Re: MPI and private methods in octave::load_save_system

[John W. Eaton]

On 2/14/19 11:15 AM, address@hidden wrote:
> > On 13 Feb 2019, at 19:09, address@hidden wrote:
> >
> > No, that version of the package is not maintained anymore, the code in there is 
> > not compatible with Octave 4.4 or later [1].
> > Current development goes on in github [2] (actually I just now moved it from 
> > bitbucket ...)
> >
> > The relevant code is here:
> >
> > https://github.com/carlodefalco/octave-mpi/blob/master/src/MPI_Send.cc#L65
> > https://github.com/carlodefalco/octave-mpi/blob/master/src/MPI_Send.cc#L67
> >
> > https://github.com/carlodefalco/octave-mpi/blob/master/src/MPI_Recv.cc#L83
> > https://github.com/carlodefalco/octave-mpi/blob/master/src/MPI_Recv.cc#L90
>
>
> BTW, if this change were to be made, would it be on some stable or default?

I'd be willing to make minor changes on stable (moving a declaration 
from private to public, for example) if necessary.  As we discussed on 
IRC, I was considering making more extensive changes to avoid making 
some messy interfaces public again, but I'm not sure that we should do 
that.  We were talking about providing functions like

  octave_value load_save_system::read_value (std::istream& is, 
std::string& name, bool& global);

and

  void load_save_system::write_value (std::ostream& os, const 
octave_value& val, const std::string& name, bool global, const 
load_save_format& fmt, bool save_as_floats);

but when I started looking at what would be required to extract the the 
read_value function from the load_vars function it was not as simple as 
I assumed.  Some of the functions that ultimately write the data require 
extra info that wouldn't be in the above interface.  We should probably 
clean all of that up and use consistent interfaces.  But I think the 
changes are too much to make just before the release.

Also on IRC, you said the functions that are required are

  write_header
  save_binary_data
  read_binary_file_header
  read_binary_data

The save_binary_data and read_binary_data functions are public and 
declared in ls-oct-binary.h.  If you know that the data on the stream is 
Octave binary data using a particular float format, do you really need 
to read and write the header, or can you omit that?  If you can omit it, 
then I don't think we need to change anything before the release.  If 
you do need it, what is it used for?  If it really is necessary, then I 
would consider making those functions public for this release.

Longer term, however, I think it would be good to come up with better 
and more consistent interfaces.

jwe