Skip to content

Kokkos::ThreadVectorRange

Christian Trott edited this page May 1, 2020 · 1 revision

Kokkos::ThreadVectorRange

Header File: Kokkos_Core.hpp

Usage:

  parallel_for(ThreadVectorRange(team,range), [=] (int i) {...});
  parallel_reduce(ThreadVectorRange(team,begin,end), 
    [=] (int i, double& lsum) {...},sum);
  parallel_scan(ThreadVectorRange(team,range), 
    [=] (int i, double& lsum, bool final) {...});

ThreadVectorRange is a nested execution policy used inside of hierarchical parallelism. In contrast to global policies, the public interface for nested policies is implemented as functions, in order to enable implicit templating on the execution space type via the team handle.

Synopsis

 template<class TeamMemberType, class iType>
  /* implementation defined */ ThreadVectorRange(TeamMemberType team, iType count);
 template<class TeamMemberType, class iType1, class iType2>
  /* implementation defined */ ThreadVectorRange(TeamMemberType team, iType1 begin, iType2 end);

Description

  • template<class TeamMemberType, class iType>
    /* Implementation defined */ ThreadVectorRange(TeamMemberType team, iType count);

    Splits the index range 0 to count-1 over the vector lanes of the calling thread.

    • Arguments

      • team: a handle to the calling team execution context.
      • count: index range length.
    • Returns

      • Implementation defined type.
    • Requirements

      • TeamMemberType is a type that models TeamHandle
      • std::is_integral<iType>::value is true.
      • count >= 0 is true;
      • This function can not be called inside a parallel operation dispatched using a TeamVectorRange policy or ThreadVectorRange policy.
  • template<class TeamMemberType, class iType1, class iType2>
    /* Implementation defined */ ThreadVectorRange(TeamMemberType team, iType1 begin, iType2 end);

    Splits the index range begin to end-1 over the vector lanes of the calling thread.

    • Arguments

      • team: a handle to the calling team execution context.
      • begin: index range begin.
      • end: index range end.
    • Returns

      • Implementation defined type.
    • Requirements

      • TeamMemberType is a type that models TeamHandle
      • std::is_integral<iType1>::value is true.
      • std::is_integral<iType2>::value is true.
      • end >= begin is true;
      • This function can not be called inside a parallel operation dispatched using a TeamVectorRange policy or ThreadVectorRange policy.

Examples

 typedef TeamPolicy<>::member_type team_handle;
 parallel_for(TeamPolicy<>(N,AUTO,4), KOKKOS_LAMBDA (const team_handle& team) {
   int n = team.league_rank();
   parallel_for(TeamThreadRange(team,M), [&] (const int i) {
     parallel_for(ThreadVectorRange(team,K), [&] (const int j) {
       A(n,i,j) = B(n,i) + j;
     });
   });
   team.team_barrier();
   int team_sum;
   parallel_reduce(TeamThreadRange(team,M), [&] (const int& i, int& threadsum) {
     int tsum = 0;
     parallel_reduce(ThreadVectorRange(team,M), [&] (const int& j, int& lsum) {
       lsum += A(n,i,j);
     },tsum);
     single(PerThread(team),[&] () {
       threadsum += tsum;
     });
   },team_sum);
     
     lsum += A(n,i);
   },team_sum);
   single(PerTeam(team),[&] () {
     A_rowsum(n) += team_sum;
   });
 });
Clone this wiki locally