[R-SIG-Mac] Building R-2.4 on MacOS X

[Simon Urbanek]

On Dec 4, 2006, at 5:21 PM, Bill Northcott wrote:

> On 05/12/2006, at 2:43 AM, Simon Urbanek wrote:
>>> 1.  The CRAN binary contains a library libBlasVeclib.dylib.
>>> Following the build method in 'The MacOS X R FAQ' this library is  
>>> not built.  What is the library for and would it be worth building?
>>>
>>
>> R can be configured to use various BLAS implementations. One of  
>> changes in 2.4.0 is that this can be done by switching  
>> libRblas.dylib (e.g. via softlink). The CRAN binary comes with two  
>> BLAS implementations - one using vecLib (libRblas.vecLib.dylib)  
>> and the internal R one (libRblas.0.dylib). They have been produced  
>> in two separate builds. This is equivalent to configuring R with -- 
>> with-blas='-framework vecLib' and --without-blas respectively. The  
>> libRblas.dylib is a symlink which can be changed depending on  
>> which implementation you want to use.
>
> Do you have any numbers about which is currently the better  
> performer on the various MacOS paltforms?
>

vecLib is often faster, especially on multi-core or multi-CPU  
machines (both Intel and PPC) and for larger problems. However, the  
performance is task-specific. For small data sizes Doug Bates had  
some examples that were much slower with vecLib - that's why I have  
opted to ship CRAN binary with vecLib by default, but left it to the  
user to switch to internal BLAS if desired. (The performance  
degradation has to do with malloc - there is a small range of  
allocation sizes that doesn't perform well on OS X and vecLib gets  
hit by that if you happen to fall into that range. This has been  
discussed on the list some time ago.).

Cheers,
Simon