|
|
Ublas : |
Subject: Re: [ublas] question about a binding of ublas
From: Rutger ter Borg (rutger_at_[hidden])
Date: 2010-11-02 17:35:30
On 2010-11-02 19:54, Kaveh Kohan wrote:
>
> I think the version of boost I am using is quite old (1.35.0) and it
> seems that there have been significant changes with respect to the
> current version (is it 1.6?). For example I came to realize and you also
> pointed out that atlas is not a binding anymore and it seems that its
> functionalities are provided via BLAS and LAPACK bindings as you said.
> Now my question is that do we need to link with libatlas.so to use the
> functionalities or libcblas.so in enough (in case one wants to use ATLAS).
Linking against licblas should be enough (if you are after the LAPACK
routines, you may need to link to an other library, depending on your
system). The added feature of the newer bindings is that you can also
try other BLAS backends besides ATLAS, such as those provided by Intel
and AMD.
> Here is my specific question:
> Please correct me if I am wrong but it seems that there is a significant
> performance difference between, for example, matrix-matrix
> multiplication if one uses ublas comparing to atlas. Here is an example
> I am using to compare them:
That's right -- ATLAS is significantly faster than uBLAS. Although you
may find that, e.g., dot products (and maybe more BLAS calls) are
executed faster by uBLAS.
Cheers,
Rutger