Hi Isuru, I've been trying to fix homebrew's openblas, and I'm running into an issue that I was hoping you might know something about. If I build openBLAS with TARGET=HASWELL then I don't have any issues with numpy. However, if I build it with TARGET=NEHALEM, which is what they want for homebrew, then I still run into the same numpy error. I see that Conda-forge uses TARGET=PRESCOTT, but if I try that I also run into the same numpy problem. Do you have any insight into why this might be the case?
Thanks, Zach On Monday, October 12, 2020 at 4:14:28 PM UTC-4 Zachary Scherr wrote: > Ah, thank you. I'll post the issue and see if I can get homebrew to > resolve it! > > On Monday, October 12, 2020 at 2:29:10 PM UTC-4 isu...@gmail.com wrote: > >> You don't need different TARGETs. `DYNAMIC_ARCH=1` means that it will >> detect at runtime what the CPU is and will use the kernel matching the CPU >> at runtime. `TARGET` gives the oldest CPU that this code will run on so >> that the common code is compiled to target this. >> >> Isuru >> >> On Mon, Oct 12, 2020 at 1:24 PM Zachary Scherr <zsc...@gmail.com> wrote: >> >>> I posted about it on their discussion page, but I'm not sure there will >>> be anything they can do. They build their libraries on several different >>> versions of MacOS, but I'm not sure they build with several different >>> TARGETS. I can try to pursue this, but the easy fix for now is that if >>> people want to install openblas via homebrew they should install from >>> source. >>> >>> On Monday, October 12, 2020 at 2:09:28 PM UTC-4 dim...@gmail.com wrote: >>> >>>> On Mon, Oct 12, 2020 at 6:17 PM Isuru Fernando <isu...@gmail.com> >>>> wrote: >>>> > >>>> > I looked at the openblas formula in homebrew and they are not passing >>>> the TARGET option. When using DYNAMIC_ARCH=1, a target should be >>>> specified. >>>> >>>> care to open an issue with them on >>>> https://github.com/Homebrew/homebrew-core ? >>>> >>>> > >>>> > Isuru >>>> > >>>> > On Mon, Oct 12, 2020 at 11:34 AM Zachary Scherr <zsc...@gmail.com> >>>> wrote: >>>> >> >>>> >> In case anyone is curious, I've tracked the problem down to >>>> homebrew's openBLAS and so it has nothing to do with sage. If I build >>>> numpy >>>> from source through homebrew then I get exactly the same error using my >>>> system python3. I'm wondering if maybe this has something to do with the >>>> fact that openBLAS is bottled with a different processor type than the one >>>> that is on my system. In any case I'll ask over there and see if I can get >>>> some help. >>>> >> >>>> >> On Sunday, October 11, 2020 at 10:17:43 PM UTC-4 Zachary Scherr >>>> wrote: >>>> >>> >>>> >>> Hi Dima, >>>> >>> >>>> >>> My system-wide numpy is installed via pip. I was just remarking >>>> that I didn't build my system-wide numpy from source so it's not a fair >>>> comparison. It's possible that the issues aren't sage related but are >>>> related to building sage's numpy against some homebrew libraries. >>>> >>> >>>> >>> On Sunday, October 11, 2020 at 8:11:13 PM UTC-4 dim...@gmail.com >>>> wrote: >>>> >>>> >>>> >>>> On Mon, Oct 12, 2020 at 1:00 AM Zachary Scherr <zsc...@gmail.com> >>>> wrote: >>>> >>>> > >>>> >>>> > That also crashes with: 86622 segmentation fault ./sage --python >>>> >>>> > >>>> >>>> > I tested this both with building sage against homebrew python >>>> 3.8.6 and against sage's default python 3.8.5 and the same error occurred >>>> in both cases. I don't have this problem using my system-wide numpy so >>>> it's >>>> possible that because sage numpy is built from source that maybe one of my >>>> homebrew libraries is causing issues. I found >>>> https://github.com/numpy/numpy/issues/11551 which could be relevant >>>> and might indicate that it's not a sage problem. >>>> >>>> >>>> >>>> If you are talking about Homebrew's numpy, it should be more or >>>> less >>>> >>>> identical with Sage's numpy (assuming you use Homebrew's openblas >>>> in >>>> >>>> Sage) >>>> >>>> cf. >>>> https://github.com/Homebrew/homebrew-core/blob/master/Formula/numpy.rb >>>> >>>> >>>> >>>> > >>>> >>>> > I should also add that if I lower the dimension of x to 100000 >>>> then it works without problem. >>>> >>>> > >>>> >>>> > On Sunday, October 11, 2020 at 3:26:11 PM UTC-4 dim...@gmail.com >>>> wrote: >>>> >>>> >> >>>> >>>> >> On Sun, Oct 11, 2020 at 8:14 PM Zachary Scherr < >>>> zsc...@gmail.com> wrote: >>>> >>>> >> > >>>> >>>> >> > Hi All, >>>> >>>> >> > >>>> >>>> >> > I was doing some bug hunting and noticed that scipy's pytests >>>> crashed pretty early. I tracked down exactly what was causing the problem >>>> and noticed it can be reproduced by the following sequence of commands: >>>> >>>> >> > >>>> >>>> >> > sage: import numpy as np >>>> >>>> >> > sage: x = np.random.randn(1000000,2) >>>> >>>> >> > sage: y = np.random.randn(2,2) >>>> >>>> >> > sage: x.dot(y) >>>> >>>> >> >>>> >>>> >> please test this in "plain" Sage python, i.e. does >>>> >>>> >> >>>> >>>> >> ./sage --python >>>> >>>> >> >>>> >>>> >> >>> import numpy as np >>>> >>>> >> >>> x = np.random.randn(1000000,2) >>>> >>>> >> >>> y = np.random.randn(2,2) >>>> >>>> >> >>> x.dot(y) >>>> >>>> >> >>>> >>>> >> reproduce the crash? >>>> >>>> >> >>>> >>>> >> >>>> >>>> >> >>>> >>>> >> > >>>> >>>> >> > produces: >>>> >>>> >> > >>>> >>>> >> > >>>> ------------------------------------------------------------------------ >>>> >>>> >> > 0 signals.cpython-38-darwin.so 0x00000001088bc62a >>>> print_backtrace + 58 >>>> >>>> >> > 1 signals.cpython-38-darwin.so 0x00000001088c0277 sigdie + >>>> 39 >>>> >>>> >> > 2 signals.cpython-38-darwin.so 0x00000001088c0212 >>>> sigdie_for_sig + 290 >>>> >>>> >> > 3 libsystem_platform.dylib 0x00007fff6c88e5fd _sigtramp + 29 >>>> >>>> >> > >>>> ------------------------------------------------------------------------ >>>> >>>> >> > Unhandled SIGBUS: A bus error occurred. >>>> >>>> >> > This probably occurred because a *compiled* module has a bug >>>> >>>> >> > in it and is not properly wrapped with sig_on(), sig_off(). >>>> >>>> >> > Python will now terminate. >>>> >>>> >> > >>>> ------------------------------------------------------------------------ >>>> >>>> >> > /Users/zscherr/sage/src/bin/sage-python: line 2: 78812 Bus >>>> error: 10 sage -python "$@" >>>> >>>> >> > >>>> >>>> >> > I haven't tested on any other platforms, but for what it's >>>> worth I have Catalina 10.15.7 with Xcode 12. >>>> >>>> >> > >>>> >>>> >> > Thanks, >>>> >>>> >> > Zach >>>> >>>> >> > >>>> >>>> >> > -- >>>> >>>> >> > You received this message because you are subscribed to the >>>> Google Groups "sage-devel" group. >>>> >>>> >> > To unsubscribe from this group and stop receiving emails from >>>> it, send an email to sage-devel+...@googlegroups.com. >>>> >>>> >> > To view this discussion on the web visit >>>> https://groups.google.com/d/msgid/sage-devel/28f58136-d3c7-40b7-9a46-36fc347acc2en%40googlegroups.com. >>>> >>>> >>>> >>>> > >>>> >>>> > -- >>>> >>>> > You received this message because you are subscribed to the >>>> Google Groups "sage-devel" group. >>>> >>>> > To unsubscribe from this group and stop receiving emails from >>>> it, send an email to sage-devel+...@googlegroups.com. >>>> >>>> > To view this discussion on the web visit >>>> https://groups.google.com/d/msgid/sage-devel/a67ca4b5-526b-40d7-a476-520d49890207n%40googlegroups.com. >>>> >>>> >>>> >> >>>> >> -- >>>> >> You received this message because you are subscribed to the Google >>>> Groups "sage-devel" group. >>>> >> To unsubscribe from this group and stop receiving emails from it, >>>> send an email to sage-devel+...@googlegroups.com. >>>> >> To view this discussion on the web visit >>>> https://groups.google.com/d/msgid/sage-devel/7ba7ef07-8460-4180-8670-397e1fb6df93n%40googlegroups.com. >>>> >>>> >>>> > >>>> > -- >>>> > You received this message because you are subscribed to the Google >>>> Groups "sage-devel" group. >>>> > To unsubscribe from this group and stop receiving emails from it, >>>> send an email to sage-devel+...@googlegroups.com. >>>> > To view this discussion on the web visit >>>> https://groups.google.com/d/msgid/sage-devel/CA%2B01voNTGybhpxULUXSuBxFFc2fAXkwyer%3DQ0pSAPpP3Am8uDQ%40mail.gmail.com. >>>> >>>> >>>> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "sage-devel" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to sage-devel+...@googlegroups.com. >>> >> To view this discussion on the web visit >>> https://groups.google.com/d/msgid/sage-devel/e0f2c485-cb29-4db3-85bc-3c2632835c90n%40googlegroups.com >>> >>> <https://groups.google.com/d/msgid/sage-devel/e0f2c485-cb29-4db3-85bc-3c2632835c90n%40googlegroups.com?utm_medium=email&utm_source=footer> >>> . >>> >> -- You received this message because you are subscribed to the Google Groups "sage-devel" group. To unsubscribe from this group and stop receiving emails from it, send an email to sage-devel+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/sage-devel/37a70675-95bb-419d-ab3c-4baa10eb632an%40googlegroups.com.
