static
int
_SCOTCH_METIS_PartGraph (
const int * const n,
const int * const xadj,
const int * const adjncy,
const int * const vwgt,
const int * const adjwgt,
const int * const numflag,
const int * const nparts,
int * const part)
{
SCOTCH_Graph grafdat; /* Scotch graph object to interface with libScotch */
SCOTCH_Strat stradat;
SCOTCH_Num baseval;
SCOTCH_Num vertnbr;
int o;
if (sizeof (SCOTCH_Num) != sizeof (int)) {
errorPrint ("METIS_PartGraph* (as of SCOTCH): SCOTCH_Num type should equate to int");
return (1);
}
SCOTCH_graphInit (&grafdat);
baseval = *numflag;
vertnbr = *n;
o = 1; /* Assume something will go wrong */
if (SCOTCH_graphBuild (&grafdat,
baseval, vertnbr, xadj, xadj + 1, vwgt, NULL,
xadj[vertnbr] - baseval, adjncy, adjwgt) == 0) {
SCOTCH_stratInit (&stradat);
#ifdef SCOTCH_DEBUG_ALL
if (SCOTCH_graphCheck (&grafdat) == 0) /* TRICK: next instruction called only if graph is consistent */
#endif /* SCOTCH_DEBUG_ALL */
o = SCOTCH_graphPart (&grafdat, *nparts, &stradat, part);
SCOTCH_stratExit (&stradat);
}
SCOTCH_graphExit (&grafdat);
if (baseval != 0) { /* MeTiS part array is based, Scotch is not */
SCOTCH_Num vertnum;
for (vertnum = 0; vertnum < vertnbr; vertnum ++)
part[vertnum] += baseval;
}
return (o);
}
void setGraph(const ordinal_type m,
const size_type_array rptr,
const ordinal_type_array cidx) {
_is_ordered = false;
_cblk = 0;
/// Scotch graph spec
/// - no diagonals, symmetric
_base = 0;
_m = m;
_nnz = rptr[m];
_rptr = rptr;
_cidx = cidx;
_perm = ordinal_type_array("Scotch::PermutationArray", _m);
_peri = ordinal_type_array("Scotch::InvPermutationArray", _m);
_range = ordinal_type_array("Scotch::RangeArray", _m);
_tree = ordinal_type_array("Scotch::TreeArray", _m);
_strat = 0;
_level = 0;
int ierr = 0;
ordinal_type *rptr_ptr = reinterpret_cast<ordinal_type*>(_rptr.ptr_on_device());
ordinal_type *cidx_ptr = reinterpret_cast<ordinal_type*>(_cidx.ptr_on_device());
ierr = SCOTCH_graphInit(&_graph);
TACHO_TEST_FOR_ABORT(ierr, "Failed in SCOTCH_graphInit");
ierr = SCOTCH_graphBuild(&_graph, // scotch graph
_base, // base value
_m, // # of vertices
rptr_ptr, // column index array pointer begin
rptr_ptr+1, // column index array pointer end
NULL, // weights on vertices (optional)
NULL, // label array on vertices (optional)
_nnz, // # of nonzeros
cidx_ptr, // column index array
NULL); // edge load array (optional)
TACHO_TEST_FOR_ABORT(ierr, "Failed in SCOTCH_graphBuild");
ierr = SCOTCH_graphCheck(&_graph);
TACHO_TEST_FOR_ABORT(ierr, "Failed in SCOTCH_graphCheck");
}
int
METISNAMEU(METIS_NodeWND) (
const SCOTCH_Num * const n,
const SCOTCH_Num * const xadj,
const SCOTCH_Num * const adjncy,
const SCOTCH_Num * const vwgt,
const SCOTCH_Num * const numflag,
const SCOTCH_Num * const options,
SCOTCH_Num * const perm,
SCOTCH_Num * const iperm)
{
SCOTCH_Graph grafdat; /* Scotch graph object to interface with libScotch */
SCOTCH_Ordering ordedat; /* Scotch ordering object to interface with libScotch */
SCOTCH_Strat stradat;
int o;
o = METIS_ERROR; /* Assume an error */
SCOTCH_graphInit (&grafdat);
if (SCOTCH_graphBuild (&grafdat,
*numflag, *n, xadj, xadj + 1, vwgt, NULL,
xadj[*n] - *numflag, adjncy, NULL) == 0) {
SCOTCH_stratInit (&stradat);
#ifdef SCOTCH_DEBUG_ALL
if (SCOTCH_graphCheck (&grafdat) == 0) /* TRICK: next instruction called only if graph is consistent */
#endif /* SCOTCH_DEBUG_ALL */
{
if (SCOTCH_graphOrderInit (&grafdat, &ordedat, iperm, perm, /* MeTiS and Scotch have opposite definitions for (inverse) permutations */
NULL, NULL, NULL) == 0) {
if (SCOTCH_graphOrderCompute (&grafdat, &ordedat, &stradat) == 0)
o = METIS_OK;
SCOTCH_graphOrderExit (&grafdat, &ordedat);
}
}
SCOTCH_stratExit (&stradat);
}
SCOTCH_graphExit (&grafdat);
return (o);
}
// Call scotch with options from dictionary.
Foam::label Foam::scotchDecomp::decompose
(
const List<int>& adjncy,
const List<int>& xadj,
const scalarField& cWeights,
List<int>& finalDecomp
)
{
// Dump graph
if (decompositionDict_.found("scotchCoeffs"))
{
const dictionary& scotchCoeffs =
decompositionDict_.subDict("scotchCoeffs");
if (scotchCoeffs.found("writeGraph"))
{
Switch writeGraph(scotchCoeffs.lookup("writeGraph"));
if (writeGraph)
{
OFstream str(mesh_.time().path() / mesh_.name() + ".grf");
Info<< "Dumping Scotch graph file to " << str.name() << endl
<< "Use this in combination with gpart." << endl;
label version = 0;
str << version << nl;
// Numer of vertices
str << xadj.size()-1 << ' ' << adjncy.size() << nl;
// Numbering starts from 0
label baseval = 0;
// Has weights?
label hasEdgeWeights = 0;
label hasVertexWeights = 0;
label numericflag = 10*hasEdgeWeights+hasVertexWeights;
str << baseval << ' ' << numericflag << nl;
for (label cellI = 0; cellI < xadj.size()-1; cellI++)
{
label start = xadj[cellI];
label end = xadj[cellI+1];
str << end-start;
for (label i = start; i < end; i++)
{
str << ' ' << adjncy[i];
}
str << nl;
}
}
}
}
// Strategy
// ~~~~~~~~
// Default.
SCOTCH_Strat stradat;
check(SCOTCH_stratInit(&stradat), "SCOTCH_stratInit");
if (decompositionDict_.found("scotchCoeffs"))
{
const dictionary& scotchCoeffs =
decompositionDict_.subDict("scotchCoeffs");
string strategy;
if (scotchCoeffs.readIfPresent("strategy", strategy))
{
if (debug)
{
Info<< "scotchDecomp : Using strategy " << strategy << endl;
}
SCOTCH_stratGraphMap(&stradat, strategy.c_str());
//fprintf(stdout, "S\tStrat=");
//SCOTCH_stratSave(&stradat, stdout);
//fprintf(stdout, "\n");
}
}
// Graph
// ~~~~~
List<int> velotab;
// Check for externally provided cellweights and if so initialise weights
scalar minWeights = gMin(cWeights);
if (cWeights.size() > 0)
{
if (minWeights <= 0)
{
WarningIn
(
"scotchDecomp::decompose"
"(const pointField&, const scalarField&)"
) << "Illegal minimum weight " << minWeights
<< endl;
}
if (cWeights.size() != xadj.size()-1)
{
FatalErrorIn
(
"scotchDecomp::decompose"
"(const pointField&, const scalarField&)"
) << "Number of cell weights " << cWeights.size()
<< " does not equal number of cells " << xadj.size()-1
<< exit(FatalError);
}
// Convert to integers.
velotab.setSize(cWeights.size());
forAll(velotab, i)
{
velotab[i] = int(cWeights[i]/minWeights);
}
}
SCOTCH_Graph grafdat;
check(SCOTCH_graphInit(&grafdat), "SCOTCH_graphInit");
check
(
SCOTCH_graphBuild
(
&grafdat,
0, // baseval, c-style numbering
xadj.size()-1, // vertnbr, nCells
xadj.begin(), // verttab, start index per cell into adjncy
&xadj[1], // vendtab, end index ,,
velotab.begin(), // velotab, vertex weights
NULL, // vlbltab
adjncy.size(), // edgenbr, number of arcs
adjncy.begin(), // edgetab
NULL // edlotab, edge weights
),
"SCOTCH_graphBuild"
);
check(SCOTCH_graphCheck(&grafdat), "SCOTCH_graphCheck");
// Architecture
// ~~~~~~~~~~~~
// (fully connected network topology since using switch)
SCOTCH_Arch archdat;
check(SCOTCH_archInit(&archdat), "SCOTCH_archInit");
List<label> processorWeights;
if (decompositionDict_.found("scotchCoeffs"))
{
const dictionary& scotchCoeffs =
decompositionDict_.subDict("scotchCoeffs");
scotchCoeffs.readIfPresent("processorWeights", processorWeights);
}
if (processorWeights.size())
{
if (debug)
{
Info<< "scotchDecomp : Using procesor weights " << processorWeights
<< endl;
}
check
(
SCOTCH_archCmpltw(&archdat, nProcessors_, processorWeights.begin()),
"SCOTCH_archCmpltw"
);
}
else
{
check
(
SCOTCH_archCmplt(&archdat, nProcessors_),
"SCOTCH_archCmplt"
);
}
//SCOTCH_Mapping mapdat;
//SCOTCH_graphMapInit(&grafdat, &mapdat, &archdat, NULL);
//SCOTCH_graphMapCompute(&grafdat, &mapdat, &stradat); /* Perform mapping */
//SCOTCH_graphMapExit(&grafdat, &mapdat);
// Hack:switch off fpu error trapping
# ifdef LINUX_GNUC
int oldExcepts = fedisableexcept
(
FE_DIVBYZERO
| FE_INVALID
| FE_OVERFLOW
);
# endif
finalDecomp.setSize(xadj.size()-1);
finalDecomp = 0;
check
(
SCOTCH_graphMap
(
&grafdat,
&archdat,
&stradat, // const SCOTCH_Strat *
finalDecomp.begin() // parttab
),
"SCOTCH_graphMap"
);
# ifdef LINUX_GNUC
feenableexcept(oldExcepts);
# endif
//finalDecomp.setSize(xadj.size()-1);
//check
//(
// SCOTCH_graphPart
// (
// &grafdat,
// nProcessors_, // partnbr
// &stradat, // const SCOTCH_Strat *
// finalDecomp.begin() // parttab
// ),
// "SCOTCH_graphPart"
//);
// Release storage for graph
SCOTCH_graphExit(&grafdat);
// Release storage for strategy
SCOTCH_stratExit(&stradat);
// Release storage for network topology
SCOTCH_archExit(&archdat);
return 0;
}
void ScotchRefineLB::work(LDStats *stats) {
/** ========================== INITIALIZATION ============================= */
ProcArray *parr = new ProcArray(stats);
ObjGraph *ogr = new ObjGraph(stats);
int cost_array[10] = {64, 256, 512, 1024, 2048, 4096, 8192, 16384, 32768, 65536};
/** ============================= STRATEGY ================================ */
// convert ObjGraph to the Scotch graph
SCOTCH_Num baseval = 0; // starting index of vertices
SCOTCH_Num vertnbr = ogr->vertices.size(); // number of vertices
SCOTCH_Num edgenbr = 0; // number of edges
SCOTCH_Num *oldpemap = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * vertnbr);
double maxLoad = 0.0;
double minLoad = 0.0;
if (vertnbr > 0) {
minLoad = ogr->vertices[baseval].getVertexLoad();
}
long maxBytes = 1;
int i, j, k, vert;
/** remove duplicate edges from recvFrom */
for(i = baseval; i < vertnbr; i++) {
for(j = 0; j < ogr->vertices[i].sendToList.size(); j++) {
vert = ogr->vertices[i].sendToList[j].getNeighborId();
for(k = 0; k < ogr->vertices[i].recvFromList.size(); k++) {
if(ogr->vertices[i].recvFromList[k].getNeighborId() == vert) {
ogr->vertices[i].sendToList[j].setNumBytes(ogr->vertices[i].sendToList[j].getNumBytes() + ogr->vertices[i].recvFromList[k].getNumBytes());
ogr->vertices[i].recvFromList.erase(ogr->vertices[i].recvFromList.begin() + k);
}
}
}
}
/** the object load is normalized to an integer between 0 and 256 */
for(i = baseval; i < vertnbr; i++) {
if(ogr->vertices[i].getVertexLoad() > maxLoad)
maxLoad = ogr->vertices[i].getVertexLoad();
if (ogr->vertices[i].getVertexLoad() < minLoad) {
minLoad = ogr->vertices[i].getVertexLoad();
}
edgenbr += ogr->vertices[i].sendToList.size() + ogr->vertices[i].recvFromList.size();
oldpemap[i] = ogr->vertices[i].getCurrentPe();
}
for(i = baseval; i < vertnbr; i++) {
for(j = 0; j < ogr->vertices[i].sendToList.size(); j++) {
if (ogr->vertices[i].sendToList[j].getNumBytes() > maxBytes) {
maxBytes = ogr->vertices[i].sendToList[j].getNumBytes();
}
}
for(j = 0; j < ogr->vertices[i].recvFromList.size(); j++) {
if (ogr->vertices[i].recvFromList[j].getNumBytes() > maxBytes) {
maxBytes = ogr->vertices[i].recvFromList[j].getNumBytes();
}
}
}
/* adjacency list */
SCOTCH_Num *verttab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * (vertnbr+1));
/* loads of vertices */
SCOTCH_Num *velotab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * vertnbr);
/* id of the neighbors */
SCOTCH_Num *edgetab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * edgenbr);
/* number of bytes exchanged */
SCOTCH_Num *edlotab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * edgenbr);
int edgeNum = 0;
double ratio = 256.0/maxLoad;
double byteRatio = 1024.0/maxBytes;
for(i = baseval; i < vertnbr; i++) {
verttab[i] = edgeNum;
velotab[i] = (int)ceil(ogr->vertices[i].getVertexLoad() * ratio);
for(j = 0; j < ogr->vertices[i].sendToList.size(); j++) {
edgetab[edgeNum] = ogr->vertices[i].sendToList[j].getNeighborId();
edlotab[edgeNum] = (int) ceil(ogr->vertices[i].sendToList[j].getNumBytes()
* byteRatio);
edgeNum++;
}
for(j = 0; j < ogr->vertices[i].recvFromList.size(); j++) {
edgetab[edgeNum] = ogr->vertices[i].recvFromList[j].getNeighborId();
edlotab[edgeNum] = (int)
ceil(ogr->vertices[i].recvFromList[j].getNumBytes() * byteRatio);
edgeNum++;
}
}
verttab[i] = edgeNum;
CkAssert(edgeNum == edgenbr);
SCOTCH_Graph graph; // Graph to partition
SCOTCH_Strat strat; // Strategy to achieve partitioning
/* Initialize data structures */
SCOTCH_graphInit (&graph);
SCOTCH_stratInit (&strat);
SCOTCH_graphBuild (&graph, baseval, vertnbr, verttab, NULL, velotab, NULL, edgenbr, edgetab, edlotab);
SCOTCH_graphCheck (&graph);
double migration_cost = 1024.0;
if (step() == 0) {
SCOTCH_stratGraphMapBuild (&strat, SCOTCH_STRATBALANCE, parr->procs.size (), 0.01);
} else {
SCOTCH_stratGraphMapBuild (&strat, SCOTCH_STRATBALANCE | SCOTCH_STRATREMAP, parr->procs.size (), 0.01);
}
SCOTCH_Num *pemap = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * vertnbr);
// Takes as input the graph, arch graph, strategy, migration cost in
// double, old mapping and new mapping
if (step() == 0) {
SCOTCH_graphPart(&graph, parr->procs.size(), &strat, pemap);
} else {
SCOTCH_graphRepart(&graph, parr->procs.size(), oldpemap, migration_cost, NULL, &strat, pemap);
}
SCOTCH_graphExit (&graph);
SCOTCH_stratExit (&strat);
free(verttab);
free(velotab);
free(edgetab);
free(edlotab);
for(i = baseval; i < vertnbr; i++) {
if(pemap[i] != ogr->vertices[i].getCurrentPe())
ogr->vertices[i].setNewPe(pemap[i]);
}
free(pemap);
free(oldpemap);
/** ============================== CLEANUP ================================ */
ogr->convertDecisions(stats);
delete parr;
delete ogr;
}
static PetscErrorCode MatPartitioningApply_PTScotch_Private(MatPartitioning part, PetscBool useND, IS *partitioning)
{
MPI_Comm pcomm,comm;
MatPartitioning_PTScotch *scotch = (MatPartitioning_PTScotch*)part->data;
PetscErrorCode ierr;
PetscMPIInt rank;
Mat mat = part->adj;
Mat_MPIAdj *adj = (Mat_MPIAdj*)mat->data;
PetscBool flg,distributed;
PetscBool proc_weight_flg;
PetscInt i,j,p,bs=1,nold;
PetscInt *NDorder = NULL;
PetscReal *vwgttab,deltval;
SCOTCH_Num *locals,*velotab,*veloloctab,*edloloctab,vertlocnbr,edgelocnbr,nparts=part->n;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)part,&pcomm);CHKERRQ(ierr);
/* Duplicate the communicator to be sure that PTSCOTCH attribute caching does not interfere with PETSc. */
ierr = MPI_Comm_dup(pcomm,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIADJ,&flg);CHKERRQ(ierr);
if (!flg) {
/* bs indicates if the converted matrix is "reduced" from the original and hence the
resulting partition results need to be stretched to match the original matrix */
nold = mat->rmap->n;
ierr = MatConvert(mat,MATMPIADJ,MAT_INITIAL_MATRIX,&mat);CHKERRQ(ierr);
if (mat->rmap->n > 0) bs = nold/mat->rmap->n;
adj = (Mat_MPIAdj*)mat->data;
}
proc_weight_flg = PETSC_TRUE;
ierr = PetscOptionsGetBool(NULL, NULL, "-mat_partitioning_ptscotch_proc_weight", &proc_weight_flg, NULL);CHKERRQ(ierr);
ierr = PetscMalloc1(mat->rmap->n+1,&locals);CHKERRQ(ierr);
if (useND) {
#if defined(PETSC_HAVE_SCOTCH_PARMETIS_V3_NODEND)
PetscInt *sizes, *seps, log2size, subd, *level, base = 0;
PetscMPIInt size;
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
log2size = PetscLog2Real(size);
subd = PetscPowInt(2,log2size);
if (subd != size) SETERRQ(comm,PETSC_ERR_SUP,"Only power of 2 communicator sizes");
ierr = PetscMalloc1(mat->rmap->n,&NDorder);CHKERRQ(ierr);
ierr = PetscMalloc3(2*size,&sizes,4*size,&seps,size,&level);CHKERRQ(ierr);
SCOTCH_ParMETIS_V3_NodeND(mat->rmap->range,adj->i,adj->j,&base,NULL,NDorder,sizes,&comm);
ierr = MatPartitioningSizesToSep_Private(subd,sizes,seps,level);CHKERRQ(ierr);
for (i=0;i<mat->rmap->n;i++) {
PetscInt loc;
ierr = PetscFindInt(NDorder[i],2*subd,seps,&loc);CHKERRQ(ierr);
if (loc < 0) {
loc = -(loc+1);
if (loc%2) { /* part of subdomain */
locals[i] = loc/2;
} else {
ierr = PetscFindInt(NDorder[i],2*(subd-1),seps+2*subd,&loc);CHKERRQ(ierr);
loc = loc < 0 ? -(loc+1)/2 : loc/2;
locals[i] = level[loc];
}
} else locals[i] = loc/2;
}
ierr = PetscFree3(sizes,seps,level);CHKERRQ(ierr);
#else
SETERRQ(pcomm,PETSC_ERR_SUP,"Need libptscotchparmetis.a compiled with -DSCOTCH_METIS_PREFIX");
#endif
} else {
velotab = NULL;
if (proc_weight_flg) {
ierr = PetscMalloc1(nparts,&vwgttab);CHKERRQ(ierr);
ierr = PetscMalloc1(nparts,&velotab);CHKERRQ(ierr);
for (j=0; j<nparts; j++) {
if (part->part_weights) vwgttab[j] = part->part_weights[j]*nparts;
else vwgttab[j] = 1.0;
}
for (i=0; i<nparts; i++) {
deltval = PetscAbsReal(vwgttab[i]-PetscFloorReal(vwgttab[i]+0.5));
if (deltval>0.01) {
for (j=0; j<nparts; j++) vwgttab[j] /= deltval;
}
}
for (i=0; i<nparts; i++) velotab[i] = (SCOTCH_Num)(vwgttab[i] + 0.5);
ierr = PetscFree(vwgttab);CHKERRQ(ierr);
}
vertlocnbr = mat->rmap->range[rank+1] - mat->rmap->range[rank];
edgelocnbr = adj->i[vertlocnbr];
veloloctab = part->vertex_weights;
edloloctab = adj->values;
/* detect whether all vertices are located at the same process in original graph */
for (p = 0; !mat->rmap->range[p+1] && p < nparts; ++p);
distributed = (mat->rmap->range[p+1] == mat->rmap->N) ? PETSC_FALSE : PETSC_TRUE;
if (distributed) {
SCOTCH_Arch archdat;
SCOTCH_Dgraph grafdat;
SCOTCH_Dmapping mappdat;
SCOTCH_Strat stradat;
ierr = SCOTCH_dgraphInit(&grafdat,comm);CHKERRQ(ierr);
ierr = SCOTCH_dgraphBuild(&grafdat,0,vertlocnbr,vertlocnbr,adj->i,adj->i+1,veloloctab,
NULL,edgelocnbr,edgelocnbr,adj->j,NULL,edloloctab);CHKERRQ(ierr);
#if defined(PETSC_USE_DEBUG)
ierr = SCOTCH_dgraphCheck(&grafdat);CHKERRQ(ierr);
#endif
ierr = SCOTCH_archInit(&archdat);CHKERRQ(ierr);
ierr = SCOTCH_stratInit(&stradat);CHKERRQ(ierr);
ierr = SCOTCH_stratDgraphMapBuild(&stradat,scotch->strategy,nparts,nparts,scotch->imbalance);CHKERRQ(ierr);
if (velotab) {
ierr = SCOTCH_archCmpltw(&archdat,nparts,velotab);CHKERRQ(ierr);
} else {
ierr = SCOTCH_archCmplt( &archdat,nparts);CHKERRQ(ierr);
}
ierr = SCOTCH_dgraphMapInit(&grafdat,&mappdat,&archdat,locals);CHKERRQ(ierr);
ierr = SCOTCH_dgraphMapCompute(&grafdat,&mappdat,&stradat);CHKERRQ(ierr);
SCOTCH_dgraphMapExit(&grafdat,&mappdat);
SCOTCH_archExit(&archdat);
SCOTCH_stratExit(&stradat);
SCOTCH_dgraphExit(&grafdat);
} else if (rank == p) {
SCOTCH_Graph grafdat;
SCOTCH_Strat stradat;
ierr = SCOTCH_graphInit(&grafdat);CHKERRQ(ierr);
ierr = SCOTCH_graphBuild(&grafdat,0,vertlocnbr,adj->i,adj->i+1,veloloctab,NULL,edgelocnbr,adj->j,edloloctab);CHKERRQ(ierr);
#if defined(PETSC_USE_DEBUG)
ierr = SCOTCH_graphCheck(&grafdat);CHKERRQ(ierr);
#endif
ierr = SCOTCH_stratInit(&stradat);CHKERRQ(ierr);
ierr = SCOTCH_stratGraphMapBuild(&stradat,scotch->strategy,nparts,scotch->imbalance);CHKERRQ(ierr);
ierr = SCOTCH_graphPart(&grafdat,nparts,&stradat,locals);CHKERRQ(ierr);
SCOTCH_stratExit(&stradat);
SCOTCH_graphExit(&grafdat);
}
ierr = PetscFree(velotab);CHKERRQ(ierr);
}
ierr = MPI_Comm_free(&comm);CHKERRQ(ierr);
if (bs > 1) {
PetscInt *newlocals;
ierr = PetscMalloc1(bs*mat->rmap->n,&newlocals);CHKERRQ(ierr);
for (i=0;i<mat->rmap->n;i++) {
for (j=0;j<bs;j++) {
newlocals[bs*i+j] = locals[i];
}
}
ierr = PetscFree(locals);CHKERRQ(ierr);
ierr = ISCreateGeneral(pcomm,bs*mat->rmap->n,newlocals,PETSC_OWN_POINTER,partitioning);CHKERRQ(ierr);
} else {
ierr = ISCreateGeneral(pcomm,mat->rmap->n,locals,PETSC_OWN_POINTER,partitioning);CHKERRQ(ierr);
}
if (useND) {
IS ndis;
if (bs > 1) {
ierr = ISCreateBlock(pcomm,bs,mat->rmap->n,NDorder,PETSC_OWN_POINTER,&ndis);CHKERRQ(ierr);
} else {
ierr = ISCreateGeneral(pcomm,mat->rmap->n,NDorder,PETSC_OWN_POINTER,&ndis);CHKERRQ(ierr);
}
ierr = ISSetPermutation(ndis);CHKERRQ(ierr);
ierr = PetscObjectCompose((PetscObject)(*partitioning),"_petsc_matpartitioning_ndorder",(PetscObject)ndis);CHKERRQ(ierr);
ierr = ISDestroy(&ndis);CHKERRQ(ierr);
}
if (!flg) {
ierr = MatDestroy(&mat);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
bool
ScotchSplitter(unsigned dim,
const int* ptRows, const int* indCols,
unsigned& nbMaxLevels, unsigned minSize,
int* loc2glob, int* glob2loc, int& nbDoms,
int*& ptOnDomains, int*& sizeOfDomains,
bool checkData, const bool verbose, FILE *fp)
{
int ierr;
// check consistency between Scotch and Dissection library
CHECK(sizeof(int) == sizeof(SCOTCH_Num),
"Incompatible integer representation between Scotch and Dissection");
if (verbose) {
int vers, rela, patc;
SCOTCH_version(&vers, &rela, &patc);
diss_printf(verbose, fp, "%s %d : Soctch version : %d.%d.%d\n",
__FILE__, __LINE__, vers, rela, patc);
}
// Allocate and initialize the Scotch graph
SCOTCH_Graph ptGraph;
ierr = SCOTCH_graphInit(&ptGraph);
CHECK(ierr==0,"Fail initializing Scotch graph");
// TRACE("Building Scotch graph\n");
ierr = SCOTCH_graphBuild(&ptGraph,
(SCOTCH_Num)0, // offset,
(SCOTCH_Num)dim,
(const SCOTCH_Num*)ptRows,
(const SCOTCH_Num*)ptRows+1,
NULL, NULL, (SCOTCH_Num)ptRows[dim],
(const SCOTCH_Num*)indCols, NULL);
// if (verbose) {
// fprintf(fp, "%s %d : 1 : ptRows = %p indCols = %p\n",
// __FILE__, __LINE__, (void *)ptRows, (void *)indCols);
// }
CHECK(ierr==0,"Scotch graph building failed !");
if (checkData) {
//TRACE("Check Scotch graph\n");
ierr = SCOTCH_graphCheck(&ptGraph);
if (ierr) {
diss_printf(verbose, fp,
"Failed the checking of the graph : bad data ?\n");
SCOTCH_graphFree(&ptGraph);
return false;
}
}
// Allocate and initialize the strategy wanted for Scotch
// TRACE("Initialize strategy for splitting\n");
SCOTCH_Strat ptStrat;
ierr = SCOTCH_stratInit(&ptStrat);
CHECK(ierr==0,
"Failed initializing Scotch strategy structure");
char *str_Strat = new char[1024];
int nbLvls = std::min(unsigned(nbMaxLevels),
highestbit(unsigned(dim/minSize)));
diss_printf(verbose, fp, "nbLevels = %d\n", nbLvls);
sprintf(str_Strat,"c{rat=0.7,cpr=n{sep=/((levl<%d)|(vert>%d))?m{type=h,rat=0.7,vert=100,low=h{pass=10},asc=b{width=3,bnd=f{bal=0.2},org=h{pass=10}f{bal=0.2}}}|m{type=h,rat=0.7,vert=100,low=h{pass=10},asc=b{width=3,bnd=f{bal=0.2},org=h{pass=10}f{bal=0.2}}};,ole=f{cmin=%d,cmax=%d,frat=0.05},ose=s},unc=n{sep=/(levl<%d)?(m{type=h,rat=0.7,vert=100,low=h{pass=10},asc=b{width=3,bnd=f{bal=0.2},org=h{pass=10}f{bal=0.2}}})|m{type=h,rat=0.7,vert=100,low=h{pass=10},asc=b{width=3,bnd=f{bal=0.2},org=h{pass=10}f{bal=0.2}}};,ole=f{cmin=%d,cmax=%d,frat=0.05},ose=s}}",
nbLvls-1,2*minSize-1,minSize,dim,nbLvls-1,minSize,dim);
// DBG_PRINT("Strategy string : %s\n", str_Strat);
ierr = SCOTCH_stratGraphOrder(&ptStrat, str_Strat);
delete [] str_Strat;
CHECK(ierr==0,
"Failed build graph ordering strategy for Scotch");
// Ordering with nested bisection :
// TRACE("Split the graph\n");
int* rangtab = new int[dim+1];
int* treetab = new int[dim];
int nbSplitDoms;
bool repeat = true;
int lastCompleteLevel;
int *levels, *nbDomsPerLevels;
SCOTCH_randomReset();
while (repeat) {
ierr = SCOTCH_graphOrder(&ptGraph, &ptStrat,
(SCOTCH_Num*)loc2glob,
(SCOTCH_Num*)glob2loc,
(SCOTCH_Num*)&nbSplitDoms,
(SCOTCH_Num*)rangtab,
(SCOTCH_Num*)treetab);
if (ierr) {
diss_printf(verbose, fp, "Failed reordering sparse matrix graph !\n");
SCOTCH_stratExit(&ptStrat);
SCOTCH_graphFree(&ptGraph);
return false;
}
levels = new int[nbSplitDoms];
// int *nbDomsPerLevels;// = new int[nbLvls];
unsigned nbLvlsScotch= compLevelOfDoms(nbSplitDoms, nbLvls, treetab, levels,
nbDomsPerLevels);
/** Search last level where number of domains is a power of two
*/
lastCompleteLevel = 0;
while ((lastCompleteLevel<nbLvlsScotch) &&
((1<<lastCompleteLevel) == nbDomsPerLevels[lastCompleteLevel]) ) {
lastCompleteLevel ++;
}
lastCompleteLevel = std::min(lastCompleteLevel, nbLvls);
nbDoms = (1<<lastCompleteLevel)-1;
// Search where start each domain per bisection level
// and compute the size of each subdomain :
// int indDom = 0;
bool flag_size_check = false;
for (int i = 0; i < nbSplitDoms; i++) {
int sz = 0;
while (levels[i]>=lastCompleteLevel) {
sz += rangtab[i+1]-rangtab[i];
i++;
}
if (sz+rangtab[i+1]-rangtab[i] <= TOO_SMALL) {
flag_size_check = true;
break;
}
//DBG_PRINT("Domain %d begin at %d\n",indDom+1,begDom);
} // loop : i
if (!flag_size_check) {
repeat = false;
break;
}
else {
delete [] levels;
delete [] nbDomsPerLevels;
}
} // while (repeat)
ptOnDomains = new int[nbDoms+1];
sizeOfDomains = new int[nbDoms];
memset(sizeOfDomains, 0, nbDoms*sizeof(int));
int* indDomPerLevel = new int[lastCompleteLevel+1];
memset(indDomPerLevel,0,(lastCompleteLevel+1)*sizeof(int));
int begDom = 0;
for (int i = 0; i < nbSplitDoms; i++) {
int sz = 0;
while (levels[i]>=lastCompleteLevel) {
sz += rangtab[i+1]-rangtab[i];
i++;
}
int indDom = (1<<levels[i])-1+indDomPerLevel[levels[i]];
// DBG_PRINT("level %d : current dom = %d\n", levels[i],indDom+1);
ptOnDomains[indDom] = begDom;
//DBG_PRINT("Domain %d begin at %d\n",indDom+1,begDom);
sizeOfDomains[indDom] = sz+rangtab[i+1]-rangtab[i];
//DBG_PRINT("Domain %d size of %d\n",indDom+1,sizeOfDomains[indDom]);
begDom += sizeOfDomains[indDom];
indDomPerLevel[levels[i]] += 1;
}
diss_printf(verbose, fp, "%s %d : indDomPerlevel[lastCompleteLevel] = %d\n",
__FILE__, __LINE__, indDomPerLevel[lastCompleteLevel]);
ptOnDomains[nbDoms] = dim;
nbMaxLevels = lastCompleteLevel;
delete [] indDomPerLevel;
delete [] nbDomsPerLevels;
delete [] levels;
delete [] treetab;
delete [] rangtab;
// Cleaning all Scotch structures :
SCOTCH_stratExit(&ptStrat);
SCOTCH_graphFree(&ptGraph);
return true;
}
static PetscErrorCode MatPartitioningApply_Scotch(MatPartitioning part, IS * partitioning)
{
PetscErrorCode ierr;
int *parttab, *locals = PETSC_NULL, rank, i, size;
size_t j;
Mat mat = part->adj, matMPI, matSeq;
int nb_locals = mat->rmap->n;
Mat_MPIAdj *adj = (Mat_MPIAdj *) mat->data;
MatPartitioning_Scotch *scotch = (MatPartitioning_Scotch *) part->data;
PetscTruth flg;
#ifdef PETSC_HAVE_UNISTD_H
int fd_stdout, fd_pipe[2], count,err;
#endif
PetscFunctionBegin;
/* check if the matrix is sequential, use MatGetSubMatrices if necessary */
ierr = MPI_Comm_size(((PetscObject)mat)->comm, &size);CHKERRQ(ierr);
ierr = PetscTypeCompare((PetscObject) mat, MATMPIADJ, &flg);CHKERRQ(ierr);
if (size > 1) {
int M, N;
IS isrow, iscol;
Mat *A;
if (flg) {
SETERRQ(0, "Distributed matrix format MPIAdj is not supported for sequential partitioners");
}
PetscPrintf(((PetscObject)part)->comm, "Converting distributed matrix to sequential: this could be a performance loss\n");CHKERRQ(ierr);
ierr = MatGetSize(mat, &M, &N);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF, M, 0, 1, &isrow);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol);CHKERRQ(ierr);
ierr = MatGetSubMatrices(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, &A);CHKERRQ(ierr);
matSeq = *A;
ierr = PetscFree(A);CHKERRQ(ierr);
ierr = ISDestroy(isrow);CHKERRQ(ierr);
ierr = ISDestroy(iscol);CHKERRQ(ierr);
} else
matSeq = mat;
/* convert the the matrix to MPIADJ type if necessary */
if (!flg) {
ierr = MatConvert(matSeq, MATMPIADJ, MAT_INITIAL_MATRIX, &matMPI);CHKERRQ(ierr);
} else {
matMPI = matSeq;
}
adj = (Mat_MPIAdj *) matMPI->data; /* finaly adj contains adjacency graph */
ierr = MPI_Comm_rank(((PetscObject)part)->comm, &rank);CHKERRQ(ierr);
{
/* definition of Scotch library arguments */
SCOTCH_Strat stratptr; /* scotch strategy */
SCOTCH_Graph grafptr; /* scotch graph */
#if defined(DOES_NOT_COMPILE_DUE_TO_BROKEN_INTERFACE)
int vertnbr = mat->rmap->N; /* number of vertices in full graph */
int *verttab = adj->i; /* start of edge list for each vertex */
int *edgetab = adj->j; /* edge list data */
int edgenbr = adj->nz; /* number of edges */
int *velotab = NULL; /* not used by petsc interface */
int *vlbltab = NULL;
int *edlotab = NULL;
int flagval = 3; /* (cf doc scotch no weight edge & vertices) */
#endif
int baseval = 0; /* 0 for C array indexing */
char strategy[256];
ierr = PetscMalloc((mat->rmap->N) * sizeof(int), &parttab);CHKERRQ(ierr);
/* redirect output to buffer scotch -> mesg_log */
#ifdef PETSC_HAVE_UNISTD_H
fd_stdout = dup(1);
pipe(fd_pipe);
close(1);
dup2(fd_pipe[1], 1);
ierr = PetscMalloc(SIZE_LOG * sizeof(char), &(scotch->mesg_log));CHKERRQ(ierr);
#endif
/* library call */
/* Construction of the scotch graph object */
ierr = SCOTCH_graphInit(&grafptr);
#if defined(DOES_NOT_COMPILE_DUE_TO_BROKEN_INTERFACE)
ierr = SCOTCH_graphBuild((SCOTCH_Graph *) &grafptr,
(const SCOTCH_Num) vertnbr,
(const SCOTCH_Num) verttab,
(const SCOTCH_Num *)velotab,
(const SCOTCH_Num *)vlbltab,
(const SCOTCH_Num *)edgenbr,
(const SCOTCH_Num *)edgetab,
(const SCOTCH_Num) edlotab,
(const SCOTCH_Num *)baseval,
(const SCOTCH_Num *)flagval);CHKERRQ(ierr);
#else
SETERRQ(PETSC_ERR_SUP,"Scotch interface currently broken");
#endif
ierr = SCOTCH_graphCheck(&grafptr);CHKERRQ(ierr);
/* Construction of the strategy */
if (scotch->strategy[0] != 0) {
ierr = PetscStrcpy(strategy, scotch->strategy);CHKERRQ(ierr);
} else {
PetscStrcpy(strategy, "b{strat=");
if (scotch->multilevel) {
/* PetscStrcat(strategy,"m{vert=");
sprintf(strategy+strlen(strategy),"%d",scotch->nbvtxcoarsed);
PetscStrcat(strategy,",asc="); */
sprintf(strategy, "b{strat=m{vert=%d,asc=",
scotch->nbvtxcoarsed);
} else
PetscStrcpy(strategy, "b{strat=");
switch (scotch->global_method) {
case MP_SCOTCH_GREEDY:
PetscStrcat(strategy, "h");
break;
case MP_SCOTCH_GPS:
PetscStrcat(strategy, "g");
break;
case MP_SCOTCH_GR_GPS:
PetscStrcat(strategy, "g|h");
}
switch (scotch->local_method) {
case MP_SCOTCH_KERNIGHAN_LIN:
if (scotch->multilevel)
PetscStrcat(strategy, ",low=f}");
else
PetscStrcat(strategy, " f");
break;
case MP_SCOTCH_NONE:
if (scotch->multilevel)
PetscStrcat(strategy, ",asc=x}");
default:
break;
}
PetscStrcat(strategy, " x}");
}
PetscPrintf(((PetscObject)part)->comm, "strategy=[%s]\n", strategy);
ierr = SCOTCH_stratInit(&stratptr);CHKERRQ(ierr);
/*
TODO: Correct this part
Commented because this doesn't exists anymore
ierr = SCOTCH_stratMap(&stratptr, strategy);CHKERRQ(ierr);
*/
/* check for option mapping */
if (!scotch->map) {
/* ********************************************
* *
* TODO: Correct this part *
* *
* Won't work with this tmp SCOTCH_Strat... *
* *
* I just modified it to make scotch compile, *
* to be able to use PaStiX... *
* *
**********************************************/
#if defined (DOES_NOT_COMPILE_DUE_TO_BROKEN_INTERFACE)
SCOTCH_Strat tmp;
ierr = SCOTCH_graphPart((const SCOTCH_Graph *)&grafptr,
(const SCOTCH_Num) &stratptr,
(const SCOTCH_Strat *)&tmp, /* The Argument changed from scotch 3.04 it was part->n, */
(SCOTCH_Num *) parttab);CHKERRQ(ierr);
#else
SETERRQ(PETSC_ERR_SUP,"Scotch interface currently broken");
#endif
ierr = PetscPrintf(PETSC_COMM_SELF, "Partition simple without mapping\n");
} else {
SCOTCH_Graph grafarch;
SCOTCH_Num *listtab;
SCOTCH_Num listnbr = 0;
SCOTCH_Arch archptr; /* file in scotch architecture format */
SCOTCH_Strat archstrat;
int arch_total_size, *parttab_tmp,err;
int cpt;
char buf[256];
FILE *file1, *file2;
char host_buf[256];
/* generate the graph that represents the arch */
file1 = fopen(scotch->arch, "r");
if (!file1) SETERRQ1(PETSC_ERR_FILE_OPEN, "Scotch: unable to open architecture file %s", scotch->arch);
ierr = SCOTCH_graphInit(&grafarch);CHKERRQ(ierr);
ierr = SCOTCH_graphLoad(&grafarch, file1, baseval, 3);CHKERRQ(ierr);
ierr = SCOTCH_graphCheck(&grafarch);CHKERRQ(ierr);
SCOTCH_graphSize(&grafarch, &arch_total_size, &cpt);
err = fclose(file1);
if (err) SETERRQ(PETSC_ERR_SYS,"fclose() failed on file");
printf("total size = %d\n", arch_total_size);
/* generate the list of nodes currently working */
ierr = PetscGetHostName(host_buf, 256);CHKERRQ(ierr);
ierr = PetscStrlen(host_buf, &j);CHKERRQ(ierr);
file2 = fopen(scotch->host_list, "r");
if (!file2) SETERRQ1(PETSC_ERR_FILE_OPEN, "Scotch: unable to open host list file %s", scotch->host_list);
i = -1;
flg = PETSC_FALSE;
while (!feof(file2) && !flg) {
i++;
fgets(buf, 256, file2);
PetscStrncmp(buf, host_buf, j, &flg);
}
err = fclose(file2);
if (err) SETERRQ(PETSC_ERR_SYS,"fclose() failed on file");
if (!flg) SETERRQ1(PETSC_ERR_LIB, "Scotch: unable to find '%s' in host list file", host_buf);
listnbr = size;
ierr = PetscMalloc(sizeof(SCOTCH_Num) * listnbr, &listtab);CHKERRQ(ierr);
ierr = MPI_Allgather(&i, 1, MPI_INT, listtab, 1, MPI_INT, ((PetscObject)part)->comm);CHKERRQ(ierr);
printf("listnbr = %d, listtab = ", listnbr);
for (i = 0; i < listnbr; i++)
printf("%d ", listtab[i]);
printf("\n");
err = fflush(stdout);
if (err) SETERRQ(PETSC_ERR_SYS,"fflush() failed on file");
ierr = SCOTCH_stratInit(&archstrat);CHKERRQ(ierr);
/**************************************************************
* *
* TODO: Correct this part *
* *
* Commented because this doesn't exists anymore *
* *
* ierr = SCOTCH_stratBipart(&archstrat, "fx");CHKERRQ(ierr); *
**************************************************************/
ierr = SCOTCH_archInit(&archptr);CHKERRQ(ierr);
ierr = SCOTCH_archBuild(&archptr, &grafarch, listnbr, listtab,
&archstrat);CHKERRQ(ierr);
ierr = PetscMalloc((mat->rmap->N) * sizeof(int), &parttab_tmp);CHKERRQ(ierr);
/************************************************************************************
* *
* TODO: Correct this part *
* *
* Commented because this doesn't exists anymore *
* *
* ierr = SCOTCH_mapInit(&mappptr, &grafptr, &archptr, parttab_tmp);CHKERRQ(ierr); *
* *
* ierr = SCOTCH_mapCompute(&mappptr, &stratptr);CHKERRQ(ierr); *
* *
* ierr = SCOTCH_mapView(&mappptr, stdout);CHKERRQ(ierr); *
************************************************************************************/
/* now we have to set in the real parttab at the good place */
/* because the ranks order are different than position in */
/* the arch graph */
for (i = 0; i < mat->rmap->N; i++) {
parttab[i] = parttab_tmp[i];
}
ierr = PetscFree(listtab);CHKERRQ(ierr);
SCOTCH_archExit(&archptr);
/*************************************************
* TODO: Correct this part *
* *
* Commented because this doesn't exists anymore *
* SCOTCH_mapExit(&mappptr); *
*************************************************/
SCOTCH_stratExit(&archstrat);
}
/* dump to mesg_log... */
#ifdef PETSC_HAVE_UNISTD_H
err = fflush(stdout);
if (err) SETERRQ(PETSC_ERR_SYS,"fflush() failed on stdout");
count = read(fd_pipe[0], scotch->mesg_log, (SIZE_LOG - 1) * sizeof(char));
if (count < 0)
count = 0;
scotch->mesg_log[count] = 0;
close(1);
dup2(fd_stdout, 1);
close(fd_stdout);
close(fd_pipe[0]);
close(fd_pipe[1]);
#endif
SCOTCH_graphExit(&grafptr);
SCOTCH_stratExit(&stratptr);
}
if (ierr)
SETERRQ(PETSC_ERR_LIB, scotch->mesg_log);
/* Creation of the index set */
ierr = MPI_Comm_rank(((PetscObject)part)->comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(((PetscObject)part)->comm, &size);CHKERRQ(ierr);
nb_locals = mat->rmap->N / size;
locals = parttab + rank * nb_locals;
if (rank < mat->rmap->N % size) {
nb_locals++;
locals += rank;
} else
locals += mat->rmap->N % size;
ierr = ISCreateGeneral(((PetscObject)part)->comm, nb_locals, locals, partitioning);CHKERRQ(ierr);
/* destroying old objects */
ierr = PetscFree(parttab);CHKERRQ(ierr);
if (matSeq != mat) {
ierr = MatDestroy(matSeq);CHKERRQ(ierr);
}
if (matMPI != mat) {
ierr = MatDestroy(matMPI);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
static
void run_scotch( Container &c,
MapContainer &mapping,
const std::size_t cores,
weight_function_t weight_func,
void *weight )
{
#if 0
static_assert( std::is_signed<
std::remove_reference< decltype( c.end() ) >::type >::value,
"Container must have signed types so that -1 may signify no mapping" );
#endif
raftgraph_t raft_graph;
get_graph_info( c,
raft_graph,
weight_func,
nullptr );
SCOTCH_Graph graph;
if( SCOTCH_graphInit( &graph ) != 0 )
{
/** TODO, add RaftLib Exception **/
std::cerr << "Failed to initialize graph!!\n";
exit( EXIT_FAILURE );
}
auto table( raft_graph.getScotchTables() );
if( SCOTCH_graphBuild(
&graph /** graph ptr **/,
0 /** base value **/,
table.num_vertices /** vertex nmbr (zero indexed) **/,
table.vtable /** vertex tab **/,
&table.vtable[ 1 ] /** vendtab **/,
nullptr /** velotab **/,
nullptr /** vlbltab **/,
table.num_edges /** edge number **/,
table.etable /** edge tab **/,
table.eweight /** edlotab **/
) != 0 )
{
/** TODO, add RaftLib Exception **/
std::cerr << "Failed to build graph\n";
exit( EXIT_FAILURE );
}
if( SCOTCH_graphCheck( &graph ) != 0 )
{
/** TODO, add RaftLib Exception **/
std::cerr << "Graph is inconsistent\n";
std::remove_reference< decltype( table ) >::type::print( std::cerr, table );
std::cerr << "\n";
raft_graph.print( std::cerr );
exit( EXIT_FAILURE );
}
/** TODO, we can do much more with this arch file **/
SCOTCH_Arch archdat;
if( SCOTCH_archInit( &archdat ) != 0 )
{
/** TODO, add RaftLib Exception **/
std::cerr << "Architecture initialization failed\n";
exit( EXIT_FAILURE );
}
/** core are equal **/
if( SCOTCH_archCmplt( &archdat, cores /** num cores **/) != 0 )
{
/** TODO, add RaftLib Exception **/
std::cerr << "Failed to create architecture file\n";
exit( EXIT_FAILURE );
}
/** strategy **/
SCOTCH_Strat stradat;
if( SCOTCH_stratInit( &stradat ) != 0 )
{
/** TODO, add RaftLib Exception **/
std::cerr << "Failed to init strategy!!\n";
exit( EXIT_FAILURE );
}
/** build recursive strategy **/
if( SCOTCH_stratGraphClusterBuild(
&stradat,
SCOTCH_STRATSPEED,
cores,
.75,
.01) != 0 )
{
/** TODO, add RaftLib Exception **/
std::cerr << "Failed to map strategy graph!!\n";
exit( EXIT_FAILURE );
}
if( SCOTCH_graphMap(
&graph /** graph ptr **/,
&archdat,
&stradat,
table.partition /** parttab **/
) != 0 )
{
/** TODO, add RaftLib Exception **/
std::cerr << "Failed to map!!\n";
exit( EXIT_FAILURE );
}
/**
* first case is for if we've mapped all vertices,
* second is for when some of the kernels are innactive
* in which case the number of vertices in the
* table will be less than the size of c in which case
* we need to get which vertices (the actual number id
* from the application) are mapped and to where, the
* returned table in mapping must include even the
* vertices that aren't active (indicated by a -1) so
* that the returning loop can be as simple as possible
*/
if( c.size() == table.num_vertices )
{
/** copy mapping **/
for( auto i( 0 ); i < table.num_vertices; i++ )
{
mapping.emplace_back( table.partition[ i ] );
}
}
else
{
const auto &vmapping( raft_graph.getVertexNumbersAtIndicies() );
auto it_map_index( vmapping.cbegin() );
auto table_index( 0 );
const auto size( c.size() );
for( auto i( 0 ); i < size; i++ )
{
if( i == (*it_map_index) && it_map_index != vmapping.cend() )
{
mapping.emplace_back( table.partition[ table_index++ ] );
++it_map_index;
}
else
{
mapping.emplace_back( -1 );
}
}
}
/** call exit graph **/
SCOTCH_graphExit( &graph );
SCOTCH_stratExit( &stradat );
SCOTCH_archExit ( &archdat );
return;
}
/* Function : renumbering
* it modifies the numbering of each node to prevent from cache missing.
*
* - boxVertNbr : number of vertices by box
* - mesh : the input mesh which is modified
*
* returning 0 if OK, 1 else
*/
int renumbering(int boxVertNbr, MMG_pMesh mesh, MMG_pSol sol) {
MMG_pPoint ppt;
MMG_pPoint points;
MMG_pTria ptri, trias;
MMG_pTetra ptet, tetras;
SCOTCH_Num edgeNbr;
SCOTCH_Num *vertTab, *vendTab, *edgeTab, *permVrtTab;
SCOTCH_Graph graf ;
int vertNbr, nodeGlbIdx, triaIdx, tetraIdx, ballTetIdx;
int i, j, k, addrNew, addrOld;
int edgeSiz;
int *vertOldTab, *permNodTab, ntreal, nereal, npreal;
int *adja,iadr;
double *metNew;
/* Computing the number of vertices and a contiguous tabular of vertices */
vertNbr = 0;
vertOldTab = (int *)M_calloc(mesh->ne + 1, sizeof(int), "renumbering");
if (!memset(vertOldTab, 0, sizeof(int)*(mesh->ne+1))) {
perror("memset");
return 1;
}
for(tetraIdx = 1 ; tetraIdx < mesh->ne + 1 ; tetraIdx++) {
/* Testing if the tetra exists */
if (!mesh->tetra[tetraIdx].v[0]) continue;
vertOldTab[tetraIdx] = vertNbr+1;
vertNbr++;
}
/* Allocating memory to compute adjacency lists */
vertTab = (SCOTCH_Num *)M_calloc(vertNbr + 1, sizeof(SCOTCH_Num), "renumbering");
if (!memset(vertTab, ~0, sizeof(SCOTCH_Num)*(vertNbr + 1))) {
perror("memset");
return 1;
}
vendTab = (SCOTCH_Num *)M_calloc(vertNbr + 1, sizeof(SCOTCH_Num), "renumbering");
edgeNbr = 1;
edgeSiz = vertNbr*2;
edgeTab = (SCOTCH_Num *)M_calloc(edgeSiz, sizeof(SCOTCH_Num), "renumbering");
/* Computing the adjacency list for each vertex */
for(tetraIdx = 1 ; tetraIdx < mesh->ne + 1 ; tetraIdx++) {
/* Testing if the tetra exists */
if (!mesh->tetra[tetraIdx].v[0]) continue;
iadr = 4*(tetraIdx-1) + 1;
adja = &mesh->adja[iadr];
for (i=0; i<4; i++) {
ballTetIdx = adja[i] >> 2;
if (!ballTetIdx) continue;
/* Testing if one neighbour of tetraIdx has already been added */
if (vertTab[vertOldTab[tetraIdx]] < 0)
vertTab[vertOldTab[tetraIdx]] = edgeNbr;
vendTab[vertOldTab[tetraIdx]] = edgeNbr+1;
/* Testing if edgeTab memory is enough */
if (edgeNbr >= edgeSiz) {
edgeSiz += EDGEGAP;
edgeTab = (SCOTCH_Num *)M_realloc(edgeTab, edgeSiz * sizeof(SCOTCH_Num), "renumbering");
}
edgeTab[edgeNbr++] = vertOldTab[ballTetIdx];
}
}
edgeNbr--;
/* Building the graph by calling Scotch functions */
SCOTCH_graphInit(&graf) ;
CHECK_SCOTCH(SCOTCH_graphBuild(&graf, (SCOTCH_Num) 1, vertNbr, vertTab+1, vendTab+1, NULL, NULL, edgeNbr, edgeTab+1, NULL), "scotch_graphbuild", 0) ;
CHECK_SCOTCH(SCOTCH_graphCheck(&graf), "scotch_graphcheck", 0) ;
permVrtTab = (SCOTCH_Num *)M_calloc(vertNbr + 1, sizeof(SCOTCH_Num), "renumbering");
CHECK_SCOTCH(kPartBoxCompute(graf, vertNbr, boxVertNbr, permVrtTab), "boxCompute", 0);
SCOTCH_graphExit(&graf) ;
M_free(vertTab);
M_free(vendTab);
M_free(edgeTab);
permNodTab = (int *)M_calloc(mesh->np + 1, sizeof(int), "renumbering");
/* Computing the new point list and modifying the sol structures*/
tetras = (MMG_pTetra)M_calloc(mesh->nemax+1,sizeof(MMG_Tetra),"renumbering");
points = (MMG_pPoint)M_calloc(mesh->npmax+1,sizeof(MMG_Point),"renumbering");
metNew = (double*)M_calloc(sol->npmax+1,sol->offset*sizeof(double),"renumbering");
nereal = 0;
npreal = 1;
for(tetraIdx = 1 ; tetraIdx < mesh->ne + 1 ; tetraIdx++) {
ptet = &mesh->tetra[tetraIdx];
/* Testing if the tetra exists */
if (!ptet->v[0]) continue;
/* Building the new point list */
tetras[permVrtTab[vertOldTab[tetraIdx]]] = *ptet;
nereal++;
for(j = 0 ; j <= 3 ; j++) {
nodeGlbIdx = mesh->tetra[tetraIdx].v[j];
if (permNodTab[nodeGlbIdx]) continue;
ppt = &mesh->point[nodeGlbIdx];
if (!(ppt->tag & M_UNUSED)) {
/* Building the new point list */
permNodTab[nodeGlbIdx] = npreal++;
points[permNodTab[nodeGlbIdx]] = *ppt;
/* Building the new sol met */
addrOld = (nodeGlbIdx-1)*sol->offset + 1;
addrNew = (permNodTab[nodeGlbIdx]-1)*sol->offset + 1;
memcpy(&metNew[addrNew], &sol->met[addrOld], sol->offset*sizeof(double));
}
}
}
M_free(mesh->tetra);
mesh->tetra = tetras;
mesh->ne = nereal;
M_free(mesh->point);
mesh->point = points;
mesh->np = npreal - 1;
M_free(sol->met);
sol->met = metNew;
trias = (MMG_pTria)M_calloc(mesh->ntmax+1,sizeof(MMG_Tria),"renumbering");
ntreal = 1;
for(triaIdx = 1 ; triaIdx < mesh->nt + 1 ; triaIdx++) {
ptri = &mesh->tria[triaIdx];
/* Testing if the tetra exists */
if (!ptri->v[0]) continue;
/* Building the new point list */
trias[ntreal] = *ptri;
ntreal++;
}
M_free(mesh->tria);
mesh->tria = trias;
mesh->nt = ntreal - 1;
mesh->npnil = mesh->np + 1;
mesh->nenil = mesh->ne + 1;
for (k=mesh->npnil; k<mesh->npmax-1; k++)
mesh->point[k].tmp = k+1;
for (k=mesh->nenil; k<mesh->nemax-1; k++)
mesh->tetra[k].v[3] = k+1;
if ( mesh->nt ) {
mesh->ntnil = mesh->nt + 1;
for (k=mesh->ntnil; k<mesh->ntmax-1; k++)
mesh->tria[k].v[2] = k+1;
}
/* Modifying the numbering of the nodes of each tetra */
for(tetraIdx = 1 ; tetraIdx < mesh->ne + 1 ; tetraIdx++) {
if (!mesh->tetra[tetraIdx].v[0]) continue;
for(j = 0 ; j <= 3 ; j++) {
mesh->tetra[tetraIdx].v[j] = permNodTab[mesh->tetra[tetraIdx].v[j]];
}
}
/* Modifying the numbering of the nodes of each triangle */
for(triaIdx = 1 ; triaIdx < mesh->nt + 1 ; triaIdx++) {
if (!mesh->tria[triaIdx].v[0]) continue;
for(j = 0 ; j <= 2 ; j++) {
mesh->tria[triaIdx].v[j] = permNodTab[mesh->tria[triaIdx].v[j]];
}
}
M_free(permVrtTab);
return 1;
}