math/openblas + poudriere + manual building

Kurt Jaeger lists at
Sun May 29 11:24:17 UTC 2016


> I am using poudriere for the first time to test some patches to math/R. One
> new feature will be the ability to link against math/openblas. My portstree
> was updated yesterday (28/05).

> For my purposes right now I only want to build math/openblas (0.2.18,1) for
> my computer, so I uncheck the DYNAMIC_ARCH option, which I suppose will
> generate code for many architectures.

Use DYNAMIC_ARCH, otherwise the port is configured to not
build in poudriere.

As far as I understand, those BLAS related codes all try either
to build for multiple CPUs or try to build for a specific CPU.

So what should the poor port maintainer do ? Either she allows
to build anyway, which produces a optimized port for the CPU
of the builder host, or she says: We can't, build it manually.

It's a mess. From my point of view, it should always build with DYNAMIC_ARCH,
which is not always the fastest binary, but at least it builds 8-}

> * After searching a bit, I have not found how to *manually* build a package
> in poudriere. Does that mean: build the port in your ports tree outside
> poudriere?


> Should that message be reworded? Is there something missing in
> the documentation? Am I missing something?

It's a difficult topic, so it's difficult to reword. Do you
have suggestions ?

> * I see some people are manually editing the Makefile of math/atlas to tune
> to their machines [2]. My solution with math/openblas has been to finally
> enable DYNAMIC_ARCH, which probably compiles more code than I need to. I am
> fairly comfortable with the package optimizing for the build machine. Isn't
> there any easy way to force poudriere to go on?

You are asking for an easy way to override MANUAL_PACKAGE_BUILD.

> Setting NO_IGNORE in the
> environment or something like that (which, surprisingly, allows building of
> forbidden, but not ignored ports)?

It would be useful, yes. Probably submit a PR for poudriere to
add some flag like this.

pi at            +49 171 3101372                         4 years to go !

More information about the freebsd-ports mailing list