DEC to support REAL*16 IEEE arithmetic

[David Hough]

DEC will be supporting the same (except for endian-ness) format for quad
as SPARC and PA-RISC, implemented in the same unfortunate way, by software.

It would be very interesting if one of the chip vendors finally implemented
quad in hardware.   But all are struggling to maximize benchmarkable performance
within the same technology constraints, so it seems unlikely in the immediate
future.   If the advance in semiconductor technology slows down enough,
added value features like this may become more important.

I'm sure it's gratifying for the Fortran groups at Sun and elsewhere that
after years of trying to keep up with VAX/VMS Fortran, the shoe is now on the
other foot.

   Newsgroups: comp.lang.fortran,comp.os.vms,comp.unix.osf.osf1,comp.sys.dec
   From: lionelaquark.enet.dec.com (Steve Lionel)
   Organization: Digital Equipment Corporation
   Date: Wed, 11 Aug 1993 01:42:12 GMT

      REAL*16 White Paper
      A Specification for Customers


      This document provides an overview of Digital's plans for REAL*16 func-
      tionality for Fortran customers on DEC OSF/1 AXP and OpenVMS AXP.


      Author:                       Beth Benoit

      Date:                         20 July 1993




      Digital Equipment Corporation
      Maynard, Massachusetts


      1  Introduction

      This document provides a high-level functional description of Digi-
      tal's plans to support IEEE extended 128-bit floating point, imple-
      mented in Fortran as REAL*16. This support will be forthcoming on both
      DEC OSF/1 AXP and OpenVMS AXP platforms before July 1994, and on Win-
      dows NT AXP some time after that.

      2  Overview

      A number of computer vendors (Digital, CRAY, IBM, and HP) have, over
      the years, offered an extended precision floating point datatype, rep-
      resented in Fortran as REAL*16. Many people in the Fortran user-community
      have come to rely on the accuracy of this datatype in their applica-
      tions. For these customers, the current lack of REAL*16 support in Dig-
      ital's AXP Fortran development environment is an impediment that keeps
      them from considering AXP when purchasing a new system. Digital rec-
      ognizes this, and is in the process of implementing support for this
      datatype.

      The requirement for 128-bit extended precision floating point datatype
      has been expressed by both DEC OSF/1 and OpenVMS Fortran customers.
      Digital expects Windows NT customers to have a similar requirement.
      Digital's customers have also made it clear that there is a require-
      ment for this support to be IEEE-conformant. For this reason, Digi-
      tal is choosing a data representation based on the IEEE standard, rather
      than the data representation used on VAX/VMS known as H-float.

      3  Operations on REAL*16

      The best way to describe which operations on REAL*16 will be allowed
      is to say that all the operations and intrinsics supported for REAL*16
      by VAX FORTRAN today will be available on Digital's AXP platforms to-
      morrow. For those who want more detail, this is described in Digital's
      DEC_Fortran_Language_Reference_Manual, which describes most of Dig-
      ital's Fortran products.

      While DEC Fortran might also support a COMPLEX*32 datatype, consist-
      ing of two REAL*16 datatypes, Digital has no plans to implement COM-
      PLEX*32 support. So far, it seems there are fewer applications that
      rely on this datatype. While the IBM RS6000, /370, and 3090 systems
      support COMPLEX*32, systems from Hewlett Packard, Sun, and Digital have
      not supported COMPLEX*32 in the past.

      On OpenVMS AXP, DEC Fortran plans to provide "foreign file support"
      in the I/O runtime library, which will enable conversion between the
      VAX H-float representation of REAL*16 and the AXP representation of
      REAL*16 for unformatted binary files. This is particularly useful for
      customers who are sharing files between both VAX and AXP OpenVMS sys-
      tems.

      [DEC Fortran for OpenVMS VAX will also provide the "foreign file
       support for conversion between VAX H-float and the AXP REAL*16
       type in its V6.1 release, due later this year. - SBL]

      4  REAL*16-the Bits and Bytes

      The ALPHA AXP hardware does not support an extended precision float-
      ing point datatype. In order to provide compiler code generation for
      the Fortran REAL*16 datatype, it is necessary to emulate this support
      in software.

      Digital's AXP implementation of the REAL*16 datatype will be 128-bit,
      with a 15-bit exponent and 113-bit mantissa. The memory layout of an
      REAL*16 on AXP will be little-endian with the sign and exponent in the
      high quadword. The register format will also be little-endian, with
      the sign and exponent in the higher numbered (odd) register of a reg-
      ister pair. The implementation of REAL*16 keeps 128-bit floating val-
      ues in even-odd pairs of fixed point registers.

      5  Performance

      Since Digital's 128-bit floating point datatype will be software-emulated
      on AXP, the performance of mathematical operations on this data-type
      will be slower than operations on 64-bit floating point numbers. This
      is also true on other systems available today; for instance, the Hewlett
      Packard PA-RISC series also supports REAL*16 in software emulation,
      as do most recent VAX-series processors from Digital. Digital expects
      the performance of REAL*16 to be competitive.

      6  Packaging and Delivery

      The addition of a new floating point datatype requires changes to a
      number of pieces of a software development system. Digital's deliv-
      ery of a software development environment which fully supports the REAL*16
      datatype will require updates of the operating system, the Fortran com-
      piler, and on OSF/1 the developer's kit which contains the dbx debug-
      ger. The I/O and math runtime libraries which are bundled with the op-
      erating system have the longest lead time, and the most restrictive
      schedules.

      Digital's policy is to not announce availability dates until the func-
      tionality is actually available. While it's too soon to announce dates,
      Digital engineering has development and testing of REAL*16 underway
      now. Based on the DEC OSF/1 and OpenVMS operating system schedules as
      they exist today, we believe REAL*16 support will be in customer hands
      before July 1994.

      7  Comments

      If you have comments about this white paper, Digital Engineering would
      like to hear from you. You can address comments to either:


      Catherine Epperson               Beth Benoit
      Fortran Product Manager          Fortran Engineering Manager
      Digital Equipment Corporation    Digital Equipment Corporation
      ZKO2-1/M11                       ZKO2-3/N30
      110 Spit Brook Road              110 Spit Brook Road
      Nashua, NH 03062-2698            Nashua, NH 03062-2698
      eppersonawereok.enet.dec.com     benoitatle.enet.dec.com

   --
   Steve Lionel                           lionelaquark.enet.dec.com
   SDT Languages Group
   Digital Equipment Corporation
   110 Spit Brook Road
   Nashua, NH 03062                       "Free advice is worth every cent"