[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: science-optimisation (Was: About the sense of removing -march=native)



Hi,

On Thu, Feb 09, 2017 at 11:47:59AM +0100, Andreas Tille wrote:
> Hi Julien,
> 
> On Wed, Feb 08, 2017 at 06:15:37PM +0100, Julien Puydt wrote:
> > > I wonder whether we could invent some mechanism that is rebuilding a
> > > package in postinst and installs the result on the machine instead of a
> > > pre-build binary.  Or we could provide some toolset which enables
> > > scientists to download a set of source packages and build these after
> > > re-activating -march=native and move the results in a local repository
> > > which just needs to be added to sources.list.
> > > 
> > > Do you consider this as feasible ideas?
> > 
> > The src:atlas package does something like this :
> > - the Debian servers ship a generic package ;
> > - the source package has a special build target for a more optimized
> > package, which you can then install.
> 
> I've read README.Debian "Building Optimized Atlas Packages on your ARCH"
> and think this is pointing very much in the same direction of my second
> rough idea.  For packages that could profit from some optimisation for
> local machines we could provide a custom target in debian/rules.  While
> I did not yet made up my mind whether it is a good idea to build inside
> the local environment instead of a pbuilder chroot, this is a detail
> that could be discussed later.

I think I read the ATLAS README.Debian and found it a bit more
convoluted than necessary; in particular, I think it creates binary
packages with different names to ones in the official archive (sorry
can't check at the moment as I am on crappy dialup)? Probably just
bumping the binary package version like we do for backports or binNMUs
would be enough.

A good start would be to both support DEB_BUILD_OPTIONS=custom for
generic performance-oriented rebuidls and
DEB_CFLAGS_MAINT_APPEND=-march=native for for the -march=native thing.

Then again, just testing whether it actually makes a difference for
end-user applications would be useful as well.

Michael


Reply to: