/*
------------------------------------------------------
transform an ETree object by
(1) merging small fronts into larger fronts
using the ETree_mergeFrontsOne() method
(2) merging small fronts into larger fronts
using the ETree_mergeFrontsAll() method
(3) split a large front into a chain of smaller fronts
using the ETree_splitFronts() method
created -- 96jun27, cca
------------------------------------------------------
*/
ETree *
ETree_transform2 (
ETree *etree,
int vwghts[],
int maxzeros,
int maxfrontsize,
int seed
) {
ETree *etree2 ;
int nfront, nvtx ;
IV *nzerosIV ;
/*
---------------
check the input
---------------
*/
if ( etree == NULL
|| (nfront = etree->nfront) <= 0
|| (nvtx = etree->nvtx) <= 0
|| maxfrontsize <= 0 ) {
fprintf(stderr, "\n fatal error in ETree_transform2(%p,%p,%d,%d,%d)"
"\n bad input\n", etree, vwghts, maxzeros, maxfrontsize,
seed) ;
spoolesFatal();
}
nzerosIV = IV_new();
IV_init(nzerosIV, nfront, NULL) ;
IV_fill(nzerosIV, 0) ;
/*
--------------------------
first, merge only children
--------------------------
*/
etree2 = ETree_mergeFrontsOne(etree, maxzeros, nzerosIV) ;
ETree_free(etree) ;
etree = etree2 ;
/*
--------------------------
second, merge all children
--------------------------
*/
etree2 = ETree_mergeFrontsAll(etree, maxzeros, nzerosIV) ;
ETree_free(etree) ;
etree = etree2 ;
/*
-----------------------------------
fourth, split large interior fronts
-----------------------------------
*/
etree2 = ETree_splitFronts(etree, vwghts, maxfrontsize, seed) ;
ETree_free(etree) ;
etree = etree2 ;
/*
------------------------
free the working storage
------------------------
*/
IV_free(nzerosIV) ;
return(etree) ; }
/*
-------------------------------------
return an IV object filled with the
weights of the component's boundaries
created -- 96oct21, cca
-------------------------------------
*/
IV *
GPart_bndWeightsIV (
GPart *gpart
) {
Graph *graph ;
int icomp, ii, ncomp, nvtx, v, vsize, vwght, w ;
int *bnd, *compids, *cweights, *mark, *vadj, *vwghts ;
IV *bndIV ;
/*
---------------
check the input
---------------
*/
if ( gpart == NULL || (graph = gpart->g) == NULL ) {
fprintf(stderr, "\n fatal error in GPart_bndWeightsIV(%p)"
"\n bad input\n", gpart) ;
exit(-1) ;
}
nvtx = gpart->nvtx ;
ncomp = gpart->ncomp ;
compids = IV_entries(&gpart->compidsIV) ;
cweights = IV_entries(&gpart->cweightsIV) ;
vwghts = graph->vwghts ;
bndIV = IV_new() ;
IV_init(bndIV, 1 + ncomp, NULL) ;
IV_fill(bndIV, 0) ;
bnd = IV_entries(bndIV) ;
mark = IVinit(ncomp+1, -1) ;
for ( v = 0 ; v < nvtx ; v++ ) {
if ( compids[v] == 0 ) {
vwght = (vwghts == NULL) ? 1 : vwghts[v] ;
Graph_adjAndSize(graph, v, &vsize, &vadj) ;
for ( ii = 0 ; ii < vsize ; ii++ ) {
w = vadj[ii] ;
if ( (icomp = compids[w]) != 0 && mark[icomp] != v ) {
mark[icomp] = v ;
bnd[icomp] += vwght ;
}
}
}
}
IVfree(mark) ;
return(bndIV) ; }
/*
-----------------------------------------------
initialize the object given the number of nodes
created -- 96mar10, cca
-----------------------------------------------
*/
void
DSTree_init1 (
DSTree *dstree,
int ndomsep,
int nvtx
) {
/*
---------------
check the input
---------------
*/
if ( dstree == NULL || ndomsep <= 0 ) {
fprintf(stderr, "\n fatal error in DSTree_init1(%p,%d,%d)"
"\n bad input\n", dstree, ndomsep, nvtx) ;
exit(-1) ;
}
DSTree_clearData(dstree) ;
dstree->tree = Tree_new() ;
Tree_init1(dstree->tree, ndomsep) ;
dstree->mapIV = IV_new() ;
IV_init(dstree->mapIV, nvtx, NULL) ;
IV_fill(dstree->mapIV, -1) ;
return ; }
/*
----------------------------------------------------------------
purpose --
if the elimination has halted before all the stages have been
eliminated, then create the schur complement graph and the map
from the original vertices those in the schur complement graph.
schurGraph -- Graph object to contain the schur complement graph
VtoPhi -- IV object to contain the map from vertices in V
to schur complement vertices in Phi
created -- 97feb01, cca
----------------------------------------------------------------
*/
void
MSMD_makeSchurComplement (
MSMD *msmd,
Graph *schurGraph,
IV *VtoPhiIV
) {
int nedge, nPhi, nvtx, totewght, totvwght ;
int *mark, *rep, *VtoPhi, *vwghts ;
int count, *list ;
int ierr, ii, size, *adj ;
int phi, psi, tag ;
IP *ip ;
IVL *adjIVL ;
MSMDvtx *u, *v, *vertices, *vfirst, *vlast, *w ;
/*
---------------
check the input
---------------
*/
if ( msmd == NULL || schurGraph == NULL || VtoPhiIV == NULL ) {
fprintf(stderr,
"\n\n fatal error in MSMD_makeSchurComplement(%p,%p,%p)"
"\n bad input\n", msmd, schurGraph, VtoPhiIV) ;
exit(-1) ;
}
vertices = msmd->vertices ;
nvtx = msmd->nvtx ;
/*
-------------------------------------
initialize the V-to-Phi map IV object
-------------------------------------
*/
IV_clearData(VtoPhiIV) ;
IV_setSize(VtoPhiIV, nvtx) ;
IV_fill(VtoPhiIV, -2) ;
VtoPhi = IV_entries(VtoPhiIV) ;
/*
---------------------------------------------
count the number of Schur complement vertices
---------------------------------------------
*/
vfirst = vertices ;
vlast = vfirst + nvtx - 1 ;
nPhi = 0 ;
for ( v = vfirst ; v <= vlast ; v++ ) {
#if MYDEBUG > 0
fprintf(stdout, "\n v->id = %d, v->status = %c", v->id, v->status) ;
fflush(stdout) ;
#endif
switch ( v->status ) {
case 'L' :
case 'E' :
case 'I' :
break ;
case 'B' :
VtoPhi[v->id] = nPhi++ ;
#if MYDEBUG > 0
fprintf(stdout, ", VtoPhi[%d] = %d", v->id, VtoPhi[v->id]) ;
fflush(stdout) ;
#endif
break ;
default :
break ;
}
}
#if MYDEBUG > 0
fprintf(stdout, "\n\n nPhi = %d", nPhi) ;
fflush(stdout) ;
#endif
/*
----------------------------------------------------
get the representative vertex id for each Phi vertex
----------------------------------------------------
*/
rep = IVinit(nPhi, -1) ;
for ( v = vfirst ; v <= vlast ; v++ ) {
if ( (phi = VtoPhi[v->id]) >= 0 ) {
#if MYDEBUG > 0
fprintf(stdout, "\n rep[%d] = %d", phi, v->id) ;
fflush(stdout) ;
#endif
rep[phi] = v->id ;
}
}
/*
------------------------------------------
set the map for indistinguishable vertices
------------------------------------------
*/
for ( v = vfirst ; v <= vlast ; v++ ) {
if ( v->status == 'I' ) {
w = v ;
while ( w->status == 'I' ) {
w = w->par ;
}
#if MYDEBUG > 0
fprintf(stdout, "\n v = %d, status = %c, w = %d, status = %c",
v->id, v->status, w->id, w->status) ;
fflush(stdout) ;
#endif
VtoPhi[v->id] = VtoPhi[w->id] ;
}
}
#if MYDEBUG > 0
fprintf(stdout, "\n\n VtoPhi") ;
IV_writeForHumanEye(VtoPhiIV, stdout) ;
fflush(stdout) ;
#endif
/*
---------------------------
initialize the Graph object
---------------------------
*/
Graph_clearData(schurGraph) ;
Graph_init1(schurGraph, 1, nPhi, 0, 0, IVL_CHUNKED, IVL_CHUNKED) ;
adjIVL = schurGraph->adjIVL ;
vwghts = schurGraph->vwghts ;
#if MYDEBUG > 0
fprintf(stdout, "\n\n schurGraph initialized, nvtx = %d",
schurGraph->nvtx) ;
fflush(stdout) ;
#endif
/*
-------------------------------
fill the vertex adjacency lists
-------------------------------
*/
mark = IVinit(nPhi, -1) ;
list = IVinit(nPhi, -1) ;
nedge = totvwght = totewght = 0 ;
for ( phi = 0 ; phi < nPhi ; phi++ ) {
/*
-----------------------------
get the representative vertex
-----------------------------
*/
v = vfirst + rep[phi] ;
#if MYDEBUG > 0
fprintf(stdout, "\n phi = %d, v = %d", phi, v->id) ;
fflush(stdout) ;
MSMDvtx_print(v, stdout) ;
fflush(stdout) ;
#endif
count = 0 ;
tag = v->id ;
/*
---------------------------
load self in adjacency list
---------------------------
*/
mark[phi] = tag ;
totewght += v->wght * v->wght ;
#if MYDEBUG > 0
fprintf(stdout, "\n mark[%d] = %d", phi, mark[phi]) ;
fflush(stdout) ;
#endif
list[count++] = phi ;
/*
----------------------------------------
load boundary lists of adjacent subtrees
----------------------------------------
*/
for ( ip = v->subtrees ; ip != NULL ; ip = ip->next ) {
u = vertices + ip->val ;
size = u->nadj ;
adj = u->adj ;
#if MYDEBUG > 0
fprintf(stdout, "\n subtree %d :", u->id) ;
IVfp80(stdout, size, adj, 15, &ierr) ;
fflush(stdout) ;
#endif
for ( ii = 0 ; ii < size ; ii++ ) {
w = vertices + adj[ii] ;
#if MYDEBUG > 0
fprintf(stdout, "\n w %d, status %c, psi %d",
w->id, w->status, VtoPhi[w->id]) ;
fflush(stdout) ;
#endif
if ( (psi = VtoPhi[w->id]) != -2 && mark[psi] != tag ) {
mark[psi] = tag ;
#if MYDEBUG > 0
fprintf(stdout, ", mark[%d] = %d", psi, mark[psi]) ;
fflush(stdout) ;
#endif
list[count++] = psi ;
totewght += v->wght * w->wght ;
}
}
}
/*
----------------------
load adjacent vertices
----------------------
*/
size = v->nadj ;
adj = v->adj ;
for ( ii = 0 ; ii < size ; ii++ ) {
w = vertices + adj[ii] ;
if ( (psi = VtoPhi[w->id]) != -2 && mark[psi] != tag ) {
mark[psi] = tag ;
list[count++] = psi ;
totewght += v->wght * w->wght ;
}
}
/*
---------------------------------------------
sort the list and inform adjacency IVL object
---------------------------------------------
*/
IVqsortUp(count, list) ;
IVL_setList(adjIVL, phi, count, list) ;
/*
--------------------------------------
set the vertex weight and increment
the total vertex weight and edge count
--------------------------------------
*/
vwghts[phi] = v->wght ;
totvwght += v->wght ;
nedge += count ;
}
schurGraph->totvwght = totvwght ;
schurGraph->nedges = nedge ;
schurGraph->totewght = totewght ;
/*
------------------------
free the working storage
------------------------
*/
IVfree(list) ;
IVfree(mark) ;
IVfree(rep) ;
return ; }
/*
-------------------------------------------------------------
purpose --- to compute a matrix-vector multiply y[] = C * x[]
where C is the identity, A or B (depending on *pprbtype).
*pnrows -- # of rows in x[]
*pncols -- # of columns in x[]
*pprbtype -- problem type
*pprbtype = 1 --> vibration problem, matrix is A
*pprbtype = 2 --> buckling problem, matrix is B
*pprbtype = 3 --> matrix is identity, y[] = x[]
x[] -- vector to be multiplied
NOTE: the x[] vector is global, not a portion
y[] -- product vector
NOTE: the y[] vector is global, not a portion
created -- 98aug28, cca & jcp
-------------------------------------------------------------
*/
void
JimMatMulMPI (
int *pnrows,
int *pncols,
double x[],
double y[],
int *pprbtype,
void *data
) {
BridgeMPI *bridge = (BridgeMPI *) data ;
int ncols, nent, nrows ;
#if MYDEBUG > 0
double t1, t2 ;
count_JimMatMul++ ;
MARKTIME(t1) ;
if ( bridge->myid == 0 ) {
fprintf(stdout, "\n (%d) JimMatMulMPI() start", count_JimMatMul) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile,
"\n (%d) JimMatMulMPI() start", count_JimMatMul) ;
fflush(bridge->msgFile) ;
#endif
nrows = *pnrows ;
ncols = *pncols ;
nent = nrows*ncols ;
if ( *pprbtype == 3 ) {
/*
--------------------------
... matrix is the identity
--------------------------
*/
DVcopy(nent, y, x) ;
} else {
BridgeMPI *bridge = (BridgeMPI *) data ;
DenseMtx *mtx, *newmtx ;
int irow, jcol, jj, kk, myid, neqns, nowned, tag = 0 ;
int *vtxmap ;
int stats[4] ;
IV *mapIV ;
/*
---------------------------------------------
slide the owned rows of x[] down in the array
---------------------------------------------
*/
vtxmap = IV_entries(bridge->vtxmapIV) ;
neqns = bridge->neqns ;
myid = bridge->myid ;
nowned = IV_size(bridge->myownedIV) ;
for ( jcol = jj = kk = 0 ; jcol < ncols ; jcol++ ) {
for ( irow = 0 ; irow < neqns ; irow++, jj++ ) {
if ( vtxmap[irow] == myid ) {
y[kk++] = x[jj] ;
}
}
}
if ( kk != nowned * ncols ) {
fprintf(stderr, "\n proc %d : kk %d, nowned %d, ncols %d",
myid, kk, nowned, ncols) ;
exit(-1) ;
}
/*
----------------------------------------
call the method that assumes local input
----------------------------------------
*/
if ( bridge->msglvl > 2 ) {
fprintf(bridge->msgFile,
"\n inside JimMatMulMPI, calling MatMulMpi"
"\n prbtype %d, nrows %d, ncols %d, nowned %d",
*pprbtype, *pnrows, *pncols, nowned) ;
fflush(bridge->msgFile) ;
}
MatMulMPI(&nowned, pncols, y, y, pprbtype, data) ;
/*
-------------------------------------------------
gather all the entries of y[] onto processor zero
-------------------------------------------------
*/
mtx = DenseMtx_new() ;
DenseMtx_init(mtx, SPOOLES_REAL, 0, 0, nowned, ncols, 1, nowned) ;
DVcopy (nowned*ncols, DenseMtx_entries(mtx), y) ;
IVcopy(nowned, mtx->rowind, IV_entries(bridge->myownedIV)) ;
mapIV = IV_new() ;
IV_init(mapIV, neqns, NULL) ;
IV_fill(mapIV, 0) ;
IVfill(4, stats, 0) ;
if ( bridge->msglvl > 2 ) {
fprintf(bridge->msgFile, "\n mtx: %d rows x %d columns",
mtx->nrow, mtx->ncol) ;
fflush(bridge->msgFile) ;
}
newmtx = DenseMtx_MPI_splitByRows(mtx, mapIV, stats, bridge->msglvl,
bridge->msgFile, tag, bridge->comm) ;
if ( bridge->msglvl > 2 ) {
fprintf(bridge->msgFile, "\n newmtx: %d rows x %d columns",
newmtx->nrow, newmtx->ncol) ;
fflush(bridge->msgFile) ;
}
DenseMtx_free(mtx) ;
mtx = newmtx ;
IV_free(mapIV) ;
if ( myid == 0 ) {
if ( mtx->nrow != neqns || mtx->ncol != ncols ) {
fprintf(bridge->msgFile,
"\n\n WHOA: mtx->nrows %d, mtx->ncols %d"
", neqns %d, ncols %d", mtx->nrow, mtx->ncol,
neqns, ncols) ;
exit(-1) ;
}
DVcopy(neqns*ncols, y, DenseMtx_entries(mtx)) ;
}
DenseMtx_free(mtx) ;
/*
---------------------------------------------
broadcast the entries to the other processors
---------------------------------------------
*/
MPI_Bcast((void *) y, neqns*ncols, MPI_DOUBLE, 0, bridge->comm) ;
if ( bridge->msglvl > 2 ) {
fprintf(bridge->msgFile, "\n after the broadcast") ;
fflush(bridge->msgFile) ;
}
}
MPI_Barrier(bridge->comm) ;
#if MYDEBUG > 0
MARKTIME(t2) ;
time_JimMatMul += t2 - t1 ;
if ( bridge->myid == 0 ) {
fprintf(stdout, "\n (%d) JimMatMulMPI() end", count_JimMatMul) ;
fprintf(stdout, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_JimMatMul) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile,
"\n (%d) JimMatMulMPI() end", count_JimMatMul) ;
fprintf(bridge->msgFile, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_JimMatMul) ;
fflush(bridge->msgFile) ;
#endif
return ; }
/*--------------------------------------------------------------------*/
int
main ( int argc, char *argv[] )
/*
------------------------------------------------------------
make ETree objects for nested dissection on a regular grid
1 -- vertex elimination tree
2 -- fundamental supernode front tree
3 -- merge only children if possible
4 -- merge all children if possible
5 -- split large non-leaf fronts
created -- 98feb05, cca
------------------------------------------------------------
*/
{
char *outETreeFileName ;
double ops[6] ;
double t1, t2 ;
EGraph *egraph ;
ETree *etree0, *etree1, *etree2, *etree3, *etree4, *etree5 ;
FILE *msgFile ;
Graph *graph ;
int nfronts[6], nfind[6], nzf[6] ;
int maxsize, maxzeros, msglvl, n1, n2, n3, nvtx, rc, v ;
int *newToOld, *oldToNew ;
IV *nzerosIV ;
if ( argc != 9 ) {
fprintf(stdout,
"\n\n usage : %s msglvl msgFile n1 n2 n3 maxzeros maxsize outFile"
"\n msglvl -- message level"
"\n msgFile -- message file"
"\n n1 -- number of points in the first direction"
"\n n2 -- number of points in the second direction"
"\n n3 -- number of points in the third direction"
"\n maxzeros -- number of points in the third direction"
"\n maxsize -- maximum number of vertices in a front"
"\n outFile -- output file, must be *.etreef or *.etreeb"
"\n", argv[0]) ;
return(0) ;
}
msglvl = atoi(argv[1]) ;
if ( strcmp(argv[2], "stdout") == 0 ) {
msgFile = stdout ;
} else if ( (msgFile = fopen(argv[2], "a")) == NULL ) {
fprintf(stderr, "\n fatal error in %s"
"\n unable to open file %s\n",
argv[0], argv[2]) ;
return(-1) ;
}
n1 = atoi(argv[3]) ;
n2 = atoi(argv[4]) ;
n3 = atoi(argv[5]) ;
maxzeros = atoi(argv[6]) ;
maxsize = atoi(argv[7]) ;
outETreeFileName = argv[8] ;
fprintf(msgFile,
"\n %s "
"\n msglvl -- %d"
"\n msgFile -- %s"
"\n n1 -- %d"
"\n n2 -- %d"
"\n n3 -- %d"
"\n maxzeros -- %d"
"\n maxsize -- %d"
"\n outFile -- %s"
"\n",
argv[0], msglvl, argv[2], n1, n2, n3,
maxzeros, maxsize, outETreeFileName) ;
fflush(msgFile) ;
/*
----------------------------
create the grid graph object
----------------------------
*/
if ( n1 == 1 ) {
egraph = EGraph_make9P(n2, n3, 1) ;
} else if ( n2 == 1 ) {
egraph = EGraph_make9P(n1, n3, 1) ;
} else if ( n3 == 1 ) {
egraph = EGraph_make9P(n1, n2, 1) ;
} else {
egraph = EGraph_make27P(n1, n2, n3, 1) ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n %d x %d x %d grid EGraph", n1, n2, n3) ;
EGraph_writeForHumanEye(egraph, msgFile) ;
fflush(msgFile) ;
}
graph = EGraph_mkAdjGraph(egraph) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n %d x %d x %d grid Graph", n1, n2, n3) ;
Graph_writeForHumanEye(graph, msgFile) ;
fflush(msgFile) ;
}
/*
----------------------------------
get the nested dissection ordering
----------------------------------
*/
nvtx = n1*n2*n3 ;
newToOld = IVinit(nvtx, -1) ;
oldToNew = IVinit(nvtx, -1) ;
mkNDperm(n1, n2, n3, newToOld, 0, n1-1, 0, n2-1, 0, n3-1) ;
for ( v = 0 ; v < nvtx ; v++ ) {
oldToNew[newToOld[v]] = v ;
}
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n %d x %d x %d nd ordering", n1, n2, n3) ;
IVfprintf(msgFile, nvtx, oldToNew) ;
fflush(msgFile) ;
}
/*
------------------------------------------
create the vertex elimination ETree object
------------------------------------------
*/
etree0 = ETree_new() ;
ETree_initFromGraphWithPerms(etree0, graph, newToOld, oldToNew) ;
nfronts[0] = ETree_nfront(etree0) ;
nfind[0] = ETree_nFactorIndices(etree0) ;
nzf[0] = ETree_nFactorEntries(etree0, SPOOLES_SYMMETRIC) ;
ops[0] = ETree_nFactorOps(etree0, SPOOLES_REAL, SPOOLES_SYMMETRIC) ;
fprintf(msgFile,
"\n vtx tree : %8d fronts, %8d indices, %8d |L|, %12.0f ops",
nfronts[0], nfind[0], nzf[0], ops[0]) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n vertex elimination tree") ;
ETree_writeForHumanEye(etree0, msgFile) ;
fflush(msgFile) ;
}
/*
---------------------------------------------
create the fundamental supernode ETree object
---------------------------------------------
*/
nzerosIV = IV_new() ;
IV_init(nzerosIV, nvtx, NULL) ;
IV_fill(nzerosIV, 0) ;
etree1 = ETree_mergeFrontsOne(etree0, 0, nzerosIV) ;
nfronts[1] = ETree_nfront(etree1) ;
nfind[1] = ETree_nFactorIndices(etree1) ;
nzf[1] = ETree_nFactorEntries(etree1, SPOOLES_SYMMETRIC) ;
ops[1] = ETree_nFactorOps(etree1, SPOOLES_REAL, SPOOLES_SYMMETRIC) ;
fprintf(msgFile,
"\n fs tree : %8d fronts, %8d indices, %8d |L|, %12.0f ops",
nfronts[1], nfind[1], nzf[1], ops[1]) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n fundamental supernode front tree") ;
ETree_writeForHumanEye(etree1, msgFile) ;
fprintf(msgFile, "\n\n nzerosIV") ;
IV_writeForHumanEye(nzerosIV, msgFile) ;
fflush(msgFile) ;
}
/*
---------------------------
try to absorb only children
---------------------------
*/
etree2 = ETree_mergeFrontsOne(etree1, maxzeros, nzerosIV) ;
nfronts[2] = ETree_nfront(etree2) ;
nfind[2] = ETree_nFactorIndices(etree2) ;
nzf[2] = ETree_nFactorEntries(etree2, SPOOLES_SYMMETRIC) ;
ops[2] = ETree_nFactorOps(etree2, SPOOLES_REAL, SPOOLES_SYMMETRIC) ;
fprintf(msgFile,
"\n merge one : %8d fronts, %8d indices, %8d |L|, %12.0f ops",
nfronts[2], nfind[2], nzf[2], ops[2]) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front tree after mergeOne") ;
ETree_writeForHumanEye(etree2, msgFile) ;
fprintf(msgFile, "\n\n nzerosIV") ;
IV_writeForHumanEye(nzerosIV, msgFile) ;
fflush(msgFile) ;
}
/*
--------------------------
try to absorb all children
--------------------------
*/
etree3 = ETree_mergeFrontsAll(etree2, maxzeros, nzerosIV) ;
nfronts[3] = ETree_nfront(etree3) ;
nfind[3] = ETree_nFactorIndices(etree3) ;
nzf[3] = ETree_nFactorEntries(etree3, SPOOLES_SYMMETRIC) ;
ops[3] = ETree_nFactorOps(etree3, SPOOLES_REAL, SPOOLES_SYMMETRIC) ;
fprintf(msgFile,
"\n merge all : %8d fronts, %8d indices, %8d |L|, %12.0f ops",
nfronts[3], nfind[3], nzf[3], ops[3]) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front tree after mergeAll") ;
ETree_writeForHumanEye(etree3, msgFile) ;
fprintf(msgFile, "\n\n nzerosIV") ;
IV_writeForHumanEye(nzerosIV, msgFile) ;
fflush(msgFile) ;
}
/*
--------------------------------
try to absorb any other children
--------------------------------
*/
etree4 = etree3 ;
/*
etree4 = ETree_mergeFrontsAny(etree3, maxzeros, nzerosIV) ;
nfronts[4] = ETree_nfront(etree4) ;
nfind[4] = ETree_nFactorIndices(etree4) ;
nzf[4] = ETree_nFactorEntries(etree4, SPOOLES_SYMMETRIC) ;
ops[4] = ETree_nFactorOps(etree4, SPOOLES_REAL, SPOOLES_SYMMETRIC) ;
fprintf(msgFile,
"\n merge any : %8d fronts, %8d indices, %8d |L|, %12.0f ops",
nfronts[4], nfind[4], nzf[4], ops[4]) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front tree after mergeAny") ;
ETree_writeForHumanEye(etree3, msgFile) ;
fprintf(msgFile, "\n\n nzerosIV") ;
IV_writeForHumanEye(nzerosIV, msgFile) ;
fflush(msgFile) ;
}
*/
/*
--------------------
split the front tree
--------------------
*/
etree5 = ETree_splitFronts(etree4, NULL, maxsize, 0) ;
nfronts[5] = ETree_nfront(etree5) ;
nfind[5] = ETree_nFactorIndices(etree5) ;
nzf[5] = ETree_nFactorEntries(etree5, SPOOLES_SYMMETRIC) ;
ops[5] = ETree_nFactorOps(etree5, SPOOLES_REAL, SPOOLES_SYMMETRIC) ;
fprintf(msgFile,
"\n split : %8d fronts, %8d indices, %8d |L|, %12.0f ops",
nfronts[5], nfind[5], nzf[5], ops[5]) ;
if ( msglvl > 2 ) {
fprintf(msgFile, "\n\n front tree after split") ;
ETree_writeForHumanEye(etree4, msgFile) ;
fflush(msgFile) ;
}
fprintf(msgFile, "\n\n complex symmetric ops %.0f",
ETree_nFactorOps(etree5, SPOOLES_COMPLEX, SPOOLES_SYMMETRIC)) ;
/*
--------------------------
write out the ETree object
--------------------------
*/
if ( strcmp(outETreeFileName, "none") != 0 ) {
MARKTIME(t1) ;
rc = ETree_writeToFile(etree5, outETreeFileName) ;
MARKTIME(t2) ;
fprintf(msgFile, "\n CPU %9.5f : write etree to file %s",
t2 - t1, outETreeFileName) ;
if ( rc != 1 ) {
fprintf(msgFile,
"\n return value %d from ETree_writeToFile(%p,%s)",
rc, etree5, outETreeFileName) ;
}
}
/*
----------------
free the objects
----------------
*/
ETree_free(etree0) ;
ETree_free(etree1) ;
ETree_free(etree2) ;
ETree_free(etree3) ;
/*
ETree_free(etree4) ;
*/
ETree_free(etree5) ;
EGraph_free(egraph) ;
Graph_free(graph) ;
IVfree(newToOld) ;
IVfree(oldToNew) ;
IV_free(nzerosIV) ;
fprintf(msgFile, "\n") ;
fclose(msgFile) ;
return(1) ; }
/*
--------------------------------------------------
purpose -- to solve a linear system
(A - sigma*B) sol[] = rhs[]
data -- pointer to bridge data object
*pnrows -- # of rows in x[] and y[]
*pncols -- # of columns in x[] and y[]
rhs[] -- vector that holds right hand sides
NOTE: the rhs[] vector is global, not a portion
sol[] -- vector to hold solutions
NOTE: the sol[] vector is global, not a portion
note: rhs[] and sol[] can be the same array.
on return, *perror holds an error code.
created -- 98aug28, cca & jcp
--------------------------------------------------
*/
void
JimSolveMPI (
int *pnrows,
int *pncols,
double rhs[],
double sol[],
void *data,
int *perror
) {
BridgeMPI *bridge = (BridgeMPI *) data ;
DenseMtx *mtx, *newmtx ;
int irow, jj, jcol, kk, myid, ncols = *pncols,
neqns, nowned, tag = 0 ;
int *vtxmap ;
int stats[4] ;
IV *mapIV ;
#if MYDEBUG > 0
double t1, t2 ;
count_JimSolve++ ;
MARKTIME(t1) ;
if ( bridge->myid == 0 ) {
fprintf(stdout, "\n (%d) JimSolve() start", count_JimSolve) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile, "\n (%d) JimSolve() start", count_JimSolve) ;
fflush(bridge->msgFile) ;
#endif
MPI_Barrier(bridge->comm) ;
/*
---------------------------------------------
slide the owned rows of rhs down in the array
---------------------------------------------
*/
vtxmap = IV_entries(bridge->vtxmapIV) ;
neqns = bridge->neqns ;
myid = bridge->myid ;
nowned = IV_size(bridge->myownedIV) ;
for ( jcol = jj = kk = 0 ; jcol < ncols ; jcol++ ) {
for ( irow = 0 ; irow < neqns ; irow++, jj++ ) {
if ( vtxmap[irow] == myid ) {
sol[kk++] = rhs[jj] ;
}
}
}
if ( kk != nowned * ncols ) {
fprintf(stderr, "\n proc %d : kk %d, nowned %d, ncols %d",
myid, kk, nowned, ncols) ;
exit(-1) ;
}
/*
----------------------------------------
call the method that assumes local input
----------------------------------------
*/
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n calling SolveMPI()") ;
fflush(bridge->msgFile) ;
}
SolveMPI(&nowned, pncols, sol, sol, data, perror) ;
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n return from SolveMPI()") ;
fflush(bridge->msgFile) ;
}
/*
------------------------------------------
gather all the entries onto processor zero
------------------------------------------
*/
mtx = DenseMtx_new() ;
DenseMtx_init(mtx, SPOOLES_REAL, 0, 0, nowned, ncols, 1, nowned) ;
DVcopy (nowned*ncols, DenseMtx_entries(mtx), sol) ;
IVcopy(nowned, mtx->rowind, IV_entries(bridge->myownedIV)) ;
mapIV = IV_new() ;
IV_init(mapIV, neqns, NULL) ;
IV_fill(mapIV, 0) ;
IVfill(4, stats, 0) ;
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n calling DenseMtx_split()()") ;
fflush(bridge->msgFile) ;
}
newmtx = DenseMtx_MPI_splitByRows(mtx, mapIV, stats, bridge->msglvl,
bridge->msgFile, tag, bridge->comm) ;
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n return from DenseMtx_split()()") ;
fflush(bridge->msgFile) ;
}
DenseMtx_free(mtx) ;
mtx = newmtx ;
IV_free(mapIV) ;
if ( myid == 0 ) {
DVcopy(neqns*ncols, sol, DenseMtx_entries(mtx)) ;
}
DenseMtx_free(mtx) ;
/*
---------------------------------------------
broadcast the entries to the other processors
---------------------------------------------
*/
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n calling MPI_Bcast()()") ;
fflush(bridge->msgFile) ;
}
MPI_Bcast((void *) sol, neqns*ncols, MPI_DOUBLE, 0, bridge->comm) ;
if ( bridge->msglvl > 1 ) {
fprintf(bridge->msgFile, "\n return from MPI_Bcast()()") ;
fflush(bridge->msgFile) ;
}
MPI_Barrier(bridge->comm) ;
/*
------------------------------------------------------------------
set the error. (this is simple since when the spooles codes detect
a fatal error, they print out a message to stderr and exit.)
------------------------------------------------------------------
*/
*perror = 0 ;
#if MYDEBUG > 0
MARKTIME(t2) ;
time_JimSolve += t2 - t1 ;
if ( bridge->myid == 0 ) {
fprintf(stdout, "\n (%d) JimSolve() end", count_JimSolve) ;
fprintf(stdout, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_JimSolve) ;
fflush(stdout) ;
}
#endif
#if MYDEBUG > 1
fprintf(bridge->msgFile, "\n (%d) JimSolve() end", count_JimSolve) ;
fprintf(bridge->msgFile, ", %8.3f seconds, %8.3f total time",
t2 - t1, time_JimSolve) ;
fflush(bridge->msgFile) ;
#endif
return ; }