KOKKOS_INLINE_FUNCTION
size_t get_thread_id(const typename Kokkos::TeamPolicy<ExecutionSpace>::member_type & teamMember) const{
#if defined( KOKKOS_HAVE_SERIAL )
if (Kokkos::Impl::is_same< Kokkos::Serial , ExecutionSpace >::value){
return 0;
}
#endif
#if defined( KOKKOS_HAVE_PTHREAD )
if (Kokkos::Impl::is_same< Kokkos::Threads , ExecutionSpace >::value){
return Kokkos::Threads::hardware_thread_id();
}
#endif
#if defined( KOKKOS_HAVE_OPENMP )
if (Kokkos::Impl::is_same< Kokkos::OpenMP, ExecutionSpace >::value){
return Kokkos::OpenMP::hardware_thread_id();
}
#endif
return teamMember.league_rank() * teamMember.team_size()+ teamMember.team_rank();
}
KOKKOS_INLINE_FUNCTION
void operator()(const typename Kokkos::TeamPolicy<ExecutionSpace>::member_type & teamMember) const {
Kokkos::single(Kokkos::PerTeam(teamMember),[=] () {
printf("teamMember teamsize:%d\n", teamMember.team_size());
});
volatile idx * myData = NULL;
size_t tid = this->get_thread_id(teamMember);
int trial = 0;
while (myData == NULL){
++trial;
Kokkos::single(Kokkos::PerThread(teamMember),[&] (volatile idx * &memptr) {
memptr = (volatile idx * )this->my_memory_pool.allocate_chunk(tid);
}, myData);
}
for (int i = 0; i < 100; ++i){
Kokkos::parallel_for(
Kokkos::ThreadVectorRange(teamMember, 32),
[&] (int j) {
myData[j] = i;
});
}
Kokkos::parallel_for(
Kokkos::ThreadVectorRange(teamMember, 32),
[&] (int j) {
myData[j] = -1;
});
Kokkos::single(Kokkos::PerThread(teamMember),[=] () {
/*printf("me:%ld lr:%d ts:%d tr:%d, Had Memory location:%ld with chunk_index:%ld in this many tries:%d\n",
tid, (int) teamMember.league_rank(), (int) teamMember.team_size(), (int) teamMember.team_rank(),
myData,my_memory_pool.get_chunk_index((idx *) myData), trial );*/
this->my_memory_pool.release_chunk((idx *) myData);
});
myData = NULL;
}
KOKKOS_INLINE_FUNCTION
void operator() (typename Kokkos::TeamPolicy<DeviceType>::member_type dev) const {
const int chunk = ((nlast - nfirst + dev.league_size() - 1 ) /
dev.league_size());
const int teamstart = chunk*dev.league_rank() + nfirst;
const int teamend = (teamstart + chunk) < nlast?(teamstart + chunk):nlast;
int mysend = 0;
for (int i=teamstart + dev.team_rank(); i<teamend; i+=dev.team_size()) {
if (x(i,dim) >= lo && x(i,dim) <= hi) mysend++;
}
const int my_store_pos = dev.team_scan(mysend,&nsend(0));
if (my_store_pos+mysend < maxsendlist) {
mysend = my_store_pos;
for(int i=teamstart + dev.team_rank(); i<teamend; i+=dev.team_size()){
if (x(i,dim) >= lo && x(i,dim) <= hi) {
sendlist(iswap,mysend++) = i;
}
}
}
}