta*y + alpha*A*x */
#if ( MACH==CRAY_PVP )
CTRSV( ftcs1, ftcs2, ftcs3, &nsupc, &lusup[luptr],
&nsupr, &lusup[ufirst], &incx );
CGEMV( ftcs2, &nrow, &nsupc, &alpha, &lusup[luptr+nsupc], &nsupr,
&lusup[ufirst], &incx, &beta, &lusup[ufirst+nsupc], &incy );
#else
ctrsv_( "L", "N", "U", &nsupc, &lusup[luptr],
&nsupr, &lusup[ufirst], &incx );
cgemv_( "N", &nrow, &nsupc, &alpha, &lusup[luptr+nsupc], &nsupr,
&lusup[ufirst], &incx, &beta, &lusup[ufirst+nsupc], &incy );
#endif
#else
clsolve ( nsupr, nsupc, &lusup[luptr], &lusup[ufirst] );
cmatvec ( nsupr, nrow, nsupc, &lusup[luptr+nsupc],
&lusup[ufirst], tempv );
/* Copy updates from tempv[*] into lusup[*] */
isub = ufirst + nsupc;
for (i = 0; i < nrow; i++) {
c_sub(&lusup[isub], &lusup[isub], &tempv[i]);
tempv[i] = zero;
++isub;
}
#endif
} /* if fst_col < jcol ... */
return 0;
}
void
pcgstrf_mark_busy_descends(int pnum, int jcol, int *etree,
//pcgstrf_shared_t *pxgstrf_shared,
pxgstrf_shared_t *pxgstrf_shared, //sj
int *bcol, int *lbusy)
{
/*
* -- SuperLU MT routine (version 1.0) --
* Univ. of California Berkeley, Xerox Palo Alto Research Center,
* and Lawrence Berkeley National Lab.
* August 15, 1997
*
* Purpose
* =======
*
* Mark busy panels in local "lbusy" array, used for linear pipelining.
*
* When jcol begins, its busy descendant panels (if any) are bcol and
* all the e-tree ancestors of bcol between bcol and jcol. This routine
* marks those columns in the array lbusy, which is local to this
* processor, to preserve a snapshot regardless of what the other
* processors are doing meanwhile.
*
* Arguments
* =========
*
* jcol (input) int
* Current panel, with leading column jcol.
*
* etree (input) int*
* Elimination tree parent pointers.
*
* bcol (input/output) int*
* Farthest busy descendant of jcol.
* On entry, it is the first column of the farthest busy panel.
* On exit, it may be adjusted to the first column of the
* farthest busy supernode.
*
* lbusy (input/output) int*
* Initially all -1, lbusy(r) = jcol means that r was busy
* at the beginning of computing jcol.
*
*/
pxgstrf_shared_t *pcgstrf_shared = pxgstrf_shared; //sj
GlobalLU_t *Glu = pcgstrf_shared->Glu;
register int w, kcol, fsupc, bcol_reg;
int *xsup;
bcol_reg = *bcol;
if ( bcol_reg < jcol ) {
/* -----------------------------------------------------------
Instead of waiting for the completion of "bcol", we can
pessimistically assume supno[bcol] == supno[bcol-1],
hence always mark as busy the supernode containing "bcol-1".
----------------------------------------------------------- */
if (pcgstrf_shared->pan_status[bcol_reg].type == RELAXED_SNODE) {
#if 0
if ( pcgstrf_shared->pan_status[bcol_reg].size < 0 )
fsupc = bcol_reg + pcgstrf_shared->pan_status[bcol_reg].size;
else fsupc = bcol_reg;
#endif
fsupc = bcol_reg;
w = pcgstrf_shared->pan_status[fsupc].size;
bcol_reg += w;
for (kcol = fsupc; kcol < bcol_reg; ++kcol)
lbusy[kcol] = jcol;
} else {
/* Find leading column "fsupc" in the supernode that
contains column "bcol-1" */
#if 0
if ( pcgstrf_shared->spin_locks[bcol_reg] ) /* WORSE PERFORMANCE!! */
await( &pcgstrf_shared->spin_locks[bcol_reg] );
#endif
xsup = Glu->xsup;
fsupc = SUPER_FSUPC ( Glu->supno[bcol_reg-1] );
for (kcol = fsupc; kcol < bcol_reg; ++kcol) lbusy[kcol] = jcol;
}
#if ( DEBUGlevel>=1 )
if (jcol >= LOCOL && jcol <= HICOL)
printf("(%d) mark_busy_descends[1] jcol %d, bcol_reg %d, fsupc %d\n",
pnum, jcol, bcol_reg, fsupc);
#endif
/* Mark as busy all columns on the path between bcol_reg and jcol */
for (kcol = bcol_reg; kcol < jcol; kcol = etree[kcol]) {
lbusy[kcol] = jcol;
}
/* INVARIANT: *bcol must be the first column of the farthest
busy supernode */
*bcol = fsupc;
} /* if bcol_reg < jcol */
}
int
pzgstrf_column_dfs(
const int pnum, /* process number */
const int m, /* number of rows in the matrix */
const int jcol, /* current column in the panel */
const int fstcol, /* first column in the panel */
int *perm_r, /* row pivotings that are done so far */
int *ispruned, /* in */
int *col_lsub, /* the RHS vector to start the dfs */
int lsub_end, /* size of col_lsub[] */
int *super_bnd,/* supernode partition by upper bound */
int *nseg, /* modified - with new segments appended */
int *segrep, /* modified - with new segments appended */
int *repfnz, /* modified */
int *xprune, /* modified */
int *marker2, /* modified */
int *parent, /* working array */
int *xplore, /* working array */
pxgstrf_shared_t *pxgstrf_shared /* modified */
)
{
/*
* -- SuperLU MT routine (version 2.0) --
* Lawrence Berkeley National Lab, Univ. of California Berkeley,
* and Xerox Palo Alto Research Center.
* September 10, 2007
*
* Purpose
* =======
* pzgstrf_column_dfs() performs a symbolic factorization on column jcol,
* and detects whether column jcol belongs in the same supernode as jcol-1.
*
* Local parameters
* ================
* A supernode representative is the last column of a supernode.
* The nonzeros in U[*,j] are segments that end at supernodal
* representatives. The routine returns a list of such supernodal
* representatives in topological order of the dfs that generates them.
* The location of the first nonzero in each such supernodal segment
* (supernodal entry location) is also returned.
*
* nseg: no of segments in current U[*,j]
* samesuper: samesuper=NO if column j does not belong in the same
* supernode as j-1. Otherwise, samesuper=YES.
*
* marker2: A-row --> A-row/col (0/1)
* repfnz: SuperA-col --> PA-row
* parent: SuperA-col --> SuperA-col
* xplore: SuperA-col --> index to L-structure
*
* Return value
* ============
* 0 success;
* > 0 number of bytes allocated when run out of space.
*
*/
GlobalLU_t *Glu = pxgstrf_shared->Glu; /* modified */
Gstat_t *Gstat = pxgstrf_shared->Gstat; /* modified */
register int jcolm1, jcolm1size, nextl, ifrom;
register int k, krep, krow, kperm, samesuper, nsuper;
register int no_lsub;
int fsupc; /* first column in a supernode */
int myfnz; /* first nonz column in a U-segment */
int chperm, chmark, chrep, kchild;
int xdfs, maxdfs, kpar;
int ito; /* Used to compress row subscripts */
int mem_error;
int *xsup, *xsup_end, *supno, *lsub, *xlsub, *xlsub_end;
static int first = 1, maxsuper;
if ( first ) {
maxsuper = sp_ienv(3);
first = 0;
}
/* Initialize pointers */
xsup = Glu->xsup;
xsup_end = Glu->xsup_end;
supno = Glu->supno;
lsub = Glu->lsub;
xlsub = Glu->xlsub;
xlsub_end = Glu->xlsub_end;
jcolm1 = jcol - 1;
nextl = lsub_end;
no_lsub = 0;
samesuper = YES;
/* Test whether the row structure of column jcol is contained
in that of column jcol-1. */
for (k = 0; k < lsub_end; ++k) {
krow = col_lsub[k];
if ( perm_r[krow] == EMPTY ) { /* krow is in L */
++no_lsub;
if (marker2[krow] != jcolm1)
samesuper = NO; /* row subset test */
marker2[krow] = jcol;
}
}
#if ( DEBUGlevel>=2 )
if (jcol == BADCOL)
printf("(%d) pzgstrf_column_dfs[1] %d, fstcol %d, lsub_end %d, no_lsub %d, samesuper? %d\n",
pnum, jcol, fstcol, lsub_end, no_lsub, samesuper);
#endif
/*
* For each nonzero in A[fstcol:n,jcol] perform DFS ...
*/
for (k = 0; k < lsub_end; ++k) {
krow = col_lsub[k];
/* if krow was visited before, go to the next nonzero */
if ( marker2[krow] == jcol ) continue;
marker2[krow] = jcol;
kperm = perm_r[krow];
#if ( DEBUGlevel>=3 )
if (jcol == BADCOL)
printf("(%d) pzgstrf_column_dfs[inner]: perm_r[krow=%d] %d\n", pnum, krow, kperm);
#endif
/* Ignore the nonzeros in U corresponding to the busy columns
during the panel DFS. */
/*if ( lbusy[kperm] != fstcol ) { xiaoye? */
if ( kperm >= fstcol ) {
/*
* krow is in U: if its supernode representative krep
* has been explored, update repfnz[*].
*/
krep = SUPER_REP(supno[kperm]);
myfnz = repfnz[krep];
#if ( DEBUGlevel>=3 )
if (jcol == BADCOL)
printf("(%d) pzgstrf_column_dfs[inner-U]: krep %d, myfnz %d, kperm %d\n",
pnum, krep, myfnz, kperm);
#endif
if ( myfnz != EMPTY ) { /* Visited before */
if ( myfnz > kperm ) repfnz[krep] = kperm;
/* continue; */
} else {
/* Otherwise, perform dfs starting at krep */
parent[krep] = EMPTY;
repfnz[krep] = kperm;
if ( ispruned[krep] ) {
if ( SINGLETON( supno[krep] ) )
xdfs = xlsub_end[krep];
else xdfs = xlsub[krep];
maxdfs = xprune[krep];
#ifdef PROFILE
Gstat->procstat[pnum].pruned++;
#endif
} else {
fsupc = SUPER_FSUPC( supno[krep] );
xdfs = xlsub[fsupc] + krep-fsupc+1;
maxdfs = xlsub_end[fsupc];
#ifdef PROFILE
Gstat->procstat[pnum].unpruned++;
#endif
}
do {
/*
* For each unmarked kchild of krep ...
*/
while ( xdfs < maxdfs ) {
kchild = lsub[xdfs];
xdfs++;
chmark = marker2[kchild];
if ( chmark != jcol ) { /* Not reached yet */
marker2[kchild] = jcol;
chperm = perm_r[kchild];
if ( chperm == EMPTY ) {
/* kchild is in L: place it in L[*,k]. */
++no_lsub;
col_lsub[nextl++] = kchild;
if (chmark != jcolm1) samesuper = NO;
} else {
/* kchild is in U: chrep = its supernode
* representative. If its rep has
* been explored, update its repfnz[*].
*/
chrep = SUPER_REP( supno[chperm] );
myfnz = repfnz[chrep];
if ( myfnz != EMPTY ) { /* Visited before */
if ( myfnz > chperm )
repfnz[chrep] = chperm;
} else {
/* Continue dfs at super-rep of kchild */
xplore[krep] = xdfs;
xplore[m + krep] = maxdfs;
parent[chrep] = krep;
krep = chrep; /* Go deeper down G(L^t) */
repfnz[krep] = chperm;
if ( ispruned[krep] ) {
if ( SINGLETON( supno[krep] ) )
xdfs = xlsub_end[krep];
else xdfs = xlsub[krep];
maxdfs = xprune[krep];
#ifdef PROFILE
Gstat->procstat[pnum].pruned++;
#endif
} else {
fsupc = SUPER_FSUPC( supno[krep] );
xdfs = xlsub[fsupc] + krep-fsupc+1;
maxdfs = xlsub_end[fsupc];
#ifdef PROFILE
Gstat->procstat[pnum].unpruned++;
#endif
}
}
} /* else */
} /* if */
} /* while */
/* krow has no more unexplored nbrs:
* place supernode-rep krep in postorder DFS,
* backtrack dfs to its parent.
*/
segrep[*nseg] = krep;
++(*nseg);
#if ( DEBUGlevel>=3 )
if (jcol == BADCOL)
printf("(%d) pzgstrf_column_dfs[inner-dfs] new nseg %d, repfnz[krep=%d] %d\n",
pnum, *nseg, krep, repfnz[krep]);
#endif
kpar = parent[krep]; /* Pop from stack, mimic recursion */
if ( kpar == EMPTY ) break; /* dfs done */
krep = kpar;
xdfs = xplore[krep];
maxdfs = xplore[m + krep];
} while ( kpar != EMPTY ); /* Do ... until empty stack */
} /* else myfnz ... */
} /* if kperm >= fstcol ... */
} /* for each nonzero ... */
#if ( DEBUGlevel>=3 )
if (jcol == BADCOL)
printf("(%d) pzgstrf_column_dfs[2]: nextl %d, samesuper? %d\n",
pnum, nextl, samesuper);
#endif
/* assert(no_lsub == nextl - no_usub);*/
/* ---------------------------------------------------------
Check to see if j belongs in the same supernode as j-1.
--------------------------------------------------------- */
/* Does it matter if jcol == 0? - xiaoye */
if ( samesuper == YES ) {
nsuper = supno[jcolm1];
jcolm1size = xlsub_end[jcolm1] - xlsub[jcolm1];
#if ( DEBUGlevel>=3 )
if (jcol == BADCOL)
printf("(%d) pzgstrf_column_dfs[YES] jcol-1 %d, jcolm1size %d, supno[%d] %d\n",
pnum, jcolm1, jcolm1size, jcolm1, nsuper);
#endif
if ( no_lsub != jcolm1size-1 )
samesuper = NO; /* Enforce T2 supernode */
else {
/* Make sure the number of columns in a supernode does not
exceed threshold. */
fsupc = xsup[nsuper];
if ( jcol - fsupc >= maxsuper )
samesuper = NO;
else {
/* start of a supernode in H (coarser partition) */
if ( super_bnd[jcol] != 0 ) samesuper = NO;
}
}
}
/* If jcol starts a new supernode, allocate storage for
* the subscript set of both first and last column of
* a previous supernode. (first for num values, last for pruning)
*/
if ( samesuper == NO ) { /* starts a new supernode */
nsuper = NewNsuper(pnum, pxgstrf_shared, &Glu->nsuper);
xsup[nsuper] = jcol;
/* Copy column jcol; also reserve space to store pruned graph */
if ((mem_error = Glu_alloc(pnum, jcol, 2*no_lsub, LSUB, &ito,
pxgstrf_shared)))
return mem_error;
xlsub[jcol] = ito;
lsub = Glu->lsub;
for (ifrom = 0; ifrom < nextl; ++ifrom) {
krow = col_lsub[ifrom];
if ( perm_r[krow] == EMPTY ) /* Filter U-subscript */
lsub[ito++] = krow;
}
k = ito;
xlsub_end[jcol] = k;
/* Make a copy in case it is a singleton supernode */
for (ifrom = xlsub[jcol]; ifrom < ito; ++ifrom)
lsub[k++] = lsub[ifrom];
} else { /* Supernode of size > 1: overwrite column jcol-1 */
k = xlsub_end[fsupc];
xlsub[jcol] = k;
xprune[fsupc] = k;
for (ifrom = 0; ifrom < nextl; ++ifrom) {
krow = col_lsub[ifrom];
if ( perm_r[krow] == EMPTY ) /* Filter U-subscript */
lsub[k++] = krow;
}
xlsub_end[jcol] = k;
}
#if ( DEBUGlevel>=3 )
if (jcol == BADCOL) {
printf("(%d) pzgstrf_column_dfs[3]: %d in prev s-node %d? %d\n",
pnum, jcol, fsupc, samesuper);
PrintInt10("lsub", xlsub_end[jcol]-xlsub[jcol], &lsub[xlsub[jcol]]);
}
#endif
/* Tidy up the pointers before exit */
xprune[jcol] = k; /* upper bound for pruning */
supno[jcol] = nsuper;
xsup_end[nsuper] = jcol + 1;
return 0;
}
void
pxgstrf_super_bnd_dfs(
const int pnum, /* process number */
const int m, /* number of rows in the matrix */
const int n, /* number of columns in the matrix */
const int jcol, /* first column of the H-supernode */
const int w, /* size of the H-supernode */
SuperMatrix *A, /* original matrix */
int *perm_r, /* in */
int *iperm_r, /* in; inverse of perm_r */
int *xprune, /* in */
int *ispruned, /* in */
int *marker, /* modified */
int *parent, /* working array */
int *xplore, /* working array */
pxgstrf_shared_t *pxgstrf_shared /* 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 a symbolic structure prediction on a supernode in the Householder
* matrix H, with jcol being the leading column.
*
*/
GlobalLU_t *Glu = pxgstrf_shared->Glu; /* modified */
register int krep, chperm, chrep, kchild;
register int invp_rep; /* "krep" numbered in the original A */
register int krow, kperm, xdfs, maxdfs, kpar;
register int fsupc, k, jj, found;
register int nrow; /* union of the nonzero rows in a supernode */
NCPformat *Astore;
int *asub, *xa_begin, *xa_end;
int *xsup, *xsup_end, *supno, *lsub, *xlsub, *xlsub_end;
/* Initialize pointers */
xsup = Glu->xsup;
xsup_end = Glu->xsup_end;
supno = Glu->supno;
lsub = Glu->lsub;
xlsub = Glu->xlsub;
xlsub_end = Glu->xlsub_end;
Astore = A->Store;
asub = Astore->rowind;
xa_begin = Astore->colbeg;
xa_end = Astore->colend;
nrow = 0;
found = n + jcol;
/* For each column in the H-supernode */
for (jj = jcol; jj < jcol + w; ++jj) {
/* For each nonz in A[*,jj] do dfs */
for (k = xa_begin[jj]; k < xa_end[jj]; ++k) {
krow = asub[k];
/* krow was visited before, go to the next nonzero. */
if ( marker[krow] == found ) continue;
/* For each unmarked nbr krow of jj ... */
kperm = perm_r[krow];
if ( kperm == EMPTY ) { /* krow is in L */
marker[krow] = found;
++nrow;
} else {
/* krow is in U: if its supernode-rep krep has been explored,
skip the search. */
krep = SUPER_REP( supno[kperm] );
invp_rep = iperm_r[krep];
/* Perform dfs starting at krep */
if ( marker[invp_rep] != found ) {
marker[invp_rep] = found;
parent[krep] = EMPTY;
if ( ispruned[krep] ) {
if ( SINGLETON( supno[krep] ) )
xdfs = xlsub_end[krep];
else xdfs = xlsub[krep];
maxdfs = xprune[krep];
} else {
fsupc = SUPER_FSUPC( supno[krep] );
xdfs = xlsub[fsupc] + krep-fsupc+1;
maxdfs = xlsub_end[fsupc];
}
do {
/* For each unmarked kchild of krep ... */
while ( xdfs < maxdfs ) {
kchild = lsub[xdfs];
xdfs++;
if (marker[kchild] != found) { /* Not reached yet */
chperm = perm_r[kchild];
if ( chperm == EMPTY ) { /* kchild is in L */
marker[kchild] = found;
++nrow;
} else {
/* kchild is in U:
* chrep = its supernode-rep. If its rep
* has been explored, skip the search.
*/
chrep = SUPER_REP( supno[chperm] );
invp_rep = iperm_r[chrep];
/* Continue dfs at snode-rep of kchild */
if ( marker[invp_rep] != found ) {
marker[invp_rep] = found;
xplore[krep] = xdfs;
xplore[m + krep] = maxdfs;
parent[chrep] = krep;
krep = chrep;/* Go deeper down G(L^t) */
xdfs = xlsub[krep];
maxdfs = xprune[krep];
if ( ispruned[krep] ) {
if ( SINGLETON( supno[krep] ) )
xdfs = xlsub_end[krep];
else xdfs = xlsub[krep];
maxdfs = xprune[krep];
} else {
fsupc = SUPER_FSUPC(supno[krep]);
xdfs = xlsub[fsupc] + krep-fsupc+1;
maxdfs = xlsub_end[fsupc];
}
} /* if */
} /* else */
} /* if... */
} /* while xdfs < maxdfs */
/* krow has no more unexplored nbrs:
* Place snode-rep krep in postorder dfs, if this
* segment is seen for the first time. Note that
* the "repfnz[krep]" may change later.
* Backtrack dfs to its parent.
*/
kpar = parent[krep]; /* Pop stack, mimic recursion */
if ( kpar == EMPTY ) break; /* dfs done */
krep = kpar;
xdfs = xplore[krep];
maxdfs = xplore[m+krep];
} while ( kpar != EMPTY ); /* do-while - until empty stack */
} /* if */
} /* else */
} /* for each nonz in A[*,jj] */
} /* for jj ... */
DynamicSetMap(pnum, jcol, nrow * w, pxgstrf_shared);
/* for (i = 1; i < w; ++i) Glu->map_in_sup[jcol + i] = -i;*/
#if ( DEBUGlevel>=1 )
printf("(%d) pxgstrf_super_bnd_dfs(): jcol= %d, w= %d, nrow= %d\n",
pnum, jcol, w, nrow);
#endif
}
void
pdgstrf_panel_dfs(
const int pnum, /* process number */
const int m, /* number of rows in the matrix */
const int w, /* current panel width */
const int jcol, /* leading column of the current panel */
SuperMatrix *A, /* original matrix */
int *perm_r, /* row pivotings that are done so far */
int *xprune, /* in */
int *ispruned, /* in */
int *lbusy, /* in; size n */
int *nseg, /* out */
int *panel_lsub, /* out */
int *w_lsub_end, /* out; values irrelevant on entry */
int *segrep, /* out */
int *repfnz, /* out */
int *marker, /* modified */
int *spa_marker, /* modified; size n-by-w */
int *parent, /* working array */
int *xplore, /* working array */
double *dense, /* out; size n-by-w */
GlobalLU_t *Glu /* modified */
)
{
/*
* -- SuperLU MT routine (version 2.0) --
* Lawrence Berkeley National Lab, Univ. of California Berkeley,
* and Xerox Palo Alto Research Center.
* September 10, 2007
*
* Purpose
* =======
*
* Performs a symbolic factorization on a panel of columns [jcol, jcol+w).
* It skips all those busy descendants that are worked on by other
* processors along the e-tree path.
*
* Notes
* =====
*
* (1) panel_lsub[0:w*n-1]: temporary for the nonzero row indices below
* the panel diagonal, which will be used later in the inner LU
* factorization. For the busy columns, some of the nonzeros in U
* may be mistakenly placed in this list, because "perm_r" is
* still "empty". Later, during dcolumn_dfs in the inner factorization,
* we must filter those nonzeros belonging in U.
*
* (2) A supernode representative is the last column of a supernode.
* The nonzeros in U[*,j] are segments that end at supernodal
* representatives.
*
* (3) The routine returns one list of the supernodal representatives
* in topological order of the DFS that generates them. This list is
* a superset of the topological order of each individual column within
* the panel. The location of the first nonzero in each supernodal
* segment (supernodal entry location) is also returned. Each column
* has a separate list for this purpose.
*
* (4) Two marker arrays are used to facilitate dfs:
* marker[i] == jj, if i was visited during dfs of current column jj;
* marker1[i] == jcol, if i was visited by earlier columns in this panel;
*
* (5) The dfs stack is the combination of xplore[2*m] and parent[m]:
* xplore[k] - pointer to k's adjancency list where search begins
* xplore[m + k] - pointer to k's adjancency list where search ends
*
* (6) Array mappings
* marker: A-row --> A-row/col (0/1)
* repfnz: SuperA-col --> PA-row
* parent: SuperA-col --> SuperA-col
* xplore: SuperA-col --> index to L-structure
*
*/
NCPformat *Astore;
double *a;
int *asub;
int *xa_begin, *xa_end;
register int krep, chperm, chmark, chrep, kchild, myfnz;
register int k, krow, kmark, kperm, fsupc;
register int xdfs, maxdfs, kpar, jj, nextp;
register int nextl_col;/* next open position in panel_lsub[*,jj] */
int *marker1; /* marker1[jj] == jcol if vertex jj was visited
by a previous column within this panel. */
int *repfnz_col; /* start of each column in the panel */
double *dense_col; /* start of each column in the panel */
int *xsup, *xsup_end, *supno, *lsub, *xlsub, *xlsub_end;
int *col_marker; /* marker array of each column in the panel */
/* Initialize pointers */
xsup = Glu->xsup;
xsup_end = Glu->xsup_end;
supno = Glu->supno;
lsub = Glu->lsub;
xlsub = Glu->xlsub;
xlsub_end = Glu->xlsub_end;
Astore = A->Store;
a = Astore->nzval;
asub = Astore->rowind;
xa_begin = Astore->colbeg;
xa_end = Astore->colend;
marker1 = marker + m;
repfnz_col = repfnz;
dense_col = dense;
nextp = 0;
*nseg = 0;
#if ( DEBUGlevel>=2 )
if (jcol == BADPAN)
printf("(%d) pdgstrf_panel_dfs[begin] jcol %d, w %d\n", pnum, jcol, w);
#endif
/*
* For each column in the panel ...
*/
for (jj = jcol; jj < jcol + w; ++jj, nextp += m) {
nextl_col = nextp;
col_marker = &spa_marker[nextp];
/*
* For each nonz in A[*,jj] perform dfs ...
*/
for (k = xa_begin[jj]; k < xa_end[jj]; ++k) {
krow = asub[k];
dense_col[krow] = a[k];
kmark = col_marker[krow];
/* if krow was visited before, go to the next nonzero */
if ( kmark == jj ) continue;
/*
* For each unmarked nbr krow of jj ...
*/
col_marker[krow] = jj;
kperm = perm_r[krow];
if ( kperm == EMPTY ) {
/* krow is in L: place it in structure of L[*,jj].
* NOTE: some entries in U may get here, because "perm_r"
* is not yet available from a preceeding busy column.
*/
panel_lsub[nextl_col++] = krow; /* krow is indexed into A */
} else {
/*
* krow is in U (0 <= kperm < jcol): if its supernode
* representative krep has been explored, update repfnz[*].
*/
if ( lbusy[kperm] == jcol ) { /* kperm is busy */
#if ( DEBUGlevel>=3 )
if (jj == BADCOL)
printf("(%d) pdgstrf_panel_dfs(%d) skip busy krow %d, kperm %d\n",
pnum, jj, krow, kperm);
#endif
continue;
}
/* Here, krep cannot possibly be "busy" */
krep = SUPER_REP( supno[kperm] );
myfnz = repfnz_col[krep];
#ifdef CHK_DFS
if (jj == BADCOL)
printf("(%d) pdgstrf_panel_dfs[1] %d, krep %d, fsupc %d, Pr[krow %d] %d, myfnz %d\n",
pnum, jj, krep, SUPER_FSUPC(supno[krep]), krow, kperm, myfnz);
#endif
if ( myfnz != EMPTY ) { /* Representative visited before */
if ( myfnz > kperm ) repfnz_col[krep] = kperm;
/* continue; */
} else {
/* Otherwise, performs dfs starting from krep */
parent[krep] = EMPTY;
repfnz_col[krep] = kperm;
if ( ispruned[krep] ) {
if ( SINGLETON( supno[krep] ) )
xdfs = xlsub_end[krep];
else xdfs = xlsub[krep];
maxdfs = xprune[krep];
#ifdef PROFILE
/*Gstat->procstat[pnum].pruned++;*/
#endif
} else {
fsupc = SUPER_FSUPC( supno[krep] );
xdfs = xlsub[fsupc] + krep-fsupc+1;
maxdfs = xlsub_end[fsupc];
#ifdef PROFILE
/*Gstat->procstat[pnum].unpruned++;*/
#endif
}
#ifdef CHK_DFS
if (jj == BADCOL)
{
register int i;
printf("(%d) pdgstrf_panel_dfs[2] %d, ispruned[%d] %d, xdfs %d, maxdfs %d\n",
pnum, jj, krep, ispruned[krep], xdfs, maxdfs);
/*for (i = xdfs; i < maxdfs; i++) printf("(%d) lsub-%d", pnum, lsub[i]);*/
printf("\n");
}
#endif
do {
while ( xdfs < maxdfs ) {
/* for each unmarked kchild of krep ... */
kchild = lsub[xdfs];
xdfs++;
chmark = col_marker[kchild];
if ( chmark != jj ) { /* Not reached yet */
col_marker[kchild] = jj;
chperm = perm_r[kchild];
if ( chperm == EMPTY ) {
/* kchild is in L: place it in L[*,j]. */
panel_lsub[nextl_col++] = kchild;
} else {
/* kchild is in U (0 <= chperm < jcol):
* chrep = its supernode-rep. If its rep
* has been explored, update its repfnz[*].
*/
if ( lbusy[chperm] == jcol ) {
#ifdef DEBUG
if (jj == BADCOL)
printf("(%d) pdgstrf_panel_dfs(%d) skip busy kchild %d, chperm %d\n",
pnum, jj, kchild, chperm);
#endif
continue;
}
chrep = SUPER_REP( supno[chperm] );
myfnz = repfnz_col[chrep];
#ifdef DEBUG
if (jj == BADCOL)
printf("(%d) pdgstrf_panel_dfs[3] %d, krep %d, Pr[kchild %d] %d, chrep %d, fsupc %d, myfnz %d\n",
pnum, jj, krep, kchild, chperm, chrep,
SUPER_FSUPC(supno[chrep]), myfnz);
#endif
if ( myfnz != EMPTY ) {/* Visited before */
if ( myfnz > chperm )
repfnz_col[chrep] = chperm;
} else {
/* Cont. dfs at snode-rep of kchild */
xplore[krep] = xdfs;
xplore[m + krep] = maxdfs;
parent[chrep] = krep;
krep = chrep; /* Go deeper down G(L) */
repfnz_col[krep] = chperm;
if ( ispruned[krep] ) {
if ( SINGLETON( supno[krep] ) )
xdfs = xlsub_end[krep];
else xdfs = xlsub[krep];
maxdfs = xprune[krep];
#ifdef PROFILE
/*procstat[pnum].pruned++;*/
#endif
} else {
fsupc = SUPER_FSUPC(supno[krep]);
xdfs = xlsub[fsupc] + krep-fsupc+1;
maxdfs = xlsub_end[fsupc];
#ifdef PROFILE
/*procstat[pnum].unpruned++;*/
#endif
}
#ifdef CHK_DFS
if (jj == BADCOL)
printf("(%d) pdgstrf_panel_dfs[4] %d, ispruned[%d] %d, xdfs %d, maxdfs %d\n",
pnum, jj, krep, ispruned[krep], xdfs, maxdfs);
#endif
} /* else */
} /* else */
} /* if... */
} /* while xdfs < maxdfs */
/* krow has no more unexplored nbrs:
* Place snode-rep krep in postorder DFS, if this
* segment is seen for the first time. (Note that
* "repfnz[krep]" may change later.)
* Backtrack dfs to its parent.
*/
if ( marker1[krep] != jcol ) {
segrep[*nseg] = krep;
++(*nseg);
marker1[krep] = jcol;
#ifdef CHK_DFS
if (jj == BADCOL)
printf("(%d) pdgstrf_panel_dfs(%d) repfnz[%d] %d added to top.list by jj %d\n",
pnum, jj, krep, repfnz_col[krep], jj);
#endif
}
kpar = parent[krep]; /* Pop stack, mimic recursion */
if ( kpar == EMPTY ) break; /* dfs done */
krep = kpar;
xdfs = xplore[krep];
maxdfs = xplore[m + krep];
#ifdef CHK_DFS
if (jj == BADCOL)
{
register int i;
printf("(%d) pdgstrf_panel_dfs[5] pop stack: %d, krep %d, xdfs %d, maxdfs %d\n",
pnum, jj, krep, xdfs, maxdfs);
/* for (i = xdfs; i < maxdfs; i++) printf("(%d) lsub-%d", pnum, lsub[i]);*/
printf("\n");
}
#endif
} while ( kpar != EMPTY ); /* until empty stack */
} /* else: myfnz == EMPTY */
} /* else: kperm != EMPTY */
} /* for each nonzero in A[*,jj] */
#if ( DEBUGlevel>=3 )
if (jj == BADCOL) {
#define REPCOL 0
krep = REPCOL;
printf("(%d) pdgstrf_panel_dfs(end) w_lsub_end[jj=%d] %d, repfnz_col[%d] %d\n",
pnum, jj, nextl_col - nextp, krep, repfnz_col[krep]);
PrintInt10("lsub", nextl_col - nextp, &panel_lsub[nextp]);
}
#endif
w_lsub_end[jj-jcol] = nextl_col - nextp;
repfnz_col += m;
dense_col += m;
} /* for jj ... */
}