parallel_reduce(0), parallel_scan(0) unit tests

[ndellingwood]

While adding unit tests for parallel_scan of size 0 (to go along with tests in #175 ) the serial.team_scan test failed with both gcc 4.9 and gcc 5.4.
Error:

[ RUN      ] serial.team_scan
Floating point exception (core dumped)

[ndellingwood]

Working on fix for the unit test.

[ndellingwood]

Fix for above was easy, found a couple more issues with cuda after adding more tests for parallel_reduce(0) and parallel_scan(0). Will track with this issue and changed issue name to better reflect this.

[ndellingwood]

Cuda test failures (runtime):
test_reduce(0) test failed with Cuda (called in TestCuda_b.hpp for example TestTeamPolicy< Kokkos::Cuda , Kokkos::Schedule<Kokkos::Static> >::test_reduce(0); ) as well as TestReduceTeam

In test_reduce, passing in 0 results in a league_size=0, and a team_size of 256 is returned by team_size_max(functor). These values are then passed in the team policy to parallel_reduce.
During execution of the parallel_reduce the body of the functor operator is never executed, however parallel_reduce returns a result of 501501 ( = 1001*1002/2 = sum(1 to 1001) in case that is relevant).

Fix (so far):
The execute() method in the RangePolicy version of ParallelReduce is protected when the amount of work is 0 i.e.

    void execute() {
      const int nwork = m_policy.end() - m_policy.begin();
      if ( nwork ) {  ...
      //stuff...
      }
     else {
       if (m_result_ptr) {
         ValueInit::init( ReducerConditional::select(m_functor , m_reducer) , m_result_ptr );
       }

I added this to the TeamPolicy version of the ParallelReduce as well, that is, within the execute() method I added the if - else check as above but replaced nwork with const int nwork = m_league_size * m_team_size
This fixed the errors I was seeing with Cuda testing with a generated makefile and make build-test (will check with test_all_sandia).

@crtrott Is this sufficient? I'm afraid I may be overlooking something within the team policy that I am not familiar with.

[crtrott]

I think this looks fine. If you put it in as a pull request instead of pushing directly I will check it before we merge it in.

[ndellingwood]

The following test added to TestThreads.hpp (parallel_for(0) type test, gcc compiler) produces a segfault:

TestTeamPolicy< Kokkos::Threads , Kokkos::Schedule<Kokkos::Dynamic> >::test_for(0);

I have a fix, but not sure if it is the proper fix or may mask something else.

Issue:
Passing in the 0 to test_for results in a league_size of 0.
In the threads' ParallelFor specialized for Dynamic type schedule, a call to exec_team is made with a for loop that calls member.valid_dynamic() (defined in ThreadsExecTeamMember )
Within this call use of m_team_base[0] results in a seg fault,
e.g. m_team_base[0]->get_work_index(m_team_alloc);
None of the body of the ThreadsExecTeamMember ctor is executed so none of the values are 'properly' initialized (whatever that means for league_size of 0).

Working solution:
In the ThreadsExecTeamMember ctor body, added an else condition that sets m_invalid_thread = 0; when the condition if ( team.league_size() ) fails. This results in the unit test passing as valid_dynamic() returns false at the beginning of the code.

[ndellingwood]

@crtrott
The following reducers test failed at runtime with Threads backend and gcc 5.4.

/Users/ndellin/Research/kokkos_items/kokkos/core/unit_test/TestReduce.hpp:1581: Failure
Value of: reference_minloc
  Actual: 3611
Expected: minmax_scalar.min_loc
Which is: 8639
[  FAILED  ] threads.reducers (4 ms)

[ndellingwood]

Pull request issued #440