SharedAllocationRecord< Kokkos::HostSpace , void >::
SharedAllocationRecord( const Kokkos::HostSpace & arg_space
, const std::string & arg_label
, const size_t arg_alloc_size
, const SharedAllocationRecord< void , void >::function_type arg_dealloc
)
// Pass through allocated [ SharedAllocationHeader , user_memory ]
// Pass through deallocation function
: SharedAllocationRecord< void , void >
(
#ifdef KOKKOS_DEBUG
& SharedAllocationRecord< Kokkos::HostSpace , void >::s_root_record,
#endif
reinterpret_cast<SharedAllocationHeader*>( arg_space.allocate( sizeof(SharedAllocationHeader) + arg_alloc_size ) )
, sizeof(SharedAllocationHeader) + arg_alloc_size
, arg_dealloc
)
, m_space( arg_space )
{
#if defined(KOKKOS_ENABLE_PROFILING)
if(Kokkos::Profiling::profileLibraryLoaded()) {
Kokkos::Profiling::allocateData(Kokkos::Profiling::SpaceHandle(arg_space.name()),arg_label,data(),arg_alloc_size);
}
#endif
// Fill in the Header information
RecordBase::m_alloc_ptr->m_record = static_cast< SharedAllocationRecord< void , void > * >( this );
strncpy( RecordBase::m_alloc_ptr->m_label
, arg_label.c_str()
, SharedAllocationHeader::maximum_label_length
);
}
// Resize thread team data scratch memory
void serial_resize_thread_team_data( size_t pool_reduce_bytes
, size_t team_reduce_bytes
, size_t team_shared_bytes
, size_t thread_local_bytes )
{
if ( pool_reduce_bytes < 512 ) pool_reduce_bytes = 512 ;
if ( team_reduce_bytes < 512 ) team_reduce_bytes = 512 ;
const size_t old_pool_reduce = g_serial_thread_team_data.pool_reduce_bytes();
const size_t old_team_reduce = g_serial_thread_team_data.team_reduce_bytes();
const size_t old_team_shared = g_serial_thread_team_data.team_shared_bytes();
const size_t old_thread_local = g_serial_thread_team_data.thread_local_bytes();
const size_t old_alloc_bytes = g_serial_thread_team_data.scratch_bytes();
// Allocate if any of the old allocation is tool small:
const bool allocate = ( old_pool_reduce < pool_reduce_bytes ) ||
( old_team_reduce < team_reduce_bytes ) ||
( old_team_shared < team_shared_bytes ) ||
( old_thread_local < thread_local_bytes );
if ( allocate ) {
Kokkos::HostSpace space ;
if ( old_alloc_bytes ) {
g_serial_thread_team_data.disband_team();
g_serial_thread_team_data.disband_pool();
space.deallocate( g_serial_thread_team_data.scratch_buffer()
, g_serial_thread_team_data.scratch_bytes() );
}
if ( pool_reduce_bytes < old_pool_reduce ) { pool_reduce_bytes = old_pool_reduce ; }
if ( team_reduce_bytes < old_team_reduce ) { team_reduce_bytes = old_team_reduce ; }
if ( team_shared_bytes < old_team_shared ) { team_shared_bytes = old_team_shared ; }
if ( thread_local_bytes < old_thread_local ) { thread_local_bytes = old_thread_local ; }
const size_t alloc_bytes =
HostThreadTeamData::scratch_size( pool_reduce_bytes
, team_reduce_bytes
, team_shared_bytes
, thread_local_bytes );
void * const ptr = space.allocate( alloc_bytes );
g_serial_thread_team_data.
scratch_assign( ((char *)ptr)
, alloc_bytes
, pool_reduce_bytes
, team_reduce_bytes
, team_shared_bytes
, thread_local_bytes );
HostThreadTeamData * pool[1] = { & g_serial_thread_team_data };
g_serial_thread_team_data.organize_pool( pool , 1 );
g_serial_thread_team_data.organize_team(1);
}
}
void Serial::finalize()
{
if ( Impl::g_serial_thread_team_data.scratch_buffer() ) {
Impl::g_serial_thread_team_data.disband_team();
Impl::g_serial_thread_team_data.disband_pool();
Kokkos::HostSpace space ;
space.deallocate( Impl::g_serial_thread_team_data.scratch_buffer()
, Impl::g_serial_thread_team_data.scratch_bytes() );
Impl::g_serial_thread_team_data.scratch_assign( (void*) 0, 0, 0, 0, 0, 0 );
}
#if defined(KOKKOS_ENABLE_PROFILING)
Kokkos::Profiling::finalize();
#endif
}
SharedAllocationRecord< Kokkos::HostSpace , void >::
SharedAllocationRecord( const Kokkos::HostSpace & arg_space
, const std::string & arg_label
, const size_t arg_alloc_size
, const SharedAllocationRecord< void , void >::function_type arg_dealloc
)
// Pass through allocated [ SharedAllocationHeader , user_memory ]
// Pass through deallocation function
: SharedAllocationRecord< void , void >
( & SharedAllocationRecord< Kokkos::HostSpace , void >::s_root_record
, reinterpret_cast<SharedAllocationHeader*>( arg_space.allocate( sizeof(SharedAllocationHeader) + arg_alloc_size ) )
, sizeof(SharedAllocationHeader) + arg_alloc_size
, arg_dealloc
)
, m_space( arg_space )
{
// Fill in the Header information
RecordBase::m_alloc_ptr->m_record = static_cast< SharedAllocationRecord< void , void > * >( this );
strncpy( RecordBase::m_alloc_ptr->m_label
, arg_label.c_str()
, SharedAllocationHeader::maximum_label_length
);
}