Everyone assembled and we began with lunch.
At 1:00pm, Jack Dongarra opened the meeting by welcoming everyone and inviting everyone to introduce themselves. He then reviewed the ``rules of order'' for the first and second readings of proposals, as well as the voting procedures for the meeting. The structure of the overall BLAS document was briefly reviewed. Each chapter is self-contained consisting of sections on overview, functionality, language independent specificiation, language dependent specification, and the reference implementation, benchmarking, verification, and validation. Debate ensued on whether Extensions should be included in the Dense chapter or the Legacy BLAS appendix/chapter. It was stressed that this list of extensions should be restricted to 2-3 routines.
A straw vote was taken on the inclusion of the Extensions in Legacy BLAS appendix or restrict it to the "Dense and Band BLAS" chapter. Mike Heroux presented the motion, and Tony seconded the motion. Of 20 attendees, 10 attendees voted in favor of its inclusion in the Legacy BLAS chapter, 7 attendees opposed, and 3 attendees abstained.
At 2:00pm, Jim Demmel presented a detailed discussion of the proposed Extended Precision BLAS. He outlined the contents of the proposal, beginning with the motivation behind the need for such routines, presenting specific examples, and proposed language bindings. Lengthy discussion took place regarding the language bindings, with particular concerns raised about the combinatorial explosion of functions required to handle all possible combinations of precisions that could occur for the arguments to functions in the extended precision BLAS. Further discussion was deferred to the upcoming break-out session, and communication via the email reflector.
A short coffee break occurred at 3:45pm.
The meeting was reconvened at 4:00pm, and broke into subgroups. Two subgroups met at 4:00-5:00pm, and two subgroups met at 5:00-6:00pm. The Functionality, Extensions, and Dense and Band BLAS, were combined into one subgroup, which met at 4:00pm. The Sparse BLAS subgroup also met at 4:00pm. At 5:00pm, the Extended Precision subgroup and C subgroup met. Due to conflict of interest, the Distributed Memory and F90 subgroups were postponed until after dinner or tomorrow.
The meeting adjourned at 6:00pm.
At 9:00am, a plenary presentation of the results of yesterday's break-out sessions occurred.
Linda Kaufmann presented a summary of discussions regarding The Functionality/Extensions/Dense BLAS session. The sub-group reviewed Chapters 1 and 2 of the document and tried to outline what should be deleted from the tables due to lack of justification, what sections should be combined, etc. Specifically, it was said that all entries in the tables of Chapter 1 must appear as part of the functionality in one of the chapters. If an entry does not appear, it will be deleted from the tables in Chapter 1.
Gary Howell of Florida Tech proposed the introduction of elementary similarity transformations into the functionality. His rationale was that implementing elementary similarity transformations in terms of matrix vector multiplies and rank one updates is possible but INEFFICIENT in that the entire active part of the matrix is repeatedly accessed unnecessarily. It was decided to continue this discussion via the mail reflectors.
Householder transformations were again suggested as a top priority new BLAS routine due to their necessity in Hessenberg reduction. Sequences of Householder transformations (some of same length, some of different length) were also proposed as a very important addition due to its application in QR bulge chasing. Sequential Givens (application: QR) were also desired, such as in LAPACK auxiliary routine xLASR. Simultaneous Givens transformations (reduction to tridiagonal form) (xLARTV, replaced by DROT) were also suggested. Linda Kaufman then presented a summary of the list of vendor extensions to the BLAS. Namely, IxAMIN and a routine to return the Index of the element of minimum magnitude.
Roldan Pozo of NIST and Mike Heroux of Cray then presented a summary of the Sparse BLAS subgroup. Roldan began by suggesting a delay in the first reading of document until the addition of user-level text. A preliminary discussion of the functionality subset occurred during the break-out session and focused on Level 3 BLAS as well as preconditioned iterative methods. As for storage formats, it was suggested to support a data-neutral interface, as well as storage-specific interfaces. These storage formats would be extendable at the User Level, and the core formats would be utilized at the Toolkit level. Example language bindings for C and Fortran77 were included in the draft of the proposal, as well as the reference implementation contained in the Toolkit. There seemed to be little support for a low-level interface.
Mike Heroux of SGI/Cray then discussed Sparse functionality, specifically in the Toolkit. Permutations had not been included in the overall functionality definitions in Chapter 1, and should now be included in Chapter1. He reiterated the principle goal for the sparse BLAS, namely, efficient computation of the most common sparse ops in the most common data structures, for preconditioned Krylov methods. He also discussed some open issues in the development of Toolkit kernels.
Jim Demmel of UC Berkeley then presented the summary of the Extended Precision subgroup. Concern was raised over the volume of routines, and it was stressed that only a subset of routines would be implemented for a given architecture. The necessity for environmental inquiry routines was discussed.
Coffee break at 11:00am.
At 11:15am, Antoine Petitet presented the summary of the Distributed Memory BLAS subgroup. The subgroup proposed to Reuse interface design of dense part of scalapackto extend sequential BLAS interface to distributed memory interface. The proposal specifically allows for implementations that can be used in a heterogeneous environment.
As a side note, a discussion began on the increased use of Fortran90, and the need for a possible "F90 Conventions" section with examples in the Introduction chapter (Chapter 1). It was then suggested that it would be a good idea to include a section in Chapter 1 outlining the programming conventions used in all language interfaces -- Fortran77, Fortran90, C, and C++. These summaries would explain the features of each language that are utilized, and if there are assumptions made, these assumptions would be stated here. In this way, the user would be better prepared for the language binding sections of the document.
At 11:30am, Tony Skjellum gave an overview of comments on the formation of the BLAS Lite document. He expected that the BLAS Lite chapter be in final form by the next meeting. He stressed the need for a fuller API into the chapter format, test for coverage, show how to do traditional BLAS in terms of the Lite/Thin BLAS, and offer a model implementation for the PPro or SPARC. The chapter will be kept in the Journal of Development due to the controversial nature of the Lite/Tiny BLAS. The Lite/Tiny BLAS are not for intended for every level of user. They are designed to get people into the BLAS mode who are now adverse to any library. Their existence should reflect what people who develop high performance BLAS actually do. They currently address vector and RISC architectures, and it is anticipated that they include programmable cache and emerging PIM. L/T is not necessarily implemented as a library -- inline assembly language, preprocessing, compiler support, etc. The term "thin" refers to no "or"s in the code, and "lite" refers to the fact that the subroutines operate on fixed block sizes. He then listed some potential environmental inquiry routines.
In summary, the three components of the L/T are: environmental routines, routines to manipulate blocks, and computational routines which operate on fixed block sizes. The Lite/Thin BLAS are now officially known as the Basic Linear Algebra Instruction Set (BLAIS).
Lunch break...
After lunch, Clint Whaley of UT presented the summary of the C interface to the BLAS subgroup, and initiated a second reading of the document. The main issue discussed was the tester. Should we have a formal definition of the tester in the BLAS document? A starting point for the C-BLAS tester was obtained from D. Manley at Digital Corporation. Theresa Do of Cray and Bruce Greer of Intel volunteered to help coordinate the tester. The second issue that was addressed was the location of the Legacy BLAS chapter. Should it appear as an appendix or as a chapter, appearing before the Journal of Development? The majority felt that the Legacy BLAS chapter should be separate and not subject to the format specifications of the main chapters. Thus, it will remain an appendix.
Voting next occurred on the modified sections of the Legacy BLAS chapter/appendix related to the C interface to the existing BLAS. There were 8 eligible voters: UT, NAG, Cray, NIST, NEC, HP/Convex, Intel, and Bell Labs.
Coffee break at 3pm.
At 3pm we began discussion of Language Issues, Data-Neutral Interface, Environmental Enquiries, and Name Space.
Jim Demmel of UC Berkeley began discussion of the Language Issues. There was general agreement that C++ and F90 should be supported. Concern was raised that data types would be specified that would not be present in certain languages, such as C and F77. Support for Java also was viewed favorably with a suggestion by Tony Skjellum that Java BLAS might be appropriate for the JOD (pending a concrete proposal). There was discussion of data neutral interfaces. At the last BLAS meeting, Andrew Lumsdasine and Tony Skjellum suggested use of data neutral interfaces to hide format, precision, etc. Several participants expressed interest in handle based operations. Iain Duff pointed out that this is already done in sparse BLAS proposal. Some outstanding issues concern translation to and from internal format, handles as an expressive abstraction, and data format hiding for performance. It was suggested that there is a role for both data neutral interfaces and data format explicit interfaces. Concern was expressed that users should be able to access lower level routines. Finally, the mechanics of language specific bindings was discussed.
The conclusion of the Language Issues discussion was that Jim will work on a Fortran90 interface, with a Fortran77 interface to a subset of the routines, as well as C and C++ interface.
Tony Skjellum led a discussion about environmental inquiry functionality. There are two primary areas of the forum concerned with environmental inquiry -- BLAIS and extended precision BLAS. It was agreed that Tony will lead the task of comprising the list of needed routines, with help from Jack Dongarra of UT.
Antoine Petitet had proposed that we next discuss Namespace issues.
A coffee break followed, and then we broke into the following break-out sessions: Functionality/Extensions/Dense BLAS, Sparse, and BLAIS.
The meeting adjourned at 6:00pm.
The meeting began at 9:00am with break-out sessions -- Functionality/Dense, Sparse, Extra Precise, Distributed-Memory, and C/F90 interfaces.
At 10:30am, the subgroups presented their progress.
The Functionality subgroup consolidated new entries for tables in Chapter 1 and Chapter 2. It was reiterated that any entry in a table in Chapter 1 must be mentioned somewhere in the document. If an entry has no use in defining the functionality of any of the chapters, then it must be removed from the tables of Chapter 1. Gary Howell of Florida Tech proposed the addition of elementary similarity transformations to the functionality tables, and volunteered to help with their inclusion. Due to the possible instability of such routines, it was felt that further discussion of their inclusion should be addressed via the mailing list. It was also asked if the functionality of the routines will be 0-based or 1-based. This offset from the beginning element is controversial but must be resolved. The first reading of Chapter 2 will occur at the next meeting (April, 1998).
Roldan Pozo of NIST and Mike Heroux of Cray then summarized the findings of the Sparse BLAS subgroup. The proposed functionality of the Sparse BLAS has been reduced to exclude external permutations. Integrity checks will be limited, and conversions will be limited. The subset of functionality will concentrate on matrix multiply and triangular solve. The example bindings in the draft will be Fortran90 and C++. A functionality table needs to be included in the Sparse chapter, and this functionality also reflected in the overall functionality of Chapter 1.
Roldan Pozo of NIST then asked the question "Should the Level 1 Sparse BLAS be included in the proposal for the sake of completeness?" He asked for a straw vote on this idea. No actual show of hands was taken but the majority felt that the Level 1 Sparse BLAS should not be mentioned in this document, mainly for fear of conflicting with the "standard" already defined for these routines.
Sven Hammarling of NAG then briefly spoke on the proposed Fortran90 interface to the BLAS. The reference implementation would accomodate the Legacy --> Thin --> Thick interfaces. The "thick" interface refers to one with error-checking. He urged the need for comments on the draft of the proposal included in the Legacy BLAS chapter of the document. It was reiterated that a "Conventions" section for Fortran90, as well as for Fortran77, C, and C++, will be added to Chapter 1 in order to better explain the features of each language that are utilized in the respective language bindings. A first reading of this chapter was proposed for the next meeting.
Jim Demmel of UCB then summarized the Precise BLAS/Extended Precision BLAS (XBLAS) proposal. The reference implementation will follow the current design of the Fortran90 interface. He then discussed specific language bindings for the routines in Fortran90, Fortran77, and C. A first reading of this chapter was proposed for the next meeting.
Antoine Petitet of NEC summarized the findings of the distributed-memory BLAS. He stressed the need to define the functionality so that a subset can be extracted. Currently, language bindings for C and Fortran77 are proposed. Concern was raised over a Fortran90 interface if this is not supported in MPI. Some of the attendees responded that a Fortran90 interface is now supported in MPI (as this is used by Visual Numerics).
In preparation for the first readings, each of these chapters should have a prioritized list of the functionality. Concern was then raised on whether the reference implementation should encompass all proposed routines or only the prioritized list? And should a reference implementation be provided for each language binding? The general concensus was that the reference implementation should be included for each language supported.
This "prioritizing" discussion also brought up the danger of defining a standard with optional sections. Participants fear proposing a standard that is not clearly-defined and has optional sections. The example cited was the HPF effort, and the fact that no two HPF compilers have the same subset of HPF directives.
The forum participants prefer to have one clearly-defined standard with no optional sections.
Jack Dongarra of UT then addressed the closing discussion for the meeting and reviewed the tentative dates for the next two meetings. He stressed that the forum is responsible for specifying the semantics and syntax of the BLAS.
The tentative date of the next forum meeting is:
and will be hosted by NIST in Washington, D.C.. The exact date of the meeting will be provided soon. First readings of the following documents will occur at this meeting
A second reading of the "C interface to the BLAS" in the Legacy BLAS chapter will also occur at the April 1998 meeting.
A preliminary deadline of mid-March 1998 is set for subgroup progress.
The next meeting would then be tentatively scheduled for August, 1998 at CRI in Eagan, MN.
The meeting was then adjourned by Jack Dongarra at 12Noon.
Attendees list for the December 3-5, 1997 BLAST Forum Meeting
Susan Blackford UT, Knoxville susan@cs.utk.edu Clay Breshears CEWES/Rice U. clay@turing.wes.hpc.mil Jim Demmel UC Berkeley demmel@cs.berkeley.edu Theresa Do SGI/Cray tdo@cray.com Jack Dongarra UT / ORNL dongarra@cs.utk.edu Iain Duff RAL/CERFACS isd@rl.ac.uk Cormac Garvey NEC garvey@atcc.necsyl.com Bruce Greer Intel bruce_s_greer@ccm.jf.intel.com Sven Hammarling NAG, UK sven@nag.co.uk Greg Henry Intel ghenry@cs.utk.edu Mike Heroux SGI/Cray mike.heroux@cray.com Gary Howell Florida Tech howell@zach.fit.edu Linda Kaufman Bell Labs lck@bell-labs.com Hsin-Ying Lin HP Convex Tech. Ctr. lin@rsn.hp.com Andrew Lumsdaine Univ. of Notre Dame Lumsdaine.1@nd.edu Antoine Petitet NEC petitet@ccrl-nece.technopark.gmd.de Roldan Pozo NIST pozo@nist.gov Tony Skjellum Miss. State Univ. tony@cs.msstate.edu Francoise Tisseur UT, Knoxville ftisseur@cs.utk.edu Clint Whaley UT, Knoxville rwhaley@cs.utk.edu
Susan Blackford and Andrew Lumsdaine agreed to take minutes for the meetings.