168.wupwise
SPEC CPU2000 Benchmark Description File
Benchmark Name
168.wupwise
Benchmark Authors
Bjoern Medeke ([email protected])
Institute of Applied Computer Science
University of Wuppertal
Gauss-Strasse 20
D-42097 Wuppertal
Germany
Benchmark Program General Category
Physics / Quantum Chromodynamics
Benchmark Description
"wupwise" is an acronym for "Wuppertal Wilson Fermion
Solver", a program in the area of lattice gauge theory (quantum
chromodynamics).
Lattice gauge theory is a discretization of quantum chromodynamics which is
generally accepted to be the fundamental physical theory of strong
interactions among the quarks as constituents of matter. The most
time-consuming part of a numerical simulation in lattice gauge theory with
Wilson fermions on the lattice is the computation of quark propagators
within a chromodynamic background gauge field. These computations use up a
major part of the world's high performance computing power.
Quark propagators are obtained by solving the inhomogeneous lattice-Dirac
equation. The Wuppertal Wilson Fermion Solver (wupwise) solves the
inhomogeneous lattice-Dirac equation via the BiCGStab iterative method
which has established itself as a method of choice.
A more detailed characterization of the program can be found in the
Postscript file "wupwise.ps" in the "docs"
subdirectory.
Input Description
The input file wupwise.in contains a few parameters from which the problem
is built, among them:
-
SEED seed value for a random number generator which supplies initial data
-
NITER number of iterations
-
KAPPA parameter characterizing the matrix M = I - KAPPA * D
Output Description
The main output file is "wupwise.out", containing a check value
(EPS) for each the NITER iterations, and a final value EPSILON
characterizing the BiCGStab result.
In addition, there is a file "te.out", containing the value
"TRUE EPSILON". Since this value (in the magnitude of 0.6E-5) has
been found to vary more between systems, it is written to a separate file
for which the SPEC-imposed accuracy requirements are different from those
required for the main output file (see values $reltol in file
168.wupwise/Spec/object.pm).
Programming Language
Fortran 77, with the change that (as in all of SPEC's Fortran
benchmarks) the type
DOUBLE PRECISION
has been replaced by
REAL*8
thus making the size of the relevant floating-point data uniform across
systems - an important consideration for fair benchmarking. While the type
REAL*8 is not part of the strict FORTRAN77 or Fortran90 standards, it is a
common language extension recognized by all Fortran compilers the SPEC CPU
subcomittee has used in its tests.
Known portability issues
The source code as prepared by SPEC contains FORTRAN77 versions of common
LINPACK BLAS functions (zaxpy.f, zcopy.f, zdotc.f, zgemm.f, zscal.f).
According to the CPU2000 Run and Reporting Rules (2.1.2), the level 1, 2,
and 3 BLAS functions may be substituted by precomputed (library-based)
subroutines (compilation for peak runs only, not for baseline runs). For
the case that this is not done (e.g., for baseline runs, or if optimized
libraries are not available), all BLAS functions needed by the benchmark
are provided in the source directory.
Note that since these source functions were taken unchanged from the
respective source code libraries, the spelling of their names has been
retained: They are defined in the files z*.f in lower case, but called in
upper case. For Fortran compilers that insist on a strict FORTRAN77
standard (which only recognizes upper-case letters in identifiers), a
suitable portability flag like
-fu Convert all lower-case identifiers to uppercase
may be necessary.
Last updated: 8 November 1999
