static result_type
apply(ViewT& v)
{
using namespace vsip_csl::dispatcher;
result_type r;
// Don't use the default dispatch list here, as that doesn't include the
// 'parallel' backend. (The latter then uses the default list for local
// dispatches.
typedef Make_type_list<
be::parallel,
be::cbe_sdk,
be::cuda,
be::cvsip,
be::mercury_sal,
be::generic>::type list_type;
if (is_expr_dense(v.block()))
{
Dispatcher<op::reduce<ReduceT>,
void(result_type&, new_block_type const&, new_order_type, new_dim_type), list_type>::
dispatch(r, new_block_type(const_cast<block_type&>(v.block())), new_order_type(), new_dim_type());
}
else
{
Dispatcher<op::reduce<ReduceT>,
void(result_type&, block_type const&, order_type, dim_type), list_type>::
dispatch(r, v.block(), order_type(), dim_type());
}
return r;
}
void
check_block_size(length_type g_size, length_type l_size, ViewT const& view)
{
test_assert(g_size == view.size(0));
test_assert(g_size == view.block().size(1, 0));
test_assert(l_size == view.local().size(0));
test_assert(l_size == view.local().block().size(1, 0));
test_assert(g_size == view.size());
test_assert(g_size == view.block().size());
test_assert(l_size == view.local().size());
test_assert(l_size == view.local().block().size());
}
typename ReduceT<typename ViewT::value_type>::result_type
reduce(ViewT v)
{
using namespace vsip_csl::dispatcher;
typedef typename ViewT::value_type T;
typedef typename ReduceT<T>::result_type result_type;
typedef typename ViewT::block_type block_type;
typedef typename get_block_layout<block_type>::order_type order_type;
typedef Int_type<ViewT::dim> dim_type;
result_type r;
#if VSIP_IMPL_REF_IMPL
Evaluator<op::reduce<ReduceT>, be::cvsip,
void(result_type&, block_type const&, order_type, dim_type)>::
exec(r, v.block(), order_type(), dim_type());
#else
// This optimization is only applicable to target platforms that provide a
// backend that uses direct data access rather than redim_get/put().
// This includes CUDA and Cell backends, but not x86 (as of August 2010).
#if defined(VSIP_IMPL_CBE_SDK) || defined(VSIP_IMPL_HAVE_CUDA)
bool const redimensionable =
(ViewT::dim != 1) && (!is_expr_block<block_type>::value);
#else
bool const redimensionable = false;
#endif
r = Reduction_dispatch_helper<ReduceT, ViewT, redimensionable>::apply(v);
#endif
return r;
}
Ps_helper(ViewT& view, Stream_pattern const& spatt, bool input)
: data_ (view.block())
, addr_ (data_.ptr())
, spatt_ (spatt)
, input_ (input)
{}