----------------------------------------------------------------------
	   Software Coordinating Committee Conference Call
		       July 27, 2001 2:00 PM EDT
			  Recorder: C. DeTar

Present:

  Mendez, Holmgren, Simone, Mawhinney, DeTar, Edwards, Brower, Pochinski

Absent: 

  Watson

** = action required

Main agenda item: Linear Algebra API

    Hierarchy of API for discussion:

    Level 1:  Single site and MPI message passing
    Level 2a: Aggregate (lattice-wide)
    Level 2b: Aggregate plus communications
    Level 3:  Inverters

    Do we want to use C++?

    Robert:
	 Will we be compiling libaries in C++ and apps in C?
	 Then structures and classes are not guaranteed to be
	 stored consistently.
    Mawhinney: Should we be adopting C++ in the first place?
	 It is the industry standard, now.
    DeTar: C++ has better software engineering features, so spares us
         the risks of writing complex arithmetic in matrix times
         vector routines using clumsy macro calls.  With templates we
         don't have to supply separate complex arithmetic primitives
         and su3_matrix and vector classes for singles and doubles.
         The real choice is dictated by what method saves labor in the
         long run.
    Pochinski: Which C++ compiler?  They differ.
    Mawhinney: ANSI 
    DeTar: At the lowest level we strive to use only limited common
         denominator C++ features.
    Brower: Bottom layers - use C; Higher layers C++
    DeTar: But it is really all C++ - use minimal C++ at lower levels
         to assure pluggability of assembly code.
    Simone: Problems with name mangling.
    DeTar: Would use extern "C", so no overloading at lower levels.
    Mawhinnery: Would like to use C++ to allow operations
         between generalized classes of matrices and vectors
    Edwards: This introduces inefficiencies.
    Mawhinney: Optimization efforts should be focused on aggregate
         operations (acting on whole lattice).  The level 1 code is
         not the main issue.  Probably need to write interfaces to
         things like the Dirac operator that involve data remapping.

    ** Questions to resolve

    1. What does the interface look like to the user?  A call with
    pointers to C++ classes or C structures?  Perhaps provide a
    different header for the two choices.

    2. Do we pass classes to the level 2 interface -- classes that
    include information about storage order, or do we assume a storage
    order?  If the latter, we will have to do remapping.

    ** Discuss these issues with our colleagues.

Next call on July 30, 31 to continue discussion, time to be announced.

Call concluded at 4:00 PM EDT.