Change mkl_info.dir_env_var from MKL to MKLROOT

[mingwandroid]

Reference: https://software.intel.com/en-us/node/528500

[charris]

Is this new and MKL was once valid, or has it always been MKLROOT?

[mingwandroid]

I don't know how to determine that as I can only find the latest documentation on the website.

[mingwandroid]

Actually, I found a reference for MKLROOT dating back to 2012:

https://software.intel.com/en-us/articles/intel-mkl-103-getting-started

[charris]

Looks like MKL 10.0 came out in 2008 and your reference is 10.3, so we are probably good. Thanks @mingwandroid .

[rgommers]

Does this actually make a difference? The mkl_info class already has a method get_mkl_rootdir that should look for MKLROOT and add it to the library and include paths.

[mingwandroid]

It does make a difference. The values from get_mkl_root get written into site.cfg (AFAIR, I am not at my computer to check this filename) only.

The case that failed for me was building scipy using a numpy built on a different system so site.cfg contained paths that don't exist.

[rgommers]

site.cfg doesn't get written, but __config__.py does get written and installed.

Scipy should not be using the paths in that file though. I can see how something would go wrong if get_info was imported from numpy.distutils instead of numpy.distutils.system_info - then you'd get the numpy build-time paths.

I have no problem with this PR, but it sounds like there's something else that needs fixing. The paths in __config__.py should never be reused for building another package.