[Debichem-devel] ACESIII-3.0.7 Availability

[Michael Banck]

Hi Ajith,

Sorry for the late reply, I was quite busy over the last days.

On Wed, Jul 18, 2012 at 12:52:37PM -0400, Ajith perera wrote:
> Hi Michael: Where is blas and lapack coming from? Are they 32 bit integer
> versions?

Bingo, after I saw suspiciously many Blas/Lapack routines in the
backtraces, I had another look at what it took us to get ACESII to
compile and run.  Turns out we skipped the blas/lapack "lb" directory
(except for dsum.f, which appears to not be shipped by regular
blas/lapack) and linked to the system-provided linear algebra routines
instead.

I never quite figured out the story with Blas/Lapack on Debian (there is
either refblas/lapack, or ATLAS), but we had recurring issues with
incompatible interfaces.

So after I move away all files in the "dup" subdirectory except for
dsum.F (see above), xscal.F, xdnrm2.f, xdscal.f and xdswap.f, ACESIII
compiles and runs fine on 32bit and 64bit.  It now builds correctly on
all Debian architectures:

https://buildd.debian.org/status/package.php?p=aces3

Note that on architectures which use MPICH instead of OpenMPI, i.e.
mips/mipsel (big/little endian MIPS) and s390/s390x (32/64-bit S/390),
the test suite fails with a segfault still.  I filed bug report to track
this ( ), and might have a look later, but this is less critical.

Thanks a lot for your help!

On another note, we had to radically downsize the test suite to test
case 1.1.1.1, as most of the other test cases proved to be too demanding
for the slower Debian architectures like ARM or MIPS.  I understand that
ACESIII is of course not targetted at embedded systems, but I think
running the test suite on very different architectures is a powerful
indicator whether the code is healthy.

I would like to propose to add some more smaller testcases maybe using
H2 or H2O and mostly smaller basis sets like VDZ, but covering all or
most methods implemented in ACESIII.  This could be a seperate testsuite
like "quicktest" and might also be a good idea to be run before/after
each checkin (unless your development machines are beefy enough to just
run the regular testsuite).

Finally, I have attached a patch we are using to add a "distclean"
Makefile target, for your consideration.  Debian packages are supposed
to build idem-potent, i.e. a "build" followed by a "clean" should leave
no trace.  By adding the distclean target and running it in clean, this
now works for us.  We can of course maintain this patch ourselves if it
does not fit your guidelines.


Best regards,

Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: makefile_distclean.patch
Type: text/x-diff
Size: 875 bytes
Desc: not available
URL: <http://lists.alioth.debian.org/pipermail/debichem-devel/attachments/20120729/ab8eaae0/attachment.patch>