int
SCOTCH_dgraphPart (
SCOTCH_Dgraph * const grafptr, /*+ Graph to map +*/
const SCOTCH_Num partnbr, /*+ Number of parts +*/
SCOTCH_Strat * const stratptr, /*+ Mapping strategy +*/
SCOTCH_Num * const termloctab) /*+ Mapping array +*/
{
SCOTCH_Arch archdat;
int o;
SCOTCH_archInit (&archdat);
SCOTCH_archCmplt (&archdat, partnbr);
o = SCOTCH_dgraphMap (grafptr, &archdat, stratptr, termloctab);
SCOTCH_archExit (&archdat);
return (o);
}
void AlgPTScotch<Adapter>::partition(
const RCP<PartitioningSolution<Adapter> > &solution
)
{
HELLO;
size_t numGlobalParts = solution->getTargetGlobalNumberOfParts();
SCOTCH_Num partnbr=0;
TPL_Traits<SCOTCH_Num, size_t>::ASSIGN_TPL_T(partnbr, numGlobalParts, env);
#ifdef HAVE_ZOLTAN2_MPI
int ierr = 0;
int me = problemComm->getRank();
const SCOTCH_Num baseval = 0; // Base value for array indexing.
// GraphModel returns GNOs from base 0.
SCOTCH_Strat stratstr; // Strategy string
// TODO: Set from parameters
SCOTCH_stratInit(&stratstr);
// Allocate and initialize PTScotch Graph data structure.
SCOTCH_Dgraph *gr = SCOTCH_dgraphAlloc(); // Scotch distributed graph
ierr = SCOTCH_dgraphInit(gr, mpicomm);
env->globalInputAssertion(__FILE__, __LINE__, "SCOTCH_dgraphInit",
!ierr, BASIC_ASSERTION, problemComm);
// Get vertex info
ArrayView<const gno_t> vtxID;
ArrayView<StridedData<lno_t, scalar_t> > xyz;
ArrayView<StridedData<lno_t, scalar_t> > vwgts;
size_t nVtx = model->getVertexList(vtxID, xyz, vwgts);
SCOTCH_Num vertlocnbr=0;
TPL_Traits<SCOTCH_Num, size_t>::ASSIGN_TPL_T(vertlocnbr, nVtx, env);
SCOTCH_Num vertlocmax = vertlocnbr; // Assumes no holes in global nums.
// Get edge info
ArrayView<const gno_t> edgeIds;
ArrayView<const int> procIds;
ArrayView<const lno_t> offsets;
ArrayView<StridedData<lno_t, scalar_t> > ewgts;
size_t nEdge = model->getEdgeList(edgeIds, procIds, offsets, ewgts);
SCOTCH_Num edgelocnbr=0;
TPL_Traits<SCOTCH_Num, size_t>::ASSIGN_TPL_T(edgelocnbr, nEdge, env);
const SCOTCH_Num edgelocsize = edgelocnbr; // Assumes adj array is compact.
SCOTCH_Num *vertloctab; // starting adj/vtx
TPL_Traits<SCOTCH_Num, lno_t>::ASSIGN_TPL_T_ARRAY(&vertloctab, offsets, env);
SCOTCH_Num *edgeloctab; // adjacencies
TPL_Traits<SCOTCH_Num, gno_t>::ASSIGN_TPL_T_ARRAY(&edgeloctab, edgeIds, env);
// We don't use these arrays, but we need them as arguments to Scotch.
SCOTCH_Num *vendloctab = NULL; // Assume consecutive storage for adj
SCOTCH_Num *vlblloctab = NULL; // Vertex label array
SCOTCH_Num *edgegsttab = NULL; // Array for ghost vertices
// Get weight info.
SCOTCH_Num *velotab = NULL; // Vertex weights
SCOTCH_Num *edlotab = NULL; // Edge weights
int nVwgts = model->getNumWeightsPerVertex();
int nEwgts = model->getNumWeightsPerEdge();
if (nVwgts > 1 && me == 0) {
std::cerr << "Warning: NumWeightsPerVertex is " << nVwgts
<< " but Scotch allows only one weight. "
<< " Zoltan2 will use only the first weight per vertex."
<< std::endl;
}
if (nEwgts > 1 && me == 0) {
std::cerr << "Warning: NumWeightsPerEdge is " << nEwgts
<< " but Scotch allows only one weight. "
<< " Zoltan2 will use only the first weight per edge."
<< std::endl;
}
if (nVwgts) {
velotab = new SCOTCH_Num[nVtx+1]; // +1 since Scotch wants all procs
// to have non-NULL arrays
scale_weights(nVtx, vwgts[0], velotab);
}
if (nEwgts) {
edlotab = new SCOTCH_Num[nEdge+1]; // +1 since Scotch wants all procs
// to have non-NULL arrays
scale_weights(nEdge, ewgts[0], edlotab);
}
// Build PTScotch distributed data structure
ierr = SCOTCH_dgraphBuild(gr, baseval, vertlocnbr, vertlocmax,
vertloctab, vendloctab, velotab, vlblloctab,
edgelocnbr, edgelocsize,
edgeloctab, edgegsttab, edlotab);
env->globalInputAssertion(__FILE__, __LINE__, "SCOTCH_dgraphBuild",
!ierr, BASIC_ASSERTION, problemComm);
// Create array for Scotch to return results in.
ArrayRCP<part_t> partList(new part_t[nVtx], 0, nVtx,true);
SCOTCH_Num *partloctab = NULL;
if (nVtx && (sizeof(SCOTCH_Num) == sizeof(part_t))) {
// Can write directly into the solution's memory
partloctab = (SCOTCH_Num *) partList.getRawPtr();
}
else {
// Can't use solution memory directly; will have to copy later.
// Note: Scotch does not like NULL arrays, so add 1 to always have non-null.
// ParMETIS has this same "feature." See Zoltan bug 4299.
partloctab = new SCOTCH_Num[nVtx+1];
}
// Get target part sizes
float *partsizes = new float[numGlobalParts];
if (!solution->criteriaHasUniformPartSizes(0))
for (size_t i=0; i<numGlobalParts; i++)
partsizes[i] = solution->getCriteriaPartSize(0, i);
else
for (size_t i=0; i<numGlobalParts; i++)
partsizes[i] = 1.0 / float(numGlobalParts);
// Allocate and initialize PTScotch target architecture data structure
SCOTCH_Arch archdat;
SCOTCH_archInit(&archdat);
SCOTCH_Num velosum = 0;
SCOTCH_dgraphSize (gr, &velosum, NULL, NULL, NULL);
SCOTCH_Num *goalsizes = new SCOTCH_Num[partnbr];
// TODO: The goalsizes are set as in Zoltan; not sure it is correct there
// or here.
// It appears velosum is global NUMBER of vertices, not global total
// vertex weight. I think we should use the latter.
// Fix this when we add vertex weights.
for (SCOTCH_Num i = 0; i < partnbr; i++)
goalsizes[i] = SCOTCH_Num(ceil(velosum * partsizes[i]));
delete [] partsizes;
SCOTCH_archCmpltw(&archdat, partnbr, goalsizes);
// Call partitioning; result returned in partloctab.
ierr = SCOTCH_dgraphMap(gr, &archdat, &stratstr, partloctab);
env->globalInputAssertion(__FILE__, __LINE__, "SCOTCH_dgraphMap",
!ierr, BASIC_ASSERTION, problemComm);
SCOTCH_archExit(&archdat);
delete [] goalsizes;
// TODO - metrics
#ifdef SHOW_ZOLTAN2_SCOTCH_MEMORY
int me = env->comm_->getRank();
#endif
#ifdef HAVE_SCOTCH_ZOLTAN2_GETMEMORYMAX
if (me == 0){
size_t scotchBytes = SCOTCH_getMemoryMax();
std::cout << "Rank " << me << ": Maximum bytes used by Scotch: ";
std::cout << scotchBytes << std::endl;
}
#endif
// Clean up PTScotch
SCOTCH_dgraphExit(gr);
free(gr);
SCOTCH_stratExit(&stratstr);
// Load answer into the solution.
if ((sizeof(SCOTCH_Num) != sizeof(part_t)) || (nVtx == 0)) {
for (size_t i = 0; i < nVtx; i++) partList[i] = partloctab[i];
delete [] partloctab;
}
solution->setParts(partList);
env->memory("Zoltan2-Scotch: After creating solution");
// Clean up copies made due to differing data sizes.
TPL_Traits<SCOTCH_Num, lno_t>::DELETE_TPL_T_ARRAY(&vertloctab);
TPL_Traits<SCOTCH_Num, gno_t>::DELETE_TPL_T_ARRAY(&edgeloctab);
if (nVwgts) delete [] velotab;
if (nEwgts) delete [] edlotab;
#else // DO NOT HAVE_MPI
// TODO: Handle serial case with calls to Scotch.
// TODO: For now, assign everything to rank 0 and assume only one part.
// TODO: Can probably use the code above for loading solution,
// TODO: instead of duplicating it here.
// TODO
// TODO: Actual logic should call Scotch when number of processes == 1.
ArrayView<const gno_t> vtxID;
ArrayView<StridedData<lno_t, scalar_t> > xyz;
ArrayView<StridedData<lno_t, scalar_t> > vwgts;
size_t nVtx = model->getVertexList(vtxID, xyz, vwgts);
ArrayRCP<part_t> partList(new part_t[nVtx], 0, nVtx, true);
for (size_t i = 0; i < nVtx; i++) partList[i] = 0;
solution->setParts(partList);
#endif // DO NOT HAVE_MPI
}
// Call scotch with options from dictionary.
Foam::label Foam::ptscotchDecomp::decompose
(
const fileName& meshPath,
const List<int>& adjncy,
const List<int>& xadj,
const scalarField& cWeights,
List<int>& finalDecomp
) const
{
if (debug)
{
Pout<< "ptscotchDecomp : entering with xadj:" << xadj.size() << endl;
}
// Dump graph
if (decompositionDict_.found("ptscotchCoeffs"))
{
const dictionary& scotchCoeffs =
decompositionDict_.subDict("ptscotchCoeffs");
if (scotchCoeffs.lookupOrDefault("writeGraph", false))
{
OFstream str
(
meshPath + "_" + Foam::name(Pstream::myProcNo()) + ".dgr"
);
Pout<< "Dumping Scotch graph file to " << str.name() << endl
<< "Use this in combination with dgpart." << endl;
globalIndex globalCells(xadj.size()-1);
// Distributed graph file (.grf)
label version = 2;
str << version << nl;
// Number of files (procglbnbr)
str << Pstream::nProcs();
// My file number (procloc)
str << ' ' << Pstream::myProcNo() << nl;
// Total number of vertices (vertglbnbr)
str << globalCells.size();
// Total number of connections (edgeglbnbr)
str << ' ' << returnReduce(xadj[xadj.size()-1], sumOp<label>())
<< nl;
// Local number of vertices (vertlocnbr)
str << xadj.size()-1;
// Local number of connections (edgelocnbr)
str << ' ' << xadj[xadj.size()-1] << nl;
// Numbering starts from 0
label baseval = 0;
// 100*hasVertlabels+10*hasEdgeWeights+1*hasVertWeighs
str << baseval << ' ' << "000" << 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<< "ptscotchDecomp : Using strategy " << strategy << endl;
}
SCOTCH_stratDgraphMap(&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
(
"ptscotchDecomp::decompose(..)"
) << "Illegal minimum weight " << minWeights
<< endl;
}
if (cWeights.size() != xadj.size()-1)
{
FatalErrorIn
(
"ptscotchDecomp::decompose(..)"
) << "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);
}
}
if (debug)
{
Pout<< "SCOTCH_dgraphInit" << endl;
}
SCOTCH_Dgraph grafdat;
check(SCOTCH_dgraphInit(&grafdat, MPI_COMM_WORLD), "SCOTCH_dgraphInit");
if (debug)
{
Pout<< "SCOTCH_dgraphBuild with:" << nl
<< "xadj.size()-1 : " << xadj.size()-1 << nl
<< "xadj : " << long(xadj.begin()) << nl
<< "velotab : " << long(velotab.begin()) << nl
<< "adjncy.size() : " << adjncy.size() << nl
<< "adjncy : " << long(adjncy.begin()) << nl
<< endl;
}
check
(
SCOTCH_dgraphBuild
(
&grafdat, // grafdat
0, // baseval, c-style numbering
xadj.size()-1, // vertlocnbr, nCells
xadj.size()-1, // vertlocmax
const_cast<SCOTCH_Num*>(xadj.begin()),
// vertloctab, start index per cell into
// adjncy
const_cast<SCOTCH_Num*>(&xadj[1]),// vendloctab, end index ,,
const_cast<SCOTCH_Num*>(velotab.begin()),// veloloctab, vtx weights
NULL, // vlblloctab
adjncy.size(), // edgelocnbr, number of arcs
adjncy.size(), // edgelocsiz
const_cast<SCOTCH_Num*>(adjncy.begin()), // edgeloctab
NULL, // edgegsttab
NULL // edlotab, edge weights
),
"SCOTCH_dgraphBuild"
);
if (debug)
{
Pout<< "SCOTCH_dgraphCheck" << endl;
}
check(SCOTCH_dgraphCheck(&grafdat), "SCOTCH_dgraphCheck");
// Architecture
// ~~~~~~~~~~~~
// (fully connected network topology since using switch)
if (debug)
{
Pout<< "SCOTCH_archInit" << endl;
}
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<< "ptscotchDecomp : Using procesor weights "
<< processorWeights
<< endl;
}
check
(
SCOTCH_archCmpltw(&archdat, nProcessors_, processorWeights.begin()),
"SCOTCH_archCmpltw"
);
}
else
{
if (debug)
{
Pout<< "SCOTCH_archCmplt" << endl;
}
check
(
SCOTCH_archCmplt(&archdat, nProcessors_),
"SCOTCH_archCmplt"
);
}
//SCOTCH_Mapping mapdat;
//SCOTCH_dgraphMapInit(&grafdat, &mapdat, &archdat, NULL);
//SCOTCH_dgraphMapCompute(&grafdat, &mapdat, &stradat); /*Perform mapping*/
//SCOTCHdgraphMapExit(&grafdat, &mapdat);
// Hack:switch off fpu error trapping
# ifdef LINUX_GNUC
int oldExcepts = fedisableexcept
(
FE_DIVBYZERO
| FE_INVALID
| FE_OVERFLOW
);
# endif
if (debug)
{
Pout<< "SCOTCH_dgraphMap" << endl;
}
finalDecomp.setSize(xadj.size()-1);
finalDecomp = 0;
check
(
SCOTCH_dgraphMap
(
&grafdat,
&archdat,
&stradat, // const SCOTCH_Strat *
finalDecomp.begin() // parttab
),
"SCOTCH_graphMap"
);
# ifdef LINUX_GNUC
feenableexcept(oldExcepts);
# endif
//finalDecomp.setSize(xadj.size()-1);
//check
//(
// SCOTCH_dgraphPart
// (
// &grafdat,
// nProcessors_, // partnbr
// &stradat, // const SCOTCH_Strat *
// finalDecomp.begin() // parttab
// ),
// "SCOTCH_graphPart"
//);
if (debug)
{
Pout<< "SCOTCH_dgraphExit" << endl;
}
// Release storage for graph
SCOTCH_dgraphExit(&grafdat);
// Release storage for strategy
SCOTCH_stratExit(&stradat);
// Release storage for network topology
SCOTCH_archExit(&archdat);
return 0;
}
