numpy/scipy and mkl

[Sean Farley]

Ryan Schmidt <ryandesign at macports.org> writes:

>> On Jun 11, 2016, at 10:11 AM, Gideon Simpson <gideon.simpson at gmail.com> wrote:
>> 
>> I have an intel mkl license, and I know that numpy/scipy can be built against it for a performance boost. I was wondering two things:
>> 
>> 1.  Is it possible that we could have a variant of numpy/scipy that supports mkl, or is that fundamentally at odds with macports, since the mkl is commercial?
>
> I wouldn't be opposed to adding that option. MacPorts already contains commercial software (like libxl, for which you must buy a license for certain uses, and oracle-instantclient, which connects to a commercial database and for which you must register and manually download the installer). I see there are a number of ways we might get a copy of mkl for free to help us create such a variant:
>
> https://software.intel.com/en-us/articles/free-mkl
>
> The question is whether all those packages contain the same or equivalent software, such that a variant created for use with one of them would work with the others as well.
>
> Another question is whether to code the variant to work with the files installed by the Intel MKL installer in their normal locations, or whether we should create a port for Intel MKL which would adjust the installation to place the files inside the MacPorts prefix.

This is quite challenging, actually. MKL would be a replacement for
ATLAS or OpenBLAS, essentially but is insane to link against. You
basically have to ask an Intel webapp for how to link with MKL:

https://software.intel.com/en-us/articles/intel-mkl-link-line-advisor

I could try to help if someone wants to go down this long rabbit-hole
but I'm quite busy this summer.