Jack Dongarra opened the meeting by welcoming everyone and inviting everyone to introduce themselves. He then reviewed the technical ``rules of order'' for the first and second readings of proposals, as well as the voting procedures for the meeting. Tony Skjellum suggested that we integrate all of the proposals into one unified document, similar in style to the MPI proposal. This issue was discussed in further detail later in the meeting.
The following presentations were made:
BLAS functionality was the first to be addressed. Sven Hammarling of NAG gave a brief overview of the Functionality proposal that has been available via the BLAST Forum homepage, as well as addressing several email messages from users pertaining to desired ``new'' routines. He stressed the areas of the proposal that still need to be written, specifically the introductory background motivation, performance issues, numerical stability, as well as justification for the new proposed functionality. Any new functionality that is proposed should have a clear justification identifying the applications where this operation is needed. Several attendees debated if some of the information contained in the proposal is not strictly ``functionality'' issues, but rather language independent specifications as well as language bindings. It was proposed that some of the information in the tables of proposed routines be reorganized into separate sections on these issues.
Linda Kaufman of Bell Labs then spoke about her recent revisions to the proposal, namely adding priorities to each of the new proposed routines. Mimi Celis of SGI and Sven Hammarling of NAG directed attention to the standardization efforts of the VSIP (Vector, Signal and Image Processing) community. Roldan Pozo of NIST proposed a study to see the use of BLAS in applications. He stressed an effort to clearly justify what is included and what is not included in order to realistically accomplish our goals. Several of the application developers present in the meeting voiced that they felt the ``Functionality'' proposal does accurately reflect the needs of their users. However, it was also pointed out that there are those in the numerical analysis community who will never rewrite their codes in terms of the BLAS. And in some cases, it is too expensive to rewrite optimized existing application programs such as NASTRAN and GAUSSIAN in terms of the BLAS. Some vendors have inserted BLAS calls inside these codes, while others have chosen to optimize them in different ways.
Tony Skjellum of Mississippi State emphasized that most programmers develop their algorithms first in ``matlab-type'' environments. The use of standards in numerical analysis programming could be increased by promoting a ``BLAST toolbox'' in Matlab, or more strongly encourage the use of LAPACK and the BLAS inside of Matlab.
After a break, Roldan Pozo of NIST presented the sparse BLAS proposal that is available on the BLAST Forum homepage. He discussed the roles and functionality of the Sparse BLAS, as well as open issues in their development. Again it was voiced that several of the issues discussed were not primarily functionality issues but language binding issues. Concern was also expressed by several attendees due to the sheer volume of routines proposed. In order for a sparse BLAS to be more readily accepted, the number of proposed routines may need to be reduced.
After Roldan's presentation, a break-out session for the meeting of subgroups was held. The following subgroups met: functionality, sparse BLAS, parallel BLAS, lite BLAS, and the C interface to the BLAS.
Jack Dongarra opened the morning session with a brief summary of the previous day's discussions.
Tony Skjellum began the morning discussions with an overview of his thoughts on the presentation of our ``proposals'' for the BLAST work, specifically organization of a unified document for the BLAST proposal instead of separate ``proposals''. The document would primarily contain separate chapters on functionality, language independent specifications, and language binding issues. Interleaved in these sections would be information pertaining to the sequential, sparse, parallel, etc. Several attendees voiced concern at this one proposal idea instead of separate proposals. Specifically, if a user or vendor is only interested in information pertaining to the sparse BLAS, he would still need to read the entire document. Richard Hanson of VNI suggested that we have an ``implementation'' chapter or appendix to ease the integration job for the vendors. Linda Kaufman of Bell Labs further suggested a ``tutorial'' or ``quick reference guide'' for each of the different standards proposed. Jack Dongarra of UT stressed the importance of a reference implementation that should be included with the standard.
Following this discussion, Tony Skjellum presented the BLAS Lite proposal that is available on the BLAST Forum homepage. Several vendors expressed concern about the volume of routines that are proposed. It was suggested that a more extensive performance evaluation of the Lite BLAS be conducted before a decision is made on their inclusion in the standardization effort. Concern was raised about the portability of the Lite BLAS since they are in effect a ``cache-based'' BLAS. Andrew Lumsdaine and Brian McCandless, who have implemented preliminary versions of the BLAS Lite routines, promised to make their results available to the BLAST Forum in a timely fashion.
Antoine Petitet of UT led discussion on the issues involved in developing a set of parallel BLAS. A note discussing the issues is available on the BLAST Forum homepage.
Chenyi Hu of the University of Houston, Downtown, presented a few slides on the needs of the Interval BLAS community. It was suggested that a separate section for the Interval BLAS be included in the BLAST proposal.
After lunch, Clint Whaley of UT summarized what is contained in the ``C Interface to the Existing BLAS'' proposal available on the BLAS Forum homepage. He also discussed what occurred in the ``break-out'' session the previous day. Jesse Bennett of Texas Instruments also gave a brief presentation on what he perceives as user needs for the CBLAS. The majority of opinions from vendors supported the view that the operations should be ``row-major'', and that character arguments should be handled as enumerated types. Mimi Celis of SGI specifically said that SGI had introduced a C interface for the BLAS which was ``column-major'' and the C users refused to use it. It was then suggested that it may be possible to present a ``row-major'' interface to the existing BLAS through careful manipulation of the transpose in calculations. An attendee thought that this is the way that Digital Corporation presented a ``row-major'' interface, and this is under investigation. Again the VSIP standardization effort was mentioned in regards to two-dimensional arrays in Fortran versus C. Mimi Celis of SGI stated that in the VSIP effort it is assumed that two-dimensional arrays in C are contiguous. Email conversation on these topics (as well as the issue of a ``complex'' data type) were encouraged.
Andrew Lumsdaine next spoke about the question of the organization of the Forum, the presentation of the proposal document, the varied user base, and his recommendations for how we should proceed. Concern was again raised for the user who is only interested in one aspect of the standard and must read the entire document. A straw vote was taken to ascertain if we should try to integrate the BLAST proposal into one unified document. Appendices for a ``tutorial'' and ``reference implementation'' will be included. The vote was unanimous to integrate the proposals into one unified document. Twenty-two people were present for the vote.
A summary of the meeting was then provided by Jack Dongarra. He focused on the specific tasks of writing several needed chapters for the BLAST proposal. Namely:
The document will be established on the net, and individual chapters will be available for downloading in latex format.
The global editors for the BLAST proposal are Andrew Lumsdaine and Tony Skjellum.
The tentative date of the next forum meeting is:
and will be hosted by Intel Corporation, with a preliminary deadline of April 30, 1997 for subgroup progress.
The meeting was then adjourned by Jack Dongarra at 3:00 PM.
Attendees list for the February 27-28, 1997 BLAST Forum Meeting
Puri Bangalore Miss. State Univ. puri@cs.msstate.edu Jesse Bennett Texas Instruments jesse@seas.smu.edu Susan Blackford Univ. of TN, Knoxville susan@cs.utk.edu Mimi Celis SGI celis@sgi.com Andrew Chapman NEC Systems Laboratory chapman@hstc.necsyl.com Isom Crawford HP/Convex isom@rsn.hp.com Dave Dodson HP/Convex dodson@rsn.hp.com Jack Dongarra Univ. of TN / ORNL dongarra@cs.utk.edu Paul Dressel HP/Convex dressel@rsn.hp.com Cormac Garvey NEC Systems Laboratory garvey@hstc.necsyl.com Ian Gladwell SMU gladwell@seas.smu.edu Bruce Greer Intel bruce_s_greer@ccm.jf.intel.com John Gunnels Univ. of TX, Austin gunnels@cs.utexas.edu Sven Hammarling NAG, UK sven@nag.co.uk Richard Hanson VNI hanson@houston.vni.com Chenyi Hu Univ. of Houston/Downtown chu@uh.edu Linda Kaufman Bell Labs lck@lucent.com Jim Koehler HP/Convex koehler@convex.hp.com Hsin-Ying Lin HP Convex Technology Ctr. lin@rsn.hp.com John Liu HP/Convex jliu@rsn.hp.com Andrew Lumsdaine Univ. of Notre Dame Lumsdaine.1@nd.edu Brian McCandless Univ. of Notre Dame bmccandl@nd.edu Joan McComb IBM Poughkeepsie mccomb@vnet.ibm.com Jim Nagy SMU jnagy@mail.smu.edu Antoine Petitet Univ. of TN petitet@cs.utk.edu Roldan Pozo NIST pozo@nist.gov Karin Remington NIST karin@cam.nist.gov Tony Skjellum Miss. State Univ. tony@cs.msstate.edu Shane Story Intel shane@ibeam.jf.intel.com Kevin Wadleigh HP/Convex wadleigh@rsn.hp.com Clint Whaley Univ. of TN, Knoxville rwhaley@cs.utk.edu Guodong Zhang HP/Convex zhang@rsn.hp.com
Susan Blackford and Andrew Lumsdaine agreed to take minutes for the meetings.