Error in single threaded positive definite inverse in OpenBLAS

[andreasnoack]

As reported at the list

Issue filed: OpenMathLib/OpenBLAS#410

[ViralBShah]

We need to have this fix in 0.3, and it seems likely that the fix will be part of openblas 0.2.10.

Cc: @xianyi @wernsaar

[tkelman]

the linalg tests are already really long, but could any additional tests have caught the difference between single and multi threaded openblas here?

[andreasnoack]

I think, it only happened for large matrices, so we can just set n=1000 instead of n=10😃

[mlubin]

It seems like the linalg tests are becoming mostly a unit test suite for OpenBLAS.

[tkelman]

Yeah, largely. Their own tests are more comprehensive than ours, but Julia gets installed on a wider variety of machines than the OpenBLAS devs have access to.

[ViralBShah]

Maybe we should have a small set of tests that are run with different problem sizes and with different numbers of threads.

[tkelman]

OpenBLAS 0.2.10 is out with the fix for this, time for another bump?

[ViralBShah]

Could you test it out and bump it? I am just boarding a flight to India and will be mostly in an aluminium tube for the next 30 hours or so.

[tkelman]

Hm, judging by OpenMathLib/OpenBLAS#415 there may be trouble on 32 bit Linux?

[ViralBShah]

Looks like it has been fixed and there is a new 0.2.10 tag.

[staticfloat]

I'm currently testing out 0.2.10 on Ubuntu and OSX. I will submit a PR when I've finished.

[ViralBShah]

@staticfloat Did this work out?

[staticfloat]

Through my adventures, we've gotten a re-tagged 0.2.10 (along with a promise from the devs that they'll try to just release new versions instead of retagging). PR is here, but it would be nice if someone could verify that the issue reported in this thread is actually solved by version bump.

[ViralBShah]

I just merged it, thinking from the comments in the PR that it was good to go. Anyways, I am building right now and will speak up if it fails.

[tkelman]

I reproduced this problem in a pre-0.2.10 version and saw it fixed in the first tagged version of 0.2.10, which differed only in a single line from the second tag, so I think we're okay

[ViralBShah]

I don't see the PosDef exception any more.

[ViralBShah]

I wonder if we can print a line in the banner encouraging everyone who downloads julia to run the tests during the RC stage and report issues.

[tkelman]

Why just during the RC stage? Should put it on the download page. Julia is under heavy development, here's how to run the test suite first thing after installing a binary, if it doesn't pass then search github for similar issues and add a comment if you see something that's already been reported, or open a new issue if you can't find anything similar.

[ViralBShah]

That is a valid point, and I actually like the idea. The downloads page can have detailed instructions, and the banner can have something just as a reminder. Perhaps the request to run tests can be made quiet after the tests are run for that particular version.

[ivarne]

Maybe a not tested/tested/test failure notice in yellow/green/red?

[ViralBShah]

Great idea! Perhaps we should have a new issue for this.