/*
* Convert a row compressed storage into a column compressed storage.
*/
void
zCompRow_to_CompCol(int m, int n, int nnz,
doublecomplex *a, int *colind, int *rowptr,
doublecomplex **at, int **rowind, int **colptr)
{
register int i, j, col, relpos;
int *marker;
/* Allocate storage for another copy of the matrix. */
*at = (doublecomplex *) doublecomplexMalloc(nnz);
*rowind = (int *) intMalloc(nnz);
*colptr = (int *) intMalloc(n+1);
marker = (int *) intCalloc(n);
/* Get counts of each column of A, and set up column pointers */
for (i = 0; i < m; ++i)
for (j = rowptr[i]; j < rowptr[i+1]; ++j) ++marker[colind[j]];
(*colptr)[0] = 0;
for (j = 0; j < n; ++j) {
(*colptr)[j+1] = (*colptr)[j] + marker[j];
marker[j] = (*colptr)[j];
}
/* Transfer the matrix into the compressed column storage. */
for (i = 0; i < m; ++i) {
for (j = rowptr[i]; j < rowptr[i+1]; ++j) {
col = colind[j];
relpos = marker[col];
(*rowind)[relpos] = i;
(*at)[relpos] = a[j];
++marker[col];
}
}
SUPERLU_FREE(marker);
}
pdgstrf_threadarg_t *
pdgstrf_thread_init(SuperMatrix *A, SuperMatrix *L, SuperMatrix *U,
pdgstrf_options_t *pdgstrf_options,
pxgstrf_shared_t *pxgstrf_shared,
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
* =======
*
* pdgstrf_thread_init() initializes the parallel data structures
* for the multithreaded routine pdgstrf_thread().
*
* Arguments
* =========
*
* A (input) SuperMatrix*
* Original matrix A, permutated by columns, of dimension
* (A->nrow, A->ncol). The type of A can be:
* Stype = NCP; Dtype = _D; Mtype = GE.
*
* L (input) SuperMatrix*
* If pdgstrf_options->refact = YES, then use the existing
* storage in L to perform LU factorization;
* Otherwise, L is not accessed. L has types:
* Stype = SCP, Dtype = _D, Mtype = TRLU.
*
* U (input) SuperMatrix*
* If pdgstrf_options->refact = YES, then use the existing
* storage in U to perform LU factorization;
* Otherwise, U is not accessed. U has types:
* Stype = NCP, Dtype = _D, Mtype = TRU.
*
* pdgstrf_options (input) pdgstrf_options_t*
* The structure contains the parameters to control how the
* factorization is performed;
* See pdgstrf_options_t structure defined in pdsp_defs.h.
*
* pxgstrf_shared (output) pxgstrf_shared_t*
* The structure contains the shared task queue and the
* synchronization variables for parallel factorization.
* See pxgstrf_shared_t structure defined in pdsp_defs.h.
*
* Gstat (output) Gstat_t*
* Record all the statistics about the factorization;
* See Gstat_t structure defined in util.h.
*
* info (output) int*
* = 0: successful exit
* > 0: if pdgstrf_options->lwork = -1, info returns the estimated
* amount of memory (in bytes) required;
* Otherwise, it returns the number of bytes allocated when
* memory allocation failure occurred, plus A->ncol.
*
*/
static GlobalLU_t Glu; /* persistent to support repeated factors. */
pdgstrf_threadarg_t *pdgstrf_threadarg;
register int n, i, nprocs;
NCPformat *Astore;
int *perm_c;
int *perm_r;
int *inv_perm_c; /* inverse of perm_c */
int *inv_perm_r; /* inverse of perm_r */
int *xprune; /* points to locations in subscript vector lsub[*].
For column i, xprune[i] denotes the point where
structural pruning begins.
I.e. only xlsub[i],..,xprune[i]-1 need to be
traversed for symbolic factorization. */
int *ispruned;/* flag to indicate whether column j is pruned */
int nzlumax;
pxgstrf_relax_t *pxgstrf_relax;
nprocs = pdgstrf_options->nprocs;
perm_c = pdgstrf_options->perm_c;
perm_r = pdgstrf_options->perm_r;
n = A->ncol;
Astore = A->Store;
inv_perm_r = (int *) intMalloc(n);
inv_perm_c = (int *) intMalloc(n);
xprune = (int *) intMalloc(n);
ispruned = (int *) intCalloc(n);
/* Pack shared data objects to each process. */
pxgstrf_shared->inv_perm_r = inv_perm_r;
pxgstrf_shared->inv_perm_c = inv_perm_c;
pxgstrf_shared->xprune = xprune;
pxgstrf_shared->ispruned = ispruned;
pxgstrf_shared->A = A;
pxgstrf_shared->Glu = &Glu;
pxgstrf_shared->Gstat = Gstat;
pxgstrf_shared->info = info;
if ( pdgstrf_options->usepr ) {
/* Compute the inverse of perm_r */
for (i = 0; i < n; ++i) inv_perm_r[perm_r[i]] = i;
}
for (i = 0; i < n; ++i) inv_perm_c[perm_c[i]] = i;
/* Initialization. */
Glu.nsuper = -1;
Glu.nextl = 0;
Glu.nextu = 0;
Glu.nextlu = 0;
ifill(perm_r, n, EMPTY);
/* Identify relaxed supernodes at the bottom of the etree. */
pxgstrf_relax = (pxgstrf_relax_t *)
SUPERLU_MALLOC((n+2) * sizeof(pxgstrf_relax_t));
pxgstrf_relax_snode(n, pdgstrf_options, pxgstrf_relax);
/* Initialize mutex variables, task queue, determine panels. */
ParallelInit(n, pxgstrf_relax, pdgstrf_options, pxgstrf_shared);
/* Set up memory image in lusup[*]. */
nzlumax = PresetMap(n, A, pxgstrf_relax, pdgstrf_options, &Glu);
if ( pdgstrf_options->refact == NO ) Glu.nzlumax = nzlumax;
SUPERLU_FREE (pxgstrf_relax);
/* Allocate global storage common to all the factor routines */
*info = pdgstrf_MemInit(n, Astore->nnz, pdgstrf_options, L, U, &Glu);
if ( *info ) return NULL;
/* Prepare arguments to all threads. */
pdgstrf_threadarg = (pdgstrf_threadarg_t *)
SUPERLU_MALLOC(nprocs * sizeof(pdgstrf_threadarg_t));
for (i = 0; i < nprocs; ++i) {
pdgstrf_threadarg[i].pnum = i;
pdgstrf_threadarg[i].info = 0;
pdgstrf_threadarg[i].pdgstrf_options = pdgstrf_options;
pdgstrf_threadarg[i].pxgstrf_shared = pxgstrf_shared;
}
#if ( DEBUGlevel==1 )
printf("** pdgstrf_thread_init() called\n");
#endif
return (pdgstrf_threadarg);
}
int
ParallelInit(int n, pxgstrf_relax_t *pxgstrf_relax,
pdgstrf_options_t *pdgstrf_options,
pxgstrf_shared_t *pxgstrf_shared)
{
int *etree = pdgstrf_options->etree;
register int w, dad, ukids, i, j, k, rs, panel_size, relax;
register int P, w_top, do_split = 0;
panel_t panel_type;
int *panel_histo = pxgstrf_shared->Gstat->panel_histo;
register int nthr, concurrency, info;
#if ( MACH==SUN )
register int sync_type = USYNC_THREAD;
/* Set concurrency level. */
nthr = sysconf(_SC_NPROCESSORS_ONLN);
thr_setconcurrency(nthr); /* number of LWPs */
concurrency = thr_getconcurrency();
#if ( PRNTlevel==1 )
printf(".. CPUs %d, concurrency (#LWP) %d, P %d\n",
nthr, concurrency, P);
#endif
/* Initialize mutex variables. */
pxgstrf_shared->lu_locks = (mutex_t *)
SUPERLU_MALLOC(NO_GLU_LOCKS * sizeof(mutex_t));
for (i = 0; i < NO_GLU_LOCKS; ++i)
mutex_init(&pxgstrf_shared->lu_locks[i], sync_type, 0);
#elif ( MACH==DEC || MACH==PTHREAD )
pxgstrf_shared->lu_locks = (pthread_mutex_t *)
SUPERLU_MALLOC(NO_GLU_LOCKS * sizeof(pthread_mutex_t));
for (i = 0; i < NO_GLU_LOCKS; ++i)
pthread_mutex_init(&pxgstrf_shared->lu_locks[i], NULL);
#else
pxgstrf_shared->lu_locks = (mutex_t *) SUPERLU_MALLOC(NO_GLU_LOCKS * sizeof(mutex_t));
#endif
#if ( PRNTlevel==1 )
printf(".. ParallelInit() ... nprocs %2d\n", pdgstrf_options->nprocs);
#endif
pxgstrf_shared->spin_locks = intCalloc(n);
pxgstrf_shared->pan_status =
(pan_status_t *) SUPERLU_MALLOC((n+1)*sizeof(pan_status_t));
pxgstrf_shared->fb_cols = intMalloc(n+1);
panel_size = pdgstrf_options->panel_size;
relax = pdgstrf_options->relax;
w = MAX(panel_size, relax) + 1;
for (i = 0; i < w; ++i) panel_histo[i] = 0;
pxgstrf_shared->num_splits = 0;
if ( (info = queue_init(&pxgstrf_shared->taskq, n)) ) {
fprintf(stderr, "ParallelInit(): %d\n", info);
ABORT("queue_init fails.");
}
/* Count children of each node in the etree. */
for (i = 0; i <= n; ++i) pxgstrf_shared->pan_status[i].ukids = 0;
for (i = 0; i < n; ++i) {
dad = etree[i];
++pxgstrf_shared->pan_status[dad].ukids;
}
/* Find the panel partitions and initialize each panel's status */
#ifdef PROFILE
num_panels = 0;
#endif
pxgstrf_shared->tasks_remain = 0;
rs = 1;
w_top = panel_size/2;
if ( w_top == 0 ) w_top = 1;
P = 12;
for (i = 0; i < n; ) {
if ( pxgstrf_relax[rs].fcol == i ) {
w = pxgstrf_relax[rs++].size;
panel_type = RELAXED_SNODE;
pxgstrf_shared->pan_status[i].state = CANGO;
} else {
w = MIN(panel_size, pxgstrf_relax[rs].fcol - i);
#ifdef SPLIT_TOP
if ( !do_split ) {
if ( (n-i) < panel_size * P ) do_split = 1;
}
if ( do_split && w > w_top ) { /* split large panel */
w = w_top;
++pxgstrf_shared->num_splits;
}
#endif
for (j = i+1; j < i + w; ++j)
/* Do not allow panel to cross a branch point in the etree. */
if ( pxgstrf_shared->pan_status[j].ukids > 1 ) break;
w = j - i; /* j should start a new panel */
panel_type = REGULAR_PANEL;
pxgstrf_shared->pan_status[i].state = UNREADY;
#ifdef DOMAINS
if ( in_domain[i] == TREE_DOMAIN ) panel_type = TREE_DOMAIN;
#endif
}
if ( panel_type == REGULAR_PANEL ) {
++pxgstrf_shared->tasks_remain;
/*printf("nondomain panel %6d -- %6d\n", i, i+w-1);
fflush(stdout);*/
}
ukids = k = 0;
for (j = i; j < i + w; ++j) {
pxgstrf_shared->pan_status[j].size = k--;
pxgstrf_shared->pan_status[j].type = panel_type;
ukids += pxgstrf_shared->pan_status[j].ukids;
}
pxgstrf_shared->pan_status[i].size = w; /* leading column */
/* only count those kids outside the panel */
pxgstrf_shared->pan_status[i].ukids = ukids - (w-1);
panel_histo[w]++;
#ifdef PROFILE
panstat[i].size = w;
++num_panels;
#endif
pxgstrf_shared->fb_cols[i] = i;
i += w;
} /* for i ... */
/* Dummy root */
pxgstrf_shared->pan_status[n].size = 1;
pxgstrf_shared->pan_status[n].state = UNREADY;
#if ( PRNTlevel==1 )
printf(".. Split: P %d, #nondomain panels %d\n", P, pxgstrf_shared->tasks_remain);
#endif
#ifdef DOMAINS
EnqueueDomains(&pxgstrf_shared->taskq, list_head, pxgstrf_shared);
#else
EnqueueRelaxSnode(&pxgstrf_shared->taskq, n, pxgstrf_relax, pxgstrf_shared);
#endif
#if ( PRNTlevel==1 )
printf(".. # tasks %d\n", pxgstrf_shared->tasks_remain);
fflush(stdout);
#endif
#ifdef PREDICT_OPT
/* Set up structure describing children */
for (i = 0; i <= n; cp_firstkid[i++] = EMPTY);
for (i = n-1; i >= 0; i--) {
dad = etree[i];
cp_nextkid[i] = cp_firstkid[dad];
cp_firstkid[dad] = i;
}
#endif
return 0;
} /* ParallelInit */
int
qrnzcnt(int neqns, int adjlen, int *xadj, int *adjncy, int *zfdperm,
int *perm, int *invp, int *etpar, int *colcnt_h,
int *nlnz, int *part_super_ata, int *part_super_h)
{
/*
o 5/20/95 Xiaoye S. Li:
Translated from fcnthn.f using f2c;
Modified to use 0-based indexing in C;
Initialize xsup = 0 as suggested by B. Peyton to handle singletons.
o 5/24/95 Xiaoye S. Li:
Modified to compute row/column counts of R in QR factorization
1. Compute row counts of A, and f(i) in a separate pass
def
2. Re-define hadj[k] === U { j | j in Struct(A_i*), j>k}
i:f(i)==k
Record supernode partition in part_super_ata[*] of size neqns:
part_super_ata[k] = size of the supernode beginning at column k;
= 0, elsewhere.
o 1/16/96 Xiaoye S. Li:
Modified to incorporate row/column counts of the Householder
Matrix H in the QR factorization A --> H , R.
Record supernode partition in part_super_h[*] of size neqns:
part_super_h[k] = size of the supernode beginning at column k;
= 0, elsewhere.
***********************************************************************
Version: 0.3
Last modified: January 12, 1995
Authors: Esmond G. Ng and Barry W. Peyton
Mathematical Sciences Section, Oak Ridge National Laboratoy
***********************************************************************
************** FCNTHN ..... FIND NONZERO COUNTS ***************
***********************************************************************
PURPOSE:
THIS SUBROUTINE DETERMINES THE ROW COUNTS AND COLUMN COUNTS IN
THE CHOLESKY FACTOR. IT USES A DISJOINT SET UNION ALGORITHM.
TECHNIQUES:
1) SUPERNODE DETECTION.
2) PATH HALVING.
3) NO UNION BY RANK.
NOTES:
1) ASSUMES A POSTORDERING OF THE ELIMINATION TREE.
INPUT PARAMETERS:
(I) NEQNS - NUMBER OF EQUATIONS.
(I) ADJLEN - LENGTH OF ADJACENCY STRUCTURE.
(I) XADJ(*) - ARRAY OF LENGTH NEQNS+1, CONTAINING POINTERS
TO THE ADJACENCY STRUCTURE.
(I) ADJNCY(*) - ARRAY OF LENGTH ADJLEN, CONTAINING
THE ADJACENCY STRUCTURE.
(I) ZFDPERM(*) - THE ROW PERMUTATION VECTOR THAT PERMUTES THE
MATRIX TO HAVE ZERO-FREE DIAGONAL.
ZFDPERM(I) = J MEANS ROW I OF THE ORIGINAL
MATRIX IS IN ROW J OF THE PERMUTED MATRIX.
(I) PERM(*) - ARRAY OF LENGTH NEQNS, CONTAINING THE
POSTORDERING.
(I) INVP(*) - ARRAY OF LENGTH NEQNS, CONTAINING THE
INVERSE OF THE POSTORDERING.
(I) ETPAR(*) - ARRAY OF LENGTH NEQNS, CONTAINING THE
ELIMINATION TREE OF THE POSTORDERED MATRIX.
OUTPUT PARAMETERS:
(I) ROWCNT(*) - ARRAY OF LENGTH NEQNS, CONTAINING THE NUMBER
OF NONZEROS IN EACH ROW OF THE FACTOR,
INCLUDING THE DIAGONAL ENTRY.
(I) COLCNT(*) - ARRAY OF LENGTH NEQNS, CONTAINING THE NUMBER
OF NONZEROS IN EACH COLUMN OF THE FACTOR,
INCLUDING THE DIAGONAL ENTRY.
(I) NLNZ - NUMBER OF NONZEROS IN THE FACTOR, INCLUDING
THE DIAGONAL ENTRIES.
(I) PART_SUPER_ATA SUPERNODE PARTITION IN THE CHOLESKY FACTOR
OF A'A.
(I) PART_SUPER_H SUPERNODE PARTITION IN THE HOUSEHOLDER
MATRIX H.
WORK PARAMETERS:
(I) SET(*) - ARRAY OF LENGTH NEQNS USED TO MAINTAIN THE
DISJOINT SETS (I.E., SUBTREES).
(I) PRVLF(*) - ARRAY OF LENGTH NEQNS USED TO RECORD THE
PREVIOUS LEAF OF EACH ROW SUBTREE.
(I) LEVEL(*) - ARRAY OF LENGTH NEQNS+1 CONTAINING THE LEVEL
(DISTANCE FROM THE ROOT).
(I) WEIGHT(*) - ARRAY OF LENGTH NEQNS+1 CONTAINING WEIGHTS
USED TO COMPUTE COLUMN COUNTS.
(I) FDESC(*) - ARRAY OF LENGTH NEQNS+1 CONTAINING THE
FIRST (I.E., LOWEST-NUMBERED) DESCENDANT.
(I) NCHILD(*) - ARRAY OF LENGTH NEQNS+1 CONTAINING THE
NUMBER OF CHILDREN.
(I) PRVNBR(*) - ARRAY OF LENGTH NEQNS USED TO RECORD THE
PREVIOUS ``LOWER NEIGHBOR'' OF EACH NODE.
FIRST CREATED ON APRIL 12, 1990.
LAST UPDATED ON JANUARY 12, 1995.
***********************************************************************
*/
/* Local variables */
int temp, last1, last2, i, j, k, lflag, pleaf, hinbr, jstop,
jstrt, ifdesc, oldnbr, parent, lownbr, lca;
int xsup; /* the ongoing supernode */
int *set, *prvlf, *level, *weight, *fdesc, *nchild, *prvnbr;
int *fnz; /* first nonzero column subscript in each row */
int *marker; /* used to remove duplicate indices */
int *fnz_hadj; /* higher-numbered neighbors of the first nonzero
(higher adjacency set of A'A) */
int *hadj_begin; /* pointers to the fnz_hadj[] structure */
int *hadj_end; /* pointers to the fnz_hadj[] structure */
/* Locally malloc'd room for QR purpose */
/* ----------------------------------------------------------
FIRST set is defined as first[j] := { i : f[i] = j } ,
which is a collection of disjoint sets of integers between
0 and n-1.
---------------------------------------------------------- */
int *first; /* header pointing to FIRST set */
int *firstset; /* linked list to describe FIRST set */
int *weight_h; /* weights for H */
int *rowcnt; /* row colunts for Lc */
int *colcnt; /* column colunts for Lc */
int *rowcnt_h; /* row colunts for H */
int nsuper; /* total number of fundamental supernodes in Lc */
int nhnz;
set = intMalloc(neqns);
prvlf = intMalloc(neqns);
level = intMalloc(neqns + 1); /* length n+1 */
weight = intMalloc(neqns + 1); /* length n+1 */
fdesc = intMalloc(neqns + 1); /* length n+1 */
nchild = intMalloc(neqns + 1); /* length n+1 */
prvnbr = intMalloc(neqns);
fnz_hadj = intMalloc(adjlen + 2*neqns + 1);
hadj_begin = fnz_hadj + adjlen; /* neqns+1 */
hadj_end = hadj_begin + neqns + 1; /* neqns */
fnz = set; /* aliasing for the time being */
marker = prvlf; /* " " " */
first = intMalloc(neqns);
firstset = intMalloc(neqns);
weight_h = intCalloc(neqns + 1); /* length n+1 */
rowcnt_h = intMalloc(neqns);
rowcnt = intMalloc(neqns);
colcnt = intMalloc(neqns);
/* -------------------------------------------------------
* Compute fnz[*], first[*], nchild[*] and row counts of A.
* Also find supernodes in H.
*
* Note that the structure of each row of H forms a simple path in
* the etree between fnz[i] and i (George, Liu & Ng (1988)).
* The "first vertices" of the supernodes in H are characterized
* by the following conditions:
* 1) first nonzero in each row of A, i.e., fnz(i);
* or 2) nchild >= 2;
* ------------------------------------------------------- */
for (k = 0; k < neqns; ++k) {
fnz[k] = first[k] = marker[k] = EMPTY;
rowcnt[k] = part_super_ata[k] = 0;
part_super_h[k] = 0;
nchild[k] = 0;
}
nchild[ROOT] = 0;
xsup = 0;
for (k = 0; k < neqns; ++k) {
parent = etpar[k];
++nchild[parent];
if ( k != 0 && nchild[k] >= 2 ) {
part_super_h[xsup] = k - xsup;
xsup = k;
}
oldnbr = perm[k];
for (j = xadj[oldnbr]; j < xadj[oldnbr+1]; ++j) {
/*
* Renumber vertices of G(A) by postorder
*/
/* i = invp[zfdperm[adjncy[j]]];*/
i = zfdperm[adjncy[j]];
++rowcnt[i];
if (fnz[i] == EMPTY) {
/*
* Build linked list to describe FIRST sets
*/
fnz[i] = k;
firstset[i] = first[k];
first[k] = i;
if ( k != 0 && xsup != k ) {
part_super_h[xsup] = k - xsup;
xsup = k;
}
}
}
}
part_super_h[xsup] = neqns - xsup;
#ifdef CHK_NZCNT
printf("%8s%8s%8s\n", "k", "fnz", "first");
for (k = 0; k < neqns; ++k)
printf("%8d%8d%8d\n", k, fnz[k], first[k]);
#endif
/* Set up fnz_hadj[*] structure. */
hadj_begin[0] = 0;
for (k = 0; k < neqns; ++k) {
temp = 0;
oldnbr = perm[k];
hadj_end[k] = hadj_begin[k];
for (j = xadj[oldnbr]; j < xadj[oldnbr+1]; ++j) {
/* hinbr = invp[zfdperm[adjncy[j]]];*/
hinbr = zfdperm[adjncy[j]];
jstrt = fnz[hinbr]; /* first nonzero must be <= k */
if ( jstrt != k && marker[jstrt] < k ) {
/* ----------------------------------
filtering k itself and duplicates
---------------------------------- */
fnz_hadj[hadj_end[jstrt]] = k;
++hadj_end[jstrt];
marker[jstrt] = k;
}
if ( jstrt == k ) temp += rowcnt[hinbr];
}
hadj_begin[k+1] = hadj_begin[k] + temp;
}
#ifdef CHK_NZCNT
printf("%8s%8s\n", "k", "hadj");
for (k = 0; k < neqns; ++k) {
printf("%8d", k);
for (j = hadj_begin[k]; j < hadj_end[k]; ++j)
printf("%8d", fnz_hadj[j]);
printf("\n");
}
#endif
/* --------------------------------------------------
COMPUTE LEVEL(*), FDESC(*), NCHILD(*).
INITIALIZE ROWCNT(*), COLCNT(*),
SET(*), PRVLF(*), WEIGHT(*), PRVNBR(*).
-------------------------------------------------- */
level[ROOT] = 0;
for (k = neqns-1; k >= 0; --k) {
rowcnt[k] = 1;
colcnt[k] = 0;
set[k] = k;
prvlf[k] = EMPTY;
level[k] = level[etpar[k]] + 1;
weight[k] = 1;
fdesc[k] = k;
prvnbr[k] = EMPTY;
}
fdesc[ROOT] = EMPTY;
for (k = 0; k < neqns; ++k) {
parent = etpar[k];
weight[parent] = 0;
colcnt_h[k] = 0;
ifdesc = fdesc[k];
if (ifdesc < fdesc[parent]) {
fdesc[parent] = ifdesc;
}
}
xsup = 0; /* BUG FIX */
nsuper = 0;
/* ------------------------------------
FOR EACH ``LOW NEIGHBOR'' LOWNBR ...
------------------------------------ */
for (lownbr = 0; lownbr < neqns; ++lownbr) {
for (i = first[lownbr]; i != EMPTY; i = firstset[i]) {
rowcnt_h[i] = 1 + ( level[lownbr] - level[i] );
++weight_h[lownbr];
parent = etpar[i];
--weight_h[parent];
}
lflag = 0;
ifdesc = fdesc[lownbr];
jstrt = hadj_begin[lownbr];
jstop = hadj_end[lownbr];
/* -----------------------------------------------
FOR EACH ``HIGH NEIGHBOR'', HINBR OF LOWNBR ...
----------------------------------------------- */
for (j = jstrt; j < jstop; ++j) {
hinbr = fnz_hadj[j];
if (hinbr > lownbr) {
if (ifdesc > prvnbr[hinbr]) {
/* -------------------------
INCREMENT WEIGHT(LOWNBR).
------------------------- */
++weight[lownbr];
pleaf = prvlf[hinbr];
/* -----------------------------------------
IF HINBR HAS NO PREVIOUS ``LOW NEIGHBOR'' THEN ...
----------------------------------------- */
if (pleaf == EMPTY) {
/* -----------------------------------------
... ACCUMULATE LOWNBR-->HINBR PATH LENGTH
IN ROWCNT(HINBR).
----------------------------------------- */
rowcnt[hinbr] = rowcnt[hinbr] +
level[lownbr] - level[hinbr];
} else {
/* -----------------------------------------
... OTHERWISE, LCA <-- FIND(PLEAF), WHICH
IS THE LEAST COMMON ANCESTOR OF PLEAF
AND LOWNBR. (PATH HALVING.)
----------------------------------------- */
last1 = pleaf;
last2 = set[last1];
lca = set[last2];
while ( lca != last2 ) {
set[last1] = lca;
last1 = lca;
last2 = set[last1];
lca = set[last2];
}
/* -------------------------------------
ACCUMULATE PLEAF-->LCA PATH LENGTH IN
ROWCNT(HINBR). DECREMENT WEIGHT(LCA).
------------------------------------- */
rowcnt[hinbr] = rowcnt[hinbr] +
level[lownbr] - level[lca];
--weight[lca];
}
/* ----------------------------------------------
LOWNBR NOW BECOMES ``PREVIOUS LEAF'' OF HINBR.
---------------------------------------------- */
prvlf[hinbr] = lownbr;
lflag = 1;
}
/* --------------------------------------------------
LOWNBR NOW BECOMES ``PREVIOUS NEIGHBOR'' OF HINBR.
-------------------------------------------------- */
prvnbr[hinbr] = lownbr;
}
} /* for j ... */
/* ----------------------------------------------------
DECREMENT WEIGHT ( PARENT(LOWNBR) ).
SET ( P(LOWNBR) ) <-- SET ( P(LOWNBR) ) + SET(XSUP).
---------------------------------------------------- */
parent = etpar[lownbr];
--weight[parent];
if (lflag == 1 || nchild[lownbr] >= 2) {
/* lownbr is detected as the beginning of the new supernode */
if ( lownbr != 0 ) part_super_ata[xsup] = lownbr - xsup;
++nsuper;
xsup = lownbr;
} else {
if ( parent == ROOT && ifdesc == lownbr ) {
/* lownbr is a singleton, and begins a new supernode
but is not detected as doing so -- BUG FIX */
part_super_ata[lownbr] = 1;
++nsuper;
xsup = lownbr;
}
}
set[xsup] = parent;
} /* for lownbr ... */
/* ---------------------------------------------------------
USE WEIGHTS TO COMPUTE COLUMN (AND TOTAL) NONZERO COUNTS.
--------------------------------------------------------- */
*nlnz = nhnz = 0;
for (k = 0; k < neqns; ++k) {
/* for R */
temp = colcnt[k] + weight[k];
colcnt[k] = temp;
*nlnz += temp;
parent = etpar[k];
if (parent != ROOT) {
colcnt[parent] += temp;
}
/* for H */
temp = colcnt_h[k] + weight_h[k];
colcnt_h[k] = temp;
nhnz += temp;
if (parent != ROOT) {
colcnt_h[parent] += temp;
}
}
part_super_ata[xsup] = neqns - xsup;
/* Fix the supernode partition in H. */
free (set);
free (prvlf);
free (level);
free (weight);
free (fdesc);
free (nchild);
free (prvnbr);
free (fnz_hadj);
free (first);
free (firstset);
free (weight_h);
free (rowcnt_h);
free (rowcnt);
free (colcnt);
#if ( PRNTlevel==1 )
printf(".. qrnzcnt() nlnz %d, nhnz %d, nlnz/nhnz %.2f\n",
*nlnz, nhnz, (float) *nlnz/nhnz);
#endif
#if ( DEBUGlevel>=2 )
print_int_vec("part_super_h", neqns, part_super_h);
#endif
return 0;
} /* qrnzcnt_ */
void
psgstrf_relax_snode(
const int n, /* number of columns in the matrix */
//psgstrf_options_t *psgstrf_options,
superlumt_options_t *psgstrf_options, //sj
pxgstrf_relax_t *pxgstrf_relax /* relaxed s-nodes */
)
{
/*
* -- SuperLU MT routine (version 2.0) --
* Lawrence Berkeley National Lab, Univ. of California Berkeley,
* and Xerox Palo Alto Research Center.
* September 10, 2007
*
* Purpose
* =======
* psgstrf_relax_snode() identifes the initial relaxed supernodes,
* assuming that the matrix has been reordered according to the postorder
* of the etree.
*
*/
register int j, parent, rs;
register int fcol; /* beginning of a snode */
int *desc; /* no of descendants of each etree node. */
int *etree = psgstrf_options->etree; /* column elimination tree */
int relax = psgstrf_options->relax; /* maximum no of columns allowed
in a relaxed s-node */
desc = intCalloc(n+1);
/* Compute the number of descendants of each node in the etree */
for (j = 0; j < n; j++) {
parent = etree[j];
desc[parent] += desc[j] + 1;
}
rs = 1;
/* Identify the relaxed supernodes by postorder traversal of the etree. */
for (j = 0; j < n; ) {
parent = etree[j];
fcol = j;
while ( parent != n && desc[parent] < relax ) {
j = parent;
parent = etree[j];
}
/* found a supernode with j being the last column. */
pxgstrf_relax[rs].fcol = fcol;
pxgstrf_relax[rs].size = j - fcol + 1;
#ifdef DOMAINS
for (i = fcol; i <= j; ++i) in_domain[i] = RELAXED_SNODE;
#endif
j++; rs++;
/* Search for a new leaf */
while ( desc[j] != 0 && j < n ) j++;
}
pxgstrf_relax[rs].fcol = n;
pxgstrf_relax[0].size = rs-1; /* number of relaxed supernodes */
#if (PRNTlevel==1)
printf(".. No of relaxed s-nodes %d\n", pxgstrf_relax[0].size);
#endif
SUPERLU_FREE (desc);
}
