ParallelReduce( const FunctorType & functor
, const Policy & policy
, const HostViewType & result_view )
: m_func( functor )
, m_policy( policy )
{
QthreadExec::resize_worker_scratch( Reduce::value_size( m_func ) , 0 );
Impl::QthreadExec::exec_all( Qthread::instance() , & ParallelReduce::execute , this );
const pointer_type data = (pointer_type) QthreadExec::exec_all_reduce_result();
Reduce::final( m_func , data );
if ( result_view.ptr_on_device() ) {
const unsigned n = Reduce::value_count( m_func );
for ( unsigned i = 0 ; i < n ; ++i ) { result_view.ptr_on_device()[i] = data[i]; }
}
}
ParallelReduce( const FunctorType & functor ,
const Policy & policy ,
const HostViewType & result_view )
: m_func( functor )
, m_policy( policy )
{
ThreadsExec::resize_reduce_scratch( Reduce::value_size( m_func ) );
ThreadsExec::start( & ParallelReduce::execute , this );
const pointer_type data = (pointer_type) ThreadsExec::root_reduce_scratch();
ThreadsExec::fence();
if ( result_view.ptr_on_device() ) {
const unsigned n = Reduce::value_count( m_func );
for ( unsigned i = 0 ; i < n ; ++i ) { result_view.ptr_on_device()[i] = data[i]; }
}
}
ParallelReduce( const FunctorType & arg_functor ,
const Policy & arg_policy ,
const HostViewType & arg_result_view )
: m_functor( arg_functor )
, m_policy( arg_policy )
, m_result_ptr( arg_result_view.ptr_on_device() )
{
static_assert( Kokkos::is_view< HostViewType >::value
, "Kokkos::Threads reduce result must be a View" );
static_assert( std::is_same< typename HostViewType::memory_space , HostSpace >::value
, "Kokkos::Threads reduce result must be a View in HostSpace" );
}