April 27-29, 1998

Cray Research Inc., Eagan, MN

April 27, 1998

Jack opened the meeting, and everyone introduced themselves. He then turned over the floor to Mike Heroux, who outlined the schedule and the meeting of subgroups. Jack then outlined the tentative agenda for organizing the breakout sessions for the morning and the group reports of sessions in the afternoon. On Tuesday morning, several attendees will be addressing the forum with issues that they feel have not been addressed yet. And on Tuesday afternoon, we will start the reading of chapters. The BLAS Extensions discussion was moved to Monday afternoon. On Wednesday, more breakout sessions and group reports, and the meeting will end around lunchtime.

Jack then reviewed the rules of order, and the goals of the forum. The main goals of the forum are to formulate ideas for the future in light of modern technology development, and to produce a document of standards to enable linear algebra to interoperate efficiently and easily. Iterative and sparse direct methods require additional functionality not in traditional BLAS. The outcome of the forum should be a document basis for the standards of implementations. -- a piece of paper not a piece of software. We will try to provide reference implementations, but the goal is not to produce software. He then outlined the number of BLAS meetings that have occurred, an overview of the original subgroups, and talked about the organization of the document. We are getting close to finishing. Two more meeting were proposed, and then the Forum will conclude with a report of our findings. We are at the end of NSF funding for travel, there is only funding for partial support for this meetings travel. We need to decide on where we should publish the document. The proposed sites for the final two meetings are Washington D.C hosted by NIST, and Starksville, Mississippi hosted by MSU.

We then discussed breaking into sessions. The morning session will encompass breakout groups for the Dense and Band chapter, the Sparse chapter, Language Bindings, and Extended Precision. And later the C interface to the BLAS, the Fortran95 interface to the BLAS, the Interval BLAS, and possibly the Distributed Memory BLAS, and BLAS Extensions, will meet.

Lunch was at noon, and a tour of the Graphics Lab at 1-1:30pm. We then continued group discussions from 1:30-2:30pm.

Group reports began at 2:30pm. Presentations were made for Dense/Band BLAS, the C interface for the BLAS, the F95 interface for the BLAS, Extended Precision, Interval Arithmetic, Language Bindings, and Sparse BLAS.

Sven Hammarling of NAG began with the summary of the Dense/Band chapter. The tables for functionality had been updated and Antoine Petitet of UT presented his Language Independent Specification (LIS) section and the proposed modifications suggested by the subgroup. Debate ensued on how precise the LIS interface should be, and what will be addressed in the Language Dependent Specification (LDS) to be derived from the LIS section. Jerzy Wasniewski of UNI-C proposed the addition of an INFO parameter which should be declared as INOUT. A lively discussion began on the issue of error-checking in this document and how much should not be specified and left to the implementation of the vendor. A straw vote was taken on the proposed LIS interface. The basic forms were:

  1. BLAS_DOT( n, x, incx, y, incy, r )
  2. BLAS_DOT( x, y, r)
  3. BLAS_DOT( n, x, y, ix, iy, r)
  4. BLAS_DOT(n, x, ix, incx, y, iy, incy, r)

A total of twenty-one people were present for the straw vote. One person voted for option (1), nine people voted for option (2), zero people voted for option (3), and eleven people voted for option (4). A straw vote was then taken on the allowance of INCX=0 in the routines. Eleven people voted yes, eight people voted no, and two people abstained. Then we started discussing the need for a straw vote on the inclusion of INFO in the LIS specification, and how this INFO would be used, in an environmental routine perhaps or directly required in the calling sequence (in the case of a Fortran77 interface). Error-checking was in strong debate and how to provide a ''light'' interface to allow for no checking if desired. The attendees were very divided on this issue. As a result, the straw vote was to be postponed until Tuesday morning.

Next Clint Whaley of UT presented briefly the C interface, and the issue of error-handling was again addressed. What should be said in the document about error-handling? Should we be precise or sufficiently vague to allow the vendor freedom? It was decided to be vague and the last sentence of section B.2.8 was deleted.

Roldan Pozo of NIST next presented the Sparse BLAS summary. Mike Heroux is adding an LIS specification to the chapter, a system handle-based representation is also being added. Linda Kaufman of Bell Labs addressed the issue of the need for gather and scatter specifications in the Sparse chapter. The jagged and ellpack storage formats were removed from the document. Linda Kaufman of Bell Labs proposed that we need a conversion ability from one storage format to another. In the Sparse BLAS chapter, only two operations are supported -- sparse matrix times a dense vector, and a triangular solve.

Sven Hammarling of NAG briefly addressed the Fortran95 interface to the Legacy BLAS.

The meeting adjourned for the day at 5:30pm.

April 28, 1998

Tuesday began at 9:00am with the conclusion of the LIS discussion for the Dense and Band Chapter, the Extended Precision summary, Interval Arithmetic, and Language Bindings. At 11:00am, there were presentations by Ed Anderson of Cray to address an issue about the Distributed Memory BLAS, Bruce Greer of Intel to talk about 64-bit architectures, Claus Bendtsen of UNI-C to talk about the C++ interface to the BLAS, Gary Howell to discuss his Level 2.5 BLAS proposals, Roldan Pozo to talk about the Java BLAS, Clint Whaley to talk about the ATLAS BLAS, and Jack Dongarra to discuss the need for Environmental routines. At 12Noon we will have lunch and then a tour of the computer room at 12:45--1:30pm.

First, Antoine Petitet of UT addressed the need to take a straw vote on the INFO parameter issue in the LIS for the Dense and Band chapter. Option (1) was that INFO does not appear in the LIS and may be accommodated through the use of an environmental routine or such mechanism. Option (2) says that INFO appears in the LIS specification of the routine. As discussion and confusion ensued, it was then pointed out that we were really addressing the question of: Should we have recoverable errors or not? Thirteen people voted for option (1), and five people voted for option (2).

And then we needed to decide on the matrix notation and multi-vector notation in LIS for the Dense and Band chapter. The following matrix notation was suggested:

        BLAS_xxx( m, n, A, ia, ja )

It was asked if we should include a stride (LDA-like) in the LIS? The majority felt that this should be specified in the LDS. A vote was taken on this proposed LIS, and one person was against, and thirteen people were in favor. A mathematical formula should be included for each routine's LIS.

Next, we voted on the LIS description for a multi-vector.

        BLAS_xxx( n, k, x, ix, incx )

Where ix is an array of size ix(k), and incx is an integer. It was decided that a stride (ix) should be allowed for each vector in the collection. A straw vote was taken and sixteen were in favor of this LIS, and two people were against.

Next, Linda Kaufman of Bell Labs presented the BLAS Extensions ideas. She wanted to stress the need for a C interface for the extensions, as well as the Legacy BLAS. She wanted an order indicating the priority for the extensions. The criterion for a BLAS routine was discussed. It must be frequently used and need to be highly optimized, and there could be a third requirement that it is difficult to write in a higher level language. She then addressed the issue of sorting and should we have a sorting parameter or a different routine name, and should it be active, passive, or active and passive. We should take a look at the LAPACK calling sequence for xLASRT. The LAPACK routine xLASRT has a character parameter to specify an increasing or decreasing sort, and the array is sorted on output from the routine. She concluded with a reiteration of the need for scatter and gather to be specified in the Sparse chapter for the Level 1 Sparse BLAS.

At 11am, Jim Demmel of the University of California, Berkeley presented a summary of the Extended Precision proposal. He presented some performance figures for some example code on various platforms. His issues for managing complexity/utility in the Extra Precision BLAS are:

  1. No new ``first class'' data types
  2. Orthogonal to rest of BLAS design
  3. PREC describes internal precision only, as runtime parameter
  4. Include a modest amount of mixed precision?
  5. Shall versus should?
  6. Environmental inquiries?

He presented a list of minimal versus maximal number of routines to include in the proposal:

  1. DOT
  2. GEMM SYRK GEMV TRSV, and other MM, MV, or SV routines
  3. all Level 2 and 3 BLAS,
  4. mixed precision
  5. BLAS extensions
  6. Sparse, Distributed Memory, or Interval routines?

Where should we draw the line on how many routines to include? Reference implementation and test code will be provided. Regarding shall versus should, Jim proposed that we supply a reference implementation for everything except the BLAS extensions and Sparse Distributed and Interval.

At 1:45pm, Chenyi presented the Interval BLAS. Only the Level 1 Interval BLAS are included in the proposal. He briefly described the language binding issues and naming scheme. Should function names be prefixed by INTERVAL, INT, or ITV, or I, or what? He will distribute a reference implementation. The functionality for the Interval BLAS need to be included in the functionality tables of Chapter 1 of the BLAS document. Should the set operation routines also be included in the Functionality tables?

Antoine Petitet of UT did not present for the distributed memory BLAS subgroup since there was not time for the group to meet, and many of the issues in the Dense and Band chapter must be resolved before they can be written for the Distributed Memory chapter. And likewise, the language bindings did not wish to address the group.

At 2:15pm, we began a series of brief 10 minute presentations. Ed Anderson spoke briefly on the continuity between dense and sparse BLAS, and proposed a different data distribution to be specified. A block major format whereby contiguous nbxnb blocks are stored contiguously in memory. Clint Whaley of UT pointed out that this is the same type of distribution that is used in ATLAS BLAS, and could be accommodated through the descriptor in ScaLAPACK. Bruce Greer of Intel next gave a presentation on 64-bit arithmetic. Claus Bendtsen of UNI-C proposed a C++ interface to the BLAS. It was decided to have a subgroup meeting the next morning to discuss this C++ implementation. Gary Howell of Florida Tech next presented the BLAS 2.5. Roldan Pozo of NIST presented a proposal for the Java BLAS. Clint Whaley of UT presented the ATLAS BLAS. Jack Dongarra of UT/ORNL then spoke about Environmental Inquiry functions. The question posed was: Should we compile a list of environmental routines to query for machine characteristics such as the number of floating point registers, number of floating point units, number of caches, cache size, type of cache, cache line size, cycle time, page size, size of TLB, cycles for floating point operations, number of processors, fused multiply/add, cycles from memory to stages in the cache, pre-fetch, bandwidth to/from memory, latency from memory/cache? The answer was affirmative.

A short break was taken at 4:30pm, and then Clint Whaley of UT presented the third reading of the C interface to the Legacy BLAS. There were fourteen eligible voters, which included Univ. of Tennessee, NAG, UNI-C, Univ. of Notre Dame, Intel, HP/Convex, Cray, Bell Labs, Univ. of California at Berkeley, Florida Tech, Univ. of Houston, Tera, NEC, and NIST. The votes were as follows:

The meeting adjourned at 6:15pm.

April 29, 1998

The first order of business was which subgroups should meet prior to continuing the readings and the date for the next Forum meeting. We need to have readings at the next meeting of the Dense chapter, Sparse chapter, Language Bindings, Distributed Memory chapter, Extended Precision chapter, and the Interval chapter. It is possible that the Interval BLAS and Distributed Memory BLAS will be put into the Journal of Development.

It was decided that the next meeting would occur sometime in the first week of October, and the final meeting would occur sometime in December. And should the next meeting be at NIST or MSU? It was reiterated that we need to finish these meetings by the end of the year. It was then suggested by Roldan Pozo of NIST that we have a ``virtual'' meeting in July since there will be so much time between this meeting and the next meeting. Perhaps we could do a ``first reading'' of some of the chapters on-line? A unanimous vote was taken on the move to hold a ``virtual'' meeting. It was then decided that all chapters would be due by July 1, a reading and commenting period of two weeks would occur from July 1-14, and then the authors of the chapter would incorporate all comments for one week, July 14-21, and voting on the chapters would occur between July 21-August 4. The number of eligible voters needs to be determined, and a majority of votes from eligible voters should be received for each chapter. Otherwise, the vote would not be valid. The chair for each subgroup can ask for straw votes on topics, and should assemble a ballot for voting. The ballot should list each section of the chapter, and give the voter space to vote in favor or against, and allow for comments. After the voting period, the chair from each subgroup will post the results to blast-comm.

A straw vote was taken on the location of the next meeting. Seventeen people voted for Washington D.C., and there were four abstentions. Three people were opposed to holding the meeting over the weekend. We will plan to hold the next meeting in Washington D. C. around the first week of October.

And at 9:30-10:30am, subgroups assembled to discuss the C++ interface, C interface, Dense and Band BLAS. After these subgroup meeting we finished the reading of the C interface section of the Legacy BLAS chapter, and then proceeded with the first reading of the Fortran95 interface section of the chapter.

A general questions about language bindings was addressed. It was proposed that the storage format for a dense matrix be defined as a multi-vector. The question was then asked how this would impact the Fortran77 interface? The consensus was that it was fine to define a matrix as a multi-vector in the LIS, and that how this was handled in the LDS was left up to the implementor. Thus, the Fortran77 interface is not impacted.

Next we resumed the reading of the C interface to the BLAS. Thirteen eligible voters were in attendance. The subgroup meeting discussed the cblas.h include file issues. So the second reading of that section was deemed ready for a reading.

Sven Hammarling of NAG next addressed the first reading of the Fortran95 interface to the BLAS. Modifications were suggested for each section, but general votes were taken.

It was suggested that we add the addition of CROT and CROTG to the Extensions section of the Legacy BLAS chapter. And then provide an interface for this ``extension'' routine. The same would be done in the C interface section of the chapter. A straw vote was taken to include these as extensions. Thirteen people voted in favor, and no one opposed or abstained.

A tentative straw vote was taken on the namespace issue of f95blas as prefix. Fourteen people voted for the prefix, one person opposed, and one person abstained.

A straw vote was then proposed by Jim Demmel. He proposed a set of language bindings for the extended precision routines. For Fortran77 and C, add _X, and PREC as final argument. Twelve people voted in favor of this language binding. There were no abstentions or oppositions.

Bruce Greer of Intel then suggested a round of applause to thank Cray Research and Mike Heroux for hosting the meeting, and Sven Hammarling officially closed the meeting at 1:30pm.

List of attendees:

Attendees list for the April 27-29, 1998 BLAST Forum Meeting

Ed Anderson          SGI/Cray   
Claus Bendtsen       UNI-C, Denmark
Susan Blackford      UT, Knoxville
Andrew Chapman       NEC        
Jim Demmel           UC Berkeley
Theresa Do           SGI/Cray   
Jack Dongarra        UT / ORNL  
Bruce Greer          Intel      
Sven Hammarling      NAG, UK    
Mike Heroux          SGI/Cray   
Gary Howell          Florida Tech
Chenyi Hu            Univ of Houston
Guangye Li           SGI/Cray   
Hsin-Ying Lin        HP Convex Tech. Ctr.
Linda Kaufman        Bell Labs  
Kristi Maschhoff     Tera Computer
Brian McCandless     SLAC/Notre Dame
Antoine Petitet      UT, Knoxville
Roldan Pozo          NIST       
Karin Remington      NIST       
Shane Story          Intel      
Jerzy Wasniewski     UNI-C, Denmark
Clint Whaley         UT, Knoxville

Susan Blackford and Brian McCandless agreed to take minutes for the meeting.