Update Givaro, FFLAS-FFPACK and LinBox

[jpflori]

Tarballs of released packages (with appropriate names) can be found at:

• ​https://github.com/linbox-team/givaro/releases/download/v4.0.1/givaro-4.0.1.tar.gz
• ​http://lig-membres.imag.fr/pernet/prereleases/linbox-1.4.1rc1.tar.gz
• ​http://lig-membres.imag.fr/pernet/prereleases/fflas_ffpack-2.2.1rc1.tar.bz2

[jpflori]

Requires to update:

• fflas-ffpack to 2.0.0, see #17636
• linbox, no release yet

[fbissey]

So we have to wait for linbox? Do we have an idea of when it will happen?

[jpflori]

Replying to fbissey:

So we have to wait for linbox? Do we have an idea of when it will happen?

Yes we have to wait for a new Linbox, expected about the end of february, Clément Pernet wrote me, together with updated Givaro and fflas-ffpack.

By the way, I started this for #17630 which might interest you as well, I'll push code there tomorrow.

[jdemeyer]

Replying to jpflori:

Replying to fbissey:

So we have to wait for linbox? Do we have an idea of when it will happen?

Yes we have to wait for a new Linbox, expected about the end of february, Clément Pernet wrote me, together with updated Givaro and fflas-ffpack.

It's June already...

Since givaro, LinBox and FFLAS-FFPACK all depend on each other, we should make one ticket.

[cpernet]

Sorry for the delay in releasing the fflas-ffpack/Givaro/LinBox environment. We're working on it now, and I'll keep you posted about the status. It should happen before the end of the year and hopefully much earlier.

I also meant to propose to replace ATLAS by OpenBLAS in sage. Looking at your clean-up work of the BLAS in #17630, I understand that we should proceed in the following order:

1. release linbox
2. upgrade linbox fflas-ffpack Givaro in Sage (this ticket #17635)
3. clean-up the BLAS usage #17630
4. replace ATLAS by OpenBLAS (ticket to be created)

I'll be happy to help in these tasks.

[jdemeyer]

Let me remind you also about #15535, which pops up regularly. I don't know if it's a bug in Linbox or in the way that Sage uses Linbox, but it would be good to have a Linbox developer look at it.

[fbissey]

Replying to cpernet:

Sorry for the delay in releasing the fflas-ffpack/Givaro/LinBox environment. We're working on it now, and I'll keep you posted about the status. It should happen before the end of the year and hopefully much earlier.

I also meant to propose to replace ATLAS by OpenBLAS in sage. Looking at your clean-up work of the BLAS in #17630, I understand that we should proceed in the following order:

1. release linbox
2. upgrade linbox fflas-ffpack Givaro in Sage (this ticket #17635)
3. clean-up the BLAS usage #17630
4. replace ATLAS by OpenBLAS (ticket to be created)

I'll be happy to help in these tasks.

I'll admit to have been a bit demotivated on the blas cleaning front. The tickets needs re-organizing and have good descriptions so we know what is supposed to go in them. Getting ATLAS to produce .pc file and using them is probably a priority task. Once that's done someone could create an optional openblas spkg or something.

[cpernet]

I'm working on it, and will release a new version of the 3 libs for the occasion, so as to ensure a smoother integration upstream. This will very likely be givaro-4.0.1 fflas-ffpack-2.2.0 linbox-1.4.1

These libraries now require c++11 support and this tickets therefore depends on #18323.

Inspired by what's propose in there, I suggest to add lines of the form

+# distutils: extra_compile_args = -std=c++11

to libs/linbox/linbox.pxd lib/linbox/fflas.pxd lib/linbox/echelonform.pxd but I understand that this is subject to a debate on sage-devel ​https://groups.google.com/forum/#!topic/sage-devel/aGwotQA9to8

If this option happens to be accepted, we would then need to add more cflags in the list: namely those produced by fflas-ffpack configure related to the target architecture: -mavx -msse4.1 and others like -fabi-version=6 -fopenmp.

I could not figure a way to append the string returned by bin/fflas-ffpack-config --cflags to these pxd files. Any hint?

[fbissey]

Replying to cpernet:

I could not figure a way to append the string returned by bin/fflas-ffpack-config --cflags to these pxd files. Any hint?

Add a definition in module_list.py, you define a LINBOX_FLAGS. It would be easier if you had a .pc file as we have a python interface to pkgconfig that we can use easily. Here you have to do your own parsing from python. Anyway you can then use LINBOX_FLAGS in the .pxd file.

[jdemeyer]

Replying to cpernet:

-mavx -msse4.1

I understand that the libraries themselves want these, but why the Cython files?

and others like -fabi-version=6

This looks like it should be used always for all C++ files then. Mixing ABI's doesn't look like a good idea.

-fopenmp

Same question as before, is this really needed for the Cython interface?

[cpernet]

Replying to jdemeyer:

Replying to cpernet:

-mavx -msse4.1

I understand that the libraries themselves want these, but why the Cython files?

fflas-ffpack and linbox are essentially source libraries, so when you include the headers, you include the whole code. This is by design, due to the ubiquity of template programming in these libraries.

Now the old linbox-sage interface compiles some specializations and offer them through a C api. These extra flags would not be required if we would stick to using this interface.

But, since then, Cython started to offer the capability to hook C++ code directly and more functionnalities have been added in sage where fflas-ffpack code is called directly, like in libs/linbox/fflas.pxd where the C++ routine FFLAS::fgemm is called.

and others like -fabi-version=6

This looks like it should be used always for all C++ files then. Mixing ABI's doesn't look like a good idea.

Sure. Maybe one should change the ABI higher up then.

-fopenmp

Same question as before, is this really needed for the Cython interface?

Same answer as above.

[cpernet]

I added my working branch to the ticket. I could not find if and where I should store the tarballs of the new spkg?

As of today (sept 28 2015); they correspond to the upstream version of the 3 libs: givaro, fflas-ffpack and linbox.

I added .pc files for givaro and fflas-ffpack (coming soon for linbox), and defined the extra_compile_flages in modules_list.py as suggested by fbissey.

This is still work in progress:

• LinBox sparse solver make the test-suite fail: I'm investigating the problem, which I could not reproduce in LinBox for the moment.
• the extra_compile_flags should be inferred from the .pc files (harcoded for the moment). I did not find how to use pkg-config through python to do it.

---

New commits:

​dd81c26

Update givaro to v4.0.1 fflas-ffpack to v2.2.0 and linbox to v1.4.1rc0.

[jdemeyer]

Replying to cpernet:

I could not find if and where I should store the tarballs of the new spkg?

Just add a link to the upstream tarballs on this ticket. If any renaming is needed, you should rename the tarball and add a link to the renamed tarball.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​c97f547

update fflas-givaro-linbox spkg version and minor chnages to configuration files

[jdemeyer]

Instead of adding flags to individual extensions, they should be added to the .pxd or .pyx files.

[cpernet]

Ok, now I get confused: fbissey suggested to define all compilation flags in module_list.py. I saw there that most spkg related extension had their flags set in this file: e.g. flint, fplll, Lfunction, m4ri, m4rie... Hence I moved all flags that I originally wrote in the linbox.pxd fflas-ffpack.pxd givaro.pxd files to module_list.py which has the advantage of defining everything at the same place.

Jeroen or François are you suggesting that the proper way to proceed is to

• define the flags givaro_extra_compile_args, fflas_ffpack_extra_compile_args, linbox_extra_compile_args in module_list.py
• use them in the pxd files with something like

  +# distutils: extra_compile_args = ${givaro_extra_compile_args}
  

? Will do, if that's the way to go.

Also, I did not find how to invoque the python interface to pkgconfig, mentionned by françois.

[fbissey]

Sorry I have been of the net for most of the week. For how to use pkgconfig you can follow the way I use it for blas in sage-on-gentoo at ​https://github.com/cschwan/sage-on-gentoo/blob/master/sci-mathematics/sage/files/sage-6.8-blas-r1.patch. Note that you probably don't need to parse the library_dir, I do it for blas in unusual location, which is expected for intel MKL or the AMD equivalent.

[jdemeyer]

Replying to cpernet:

I moved all flags that I originally wrote in the linbox.pxd fflas-ffpack.pxd givaro.pxd files to module_list.py

Please don't!

which has the advantage of defining everything at the same place.

...but the disadvantage of requiring to add the flags to every single extension, which is much less scalable.

Jeroen or François are you suggesting that the proper way to proceed is to

• define the flags givaro_extra_compile_args, fflas_ffpack_extra_compile_args, linbox_extra_compile_args in module_list.py
• use them in the pxd files with something like

  +# distutils: extra_compile_args = ${givaro_extra_compile_args}
  

Essentially yes. You can grep for GSL_LIBRARIES to see how it's supposed to be done.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​fb16751

Pass the library specific cflags in the pxd files instead of extensions in module_list.py

[jdemeyer]

This is still not quite what I meant.

For example, why does src/sage/rings/polynomial/multi_polynomial_ideal_libsingular.pxd need GIVARO_CFLAGS? Looking at that file, there is nothing related to givaro.

I assume you added GIVARO_CFLAGS because it still somehow indirectly includes Givaro code. But then you should add GIVARO_CFLAGS in the place where the Givaro headers are actually included and Cython's dependency tracking should pick it up.

See again the GSL_LIBRARIES example: I don't add GSL_LIBRARIES to every module which somehow indirectly uses GSL, only where the library functions are actually declared.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​c25de33

Remove all unecessary extra_compile_args for GIVARO and FFLASFFPACK.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​73cc6b3

Finds GIVARO and FFLASFFPACK CFLAGS in the .pc pkgconfig file.

[cpernet]

Replying to jdemeyer:

I assume you added GIVARO_CFLAGS because it still somehow indirectly includes Givaro code. But then you should add GIVARO_CFLAGS in the place where the Givaro headers are actually included and Cython's dependency tracking should pick it up.

Yes. It was necessary to add these flags to extensions when everything was gathered in module_list.py, but no longer now that it is in each pxd files. I now see the point! Sorry for messing this up, it's fixed. I also made the CFLAGS variables set by the pkgconfig .pc files, as suggested by François.

One pb remains: there is no gmp.pc for the moment, and the givaro.pc lists gmp as a dependency. Similarly fflas-ffpack.pc should list blas as dependency, but we still don't have a blas.pc file (see #17075). I suggest to remove these dependencies for the moment, and I'll add them once we have the proper .pc files. If you confirm, I'll produce new tar.gz for givaro and fflas-ffpack accordingly.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​7fd9523

Add link flags for fflas-ffpack.

[jdemeyer]

Are there any significant API changes in the new linbox? I would like to add a linbox interface for sparse matrices over ZZ. I could do that using the current linbox-1.3.2 but I'd like to know that it will still work after upgrading linbox.

[jdemeyer]

Merged latest beta.

---

New commits:

​c447352

Merge tag '7.1.beta3' into t/17635/givaro_fflasffpack_linbox

[git]

Branch pushed to git repo; I updated commit sha1. This was a forced push. New commits:

​f31752c

Merge tag '7.1.beta3' into t/17635/givaro_fflasffpack_linbox

[jdemeyer]

fflas_ffpack doesn't build:

libtool: compile:  g++ -DHAVE_CONFIG_H -I. -I../.. -I../.. -I../../fflas-ffpack/utils/ -I../../fflas-ffpack/fflas/ -I../../fflas-ffpack/ffpack -I../../fflas-ffpack/field -I/usr/local/src/sage-git/local/include -I/usr/local/src/sage-git/local//include -D__FFLASFFPACK_HAVE_CBLAS -msse4.1 -mavx -fabi-version=6 -fopenmp -O2 -Wall -DNDEBUG -UFFLASFFPACK_DEBUG -g -O2 -fPIC -std=gnu++11 -c fflas_L1_inst.C  -fPIC -DPIC -o .libs/fflas_L1_inst.o
libtool: compile:  g++ -DHAVE_CONFIG_H -I. -I../.. -I../.. -I../../fflas-ffpack/utils/ -I../../fflas-ffpack/fflas/ -I../../fflas-ffpack/ffpack -I../../fflas-ffpack/field -I/usr/local/src/sage-git/local/include -I/usr/local/src/sage-git/local//include -D__FFLASFFPACK_HAVE_CBLAS -msse4.1 -mavx -fabi-version=6 -fopenmp -O2 -Wall -DNDEBUG -UFFLASFFPACK_DEBUG -g -O2 -fPIC -std=gnu++11 -c fflas_L3_inst.C  -fPIC -DPIC -o .libs/fflas_L3_inst.o
In file included from ../../fflas-ffpack/fflas/fflas_sparse/csr.h:89:0,
                 from ../../fflas-ffpack/fflas/fflas_sparse.h:126,
                 from ../../fflas-ffpack/fflas/fflas.h:164,
                 from fflas_L1_inst.C:34:
../../fflas-ffpack/fflas/fflas_sparse/csr/csr_pspmm.inl:48:2: error: stray '#' in program
   PAR_BLOCK{
  ^

[cpernet]

I've updated the tarballs of the 3 libraries to the latests rc1 (the failing code in longer there) and updated the checksums. Everything builds on my box.

---

New commits:

​db82895	Patch merged upstream in fflas-ffpack
​a45e379	missing coma
​4cd2289	updating givaro, fflas-ffpack and linbox to latest rc1 version

[jdemeyer]

On powerpc64le, I had to kill the Givaro test test-rmint_inv_ after 11 hours. I don't think the test is supposed to take that long...

[cpernet]

Thanks for raising this one. Bug reported on givaro's issue tracker: ​https://github.com/linbox-team/givaro/issues/6

Currently looking for an acces to a ppc64el to debug it.

The part of givaro that fails, recint, does small lenght multi-precision arithmetic, and very likely triggers endianness issues. Note that this part of the library is not currently used in sage neither directly nor through fflas-ffpack and linbox. However we'd like to have it fixed anyway, in order to update givaro's debian package.

[jdemeyer]

On the same machine, fflas-ffpack also fails: ​https://github.com/linbox-team/fflas-ffpack/issues/16

[jdemeyer]

Apart from the build failures on some systems, are there any further obstructions to this ticket?

[cpernet]

Replying to jdemeyer:

On the same machine, fflas-ffpack also fails: ​https://github.com/linbox-team/fflas-ffpack/issues/16

Fixed in ​https://github.com/linbox-team/fflas-ffpack/commit/8634b80bd0c20c538faf9471c8e2289c5a50a925

Replying to jdemeyer:

Apart from the build failures on some systems, are there any further obstructions to this ticket?

Apart from these, no obstruction on our side.

[jdemeyer]

Replying to cpernet:

Apart from these, no obstruction on our side.

So why was this ticket never set to "needs_review" then?

[jdemeyer]

Replying to cpernet:

Replying to jdemeyer:

On the same machine, fflas-ffpack also fails: ​https://github.com/linbox-team/fflas-ffpack/issues/16

Fixed in ​https://github.com/linbox-team/fflas-ffpack/commit/8634b80bd0c20c538faf9471c8e2289c5a50a925

Do you plan to release fflas_ffpack-2.2.0rc2 for this? Or should we just use git master?

[cpernet]

• the bug of Givaro on powerpc64le has been fixed upstream : ​https://github.com/linbox-team/givaro/issues/6
• yes I will cut a rc2 for both givaro and fflas-ffpack, if not a release.
• I did not put the need_review as there was still a hanging question about the blas.pc file and gmp.pc. I assume we will do without these files for the moment.

[jdemeyer]

Replying to cpernet:

• I did not put the need_review as there was still a hanging question about the blas.pc file and gmp.pc.

I see. However, did anything change compared to earlier versions of these packages regarding compilation? If not and it worked before, it should still work...

[cpernet]

As far as build and compilation is concerned, nothing changed. I'm checking a last few things on the linbox side (test suite fails badly on power8). Then I'd like to release this evening or tomorow and put this ticket in needs_review status.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​f82a77d

Update checksums to the latests released packages

[cpernet]

Givaro, fflas-ffpack and linbox have just been released: ​https://groups.google.com/forum/#!topic/linbox-devel/6dIj_HW3k38

I updated the link in the ticket descriptions to the latest tarball and updated the checksums.

This ticket is now ready for review.

[fbissey]

Before passing to other details for this ticket, is everyone ok with linbox having openmp enabled?

[fbissey]

Actually both fflas-ffpack and linbox, fflas-ffpack has openmp enabled by default and it is stated explicitly in linbox. Our minmal compiler should have openmp support, I am wondering about unintended consequences. In sage-on-gentoo I am now using openblas with threads rather than openmp as I got into trouble with it.

[cpernet]

Eventhough openmp is enabled in fflas-ffpack̀ and linbox, the current code called by sage does not use any OpenMP parallelism.

If you think it is necessary, I could pass the --disable-open argument to the configure scripts. Thanks for pointing out that linbox still forces the -fopenmp evenif --disable-omp was passed to configure. We'll fix it.

Regarding the use of a threaded BLAS: it is possible to run the sequential code of fflas-ffpack on top of a parallel BLAS, but one will achieve much better performances by running the parallel code of fflas-ffpack (with OpenMP enabled) on top of a sequential BLAS. This will need further discussions when we'll want to have sage run its linear algebra kernels in parallel.

[fbissey]

Replying to cpernet:

Eventhough openmp is enabled in fflas-ffpack̀ and linbox, the current code called by sage does not use any OpenMP parallelism.

Yes

If you think it is necessary, I could pass the --disable-open argument to the configure scripts.

At this stage I think that's an unknown that will only be determined by experience.

Thanks for pointing out that linbox still forces the -fopenmp evenif --disable-omp was passed to configure. We'll fix it.

I haven't looked at linbox' configure.ac yet so I don't know about that. I only noticed that it for fflas-ffpack it is the default if you don't pass anything.

Regarding the use of a threaded BLAS: it is possible to run the sequential code of fflas-ffpack on top of a parallel BLAS, but one will achieve much better performances by running the parallel code of fflas-ffpack (with OpenMP enabled) on top of a sequential BLAS. This will need further discussions when we'll want to have sage run its linear algebra kernels in parallel.

Sure.

[fbissey]

I had a concern that there may be linking issue on cygwin because of libgomp. But -fopenmp should be in the CFLAGS provided by fflas-ffpack.pc. So long as those flags are passed to the linker I guess we should be ok.

@jpflori can you comment on the linking needs on cygwin?

[fbissey]

OK, more thoughts about fflas-ffpack. By default usage of sse4.1, avx, avx2 and fma instructions are enabled if found. This is a problem when building a binary for redistribution as we cannot guarantee that the target machine will support any instructions found on the building machine. We need to be able to disable those for binary build (SAGE_FAT_BINARY=yes?).

Going through linbox now. Like fflas-ffpack, openmp is enabled by default even without --enable-openmp in spkg-install but it looks like it can be disable the standard way to me. linbox looks for

LB_CHECK_IML
LB_CHECK_M4RI
LB_CHECK_M4RIE
LB_CHECK_MPFR
# we nedd mpfr next :
LB_CHECK_FPLLL
LB_CHECK_FLINT

LB_CHECK_NTL

but none of these are in linbox's dependencies. Are any of those actually used? Actually the current version should depend on ntl and iml

(sage-sh) fbissey@QCD-nzi3:sage-7.0$ ldd -r local/lib/liblinboxsage.so.0.0.0 
	linux-vdso.so.1 (0x00007ffcaf131000)
	libgmpxx.so.8 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libgmpxx.so.8 (0x00007f3f89ea8000)
	libgmp.so.16 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libgmp.so.16 (0x00007f3f89c31000)
	libntl.so.19 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libntl.so.19 (0x00007f3f89810000)
	libcblas.so.3 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libcblas.so.3 (0x00007f3f895ed000)
	libatlas.so.3 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libatlas.so.3 (0x00007f3f88d01000)
	liblinbox.so.0 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/liblinbox.so.0 (0x00007f3f88afa000)
	libgivaro.so.0 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libgivaro.so.0 (0x00007f3f8889d000)
	libstdc++.so.6 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libstdc++.so.6 (0x00007f3f88557000)
	libm.so.6 => /lib64/libm.so.6 (0x00007f3f88253000)
	libc.so.6 => /lib64/libc.so.6 (0x00007f3f87eb8000)
	libgcc_s.so.1 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libgcc_s.so.1 (0x00007f3f87ca0000)
	libgf2x.so.1 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libgf2x.so.1 (0x00007f3f87a8a000)
	libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f3f8786d000)
	/lib64/ld-linux-x86-64.so.2 (0x0000561bc940b000)
undefined symbol: maxMagnMP	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: findRNS	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: RNSbound	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: certSolveMP	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: certSolveRedMP	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: nonsingSolvLlhsMM	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: nonsingSolvRNSMM	(local/lib/liblinboxsage.so.0.0.0)

compared to sage-on-gentoo on the same machine

ldd -r /usr/lib64/liblinboxsage.so.0.0.0
	linux-vdso.so.1 (0x00007ffe013bb000)
	libgmpxx.so.4 => /usr/lib64/libgmpxx.so.4 (0x00007f8a84d7a000)
	libgmp.so.10 => /usr/lib64/libgmp.so.10 (0x00007f8a84b09000)
	libntl-9.6.2.so => /usr/lib64/libntl-9.6.2.so (0x00007f8a846e5000)
	libreflapack.so => /usr/lib64/libreflapack.so (0x00007f8a83f03000)
	libopenblas_threads.so.0 => /usr/lib64/libopenblas_threads.so.0 (0x00007f8a8396b000)
	liblinbox.so.0 => /usr/lib64/liblinbox.so.0 (0x00007f8a83763000)
	libgivaro.so.0 => /usr/lib64/libgivaro.so.0 (0x00007f8a83507000)
	libiml.so.0 => /usr/lib64/libiml.so.0 (0x00007f8a832e6000)
	libstdc++.so.6 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libstdc++.so.6 (0x00007f8a82fd6000)
	libm.so.6 => /lib64/libm.so.6 (0x00007f8a82cd2000)
	libc.so.6 => /lib64/libc.so.6 (0x00007f8a82937000)
	libgcc_s.so.1 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libgcc_s.so.1 (0x00007f8a8271f000)
	libgf2x.so.1 => /usr/lib64/libgf2x.so.1 (0x00007f8a82509000)
	libgfortran.so.3 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libgfortran.so.3 (0x00007f8a821e0000)
	libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f8a81fc4000)
	/lib64/ld-linux-x86-64.so.2 (0x00005557ec4cf000)
	libquadmath.so.0 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libquadmath.so.0 (0x00007f8a81d85000)

Where I had to add LIBS="-liml" for things to be completely resolved.

[cpernet]

Replying to fbissey:

OK, more thoughts about fflas-ffpack. By default usage of sse4.1, avx, avx2 and fma instructions are enabled if found. This is a problem when building a binary for redistribution as we cannot guarantee that the target machine will support any instructions found on the building machine. We need to be able to disable those for binary build (SAGE_FAT_BINARY=yes?).

Sure, we could do that. The alternative option would be to do as for the openblas deb package: have it include all kernel variants and let the switch happen dynamically depending on the variant of x86 it is run on. I am not very familiar with this technique and need to investigate how this can be built.

Going through linbox now. Like fflas-ffpack, openmp is enabled by default even without --enable-openmp in spkg-install but it looks like it can be disable the standard way to me.

Yes it can, but in linbox it will still pass a -fopenmp argument to g++ even if you passed the --disable-openmp option at configure time. You made me find it yesterday, and this is now ​https://github.com/linbox-team/linbox/issues/24.

linbox looks for

LB_CHECK_IML
LB_CHECK_M4RI
LB_CHECK_M4RIE
LB_CHECK_MPFR
# we nedd mpfr next :
LB_CHECK_FPLLL
LB_CHECK_FLINT

LB_CHECK_NTL

but none of these are in linbox's dependencies. Are any of those actually used?

These are optional dependencies. The long term goal is to have linbox code that uses these libs when available. For the moment, I think they are mostly used for benchmarking comparisons, except for NTL, IML and fplll used here and there when available.

Passing the --with-all= option to configure, as we do, will automatically enable all these optional dependencies if they are found. We could be more picky if this causes problems.

Actually the current version should depend on ntl and iml

(sage-sh) fbissey@QCD-nzi3:sage-7.0$ ldd -r local/lib/liblinboxsage.so.0.0.0 
	linux-vdso.so.1 (0x00007ffcaf131000)
	libgmpxx.so.8 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libgmpxx.so.8 (0x00007f3f89ea8000)
	libgmp.so.16 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libgmp.so.16 (0x00007f3f89c31000)
	libntl.so.19 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libntl.so.19 (0x00007f3f89810000)
	libcblas.so.3 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libcblas.so.3 (0x00007f3f895ed000)
	libatlas.so.3 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libatlas.so.3 (0x00007f3f88d01000)
	liblinbox.so.0 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/liblinbox.so.0 (0x00007f3f88afa000)
	libgivaro.so.0 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libgivaro.so.0 (0x00007f3f8889d000)
	libstdc++.so.6 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libstdc++.so.6 (0x00007f3f88557000)
	libm.so.6 => /lib64/libm.so.6 (0x00007f3f88253000)
	libc.so.6 => /lib64/libc.so.6 (0x00007f3f87eb8000)
	libgcc_s.so.1 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libgcc_s.so.1 (0x00007f3f87ca0000)
	libgf2x.so.1 => /home/fbissey/sandbox/git-fork/sage-7.0/local/lib/libgf2x.so.1 (0x00007f3f87a8a000)
	libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f3f8786d000)
	/lib64/ld-linux-x86-64.so.2 (0x0000561bc940b000)
undefined symbol: maxMagnMP	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: findRNS	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: RNSbound	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: certSolveMP	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: certSolveRedMP	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: nonsingSolvLlhsMM	(local/lib/liblinboxsage.so.0.0.0)
undefined symbol: nonsingSolvRNSMM	(local/lib/liblinboxsage.so.0.0.0)

compared to sage-on-gentoo on the same machine

ldd -r /usr/lib64/liblinboxsage.so.0.0.0
	linux-vdso.so.1 (0x00007ffe013bb000)
	libgmpxx.so.4 => /usr/lib64/libgmpxx.so.4 (0x00007f8a84d7a000)
	libgmp.so.10 => /usr/lib64/libgmp.so.10 (0x00007f8a84b09000)
	libntl-9.6.2.so => /usr/lib64/libntl-9.6.2.so (0x00007f8a846e5000)
	libreflapack.so => /usr/lib64/libreflapack.so (0x00007f8a83f03000)
	libopenblas_threads.so.0 => /usr/lib64/libopenblas_threads.so.0 (0x00007f8a8396b000)
	liblinbox.so.0 => /usr/lib64/liblinbox.so.0 (0x00007f8a83763000)
	libgivaro.so.0 => /usr/lib64/libgivaro.so.0 (0x00007f8a83507000)
	libiml.so.0 => /usr/lib64/libiml.so.0 (0x00007f8a832e6000)
	libstdc++.so.6 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libstdc++.so.6 (0x00007f8a82fd6000)
	libm.so.6 => /lib64/libm.so.6 (0x00007f8a82cd2000)
	libc.so.6 => /lib64/libc.so.6 (0x00007f8a82937000)
	libgcc_s.so.1 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libgcc_s.so.1 (0x00007f8a8271f000)
	libgf2x.so.1 => /usr/lib64/libgf2x.so.1 (0x00007f8a82509000)
	libgfortran.so.3 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libgfortran.so.3 (0x00007f8a821e0000)
	libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f8a81fc4000)
	/lib64/ld-linux-x86-64.so.2 (0x00005557ec4cf000)
	libquadmath.so.0 => /usr/lib/gcc/x86_64-pc-linux-gnu/4.9.3/libquadmath.so.0 (0x00007f8a81d85000)

Where I had to add LIBS="-liml" for things to be completely resolved.

This seems correct: the sage-linbox interface only uses these 2 libs when found. Are you saying that the -liml option is not passed to gcc even when iml was detected at configure time? This would be another bug then.

[fbissey]

Replying to cpernet:

This seems correct: the sage-linbox interface only uses these 2 libs when found. Are you saying that the -liml option is not passed to gcc even when iml was detected at configure time? This would be another bug then.

Yes, at least it is the case in 1.3.6, I haven't confirmed yet it is the case in 1.4.0, but if you weren't aware of it then it is probably still the case. I would rather have ntl and iml at least be in the dependencies as it makes the build reproducible. Otherwise the libraries linked to linboxsage will be random.

[jdemeyer]

On POWER8, I get an internal compiler error while compiling fflas-ffpack using GCC-4.9.3.

Clement, did you manage to compile fflas-ffpack on sardonis? Do you use the system-installed GCC 4.9.2 or did you compile within Sage?

[cpernet]

Replying to jdemeyer:

On POWER8, I get an internal compiler error while compiling fflas-ffpack using GCC-4.9.3.

Clement, did you manage to compile fflas-ffpack on sardonis? Do you use the system-installed GCC 4.9.2 or did you compile within Sage?

Outch! I did compile with the system-installed gcc v 4.9.2 and did not hit any internal compiler error.

[fbissey]

linbox again, LB_OPT is commented out in configure.ac therefore --enable-optimization which is still in spkg-install is no longer a recognised option and should be removed unless it is an error upstream.

[jdemeyer]

src/sage/rings/finite_rings/element_givaro.pxd refers to givaro/givgfq.h, which no longer exists.

[jdemeyer]

On powerpc64le with GCC 5.3.0, all 3 packages build and pass their testsuite.

[jdemeyer]

On my laptop (Linux tamiyo 3.17.7-gentoo #1 SMP PREEMPT Wed Dec 31 20:06:39 CET 2014 x86_64 Intel(R) Core(TM) i7-2640M CPU @ 2.80GHz GenuineIntel GNU/Linux), fflas_ffpack fails to build with GCC 4.8.4:

make[6]: Entering directory '/usr/local/src/sage-config/local/var/tmp/sage/build/fflas_ffpack-2.2.0/src/fflas-ffpack/interfaces/libs'
/bin/bash ../../../libtool  --tag=CXX   --mode=compile g++ -DHAVE_CONFIG_H -I. -I../../..  -I../../..  -I../../../fflas-ffpack/utils/ -I../../../fflas-ffpack/fflas/ -I../../../fflas-ffpack/ffpack -I../../../fflas-ffpack/field -I/usr/local/src/sage-config/local/include -I/usr/local/src/sage-config/local//include -D__FFLASFFPACK_HAVE_CBLAS  -msse4.1 -mavx -fabi-version=6 -fopenmp  -O2  -Wall -DNDEBUG -UFFLASFFPACK_DEBUG -g -O2 -fPIC  -std=gnu++11 -c -o fflas_L1_inst.lo fflas_L1_inst.C
libtool: compile:  g++ -DHAVE_CONFIG_H -I. -I../../.. -I../../.. -I../../../fflas-ffpack/utils/ -I../../../fflas-ffpack/fflas/ -I../../../fflas-ffpack/ffpack -I../../../fflas-ffpack/field -I/usr/local/src/sage-config/local/include -I/usr/local/src/sage-config/local//include -D__FFLASFFPACK_HAVE_CBLAS -msse4.1 -mavx -fabi-version=6 -fopenmp -O2 -Wall -DNDEBUG -UFFLASFFPACK_DEBUG -g -O2 -fPIC -std=gnu++11 -c fflas_L1_inst.C  -fPIC -DPIC -o .libs/fflas_L1_inst.o
In file included from ../../../fflas-ffpack/fflas/fflas_fger_mp.inl:40:0,
                 from ../../../fflas-ffpack/fflas/fflas.h:126,
                 from fflas_L1_inst.C:34:
../../../fflas-ffpack/fflas/fflas_fgemm/fgemm_classical_mp.inl:204:2: error: stray '#' in program
   WAIT;
  ^
../../../fflas-ffpack/fflas/fflas_fgemm/fgemm_classical_mp.inl: In function 'typename FFPACK::RNSInteger<T>::Element_ptr FFLAS::fgemm(const FFPACK::RNSInteger<T>&, FFLAS::FFLAS_TRANSPOSE, FFLAS::FFLAS_TRANSPOSE, size_t, size_t, size_t, typename FFPACK::RNSInteger<T>::Element, typename FFPACK::RNSInteger<T>::ConstElement_ptr, size_t, typename FFPACK::RNSInteger<T>::ConstElement_ptr, size_t, typename FFPACK::RNSInteger<T>::Element, typename FFPACK::RNSInteger<T>::Element_ptr, size_t, FFLAS::MMHelper<FFPACK::RNSInteger<T>, FFLAS::MMHelperAlgo::Classic, FFLAS::ModeCategories::DefaultTag, FFLAS::ParSeqHelper::Parallel<Cut, Param> >&)':
../../../fflas-ffpack/fflas/fflas_fgemm/fgemm_classical_mp.inl:204:8: error: expected ';' before string constant
   WAIT;
        ^
Makefile:576: recipe for target 'fflas_L1_inst.lo' failed
make[6]: *** [fflas_L1_inst.lo] Error 1

Full log attached.

[jdemeyer]

fflas_ffpack build failure on x86_64 GNU/linux GCC 4.8.4

[jdemeyer]

Some of the preprocessed code looks like

#pragma omp task shared(F)
 {fadd(F, rowsize, N, A+iter.begin()*lda, lda, B+iter.begin()*ldb, ldb, C+iter.begin()*ldc, ldc);;};;} } };};} #pragma omp taskwait
;;        }

so I guess you're missing some newlines in certain macros.

[jdemeyer]

Due to ​https://github.com/linbox-team/linbox/issues/24 this does not actually work...

---

New commits:

​6c253a0

Disable openmp in fflas_ffpack

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​5dc9f12

Disable openmp in linbox

[jdemeyer]

Note that this ticket conflicts badly with #20133.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​26fc91b	Cosmetic changes
​35c5190	Fix Givaro includes

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​b153b93

Put FFLASFFPACK_CFLAGS where the linbox headers are included

[jdemeyer]

I got this to build properly now, but there are still a lot of doctest failures in finite field arithmetic :-(

Are there any significant API changes in Givaro that we should know about?

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​402dd10

Reduce modulo characteristic before calling GFqDom::init()

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​186c233

If SAGE_FAT_BINARY is set, disable all processor-specific optimizations

[fbissey]

Replying to jdemeyer:

Note that this ticket conflicts badly with #20133.

Will have conflicts with #20140 if I add bits to the dependencies of linbox. As mentioned earlier it has automagic dependencies on iml, ntl and now fplll (to check) so you potentially get slightly different linbox depending on the build order. I am uncomfortable with that.

[jdemeyer]

I agree with your last comment.

[fbissey]

I spent quite some today working on fflas-ffpack-2.2.0 for sage-on-gentoo, expect a PR shortly to fix the tests when fflas-ffpack has been configured with --enable-precompilation (not that the test will work if you previously installed fflas-ffpack with --enable-precompilation in a standard location - do I need to say more).

lapack, and even probably ATLAS' clapack interface, is detected properly but you have to pass all the libraries for lapack and cblas in the --with-blas-libs switch. With our new .pc files that should translate to

--with-blas-cflags="$(pkg-config cblas --cflags)" --with-blas-libs="$(pkg-config lapack --libs)"

Although it would probably be safer to put

--with-blas-cflags="$(pkg-config cblas --cflags)" --with-blas-libs="$(pkg-config lapack  cblas blas --libs)"

in case lapack.pc doesn't pull either or both cblas and blas.

sage run spkg-check after install (a bit daft but hard to fix with the current system) so we should also be ok to add a spkg-check to fflas-ffpack without including my upcoming PR.

[fbissey]

Started on linbox. iml properly linked in this version if found. fplll detected but nowhere to be seen in the final library and no missing fplll symbols. On the other I have just been wacked by the detection of opencl that I wasn't expecting. No -lOpenCL are passed to liblinbox or liblinboxsage` when linking when it should be. Let's disable it explicitly for the time being.

[fbissey]

I have sent a request for some explanations to Clément for the linbox tarball, I have patched something and run autoreconf on it and the libraries I get out are completely different.

[jdemeyer]

Any news here? There have been various changes in how Sage uses pkgconfig, should we try again to get this working?

[fbissey]

Clement told me a couple of weeks ago he was on holidays. I also send a couple of PR for fflas-ffpack and linbox. He is supposed to cut new releases. I would prefer to wait for a new cut of linbox at least as patching the one in the description and running autoreconf on it lead to unexpected results (liblinbox and liblinboxsage were linked completely differently, WTF?).

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​0e5366f	Drop -std=c99 from C++ extensions
​02afed2	Merge remote-tracking branch 'trac/u/jdemeyer/drop__std_c99_from_c___extensions' into t/17635/givaro_fflasffpack_linbox

[jdemeyer]

Note that this merge is really more than just a merge: I also started using pkgconfig similarly to the way how we started using it in Sage 7.1.

[jdemeyer]

This branch builds fine and passes make ptestlong for me.

[cpernet]

Hi, Sorry for the delay. Just for the status report. I've produced fflas-ffpack-2.2.1rc0 and linbox-1.4.1rc0 (links available in the description of this ticket), where I tried to fix the remaining issues:

• fixing the disable-omp problem (issue ​https://github.com/linbox-team/linbox/issues/24)
• the missing line break in OMP macro (but it should not be a problem here, as we disable omp) (comment 67_68)
• removing the obsolete --enable-optimization of linbox (comment 64)

I still did not touch the opencl issue.

[jdemeyer]

Replying to cpernet:

links available in the description of this ticket

No, they are not.

---

New commits:

​608717f	remove obsolete --enable-optimization
​b6b60be	Merge branch 'u/jdemeyer/givaro_fflasffpack_linbox' of git://trac.sagemath.org/sage into t/17635/givaro_fflasffpack_linbox
​3389f1f	remove obsolete --disable-optimization
​ce91fb5	Update to fflas-ffpack-2.2.1rc0 and linbox-1.4.1rc0.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​94d4922	update to the checksum of the released givaro-4.0.1 tarball
​fcbcde4	update to linbox-1.4.1rc1 and fflas-ffpack-2.2.1rc1.
​7823566	update checksums

[cpernet]

I updated linbox and fflas-ffpack release candidates to address the several issues on dynamic lib linkage (including francois PR). I took this opportunity to create a linbox.pc file and update the module_list.py accordingly.

all libraries compile and pass their test suites on my side, and sage builds and passes the test suite with these new rc1's. I also tested the libs but not sage compilation on sardonis (power8).

[fbissey]

OK, so I have done my inspection:

1) list of linbox dependency that we want to be there: ntl, fplll, flint, mpfr, iml. Those are direct build dependencies

readelf -d liblinbox.so.0.0.0

Dynamic section at offset 0x44a80 contains 42 entries:
  Tag        Type                         Name/Value
 0x0000000000000001 (NEEDED)             Shared library: [libfflas.so.1]
 0x0000000000000001 (NEEDED)             Shared library: [libffpack.so.1]
 0x0000000000000001 (NEEDED)             Shared library: [libreflapack.so.3]
 0x0000000000000001 (NEEDED)             Shared library: [libopenblas_threads.so.0]
 0x0000000000000001 (NEEDED)             Shared library: [libgivaro.so.8]
 0x0000000000000001 (NEEDED)             Shared library: [libgmpxx.so.4]
 0x0000000000000001 (NEEDED)             Shared library: [libgmp.so.10]
 0x0000000000000001 (NEEDED)             Shared library: [libntl-9.7.0.so]
 0x0000000000000001 (NEEDED)             Shared library: [libfplll.so.0]
 0x0000000000000001 (NEEDED)             Shared library: [libflint.so.13]
 0x0000000000000001 (NEEDED)             Shared library: [libmpfr.so.4]
 0x0000000000000001 (NEEDED)             Shared library: [libiml.so.0]
 0x0000000000000001 (NEEDED)             Shared library: [libOpenCL.so.1]
 0x0000000000000001 (NEEDED)             Shared library: [libpthread.so.0]
 0x0000000000000001 (NEEDED)             Shared library: [libstdc++.so.6]
 0x0000000000000001 (NEEDED)             Shared library: [libm.so.6]
 0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]
 0x0000000000000001 (NEEDED)             Shared library: [libgcc_s.so.1]
 0x000000000000000e (SONAME)             Library soname: [liblinbox.so.0]

flint is not in the current list. m4ri and m4rie can be removed as they are looked for by configure but not used.

2) we want to have opencl (--without-ocl) to be turned off, at least for binary build. 3) because sage/libs/linbox/linbox.pyx has been given a block

# distutils: extra_compile_args = LINBOX_CFLAGS
# distutils: libraries = LINBOX_LIBRARIES
# distutils: library_dirs = LINBOX_LIBDIR

the entry in modules_list.py doesn't need libraries = linbox_libs. cblas_include_dirs can stay as its content is somehow not included in fflas-ffpack.pc (but it is in fflas-ffpack-config --cflags). Similarly fflas-ffpack.pc has no information on the blas/lapack libraries - may be it should - you would inherit the correct stuff without having to add anything. Someone check this on a vanilla install please.

4) Do we want lapack support in fflas-ffpack? That would mean changing LINBOX_BLAS="$(pkg-config --libs cblas)" to LINBOX_BLAS="$(pkg-config --libs lapack cblas)" this has an impact on the previous item as you really need to know which libraries are used (especially if your linker is dumb). Note: you may want at least LINBOX_BLAS="$(pkg-config --cflags --libs cblas)" in case the headers are in a funny location.

[fbissey]

@ Clement for future releases of fflas-ffpack and linbox. Now that you have committed to provide .pc files, you could insist on pkg-config being a building requirement. Once you do that, macros like FF_CHECK_GIVARO and LB_CHECK_FFLAS_FFPACK can die in peace and be replaced by

PKG_PROG_PKG_CONFIG

PKG_CHECK_MODULES(...)

leveraging pkg.m4 which is installed as part of ​http://pkgconfig.freedesktop.org's pkg-config. I can probably send you some PR for that in time.

[fbissey]

Acted upon my comment in (1) and (2). Added some stuff in fflas_ffpack's spkg-install to address my final note in section (4). I would rather have the rest of my comments be checked/discussed before they are acted upon.

---

New commits:

​942002d	Merge branch 'develop' into givaro_fflasffpack_linbox
​8cd62e2	Addressing issues in in my comments section (1), (2) and final note section (4)

[fbissey]

Is it just me or the givaro checksum is wrong?

[fbissey]

Somewhat tangential to this ticket, can someone explain to me why m4rie depends on givaro? I cannot find any reference to givaro in it, but because it is marked as depending on givaro it got rebuilt when I used make.

Of course if the dependency is bogus it shouldn't.

[cpernet]

Replying to fbissey:

Is it just me or the givaro checksum is wrong?

My bad. I'll revert it.

[cpernet]

Replying to fbissey:

Somewhat tangential to this ticket, can someone explain to me why m4rie depends on givaro? I cannot find any reference to givaro in it, but because it is marked as depending on givaro it got rebuilt when I used make.

Of course if the dependency is bogus it shouldn't.

It seems bogus to me. There is no reason for m4rie to depend on givaro and a grep confirms there is no mention of givaro in m4rie. The dependency can be removed.

[cpernet]

I'm fine with your patch on (1) and (2).

(4) Regarding the use of BLAS CFLAGS. The headers of a BLAS are standard both in C and Fortran. Hence fflas-ffpack has the headers for both C and Fortran BLAS (at least for the part of it that it uses). Hence, I did not feel the need to deal with include dirs (also in the .pc file). But ok, there is no harm in adding it and it'll probably prevent some future issues.

(3) Ok to remove these entries in module_list.py. Thanks for pointing out that fflas-ffpack.pc does not mention the BLAS flags. I'll fix this.

[fbissey]

Yes the API to cblas is "standard" that doesn't mean that the headers cannot be slightly different between implementations. You can indeed choose to include your own standard headers but if you include "system headers" for cblas you need to be more careful.

[jdemeyer]

Replying to fbissey:

Is it just me or the givaro checksum is wrong?

I remember it being wrong before too.

@cpernet: are you using a different version from the officially released one or are you changing released tarballs?

[jdemeyer]

The givaro tarball is missing ./configure. It looks like autogen.sh was not run...

[cpernet]

Replying to jdemeyer:

The givaro tarball is missing ./configure. It looks like autogen.sh was not run...

Indeed, I just realized that Github release tool (simply based on a tag) does not work nicely with autotools and make dist targets generating the configure file. I added the proper tar.gz generated by make dist as a binary and will fix the link on this ticket.