matrix_map<T, ptrdiff_t>
map(const boost::numeric::ublas::compressed_matrix<T, boost::numeric::ublas::row_major> &A) {
return matrix_map<T, ptrdiff_t>(
A.size1(),
A.size2(),
// amgcl expects signed type for row and col data. This should work
// for decently sized matrices:
reinterpret_cast<const ptrdiff_t*>(A.index1_data().begin()),
reinterpret_cast<const ptrdiff_t*>(A.index2_data().begin()),
A.value_data().begin()
);
}
std::tuple<
size_t,
boost::iterator_range<const size_t*>,
boost::iterator_range<const size_t*>,
boost::iterator_range<const T*>
>
map(const boost::numeric::ublas::compressed_matrix<T, boost::numeric::ublas::row_major> &A) {
return std::make_tuple(
A.size1(),
boost::make_iterator_range(
A.index1_data().begin(), A.index1_data().end()
),
boost::make_iterator_range(
A.index2_data().begin(), A.index2_data().end()
),
boost::make_iterator_range(
A.value_data().begin(), A.value_data().end()
)
);
}
void axpy(
const boost::numeric::ublas::compressed_matrix<real, boost::numeric::ublas::row_major> &A,
const boost::numeric::ublas::vector<real> &x,
boost::numeric::ublas::vector<real> &y
)
{
const ptrdiff_t n = A.size1();
const size_t *Arow = A.index1_data().begin();
const size_t *Acol = A.index2_data().begin();
const real *Aval = A.value_data().begin();
#pragma omp parallel for schedule(dynamic, 1024)
for(ptrdiff_t i = 0; i < n; ++i) {
real buf = 0;
for(size_t j = Arow[i], e = Arow[i + 1]; j < e; ++j)
buf += Aval[j] * x[Acol[j]];
y[i] = buf;
}
}