boolean reduction value_type changes answer

[ibaned]

I'm trying to write some custom parallel_reduce functors which will compute whether a condition is true for all items in an array. I tried to use "bool" as the value_type and implement methods based on the && operator. Strangely, it gives the wrong output when value_type is bool but I can fix it by using an integer value_type:

struct SameContent {
#if 0
  typedef int value_type;
#else
  typedef bool value_type;
#endif
  Kokkos::View<int*> a_;
  Kokkos::View<int*> b_;
  SameContent(Kokkos::View<int*> a, Kokkos::View<int*> b):a_(a),b_(b) {}
  KOKKOS_INLINE_FUNCTION
  void init(value_type& update) const
  {
    update = 1;
  }
  KOKKOS_INLINE_FUNCTION
  void join(volatile value_type& update,
      const volatile value_type& input) const
  {
    update = update && input;
  }
  KOKKOS_INLINE_FUNCTION
  void operator()(int i, value_type& update) const
  {
    update = update && (a_[i] == b_[i]);
  }
};

Am I misusing Kokkos here ?
If this is something that should work, I have a two-file standalone example that replicates the issue. All the above is using Kokkos with CUDA on Shannon.

[hcedwar]

Reduction value_type needs to be a word-aligned plain-old-data types. Unfortunately 'bool' can be compiled to a fraction of a word, which is probably the source of the error. We should introduce a static_assert to catch / enforce this requirement.

[ibaned]

Thank you for the explanation. I'll switch to word-sized value_types in my code. A compile-time check would be much appreciated.

[hcedwar]

The types can be larger than a word (i.e., int) just not smaller. Re-opening until we get the static_assert into the code, then will close.

[crtrott]

Pushed to Master and Trilinos: you can not do reductions with types smaller than 32bit now

[ibaned]

Just FYI, I had to increase this further from int32_t to int64_t to use an NVidia GTX 980 Ti with the CUDA 8.0 RC.

ibaned/omega_h@18110db