[R] Installing OpenBLAS in R

[Prof Brian Ripley]

On 11/07/2013 19:26, Peter Langfelder wrote:
> [cc-ing the list]
>
> On Wed, Jul 10, 2013 at 11:36 PM, Christofer Bogaso
> <bogaso.christofer at gmail.com> wrote:
>> Hello again,
>>
>> I was wondering if it is possible to install OpenBLAS linear algebra
>> library in R running under windows.
>
> You will have to re-compile R from source and re-install it.

No, all you need to do on Windows is to replace Rblas.dll.  On Linux you 
most likely only need to link to libRblas.so.

However

- GotoBLAS contained alignment bugs (he called them bugs in R ...) that 
stopped 32-bit versions working (or working well). OpenBLAS is a 
descendant, and I have not heard this has been fixed.

- I suspect any Windows user who needs to ask does not know enough to do 
this.  In any case, the posting guide makes clear this is an R-devel topic.

> http://cran.r-project.org/doc/manuals/r-devel/R-admin.html
>
> and in particular
>
> http://cran.r-project.org/doc/manuals/r-devel/R-admin.html#BLAS
>
> For what it's worth, on my Core i5 processor running linux I found
> that ATLAS BLAS is somewhat faster than OpenBLAS.

Depends on the precise CPU and the task.  For SandyBridge Core i7's (and 
Xeons) MKL seems better than anything else (and IvyBridge ones are similar).

Now we commonly have systems with 16 cores (and 32 virtual CPUs) the 
game can be very different from a dual-core Core i5.  We do need to know 
what sort of system we are talking about.

>
> HTH,
>
> Peter



-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595