Install ATLAS header files and cleanup IML spkg

[John Palmieri]

With the new ATLAS spkg from #10508, iml fails its self tests. See ​http://trac.sagemath.org/sage_trac/ticket/10508#comment:518 and later comments.

Use spkgs at:

• ​http://boxen.math.washington.edu/home/vbraun/spkg/atlas-3.10.1.p1.spkg (spkg diff​)
• ​http://boxen.math.washington.edu/home/jdemeyer/spkg/iml-1.0.1.p15.spkg (spkg diff without deletes​)

Apply 14699_env_version_bump.patch​ to SAGE_ROOT

[Jean-Pierre Flori]

Surely an ATLAS problem, see ​http://trac.sagemath.org/sage_trac/ticket/10508#comment:526 for a solution.

I suggest we call at least install_inc and potentially install_lib from ATLAS build system to install headers and static libs.

[Jean-Pierre Flori]

Or just install which should call both of them.

[Leif Leonhardy]

John, U sure about sage.math?

128.208.160.191	sage.math.washington.edu sage sage.math

[Leif Leonhardy]

Replying to leif:

John, U sure about sage.math?

128.208.160.191	sage.math.washington.edu sage sage.math

(As opposed to

;; ANSWER SECTION:
sage.math.washington.edu. 600	IN	A	128.208.160.197

which is boxen.)

[Jean-Pierre Flori]

It would be nice to fix that on top of #14605 which I've tested quite extensively and is finished since some time.

[Leif Leonhardy]

+1 to also again install static libs.

[Jean-Pierre Flori]

In fact disregard my comment about #14605. We should fix the issue here quickly, hopefully before 5.10 official release if it's not too late. Although as it does not seem to prevent anything from working except for IML testsuite, 5.11 should be ok as well.

[Koen van de Sande]

A side-effect of this ticket is probably that numpy gets built without ATLAS acceleration.

[Leif Leonhardy]

For the record, I did get

ATLAS_CFLAGS=''
ATLAS_LIBS='-lcblas -latlas'

-L${SAGE_LOCAL}/lib "incidentally" came from / for GMP, although after -lcblas -latlas, but we meanwhile add ${SAGE_LOCAL}/lib to LIBRARY_PATH (in sage-env), so that it worked.

[Leif Leonhardy]

P.S.: Although IIRC according to POSIX, -Ldir1 -lfoo -Ldir2 -lbar is the same as -Ldir1 -Ldir2 -lfoo -lbar.

[Leif Leonhardy]

Replying to leif:

P.S.: Although IIRC according to POSIX, -Ldir1 -lfoo -Ldir2 -lbar is the same as -Ldir1 -Ldir2 -lfoo -lbar.

At least GNU ld docs agree with this:

All -L options apply to all -l options, regardless of the order in which the options appear.

[Koen van de Sande]

You'd still need a line

    cmd += ' --incdir=' + os.path.join(conf['SAGE_LOCAL'],'include') 

in configure_atlas_library() for the header files to go the correct place?

And currently numpy is built without Atlas (it uses another blas): local/lib/python2.7/site-packages/numpy/core/_dotblas.so normally contains ATL_* symbols (for an old Atlas 3.8.3 version, at least).

[Jean-Pierre Flori]

Replying to Koen:

You'd still need a line

    cmd += ' --incdir=' + os.path.join(conf['SAGE_LOCAL'],'include') 

in configure_atlas_library() for the header files to go the correct place?

I don't think so. We use prefix and that should be enough from what I've seen in ATLAS configure script.

[Jean-Pierre Flori]

Replying to Koen:

And currently numpy is built without Atlas (it uses another blas): local/lib/python2.7/site-packages/numpy/core/_dotblas.so normally contains ATL_* symbols (for an old Atlas 3.8.3 version, at least).

On my gcc110 install (where IML testsuite fails), numpy uses ATLAS correctly. There is no ATLAS symbols in the file you pointed because we now use a shared version of ATLAS. But if I run ldd on the file it points to the ATLAS libs.

And I seem to remember we had lots of troubles correctly biulding numpy and scipy with the new ATLAS so hopefully it was actually used.

[Leif Leonhardy]

s/inlcude/include/

and IMHO (=WTF?) also s/serial/single-threaded/g s/parallel/multi-threaded/g ...

(but the latter is not JP's invention I guess)

[Jean-Pierre Flori]

Replying to leif:

s/inlcude/include/

Done.

and IMHO (=WTF?) also s/serial/single-threaded/g s/parallel/multi-threaded/g ...

(but the latter is not JP's invention I guess)

Not mine indeed.

[Leif Leonhardy]

Replying to jpflori:

Replying to Koen:

And currently numpy is built without Atlas (it uses another blas): local/lib/python2.7/site-packages/numpy/core/_dotblas.so normally contains ATL_* symbols (for an old Atlas 3.8.3 version, at least).

On my gcc110 install (where IML testsuite fails), numpy uses ATLAS correctly. There is no ATLAS symbols in the file you pointed because we now use a shared version of ATLAS. But if I run ldd on the file it points to the ATLAS libs.

Yes.

$ readelf -d local/lib/python2.7/site-packages/numpy/core/_dotblas.so

Dynamic section at offset 0x4dc0 contains 26 entries:
  Tag        Type                         Name/Value
 0x0000000000000001 (NEEDED)             Shared library: [libptf77blas.so.2]
 0x0000000000000001 (NEEDED)             Shared library: [libptcblas.so.2]
 0x0000000000000001 (NEEDED)             Shared library: [libatlas.so.2]
 0x0000000000000001 (NEEDED)             Shared library: [libpython2.7.so.1.0]
 0x0000000000000001 (NEEDED)             Shared library: [libpthread.so.0]
 0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]
 ...

[Jean-Pierre Flori]

make_atlas() takes no argument.

[John Palmieri]

I saw the error on boxen. I thought that sage.math was dead? In any case, it's the machine I get to by typing ssh sage.math.washington.edu, and when I log off it says Connection to sage.math.washington.edu closed. The $HOSTNAME is boxen, though.

[John Palmieri]

On boxen, after installing the new ATLAS spkg, the current iml spkg installs and passes its tests. Same for an iml spkg which doesn't copy over the patched version of configure.

[Leif Leonhardy]

P.S.: I guess the odd "patch" to IML's configure dates back to the times when ATLAS was somehow optional, or at least did not get properly detected on all systems. Running / building IML's test suite on such systems was always broken then.

(In other parts of Sage using IML, BLAS libraries get explicitly added by us when linking.)

[Volker Braun]

Do not install static libraries, its a feature that we don't.

Also, nobody in his right mind expects blas headers. This is why the iml testsuite, which the authors probably didn't expect to be distributed, is the only place that trips over it. For starters, there is no universal naming standard (blas.h vs. cblas.h vs. foo_cblas.h vs. foo/cblas.h). There are multiple blas implementations and if you all install them at the same place then they would overwrite each other.

[Volker Braun]

And if you need a reason for not installing static libraries you just have to look at this ticket where you got confused about whether something links against atlas or not. Shared libraries are the only way that make it clear which particular atlas version you are linking against and that you didn't accidentally pick up some system atlas.

[Koen van de Sande]

I disagree that nobody expects the blas headers. If I try to install a package on top of Sage, e.g. the Shogun machine learning toolbox, then it will only compile in certain features if it can find a blas - and for that it needs blas headers. With older sage versions this works just fine.

[Volker Braun]

Shogun just makes it unnecessarily complicated for themselves, you need to define preprocessor constants to tell it which blas header to look for...

I agree that we should have blas/cblas/f77blas headers since its the right thing to do. But when we fix the IML testsuite we shouldn't introduce that as a dependency, because very often the headers are either not installed or installed with a funky name.

[Leif Leonhardy]

Replying to vbraun:

Do not install static libraries, its a feature that we don't.

At least they shouldn't get discarded, which is simply stupid.

[Leif Leonhardy]

Btw., the new spkg and afterwards IML installed without problems, and both passed their test suites on Linux x86_64 (Ubuntu 10.04.4, Sage 5.10.rc0).

[Volker Braun]

Replying to leif:

At least they shouldn't get discarded, which is simply stupid.

Do you also object to discarding the *.o files? We delete what we don't want around in the install, thats the way it works. Ideally upstream would have an --disable-static switch, but they don't.

[Volker Braun]

Removed the static library install bit from the spkg

[Leif Leonhardy]

Replying to vbraun:

Replying to leif:

At least they shouldn't get discarded, which is simply stupid.

Do you also object to discarding the *.o files?

man ar?

[Volker Braun]

Replying to leif:

man ar?

I take it that this means that you agree with me, and we should neither install object files nor static libraries ;-)

[Leif Leonhardy]

Probably vote on sage-devel on whether to discard the static libraries or not.

[Jean-Pierre Flori]

Replying to leif:

Replying to vbraun:

Do not install static libraries, its a feature that we don't.

At least they shouldn't get discarded, which is simply stupid.

Having static libraries rather than nothing on systems where building shared ones is problematic might be nice... Also note it's the only option upstream really supports.

[Leif Leonhardy]

Replying to leif:

Probably vote on sage-devel on whether to discard the static libraries or not.

... or make it optional.

[Volker Braun]

Done, see ​https://groups.google.com/d/msg/sage-devel/mthuvh5SRpQ/nDekknXc76IJ

[Jeroen Demeyer]

Can we please concentrate on the actual ticket issue here, which is about header files and iml failing? I don't think this has anything to do with static libraries (but please correct me if I'm wrong).

[Jean-Pierre Flori]

Ok. Only concentrating on that, I'm ok with Volker changes to my changes.

[Volker Braun]

I'm ok with Jean-Pierre changes to the extent that I haven't changed them, of course.

[Leif Leonhardy]

Replying to jdemeyer:

Can we please concentrate on the actual ticket issue here, which is about header files and iml failing? I don't think this has anything to do with static libraries (but please correct me if I'm wrong).

Well, if we used ATLAS' install rather than further hacking, we'd get both the headers and the static libraries.

I don't see any reason why discarding static libraries should be done by default; if someone feels they're dangerous to him/her, we could add some option to automatically remove them, although that's IMHO a pretty unimportant feature.

[Leif Leonhardy]

P.S.: Removing the "patch" to IML's configure from the IML spkg IMHO belongs to this ticket as well, as that was hiding (or rather obscuring) the real problem.

[Volker Braun]

Replying to leif:

P.S.: Removing the "patch" to IML's configure from the IML spkg IMHO belongs to this ticket as well, as that was hiding (or rather obscuring) the real problem.

Then either attach a fixed iml spkg to this packet real soon (TM) or open a ticket for it. But don't leave footnotes on finished tickets.

[Leif Leonhardy]

Replying to vbraun:

Replying to leif:

P.S.: Removing the "patch" to IML's configure from the IML spkg IMHO belongs to this ticket as well, as that was hiding (or rather obscuring) the real problem.

Then either attach a fixed iml spkg to this packet real soon (TM) or open a ticket for it. But don't leave footnotes on finished tickets.

I was saying it's IMHO not finished.

[Jean-Pierre Flori]

Indeed, the ticket original title was about IML testsuite, so we should also provide an IML spkg (quite trivial just remove the horrible patch disabling test for cblas.h).

leif: could you make the spkg? I'll quickly review it.

[Jean-Pierre Flori]

Or at least we should provide this fixed IML spkg in a follow up ticket mentioned here and put as blocker for 5.10.

[Leif Leonhardy]

This doesn't at all work with SAGE_ATLAS_LIB [unless the headers happen to be installed system-wide, which is the same situation as before], since headers don't get copied / symlinked.

What happens on MacOS X (where Sage's ATLAS by default doesn't get built)?

(Similar for Cygwin perhaps.)

Replying to jpflori:

Indeed, the ticket original title was about IML testsuite, so we should also provide an IML spkg (quite trivial just remove the horrible patch disabling test for cblas.h).

So I'm not sure if it's really that easy.

[Jean-Pierre Flori]

Yeah, I was just looking into that... Not really anything useful to do when a system-wide ATLAS is used.

Does really no software linking to ATLAS expect to find headers in some standard or provided place? Do they all provide their own BLAS headers and expect everything to be fine?

[Volker Braun]

Its pretty much common practice to use your own blas/lapack headers. For example, linbox/fflas: ​http://linalg.org/projects/fflas-ffpack/browser/fflas-ffpack/config-blas.h Definitely weird, but then its probably the most stable API ever. And the lack of uniform naming of the header across vendors doesn't help.

[Jean-Pierre Flori]

A test spkg is at:

• ​http://boxen.math.washington.edu/home/jpflori/spkg/iml-1.0.1.p15.spkg

Not really wonderful as the src directory still includes a modified upstream source, but maybe a little bit better looking. Not tested yet, I'm doing it right now.

[Jeroen Demeyer]

The use of "$SAGE_LOCAL"/lib/libatlas.so.*.* seems to assume that there is exactly one file matching "$SAGE_LOCAL"/lib/libatlas.so.*.*

I doubt that this assumption is true in general.

[Jean-Pierre Flori]

Quite true (just wondering if we really want to support Sage's installs with multiple ATLAS version installed?).

[Jeroen Demeyer]

Working on IML spkg, hang on...

[Jean-Pierre Flori]

Of course I guess the IML testsuite will still fail on OS X or Cygwin.

[Jeroen Demeyer]

IML spkg needs review, I restored upstream sources and added patches in the patches/ directory. I know the patches can be catalogued and documented better, but let's leave that for next time...

Tested on Linux and OS X x86_64, IML (with testsuite) works.

[Jeroen Demeyer]

I'm adding 14699_env_version_bump.patch​ because ATLAS now requires the FC environment variable to be set, so it needs an up-to-date sage-env.

[Jeroen Demeyer]

Spkg diff, for review only.

[Jean-Pierre Flori]

Replying to jdemeyer:

I'm adding 14699_env_version_bump.patch​ because ATLAS now requires the FC environment variable to be set, so it needs an up-to-date sage-env.

What does that achieve exactly? Just looking at the patch, I cannot really guess. Does it force upgrading the sage-env script?

A question for Volker: How does the SO_VERSION variable in the ATLAS autotools project be updated?

[Jeroen Demeyer]

I'll copy the comments from sage-env:

# Don't execute the commands more than once for the same version of
# sage-env.  Check this after checking the validity of SAGE_ROOT, but
# before modifying its value.
#
# The idea of versioning is that SAGE_ENV_SOURCED will be set to a
# "version number" of sage-env.  If a different version of sage-env was
# sourced earlier (when upgrading), we still source the new version.
# The sage-env version should be increased whenever a newer sage-env is
# required for upgrading.  Note that spkg/standard/deps might also need
# to be changed: the packages which need the new sage-env must depend on
# SAGE_ROOT_REPO (the root repo contains sage-env).
#
# sage-env version history:
# - sage-4.7.1: version 1 (#10469)
# - sage-5.4:   version 2 (#13395)
#
SAGE_ENV_VERSION=2
if [ "$SAGE_ENV_SOURCED" = "$SAGE_ENV_VERSION" ]; then
    # Already sourced, nothing to do.
    return 0
fi
export SAGE_ENV_SOURCED=$SAGE_ENV_VERSION

[Jean-Pierre Flori]

In the IMl spkg, you don't reexport CFLAGS after checking for SAGE64 (both in spkg-check and spkg-install).

And when looking through trac there is a strange .patch part about renaming sage3_memleak, but that may just be a trac issue.

[Jeroen Demeyer]

Replying to jpflori:

In the IMl spkg, you don't reexport CFLAGS after checking for SAGE64 (both in spkg-check and spkg-install).

It's already exported before. No need to "reexport" it.

And when looking through trac there is a strange .patch part about renaming sage3_memleak, but that may just be a trac issue.

Indeed, that's a Trac bug.

[Jeroen Demeyer]

I should also note that the only new patch in IML is patches/remove_repl.patch, all the others were already (pre-patched) in the spkg.

[Jean-Pierre Flori]

Replying to jdemeyer:

Replying to jpflori:

In the IMl spkg, you don't reexport CFLAGS after checking for SAGE64 (both in spkg-check and spkg-install).

It's already exported before. No need to "reexport" it.

Damn, I was not aware of that!

And when looking through trac there is a strange .patch part about renaming sage3_memleak, but that may just be a trac issue.

Indeed, that's a Trac bug.

[Jean-Pierre Flori]

The spkg looks good, seems to work on my systems.

[Jeroen Demeyer]

Thanks for the review, do you by chance feel like reviewing the IML upgrade at #748?

[Jeroen Demeyer]

Reverting the ATLAS package: #14753.

[Jeroen Demeyer]

Keeping "merged in" information for future reference, we should not change this ticket further.