Another attempt at h2d

[jgfouca]

Fixes the case where user is trying to use unpacked pre-allocated device views.

host_to_device has so many overloads and template that it was getting a bit hard to understand what was going on. Separate the actual implementation out and have the user-facing functions call it. Some of the user-facing functions have multiple levels of syntactic sugar.

Testing

Added this case onto the existing test.

[bartgol]

Is this function reinterpreting, say, a vector<View<T*>> as vector<View<S*>>? This seems dangerous. I see that you use it in a safe way in the utils, as you interpret View<Real**> as View<Pack<Real,1>**, which has the same extent. But it is somewhat tricky.

We should at the very least verify that sizeof(KokkosT)==sizeof(ViewT::value_type), to avoid bad usage.

Even better would be to make this fcn a bit more "private", like putting it in an impl namespace to discourage usage. I'm weary of exposing a hacky-ish function...

[bartgol]

Another option would be to call this repack_views and do something like

template<int N, typename ViewT>
auto repack_views(std::vector<ViewT>& views) {
  using packed_view = decltype(repack<N>(views[0]));
  std::vector<packed_view> ret;
  for (auto v : views)
    ret.push_back(repack<N>(v));
  return ret;
}

[jgfouca]

Good idea

[jgfouca]

Ok, done