Note: This is a public test instance of Red Hat Bugzilla. The data contained within is a snapshot of the live data so any changes you make will not be reflected in the production Bugzilla. Email is disabled so feel free to test any aspect of the site that you want. File any problems you find or give feedback at bugzilla.redhat.com.
Bug 1080073
Comment on attachment 878111 [details]
atlas.fedora.ppc64le.tempo.patch tempo patch (stage2 only)
obsoleted my next attachments to follow.
Created attachment 898654 [details] atlas.spec-ppc64le.patch the one to update atlas.spec I am providing in successive attachments 5 patch files and a tarball that allow atlas to be built for ppc64le archi on Power7. (now that gcc 4.9 support vsx/altivec on ppc64le) I will have to provide later on other patches for Power8 when I will have access to one. But at least those one are sufficient for Power7. Refer to comments in patch headers for details. atlas.spec-ppc64le.patch the one to update atlas.spec using those below xlf.command.not.found.patch getdoublearr.stripwhite.patch initialize_malloc_memory.invtrsm.wms.oct23.patch atlas.ppc64le-abiv2.patch POWER764LEVSX.tar.bz2 tarball used as new Source15 in atlas.spec Created attachment 898655 [details]
xlf.command.not.found.patch
Created attachment 898656 [details]
getdoublearr.stripwhite.patch
Created attachment 898657 [details]
initialize_malloc_memory.invtrsm.wms.oct23.patch
Created attachment 898658 [details]
atlas.ppc64le-abiv2.patch
Created attachment 898659 [details] POWER764LEVSX.tar.bz2 tarball set of patches validated on ppcle koji scratchbuild: http://ppc-le.koji.fedoraproject.org/koji/taskinfo?taskID=65415 related patches submitted upstream in atlas-devel ML on June 17th but no answer yet: http://sourceforge.net/p/math-atlas/mailman/message/32471499/ Comment on attachment 898655 [details] xlf.command.not.found.patch after discussion upstream via https://sourceforge.net/p/math-atlas/patches/62/ mark this patch as obsolete. Created attachment 917277 [details] issue_64.patch after discussion via https://sourceforge.net/p/math-atlas/patches/64/ replace the initialize_malloc_memory.invtrsm.wms.oct23.patch by this new issue_64.patch Created attachment 917280 [details] atlas.ppc64le-abiv2.patch after discussion via https://sourceforge.net/p/math-atlas/patches/65/ update the atlas.ppc64le-abiv2.patch Comment on attachment 898656 [details] getdoublearr.stripwhite.patch I still have to work on it based on https://sourceforge.net/p/math-atlas/patches/63/ Created attachment 917284 [details]
atlas.spec-ppc64le.patch
update the atlas.spec-ppc64le.patch to force archdef=0 and use only last specified patches.
With above patches I am able to build/check atlas package for ppc64le archi on POWER7 (and generate rpms) there are still segfaults reported in console but not identified in build.log. last pending scratch build: http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=1941601 Created attachment 919654 [details] atlas-new_archdef_for_ppc64le.patch I forgot to save as attachment the atlas-new_archdef_for_ppc64le.patch already used in previous scratch build of comment 15. This patch is submitted upstream via https://sourceforge.net/p/math-atlas/patches/66/ This package has changed ownership in the Fedora Package Database. Reassigning to the new owner of this component. This package has changed ownership in the Fedora Package Database. Reassigning to the new owner of this component. This package has changed ownership in the Fedora Package Database. Reassigning to the new owner of this component. A subset of above patches are embeded in new atlas 3.10.2 that is sufficient to have atlas build to succeed on ppc64le archi as per scratch build http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=1965167 So close this bug as duplicate of rhbz #1118596 *** This bug has been marked as a duplicate of bug 1118596 *** (In reply to Michel Normand from comment #20) > A subset of above patches are embeded in new atlas 3.10.2 that is sufficient > to have atlas build to succeed on ppc64le archi as per scratch build > http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=1965167 > > So close this bug as duplicate of rhbz #1118596 > > *** This bug has been marked as a duplicate of bug 1118596 *** After discussion I will keep this bug open to track the atlas-new_archdef_for_ppc64le.patch (issue 66 still open upstream) that is still missing in atlas-3.10.2 version. (In reply to Michel Normand from comment #21) > (In reply to Michel Normand from comment #20) > > A subset of above patches are embeded in new atlas 3.10.2 that is sufficient > > to have atlas build to succeed on ppc64le archi as per scratch build > > http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=1965167 > > > > So close this bug as duplicate of rhbz #1118596 > > > > *** This bug has been marked as a duplicate of bug 1118596 *** > > After discussion I will keep this bug open to track the > atlas-new_archdef_for_ppc64le.patch (issue 66 still open upstream) that is > still missing in atlas-3.10.2 version. current_status: * with atlas 3.10.2 able to build one time on koji for fedora 21 (Id 254085) but still have failing tests depending of builder (Id 1965860) http://ppc.koji.fedoraproject.org/koji/buildinfo?buildID=1965860 http://ppc.koji.fedoraproject.org/koji/buildinfo?buildID=254085 * I have a set of additionnal patches for 3.10.1 or 3.10.2 that allow to build, and ignore the make check errors for ppc64le. as reported in specified scratch builds. for atlas 3.10.1: atlas.3.10.1_spec_with_ppc64le_patches.patch issue_64.patch (in 3.10.2 upstream) atlas.3.10.1-ppc64le_abiv2_step1.patch (in 3.10.2 upstream) atlas.3.10.2-ppc64le_abiv2_step2.patch atlas.3.10.2-ppc64le_abiv2_step3.patch http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=2001446 for atlas 3.10.2: atlas.3.10.2-ppc64le_abiv2_step2.patch atlas.3.10.2-ppc64le_abiv2_step3.patch * The step2 step3 patches are already submitted upstream at https://sourceforge.net/p/math-atlas/patches/65 * for make check error, still need more work related to vector assembly code that assumes big-endian env (eg: ATL_cmm4x4x128_av.c and ATL_smm4x4x128_av.c) Would need significant work to support little-endian as per https://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/FBFA164F824370F987256D6A006F424D/$file/vector_simd_pem.ppc.2005AUG23.pdf Comment on attachment 898656 [details]
getdoublearr.stripwhite.patch
not mandatory patch after discussion upstream, so change to obsolete.
Comment on attachment 898659 [details] POWER764LEVSX.tar.bz2 tarball not mandatory patch after discussion upstream, so change to obsolete. Comment on attachment 919654 [details]
atlas-new_archdef_for_ppc64le.patch
not mandatory patch after discussion upstream, so change to obsolete.
Created attachment 927912 [details]
atlas.3.10.1_spec_with_ppc64le_patches.patch
Created attachment 927913 [details]
atlas.3.10.1-ppc64le_abiv2_step1.patch
Created attachment 927915 [details]
atlas.3.10.2-ppc64le_abiv2_step2.patch
Created attachment 927916 [details]
atlas.3.10.2-ppc64le_abiv2_step3.patch
> * for make check error, still need more work
> related to vector assembly code that assumes big-endian env
> (eg: ATL_cmm4x4x128_av.c and ATL_smm4x4x128_av.c)
> Would need significant work to support little-endian as per
>
> https://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/
> FBFA164F824370F987256D6A006F424D/$file/vector_simd_pem.ppc.2005AUG23.pdf
Michel, could we somehow disable those 2 files and use a generic implementation instead until they are updated for LE?
BTW are you working on updating the config mechanisms for Power8? I guess it would be an unknown CPU with new problems ... (In reply to Dan Horák from comment #30) > > * for make check error, still need more work > > related to vector assembly code that assumes big-endian env > > (eg: ATL_cmm4x4x128_av.c and ATL_smm4x4x128_av.c) > > Would need significant work to support little-endian as per > > > > https://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/ > > FBFA164F824370F987256D6A006F424D/$file/vector_simd_pem.ppc.2005AUG23.pdf > > Michel, could we somehow disable those 2 files and use a generic > implementation instead until they are updated for LE? A patch disabling those 2 files is not sufficient to avoid the make check error. For the time beeing I only have the solution brought by already attached patches that ignore make check error for ppc64le. (In reply to Dan Horák from comment #31) > BTW are you working on updating the config mechanisms for Power8? I guess it > would be an unknown CPU with new problems ... Yes I know, but I do not have yet easy access to a P8 machine to test new related patches. Created attachment 939224 [details]
atlas.3.10.1-add_power8_cpu.patch
Created attachment 939225 [details]
atlas.3.10.2-add_power8_cpu.patch
The two previous patches atlas.3.10.1-add_power8_cpu.patch and atlas.3.10.2-add_power8_cpu.patch are adding the support of Power8 on atlas for the two release atlas.3.10.1 and atlas.3.10.2. I am just on the way to build with them on a Power 8 machine. Created attachment 939357 [details]
atlas.3.10.2-add_power8_cpu.patch
correction of a small typo in previous atlas.3.10.2-add_power8_cpu.patch
Created attachment 940404 [details]
atlas.3.10.1_spec_add_power8_archdef.patch
atlas.3.10.1_spec_add_power8_archdef.patch as attachment as suggested patch of atlas.spec to speedup build on Power8. I previously verified that the archdef for Power8 has same content as for Power7.
Created attachment 942268 [details]
atlas.3.10.1_spec_with_ppc64le_patches.patch
new status versus previous one in comment #22 current_status: * have a set of patches for 3.10.1 or 3.10.2 atlas releases that allow to build atlas for power7 for ppc64 and ppc64le archi with no make check error, as reported in specified scratch builds. * I verified with mock build that work also on power8 hdw. for atlas 3.10.1: atlas.3.10.1_spec_with_ppc64le_patches.patch issue_64.patch (in 3.10.2 upstream) atlas.3.10.1-ppc64le_abiv2_step1.patch (in 3.10.2 upstream) atlas.3.10.2-ppc64le_abiv2_step2.patch atlas.3.10.2-ppc64le_abiv2_step3.patch atlas.3.10.2-ppc64le_do_not_use_files_with_lvx.patch atlas-new_archdef_for_ppc64le.patch atlas.3.10.1-add_power8_cpu.patch http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=2123725 for atlas 3.10.2: atlas.3.10.2-ppc64le_abiv2_step2.patch atlas.3.10.2-ppc64le_abiv2_step3.patch atlas.3.10.2-ppc64le_do_not_use_files_with_lvx.patch atlas-new_archdef_for_ppc64le.patch atlas.3.10.2-add_power8_cpu.patch http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=2123734 * The step2 step3 patches are already submitted upstream but not yet pushed completely in 3.11.30 https://sourceforge.net/p/math-atlas/patches/65 * The atlas-new_archdef_for_ppc64le.patch already upstream in 3.11.30 https://sourceforge.net/p/math-atlas/patches/66 * the atlas.3.10.1-add_power8_cpu.patch submitted upstream at https://sourceforge.net/p/math-atlas/patches/67/ * The atlas.3.10.2-ppc64le_do_not_use_files_with_lvx.patch is a bypass removing some PowerPC files related to vector assembly code that assumes big-endian env (eg: ATL_cmm4x4x128_av.c and ATL_smm4x4x128_av.c) Would need significant work to support little-endian as per https://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/FBFA164F824370F987256D6A006F424D/$file/vector_simd_pem.ppc.2005AUG23.pdf This is not yet solved upstream and tracked by upstream issue 65 (same as above) * The above patches are ready to be pushed in fedora db for atlas 3.10.1 (atlas 3.10.2 still failing arm as tracked in https://bugzilla.redhat.com/show_bug.cgi?id=1118596#c6) Comment on attachment 940404 [details]
atlas.3.10.1_spec_add_power8_archdef.patch
atlas.3.10.1_spec_add_power8_archdef.patch removed as obsoleted by atlas.3.10.1_spec_with_ppc64le_patches.patch
Created attachment 942576 [details] atlas.3.10.2-ppc64le_do_not_use_files_with_lvx.patch atlas.3.10.2-ppc64le_do_not_use_files_with_lvx.patch referenced in comment #40 (same for 3.10.1 and 3.10.2) Created attachment 942578 [details] atlas-new_archdef_for_ppc64le.patch atlas-new_archdef_for_ppc64le.patch referenced in comment #40 (same for 3.10.1 and 3.10.2) Created attachment 943255 [details] POWER764LEVSX.tar.bz2 POWER764LEVSX.tar.bz2 is the archdef file referenced in source15 of atlas.3.10.1_spec_with_ppc64le_patches.patch. Its presence allow to speed up the build process. This bug appears to have been reported against 'rawhide' during the Fedora 22 development cycle. Changing version to '22'. More information and reason for this action is here: https://fedoraproject.org/wiki/Fedora_Program_Management/HouseKeeping/Fedora22 Is this bug still valid? How is it different from [1]? [1] https://bugzilla.redhat.com/show_bug.cgi?id=1111650 (In reply to Rafael Fonseca from comment #46) > Is this bug still valid? How is it different from [1]? Yes this bug is still valid and related patches detailed in comment #40 were never accepted by fedora maintainer Those patches related to 3.10.1/3.10.2 versions are supposed to be embedded upstream for version 3.11.35 (not stable) as per issues status (2) But I did not took the time to tried it. The patch of comment #42 is not upstream If that could help I recently applied 3.10.2 patches in opensuse tumbleweed at (3) (2) https://sourceforge.net/p/math-atlas/patches/65/ https://sourceforge.net/p/math-atlas/patches/66/ https://sourceforge.net/p/math-atlas/patches/67/ (3) https://build.opensuse.org/package/show/openSUSE:Factory:PowerPC/libatlas3 > > [1] https://bugzilla.redhat.com/show_bug.cgi?id=1111650 This bug [1] is addressing a specific error reporting during build but that is not fatal to the build process. it's now fixed in rawhide. thanks to Michel for the fixes! |
Created attachment 878111 [details] atlas.fedora.ppc64le.tempo.patch tempo patch (stage2 only) Description of problem: atlas do not build on ppc64le archi Version-Release number of selected component (if applicable): atlas-3.10.1-12.fc21 Steps to Reproduce: 1. fedpkg clone -a atlas 2. fedpkg prep 3. fedpkg mockbuild Additional info: In attachement there is the temporary patch allowing to build on ppc64le archi. Warning: need rework because has bypasses identified during stage2 build.