void L1DistanceMatrixTU(El::UpperOrLower uplo,
direction_t dirA, direction_t dirB, T alpha,
const El::DistMatrix<T, El::STAR, El::MC> &A,
const El::DistMatrix<T, El::STAR, El::MR> &B,
T beta, El::DistMatrix<T> &C) {
// TODO verify sizes
const T *a = A.LockedBuffer();
El::Int ldA = A.LDim();
const T *b = B.LockedBuffer();
El::Int ldB = B.LDim();
T *c = C.Buffer();
El::Int ldC = C.LDim();
El::Int d = A.Height();
/* Not the most efficient way... but mimicking BLAS is too much work! */
if (dirA == base::COLUMNS && dirB == base::COLUMNS) {
El::Int n = C.LocalWidth();
El::Int m = C.LocalHeight();
for (El::Int j = 0; j < n; j++)
for(El::Int i =
((uplo == El::UPPER) ? 0 : C.LocalRowOffset(A.GlobalCol(j)));
i < ((uplo == El::UPPER) ? C.LocalRowOffset(A.GlobalCol(j) + 1) : m); i++) {
T v = 0.0;
for (El::Int k = 0; k < d; k++)
v += std::abs(b[j * ldB + k] - a[i * ldA + k]);
c[j * ldC + i] = beta * c[j * ldC + i] + alpha * v;
}
}
// TODO the rest of the cases.
}
inline void outer_panel_mixed_gemm_impl_tn(
const double alpha,
const SpParMat<index_type, value_type, SpDCCols<index_type, value_type> > &A,
const El::DistMatrix<value_type, col_d, El::STAR> &S,
const double beta,
El::DistMatrix<value_type, El::STAR, El::STAR> &C) {
El::DistMatrix<value_type, El::STAR, El::STAR>
tmp_C(C.Height(), C.Width());
El::Zero(tmp_C);
utility::combblas_slab_view_t<index_type, value_type> cbview(A, false);
//FIXME: factor
size_t slab_size = 2 * S.Grid().Height();
for(size_t cur_row_idx = 0; cur_row_idx < cbview.ncols();
cur_row_idx += slab_size) {
size_t cur_slab_size =
std::min(slab_size, cbview.ncols() - cur_row_idx);
// get the next slab_size columns of B
El::DistMatrix<value_type, El::STAR, El::STAR>
A_row(cur_slab_size, S.Height());
// transpose is column
//cbview.extract_elemental_column_slab_view(A_row, cur_slab_size);
cbview.extract_full_slab_view(cur_slab_size);
// matrix mult (FIXME only iter nz)
for(size_t l_row_idx = 0; l_row_idx < A_row.LocalHeight();
++l_row_idx) {
size_t g_row_idx = l_row_idx * A_row.ColStride()
+ A_row.ColShift() + cur_row_idx;
for(size_t l_col_idx = 0; l_col_idx < A_row.LocalWidth();
l_col_idx++) {
//XXX: should be the same as l_col_idx
size_t g_col_idx = l_col_idx * A_row.RowStride()
+ A_row.RowShift();
// continue if we don't own values in S in this row
if(!S.IsLocalRow(g_col_idx))
continue;
//get transposed value
value_type val = alpha * cbview(g_col_idx, g_row_idx);
for(size_t s_col_idx = 0; s_col_idx < S.LocalWidth();
s_col_idx++) {
tmp_C.UpdateLocal(g_row_idx, s_col_idx,
val * S.GetLocal(S.LocalRow(g_col_idx), s_col_idx));
}
}
}
}
//FIXME: scaling
if(A.getcommgrid()->GetRank() == 0) {
for(size_t col_idx = 0; col_idx < C.Width(); col_idx++)
for(size_t row_idx = 0; row_idx < C.Height(); row_idx++)
tmp_C.UpdateLocal(row_idx, col_idx,
beta * C.GetLocal(row_idx, col_idx));
}
//FIXME: Use utility getter
boost::mpi::communicator world(
A.getcommgrid()->GetWorld(), boost::mpi::comm_duplicate);
boost::mpi::all_reduce (world,
tmp_C.LockedBuffer(),
C.Height() * C.Width(),
C.Buffer(),
std::plus<value_type>());
}