allow an Accessor type to return a known receive size

[jjwilke]

to receive values, you need the following steps:

MPI_Probe(); //might not be called until quite late in the progress engine
getSize();
MPI_Irecv();
....
MPI_Wait(...);
unpack();

Even if the application knows ahead of time how large the receive should be, it just has to wait for the packed message to arrive.

if would be very useful to be able to do this:

if (Accessor::canPrecomputeSize()){
  int size = Accessor::precomputeSize(...);
  void* buf = allocate_temp_buf(size);
  MPI_Irecv(...);
}
...
MPI_Wait(...);
unpack...);

This probably makes the most sense to do using SFINAE on the Accessor to see if it has a static function "precompute_size()"

[jjwilke]

see branch index-lookup-8 for how the send/recv system is set up

[dhollman]

This should take 10 minutes. It’s the kind of thing we’ve gotten very good at over the past couple of years. I’ll do it when I get in

[dhollman]

@jjwilke By making precompute_size() a static method that takes no arguments, you're implying that it's statically known information, right? Wouldn't this better be expressed with a static data member static_size or something like that?

It might actually make more sense to somehow allow instances of accessors and use instance methods for all of these rather than just the types and static methods, since the receiving side might know ahead of time what the size is, but it may depend on the mesh or some other locally-available information.

[giulioborghesi]

I have already written a metafunction to do this. David, are you still doing it, or should I just go ahead and push my code?

[dhollman]

I mean, using tinympl, it's just:

template <typename T>
using _has_precompute_size_archetype = decltype(T::precompute_size());
template <typename T>
static constexpr auto has_precompute_size = tinympl::is_detected_convertible<size_t,
  _has_precompute_size_archetype, T
>::value;

But we need to work out the kinks of the Accessor concept before we go too far down this road.

[jjwilke]

Yes, it needs to take an argument just like pack or unpack. The information is state-dependent and dynamic - but the the fact that the incoming size is known is statically known.

[jjwilke]

I would probably like to see a similar function like "preexisting_buffer()" that means the object also maintains its own, permanent buffers that can be used.