[SciPy-User] Modify package dependencies, or issue a runtime warning, if a fast BLAS implementation not found?

Hugh Perkins hughperkins@gmail....
Thu Sep 13 10:18:16 CDT 2012


Sounds good.  By the way, some additional information: I checked with
openblas why the multithreading doesn't seem very good, and apparently
the maximum number of threads is hard-coded at compile time to be
equal to the number of threads on the build machine.

Therefore, on a machine with more than one or two cores, you probably
want to build openblas from source to get the best performance.

it still runs four times slower than matlab in my tests (on a 12-core
machine), but that's better than the six times slower I get with stock
openblas (or 15-20 times slower or so with atlas/reference blas et
al).

On Thu, Sep 13, 2012 at 8:14 PM, Daπid <davidmenhur@gmail.com> wrote:
> Another option, as suggested recently elsewhere, would be to write a
> nice installation tutorial with both  the "quick and dirty" way
> (plainly install the distributed version) and the optimized way
> (linking to Blas, Blas goto and ATLAS). As far as  I know, there are
> no good simple explanations on how to build numpy with these links.


More information about the SciPy-User mailing list