inline void inner_panel_mixed_gemm_impl_nn(
const double alpha,
const SpParMat<index_type, value_type, SpDCCols<index_type, value_type> > &A,
const El::DistMatrix<value_type, El::STAR, El::STAR> &S,
const double beta,
El::DistMatrix<value_type, col_d, El::STAR> &C) {
int n_proc_side = A.getcommgrid()->GetGridRows();
int output_width = S.Width();
int output_height = A.getnrow();
size_t rank = A.getcommgrid()->GetRank();
size_t cb_row_offset = utility::cb_my_row_offset(A);
typedef SpDCCols< index_type, value_type > col_t;
typedef SpParMat< index_type, value_type, col_t > matrix_type;
matrix_type &_A = const_cast<matrix_type&>(A);
col_t &data = _A.seq();
// 1) compute the local values still using the CombBLAS distribution (2D
// processor grid). We assume the result is dense.
std::vector<double> local_matrix;
mixed_gemm_local_part_nn(alpha, A, S, 0.0, local_matrix);
// 2) reduce first along rows so that each processor owns the values in
// the output row of the SOMETHING/* matrix and values for processors in
// the same processor column.
boost::mpi::communicator my_row_comm(
A.getcommgrid()->GetRowWorld(), boost::mpi::comm_duplicate);
// storage for other procs in same row communicator: rank -> (row, values)
typedef std::vector<std::pair<int, std::vector<double> > > for_rank_t;
std::vector<for_rank_t> for_rank(n_proc_side);
for(size_t local_row = 0; local_row < data.getnrow(); ++local_row) {
size_t row = local_row + cb_row_offset;
// the owner for VR/* and VC/* matrices is independent of the column
size_t target_proc = utility::owner(C, row, static_cast<size_t>(0));
// if the target processor is not in the current row communicator, get
// the value in the processor grid sharing the same row.
if(!A.getcommgrid()->OnSameProcRow(target_proc))
target_proc = static_cast<int>(rank / n_proc_side) *
n_proc_side + target_proc % n_proc_side;
size_t target_row_rank = A.getcommgrid()->GetRankInProcRow(target_proc);
// reduce partial row (FIXME: if the resulting matrix is still
// expected to be sparse, change this to communicate only nnz).
// Working on local_width columns concurrently per column processing
// group.
size_t local_width = S.Width();
const value_type* buffer = &local_matrix[local_row * local_width];
std::vector<value_type> new_values(local_width);
boost::mpi::reduce(my_row_comm, buffer, local_width,
&new_values[0], std::plus<value_type>(), target_row_rank);
// processor stores result directly if it is the owning rank of that
// row, save for subsequent communication along rows otherwise
if(rank == utility::owner(C, row, static_cast<size_t>(0))) {
int elem_lrow = C.LocalRow(row);
for(size_t idx = 0; idx < local_width; ++idx) {
int elem_lcol = C.LocalCol(idx);
C.SetLocal(elem_lrow, elem_lcol,
new_values[idx] + beta * C.GetLocal(elem_lrow, elem_lcol));
}
} else if (rank == target_proc) {
// store for later comm across rows
for_rank[utility::owner(C, row, static_cast<size_t>(0)) / n_proc_side].push_back(
std::make_pair(row, new_values));
}
}
// 3) gather remaining values along rows: we exchange all the values with
// other processors in the same communicator row and then add them to
// our local part.
boost::mpi::communicator my_col_comm(
A.getcommgrid()->GetColWorld(), boost::mpi::comm_duplicate);
std::vector<for_rank_t> new_values;
for(int i = 0; i < n_proc_side; ++i)
boost::mpi::gather(my_col_comm, for_rank[i], new_values, i);
// insert new values
for(size_t proc = 0; proc < new_values.size(); ++proc) {
const for_rank_t &cur = new_values[proc];
for(size_t i = 0; i < cur.size(); ++i) {
int elem_lrow = C.LocalRow(cur[i].first);
for(size_t j = 0; j < cur[i].second.size(); ++j) {
size_t elem_lcol = C.LocalCol(j);
C.SetLocal(elem_lrow, elem_lcol,
cur[i].second[j] + beta *
C.GetLocal(elem_lrow, elem_lcol));
}
}
}
}
IT Height(const SpParMat<IT, VT, SpDCCols<IT, VT> >& A) {
return A.getnrow();
}
