carry upstream patch for using specific kernels via OPENBLAS_CORETYPE

[tkelman]

This is cherry-picked from OpenMathLib/OpenBLAS#386
and provides at least a workaround for Haswell and AMD problems seen in #6504, JuliaLang/LinearAlgebra.jl#117

[ViralBShah]

Ideally, it would be great if @xianyi can make a new release. If not, and this is the only thing that is blocking 0.3, we can include the patch.

[ViralBShah]

BTW, we should also include the tests that exercise both those issues, either as part of this patch or bumping openblas to a new release. At least regressions will be caught on the respective platforms.

[ViralBShah]

Cc: @andreasnoackjensen

[tkelman]

Yeah we are starting to pretty heavily patch things up here.

The tests are processor-dependent, which is part of what makes them so hard to debug. The users with the specific types of AMD or Haswell processors haven't been responsive enough to make debugging this go any faster. I'll hopefully be able to get my hands on a Windows Haswell laptop starting next week.

[tkelman]

I think Andreas is somewhere with poor connectivity lately.

I'm making an effort to clean up the last known failures in the test suite on Windows, my goal is that everyone who installs a Julia binary should immediately run the test suite (there's a test-julia.bat for making this easier on Windows, but I guess julia runtests.jl all isn't too hard).

[ViralBShah]

That would greatly help! It would be great if the AppVeyor build results could be featured on every commit and PR the way we do for Travis.

[tkelman]

Absolutely agree. Status there is Julia codegen got slower and AppVeyor switched to smaller VM's, so the tests can no longer finish in half an hour. There might also be a Win64-specific memory leak somewhere, see #6881

Also more-minor concerns of Travis and AppVeyor overwrite each other on Github PR status indicators at the moment, hopefully Github will eventually fix their API so there's a better solution for that. Maybe eventually Travis could support Windows themselves, that might be further off.

[ViralBShah]

@Keno had submitted Windows support patches to Travis a long time back, IIRC. Travis also has mac, but you can use only mac or linux, the last I looked.

[Keno]

You can use both now.

[staticfloat]

Yep, I had some preliminary work on getting OSX travis builds running, but that is also waiting on our buildfarm since it requires bottling all of the dependencies so we're not building LLVM and OpenBLAS every time we want to test a julia build. ;)

[staticfloat]

Yet another thing on my todo list this summer!

[tkelman]

@staticfloat can't you install from the binaries? just gotta figure out the command line to extract and install from a dmg, right? err nevermind that works for packages but not base julia

[staticfloat]

@tkelman right. I actually got that working for Nettle a while back, which is pretty neat (Note that the failure for the 0.2.1 binary is because that commit is 0.3+ code, and unfortunately I haven't thought through a good way to handle that situation properly. Perhaps just giving up gracefully if the Julia version doesn't conform to what the package wants....).

Eventually, I'd like to have a standard Julia installation and package setup script that just gets run by every package under the sun. :)

[tkelman]

Yes, that would be a good idea. I've been turning on AppVeyor on a decent number of packages already. Want to be unnecessarily clever and make the script a cmd/sh polyglot?

[staticfloat]

Want to be unnecessarily clever and make the script a cmd/sh polyglot?

That sounds horrifying. You've got my attention. :) I have no idea how you'd do that, since sh and cmd syntax is pretty different, but I'm willing to try!

My scripts are pretty easy; I've got one for linux, and one for OSX, but we can trivially combine the two in an if statement based on $TRAVIS_OS_NAME

[staticfloat]

Also, we're pretty far off-topic. Let's move the discussion over here.

[tkelman]

closing in favor of #7459