void inner_prod_impl(vector_base<NumericT> const & x,
vector_tuple<NumericT> const & vec_tuple,
vector_base<NumericT> & result)
{
typedef NumericT value_type;
value_type const * data_x = detail::extract_raw_pointer<value_type>(x);
vcl_size_t start_x = viennacl::traits::start(x);
vcl_size_t inc_x = viennacl::traits::stride(x);
vcl_size_t size_x = viennacl::traits::size(x);
std::vector<value_type> temp(vec_tuple.const_size());
std::vector<value_type const *> data_y(vec_tuple.const_size());
std::vector<vcl_size_t> start_y(vec_tuple.const_size());
std::vector<vcl_size_t> stride_y(vec_tuple.const_size());
for (vcl_size_t j=0; j<vec_tuple.const_size(); ++j)
{
data_y[j] = detail::extract_raw_pointer<value_type>(vec_tuple.const_at(j));
start_y[j] = viennacl::traits::start(vec_tuple.const_at(j));
stride_y[j] = viennacl::traits::stride(vec_tuple.const_at(j));
}
// Note: No OpenMP here because it cannot perform a reduction on temp-array. Savings in memory bandwidth are expected to still justify this approach...
for (vcl_size_t i = 0; i < size_x; ++i)
{
value_type entry_x = data_x[i*inc_x+start_x];
for (vcl_size_t j=0; j < vec_tuple.const_size(); ++j)
temp[j] += entry_x * data_y[j][i*stride_y[j]+start_y[j]];
}
for (vcl_size_t j=0; j < vec_tuple.const_size(); ++j)
result[j] = temp[j]; //Note: Assignment to result might be expensive, thus 'temp' is used for accumulation
}
void inner_prod_impl(vector_base<T> const & x,
vector_tuple<T> const & y_tuple,
vector_base<T> & result)
{
assert( x.size() == y_tuple.const_at(0).size() && bool("Size mismatch") );
assert( result.size() == y_tuple.const_size() && bool("Number of elements does not match result size") );
switch (viennacl::traits::handle(x).get_active_handle_id())
{
case viennacl::MAIN_MEMORY:
viennacl::linalg::host_based::inner_prod_impl(x, y_tuple, result);
break;
#ifdef VIENNACL_WITH_OPENCL
case viennacl::OPENCL_MEMORY:
viennacl::linalg::opencl::inner_prod_impl(x, y_tuple, result);
break;
#endif
#ifdef VIENNACL_WITH_CUDA
case viennacl::CUDA_MEMORY:
viennacl::linalg::cuda::inner_prod_impl(x, y_tuple, result);
break;
#endif
default:
throw "not implemented";
}
}