Re: Bug#875618: openblas: please enable build on s390x
[CC’ing the debian-s390@lists.debian.org list; s390 folks, please keep the bug in CC on replies] Dear Graham, On Tue, Sep 12, 2017 at 07:28:04PM +0200, Graham Inggs wrote: > Source: openblas > Version: 0.2.20+ds-1 > Severity: wishlist > From Changelog.txt in OpenBLAS 0.2.20: > > IBM Z: * New target z13 with BLAS3 optimizations > > I have just checked, and openblas/0.2.20-3 builds successfully on > zelenka.debian.org. > Please enable building on s390x. Unfortunately it does not look that simple. OpenBLAS is optimized for z13, but our s390x port is supposed to support all the z systems (see [1]). In particular, the OpenBLAS build system adds the "-march=z13 -mzvector" compilation flags. If I remove them, then the package fails to build on zelenka: it complains about unknown assembly instructions, which are not present on old z-systems. This is the proof that OpenBLAS cannot produce a binary generic enough for our s390x port. So unless I am missing something, it’s not possible to enable building for s390x until 1) OpenBLAS supports older z-systems or 2) the hardware requirements for the Debian s390x port are upgraded. Best, [1] https://www.debian.org/releases/stable/s390x/ch02s01.html.en#idm45373715987328 -- ⢀⣴⠾⠻⢶⣦⠀ Sébastien Villemot ⣾⠁⢠⠒⠀⣿⡁ Debian Developer ⢿⡄⠘⠷⠚⠋⠀ http://sebastien.villemot.name ⠈⠳⣄ http://www.debian.org signature.asc Description: PGP signature
Re: Bug#875618: openblas: please enable build on s390x
On Wed, Sep 13, 2017 at 08:33:08PM +, PICCA Frederic-Emmanuel wrote: > Hello > > > Unfortunately it does not look that simple. OpenBLAS is optimized for z13, > > but > > our s390x port is supposed to support all the z systems (see [1]). > > what about asking for a a z13-support package to the isa-support (source > package) maintainer. This way it could be possible to upload an optimise > vesion of openblas which can install on recent enought s390x machines. I am not totally convinced by this solution. If we adopt it, somebody who installs e.g. octave on an old system-z machine will be hit by a failure in the dpkg installation process, which needs manual intervention. This is likely to generate problems in automated installers (and also confuse and annoy system admins). > the question will be then : does the buildd support these instructions ? I leave that to the s390 porters to answer. -- ⢀⣴⠾⠻⢶⣦⠀ Sébastien Villemot ⣾⠁⢠⠒⠀⣿⡁ Debian Developer ⢿⡄⠘⠷⠚⠋⠀ http://sebastien.villemot.name ⠈⠳⣄ http://www.debian.org signature.asc Description: PGP signature
Re: Bug#875618: openblas: please enable build on s390x
Control: tags -1 upstream Control: forwarded -1 https://github.com/xianyi/OpenBLAS/issues/1307 On Thu, Sep 14, 2017 at 12:34:34PM +0200, Christian Borntraeger wrote: > On 09/14/2017 09:58 AM, Sébastien Villemot wrote: > > On Wed, Sep 13, 2017 at 08:33:08PM +, PICCA Frederic-Emmanuel wrote: > >>> Unfortunately it does not look that simple. OpenBLAS is optimized for > >>> z13, but > >>> our s390x port is supposed to support all the z systems (see [1]). > >> > >> what about asking for a a z13-support package to the isa-support (source > >> package) maintainer. This way it could be possible to upload an optimise > >> vesion of openblas which can install on recent enought s390x machines. > > > > I am not totally convinced by this solution. If we adopt it, somebody who > > installs e.g. octave on an old system-z machine will be hit by a failure in > > the > > dpkg installation process, which needs manual intervention. This is likely > > to > > generate problems in automated installers (and also confuse and annoy system > > admins). > > > >> the question will be then : does the buildd support these instructions ? > > > > I leave that to the s390 porters to answer. > > FWIW, some years ago I did the atlas port for s390x. For dynamic linking the > atlas > build/package process did support the exploitation of ELF HW_CAPS. So you > could > build a z900 (generic) and a z13 variant which is then picked by the linker > at > runtime. No idea if openblas allows the same. Of course the static variant > (.a) > must be the generic one. Thanks for your feedback. I have opened a request upstream about the need for a z900 kernel, and for a dynamic selection between the z900 and z13 kernels (as OpenBLAS currently does on x86). -- ⢀⣴⠾⠻⢶⣦⠀ Sébastien Villemot ⣾⠁⢠⠒⠀⣿⡁ Debian Developer ⢿⡄⠘⠷⠚⠋⠀ http://sebastien.villemot.name ⠈⠳⣄ http://www.debian.org signature.asc Description: PGP signature
Re: Bug#875618: openblas: please enable build on s390x
Control: tags -1 = pending Control: notforwarded -1 On Mon, Sep 18, 2017 at 01:56:23PM +0200, Sébastien Villemot wrote: > On Thu, Sep 14, 2017 at 12:34:34PM +0200, Christian Borntraeger wrote: > > On 09/14/2017 09:58 AM, Sébastien Villemot wrote: > > > On Wed, Sep 13, 2017 at 08:33:08PM +, PICCA Frederic-Emmanuel wrote: > > >>> Unfortunately it does not look that simple. OpenBLAS is optimized for > > >>> z13, but > > >>> our s390x port is supposed to support all the z systems (see [1]). > > >> > > >> what about asking for a a z13-support package to the isa-support (source > > >> package) maintainer. This way it could be possible to upload an optimise > > >> vesion of openblas which can install on recent enought s390x machines. > > > > > > I am not totally convinced by this solution. If we adopt it, somebody who > > > installs e.g. octave on an old system-z machine will be hit by a failure > > > in the > > > dpkg installation process, which needs manual intervention. This is > > > likely to > > > generate problems in automated installers (and also confuse and annoy > > > system > > > admins). > > > > > >> the question will be then : does the buildd support these instructions ? > > > > > > I leave that to the s390 porters to answer. > > > > FWIW, some years ago I did the atlas port for s390x. For dynamic linking > > the atlas > > build/package process did support the exploitation of ELF HW_CAPS. So you > > could > > build a z900 (generic) and a z13 variant which is then picked by the linker > > at > > runtime. No idea if openblas allows the same. Of course the static variant > > (.a) > > must be the generic one. > > Thanks for your feedback. I have opened a request upstream about the need for > a > z900 kernel, and for a dynamic selection between the z900 and z13 kernels > (as OpenBLAS currently does on x86). It turns out that there is already a support for generic System z, I had overlooked that. I have therefore pushed a changeset that builds a generic s390x binary. However, upstream does not currently provide runtime detection, so owners of a z13 system will have to recompile locally (as explained in README.Debian) in order to get the best out of OpenBLAS (note that this is the same situation as all the non-x86 archs). -- ⢀⣴⠾⠻⢶⣦⠀ Sébastien Villemot ⣾⠁⢠⠒⠀⣿⡁ Debian Developer ⢿⡄⠘⠷⠚⠋⠀ http://sebastien.villemot.name ⠈⠳⣄ http://www.debian.org signature.asc Description: PGP signature
Re: pandas -- I think we should drop BE platforms
On Sat, Feb 24, 2018 at 02:47:19PM -0500, Yaroslav Halchenko wrote: > 1. remove all patches which disable tests on big endians (I would have > even patched to remove upstream skips on those archs, so we do not fall > into the trap of missing smth) I agree with that. > 2. Please correct me if I am wrong, but I do not think that there > is some flag to say smth like > > Architecture: any [!big-endian] > > so we would need to unlist them explicitly > > Architecture: any [!s390x, !powerpc, !mips] > > (or powerpc is not BE?) No, leave "Architecture: any". Then the builds will fail on those arches. This is what porters generally prefer, because they have a clear information about what is wrong with the package on their arch (and possibly they can act and send patches). -- ⢀⣴⠾⠻⢶⣦⠀ Sébastien Villemot ⣾⠁⢠⠒⠀⣿⡁ Debian Developer ⢿⡄⠘⠷⠚⠋⠀ http://sebastien.villemot.name ⠈⠳⣄ http://www.debian.org signature.asc Description: PGP signature