Hello,
Your quadruple version of LAPACK looks great, you should not have concerns.
You have different numerical failures than double precision LAPACK, but nothing
worrisome.
Julien.
On Oct 2, 2011, at 4:15 PM, Alan W. Irwin wrote:
On 2011-09-29 14:40-0700 Alan W. Irwin wrote:
export FC="gfortran-4.6"
export FFLAGS="-O3 -fPIC -ffixed-line-length-132"
mkdir -p build_double_dir
cd build_double_dir
cmake ../lapack-3.3.1_double_complex >& cmake.out
make VERBOSE=1 -j4 >& make.out
ctest >& ctest_double.out
export FFLAGS="-O3 -fPIC -fdefault-real-8 -ffixed-line-length-132"
As in the default precision case, the cmake and make steps completed
without any warning or error messages, but the ctest stage took 6000
seconds with 9 tests taking more than 100 seconds each and two of
those 9 timing out at 1500 seconds (!) On the other hand, many of the
test results completed in times comparable to the corresponding
default test times. (See the attached ctest_quadruple.out results).
So I think this approach for implementing a quadruple-precision
lapack/blas might be close to success.
Actually, the fundamental issue is that on this Intel 64-bit hardware
(uname -a identifies it as
Linux raven 2.6.32-5-amd64 #1 SMP Fri Sep 9 20:23:16 UTC 2011 x86_64
GNU/Linux
),
gfortran quadruple precision is implemented with software rather than
hardware so all the quadruple-precision tests consume roughly 100
times the cpu time of the corresponding double-precision tests. This
slowdown caused timeouts on two of the tests. If I specify a much
longer timeout limit and the verbose option to get detailed test
details, e.g.,
ctest --verbose --timeout 36000
then the 128-bit floating point and 256-bit complex tests all pass
in roughly 2 hours on this hardware.
However, "pass" is defined as a certain fraction of the tests
succeeding, and the verbose results (attached as
ctest_double_verbose.out and ctest_quadruple_verbose.out) show some
individual test failures in detail for both the double and
quadruple case.
Could somebody take a look at those detailed test results to see if
there is anything of particular concern or have I achieved a pretty
high quality (albeit ~100 times slower than the corresponding
64-bit/128-bit float/complex results because of hardware limitations)
quadruple-precision lapack/blas?
Alan
__________________________
Alan W. Irwin
Astronomical research affiliation with Department of Physics and Astronomy,
University of Victoria (astrowww.phys.uvic.ca).
Programming affiliations with the FreeEOS equation-of-state
implementation for stellar interiors (freeeos.sf.net); the Time
Ephemerides project (timeephem.sf.net); PLplot scientific plotting
software package (plplot.sf.net); the libLASi project
(unifont.org/lasi); the Loads of Linux Links project (loll.sf.net);
and the Linux Brochure Project (lbproject.sf.net).
__________________________
Linux-powered Science
__________________________<ctest_double_verbose.out><ctest_quadruple_verbose.out>_______________________________________________
Lapack mailing list
Lapack@Domain.Removed
http://lists.eecs.utk.edu/mailman/listinfo/lapack
|