Update ATLAS to version 3.9.x

[vbraun]

Update to the developer release atlas-3.9.x according to advice of Clint Whaley. The new atlas release now builds netlib lapack itself, so the lapack tarball is now included in /patches. The lapack spkg is no longer needed and will be removed.

• Updated to newest upstream source, various patches are no longer required
• SAGE_ATLAS_LIB=path now searches in path/libatlas.so instead of path/lib/libatlas.so so it works for people with atlas in /lib64, too.
• Threading is now enabled by default
• Flush before os.system (#13210)

By default, ATLAS will now try twice to get timings and fail immediately if throttling is enabled. If auto-tuning fails build with SAGE_ATLAS_ARCH=fast, and if that fails with SAGE_ATLAS_ARCH=base. On x86, the fast and base targets are the new ATLAS generic targets x86SSE3 and x86SSE2/x86x87.

The current spkg version is at

​http://www.stp.dias.ie/~vbraun/Sage/spkg/atlas-3.9.84.spkg

TODO:

• Setting the Fortran compiler to sage_fortran doesn't work currently, see ​https://sourceforge.net/tracker/?func=detail&atid=379483&aid=3541376&group_id=23725. But as long as you are using gfortran the current spkg works.

[vbraun]

Note that you cannot just use sage -f atlas-3.9.32.spkg to update atlas only. Many other spkgs use blas/lapack and must be rebuilt. The easiest way is to do a separate Sage installation...

The cvxopt spkg needs to be updated to link correctly with this atlas release, see #10509.

[dimpase]

Does it mean that LAPACK spkg can be removed, too?

[vbraun]

Yes, the lapack spkg can be removed.

I'm still trying to debug some issues with linbox...

[fbissey]

BLAS can also be removed if we go with this. f2c which was used before we got ATLAS to provide cblas by f2c-ing the BLAS package should also be removed (I think it listed in scipy's dependency only).

[dimpase]

on 32-bit x86 Linux (Debian squeeze) I get the following, when trying to install the spkg (applied the patches to a pristine Sage 4.6.1.alpha3):

...
make -j1 libatlas.so libptf77blas.so libf77blas.so \
                libptcblas.so libcblas.so liblapack.so
make[3]: Entering directory `/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32/ATLAS-build/lib'
ld -melf_i386 -shared -soname /usr/local/src/sage/sage-4.6.1.alpha3/local/lib/libatlas.so -o libatlas.so \
           -rpath-link /usr/local/src/sage/sage-4.6.1.alpha3/local/lib \
           --whole-archive libatlas.a --no-whole-archive -lc -lm
make[3]: *** No rule to make target `libptf77blas.a', needed by `libptf77blas.so'.  Stop.
make[3]: Leaving directory `/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/at
las-3.9.32/ATLAS-build/lib'
make[2]: *** [ptshared] Error 2
make[2]: Leaving directory `/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/at
las-3.9.32/ATLAS-build/lib'
Configuration:
    SAGE_LOCAL: /usr/local/src/sage/sage-4.6.1.alpha3/local
    linker_Solaris?: False
    PPC?: False
    SPKG_DIR: /usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32
    linker_GNU?: True
    ld: GNU
    system: Linux
    Darwin?: False
    machine: i686
    fortran: gfortran
    Solaris?: False
    fortran_g95?: False
    bits: 32bit
    CYGWIN?: False
    SPARC?: False
    fortran_GNU?: True
    FreeBSD?: False
    32bit?: True
    Linux?: True
    64bit?: False
    Intel?: False
    processor: 

[dimpase]

Replying to dimpase:

on 32-bit x86 Linux (Debian squeeze) I get the following, when trying to install the spkg (applied the patches to a pristine Sage 4.6.1.alpha3):

complete install.log is here: ​http://boxen.math.washington.edu/home/dima/tmp/install-alt3.9.log.gz

[vbraun]

Hi Dima,

I think the problem is

make[6]: Entering directory `/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32/ATLAS-build/tune/sysinfo'
gcc -c -DL2SIZE=4194304 -I/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32/ATLAS-build/include -I/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32/ATLAS-build/../src//include -I/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32/ATLAS-build/../src//include/contrib -DAdd_ -DF77_INTEGER=int -DStringSunStyle -DATL_OS_Linux -DATL_ARCH_CoreDuo -DATL_CPUMHZ=800 -DATL_SSE3 -DATL_SSE2 -DATL_SSE1 -DATL_GAS_x8632  -fomit-frame-pointer -O3 -mfpmath=387 -fPIC -m32 /usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32/ATLAS-build/../src//tune/sysinfo/findNT.c
gcc -DL2SIZE=4194304 -I/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32/ATLAS-build/include -I/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32/ATLAS-build/../src//include -I/usr/local/src/sage/sage-4.6.1.alpha3/spkg/build/atlas-3.9.32/ATLAS-build/../src//include/contrib -DAdd_ -DF77_INTEGER=int -DStringSunStyle -DATL_OS_Linux -DATL_ARCH_CoreDuo -DATL_CPUMHZ=800 -DATL_SSE3 -DATL_SSE2 -DATL_SSE1 -DATL_GAS_x8632  -fomit-frame-pointer -O3 -mfpmath=387 -fPIC -m32 -o xfindNT findNT.o ATL_walltime.o -lm
/usr/lib/crt1.o: In function `_start':
(.text+0x18): undefined reference to `main'
collect2: ld returned 1 exit status
make[6]: *** [xfindNT] Error 1

This fails because ATL_NCPU is not set. Do you have a single-core processor? I guess the threaded libraries are not built in that case.

[dimpase]

Replying to vbraun:

This fails because ATL_NCPU is not set. Do you have a single-core processor? I guess the threaded libraries are not built in that case.

yes, it's single-core. An old Pentium M. (Atlas 3.8 spkg does not build on it, at all)

[vbraun]

I changed the spkg-install to make only single-threaded shared libraries if necessary. Should work now.

For simplicity, I made spkgs for cvxopts and sage_scripts. So you need to add

​http://www.stp.dias.ie/~vbraun/Sage/spkg/atlas-3.9.32.spkg ​http://www.stp.dias.ie/~vbraun/Sage/spkg/cvxopt-1.1.3.p0.spkg ​http://www.stp.dias.ie/~vbraun/Sage/spkg/sage_scripts-4.6.1.alpha3.p0.spkg

to $SAGE_ROOT/spkg/standard and then

• replace spkg/install with the attached version (Note that sage_scripts overwrites this file during installation, aargh)
• replace spkg/standard/deps with the attached version.

I'm still having doctest errors with linbox...

[dimpase]

Replying to vbraun:

I changed the spkg-install to make only single-threaded shared libraries if necessary. Should work now.

It builds OK, but then testlong gives quite a bit of failures:

The following tests failed:
        sage -t  -long -force_lib "devel/sage/doc/en/bordeaux_2008/elliptic_curves.rst"
        sage -t  -long -force_lib "devel/sage/sage/modular/modsym/space.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/modsym/tests.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/modsym/subspace.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/modsym/ambient.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/modform/space.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/modform/element.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/modform/ambient.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/hecke/submodule.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/abvar/abvar.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/abvar/torsion_subgroup.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/abvar/cuspidal_subgroup.py"
        sage -t  -long -force_lib "devel/sage/sage/modular/abvar/finite_subgroup.py"
        sage -t  -long -force_lib "devel/sage/sage/matrix/matrix_integer_dense.pyx"
        sage -t  -long -force_lib "devel/sage/sage/matrix/matrix_integer_dense_hnf.py"
        sage -t  -long -force_lib "devel/sage/sage/rings/qqbar.py"
        sage -t  -long -force_lib "devel/sage/sage/finance/time_series.pyx"
        sage -t  -long -force_lib "devel/sage/sage/schemes/elliptic_curves/padic_lseries.py"
        sage -t  -long -force_lib "devel/sage/sage/schemes/elliptic_curves/ell_modular_symbols.py"
        sage -t  -long -force_lib "devel/sage/sage/schemes/elliptic_curves/ell_rational_field.py"
        sage -t  -long -force_lib "devel/sage/sage/schemes/elliptic_curves/sha_tate.py"
Total time for all tests: 30045.4 seconds

I do not know how many of them are Atlas related, though. The log is here: ​http://boxen.math.washington.edu/home/dima/tmp/testlong-atl3.9.log

[fbissey]

A lot of it is probably related. Was this a build from scratch? linbox calls seem to be particularly affected, there are a lot of failures in that path: sage.matrix.matrix_integer_dense.Matrix_integer_dense._charpoly_linbox stuff going through: sage.matrix.matrix_rational_dense.Matrix_rational_dense.right_kernel is also affected and that smells like atlas at work too.

I must say we have observed a number of failures related to ATLAS-3.9.23 in sage-on-gentoo: ​https://github.com/cschwan/sage-on-gentoo/issues#issue/3 you have failure there too but not due to iml/ATLAS as far as I can see. ​https://github.com/cschwan/sage-on-gentoo/issues#issue/6 more subtle. I see you have the failure with devel/sage/sage/finance/time_series.pyx you may want to look the comment I wrote about it in the section "known test failures for Sage on Gentoo": ​https://github.com/cschwan/sage-on-gentoo/wiki/Known-test-failures note that sage-on-gentoo can use (c)blas-reference, ATLAS or gsl-cblas or some combinations - in fact amd and intel libraries could in principle be used but no one I know has tried. ​http://www.gentoo.org/proj/en/science/blas-lapack.xml

Not sure about the SIGFPE, it could come from a result rounded to 0 or the like.

[dimpase]

Replying to fbissey:

A lot of it is probably related. Was this a build from scratch?

yes, it's a build from scratch, on the same machine (old Pentium M) as already discussed above.

linbox calls seem to be particularly affected, there are a lot of failures in that path:

I wonder if these are Atlas bugs, or Linbox bugs...

One should try an OSX build.

[fbissey]

why OS X in particular? I could do that but not before the 5th of January, when my university reopens. Ok I could do it before that but I'll enjoy the break a little bit more :)

[dimpase]

Replying to fbissey:

why OS X in particular? I could do that but not before the 5th of January, when my university reopens. Ok I could do it before that but I'll enjoy the break a little bit more :)

Or does OSX remain disabled for Atlas, i.e. it's not built?

[fbissey]

Replying to dimpase:

Replying to fbissey:

why OS X in particular? I could do that but not before the 5th of January, when my university reopens. Ok I could do it before that but I'll enjoy the break a little bit more :)

Or does OSX remain disabled for Atlas, i.e. it's not built?

I had forgotten about that. I checked the spkg and ATLAS is not built on cygwin and OS X. So it makes sense.

[vbraun]

I get the same doctest errors. Most of them are linbox related. The SIGFPE is from converting a NAN into a GMP integer, but I haven't gotten to the root of the NAN yet. In trying to debug this I've noticed that there are a bunch of valgrind warnings in the linbox code path we are using. I've asked some more specific questions on the linbox-use mailinglist:

​https://groups.google.com/d/topic/linbox-use/N3QNNOQuTAc/discussion

But so far no final conclusion.

[fbissey]

Replying to dimpase:

Replying to vbraun:

This fails because ATL_NCPU is not set. Do you have a single-core processor? I guess the threaded libraries are not built in that case.

yes, it's single-core. An old Pentium M. (Atlas 3.8 spkg does not build on it, at all)

I notice you say it is a Pentium M, yet ATLAS is compiling things with the assumption that it is a coreduo: -DATL_ARCH_CoreDuo -DATL_CPUMHZ=800 -DATL_SSE3 -DATL_SSE2 -DATL_SSE1 -DATL_GAS_x8632 I am guessing the speed is right but the rest is not. Was your successful build using these parameters? I had in one instance a non-working ATLAS because the cpu type was misdetected (wanted to use sse3 when it didn't have them). The build was curiously successful but the library was unusable.

[dimpase]

Replying to fbissey:

Replying to dimpase:

Replying to vbraun:

This fails because ATL_NCPU is not set. Do you have a single-core processor? I guess the threaded libraries are not built in that case.

yes, it's single-core. An old Pentium M. (Atlas 3.8 spkg does not build on it, at all)

I notice you say it is a Pentium M, yet ATLAS is compiling things with the assumption that it is a coreduo: -DATL_ARCH_CoreDuo -DATL_CPUMHZ=800 -DATL_SSE3 -DATL_SSE2 -DATL_SSE1 -DATL_GAS_x8632 I am guessing the speed is right but the rest is not. Was your successful build using these parameters? I had in one instance a non-working ATLAS because the cpu type was misdetected (wanted to use sse3 when it didn't have them). The build was curiously successful but the library was unusable.

The processor is Pentium M Banias 1.1GHz, ​http://ark.intel.com/Product.aspx?id=27600 (according to GotoBlas? installation procedure :-))

The Atlas built is not totally useless, it works for many doctests. By the way, 3.8 also thinks it's a CoreDuo?. And it works. I saw somewhere a note that it's OK.

[fbissey]

Replying to vbraun:

I get the same doctest errors. Most of them are linbox related. The SIGFPE is from converting a NAN into a GMP integer, but I haven't gotten to the root of the NAN yet. In trying to debug this I've noticed that there are a bunch of valgrind warnings in the linbox code path we are using. I've asked some more specific questions on the linbox-use mailinglist:

​https://groups.google.com/d/topic/linbox-use/N3QNNOQuTAc/discussion

But so far no final conclusion.

Converting a NAN into a GMP integer is exactly what was happening in ​https://github.com/cschwan/sage-on-gentoo/issues/3 and it didn't happen when using gslcblas. I will do a full build of sage-on-gentoo with 3.9.40 (or 41) and see if I can see anything.

[fbissey]

Yuck, still got problems leading to linbox, I got the ones leading to iml too. Note that using another cblas like gslcblas/openblas/reference(netlib) all these problems disappear which seem to indicate that there is something going on in ATLAS itself or that all the others gets it wrong (I realise that on a stock sage you may have trouble compiling iml with anything else than ATLAS, it requires some patching to the configure script to be able to do so).

[timdumol]

It may be worth noting that new released are available for both stable and development:

​http://sourceforge.net/projects/math-atlas/files/Stable/3.8.4/|Stable ​http://sourceforge.net/projects/math-atlas/files/Developer%20%28unstable%29/3.9.41/|Unstable

[fbissey]

I have 3.9.41 now, the only difference was it compiled faster because there was tuning for my cpu. But we could investigate 3.8.4. Thanks for pointing it is out, some of us (me at least) didn't think it would ever happen. We may have a stable ATLAS supporting newer CPUs at last, and it looks like I could test it quickly.

[fbissey]

OK 3.8.4 doesn't suffer from any of the drawbacks that are apparent in 3.9.23+.

The big drawback is that it doesn't build lapack nicely on its own, provided we point to the sources like the latest 3.9.xx do. That means spkg-install would need a little bit more work.

Opinions?

[vbraun]

Can we keep this ticket focused on the development version of atlas and discuss the stable ATLAS on #10226? I updated the spkg on the latter ticket to the new stable ATLAS release.

[fbissey]

Sure. My current opinion and that's something I am pushing to sage-on-gentoo users is to avoid ATLAS 3.9.xx for the time being. It is possible that ATLAS-3.9.xx is doing something permissible that iml and linbox are not ready to catch. My line of thought is that I remember that some algorithm in iml takes the inverse of a result returned by cblas. If the result is 0 instead of a small value we may have our NaN, more likely some result from ATLAS is a NaN.

[leif]

Note that you can copy updated spkgs into spkg/standard/ and then do

env SAGE_UPGRADING=yes make build

to rebuild all dependent packages.

(We should perhaps add make targets for that to the top-level Makefile and document them in the Developer's Guide, as this is currently just a side-effect of the effort to make upgrading work.)

If you at the same time need to apply patches to the Sage library, things get a bit more complicated, as e.g. Sage switches to the main branch before reinstalling the Sage library package. One safe way is to first apply the patches, create a new sage-x.y.z-whatever spkg (with devel/sage/spkg-dist) and replace the one in spkg/standard/ with that one (or at least make sure newest_version will pick up the right one).

Note that the extension modules' dependencies in module_list.py are currently far from complete. #8664 adds some in a generic way by adding them automatically in setup.py, i.e. lets modules also depend on the headers of the libraries they use (which [only] works if the headers' mtimes get modified / updated during installation of their corresponding libraries). The dumb alternative is to run sage -ba-force after an "upgrade" process.

[leif]

Ping.

[dimpase]

Replying to leif:

Ping.

just fired up my vintage MacOSX 32-bit Powerbbok PPC... Will know more in, like, 12 hours...

[fbissey]

Several issues:

• atlas is now at version 3.9.49 (last I checked)
• I have reports of it failling to build on some old opterons
• @dimpase: do you need atlas on OS X ppc? Don't you use vectorize like the other OS X?
• I haven't checked but I am quite sure from other reports that cblas_dgem{m,v} is still buggy (someone posted that R built against that Atlas lapack was giving them trouble and R upstream pointed the finger to Atlas).

[dimpase]

Replying to fbissey:

Several issues:

• atlas is now at version 3.9.49 (last I checked)
• I have reports of it failling to build on some old opterons

so what? I have a 5-year old 32-bit Intel laptop on which the sage-current Atlas does not build.

• @dimpase: do you need atlas on OS X ppc? Don't you use vectorize like the other OS X?

indeed. But that's for #10509. Oops, was posting on the wrong ticket...

[kcrisman]

With respect to comment:33:ticket:12011, should this be closed as a dup of #12011?

[dimpase]

propose to close this one and refer to #12011 the the continuation of the upgrade work.

[vbraun]

I would say #12011 is a duplicate of my ticket but oh well ;-)

[jdemeyer]

Seems like #12011 isn't a duplicate after all...