Jack, Well, I'm back to Xcode 7, and the build works. It's a shame that atlas can't use the AVX instructions, because the FMA ops are supposed to provide pretty big performance wins for a lot of algorithms.
Thanks, Sean > On Sep 19, 2015, at 12:52, Sean Lake <odysseus9...@gmail.com> wrote: > > Jack, > > Thanks for your hard work. I had to downgrade to Xcode 6.4 because I need > matplotlib-py27 and the QT5 build is failing. > > Sean > >> On Sep 19, 2015, at 11:38, Jack Howarth <howarth.at.f...@gmail.com> wrote: >> >> Sean, >> Try current fink cvs for the 10.7 tree. I've added an >> atlas-xcode7.patch to the current atlas.info which suppresses the probing >> for AVX support so that the build stays at SSE3. This builds fine here on >> 10.10 against Xcode 7 on a MacPro 3,1. >> Jack >> >> On Fri, Sep 18, 2015 at 11:42 PM, Sean Lake <odysseus9...@gmail.com> wrote: >> Ok, trying again to send to fink-users, not fink-core. >> >> Sean Lake >> >>> On Sep 18, 2015, at 20:29, Jack Howarth <howarth.at.f...@gmail.com> wrote: >>> >>> >>> >>> On Fri, Sep 18, 2015 at 11:05 PM, Sean Lake <odysseus9...@gmail.com> wrote: >>> I've attached the log. Perhaps you can find the source of the error in >>> there. >>> >>> Thanks, >>> Sean Lake >>> >>> >>> Sean, >>> This is unlikely to be fixed anytime soon. The atlas build has always >>> be very fragile and the new clang assembler is now allowing it to explore >>> using the AVX opcodes. Previously on darwin the legacy GNU based assembler >>> was limited to SSE3 opcodes. Notice that the errors are starting here... >>> >>> /sw/src/fink.build/atlas-3.10.2-2/darwin_bld/../ATLAS//src/blas/level3/kernel/ATL_trsmKL_rk4.c:39:9: >>> warning: AVX vector return without AVX enabled changes the ABI [-Wpsabi] >>> rB00 = _mm256_broadcast_pd((void*)pB0); /* B10 B00 B10 B00 >>> */ >>> ^ >>> /sw/src/fink.build/atlas-3.10.2-2/darwin_bld/../ATLAS//src/blas/level3/kernel/ATL_trsmKR_rk4.c: >>> In function 'ATL_rk4': >>> /sw/src/fink.build/atlas-3.10.2-2/darwin_bld/../ATLAS//src/blas/level3/kernel/ATL_trsmKR_rk4.c:39:9: >>> warning: AVX vector return without AVX enabled changes the ABI [-Wpsabi] >>> rB00 = _mm256_broadcast_pd((void*)pB0); /* B10 B00 B10 B00 >>> */ >>> ^ >>> In file included from >>> /sw/lib/gcc5/lib/gcc/x86_64-apple-darwin14.5.0/5.2.0/include/immintrin.h:41:0, >>> from >>> /sw/src/fink.build/atlas-3.10.2-2/darwin_bld/../ATLAS//src/blas/level3/kernel/ATL_trsmKL_rk4.c:14: >>> /sw/lib/gcc5/lib/gcc/x86_64-apple-darwin14.5.0/5.2.0/include/avxintrin.h:834:1: >>> error: inlining failed in call to always_inline '_mm256_load_pd': target >>> specific option mismatch >>> _mm256_load_pd (double const *__P) >>> ^ >>> >>> Upstream will have to revise their in-line assembly code to be compatible >>> with the clang-based assembler or they will have to provide a mechanism to >>> override the selection of the AVX-opcodes. Note that the legacy GNU-based >>> assembler never allowed the optimizations to expand beyond... >>> >>> -DATL_ARCH_Corei1 -DATL_CPUMHZ=2260 -DATL_SSE3 -DATL_SSE2 -DATL_SSE1 >>> >>> but the new clang-based assembler is allowing totally new optimization >>> levels to be attempted. >>> >>> -DATL_ARCH_Corei3 -DATL_CPUMHZ=3500 -DATL_AVXMAC -DATL_AVX -DATL_SSE3 >>> -DATL_SSE2 -DATL_SSE1 >>> >>> Jack >>> >>>> On Sep 18, 2015, at 19:16, Jack Howarth <howarth.at.f...@gmail.com> wrote: >>>> >>>> >>>> >>>> On Fri, Sep 18, 2015 at 10:14 PM, Sean Lake <odysseus9...@gmail.com> wrote: >>>> Same error. >>>> >>>> Then I am out of ideas here. >>>> >>>> Thanks, >>>> Sean >>>> >>>>> On Sep 18, 2015, at 18:36, Jack Howarth <howarth.at.f...@gmail.com> wrote: >>>>> >>>>> >>>>> >>>>> On Fri, Sep 18, 2015 at 9:35 PM, Sean Lake <odysseus9...@gmail.com> wrote: >>>>> fink's make is installed. I also have cmake. >>>>> >>>>> i make 4.1-2 Software build tool >>>>> i cmake 3.3.0-1 Cross-platform Make >>>>> >>>>> Try it again without fink make installed. >>>>> >>>>> >>>>>> On Sep 18, 2015, at 18:33, Jack Howarth <howarth.at.f...@gmail.com> >>>>>> wrote: >>>>>> >>>>>> >>>>>> >>>>>> On Fri, Sep 18, 2015 at 9:17 PM, Sean Lake <odysseus9...@gmail.com> >>>>>> wrote: >>>>>> >>>>>>> On Sep 18, 2015, at 17:22, Jack Howarth <howarth.at.f...@gmail.com> >>>>>>> wrote: >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Fri, Sep 18, 2015 at 8:12 PM, Sean Lake <odysseus9...@gmail.com> >>>>>>> wrote: >>>>>>> Hello all, >>>>>>> >>>>>>> Not sure why the install is failing on this one. >>>>>>> >>>>>>> Sean >>>>>>> >>>>>>> Error text: >>>>>>> make[1]: Entering directory >>>>>>> '/sw/src/fink.build/atlas-3.10.2-2/darwin_bld' >>>>>>> cp /sw/src/fink.build/atlas-3.10.2-2/darwin_bld/lib/libatlas.a >>>>>>> /sw/src/fink.build/root-atlas-3.10.2-2/sw/lib/. >>>>>>> cp /sw/src/fink.build/atlas-3.10.2-2/darwin_bld/lib/libcblas.a >>>>>>> /sw/src/fink.build/root-atlas-3.10.2-2/sw/lib/. >>>>>>> cp /sw/src/fink.build/atlas-3.10.2-2/darwin_bld/lib/liblapack.a >>>>>>> /sw/src/fink.build/root-atlas-3.10.2-2/sw/lib/. >>>>>>> cp: /sw/src/fink.build/atlas-3.10.2-2/darwin_bld/lib/liblapack.a: No >>>>>>> such file or directory >>>>>>> Make.top:658: recipe for target 'install_lib' failed >>>>>>> make[1]: *** [install_lib] Error 1 >>>>>>> >>>>>>> >>>>>>> Was fink make installed during this build? >>>>>> >>>>>> I don't believe so. I ran the command: >>>>>> sudo fink install atlas atlas-shlibs >>>>>> >>>>>> I get the same error when I run: >>>>>> sudo fink rebuild atlas atlas-shlibs >>>>>> >>>>>> Does it behave differently if fink make is installed? >>>>>> >>>>>> >>>>>>> >>>>>>> System info: >>>>>>> Package manager version: 0.38.7 >>>>>>> Distribution version: selfupdate-rsync Fri Sep 18 16:34:04 2015, 10.10, >>>>>>> x86_64 >>>>>>> Trees: local/main stable/main stable/crypto >>>>>>> Xcode.app: 7.0 >>>>>>> Xcode command-line tools: 7.0.0.0.1.1441394355 >>>>>>> atlas-3.10.2-2 is set to build with only one job. >>>>> >>>>> >>>> >>>> >>> >>> >>> >> >> > ------------------------------------------------------------------------------ _______________________________________________ Fink-users mailing list Fink-users@lists.sourceforge.net List archive: http://news.gmane.org/gmane.os.macosx.fink.user Subscription management: https://lists.sourceforge.net/lists/listinfo/fink-users
