atlas depends on gcc43 *and* gcc44 ???

[vincent habchi]

Le 2 juin 2010 à 23:13, Stephen Langer a écrit :

> I was wondering why anything depends on atlas at all.  Is atlas noticeably better than the lapack and blas routines in the Accelerate framework?  I couldn't find any comparisons on-line.

I am unsure. But, as long as Apple does not state that its Blas and Lapack libraries are OpenCL based, in which case they may be hundred times quicker that CPU-thread-based Atlas, I am lead to believe that Atlas compiled with the latest gcc45 should be more efficient than Apple blas or lapack, most probably generated with gcc42 (the last GPL2 version) and backward-compatible with MacOS 10.4 or 10.5; Maybe they can get better efficiency with Clang/LLVM. However, I faintly remember comparing the Atlas built-in tests, and there was a clear gap between gcc43 and gcc44. 

It should not be very difficult to build up a test, like finding the eigenvalues of a random but symmetric definite positive matrix of great size (say : 50,000 x 50,000 or more) in Xcode, and link either with the built-in framework or our Atlas.

Cheers
Vincent

PS : Does Apple provide scipy in its Python distribution?