[R-sig-hpc] R with FlexiBLAS on Mac and Windows

[George Ostrouchov]

After some help from the FlexiBLAS team, it now builds on macOS. Via Inaki’s flexiblas package, it gives run-time switchable OpenBLAS and vecLib, and for OpenBLAS also run-time core use control. The following is the complete recipe

* Install Xcode and command line tools
* Install Homebrew: https://brew.sh/
* In a terminal window:
    * brew install cmake
    * brew install openblas
* cmake needs to be told about OpenBLAS: 
     * export CMAKE_PREFIX_PATH=/usr/local/opt/openblas:$CMAKE_PREFIX_PATH
* Install FlexiBLAS: https://www.mpi-magdeburg.mpg.de/projects/flexiblas
    * See the Install section in its README.md
* After installation, link to R (terminal window):
    * ln -sf /usr/local/lib/libflexiblas.dylib /Library/Frameworks/R.framework/Resources/lib/libRblas.dylib
* In R, install.packages("flexiblas") and test if it works:
    * flexiblas_avail()
    * flexiblas_list()

vecLib threads can be controlled with VECLIB_MAXIMUM_THREADS at load time but not at run-time via flexiblas yet.

George


> On May 30, 2022, at 10:32 AM, George Ostrouchov <georgeost using gmail.com> wrote:
> 
> Hi Simon, Hi Dirk,
> 
> Thanks for the pointers and checking on the Mac FlexiBLAS compile. I too have not been able to compile it on macOS yet. Let me give a bit more background.
> 
> I just finished teaching an R on clusters class on IT4I's Karolina, where I found FlexiBLAS as the R default installed with the EasyBuild foss toolchain: https://docs.easybuild.io/en/latest/Common-toolchains.html#foss-toolchain and I used Inaki’s package to manage cores. I’ll be teaching one in July on ORNL clusters, where we successfully made a spack recipe for the same. So I see this eventually becoming standard for R on clusters.
> 
> Having the same interface on a laptop would allow teaching the same thing on a laptop as on a cluster, making the transition much easier.  There are broader reasons (wrt python, Julia, and standardizing the whole data science toolchain) why going with FlexiBLAS is compelling.
> 
> Any pointers you may have on compiling FlexiBLAS on a Mac will be helpful. I will continue trying and if successful swapping it in as Simon suggests. A lot of people use Windows in my classes so I’m also looking for a Windows pathway.
> 
> Thanks,
> George
> 
> 
>> On May 30, 2022, at 1:13 AM, Simon Urbanek <simon.urbanek using R-project.org> wrote:
>> 
>> George,
>> 
>> the macOS R release uses shared BLAS, and provides two umbrellas: one with native BLAS and the other with Apple Accelerate - both of which can be switched at load time. Analogously, it should be in principle trivial to provide one for FlexiBLAS (since you don't even need to re-compile R if the ABI is GNU Fortran-compatible which according to the docs FlexiBLAS is supposed to be). However, based on a quick test it seems current FlexiBLAS doesn't build on macOS so there the users are probably better off using the supplied BLAS options unless someone provides a compatible binary.
>> 
>> Cheers,
>> Simon
>> 
>> 
>> 
>>> On 30/05/2022, at 8:02 AM, George Ostrouchov <georgeost using gmail.com> wrote:
>>> 
>>> Are there good recipes for building R with FlexiBLAS on a Mac and on Windows?
>>> 
>>> I have found that R with FlexiBLAS provides an easy way to dynamically switch BLAS backends and have dynamic control over core use from R.  This is now becoming more common on clusters, where I have first seen it on systems that use EasyBuild. We have also now done it with spack. I think this has a future, with plans to tackle GPU backend switching eventually.
>>> 
>>> Having the same control on laptops would provide a consistent pathway to core control from laptop to cluster. This would remove one more barrier in a move from a laptop to a cluster. I would like to use it in my tutorials that aim to take R users to clusters and supercomputers.
>>> 
>>> Thanks,
>>> George
>>> _______________________________________________
>>> R-sig-hpc mailing list
>>> R-sig-hpc using r-project.org
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-hpc
>>> 
>> 
>