ARSC T3D Users' Newsletter 49, August 25, 1995
Programming Environment 1.2.2 Now Default
As of 8/23/95, ARSC is running the 1.2.2 Programming Environment as the default. If you have any problems with this release on the T3D contact Mike Ess, email@example.com. This release has many new features and some of them were described in the CRI release letter that appeared in T3D newsletter #47 (8/11/95). Below is an additional description of what's in the 1.2.2 release that appeared in the Cray Research Service Bulletin:
> > Systems: PVP, MPP > OS: UNICOS 8.0, UNICOS 9.0, > UNICOS MAX 220.127.116.11 or later > Product: CF77 6.2.2 PE, > CF77 6.2.2 PE for MPP systems, > CF90 1.2.2 PE, > CF90 1.2.2 PE for MPP systems, > Cray Standard C 1.2.2 PE, > Cray Standard C 1.2.2 PE for MPP systems > Audience: All > Date: July 1995 > > The following Cray Programming Environments were released in June: > > CF90 1.2.2 Programming Environment for both Cray PVP and Cray MPP > systems > Cray Standard C 1.2.2 Programming Environment for both Cray PVP and > Cray MPP systems > CF77 6.2.2 Programming Environment for Cray PVP and Cray MPP systems > > The programming environments provide an integrated development environment > and combine compilers with development tools and libraries. Each of these > programming environments includes the CrayTools 1.3.2 package, the > CrayLibs 1.2.2 package for either Cray PVP or Cray MPP systems, and one of > the following compilers: > > CF90 1.0.3 > CF90_M 0.1.2 > SCC 4.0.3 > SCC_M 4.0.4 > CF77 6.0.4 > CF77_M 6.2.1 > > You cannot order programming environment components separately; you must > order a complete programming environment. To order a programming environment, > contact your Cray Research service representative or complete a software > order form in the CRInform service. > > Programming environment components discussed in this article are supported on > Cray PVP systems that run the UNICOS 8.0 release or later and on Cray MPP > systems that run both the UNICOS MAX 18.104.22.168 release or later and the > UNICOS 8.0 release or later. > > Note: You must install the CF77_M 6.2.1 compiling system and CrayLibs_M 1.2.2 > at the same time because dependencies exist between these products in the > following shared array intrinsics, which are introduced for the first time in > this release: > > PRESUM, PREPROD, PREMIN, and PREMAX parallel prefix operations > SCANNSUM, SCANPROD, SCANMIN, and SCANMAX segmented scan operations > > If programs that use these array intrinsics are compiled with CF77_M 6.2.1, > you also must use CrayLibs_M 1.2.2 or later. > > Enhancements > > For both PVP and MPP systems, the libsci LAPACK libraries contain additional > support for the following (LAPACK is a software package for solving linear > systems and matrix eigenvalue problems): > > Factorization variant QR with pivoting > Linear least squares driver routines > Reductions to condensed form for eigenvalue computations > > The libraries also contain performance improvements to xLARTG and xLARFG, the > auxiliary routines that generate elementary transformations. > > For MPP systems, the programming environment 1.2.2 releases contain the > following enhancements: > > Performance in the CF77_M compiling system was enhanced by adding > optimizations that improve cache performance for stack data. > Performance was enhanced for all MPP 1.2.2 programming environments > through optimizations that streamline inner loop computations. > The SHMEM_MY_PE() library routine, which returns the number of the > processor elements (PEs) executing, and the SHMEM_N_PES() library > routine, which returns the number of PEs available to the program, were > added. These routines are especially useful with the CF90 compiler, > which does not have intrinsics for this purpose. > Performance was improved in the SHMEM_GET library routine when transfer > sizes of one to three words are achieved. > Performance was improved for the Parallel Virtual Machine (PVM) global > operation function, pvm_reduce(). Previously, pvm_reduce() was layered > on the pvm_psend() and pvm_precv() functions, and it used the same > linear algorithm as is used in the public domain version of PVM. The > optimized version uses more direct shared memory and achieves more > logarithmic scaling. > The following Fortran shared array intrinsics were added: > PRESUM, PREPROD, PREMIN, and PREMAX parallel prefix operations > SCANNSUM, SCANPROD, SCANMIN, and SCANMAX segmented scan operations > > These intrinsic functions behave similarly to reduction functions, but > partial results are retained in the result array. The segmented scan > intrinsics also carry a mask of stop bits that tells the operation to stop > and start over again. > > Enhancements to the libsci libraries for Cray MPP systems include the > following: > > The addition of an optimized implementation of ScaLAPACK, a set of > scalable algorithms for solving linear algebra problems. ScaLAPACK > contains several linear solvers, similarity and equivalence > reduction routines, and an Eigensystem solver for real symmetric > matrices. These algorithms use the Basic Linear Algebra Communication > Subprograms (BLACS) for communication and the Parallel Basic Linear > Algebraic > Subroutines (PBLAS) for computation. The PBLAS, new in this release, > provide distributed-memory software comparable to the BLAS in > shared-memory codes. > New Fast Fourier Transform (FFT) routines were added to perform > parallel 2-D and 3-D real-to-complex and complex-to-real FFTs. > The performance of the FFT subroutine, CCFFT, was improved for > prime-length vectors and for lengths that cannot be factored into > powers of 2, 3, and 5. The single-PE routines > SCFFT and CSFFT have been extended to work with odd-length vectors. > Newly optimized versions of the following routines were added: > SNRM2, SCNRM2, SROTG, and CROTG from the Level 1 BLAS routines and > STRSM, STRMM, and the other Level 3 BLAS routines. > > Compatibilities issues > > The user interface for loop unrolling in the CF77_M compiling system has > changed. In previous releases, when a user specified the following, a total > of 5 copies of the loop body were created; the original loop plus 4: > > CDIR$ UNROLL  > > This implementation was not compatible with the FPP implementation of the > directive and was confusing to users. With this release, the previous example > results in a total of 4 copies of the loop body; the original plus 3. > > Documentation changes > > In addition to numerous man pages updated for this release, the following > manuals have also been updated: > > Cray MPP Fortran Reference Manual, publication SR-2504 6.2.2 (available > online in Docview form only) > Programming Environment Installation Guide, publication SG-5217 1.2.2 > (available in printed form only)
One Year Anniversary of the ARSC T3D NewsletterThe first ARSC T3D Newsletter was mailed out on 8/25/94 and since then there have been 49 issues (almost weekly). I thought about providing an index but that would be a lot of work and would be incomplete no matter how much effort was put into it.
What has been useful for me is to accumulate all of the newsletters in a single directory (~.5MB) and when a query comes up I just grep for the keywords. With grep there is no need for an index.
PATP Meeting at JPLThis week Ming Jiang and I will be at the PATP Scientific Conference held at JPL, August 24th and 25th. I'll have details of that meeting in future newsletters.
List of Differences Between T3D and Y-MPThe current list of differences between the T3D and the Y-MP is:
- Data type sizes are not the same (Newsletter #5)
- Uninitialized variables are different (Newsletter #6)
- The effect of the -a static compiler switch (Newsletter #7)
- There is no GETENV on the T3D (Newsletter #8)
- Missing routine SMACH on T3D (Newsletter #9)
- Different Arithmetics (Newsletter #9)
- Different clock granularities for gettimeofday (Newsletter #11)
- Restrictions on record length for direct I/O files (Newsletter #19)
- Implied DO loop is not "vectorized" on the T3D (Newsletter #20)
- Missing Linpack and Eispack routines in libsci (Newsletter #25)
- F90 manual for Y-MP, no manual for T3D (Newsletter #31)
- RANF() and its manpage differ between machines (Newsletter #37)
- CRAY2IEG is available only on the Y-MP (Newsletter #40)
- Missing sort routines on the T3D (Newsletter #41)
Ed Kornkven ARSC HPC Specialist ph: 907-450-8669 Kate Hedstrom ARSC Oceanographic Specialist ph: 907-450-8678 Arctic Region Supercomputing Center University of Alaska Fairbanks PO Box 756020 Fairbanks AK 99775-6020
Subscribe to (or unsubscribe from) the e-mail edition of the
ARSC HPC Users' Newsletter.
Back issues of the ASCII e-mail edition of the ARSC T3D/T3E/HPC Users' Newsletter are available by request. Please contact the editors.