GraphHelper_Scotch(const CrsMatrixType& A,
const int seed = GraphHelper::DefaultRandomSeed) {
_label = "GraphHelper_Scotch::" + A.Label();
_is_ordered = false;
_cblk = 0;
// scotch does not allow self-contribution (diagonal term in sparse matrix)
_base = 0; //A.BaseVal();
_m = A.NumRows();
_nnz = A.NumNonZeros();
_rptr = size_type_array(_label+"::RowPtrArray", _m+1);
_cidx = ordinal_type_array(_label+"::ColIndexArray", _nnz);
_perm = ordinal_type_array(_label+"::PermutationArray", _m);
_peri = ordinal_type_array(_label+"::InvPermutationArray", _m);
_range = ordinal_type_array(_label+"::RangeArray", _m);
_tree = ordinal_type_array(_label+"::TreeArray", _m);
// create a graph structure without diagonals
A.convertGraph(_nnz, _rptr, _cidx);
int ierr = 0;
ordinal_type *rptr = reinterpret_cast<ordinal_type*>(_rptr.ptr_on_device());
ordinal_type *cidx = reinterpret_cast<ordinal_type*>(_cidx.ptr_on_device());
if (seed != GraphHelper::DefaultRandomSeed) {
SCOTCH_randomSeed(seed);
SCOTCH_randomReset();
}
ierr = SCOTCH_graphInit(&_graph);
CHKERR(ierr);
ierr = SCOTCH_graphBuild(&_graph, // scotch graph
_base, // base value
_m, // # of vertices
rptr, // column index array pointer begin
rptr+1, // column index array pointer end
NULL, // weights on vertices (optional)
NULL, // label array on vertices (optional)
_nnz, // # of nonzeros
cidx, // column index array
NULL);
CHKERR(ierr); // edge load array (optional)
ierr = SCOTCH_graphCheck(&_graph);
CHKERR(ierr);
}
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 computeOrdering() {
int r_val = 0;
camd_defaults(_control);
camd_control(_control);
ordinal_type *rptr = reinterpret_cast<ordinal_type*>(_rptr.ptr_on_device());
ordinal_type *cidx = reinterpret_cast<ordinal_type*>(_cidx.ptr_on_device());
ordinal_type *cnst = reinterpret_cast<ordinal_type*>(_cnst.ptr_on_device());
ordinal_type *next = reinterpret_cast<ordinal_type*>(_next.ptr_on_device());
ordinal_type *perm = reinterpret_cast<ordinal_type*>(_perm.ptr_on_device());
// length array
// assume there always is diagonal and the given matrix is symmetry
ordinal_type_array lwork(_label+"::LWorkArray", _m);
ordinal_type *lwork_ptr = reinterpret_cast<ordinal_type*>(lwork.ptr_on_device());
for (ordinal_type i=0;i<_m;++i)
lwork_ptr[i] = rptr[i+1] - rptr[i];
// workspace
const size_type swlen = _nnz + _nnz/5 + 5*(_m+1);;
ordinal_type_array swork(_label+"::SWorkArray", swlen);
ordinal_type *swork_ptr = reinterpret_cast<ordinal_type*>(swork.ptr_on_device());
ordinal_type *pe_ptr = reinterpret_cast<ordinal_type*>(_pe.ptr_on_device()); // 1) Pe
size_type pfree = 0;
for (ordinal_type i=0;i<_m;++i) {
pe_ptr[i] = pfree;
pfree += lwork_ptr[i];
}
if (_nnz != pfree)
ERROR(">> nnz in the graph does not match to nnz count (pfree)");
ordinal_type *nv_ptr = reinterpret_cast<ordinal_type*>(_nv.ptr_on_device()); // 2) Nv
ordinal_type *hd_ptr = swork_ptr; swork_ptr += (_m+1); // 3) Head
ordinal_type *el_ptr = reinterpret_cast<ordinal_type*>(_el.ptr_on_device()); // 4) Elen
ordinal_type *dg_ptr = swork_ptr; swork_ptr += _m; // 5) Degree
ordinal_type *wk_ptr = swork_ptr; swork_ptr += (_m+1); // 6) W
ordinal_type *bk_ptr = swork_ptr; swork_ptr += _m; // 7) BucketSet
const size_type iwlen = swlen - (4*_m+2);
ordinal_type *iw_ptr = swork_ptr; swork_ptr += iwlen; // Iw
for (ordinal_type i=0;i<pfree;++i)
iw_ptr[i] = cidx[i];
CAMD::run(_m, pe_ptr, iw_ptr, lwork_ptr, iwlen, pfree,
// output
nv_ptr, next, perm, hd_ptr, el_ptr, dg_ptr, wk_ptr,
_control, _info, cnst, bk_ptr);
r_val = (_info[CAMD_STATUS] != CAMD_OK);
for (ordinal_type i=0;i<_m;++i)
_peri[_perm[i]] = i;
_is_ordered = true;
return r_val;
}
