DGESVD gives incorrect result

[vladimir-ch]

Our Gonum test for Dgesvd has recently started to fail when using OpenBLAS as the underlying implementation. I've bisected it to 0144068. As a data point, out of the various test cases that we have it seems to fail only for the matrix sizes 300x600 and 600x300.

[martin-frbg]

Oh great. :-(
Could you check if this can be repaired by restoring the two instances of &= 0 (lines 428 and 472) that supposedly were no better than the plain =0 that replaced them in the PR ?

[sandwichmaker]

@oon3m0oo

[vladimir-ch]

Reverting that commit fixes it but just restoring the &= 0 does not.

[martin-frbg]

Reverted now, thanks again for testing.

[oon3m0oo]

Interesting. I wonder if it was the movement of the top block of code. There didn't seem to be a dependency there, but I guess there was!

[ViralBShah]

Will there be a quick 0.3.2 with this hotfix?

[martin-frbg]

Depends on how quick you would want it to be - I am very busy this week, and I would also like to see at least #1673 resolved so that we do not end up with a new single-fix release every two weeks.

[ViralBShah]

Something in a couple of weeks would be great to line up with Julia 0.7. But otherwise we can always pick up the patches too.

[timokau]

I personally think there is nothing wrong with a patch release every two weeks. Better than having a lot of people re-discover the issue and spend time fixing it.

[martin-frbg]

Closing as 0.3.2 is released now.