Build atlas with the generic options recommended by the upstream
documentation for distributions. The expression now takes the parameter
'threads' which configures the number of threads atlas will use. The
default is to build serial atlas ('threads = "0"'). The expression also
takes the parameter 'cacheEdge' which is the L2 cache per core, in
bytes. This reduces build time because the cache size doesn't need to be
detected. It also reduces impurity, since different build nodes on Hydra
may have different hardware. It is set to 256k by default, which is
recommended for distributions by the upstream documentation.
We saw a crash in many computers, in the octave check phase, where octave crashed.
It was due to atlas being built for AMD Family 10h, which has a special SSE
trick that others computer don't have.
For x86_64, atlas is for K7. And for i686, PII.
svn path=/nixpkgs/trunk/; revision=33780
build machine because that feature offsets the performance timings. We ignore
that check, however, because with binaries being pre-built on Hydra those
timings aren't accurate for the local machine in the first place. The build log
might show something such as the following:
| It appears you have cpu throttling enabled, which makes timings
| unreliable and an ATLAS install nonsensical. Aborting.
| See ATLAS/INSTALL.txt for further information
| Ignoring CPU throttling by user override!
svn path=/nixpkgs/trunk/; revision=32506
