ParallelReduce( const FunctorType & arg_functor
, const Policy & arg_policy
, const ViewType & arg_result )
: m_functor( arg_functor )
, m_policy( arg_policy )
, m_result_ptr( arg_result.ptr_on_device() )
, m_shared( arg_policy.scratch_size() + FunctorTeamShmemSize< FunctorType >::value( arg_functor , arg_policy.team_size() ) )
{ }
Teuchos::ArrayRCP<typename ViewType::value_type>
persistingView(
const ViewType& view,
typename Teuchos::ArrayRCP<typename ViewType::value_type>::size_type offset,
typename Teuchos::ArrayRCP<typename ViewType::value_type>::size_type size) {
return Teuchos::arcp(view.ptr_on_device()+offset, 0, size,
deallocator(view), false);
}
ParallelReduce( const FunctorType & functor
, const Policy & policy
, const ViewType & result )
: m_func( functor )
, m_policy( policy )
{
ThreadsExec::resize_shared_scratch( FunctorShmemSize< FunctorType >::value( functor ) );
ThreadsExec::resize_reduce_scratch( Reduce::value_size( m_func ) );
ThreadsExec::start( & ParallelReduce::execute , this , policy.league_size() , policy.team_size() );
const pointer_type data = (pointer_type) ThreadsExec::root_reduce_scratch();
ThreadsExec::fence();
const unsigned n = Reduce::value_count( m_func );
for ( unsigned i = 0 ; i < n ; ++i ) { result.ptr_on_device()[i] = data[i]; }
}
Teuchos::ArrayRCP<typename ViewType::value_type>
persistingView(const ViewType& view) {
// mfh (05 Sep 2013) We use a custom deallocator that keeps the
// view, but otherwise does nothing. It doesn't deallocate the
// view; it just maintains the reference count for the lifetime
// of the deallocator object (and therefore, for the lifetime of
// the ArrayRCP). It seems weird that this is a nonowning
// ArrayRCP (the last argument is false). However, that's OK,
// because (a) the ArrayRCP (actually, the underlying
// RCPNodeTmpl; see Teuchos_RCPNode.hpp) keeps the deallocator
// object regardless of whether the ArrayRCP is owning, and (b)
// if we make this an owning ArrayRCP, the RCPNode tracking
// system (in a debug build: Teuchos_ENABLE_DEBUG:BOOL=ON) will
// report problems, because multiple owning ArrayRCPs have the
// same raw pointer but not the same underlying RCPNode. That
// would be a problem if those ArrayRCPs shared memory that was
// allocated (e.g.,) using new or malloc, but it's not a problem
// here, because the custom deallocator does not free anything.
// Nevertheless, it's better not to trouble the tracking system.
return Teuchos::arcp(view.ptr_on_device(), 0, view.capacity(),
deallocator(view), false);
}
Teuchos::ArrayView<const typename ViewType::value_type>
getConstArrayView(const ViewType& a) {
return Teuchos::ArrayView<const typename ViewType::value_type>(
a.ptr_on_device(), a.size());
}
