Jack Dongarra opened the meeting by welcoming everyone and inviting everyone to introduce themselves. He then gave a brief summary of the previous BLAS workshops, and asked which subgroups should meet at this workshop. It was decided that the following subgroups would be discussed:
BLAS functionality was the first to be addressed. The discussion focused on the draft of the Functionality proposal that Barry Smith of ANL and Sven Hammarling of NAG had written and made available via the BLAS Forum homepage. Sven gave a brief overview of the proposal and invited comments and suggestions. Several attendees voiced their opinions on the proposal, and each of the lists of proposed routines were discussed, as well as possible additions. Revised lists were to be prepared that night and presented the following day. Anne Trefethen volunteered to expand the ``multiple instances'' section for the proposal.
No particular programming language will be primary in the discussion of functionality, instead we will discuss syntax and semantics for Fortran 77, Fortran 90, C, and C++. We shall provide reference implementations for the proposed BLAS routines in each language.
There are different ways of classifying functionality -- BLAS-like and tensor notation. Andrew Lumsdaine led a discussion explaining the proposed tensor notation. An open question for the forum is whether one or both of the classification schemes should be used.
After a break, the BLAS Lite were discussed. Two versions of the BLAS Lite were proposed. One (debugging) version with error-checking, and another (performance) version without error-checking. To provide good performance on all problem sizes, separate interfaces for stride 1 operations may be provided, as well as separate routines for small (block) problems, e.g., matrix multiplies of size 2, 3, 4, and 5.
Puri Bangalore raised the issue of specifying time and space (CPU and memory) complexity for various routines. This sort of specification is clearly useful for users of the routines, especially when the user might want to choose between various routines that provide similar functionality but which have different CPU/memory usage. However, it is not clear that many of the proposed BLAS functions admit implementations that have differing CPU/memory complexity. The obvious exceptions are xNRM2 (Euclidean vector norm) and matrix multiply. An open question here is whether we should give options for varying degrees of accuracy depending upon time and space considerations. Specifically, should a fast version of xNRM2 or a Strassen matrix multiply be provided?
Jack Dongarra opened the morning session with a summary of the previous day's discussions. The first topic of the morning was the discussion of the object-based BLAS interface, led by Andrew Lumsdaine.
During the discussion of the object-based interface, Andrew Lumsdaine mentioned that it may be important to consider the effects of new parts of the ANSI C++ standard, especially the Standard Template Library. He stressed the composition (templates) approach to building the library instead of the inheritance approach.
After a short break, Sven Hammarling spoke on the proposed revisions to the Functionality proposal. Specifically, updated tables of routines were presented.
Anne Trefethen gave input on the needed ``multiple instance'' routines.
The Lite BLAS were again discussed. Puri Bangalore opened the discussion by asking which data types, matrix types, and storage formats should be supported. Puri provided a list as a starting point and several attendees suggested additions.
The issue of error checking was introduced. Puri suggested that all functions should return an error code. Barry Smith pointed out that in the BLAS Lite there are very few errors that can occur. It was then suggested that each BLAS function might have two entry points, in similar fashion to MPI.
Puri then raised the thorny issue of indexing, in particular, whether indexing should be 0 based (as in C) or 1 based (as in Fortran). For dense linear algebra this is not a real problem, but it is a problem when user data maintains explicit indexing information (as in compressed sparse storage schemes). Suggested options for indexing were to support both via run-time or compile-time schemes or to make the index support language-dependent.
Next, Puri asked whether a stride argument is needed. Sven indicated that stride arguments are convenient and are required by some algorithms.
Puri also addressed guidelines to implementors and discussed these four issues:
The vendors expressed concern about the amount of work that these guidelines might impose upon them. It was clarified that this was intended to be mostly documentation, rather than additional implementation work. Puri suggested that the implementors were free to do what is best for their architecture, but as a library writer or user one would like to know what has been done, at least in terms of memory usage if extra memory is allocated within a function.
After a lunch break, Barry Smith spoke of sample implementations of the BLAS Lite. Proposed calling sequences for routines were discussed, as well as the necessity for reference implementations of some of the routines. Barry volunteered for ANL to implement the C and C++ versions of the routines. The University of Tennessee will provide Fortran 90 reference implementations for some of the routines.
Jack Dongarra wrapped up the meeting by suggesting that the Functionality proposal be finished by summer. Andrew Lumsdaine will also draft an Object-based BLAS proposal. Both proposals will be available on the Web page, and comments/suggestions from the user community are welcome.
The dates of the next forum meetings are:
The meeting was then adjourned by Jack Dongarra at 3:30 PM.
Attendees list for the May 20-21, 1996 BLAST Forum Meeting
Satish Balay ANL balay@mcs.anl.gov Puri Bangalore Miss. State Univ. puri@cs.msstate.edu Andrew Cleary Univ. of TN cleary@cs.utk.edu Jack Dongarra Univ. of TN / ORNL dongarra@cs.utk.edu Sven Hammarling NAG, UK / Univ. of TN hammarli@cs.utk.edu Hidehiko Hasegawa ULIS, Tsukuba, Japan hasegawa@ulis.ac.jp Satomi Hasegawa Hitachi 75207.2076@Compuserve.com Naoki Iwata NEC Systems Laboratory iwata@hstc.necsyl.com Chandrika Kamath DEC kamath@caldec.enet.dec.com Guangye Li Cray Research gli@cray.com Hsin-Ying Lin HP Convex Technology Ctr. lin@rsn.hp.com Andrew Lumsdaine Univ. of Notre Dame Andrew.Lumsdaine@nd.edu Joan McComb IBM Poughkeepsie mccomb@vnet.ibm.com Susan Ostrouchov Univ. of TN susan@cs.utk.edu Antoine Petitet Univ. of TN petitet@cs.utk.edu Barry Smith ANL bsmith@mcs.anl.gov Anne Trefethen Cornell Theory Center aet@tc.cornell.edu
Susan Ostrouchov and Andrew Lumsdaine agreed to take minutes for the meetings.