Re: [PATCH 2/2] gnu: zita-convolver: Drop the '-march=native' optimization.

[iyzsong@member.fsf.org (宋文武)]

ludo@gnu.org (Ludovic Courtès) writes:

> Ricardo Wurmus <rekado@elephly.net> skribis:
>
>> 宋文武 <iyzsong@member.fsf.org> writes:
>>
>>> Ricardo Wurmus <rekado@elephly.net> writes:
>>>
>>>>> +              (snippet
>>>>> +               ;; Don't optimize for a specific processor architecture.
>>>>> +               '(substitute* "libs/Makefile"
>>>>> +                  (("^CXXFLAGS \\+= -march=native") "")))
>>>>> +              (modules '((guix build utils)))
>>>>
>>>> Is this to avoid that packages are optimised for the CPU of the build
>>>> slave?
>>> Yes, using that from hydra will crash guitarix for me.
>>> Same as: <https://bugs.archlinux.org/task/36248>.
>>
>> Guitarix also crashes for me on i686 — and I thought it was something to
>> do with my machine or the broken build of eigen.
Pushed, hopefully this will fix it on i686 too.
>
> Eigen is a header-only library, so what matters is the optimization
> flags passed when building the software that uses Eigen.
>
>>>> If so, could we instead pass different optimisation flags for
>>>> different architectures?  For a convolver I’d prefer to have *some*
>>>> optimisation, even if it’s just the greatest common divisor.
>
> It’s likely that zita-convolver has -O2 or similar *in addition* to
> -march=native, because the latter is no substitute for the former.
>
> Could you check that in the makefiles?
Yes, it actually has:
  CXXFLAGS =+ -ffast-math -funroll-loops -O3