On Thu, Feb 16, 2017 at 10:47 PM, Matthias Bodenbinder
<[email protected]> wrote:
>
> Am 17.02.2017 um 06:57 schrieb Matthias Bodenbinder:
> > Am 16.02.2017 um 23:27 schrieb Tobias Ellinghaus:
> >
> >> Neither module was changed between those two releases. Did you compile both
> >> versions yourself and are you sure that the compile flags are identical?
> >>
> >>> Good job! Thank you!
> >>>
> >>> Matthias
> >>
> >
> > I compiled 2.2.3 myself and took 2.2.1 from debian testing .
> >
> > I will compile 2.2.1 myself and see.
>
> I checked it. This is indeed an issue with the self-compilation.
>
> If I pull the source from debian testing (apt-get source darktable) and
> compile it by myself the benchmark is fast (15 s) compared to 22 s when I use
> the debian testing binaries.
>
> This is somewhat shocking to me. Why is that?
Pre-compiled binaries that you get from repositories have to be compiled for the
lowest common denominator. As a result, they're generally compiled with the
-march=generic/-mtune=generic flags. If you're compiling it yourself, then you
can target the specific architecture of the machine you're on, which lets the
compiler make better optimization choices.
If you look at the cmake files for Darktable, the march and mtune settings have
been split into their own file:
https://raw.githubusercontent.com/darktable-org/darktable/master/cmake/march-mtune.cmake
And indeed, it looks to see whether or not you're building a binary package
(i.e. .deb or .rpm) to figure out what -march/-mtune setting to use.
> And could I eventually see similar performance improvements with other apps?
Maybe. It depends on the software, and the build configuration. Some software
might have multiple versions of the processor-intensive routines using different
levels of intrinsics (SSE2 vs. AVX2, etc), and use runtime dynamic dispatching
to decide which routine to use. Compiling your own won't help here. Other
software might be IO bound, for example, so the extra optimizations gained from
compiling with -march=native are overshadowed by other factors. Compiling your
own won't help here much, either. Many software packages don't take into
consideration whether they're being built for a inclusion in a repository, so
their build system doesn't make any effort at picking the correct -march/-mtune
setting the way Darktable does, leaving it to the compiler default
(-mtune=generic).
Compiling your own may only help here if you go to the trouble of figuring out
what it's doing, and forcing the -march/-mtune setting through command-line
flags -- and then hope that the software doesn't fall into one of the
earlier categories.
Cheers,
Ammon
___________________________________________________________________________
darktable developer mailing list
to unsubscribe send a mail to [email protected]
