Re: [Chasm-interop-devel] The future of Chasm?

[Tom E. <tep...@ll...>]

ber...@or... wrote:
> 2) The Babel folks need to understand whether what's been proposed for
>    F2008 really covers everything they've been getting from Chasm thus
>    far. (Also, I guess the previous draft stuff that some vendors have
>    already implemented.)
>   
I've downloaded the latest draft for the next Fortran standard 
(2007-Sep-27), and I've read the relevant section on C interoperability 
(15). It seems largely like what I read in the F2003 standard. *The 
Fortran 2008 standard as currently proposed DOES NOT provide the level 
of interoperability that Babel requires and that Chasm provides.* Even a 
complete implementation of the F2003 or F2008 standard would not be 
sufficient to provide the level of interoperability that Babel needs. In 
particular, the existing F2003 standard *and *the proposed F2008 
standard do not provide sufficient capabilities to:

   1. Allow passing array slices without copying data. For example, if I
      take a 2-D slice of a 3-D array in Babel, I can end up with an
      array where no dimension is stride one. With CHASM and Babel, I
      can pass that array back and forth between Fortran 90/95 and any
      other supported language without requiring copying. As far as I
      can see, the bind C approach requires arrays that are passed
      between Fortran and C to be dense and column-major ordered. Babel
      needs access to the internals of the array descriptor to pass
      other kinds of arrays without copying.
   2. The C_F_POINTER subroutine forces every array lower bound to be 1
      regardless of what the Babel array's lower bound is. In general,
      the bind C feature doesn't give us control of the lower and upper
      bounds like CHASM does.
   3. Note 15.17 says "A polymorphic, allocatable, or pointer array is
      never interoperable."  This seems to be a problem too because
      that's basically what Babel arrays are.


> 3) Assuming the F2008 stuff is an appropriate replacement, Babel needs
>    to extended to support it -- testing the compiler for support in
>    autoconf, and generating the appropriate code.
>   
This assumption is completely and utterly incorrect. The 
interoperability provided by the F2003/F2008 standards (as they 
currently exist) is insufficient for (at least) the reasons enumerated 
above. To be utterly clear, we cannot provide Babel users the level of 
Fortran interoperability that they're accustomed to in Babel's F90/95 
binding using the features available in F2003/F2008 standard as they 
exist today.
> 4) We need for Chasm to be maintained until users upgrade to compiler
>    versions supporting the F2008 features.  Just because a site has a
>    compiler version that does support the needed features, doesn't
>    mean the users are actually using that version.  Many sites,
>    especially large centers, maintain multiple compiler versions
>    simultaneously.
>   
We need a library, such as Chasm, to provide access to internals of the 
Fortran dope vectors, so we have direct access to the lower bound, upper 
bound, and stride information about arrays. Until a Fortran standard 
comes out and is implemented to provide that level of access to the 
internals of Fortran arrays, we will need the capabilities that Chasm 
currently provides.

Tom Epperly
Babel Architect
https://computation.llnl.gov/casc/components/