NumPy, SciPy, universal builds, and default fortran compiler for scientific ports

[vincent habchi]

Le 17 juin 2010 à 17:48, Joshua Root a écrit :

> I don't know why anyone would need a universal scipy or numpy. Do they
> have any dependents that only build for a particular arch? It might be
> best to just mark all the dependents as non-universal too.

Well, that's not a question especially tied to scipy or numpy. Question is: with the rapid obsolescence of ppc* and i386 machines, what's the point in compiling universal apps ? If the response is: compile on a fast machine, execute also on a slow one, then I think it's worth also for scipy/numpy.

Yet, I am skeptical, because Atlas configuration process use aggressive optimization based on processor discovery: if you compile, let's say, on a Core 2 duo machine, even in 32-bit mode, I bet the optimizer will embed SSE3/SSE4 instructions, that would not execute on old 32-bit processors. I wonder if someone has tried to execute a universal Atlas build on a i5 machine on a first generation MacBook with a Core Duo processor. If I am not mistaken, it might well crash.

> It's high time this happened. There has been no movement on these
> tickets, so maintainer timeout is going to be invoked.

So, what are we going to do practically?
Vincent